blob: 08fcaebc61bdb52b92f8ba37ccd4d2b3dac41779 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 md.c : Multiple Devices driver for Linux
NeilBrownf72ffdd2014-09-30 14:23:59 +10003 Copyright (C) 1998, 1999, 2000 Ingo Molnar
Linus Torvalds1da177e2005-04-16 15:20:36 -07004
5 completely rewritten, based on the MD driver code from Marc Zyngier
6
7 Changes:
8
9 - RAID-1/RAID-5 extensions by Miguel de Icaza, Gadi Oxman, Ingo Molnar
10 - RAID-6 extensions by H. Peter Anvin <hpa@zytor.com>
11 - boot support for linear and striped mode by Harald Hoyer <HarryH@Royal.Net>
12 - kerneld support by Boris Tobotras <boris@xtalk.msk.su>
13 - kmod support by: Cyrus Durgin
14 - RAID0 bugfixes: Mark Anthony Lisher <markal@iname.com>
15 - Devfs support by Richard Gooch <rgooch@atnf.csiro.au>
16
17 - lots of fixes and improvements to the RAID1/RAID5 and generic
18 RAID code (such as request based resynchronization):
19
20 Neil Brown <neilb@cse.unsw.edu.au>.
21
NeilBrown32a76272005-06-21 17:17:14 -070022 - persistent bitmap code
23 Copyright (C) 2003-2004, Paul Clements, SteelEye Technology, Inc.
24
Linus Torvalds1da177e2005-04-16 15:20:36 -070025 This program is free software; you can redistribute it and/or modify
26 it under the terms of the GNU General Public License as published by
27 the Free Software Foundation; either version 2, or (at your option)
28 any later version.
29
30 You should have received a copy of the GNU General Public License
31 (for example /usr/src/linux/COPYING); if not, write to the Free
32 Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
NeilBrown9d487392016-11-02 14:16:49 +110033
34 Errors, Warnings, etc.
35 Please use:
36 pr_crit() for error conditions that risk data loss
37 pr_err() for error conditions that are unexpected, like an IO error
38 or internal inconsistency
39 pr_warn() for error conditions that could have been predicated, like
40 adding a device to an array when it has incompatible metadata
41 pr_info() for every interesting, very rare events, like an array starting
42 or stopping, or resync starting or stopping
43 pr_debug() for everything else.
44
Linus Torvalds1da177e2005-04-16 15:20:36 -070045*/
46
Ingo Molnar3f07c012017-02-08 18:51:30 +010047#include <linux/sched/signal.h>
NeilBrowna6fb0932005-09-09 16:23:56 -070048#include <linux/kthread.h>
NeilBrownbff61972009-03-31 14:33:13 +110049#include <linux/blkdev.h>
Vishal Vermafc974ee2015-12-24 19:20:34 -070050#include <linux/badblocks.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070051#include <linux/sysctl.h>
NeilBrownbff61972009-03-31 14:33:13 +110052#include <linux/seq_file.h>
Al Viroff01bb42011-09-16 02:31:11 -040053#include <linux/fs.h>
NeilBrownd7603b72006-01-06 00:20:30 -080054#include <linux/poll.h>
NeilBrown16f17b32006-06-26 00:27:37 -070055#include <linux/ctype.h>
André Goddard Rosae7d28602009-12-14 18:01:06 -080056#include <linux/string.h>
NeilBrownfb4d8c72008-10-13 11:55:12 +110057#include <linux/hdreg.h>
58#include <linux/proc_fs.h>
59#include <linux/random.h>
Paul Gortmaker056075c2011-07-03 13:58:33 -040060#include <linux/module.h>
NeilBrownfb4d8c72008-10-13 11:55:12 +110061#include <linux/reboot.h>
NeilBrown32a76272005-06-21 17:17:14 -070062#include <linux/file.h>
Arnd Bergmannaa98aa32009-12-14 12:50:05 +110063#include <linux/compat.h>
Stephen Rothwell25570722008-10-15 09:09:21 +110064#include <linux/delay.h>
NeilBrownbff61972009-03-31 14:33:13 +110065#include <linux/raid/md_p.h>
66#include <linux/raid/md_u.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090067#include <linux/slab.h>
NeilBrown4ad23a972017-03-15 14:05:14 +110068#include <linux/percpu-refcount.h>
69
Shaohua Li504634f2016-11-18 09:44:08 -080070#include <trace/events/block.h>
NeilBrown43b2e5d2009-03-31 14:33:13 +110071#include "md.h"
Christoph Hellwigef740c32009-03-31 14:27:03 +110072#include "bitmap.h"
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -050073#include "md-cluster.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
Linus Torvalds1da177e2005-04-16 15:20:36 -070075#ifndef MODULE
NeilBrownd710e132008-10-13 11:55:12 +110076static void autostart_arrays(int part);
Linus Torvalds1da177e2005-04-16 15:20:36 -070077#endif
78
NeilBrown01f96c02011-09-21 15:30:20 +100079/* pers_list is a list of registered personalities protected
80 * by pers_lock.
81 * pers_lock does extra service to protect accesses to
82 * mddev->thread when the mutex cannot be held.
83 */
NeilBrown2604b702006-01-06 00:20:36 -080084static LIST_HEAD(pers_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -070085static DEFINE_SPINLOCK(pers_lock);
86
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -050087struct md_cluster_operations *md_cluster_ops;
Goldwyn Rodrigues589a1c42014-06-07 02:39:37 -050088EXPORT_SYMBOL(md_cluster_ops);
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -050089struct module *md_cluster_mod;
90EXPORT_SYMBOL(md_cluster_mod);
91
Bernd Schubert90b08712008-05-23 13:04:38 -070092static DECLARE_WAIT_QUEUE_HEAD(resync_wait);
Tejun Heoe804ac72010-10-15 15:36:08 +020093static struct workqueue_struct *md_wq;
94static struct workqueue_struct *md_misc_wq;
Bernd Schubert90b08712008-05-23 13:04:38 -070095
NeilBrown746d3202013-04-24 11:42:41 +100096static int remove_and_add_spares(struct mddev *mddev,
97 struct md_rdev *this);
NeilBrown5aa61f42014-12-15 12:56:57 +110098static void mddev_detach(struct mddev *mddev);
NeilBrown746d3202013-04-24 11:42:41 +100099
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100/*
Robert Becker1e509152009-12-14 12:49:58 +1100101 * Default number of read corrections we'll attempt on an rdev
102 * before ejecting it from the array. We divide the read error
103 * count by 2 for every hour elapsed between read errors.
104 */
105#define MD_DEFAULT_MAX_CORRECTED_READ_ERRORS 20
106/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107 * Current RAID-1,4,5 parallel reconstruction 'guaranteed speed limit'
108 * is 1000 KB/sec, so the extra system load does not show up that much.
109 * Increase it if you want to have more _guaranteed_ speed. Note that
Adrian Bunk338cec32005-09-10 00:26:54 -0700110 * the RAID driver will use the maximum available bandwidth if the IO
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111 * subsystem is idle. There is also an 'absolute maximum' reconstruction
112 * speed limit - in case reconstruction slows down your system despite
113 * idle IO detection.
114 *
115 * you can change it via /proc/sys/dev/raid/speed_limit_min and _max.
NeilBrown88202a02006-01-06 00:21:36 -0800116 * or /sys/block/mdX/md/sync_speed_{min,max}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117 */
118
119static int sysctl_speed_limit_min = 1000;
120static int sysctl_speed_limit_max = 200000;
NeilBrownfd01b882011-10-11 16:47:53 +1100121static inline int speed_min(struct mddev *mddev)
NeilBrown88202a02006-01-06 00:21:36 -0800122{
123 return mddev->sync_speed_min ?
124 mddev->sync_speed_min : sysctl_speed_limit_min;
125}
126
NeilBrownfd01b882011-10-11 16:47:53 +1100127static inline int speed_max(struct mddev *mddev)
NeilBrown88202a02006-01-06 00:21:36 -0800128{
129 return mddev->sync_speed_max ?
130 mddev->sync_speed_max : sysctl_speed_limit_max;
131}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132
133static struct ctl_table_header *raid_table_header;
134
Joe Perches82592c32013-11-14 15:16:18 +1100135static struct ctl_table raid_table[] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136 {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 .procname = "speed_limit_min",
138 .data = &sysctl_speed_limit_min,
139 .maxlen = sizeof(int),
NeilBrown80ca3a42006-07-10 04:44:18 -0700140 .mode = S_IRUGO|S_IWUSR,
Eric W. Biederman6d456112009-11-16 03:11:48 -0800141 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142 },
143 {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 .procname = "speed_limit_max",
145 .data = &sysctl_speed_limit_max,
146 .maxlen = sizeof(int),
NeilBrown80ca3a42006-07-10 04:44:18 -0700147 .mode = S_IRUGO|S_IWUSR,
Eric W. Biederman6d456112009-11-16 03:11:48 -0800148 .proc_handler = proc_dointvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 },
Eric W. Biederman894d2492009-11-05 14:34:02 -0800150 { }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151};
152
Joe Perches82592c32013-11-14 15:16:18 +1100153static struct ctl_table raid_dir_table[] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155 .procname = "raid",
156 .maxlen = 0,
NeilBrown80ca3a42006-07-10 04:44:18 -0700157 .mode = S_IRUGO|S_IXUGO,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158 .child = raid_table,
159 },
Eric W. Biederman894d2492009-11-05 14:34:02 -0800160 { }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161};
162
Joe Perches82592c32013-11-14 15:16:18 +1100163static struct ctl_table raid_root_table[] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165 .procname = "dev",
166 .maxlen = 0,
167 .mode = 0555,
168 .child = raid_dir_table,
169 },
Eric W. Biederman894d2492009-11-05 14:34:02 -0800170 { }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171};
172
Alexey Dobriyan83d5cde2009-09-21 17:01:13 -0700173static const struct block_device_operations md_fops;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174
NeilBrownf91de922005-11-08 21:39:36 -0800175static int start_readonly;
176
NeilBrown78b63502017-04-12 16:26:13 +1000177/*
178 * The original mechanism for creating an md device is to create
179 * a device node in /dev and to open it. This causes races with device-close.
180 * The preferred method is to write to the "new_array" module parameter.
181 * This can avoid races.
182 * Setting create_on_open to false disables the original mechanism
183 * so all the races disappear.
184 */
185static bool create_on_open = true;
186
NeilBrowna167f662010-10-26 18:31:13 +1100187/* bio_clone_mddev
NeilBrown9b10f6a2017-06-18 14:38:59 +1000188 * like bio_clone_bioset, but with a local bio set
NeilBrowna167f662010-10-26 18:31:13 +1100189 */
190
NeilBrowna167f662010-10-26 18:31:13 +1100191struct bio *bio_alloc_mddev(gfp_t gfp_mask, int nr_iovecs,
NeilBrownfd01b882011-10-11 16:47:53 +1100192 struct mddev *mddev)
NeilBrowna167f662010-10-26 18:31:13 +1100193{
194 struct bio *b;
NeilBrowna167f662010-10-26 18:31:13 +1100195
196 if (!mddev || !mddev->bio_set)
197 return bio_alloc(gfp_mask, nr_iovecs);
198
Kent Overstreet395c72a2012-09-06 15:34:55 -0700199 b = bio_alloc_bioset(gfp_mask, nr_iovecs, mddev->bio_set);
NeilBrowna167f662010-10-26 18:31:13 +1100200 if (!b)
201 return NULL;
NeilBrowna167f662010-10-26 18:31:13 +1100202 return b;
203}
204EXPORT_SYMBOL_GPL(bio_alloc_mddev);
205
NeilBrown5a850712017-06-21 09:12:21 +1000206static struct bio *md_bio_alloc_sync(struct mddev *mddev)
207{
Shaohua Li7f053a6a2017-06-23 09:19:49 -0700208 if (!mddev || !mddev->sync_set)
NeilBrown5a850712017-06-21 09:12:21 +1000209 return bio_alloc(GFP_NOIO, 1);
210
211 return bio_alloc_bioset(GFP_NOIO, 1, mddev->sync_set);
212}
213
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214/*
NeilBrownd7603b72006-01-06 00:20:30 -0800215 * We have a system wide 'event count' that is incremented
216 * on any 'interesting' event, and readers of /proc/mdstat
217 * can use 'poll' or 'select' to find out when the event
218 * count increases.
219 *
220 * Events are:
221 * start array, stop array, error, add device, remove device,
222 * start build, activate spare
223 */
NeilBrown2989ddb2006-01-06 00:20:43 -0800224static DECLARE_WAIT_QUEUE_HEAD(md_event_waiters);
NeilBrownd7603b72006-01-06 00:20:30 -0800225static atomic_t md_event_count;
NeilBrownfd01b882011-10-11 16:47:53 +1100226void md_new_event(struct mddev *mddev)
NeilBrownd7603b72006-01-06 00:20:30 -0800227{
228 atomic_inc(&md_event_count);
229 wake_up(&md_event_waiters);
230}
NeilBrown29269552006-03-27 01:18:10 -0800231EXPORT_SYMBOL_GPL(md_new_event);
NeilBrownd7603b72006-01-06 00:20:30 -0800232
233/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 * Enables to iterate over all existing md arrays
235 * all_mddevs_lock protects this list.
236 */
237static LIST_HEAD(all_mddevs);
238static DEFINE_SPINLOCK(all_mddevs_lock);
239
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240/*
241 * iterates through all used mddevs in the system.
242 * We take care to grab the all_mddevs_lock whenever navigating
243 * the list, and to always hold a refcount when unlocked.
244 * Any code which breaks out of this loop while own
245 * a reference to the current mddev and must mddev_put it.
246 */
NeilBrownfd01b882011-10-11 16:47:53 +1100247#define for_each_mddev(_mddev,_tmp) \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248 \
NeilBrownf72ffdd2014-09-30 14:23:59 +1000249 for (({ spin_lock(&all_mddevs_lock); \
NeilBrownfd01b882011-10-11 16:47:53 +1100250 _tmp = all_mddevs.next; \
251 _mddev = NULL;}); \
252 ({ if (_tmp != &all_mddevs) \
253 mddev_get(list_entry(_tmp, struct mddev, all_mddevs));\
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 spin_unlock(&all_mddevs_lock); \
NeilBrownfd01b882011-10-11 16:47:53 +1100255 if (_mddev) mddev_put(_mddev); \
256 _mddev = list_entry(_tmp, struct mddev, all_mddevs); \
257 _tmp != &all_mddevs;}); \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 ({ spin_lock(&all_mddevs_lock); \
NeilBrownfd01b882011-10-11 16:47:53 +1100259 _tmp = _tmp->next;}) \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 )
261
NeilBrown409c57f2009-03-31 14:39:39 +1100262/* Rather than calling directly into the personality make_request function,
263 * IO requests come here first so that we can check if the device is
264 * being suspended pending a reconfiguration.
265 * We hold a refcount over the call to ->make_request. By the time that
266 * call has finished, the bio has been linked into some internal structure
267 * and so is visible to ->quiesce(), so we don't need the refcount any more.
268 */
Jens Axboedece1632015-11-05 10:41:16 -0700269static blk_qc_t md_make_request(struct request_queue *q, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270{
NeilBrown49077322010-03-25 16:20:56 +1100271 const int rw = bio_data_dir(bio);
NeilBrownfd01b882011-10-11 16:47:53 +1100272 struct mddev *mddev = q->queuedata;
Chris Masone91ece52011-02-07 19:21:48 -0500273 unsigned int sectors;
Gu Zheng74672d02015-04-03 08:44:47 +0800274 int cpu;
NeilBrown49077322010-03-25 16:20:56 +1100275
NeilBrownaf67c312017-06-18 14:38:57 +1000276 blk_queue_split(q, &bio);
Kent Overstreet54efd502015-04-23 22:37:18 -0700277
NeilBrown274d8cb2016-01-04 16:16:58 +1100278 if (mddev == NULL || mddev->pers == NULL) {
NeilBrown409c57f2009-03-31 14:39:39 +1100279 bio_io_error(bio);
Jens Axboedece1632015-11-05 10:41:16 -0700280 return BLK_QC_T_NONE;
NeilBrown409c57f2009-03-31 14:39:39 +1100281 }
Sebastian Riemerbbfa57c2013-02-21 13:28:09 +1100282 if (mddev->ro == 1 && unlikely(rw == WRITE)) {
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200283 if (bio_sectors(bio) != 0)
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200284 bio->bi_status = BLK_STS_IOERR;
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200285 bio_endio(bio);
Jens Axboedece1632015-11-05 10:41:16 -0700286 return BLK_QC_T_NONE;
Sebastian Riemerbbfa57c2013-02-21 13:28:09 +1100287 }
NeilBrowncc27b0c2017-06-05 16:49:39 +1000288check_suspended:
NeilBrown409c57f2009-03-31 14:39:39 +1100289 rcu_read_lock();
Tejun Heoe9c74692010-09-03 11:56:18 +0200290 if (mddev->suspended) {
NeilBrown409c57f2009-03-31 14:39:39 +1100291 DEFINE_WAIT(__wait);
292 for (;;) {
293 prepare_to_wait(&mddev->sb_wait, &__wait,
294 TASK_UNINTERRUPTIBLE);
Tejun Heoe9c74692010-09-03 11:56:18 +0200295 if (!mddev->suspended)
NeilBrown409c57f2009-03-31 14:39:39 +1100296 break;
297 rcu_read_unlock();
298 schedule();
299 rcu_read_lock();
300 }
301 finish_wait(&mddev->sb_wait, &__wait);
302 }
303 atomic_inc(&mddev->active_io);
304 rcu_read_unlock();
NeilBrown49077322010-03-25 16:20:56 +1100305
Chris Masone91ece52011-02-07 19:21:48 -0500306 /*
307 * save the sectors now since our bio can
308 * go away inside make_request
309 */
310 sectors = bio_sectors(bio);
Shaohua Li9c573de2016-04-25 16:52:38 -0700311 /* bio could be mergeable after passing to underlayer */
Jens Axboe1eff9d32016-08-05 15:35:16 -0600312 bio->bi_opf &= ~REQ_NOMERGE;
NeilBrowncc27b0c2017-06-05 16:49:39 +1000313 if (!mddev->pers->make_request(mddev, bio)) {
314 atomic_dec(&mddev->active_io);
315 wake_up(&mddev->sb_wait);
316 goto check_suspended;
317 }
NeilBrown49077322010-03-25 16:20:56 +1100318
Gu Zheng74672d02015-04-03 08:44:47 +0800319 cpu = part_stat_lock();
320 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
321 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw], sectors);
322 part_stat_unlock();
NeilBrown49077322010-03-25 16:20:56 +1100323
NeilBrown409c57f2009-03-31 14:39:39 +1100324 if (atomic_dec_and_test(&mddev->active_io) && mddev->suspended)
325 wake_up(&mddev->sb_wait);
Jens Axboedece1632015-11-05 10:41:16 -0700326
327 return BLK_QC_T_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328}
329
NeilBrown9e35b992010-04-06 14:23:02 +1000330/* mddev_suspend makes sure no new requests are submitted
331 * to the device, and that any requests that have been submitted
332 * are completely handled.
NeilBrownafa0f552014-12-15 12:56:58 +1100333 * Once mddev_detach() is called and completes, the module will be
334 * completely unused.
NeilBrown9e35b992010-04-06 14:23:02 +1000335 */
NeilBrownfd01b882011-10-11 16:47:53 +1100336void mddev_suspend(struct mddev *mddev)
NeilBrown409c57f2009-03-31 14:39:39 +1100337{
Heinz Mauelshagen092398d2016-05-03 19:43:57 +0200338 WARN_ON_ONCE(mddev->thread && current == mddev->thread->tsk);
Mikulas Patocka0dc10e52015-12-18 15:19:16 +1100339 if (mddev->suspended++)
340 return;
NeilBrown409c57f2009-03-31 14:39:39 +1100341 synchronize_rcu();
NeilBrowncc27b0c2017-06-05 16:49:39 +1000342 wake_up(&mddev->sb_wait);
NeilBrown409c57f2009-03-31 14:39:39 +1100343 wait_event(mddev->sb_wait, atomic_read(&mddev->active_io) == 0);
344 mddev->pers->quiesce(mddev, 1);
Jonathan Brassow0d9f4f12012-05-16 04:06:14 -0500345
346 del_timer_sync(&mddev->safemode_timer);
NeilBrown409c57f2009-03-31 14:39:39 +1100347}
NeilBrown390ee602010-06-01 19:37:27 +1000348EXPORT_SYMBOL_GPL(mddev_suspend);
NeilBrown409c57f2009-03-31 14:39:39 +1100349
NeilBrownfd01b882011-10-11 16:47:53 +1100350void mddev_resume(struct mddev *mddev)
NeilBrown409c57f2009-03-31 14:39:39 +1100351{
Mikulas Patocka0dc10e52015-12-18 15:19:16 +1100352 if (--mddev->suspended)
353 return;
NeilBrown409c57f2009-03-31 14:39:39 +1100354 wake_up(&mddev->sb_wait);
355 mddev->pers->quiesce(mddev, 0);
Jonathan Brassow0fd018a2011-06-07 17:49:36 -0500356
Jonathan Brassow47525e52012-05-22 13:55:29 +1000357 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
Jonathan Brassow0fd018a2011-06-07 17:49:36 -0500358 md_wakeup_thread(mddev->thread);
359 md_wakeup_thread(mddev->sync_thread); /* possibly kick off a reshape */
NeilBrown409c57f2009-03-31 14:39:39 +1100360}
NeilBrown390ee602010-06-01 19:37:27 +1000361EXPORT_SYMBOL_GPL(mddev_resume);
NeilBrown409c57f2009-03-31 14:39:39 +1100362
NeilBrownfd01b882011-10-11 16:47:53 +1100363int mddev_congested(struct mddev *mddev, int bits)
NeilBrown3fa841d2009-09-23 18:10:29 +1000364{
NeilBrown5c675f82014-12-15 12:56:56 +1100365 struct md_personality *pers = mddev->pers;
366 int ret = 0;
367
368 rcu_read_lock();
369 if (mddev->suspended)
370 ret = 1;
371 else if (pers && pers->congested)
372 ret = pers->congested(mddev, bits);
373 rcu_read_unlock();
374 return ret;
NeilBrown3fa841d2009-09-23 18:10:29 +1000375}
NeilBrown5c675f82014-12-15 12:56:56 +1100376EXPORT_SYMBOL_GPL(mddev_congested);
377static int md_congested(void *data, int bits)
378{
379 struct mddev *mddev = data;
380 return mddev_congested(mddev, bits);
381}
NeilBrown3fa841d2009-09-23 18:10:29 +1000382
NeilBrowna2826aa2009-12-14 12:49:49 +1100383/*
Tejun Heoe9c74692010-09-03 11:56:18 +0200384 * Generic flush handling for md
NeilBrowna2826aa2009-12-14 12:49:49 +1100385 */
386
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200387static void md_end_flush(struct bio *bio)
NeilBrowna2826aa2009-12-14 12:49:49 +1100388{
NeilBrown3cb03002011-10-11 16:45:26 +1100389 struct md_rdev *rdev = bio->bi_private;
NeilBrownfd01b882011-10-11 16:47:53 +1100390 struct mddev *mddev = rdev->mddev;
NeilBrowna2826aa2009-12-14 12:49:49 +1100391
392 rdev_dec_pending(rdev, mddev);
393
394 if (atomic_dec_and_test(&mddev->flush_pending)) {
Tejun Heoe9c74692010-09-03 11:56:18 +0200395 /* The pre-request flush has finished */
Tejun Heoe804ac72010-10-15 15:36:08 +0200396 queue_work(md_wq, &mddev->flush_work);
NeilBrowna2826aa2009-12-14 12:49:49 +1100397 }
398 bio_put(bio);
399}
400
NeilBrowna7a07e62010-12-09 16:04:25 +1100401static void md_submit_flush_data(struct work_struct *ws);
402
NeilBrowna035fc32010-12-09 16:17:51 +1100403static void submit_flushes(struct work_struct *ws)
NeilBrowna2826aa2009-12-14 12:49:49 +1100404{
NeilBrownfd01b882011-10-11 16:47:53 +1100405 struct mddev *mddev = container_of(ws, struct mddev, flush_work);
NeilBrown3cb03002011-10-11 16:45:26 +1100406 struct md_rdev *rdev;
NeilBrowna2826aa2009-12-14 12:49:49 +1100407
NeilBrowna7a07e62010-12-09 16:04:25 +1100408 INIT_WORK(&mddev->flush_work, md_submit_flush_data);
409 atomic_set(&mddev->flush_pending, 1);
NeilBrowna2826aa2009-12-14 12:49:49 +1100410 rcu_read_lock();
NeilBrowndafb20f2012-03-19 12:46:39 +1100411 rdev_for_each_rcu(rdev, mddev)
NeilBrowna2826aa2009-12-14 12:49:49 +1100412 if (rdev->raid_disk >= 0 &&
413 !test_bit(Faulty, &rdev->flags)) {
414 /* Take two references, one is dropped
415 * when request finishes, one after
416 * we reclaim rcu_read_lock
417 */
418 struct bio *bi;
419 atomic_inc(&rdev->nr_pending);
420 atomic_inc(&rdev->nr_pending);
421 rcu_read_unlock();
Shaohua Lib5e1b8c2012-05-21 09:26:59 +1000422 bi = bio_alloc_mddev(GFP_NOIO, 0, mddev);
Tejun Heoe9c74692010-09-03 11:56:18 +0200423 bi->bi_end_io = md_end_flush;
NeilBrowna2826aa2009-12-14 12:49:49 +1100424 bi->bi_private = rdev;
Christoph Hellwig74d46992017-08-23 19:10:32 +0200425 bio_set_dev(bi, rdev->bdev);
Christoph Hellwig70fd7612016-11-01 07:40:10 -0600426 bi->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
NeilBrowna2826aa2009-12-14 12:49:49 +1100427 atomic_inc(&mddev->flush_pending);
Mike Christie4e49ea42016-06-05 14:31:41 -0500428 submit_bio(bi);
NeilBrowna2826aa2009-12-14 12:49:49 +1100429 rcu_read_lock();
430 rdev_dec_pending(rdev, mddev);
431 }
432 rcu_read_unlock();
NeilBrowna7a07e62010-12-09 16:04:25 +1100433 if (atomic_dec_and_test(&mddev->flush_pending))
434 queue_work(md_wq, &mddev->flush_work);
NeilBrowna2826aa2009-12-14 12:49:49 +1100435}
436
Tejun Heoe9c74692010-09-03 11:56:18 +0200437static void md_submit_flush_data(struct work_struct *ws)
NeilBrowna2826aa2009-12-14 12:49:49 +1100438{
NeilBrownfd01b882011-10-11 16:47:53 +1100439 struct mddev *mddev = container_of(ws, struct mddev, flush_work);
Tejun Heoe9c74692010-09-03 11:56:18 +0200440 struct bio *bio = mddev->flush_bio;
NeilBrowna2826aa2009-12-14 12:49:49 +1100441
Kent Overstreet4f024f32013-10-11 15:44:27 -0700442 if (bio->bi_iter.bi_size == 0)
NeilBrowna2826aa2009-12-14 12:49:49 +1100443 /* an empty barrier - all done */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200444 bio_endio(bio);
NeilBrowna2826aa2009-12-14 12:49:49 +1100445 else {
Jens Axboe1eff9d32016-08-05 15:35:16 -0600446 bio->bi_opf &= ~REQ_PREFLUSH;
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +0200447 mddev->pers->make_request(mddev, bio);
NeilBrowna2826aa2009-12-14 12:49:49 +1100448 }
NeilBrown2b74e122010-12-09 15:59:01 +1100449
450 mddev->flush_bio = NULL;
451 wake_up(&mddev->sb_wait);
NeilBrowna2826aa2009-12-14 12:49:49 +1100452}
453
NeilBrownfd01b882011-10-11 16:47:53 +1100454void md_flush_request(struct mddev *mddev, struct bio *bio)
NeilBrowna2826aa2009-12-14 12:49:49 +1100455{
NeilBrown85572d72014-12-15 12:56:56 +1100456 spin_lock_irq(&mddev->lock);
NeilBrowna2826aa2009-12-14 12:49:49 +1100457 wait_event_lock_irq(mddev->sb_wait,
Tejun Heoe9c74692010-09-03 11:56:18 +0200458 !mddev->flush_bio,
NeilBrown85572d72014-12-15 12:56:56 +1100459 mddev->lock);
Tejun Heoe9c74692010-09-03 11:56:18 +0200460 mddev->flush_bio = bio;
NeilBrown85572d72014-12-15 12:56:56 +1100461 spin_unlock_irq(&mddev->lock);
NeilBrowna2826aa2009-12-14 12:49:49 +1100462
NeilBrowna035fc32010-12-09 16:17:51 +1100463 INIT_WORK(&mddev->flush_work, submit_flushes);
464 queue_work(md_wq, &mddev->flush_work);
NeilBrowna2826aa2009-12-14 12:49:49 +1100465}
Tejun Heoe9c74692010-09-03 11:56:18 +0200466EXPORT_SYMBOL(md_flush_request);
NeilBrown409c57f2009-03-31 14:39:39 +1100467
NeilBrownfd01b882011-10-11 16:47:53 +1100468static inline struct mddev *mddev_get(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469{
470 atomic_inc(&mddev->active);
471 return mddev;
472}
473
Dan Williams5fd3a172009-03-04 00:57:25 -0700474static void mddev_delayed_delete(struct work_struct *ws);
NeilBrownd3374822009-01-09 08:31:10 +1100475
NeilBrownfd01b882011-10-11 16:47:53 +1100476static void mddev_put(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477{
Shaohua Li7184ef82017-07-03 14:34:23 -0700478 struct bio_set *bs = NULL, *sync_bs = NULL;
NeilBrowna167f662010-10-26 18:31:13 +1100479
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480 if (!atomic_dec_and_lock(&mddev->active, &all_mddevs_lock))
481 return;
NeilBrownd3374822009-01-09 08:31:10 +1100482 if (!mddev->raid_disks && list_empty(&mddev->disks) &&
NeilBrowncbd19982009-12-30 12:08:49 +1100483 mddev->ctime == 0 && !mddev->hold_active) {
484 /* Array is not configured at all, and not held active,
485 * so destroy it */
NeilBrownaf8a2432011-12-08 15:49:46 +1100486 list_del_init(&mddev->all_mddevs);
NeilBrowna167f662010-10-26 18:31:13 +1100487 bs = mddev->bio_set;
Shaohua Li7184ef82017-07-03 14:34:23 -0700488 sync_bs = mddev->sync_set;
NeilBrowna167f662010-10-26 18:31:13 +1100489 mddev->bio_set = NULL;
NeilBrown5a850712017-06-21 09:12:21 +1000490 mddev->sync_set = NULL;
NeilBrownd3374822009-01-09 08:31:10 +1100491 if (mddev->gendisk) {
Tejun Heoe804ac72010-10-15 15:36:08 +0200492 /* We did a probe so need to clean up. Call
493 * queue_work inside the spinlock so that
494 * flush_workqueue() after mddev_find will
495 * succeed in waiting for the work to be done.
NeilBrownd3374822009-01-09 08:31:10 +1100496 */
497 INIT_WORK(&mddev->del_work, mddev_delayed_delete);
Tejun Heoe804ac72010-10-15 15:36:08 +0200498 queue_work(md_misc_wq, &mddev->del_work);
NeilBrownd3374822009-01-09 08:31:10 +1100499 } else
500 kfree(mddev);
501 }
502 spin_unlock(&all_mddevs_lock);
NeilBrowna167f662010-10-26 18:31:13 +1100503 if (bs)
504 bioset_free(bs);
Shaohua Li7184ef82017-07-03 14:34:23 -0700505 if (sync_bs)
506 bioset_free(sync_bs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507}
508
Sasha Levin25b2edf2015-07-24 18:19:58 -0400509static void md_safemode_timeout(unsigned long data);
510
NeilBrownfd01b882011-10-11 16:47:53 +1100511void mddev_init(struct mddev *mddev)
NeilBrownfafd7fb2010-04-01 15:55:30 +1100512{
513 mutex_init(&mddev->open_mutex);
514 mutex_init(&mddev->reconfig_mutex);
515 mutex_init(&mddev->bitmap_info.mutex);
516 INIT_LIST_HEAD(&mddev->disks);
517 INIT_LIST_HEAD(&mddev->all_mddevs);
Sasha Levin25b2edf2015-07-24 18:19:58 -0400518 setup_timer(&mddev->safemode_timer, md_safemode_timeout,
519 (unsigned long) mddev);
NeilBrownfafd7fb2010-04-01 15:55:30 +1100520 atomic_set(&mddev->active, 1);
521 atomic_set(&mddev->openers, 0);
522 atomic_set(&mddev->active_io, 0);
NeilBrown85572d72014-12-15 12:56:56 +1100523 spin_lock_init(&mddev->lock);
NeilBrownfafd7fb2010-04-01 15:55:30 +1100524 atomic_set(&mddev->flush_pending, 0);
525 init_waitqueue_head(&mddev->sb_wait);
526 init_waitqueue_head(&mddev->recovery_wait);
527 mddev->reshape_position = MaxSector;
NeilBrown2c810cd2012-05-21 09:27:00 +1000528 mddev->reshape_backwards = 0;
Jonathan Brassowc4a39552013-06-25 01:23:59 -0500529 mddev->last_sync_action = "none";
NeilBrownfafd7fb2010-04-01 15:55:30 +1100530 mddev->resync_min = 0;
531 mddev->resync_max = MaxSector;
532 mddev->level = LEVEL_NONE;
533}
NeilBrown390ee602010-06-01 19:37:27 +1000534EXPORT_SYMBOL_GPL(mddev_init);
NeilBrownfafd7fb2010-04-01 15:55:30 +1100535
NeilBrownf72ffdd2014-09-30 14:23:59 +1000536static struct mddev *mddev_find(dev_t unit)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537{
NeilBrownfd01b882011-10-11 16:47:53 +1100538 struct mddev *mddev, *new = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539
NeilBrown8f5f02c2011-02-16 13:58:51 +1100540 if (unit && MAJOR(unit) != MD_MAJOR)
541 unit &= ~((1<<MdpMinorShift)-1);
542
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 retry:
544 spin_lock(&all_mddevs_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545
NeilBrownefeb53c2009-01-09 08:31:10 +1100546 if (unit) {
547 list_for_each_entry(mddev, &all_mddevs, all_mddevs)
548 if (mddev->unit == unit) {
549 mddev_get(mddev);
550 spin_unlock(&all_mddevs_lock);
551 kfree(new);
552 return mddev;
553 }
554
555 if (new) {
556 list_add(&new->all_mddevs, &all_mddevs);
557 spin_unlock(&all_mddevs_lock);
558 new->hold_active = UNTIL_IOCTL;
559 return new;
560 }
561 } else if (new) {
562 /* find an unused unit number */
563 static int next_minor = 512;
564 int start = next_minor;
565 int is_free = 0;
566 int dev = 0;
567 while (!is_free) {
568 dev = MKDEV(MD_MAJOR, next_minor);
569 next_minor++;
570 if (next_minor > MINORMASK)
571 next_minor = 0;
572 if (next_minor == start) {
573 /* Oh dear, all in use. */
574 spin_unlock(&all_mddevs_lock);
575 kfree(new);
576 return NULL;
577 }
NeilBrownf72ffdd2014-09-30 14:23:59 +1000578
NeilBrownefeb53c2009-01-09 08:31:10 +1100579 is_free = 1;
580 list_for_each_entry(mddev, &all_mddevs, all_mddevs)
581 if (mddev->unit == dev) {
582 is_free = 0;
583 break;
584 }
585 }
586 new->unit = dev;
587 new->md_minor = MINOR(dev);
588 new->hold_active = UNTIL_STOP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589 list_add(&new->all_mddevs, &all_mddevs);
590 spin_unlock(&all_mddevs_lock);
591 return new;
592 }
593 spin_unlock(&all_mddevs_lock);
594
NeilBrown9ffae0c2006-01-06 00:20:32 -0800595 new = kzalloc(sizeof(*new), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596 if (!new)
597 return NULL;
598
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599 new->unit = unit;
600 if (MAJOR(unit) == MD_MAJOR)
601 new->md_minor = MINOR(unit);
602 else
603 new->md_minor = MINOR(unit) >> MdpMinorShift;
604
NeilBrownfafd7fb2010-04-01 15:55:30 +1100605 mddev_init(new);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607 goto retry;
608}
609
NeilBrownb6eb1272010-04-15 10:13:47 +1000610static struct attribute_group md_redundancy_group;
611
NeilBrown5c47daf2014-12-15 12:57:01 +1100612void mddev_unlock(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613{
NeilBrowna64c8762010-04-14 17:15:37 +1000614 if (mddev->to_remove) {
NeilBrownb6eb1272010-04-15 10:13:47 +1000615 /* These cannot be removed under reconfig_mutex as
616 * an access to the files will try to take reconfig_mutex
617 * while holding the file unremovable, which leads to
618 * a deadlock.
NeilBrownbb4f1e92010-08-08 21:18:03 +1000619 * So hold set sysfs_active while the remove in happeing,
620 * and anything else which might set ->to_remove or my
621 * otherwise change the sysfs namespace will fail with
622 * -EBUSY if sysfs_active is still set.
623 * We set sysfs_active under reconfig_mutex and elsewhere
624 * test it under the same mutex to ensure its correct value
625 * is seen.
NeilBrownb6eb1272010-04-15 10:13:47 +1000626 */
NeilBrowna64c8762010-04-14 17:15:37 +1000627 struct attribute_group *to_remove = mddev->to_remove;
628 mddev->to_remove = NULL;
NeilBrownbb4f1e92010-08-08 21:18:03 +1000629 mddev->sysfs_active = 1;
NeilBrownb6eb1272010-04-15 10:13:47 +1000630 mutex_unlock(&mddev->reconfig_mutex);
631
NeilBrown00bcb4a2010-06-01 19:37:23 +1000632 if (mddev->kobj.sd) {
633 if (to_remove != &md_redundancy_group)
634 sysfs_remove_group(&mddev->kobj, to_remove);
635 if (mddev->pers == NULL ||
636 mddev->pers->sync_request == NULL) {
637 sysfs_remove_group(&mddev->kobj, &md_redundancy_group);
638 if (mddev->sysfs_action)
639 sysfs_put(mddev->sysfs_action);
640 mddev->sysfs_action = NULL;
641 }
NeilBrowna64c8762010-04-14 17:15:37 +1000642 }
NeilBrownbb4f1e92010-08-08 21:18:03 +1000643 mddev->sysfs_active = 0;
NeilBrownb6eb1272010-04-15 10:13:47 +1000644 } else
645 mutex_unlock(&mddev->reconfig_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646
Chris Dunlop751e67c2011-10-19 16:48:26 +1100647 /* As we've dropped the mutex we need a spinlock to
648 * make sure the thread doesn't disappear
NeilBrown01f96c02011-09-21 15:30:20 +1000649 */
650 spin_lock(&pers_lock);
NeilBrown005eca52005-08-22 13:11:08 -0700651 md_wakeup_thread(mddev->thread);
NeilBrown01f96c02011-09-21 15:30:20 +1000652 spin_unlock(&pers_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653}
NeilBrown5c47daf2014-12-15 12:57:01 +1100654EXPORT_SYMBOL_GPL(mddev_unlock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655
Goldwyn Rodrigues57d051d2015-04-14 10:43:55 -0500656struct md_rdev *md_find_rdev_nr_rcu(struct mddev *mddev, int nr)
NeilBrown1ca69c42012-10-11 13:37:33 +1100657{
658 struct md_rdev *rdev;
659
660 rdev_for_each_rcu(rdev, mddev)
661 if (rdev->desc_nr == nr)
662 return rdev;
663
664 return NULL;
665}
Goldwyn Rodrigues57d051d2015-04-14 10:43:55 -0500666EXPORT_SYMBOL_GPL(md_find_rdev_nr_rcu);
NeilBrown1ca69c42012-10-11 13:37:33 +1100667
668static struct md_rdev *find_rdev(struct mddev *mddev, dev_t dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669{
NeilBrown3cb03002011-10-11 16:45:26 +1100670 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671
NeilBrowndafb20f2012-03-19 12:46:39 +1100672 rdev_for_each(rdev, mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 if (rdev->bdev->bd_dev == dev)
674 return rdev;
Cheng Renquan159ec1f2009-01-09 08:31:08 +1100675
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 return NULL;
677}
678
NeilBrown1ca69c42012-10-11 13:37:33 +1100679static struct md_rdev *find_rdev_rcu(struct mddev *mddev, dev_t dev)
680{
681 struct md_rdev *rdev;
682
683 rdev_for_each_rcu(rdev, mddev)
684 if (rdev->bdev->bd_dev == dev)
685 return rdev;
686
687 return NULL;
688}
689
NeilBrown84fc4b52011-10-11 16:49:58 +1100690static struct md_personality *find_pers(int level, char *clevel)
NeilBrown2604b702006-01-06 00:20:36 -0800691{
NeilBrown84fc4b52011-10-11 16:49:58 +1100692 struct md_personality *pers;
NeilBrownd9d166c2006-01-06 00:20:51 -0800693 list_for_each_entry(pers, &pers_list, list) {
694 if (level != LEVEL_NONE && pers->level == level)
NeilBrown2604b702006-01-06 00:20:36 -0800695 return pers;
NeilBrownd9d166c2006-01-06 00:20:51 -0800696 if (strcmp(pers->name, clevel)==0)
697 return pers;
698 }
NeilBrown2604b702006-01-06 00:20:36 -0800699 return NULL;
700}
701
Andre Nollb73df2d2008-07-11 22:02:23 +1000702/* return the offset of the super block in 512byte sectors */
NeilBrown3cb03002011-10-11 16:45:26 +1100703static inline sector_t calc_dev_sboffset(struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704{
Jonathan Brassow57b2caa2011-01-14 09:14:33 +1100705 sector_t num_sectors = i_size_read(rdev->bdev->bd_inode) / 512;
Andre Nollb73df2d2008-07-11 22:02:23 +1000706 return MD_NEW_SIZE_SECTORS(num_sectors);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707}
708
NeilBrownf72ffdd2014-09-30 14:23:59 +1000709static int alloc_disk_sb(struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 rdev->sb_page = alloc_page(GFP_KERNEL);
NeilBrown7f0f0d82016-11-02 14:16:49 +1100712 if (!rdev->sb_page)
Andre Nollebc24332008-07-11 22:02:20 +1000713 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 return 0;
715}
716
NeilBrown545c8792012-05-22 13:54:30 +1000717void md_rdev_clear(struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718{
719 if (rdev->sb_page) {
NeilBrown2d1f3b52006-01-06 00:20:31 -0800720 put_page(rdev->sb_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 rdev->sb_loaded = 0;
722 rdev->sb_page = NULL;
Andre Noll0f420352008-07-11 22:02:23 +1000723 rdev->sb_start = 0;
Andre Nolldd8ac332009-03-31 14:33:13 +1100724 rdev->sectors = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 }
NeilBrown2699b672011-07-28 11:31:47 +1000726 if (rdev->bb_page) {
727 put_page(rdev->bb_page);
728 rdev->bb_page = NULL;
729 }
Dan Williamsd3b407fb2016-01-06 12:19:22 -0800730 badblocks_exit(&rdev->badblocks);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731}
NeilBrown545c8792012-05-22 13:54:30 +1000732EXPORT_SYMBOL_GPL(md_rdev_clear);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200734static void super_written(struct bio *bio)
NeilBrown7bfa19f2005-06-21 17:17:28 -0700735{
NeilBrown3cb03002011-10-11 16:45:26 +1100736 struct md_rdev *rdev = bio->bi_private;
NeilBrownfd01b882011-10-11 16:47:53 +1100737 struct mddev *mddev = rdev->mddev;
NeilBrown7bfa19f2005-06-21 17:17:28 -0700738
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200739 if (bio->bi_status) {
740 pr_err("md: super_written gets error=%d\n", bio->bi_status);
NeilBrowna9701a32005-11-08 21:39:34 -0800741 md_error(mddev, rdev);
NeilBrown46533ff2016-11-18 16:16:11 +1100742 if (!test_bit(Faulty, &rdev->flags)
743 && (bio->bi_opf & MD_FAILFAST)) {
Shaohua Li29530792016-12-08 15:48:19 -0800744 set_bit(MD_SB_NEED_REWRITE, &mddev->sb_flags);
NeilBrown46533ff2016-11-18 16:16:11 +1100745 set_bit(LastDev, &rdev->flags);
746 }
747 } else
748 clear_bit(LastDev, &rdev->flags);
NeilBrown7bfa19f2005-06-21 17:17:28 -0700749
NeilBrowna9701a32005-11-08 21:39:34 -0800750 if (atomic_dec_and_test(&mddev->pending_writes))
751 wake_up(&mddev->sb_wait);
Shaohua Lied3b98c2016-03-29 14:00:19 -0700752 rdev_dec_pending(rdev, mddev);
Neil Brownf8b58ed2005-06-27 22:29:34 -0700753 bio_put(bio);
NeilBrown7bfa19f2005-06-21 17:17:28 -0700754}
755
NeilBrownfd01b882011-10-11 16:47:53 +1100756void md_super_write(struct mddev *mddev, struct md_rdev *rdev,
NeilBrown7bfa19f2005-06-21 17:17:28 -0700757 sector_t sector, int size, struct page *page)
758{
759 /* write first size bytes of page to sector of rdev
760 * Increment mddev->pending_writes before returning
761 * and decrement it on completion, waking up sb_wait
762 * if zero is reached.
763 * If an error occurred, call md_error
764 */
NeilBrown46533ff2016-11-18 16:16:11 +1100765 struct bio *bio;
766 int ff = 0;
767
768 if (test_bit(Faulty, &rdev->flags))
769 return;
770
NeilBrown5a850712017-06-21 09:12:21 +1000771 bio = md_bio_alloc_sync(mddev);
NeilBrown7bfa19f2005-06-21 17:17:28 -0700772
Shaohua Lied3b98c2016-03-29 14:00:19 -0700773 atomic_inc(&rdev->nr_pending);
774
Christoph Hellwig74d46992017-08-23 19:10:32 +0200775 bio_set_dev(bio, rdev->meta_bdev ? rdev->meta_bdev : rdev->bdev);
Kent Overstreet4f024f32013-10-11 15:44:27 -0700776 bio->bi_iter.bi_sector = sector;
NeilBrown7bfa19f2005-06-21 17:17:28 -0700777 bio_add_page(bio, page, size, 0);
778 bio->bi_private = rdev;
779 bio->bi_end_io = super_written;
NeilBrown46533ff2016-11-18 16:16:11 +1100780
781 if (test_bit(MD_FAILFAST_SUPPORTED, &mddev->flags) &&
782 test_bit(FailFast, &rdev->flags) &&
783 !test_bit(LastDev, &rdev->flags))
784 ff = MD_FAILFAST;
Jan Kara5a8948f2017-05-31 09:44:33 +0200785 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_PREFLUSH | REQ_FUA | ff;
NeilBrowna9701a32005-11-08 21:39:34 -0800786
NeilBrown7bfa19f2005-06-21 17:17:28 -0700787 atomic_inc(&mddev->pending_writes);
Mike Christie4e49ea42016-06-05 14:31:41 -0500788 submit_bio(bio);
NeilBrowna9701a32005-11-08 21:39:34 -0800789}
790
NeilBrown46533ff2016-11-18 16:16:11 +1100791int md_super_wait(struct mddev *mddev)
NeilBrowna9701a32005-11-08 21:39:34 -0800792{
Tejun Heoe9c74692010-09-03 11:56:18 +0200793 /* wait for all superblock writes that were scheduled to complete */
NeilBrown1967cd52014-09-09 14:20:28 +1000794 wait_event(mddev->sb_wait, atomic_read(&mddev->pending_writes)==0);
Shaohua Li29530792016-12-08 15:48:19 -0800795 if (test_and_clear_bit(MD_SB_NEED_REWRITE, &mddev->sb_flags))
NeilBrown46533ff2016-11-18 16:16:11 +1100796 return -EAGAIN;
797 return 0;
NeilBrown7bfa19f2005-06-21 17:17:28 -0700798}
799
NeilBrown3cb03002011-10-11 16:45:26 +1100800int sync_page_io(struct md_rdev *rdev, sector_t sector, int size,
Mike Christie796a5cf2016-06-05 14:32:07 -0500801 struct page *page, int op, int op_flags, bool metadata_op)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802{
NeilBrown5a850712017-06-21 09:12:21 +1000803 struct bio *bio = md_bio_alloc_sync(rdev->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 int ret;
805
Christoph Hellwig74d46992017-08-23 19:10:32 +0200806 if (metadata_op && rdev->meta_bdev)
807 bio_set_dev(bio, rdev->meta_bdev);
808 else
809 bio_set_dev(bio, rdev->bdev);
Mike Christie796a5cf2016-06-05 14:32:07 -0500810 bio_set_op_attrs(bio, op, op_flags);
Jonathan Brassowccebd4c2011-01-14 09:14:33 +1100811 if (metadata_op)
Kent Overstreet4f024f32013-10-11 15:44:27 -0700812 bio->bi_iter.bi_sector = sector + rdev->sb_start;
NeilBrown1fdd6fc92012-05-21 09:28:32 +1000813 else if (rdev->mddev->reshape_position != MaxSector &&
814 (rdev->mddev->reshape_backwards ==
815 (sector >= rdev->mddev->reshape_position)))
Kent Overstreet4f024f32013-10-11 15:44:27 -0700816 bio->bi_iter.bi_sector = sector + rdev->new_data_offset;
Jonathan Brassowccebd4c2011-01-14 09:14:33 +1100817 else
Kent Overstreet4f024f32013-10-11 15:44:27 -0700818 bio->bi_iter.bi_sector = sector + rdev->data_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 bio_add_page(bio, page, size, 0);
Mike Christie4e49ea42016-06-05 14:31:41 -0500820
821 submit_bio_wait(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200823 ret = !bio->bi_status;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 bio_put(bio);
825 return ret;
826}
NeilBrowna8745db2006-01-06 00:20:34 -0800827EXPORT_SYMBOL_GPL(sync_page_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828
NeilBrownf72ffdd2014-09-30 14:23:59 +1000829static int read_disk_sb(struct md_rdev *rdev, int size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830{
831 char b[BDEVNAME_SIZE];
NeilBrown403df472014-09-30 15:52:29 +1000832
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 if (rdev->sb_loaded)
834 return 0;
835
Mike Christie796a5cf2016-06-05 14:32:07 -0500836 if (!sync_page_io(rdev, 0, size, rdev->sb_page, REQ_OP_READ, 0, true))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 goto fail;
838 rdev->sb_loaded = 1;
839 return 0;
840
841fail:
NeilBrown9d487392016-11-02 14:16:49 +1100842 pr_err("md: disabled device %s, could not read superblock.\n",
843 bdevname(rdev->bdev,b));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844 return -EINVAL;
845}
846
Amir Goldsteine6fd2092017-05-04 16:26:20 +0300847static int md_uuid_equal(mdp_super_t *sb1, mdp_super_t *sb2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848{
NeilBrownf72ffdd2014-09-30 14:23:59 +1000849 return sb1->set_uuid0 == sb2->set_uuid0 &&
Andre Noll05710462008-07-11 22:02:20 +1000850 sb1->set_uuid1 == sb2->set_uuid1 &&
851 sb1->set_uuid2 == sb2->set_uuid2 &&
852 sb1->set_uuid3 == sb2->set_uuid3;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853}
854
Amir Goldsteine6fd2092017-05-04 16:26:20 +0300855static int md_sb_equal(mdp_super_t *sb1, mdp_super_t *sb2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856{
857 int ret;
858 mdp_super_t *tmp1, *tmp2;
859
860 tmp1 = kmalloc(sizeof(*tmp1),GFP_KERNEL);
861 tmp2 = kmalloc(sizeof(*tmp2),GFP_KERNEL);
862
863 if (!tmp1 || !tmp2) {
864 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865 goto abort;
866 }
867
868 *tmp1 = *sb1;
869 *tmp2 = *sb2;
870
871 /*
872 * nr_disks is not constant
873 */
874 tmp1->nr_disks = 0;
875 tmp2->nr_disks = 0;
876
Andre Nollce0c8e02008-07-11 22:02:20 +1000877 ret = (memcmp(tmp1, tmp2, MD_SB_GENERIC_CONSTANT_WORDS * 4) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878abort:
Jesper Juhl990a8ba2005-06-21 17:17:30 -0700879 kfree(tmp1);
880 kfree(tmp2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 return ret;
882}
883
NeilBrown4d167f02007-05-09 02:35:37 -0700884static u32 md_csum_fold(u32 csum)
885{
886 csum = (csum & 0xffff) + (csum >> 16);
887 return (csum & 0xffff) + (csum >> 16);
888}
889
NeilBrownf72ffdd2014-09-30 14:23:59 +1000890static unsigned int calc_sb_csum(mdp_super_t *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891{
NeilBrown4d167f02007-05-09 02:35:37 -0700892 u64 newcsum = 0;
893 u32 *sb32 = (u32*)sb;
894 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 unsigned int disk_csum, csum;
896
897 disk_csum = sb->sb_csum;
898 sb->sb_csum = 0;
NeilBrown4d167f02007-05-09 02:35:37 -0700899
900 for (i = 0; i < MD_SB_BYTES/4 ; i++)
901 newcsum += sb32[i];
902 csum = (newcsum & 0xffffffff) + (newcsum>>32);
903
NeilBrown4d167f02007-05-09 02:35:37 -0700904#ifdef CONFIG_ALPHA
905 /* This used to use csum_partial, which was wrong for several
906 * reasons including that different results are returned on
907 * different architectures. It isn't critical that we get exactly
908 * the same return value as before (we always csum_fold before
909 * testing, and that removes any differences). However as we
910 * know that csum_partial always returned a 16bit value on
911 * alphas, do a fold to maximise conformity to previous behaviour.
912 */
913 sb->sb_csum = md_csum_fold(disk_csum);
914#else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 sb->sb_csum = disk_csum;
NeilBrown4d167f02007-05-09 02:35:37 -0700916#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 return csum;
918}
919
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920/*
921 * Handle superblock details.
922 * We want to be able to handle multiple superblock formats
923 * so we have a common interface to them all, and an array of
924 * different handlers.
925 * We rely on user-space to write the initial superblock, and support
926 * reading and updating of superblocks.
927 * Interface methods are:
NeilBrown3cb03002011-10-11 16:45:26 +1100928 * int load_super(struct md_rdev *dev, struct md_rdev *refdev, int minor_version)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 * loads and validates a superblock on dev.
930 * if refdev != NULL, compare superblocks on both devices
931 * Return:
932 * 0 - dev has a superblock that is compatible with refdev
933 * 1 - dev has a superblock that is compatible and newer than refdev
934 * so dev should be used as the refdev in future
935 * -EINVAL superblock incompatible or invalid
936 * -othererror e.g. -EIO
937 *
NeilBrownfd01b882011-10-11 16:47:53 +1100938 * int validate_super(struct mddev *mddev, struct md_rdev *dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 * Verify that dev is acceptable into mddev.
940 * The first time, mddev->raid_disks will be 0, and data from
941 * dev should be merged in. Subsequent calls check that dev
942 * is new enough. Return 0 or -EINVAL
943 *
NeilBrownfd01b882011-10-11 16:47:53 +1100944 * void sync_super(struct mddev *mddev, struct md_rdev *dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 * Update the superblock for rdev with data in mddev
946 * This does not write to disc.
947 *
948 */
949
950struct super_type {
Chris Webb0cd17fe2008-06-28 08:31:46 +1000951 char *name;
952 struct module *owner;
NeilBrownc6563a82012-05-21 09:27:00 +1000953 int (*load_super)(struct md_rdev *rdev,
954 struct md_rdev *refdev,
Chris Webb0cd17fe2008-06-28 08:31:46 +1000955 int minor_version);
NeilBrownc6563a82012-05-21 09:27:00 +1000956 int (*validate_super)(struct mddev *mddev,
957 struct md_rdev *rdev);
958 void (*sync_super)(struct mddev *mddev,
959 struct md_rdev *rdev);
NeilBrown3cb03002011-10-11 16:45:26 +1100960 unsigned long long (*rdev_size_change)(struct md_rdev *rdev,
Andre Noll15f4a5f2008-07-21 14:42:12 +1000961 sector_t num_sectors);
NeilBrownc6563a82012-05-21 09:27:00 +1000962 int (*allow_new_offset)(struct md_rdev *rdev,
963 unsigned long long new_offset);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964};
965
966/*
Andre Noll0894cc32009-06-18 08:49:23 +1000967 * Check that the given mddev has no bitmap.
968 *
969 * This function is called from the run method of all personalities that do not
970 * support bitmaps. It prints an error message and returns non-zero if mddev
971 * has a bitmap. Otherwise, it returns 0.
972 *
973 */
NeilBrownfd01b882011-10-11 16:47:53 +1100974int md_check_no_bitmap(struct mddev *mddev)
Andre Noll0894cc32009-06-18 08:49:23 +1000975{
NeilBrownc3d97142009-12-14 12:49:52 +1100976 if (!mddev->bitmap_info.file && !mddev->bitmap_info.offset)
Andre Noll0894cc32009-06-18 08:49:23 +1000977 return 0;
NeilBrown9d487392016-11-02 14:16:49 +1100978 pr_warn("%s: bitmaps are not supported for %s\n",
Andre Noll0894cc32009-06-18 08:49:23 +1000979 mdname(mddev), mddev->pers->name);
980 return 1;
981}
982EXPORT_SYMBOL(md_check_no_bitmap);
983
984/*
NeilBrownf72ffdd2014-09-30 14:23:59 +1000985 * load_super for 0.90.0
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 */
NeilBrown3cb03002011-10-11 16:45:26 +1100987static int super_90_load(struct md_rdev *rdev, struct md_rdev *refdev, int minor_version)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988{
989 char b[BDEVNAME_SIZE], b2[BDEVNAME_SIZE];
990 mdp_super_t *sb;
991 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992
993 /*
Andre Noll0f420352008-07-11 22:02:23 +1000994 * Calculate the position of the superblock (512byte sectors),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995 * it's at the end of the disk.
996 *
997 * It also happens to be a multiple of 4Kb.
998 */
Jonathan Brassow57b2caa2011-01-14 09:14:33 +1100999 rdev->sb_start = calc_dev_sboffset(rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000
NeilBrown0002b272005-09-09 16:23:53 -07001001 ret = read_disk_sb(rdev, MD_SB_BYTES);
NeilBrown9d487392016-11-02 14:16:49 +11001002 if (ret)
1003 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004
1005 ret = -EINVAL;
1006
1007 bdevname(rdev->bdev, b);
Namhyung Kim65a06f062011-07-27 11:00:36 +10001008 sb = page_address(rdev->sb_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009
1010 if (sb->md_magic != MD_SB_MAGIC) {
NeilBrown9d487392016-11-02 14:16:49 +11001011 pr_warn("md: invalid raid superblock magic on %s\n", b);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 goto abort;
1013 }
1014
1015 if (sb->major_version != 0 ||
NeilBrownf6705572006-03-27 01:18:11 -08001016 sb->minor_version < 90 ||
1017 sb->minor_version > 91) {
NeilBrown9d487392016-11-02 14:16:49 +11001018 pr_warn("Bad version number %d.%d on %s\n",
1019 sb->major_version, sb->minor_version, b);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020 goto abort;
1021 }
1022
1023 if (sb->raid_disks <= 0)
1024 goto abort;
1025
NeilBrown4d167f02007-05-09 02:35:37 -07001026 if (md_csum_fold(calc_sb_csum(sb)) != md_csum_fold(sb->sb_csum)) {
NeilBrown9d487392016-11-02 14:16:49 +11001027 pr_warn("md: invalid superblock checksum on %s\n", b);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 goto abort;
1029 }
1030
1031 rdev->preferred_minor = sb->md_minor;
1032 rdev->data_offset = 0;
NeilBrownc6563a82012-05-21 09:27:00 +10001033 rdev->new_data_offset = 0;
NeilBrown0002b272005-09-09 16:23:53 -07001034 rdev->sb_size = MD_SB_BYTES;
NeilBrown9f2f3832011-07-28 11:31:47 +10001035 rdev->badblocks.shift = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036
1037 if (sb->level == LEVEL_MULTIPATH)
1038 rdev->desc_nr = -1;
1039 else
1040 rdev->desc_nr = sb->this_disk.number;
1041
Harvey Harrison9a7b2b02008-04-28 02:15:49 -07001042 if (!refdev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 ret = 1;
Harvey Harrison9a7b2b02008-04-28 02:15:49 -07001044 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 __u64 ev1, ev2;
Namhyung Kim65a06f062011-07-27 11:00:36 +10001046 mdp_super_t *refsb = page_address(refdev->sb_page);
Amir Goldsteine6fd2092017-05-04 16:26:20 +03001047 if (!md_uuid_equal(refsb, sb)) {
NeilBrown9d487392016-11-02 14:16:49 +11001048 pr_warn("md: %s has different UUID to %s\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049 b, bdevname(refdev->bdev,b2));
1050 goto abort;
1051 }
Amir Goldsteine6fd2092017-05-04 16:26:20 +03001052 if (!md_sb_equal(refsb, sb)) {
NeilBrown9d487392016-11-02 14:16:49 +11001053 pr_warn("md: %s has same UUID but different superblock to %s\n",
1054 b, bdevname(refdev->bdev, b2));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055 goto abort;
1056 }
1057 ev1 = md_event(sb);
1058 ev2 = md_event(refsb);
1059 if (ev1 > ev2)
1060 ret = 1;
NeilBrownf72ffdd2014-09-30 14:23:59 +10001061 else
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 ret = 0;
1063 }
NeilBrown8190e752009-06-18 08:48:58 +10001064 rdev->sectors = rdev->sb_start;
NeilBrown667a5312012-08-16 16:46:12 +10001065 /* Limit to 4TB as metadata cannot record more than that.
1066 * (not needed for Linear and RAID0 as metadata doesn't
1067 * record this size)
1068 */
Arnd Bergmann3312c952015-12-21 10:51:01 +11001069 if (IS_ENABLED(CONFIG_LBDAF) && (u64)rdev->sectors >= (2ULL << 32) &&
1070 sb->level >= 1)
1071 rdev->sectors = (sector_t)(2ULL << 32) - 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072
NeilBrown27a7b262011-09-10 17:21:28 +10001073 if (rdev->sectors < ((sector_t)sb->size) * 2 && sb->level >= 1)
NeilBrown2bf071b2006-01-06 00:20:55 -08001074 /* "this cannot possibly happen" ... */
1075 ret = -EINVAL;
1076
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077 abort:
1078 return ret;
1079}
1080
1081/*
1082 * validate_super for 0.90.0
1083 */
NeilBrownfd01b882011-10-11 16:47:53 +11001084static int super_90_validate(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085{
1086 mdp_disk_t *desc;
Namhyung Kim65a06f062011-07-27 11:00:36 +10001087 mdp_super_t *sb = page_address(rdev->sb_page);
NeilBrown07d84d102006-06-26 00:27:56 -07001088 __u64 ev1 = md_event(sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089
NeilBrown41158c72005-06-21 17:17:25 -07001090 rdev->raid_disk = -1;
NeilBrownc5d79ad2008-02-06 01:39:54 -08001091 clear_bit(Faulty, &rdev->flags);
1092 clear_bit(In_sync, &rdev->flags);
NeilBrown8313b8e2013-12-12 10:13:33 +11001093 clear_bit(Bitmap_sync, &rdev->flags);
NeilBrownc5d79ad2008-02-06 01:39:54 -08001094 clear_bit(WriteMostly, &rdev->flags);
NeilBrownc5d79ad2008-02-06 01:39:54 -08001095
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 if (mddev->raid_disks == 0) {
1097 mddev->major_version = 0;
1098 mddev->minor_version = sb->minor_version;
1099 mddev->patch_version = sb->patch_version;
NeilBrowne6910632008-02-06 01:39:51 -08001100 mddev->external = 0;
Andre Noll9d8f0362009-06-18 08:45:01 +10001101 mddev->chunk_sectors = sb->chunk_size >> 9;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 mddev->ctime = sb->ctime;
1103 mddev->utime = sb->utime;
1104 mddev->level = sb->level;
NeilBrownd9d166c2006-01-06 00:20:51 -08001105 mddev->clevel[0] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106 mddev->layout = sb->layout;
1107 mddev->raid_disks = sb->raid_disks;
NeilBrown27a7b262011-09-10 17:21:28 +10001108 mddev->dev_sectors = ((sector_t)sb->size) * 2;
NeilBrown07d84d102006-06-26 00:27:56 -07001109 mddev->events = ev1;
NeilBrownc3d97142009-12-14 12:49:52 +11001110 mddev->bitmap_info.offset = 0;
NeilBrown6409bb02012-05-22 13:55:07 +10001111 mddev->bitmap_info.space = 0;
1112 /* bitmap can use 60 K after the 4K superblocks */
NeilBrownc3d97142009-12-14 12:49:52 +11001113 mddev->bitmap_info.default_offset = MD_SB_BYTES >> 9;
NeilBrown6409bb02012-05-22 13:55:07 +10001114 mddev->bitmap_info.default_space = 64*2 - (MD_SB_BYTES >> 9);
NeilBrown2c810cd2012-05-21 09:27:00 +10001115 mddev->reshape_backwards = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116
NeilBrownf6705572006-03-27 01:18:11 -08001117 if (mddev->minor_version >= 91) {
1118 mddev->reshape_position = sb->reshape_position;
1119 mddev->delta_disks = sb->delta_disks;
1120 mddev->new_level = sb->new_level;
1121 mddev->new_layout = sb->new_layout;
Andre Noll664e7c42009-06-18 08:45:27 +10001122 mddev->new_chunk_sectors = sb->new_chunk >> 9;
NeilBrown2c810cd2012-05-21 09:27:00 +10001123 if (mddev->delta_disks < 0)
1124 mddev->reshape_backwards = 1;
NeilBrownf6705572006-03-27 01:18:11 -08001125 } else {
1126 mddev->reshape_position = MaxSector;
1127 mddev->delta_disks = 0;
1128 mddev->new_level = mddev->level;
1129 mddev->new_layout = mddev->layout;
Andre Noll664e7c42009-06-18 08:45:27 +10001130 mddev->new_chunk_sectors = mddev->chunk_sectors;
NeilBrownf6705572006-03-27 01:18:11 -08001131 }
1132
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 if (sb->state & (1<<MD_SB_CLEAN))
1134 mddev->recovery_cp = MaxSector;
1135 else {
NeilBrownf72ffdd2014-09-30 14:23:59 +10001136 if (sb->events_hi == sb->cp_events_hi &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137 sb->events_lo == sb->cp_events_lo) {
1138 mddev->recovery_cp = sb->recovery_cp;
1139 } else
1140 mddev->recovery_cp = 0;
1141 }
1142
1143 memcpy(mddev->uuid+0, &sb->set_uuid0, 4);
1144 memcpy(mddev->uuid+4, &sb->set_uuid1, 4);
1145 memcpy(mddev->uuid+8, &sb->set_uuid2, 4);
1146 memcpy(mddev->uuid+12,&sb->set_uuid3, 4);
1147
1148 mddev->max_disks = MD_SB_DISKS;
NeilBrowna654b9d82005-06-21 17:17:27 -07001149
1150 if (sb->state & (1<<MD_SB_BITMAP_PRESENT) &&
NeilBrown6409bb02012-05-22 13:55:07 +10001151 mddev->bitmap_info.file == NULL) {
NeilBrownc3d97142009-12-14 12:49:52 +11001152 mddev->bitmap_info.offset =
1153 mddev->bitmap_info.default_offset;
NeilBrown6409bb02012-05-22 13:55:07 +10001154 mddev->bitmap_info.space =
Dave Jonesc9ad0202013-08-19 22:26:32 -04001155 mddev->bitmap_info.default_space;
NeilBrown6409bb02012-05-22 13:55:07 +10001156 }
NeilBrowna654b9d82005-06-21 17:17:27 -07001157
NeilBrown41158c72005-06-21 17:17:25 -07001158 } else if (mddev->pers == NULL) {
NeilBrownbe6800a2010-05-18 10:17:09 +10001159 /* Insist on good event counter while assembling, except
1160 * for spares (which don't need an event count) */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 ++ev1;
NeilBrownbe6800a2010-05-18 10:17:09 +10001162 if (sb->disks[rdev->desc_nr].state & (
1163 (1<<MD_DISK_SYNC) | (1 << MD_DISK_ACTIVE)))
NeilBrownf72ffdd2014-09-30 14:23:59 +10001164 if (ev1 < mddev->events)
NeilBrownbe6800a2010-05-18 10:17:09 +10001165 return -EINVAL;
NeilBrown41158c72005-06-21 17:17:25 -07001166 } else if (mddev->bitmap) {
1167 /* if adding to array with a bitmap, then we can accept an
1168 * older device ... but not too old.
1169 */
NeilBrown41158c72005-06-21 17:17:25 -07001170 if (ev1 < mddev->bitmap->events_cleared)
1171 return 0;
NeilBrown8313b8e2013-12-12 10:13:33 +11001172 if (ev1 < mddev->events)
1173 set_bit(Bitmap_sync, &rdev->flags);
NeilBrown07d84d102006-06-26 00:27:56 -07001174 } else {
1175 if (ev1 < mddev->events)
1176 /* just a hot-add of a new device, leave raid_disk at -1 */
1177 return 0;
1178 }
NeilBrown41158c72005-06-21 17:17:25 -07001179
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 if (mddev->level != LEVEL_MULTIPATH) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 desc = sb->disks + rdev->desc_nr;
1182
1183 if (desc->state & (1<<MD_DISK_FAULTY))
NeilBrownb2d444d2005-11-08 21:39:31 -08001184 set_bit(Faulty, &rdev->flags);
NeilBrown7c7546c2006-06-26 00:27:41 -07001185 else if (desc->state & (1<<MD_DISK_SYNC) /* &&
1186 desc->raid_disk < mddev->raid_disks */) {
NeilBrownb2d444d2005-11-08 21:39:31 -08001187 set_bit(In_sync, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 rdev->raid_disk = desc->raid_disk;
NeilBrownf4667222013-12-09 12:04:56 +11001189 rdev->saved_raid_disk = desc->raid_disk;
NeilBrown0261cd9f2009-11-13 17:40:48 +11001190 } else if (desc->state & (1<<MD_DISK_ACTIVE)) {
1191 /* active but not in sync implies recovery up to
1192 * reshape position. We don't know exactly where
1193 * that is, so set to zero for now */
1194 if (mddev->minor_version >= 91) {
1195 rdev->recovery_offset = 0;
1196 rdev->raid_disk = desc->raid_disk;
1197 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 }
NeilBrown8ddf9ef2005-09-09 16:23:45 -07001199 if (desc->state & (1<<MD_DISK_WRITEMOSTLY))
1200 set_bit(WriteMostly, &rdev->flags);
NeilBrown688834e2016-11-18 16:16:11 +11001201 if (desc->state & (1<<MD_DISK_FAILFAST))
1202 set_bit(FailFast, &rdev->flags);
NeilBrown41158c72005-06-21 17:17:25 -07001203 } else /* MULTIPATH are always insync */
NeilBrownb2d444d2005-11-08 21:39:31 -08001204 set_bit(In_sync, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 return 0;
1206}
1207
1208/*
1209 * sync_super for 0.90.0
1210 */
NeilBrownfd01b882011-10-11 16:47:53 +11001211static void super_90_sync(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212{
1213 mdp_super_t *sb;
NeilBrown3cb03002011-10-11 16:45:26 +11001214 struct md_rdev *rdev2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 int next_spare = mddev->raid_disks;
NeilBrown19133a42005-11-08 21:39:35 -08001216
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217 /* make rdev->sb match mddev data..
1218 *
1219 * 1/ zero out disks
1220 * 2/ Add info for each disk, keeping track of highest desc_nr (next_spare);
1221 * 3/ any empty disks < next_spare become removed
1222 *
1223 * disks[0] gets initialised to REMOVED because
1224 * we cannot be sure from other fields if it has
1225 * been initialised or not.
1226 */
1227 int i;
1228 int active=0, working=0,failed=0,spare=0,nr_disks=0;
1229
NeilBrown61181562005-09-09 16:24:02 -07001230 rdev->sb_size = MD_SB_BYTES;
1231
Namhyung Kim65a06f062011-07-27 11:00:36 +10001232 sb = page_address(rdev->sb_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233
1234 memset(sb, 0, sizeof(*sb));
1235
1236 sb->md_magic = MD_SB_MAGIC;
1237 sb->major_version = mddev->major_version;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238 sb->patch_version = mddev->patch_version;
1239 sb->gvalid_words = 0; /* ignored */
1240 memcpy(&sb->set_uuid0, mddev->uuid+0, 4);
1241 memcpy(&sb->set_uuid1, mddev->uuid+4, 4);
1242 memcpy(&sb->set_uuid2, mddev->uuid+8, 4);
1243 memcpy(&sb->set_uuid3, mddev->uuid+12,4);
1244
Deepa Dinamani9ebc6ef2015-12-21 10:51:01 +11001245 sb->ctime = clamp_t(time64_t, mddev->ctime, 0, U32_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 sb->level = mddev->level;
Andre Noll58c0fed2009-03-31 14:33:13 +11001247 sb->size = mddev->dev_sectors / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 sb->raid_disks = mddev->raid_disks;
1249 sb->md_minor = mddev->md_minor;
NeilBrowne6910632008-02-06 01:39:51 -08001250 sb->not_persistent = 0;
Deepa Dinamani9ebc6ef2015-12-21 10:51:01 +11001251 sb->utime = clamp_t(time64_t, mddev->utime, 0, U32_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 sb->state = 0;
1253 sb->events_hi = (mddev->events>>32);
1254 sb->events_lo = (u32)mddev->events;
1255
NeilBrownf6705572006-03-27 01:18:11 -08001256 if (mddev->reshape_position == MaxSector)
1257 sb->minor_version = 90;
1258 else {
1259 sb->minor_version = 91;
1260 sb->reshape_position = mddev->reshape_position;
1261 sb->new_level = mddev->new_level;
1262 sb->delta_disks = mddev->delta_disks;
1263 sb->new_layout = mddev->new_layout;
Andre Noll664e7c42009-06-18 08:45:27 +10001264 sb->new_chunk = mddev->new_chunk_sectors << 9;
NeilBrownf6705572006-03-27 01:18:11 -08001265 }
1266 mddev->minor_version = sb->minor_version;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 if (mddev->in_sync)
1268 {
1269 sb->recovery_cp = mddev->recovery_cp;
1270 sb->cp_events_hi = (mddev->events>>32);
1271 sb->cp_events_lo = (u32)mddev->events;
1272 if (mddev->recovery_cp == MaxSector)
1273 sb->state = (1<< MD_SB_CLEAN);
1274 } else
1275 sb->recovery_cp = 0;
1276
1277 sb->layout = mddev->layout;
Andre Noll9d8f0362009-06-18 08:45:01 +10001278 sb->chunk_size = mddev->chunk_sectors << 9;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279
NeilBrownc3d97142009-12-14 12:49:52 +11001280 if (mddev->bitmap && mddev->bitmap_info.file == NULL)
NeilBrowna654b9d82005-06-21 17:17:27 -07001281 sb->state |= (1<<MD_SB_BITMAP_PRESENT);
1282
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283 sb->disks[0].state = (1<<MD_DISK_REMOVED);
NeilBrowndafb20f2012-03-19 12:46:39 +11001284 rdev_for_each(rdev2, mddev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 mdp_disk_t *d;
NeilBrown86e6ffd2005-11-08 21:39:24 -08001286 int desc_nr;
NeilBrown0261cd9f2009-11-13 17:40:48 +11001287 int is_active = test_bit(In_sync, &rdev2->flags);
1288
1289 if (rdev2->raid_disk >= 0 &&
1290 sb->minor_version >= 91)
1291 /* we have nowhere to store the recovery_offset,
1292 * but if it is not below the reshape_position,
1293 * we can piggy-back on that.
1294 */
1295 is_active = 1;
1296 if (rdev2->raid_disk < 0 ||
1297 test_bit(Faulty, &rdev2->flags))
1298 is_active = 0;
1299 if (is_active)
NeilBrown86e6ffd2005-11-08 21:39:24 -08001300 desc_nr = rdev2->raid_disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 else
NeilBrown86e6ffd2005-11-08 21:39:24 -08001302 desc_nr = next_spare++;
NeilBrown19133a42005-11-08 21:39:35 -08001303 rdev2->desc_nr = desc_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 d = &sb->disks[rdev2->desc_nr];
1305 nr_disks++;
1306 d->number = rdev2->desc_nr;
1307 d->major = MAJOR(rdev2->bdev->bd_dev);
1308 d->minor = MINOR(rdev2->bdev->bd_dev);
NeilBrown0261cd9f2009-11-13 17:40:48 +11001309 if (is_active)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 d->raid_disk = rdev2->raid_disk;
1311 else
1312 d->raid_disk = rdev2->desc_nr; /* compatibility */
NeilBrown1be78922006-03-27 01:18:03 -08001313 if (test_bit(Faulty, &rdev2->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 d->state = (1<<MD_DISK_FAULTY);
NeilBrown0261cd9f2009-11-13 17:40:48 +11001315 else if (is_active) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316 d->state = (1<<MD_DISK_ACTIVE);
NeilBrown0261cd9f2009-11-13 17:40:48 +11001317 if (test_bit(In_sync, &rdev2->flags))
1318 d->state |= (1<<MD_DISK_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319 active++;
1320 working++;
1321 } else {
1322 d->state = 0;
1323 spare++;
1324 working++;
1325 }
NeilBrown8ddf9ef2005-09-09 16:23:45 -07001326 if (test_bit(WriteMostly, &rdev2->flags))
1327 d->state |= (1<<MD_DISK_WRITEMOSTLY);
NeilBrown688834e2016-11-18 16:16:11 +11001328 if (test_bit(FailFast, &rdev2->flags))
1329 d->state |= (1<<MD_DISK_FAILFAST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331 /* now set the "removed" and "faulty" bits on any missing devices */
1332 for (i=0 ; i < mddev->raid_disks ; i++) {
1333 mdp_disk_t *d = &sb->disks[i];
1334 if (d->state == 0 && d->number == 0) {
1335 d->number = i;
1336 d->raid_disk = i;
1337 d->state = (1<<MD_DISK_REMOVED);
1338 d->state |= (1<<MD_DISK_FAULTY);
1339 failed++;
1340 }
1341 }
1342 sb->nr_disks = nr_disks;
1343 sb->active_disks = active;
1344 sb->working_disks = working;
1345 sb->failed_disks = failed;
1346 sb->spare_disks = spare;
1347
1348 sb->this_disk = sb->disks[rdev->desc_nr];
1349 sb->sb_csum = calc_sb_csum(sb);
1350}
1351
1352/*
Chris Webb0cd17fe2008-06-28 08:31:46 +10001353 * rdev_size_change for 0.90.0
1354 */
1355static unsigned long long
NeilBrown3cb03002011-10-11 16:45:26 +11001356super_90_rdev_size_change(struct md_rdev *rdev, sector_t num_sectors)
Chris Webb0cd17fe2008-06-28 08:31:46 +10001357{
Andre Noll58c0fed2009-03-31 14:33:13 +11001358 if (num_sectors && num_sectors < rdev->mddev->dev_sectors)
Chris Webb0cd17fe2008-06-28 08:31:46 +10001359 return 0; /* component must fit device */
NeilBrownc3d97142009-12-14 12:49:52 +11001360 if (rdev->mddev->bitmap_info.offset)
Chris Webb0cd17fe2008-06-28 08:31:46 +10001361 return 0; /* can't move bitmap */
Jonathan Brassow57b2caa2011-01-14 09:14:33 +11001362 rdev->sb_start = calc_dev_sboffset(rdev);
Andre Noll15f4a5f2008-07-21 14:42:12 +10001363 if (!num_sectors || num_sectors > rdev->sb_start)
1364 num_sectors = rdev->sb_start;
NeilBrown27a7b262011-09-10 17:21:28 +10001365 /* Limit to 4TB as metadata cannot record more than that.
1366 * 4TB == 2^32 KB, or 2*2^32 sectors.
1367 */
Arnd Bergmann3312c952015-12-21 10:51:01 +11001368 if (IS_ENABLED(CONFIG_LBDAF) && (u64)num_sectors >= (2ULL << 32) &&
1369 rdev->mddev->level >= 1)
1370 num_sectors = (sector_t)(2ULL << 32) - 2;
NeilBrown46533ff2016-11-18 16:16:11 +11001371 do {
1372 md_super_write(rdev->mddev, rdev, rdev->sb_start, rdev->sb_size,
Chris Webb0cd17fe2008-06-28 08:31:46 +10001373 rdev->sb_page);
NeilBrown46533ff2016-11-18 16:16:11 +11001374 } while (md_super_wait(rdev->mddev) < 0);
Justin Maggardc26a44e2010-11-24 16:36:17 +11001375 return num_sectors;
Chris Webb0cd17fe2008-06-28 08:31:46 +10001376}
1377
NeilBrownc6563a82012-05-21 09:27:00 +10001378static int
1379super_90_allow_new_offset(struct md_rdev *rdev, unsigned long long new_offset)
1380{
1381 /* non-zero offset changes not possible with v0.90 */
1382 return new_offset == 0;
1383}
Chris Webb0cd17fe2008-06-28 08:31:46 +10001384
1385/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 * version 1 superblock
1387 */
1388
NeilBrownf72ffdd2014-09-30 14:23:59 +10001389static __le32 calc_sb_1_csum(struct mdp_superblock_1 *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390{
NeilBrown1c05b4b2006-10-21 10:24:08 -07001391 __le32 disk_csum;
1392 u32 csum;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393 unsigned long long newcsum;
1394 int size = 256 + le32_to_cpu(sb->max_dev)*2;
NeilBrown1c05b4b2006-10-21 10:24:08 -07001395 __le32 *isuper = (__le32*)sb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396
1397 disk_csum = sb->sb_csum;
1398 sb->sb_csum = 0;
1399 newcsum = 0;
NeilBrown1f3c9902012-12-11 13:09:00 +11001400 for (; size >= 4; size -= 4)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401 newcsum += le32_to_cpu(*isuper++);
1402
1403 if (size == 2)
NeilBrown1c05b4b2006-10-21 10:24:08 -07001404 newcsum += le16_to_cpu(*(__le16*) isuper);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405
1406 csum = (newcsum & 0xffffffff) + (newcsum >> 32);
1407 sb->sb_csum = disk_csum;
1408 return cpu_to_le32(csum);
1409}
1410
NeilBrown3cb03002011-10-11 16:45:26 +11001411static int super_1_load(struct md_rdev *rdev, struct md_rdev *refdev, int minor_version)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412{
1413 struct mdp_superblock_1 *sb;
1414 int ret;
Andre Noll0f420352008-07-11 22:02:23 +10001415 sector_t sb_start;
NeilBrownc6563a82012-05-21 09:27:00 +10001416 sector_t sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417 char b[BDEVNAME_SIZE], b2[BDEVNAME_SIZE];
NeilBrown0002b272005-09-09 16:23:53 -07001418 int bmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419
1420 /*
Andre Noll0f420352008-07-11 22:02:23 +10001421 * Calculate the position of the superblock in 512byte sectors.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 * It is always aligned to a 4K boundary and
1423 * depeding on minor_version, it can be:
1424 * 0: At least 8K, but less than 12K, from end of device
1425 * 1: At start of device
1426 * 2: 4K from start of device.
1427 */
1428 switch(minor_version) {
1429 case 0:
Mike Snitzer77304d22010-11-08 14:39:12 +01001430 sb_start = i_size_read(rdev->bdev->bd_inode) >> 9;
Andre Noll0f420352008-07-11 22:02:23 +10001431 sb_start -= 8*2;
1432 sb_start &= ~(sector_t)(4*2-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 break;
1434 case 1:
Andre Noll0f420352008-07-11 22:02:23 +10001435 sb_start = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 break;
1437 case 2:
Andre Noll0f420352008-07-11 22:02:23 +10001438 sb_start = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 break;
1440 default:
1441 return -EINVAL;
1442 }
Andre Noll0f420352008-07-11 22:02:23 +10001443 rdev->sb_start = sb_start;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444
NeilBrown0002b272005-09-09 16:23:53 -07001445 /* superblock is rarely larger than 1K, but it can be larger,
1446 * and it is safe to read 4k, so we do that
1447 */
1448 ret = read_disk_sb(rdev, 4096);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449 if (ret) return ret;
1450
Namhyung Kim65a06f062011-07-27 11:00:36 +10001451 sb = page_address(rdev->sb_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452
1453 if (sb->magic != cpu_to_le32(MD_SB_MAGIC) ||
1454 sb->major_version != cpu_to_le32(1) ||
1455 le32_to_cpu(sb->max_dev) > (4096-256)/2 ||
Andre Noll0f420352008-07-11 22:02:23 +10001456 le64_to_cpu(sb->super_offset) != rdev->sb_start ||
NeilBrown71c08052005-09-09 16:23:51 -07001457 (le32_to_cpu(sb->feature_map) & ~MD_FEATURE_ALL) != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458 return -EINVAL;
1459
1460 if (calc_sb_1_csum(sb) != sb->sb_csum) {
NeilBrown9d487392016-11-02 14:16:49 +11001461 pr_warn("md: invalid superblock checksum on %s\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462 bdevname(rdev->bdev,b));
1463 return -EINVAL;
1464 }
1465 if (le64_to_cpu(sb->data_size) < 10) {
NeilBrown9d487392016-11-02 14:16:49 +11001466 pr_warn("md: data_size too small on %s\n",
1467 bdevname(rdev->bdev,b));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 return -EINVAL;
1469 }
NeilBrownc6563a82012-05-21 09:27:00 +10001470 if (sb->pad0 ||
1471 sb->pad3[0] ||
1472 memcmp(sb->pad3, sb->pad3+1, sizeof(sb->pad3) - sizeof(sb->pad3[1])))
1473 /* Some padding is non-zero, might be a new feature */
1474 return -EINVAL;
NeilBrowne11e93f2007-05-09 02:35:36 -07001475
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476 rdev->preferred_minor = 0xffff;
1477 rdev->data_offset = le64_to_cpu(sb->data_offset);
NeilBrownc6563a82012-05-21 09:27:00 +10001478 rdev->new_data_offset = rdev->data_offset;
1479 if ((le32_to_cpu(sb->feature_map) & MD_FEATURE_RESHAPE_ACTIVE) &&
1480 (le32_to_cpu(sb->feature_map) & MD_FEATURE_NEW_OFFSET))
1481 rdev->new_data_offset += (s32)le32_to_cpu(sb->new_offset);
NeilBrown4dbcdc72006-01-06 00:20:52 -08001482 atomic_set(&rdev->corrected_errors, le32_to_cpu(sb->cnt_corrected_read));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483
NeilBrown0002b272005-09-09 16:23:53 -07001484 rdev->sb_size = le32_to_cpu(sb->max_dev) * 2 + 256;
Martin K. Petersene1defc42009-05-22 17:17:49 -04001485 bmask = queue_logical_block_size(rdev->bdev->bd_disk->queue)-1;
NeilBrown0002b272005-09-09 16:23:53 -07001486 if (rdev->sb_size & bmask)
NeilBrowna1801f82008-03-04 14:29:31 -08001487 rdev->sb_size = (rdev->sb_size | bmask) + 1;
1488
1489 if (minor_version
Andre Noll0f420352008-07-11 22:02:23 +10001490 && rdev->data_offset < sb_start + (rdev->sb_size/512))
NeilBrowna1801f82008-03-04 14:29:31 -08001491 return -EINVAL;
NeilBrownc6563a82012-05-21 09:27:00 +10001492 if (minor_version
1493 && rdev->new_data_offset < sb_start + (rdev->sb_size/512))
1494 return -EINVAL;
NeilBrown0002b272005-09-09 16:23:53 -07001495
NeilBrown31b65a02006-07-10 04:44:14 -07001496 if (sb->level == cpu_to_le32(LEVEL_MULTIPATH))
1497 rdev->desc_nr = -1;
1498 else
1499 rdev->desc_nr = le32_to_cpu(sb->dev_number);
1500
NeilBrown2699b672011-07-28 11:31:47 +10001501 if (!rdev->bb_page) {
1502 rdev->bb_page = alloc_page(GFP_KERNEL);
1503 if (!rdev->bb_page)
1504 return -ENOMEM;
1505 }
1506 if ((le32_to_cpu(sb->feature_map) & MD_FEATURE_BAD_BLOCKS) &&
1507 rdev->badblocks.count == 0) {
1508 /* need to load the bad block list.
1509 * Currently we limit it to one page.
1510 */
1511 s32 offset;
1512 sector_t bb_sector;
1513 u64 *bbp;
1514 int i;
1515 int sectors = le16_to_cpu(sb->bblog_size);
1516 if (sectors > (PAGE_SIZE / 512))
1517 return -EINVAL;
1518 offset = le32_to_cpu(sb->bblog_offset);
1519 if (offset == 0)
1520 return -EINVAL;
1521 bb_sector = (long long)offset;
1522 if (!sync_page_io(rdev, bb_sector, sectors << 9,
Mike Christie796a5cf2016-06-05 14:32:07 -05001523 rdev->bb_page, REQ_OP_READ, 0, true))
NeilBrown2699b672011-07-28 11:31:47 +10001524 return -EIO;
1525 bbp = (u64 *)page_address(rdev->bb_page);
1526 rdev->badblocks.shift = sb->bblog_shift;
1527 for (i = 0 ; i < (sectors << (9-3)) ; i++, bbp++) {
1528 u64 bb = le64_to_cpu(*bbp);
1529 int count = bb & (0x3ff);
1530 u64 sector = bb >> 10;
1531 sector <<= sb->bblog_shift;
1532 count <<= sb->bblog_shift;
1533 if (bb + 1 == 0)
1534 break;
Vishal Vermafc974ee2015-12-24 19:20:34 -07001535 if (badblocks_set(&rdev->badblocks, sector, count, 1))
NeilBrown2699b672011-07-28 11:31:47 +10001536 return -EINVAL;
1537 }
NeilBrown486adf72013-04-24 11:42:44 +10001538 } else if (sb->bblog_offset != 0)
1539 rdev->badblocks.shift = 0;
NeilBrown2699b672011-07-28 11:31:47 +10001540
Pawel Baldysiakddc08822017-08-16 17:13:45 +02001541 if ((le32_to_cpu(sb->feature_map) &
1542 (MD_FEATURE_PPL | MD_FEATURE_MULTIPLE_PPLS))) {
Artur Paszkiewiczea0213e2017-03-09 09:59:57 +01001543 rdev->ppl.offset = (__s16)le16_to_cpu(sb->ppl.offset);
1544 rdev->ppl.size = le16_to_cpu(sb->ppl.size);
1545 rdev->ppl.sector = rdev->sb_start + rdev->ppl.offset;
1546 }
1547
Harvey Harrison9a7b2b02008-04-28 02:15:49 -07001548 if (!refdev) {
NeilBrown8ed75462006-02-03 03:03:41 -08001549 ret = 1;
Harvey Harrison9a7b2b02008-04-28 02:15:49 -07001550 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551 __u64 ev1, ev2;
Namhyung Kim65a06f062011-07-27 11:00:36 +10001552 struct mdp_superblock_1 *refsb = page_address(refdev->sb_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553
1554 if (memcmp(sb->set_uuid, refsb->set_uuid, 16) != 0 ||
1555 sb->level != refsb->level ||
1556 sb->layout != refsb->layout ||
1557 sb->chunksize != refsb->chunksize) {
NeilBrown9d487392016-11-02 14:16:49 +11001558 pr_warn("md: %s has strangely different superblock to %s\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559 bdevname(rdev->bdev,b),
1560 bdevname(refdev->bdev,b2));
1561 return -EINVAL;
1562 }
1563 ev1 = le64_to_cpu(sb->events);
1564 ev2 = le64_to_cpu(refsb->events);
1565
1566 if (ev1 > ev2)
NeilBrown8ed75462006-02-03 03:03:41 -08001567 ret = 1;
1568 else
1569 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 }
NeilBrownc6563a82012-05-21 09:27:00 +10001571 if (minor_version) {
1572 sectors = (i_size_read(rdev->bdev->bd_inode) >> 9);
1573 sectors -= rdev->data_offset;
1574 } else
1575 sectors = rdev->sb_start;
1576 if (sectors < le64_to_cpu(sb->data_size))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 return -EINVAL;
Andre Nolldd8ac332009-03-31 14:33:13 +11001578 rdev->sectors = le64_to_cpu(sb->data_size);
NeilBrown8ed75462006-02-03 03:03:41 -08001579 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580}
1581
NeilBrownfd01b882011-10-11 16:47:53 +11001582static int super_1_validate(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583{
Namhyung Kim65a06f062011-07-27 11:00:36 +10001584 struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
NeilBrown07d84d102006-06-26 00:27:56 -07001585 __u64 ev1 = le64_to_cpu(sb->events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586
NeilBrown41158c72005-06-21 17:17:25 -07001587 rdev->raid_disk = -1;
NeilBrownc5d79ad2008-02-06 01:39:54 -08001588 clear_bit(Faulty, &rdev->flags);
1589 clear_bit(In_sync, &rdev->flags);
NeilBrown8313b8e2013-12-12 10:13:33 +11001590 clear_bit(Bitmap_sync, &rdev->flags);
NeilBrownc5d79ad2008-02-06 01:39:54 -08001591 clear_bit(WriteMostly, &rdev->flags);
NeilBrownc5d79ad2008-02-06 01:39:54 -08001592
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 if (mddev->raid_disks == 0) {
1594 mddev->major_version = 1;
1595 mddev->patch_version = 0;
NeilBrowne6910632008-02-06 01:39:51 -08001596 mddev->external = 0;
Andre Noll9d8f0362009-06-18 08:45:01 +10001597 mddev->chunk_sectors = le32_to_cpu(sb->chunksize);
Deepa Dinamani9ebc6ef2015-12-21 10:51:01 +11001598 mddev->ctime = le64_to_cpu(sb->ctime);
1599 mddev->utime = le64_to_cpu(sb->utime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600 mddev->level = le32_to_cpu(sb->level);
NeilBrownd9d166c2006-01-06 00:20:51 -08001601 mddev->clevel[0] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 mddev->layout = le32_to_cpu(sb->layout);
1603 mddev->raid_disks = le32_to_cpu(sb->raid_disks);
Andre Noll58c0fed2009-03-31 14:33:13 +11001604 mddev->dev_sectors = le64_to_cpu(sb->size);
NeilBrown07d84d102006-06-26 00:27:56 -07001605 mddev->events = ev1;
NeilBrownc3d97142009-12-14 12:49:52 +11001606 mddev->bitmap_info.offset = 0;
NeilBrown6409bb02012-05-22 13:55:07 +10001607 mddev->bitmap_info.space = 0;
1608 /* Default location for bitmap is 1K after superblock
1609 * using 3K - total of 4K
1610 */
NeilBrownc3d97142009-12-14 12:49:52 +11001611 mddev->bitmap_info.default_offset = 1024 >> 9;
NeilBrown6409bb02012-05-22 13:55:07 +10001612 mddev->bitmap_info.default_space = (4096-1024) >> 9;
NeilBrown2c810cd2012-05-21 09:27:00 +10001613 mddev->reshape_backwards = 0;
1614
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615 mddev->recovery_cp = le64_to_cpu(sb->resync_offset);
1616 memcpy(mddev->uuid, sb->set_uuid, 16);
1617
1618 mddev->max_disks = (4096-256)/2;
NeilBrowna654b9d82005-06-21 17:17:27 -07001619
NeilBrown71c08052005-09-09 16:23:51 -07001620 if ((le32_to_cpu(sb->feature_map) & MD_FEATURE_BITMAP_OFFSET) &&
NeilBrown6409bb02012-05-22 13:55:07 +10001621 mddev->bitmap_info.file == NULL) {
NeilBrownc3d97142009-12-14 12:49:52 +11001622 mddev->bitmap_info.offset =
1623 (__s32)le32_to_cpu(sb->bitmap_offset);
NeilBrown6409bb02012-05-22 13:55:07 +10001624 /* Metadata doesn't record how much space is available.
1625 * For 1.0, we assume we can use up to the superblock
1626 * if before, else to 4K beyond superblock.
1627 * For others, assume no change is possible.
1628 */
1629 if (mddev->minor_version > 0)
1630 mddev->bitmap_info.space = 0;
1631 else if (mddev->bitmap_info.offset > 0)
1632 mddev->bitmap_info.space =
1633 8 - mddev->bitmap_info.offset;
1634 else
1635 mddev->bitmap_info.space =
1636 -mddev->bitmap_info.offset;
1637 }
NeilBrowne11e93f2007-05-09 02:35:36 -07001638
NeilBrownf6705572006-03-27 01:18:11 -08001639 if ((le32_to_cpu(sb->feature_map) & MD_FEATURE_RESHAPE_ACTIVE)) {
1640 mddev->reshape_position = le64_to_cpu(sb->reshape_position);
1641 mddev->delta_disks = le32_to_cpu(sb->delta_disks);
1642 mddev->new_level = le32_to_cpu(sb->new_level);
1643 mddev->new_layout = le32_to_cpu(sb->new_layout);
Andre Noll664e7c42009-06-18 08:45:27 +10001644 mddev->new_chunk_sectors = le32_to_cpu(sb->new_chunk);
NeilBrown2c810cd2012-05-21 09:27:00 +10001645 if (mddev->delta_disks < 0 ||
1646 (mddev->delta_disks == 0 &&
1647 (le32_to_cpu(sb->feature_map)
1648 & MD_FEATURE_RESHAPE_BACKWARDS)))
1649 mddev->reshape_backwards = 1;
NeilBrownf6705572006-03-27 01:18:11 -08001650 } else {
1651 mddev->reshape_position = MaxSector;
1652 mddev->delta_disks = 0;
1653 mddev->new_level = mddev->level;
1654 mddev->new_layout = mddev->layout;
Andre Noll664e7c42009-06-18 08:45:27 +10001655 mddev->new_chunk_sectors = mddev->chunk_sectors;
NeilBrownf6705572006-03-27 01:18:11 -08001656 }
1657
Song Liu486b0f72016-08-19 15:34:01 -07001658 if (le32_to_cpu(sb->feature_map) & MD_FEATURE_JOURNAL)
Shaohua Lia62ab492016-01-06 14:37:13 -08001659 set_bit(MD_HAS_JOURNAL, &mddev->flags);
Artur Paszkiewiczea0213e2017-03-09 09:59:57 +01001660
Pawel Baldysiakddc08822017-08-16 17:13:45 +02001661 if (le32_to_cpu(sb->feature_map) &
1662 (MD_FEATURE_PPL | MD_FEATURE_MULTIPLE_PPLS)) {
Artur Paszkiewiczea0213e2017-03-09 09:59:57 +01001663 if (le32_to_cpu(sb->feature_map) &
1664 (MD_FEATURE_BITMAP_OFFSET | MD_FEATURE_JOURNAL))
1665 return -EINVAL;
Pawel Baldysiakddc08822017-08-16 17:13:45 +02001666 if ((le32_to_cpu(sb->feature_map) & MD_FEATURE_PPL) &&
1667 (le32_to_cpu(sb->feature_map) &
1668 MD_FEATURE_MULTIPLE_PPLS))
1669 return -EINVAL;
Artur Paszkiewiczea0213e2017-03-09 09:59:57 +01001670 set_bit(MD_HAS_PPL, &mddev->flags);
1671 }
NeilBrown41158c72005-06-21 17:17:25 -07001672 } else if (mddev->pers == NULL) {
NeilBrownbe6800a2010-05-18 10:17:09 +10001673 /* Insist of good event counter while assembling, except for
1674 * spares (which don't need an event count) */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675 ++ev1;
NeilBrownbe6800a2010-05-18 10:17:09 +10001676 if (rdev->desc_nr >= 0 &&
1677 rdev->desc_nr < le32_to_cpu(sb->max_dev) &&
Song Liua3dfbda2015-10-08 21:54:11 -07001678 (le16_to_cpu(sb->dev_roles[rdev->desc_nr]) < MD_DISK_ROLE_MAX ||
1679 le16_to_cpu(sb->dev_roles[rdev->desc_nr]) == MD_DISK_ROLE_JOURNAL))
NeilBrownbe6800a2010-05-18 10:17:09 +10001680 if (ev1 < mddev->events)
1681 return -EINVAL;
NeilBrown41158c72005-06-21 17:17:25 -07001682 } else if (mddev->bitmap) {
1683 /* If adding to array with a bitmap, then we can accept an
1684 * older device, but not too old.
1685 */
NeilBrown41158c72005-06-21 17:17:25 -07001686 if (ev1 < mddev->bitmap->events_cleared)
1687 return 0;
NeilBrown8313b8e2013-12-12 10:13:33 +11001688 if (ev1 < mddev->events)
1689 set_bit(Bitmap_sync, &rdev->flags);
NeilBrown07d84d102006-06-26 00:27:56 -07001690 } else {
1691 if (ev1 < mddev->events)
1692 /* just a hot-add of a new device, leave raid_disk at -1 */
1693 return 0;
1694 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695 if (mddev->level != LEVEL_MULTIPATH) {
1696 int role;
NeilBrown3673f302009-08-03 10:59:56 +10001697 if (rdev->desc_nr < 0 ||
1698 rdev->desc_nr >= le32_to_cpu(sb->max_dev)) {
Song Liuc4d4c912015-08-13 14:31:54 -07001699 role = MD_DISK_ROLE_SPARE;
NeilBrown3673f302009-08-03 10:59:56 +10001700 rdev->desc_nr = -1;
1701 } else
1702 role = le16_to_cpu(sb->dev_roles[rdev->desc_nr]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 switch(role) {
Song Liuc4d4c912015-08-13 14:31:54 -07001704 case MD_DISK_ROLE_SPARE: /* spare */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 break;
Song Liuc4d4c912015-08-13 14:31:54 -07001706 case MD_DISK_ROLE_FAULTY: /* faulty */
NeilBrownb2d444d2005-11-08 21:39:31 -08001707 set_bit(Faulty, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 break;
Song Liubac624f2015-08-13 14:31:55 -07001709 case MD_DISK_ROLE_JOURNAL: /* journal device */
1710 if (!(le32_to_cpu(sb->feature_map) & MD_FEATURE_JOURNAL)) {
1711 /* journal device without journal feature */
NeilBrown9d487392016-11-02 14:16:49 +11001712 pr_warn("md: journal device provided without journal feature, ignoring the device\n");
Song Liubac624f2015-08-13 14:31:55 -07001713 return -EINVAL;
1714 }
1715 set_bit(Journal, &rdev->flags);
Shaohua Li3069aa82015-08-13 14:31:56 -07001716 rdev->journal_tail = le64_to_cpu(sb->journal_tail);
Shaohua Li9b156032015-12-18 15:19:16 +11001717 rdev->raid_disk = 0;
Song Liubac624f2015-08-13 14:31:55 -07001718 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 default:
NeilBrownf4667222013-12-09 12:04:56 +11001720 rdev->saved_raid_disk = role;
NeilBrown5fd6c1d2006-06-26 00:27:40 -07001721 if ((le32_to_cpu(sb->feature_map) &
NeilBrownf4667222013-12-09 12:04:56 +11001722 MD_FEATURE_RECOVERY_OFFSET)) {
NeilBrown5fd6c1d2006-06-26 00:27:40 -07001723 rdev->recovery_offset = le64_to_cpu(sb->recovery_offset);
NeilBrownf4667222013-12-09 12:04:56 +11001724 if (!(le32_to_cpu(sb->feature_map) &
1725 MD_FEATURE_RECOVERY_BITMAP))
1726 rdev->saved_raid_disk = -1;
1727 } else
NeilBrown5fd6c1d2006-06-26 00:27:40 -07001728 set_bit(In_sync, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 rdev->raid_disk = role;
1730 break;
1731 }
NeilBrown8ddf9ef2005-09-09 16:23:45 -07001732 if (sb->devflags & WriteMostly1)
1733 set_bit(WriteMostly, &rdev->flags);
NeilBrown688834e2016-11-18 16:16:11 +11001734 if (sb->devflags & FailFast1)
1735 set_bit(FailFast, &rdev->flags);
NeilBrown2d78f8c2011-12-23 10:17:51 +11001736 if (le32_to_cpu(sb->feature_map) & MD_FEATURE_REPLACEMENT)
1737 set_bit(Replacement, &rdev->flags);
NeilBrown41158c72005-06-21 17:17:25 -07001738 } else /* MULTIPATH are always insync */
NeilBrownb2d444d2005-11-08 21:39:31 -08001739 set_bit(In_sync, &rdev->flags);
NeilBrown41158c72005-06-21 17:17:25 -07001740
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 return 0;
1742}
1743
NeilBrownfd01b882011-10-11 16:47:53 +11001744static void super_1_sync(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745{
1746 struct mdp_superblock_1 *sb;
NeilBrown3cb03002011-10-11 16:45:26 +11001747 struct md_rdev *rdev2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 int max_dev, i;
1749 /* make rdev->sb match mddev and rdev data. */
1750
Namhyung Kim65a06f062011-07-27 11:00:36 +10001751 sb = page_address(rdev->sb_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752
1753 sb->feature_map = 0;
1754 sb->pad0 = 0;
NeilBrown5fd6c1d2006-06-26 00:27:40 -07001755 sb->recovery_offset = cpu_to_le64(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 memset(sb->pad3, 0, sizeof(sb->pad3));
1757
1758 sb->utime = cpu_to_le64((__u64)mddev->utime);
1759 sb->events = cpu_to_le64(mddev->events);
1760 if (mddev->in_sync)
1761 sb->resync_offset = cpu_to_le64(mddev->recovery_cp);
Shaohua Libd18f642015-09-02 13:49:50 -07001762 else if (test_bit(MD_JOURNAL_CLEAN, &mddev->flags))
1763 sb->resync_offset = cpu_to_le64(MaxSector);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764 else
1765 sb->resync_offset = cpu_to_le64(0);
1766
NeilBrown1c05b4b2006-10-21 10:24:08 -07001767 sb->cnt_corrected_read = cpu_to_le32(atomic_read(&rdev->corrected_errors));
NeilBrown4dbcdc72006-01-06 00:20:52 -08001768
NeilBrownf0ca3402006-02-02 14:28:04 -08001769 sb->raid_disks = cpu_to_le32(mddev->raid_disks);
Andre Noll58c0fed2009-03-31 14:33:13 +11001770 sb->size = cpu_to_le64(mddev->dev_sectors);
Andre Noll9d8f0362009-06-18 08:45:01 +10001771 sb->chunksize = cpu_to_le32(mddev->chunk_sectors);
NeilBrown62e1e382009-05-26 09:40:59 +10001772 sb->level = cpu_to_le32(mddev->level);
1773 sb->layout = cpu_to_le32(mddev->layout);
NeilBrown688834e2016-11-18 16:16:11 +11001774 if (test_bit(FailFast, &rdev->flags))
1775 sb->devflags |= FailFast1;
1776 else
1777 sb->devflags &= ~FailFast1;
NeilBrownf0ca3402006-02-02 14:28:04 -08001778
NeilBrownaeb9b212011-08-25 14:43:08 +10001779 if (test_bit(WriteMostly, &rdev->flags))
1780 sb->devflags |= WriteMostly1;
1781 else
1782 sb->devflags &= ~WriteMostly1;
NeilBrownc6563a82012-05-21 09:27:00 +10001783 sb->data_offset = cpu_to_le64(rdev->data_offset);
1784 sb->data_size = cpu_to_le64(rdev->sectors);
NeilBrownaeb9b212011-08-25 14:43:08 +10001785
NeilBrownc3d97142009-12-14 12:49:52 +11001786 if (mddev->bitmap && mddev->bitmap_info.file == NULL) {
1787 sb->bitmap_offset = cpu_to_le32((__u32)mddev->bitmap_info.offset);
NeilBrown71c08052005-09-09 16:23:51 -07001788 sb->feature_map = cpu_to_le32(MD_FEATURE_BITMAP_OFFSET);
NeilBrowna654b9d82005-06-21 17:17:27 -07001789 }
NeilBrown5fd6c1d2006-06-26 00:27:40 -07001790
Shaohua Lif2076e72015-10-08 21:54:12 -07001791 if (rdev->raid_disk >= 0 && !test_bit(Journal, &rdev->flags) &&
NeilBrown97e4f422009-03-31 14:33:13 +11001792 !test_bit(In_sync, &rdev->flags)) {
NeilBrown93be75f2009-12-14 12:50:06 +11001793 sb->feature_map |=
1794 cpu_to_le32(MD_FEATURE_RECOVERY_OFFSET);
1795 sb->recovery_offset =
1796 cpu_to_le64(rdev->recovery_offset);
NeilBrownf4667222013-12-09 12:04:56 +11001797 if (rdev->saved_raid_disk >= 0 && mddev->bitmap)
1798 sb->feature_map |=
1799 cpu_to_le32(MD_FEATURE_RECOVERY_BITMAP);
NeilBrown5fd6c1d2006-06-26 00:27:40 -07001800 }
Shaohua Li3069aa82015-08-13 14:31:56 -07001801 /* Note: recovery_offset and journal_tail share space */
1802 if (test_bit(Journal, &rdev->flags))
1803 sb->journal_tail = cpu_to_le64(rdev->journal_tail);
NeilBrown2d78f8c2011-12-23 10:17:51 +11001804 if (test_bit(Replacement, &rdev->flags))
1805 sb->feature_map |=
1806 cpu_to_le32(MD_FEATURE_REPLACEMENT);
NeilBrown5fd6c1d2006-06-26 00:27:40 -07001807
NeilBrownf6705572006-03-27 01:18:11 -08001808 if (mddev->reshape_position != MaxSector) {
1809 sb->feature_map |= cpu_to_le32(MD_FEATURE_RESHAPE_ACTIVE);
1810 sb->reshape_position = cpu_to_le64(mddev->reshape_position);
1811 sb->new_layout = cpu_to_le32(mddev->new_layout);
1812 sb->delta_disks = cpu_to_le32(mddev->delta_disks);
1813 sb->new_level = cpu_to_le32(mddev->new_level);
Andre Noll664e7c42009-06-18 08:45:27 +10001814 sb->new_chunk = cpu_to_le32(mddev->new_chunk_sectors);
NeilBrown2c810cd2012-05-21 09:27:00 +10001815 if (mddev->delta_disks == 0 &&
1816 mddev->reshape_backwards)
1817 sb->feature_map
1818 |= cpu_to_le32(MD_FEATURE_RESHAPE_BACKWARDS);
NeilBrownc6563a82012-05-21 09:27:00 +10001819 if (rdev->new_data_offset != rdev->data_offset) {
1820 sb->feature_map
1821 |= cpu_to_le32(MD_FEATURE_NEW_OFFSET);
1822 sb->new_offset = cpu_to_le32((__u32)(rdev->new_data_offset
1823 - rdev->data_offset));
1824 }
NeilBrownf6705572006-03-27 01:18:11 -08001825 }
NeilBrowna654b9d82005-06-21 17:17:27 -07001826
Goldwyn Rodrigues3c462c82015-08-19 07:35:54 +10001827 if (mddev_is_clustered(mddev))
1828 sb->feature_map |= cpu_to_le32(MD_FEATURE_CLUSTERED);
1829
NeilBrown2699b672011-07-28 11:31:47 +10001830 if (rdev->badblocks.count == 0)
1831 /* Nothing to do for bad blocks*/ ;
1832 else if (sb->bblog_offset == 0)
1833 /* Cannot record bad blocks on this device */
1834 md_error(mddev, rdev);
1835 else {
1836 struct badblocks *bb = &rdev->badblocks;
1837 u64 *bbp = (u64 *)page_address(rdev->bb_page);
1838 u64 *p = bb->page;
1839 sb->feature_map |= cpu_to_le32(MD_FEATURE_BAD_BLOCKS);
1840 if (bb->changed) {
1841 unsigned seq;
1842
1843retry:
1844 seq = read_seqbegin(&bb->lock);
1845
1846 memset(bbp, 0xff, PAGE_SIZE);
1847
1848 for (i = 0 ; i < bb->count ; i++) {
majianpeng35f9ac22012-11-08 08:56:27 +08001849 u64 internal_bb = p[i];
NeilBrown2699b672011-07-28 11:31:47 +10001850 u64 store_bb = ((BB_OFFSET(internal_bb) << 10)
1851 | BB_LEN(internal_bb));
majianpeng35f9ac22012-11-08 08:56:27 +08001852 bbp[i] = cpu_to_le64(store_bb);
NeilBrown2699b672011-07-28 11:31:47 +10001853 }
NeilBrownd0962932012-03-19 12:46:41 +11001854 bb->changed = 0;
NeilBrown2699b672011-07-28 11:31:47 +10001855 if (read_seqretry(&bb->lock, seq))
1856 goto retry;
1857
1858 bb->sector = (rdev->sb_start +
1859 (int)le32_to_cpu(sb->bblog_offset));
1860 bb->size = le16_to_cpu(sb->bblog_size);
NeilBrown2699b672011-07-28 11:31:47 +10001861 }
1862 }
1863
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 max_dev = 0;
NeilBrowndafb20f2012-03-19 12:46:39 +11001865 rdev_for_each(rdev2, mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 if (rdev2->desc_nr+1 > max_dev)
1867 max_dev = rdev2->desc_nr+1;
NeilBrowna778b732007-05-23 13:58:10 -07001868
NeilBrown70471da2009-08-03 10:59:57 +10001869 if (max_dev > le32_to_cpu(sb->max_dev)) {
1870 int bmask;
NeilBrowna778b732007-05-23 13:58:10 -07001871 sb->max_dev = cpu_to_le32(max_dev);
NeilBrown70471da2009-08-03 10:59:57 +10001872 rdev->sb_size = max_dev * 2 + 256;
1873 bmask = queue_logical_block_size(rdev->bdev->bd_disk->queue)-1;
1874 if (rdev->sb_size & bmask)
1875 rdev->sb_size = (rdev->sb_size | bmask) + 1;
NeilBrownddcf3522010-09-08 16:48:17 +10001876 } else
1877 max_dev = le32_to_cpu(sb->max_dev);
1878
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879 for (i=0; i<max_dev;i++)
Lidong Zhong8df72022017-06-12 10:45:55 +08001880 sb->dev_roles[i] = cpu_to_le16(MD_DISK_ROLE_SPARE);
NeilBrownf72ffdd2014-09-30 14:23:59 +10001881
Song Liua97b7892015-10-08 21:54:09 -07001882 if (test_bit(MD_HAS_JOURNAL, &mddev->flags))
1883 sb->feature_map |= cpu_to_le32(MD_FEATURE_JOURNAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884
Artur Paszkiewiczea0213e2017-03-09 09:59:57 +01001885 if (test_bit(MD_HAS_PPL, &mddev->flags)) {
Pawel Baldysiakddc08822017-08-16 17:13:45 +02001886 if (test_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags))
1887 sb->feature_map |=
1888 cpu_to_le32(MD_FEATURE_MULTIPLE_PPLS);
1889 else
1890 sb->feature_map |= cpu_to_le32(MD_FEATURE_PPL);
Artur Paszkiewiczea0213e2017-03-09 09:59:57 +01001891 sb->ppl.offset = cpu_to_le16(rdev->ppl.offset);
1892 sb->ppl.size = cpu_to_le16(rdev->ppl.size);
1893 }
1894
NeilBrowndafb20f2012-03-19 12:46:39 +11001895 rdev_for_each(rdev2, mddev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 i = rdev2->desc_nr;
NeilBrownb2d444d2005-11-08 21:39:31 -08001897 if (test_bit(Faulty, &rdev2->flags))
Song Liuc4d4c912015-08-13 14:31:54 -07001898 sb->dev_roles[i] = cpu_to_le16(MD_DISK_ROLE_FAULTY);
NeilBrownb2d444d2005-11-08 21:39:31 -08001899 else if (test_bit(In_sync, &rdev2->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 sb->dev_roles[i] = cpu_to_le16(rdev2->raid_disk);
Song Liua97b7892015-10-08 21:54:09 -07001901 else if (test_bit(Journal, &rdev2->flags))
Song Liubac624f2015-08-13 14:31:55 -07001902 sb->dev_roles[i] = cpu_to_le16(MD_DISK_ROLE_JOURNAL);
NeilBrown93be75f2009-12-14 12:50:06 +11001903 else if (rdev2->raid_disk >= 0)
NeilBrown5fd6c1d2006-06-26 00:27:40 -07001904 sb->dev_roles[i] = cpu_to_le16(rdev2->raid_disk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 else
Song Liuc4d4c912015-08-13 14:31:54 -07001906 sb->dev_roles[i] = cpu_to_le16(MD_DISK_ROLE_SPARE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 }
1908
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909 sb->sb_csum = calc_sb_1_csum(sb);
1910}
1911
Chris Webb0cd17fe2008-06-28 08:31:46 +10001912static unsigned long long
NeilBrown3cb03002011-10-11 16:45:26 +11001913super_1_rdev_size_change(struct md_rdev *rdev, sector_t num_sectors)
Chris Webb0cd17fe2008-06-28 08:31:46 +10001914{
1915 struct mdp_superblock_1 *sb;
Andre Noll15f4a5f2008-07-21 14:42:12 +10001916 sector_t max_sectors;
Andre Noll58c0fed2009-03-31 14:33:13 +11001917 if (num_sectors && num_sectors < rdev->mddev->dev_sectors)
Chris Webb0cd17fe2008-06-28 08:31:46 +10001918 return 0; /* component must fit device */
NeilBrownc6563a82012-05-21 09:27:00 +10001919 if (rdev->data_offset != rdev->new_data_offset)
1920 return 0; /* too confusing */
Andre Noll0f420352008-07-11 22:02:23 +10001921 if (rdev->sb_start < rdev->data_offset) {
Chris Webb0cd17fe2008-06-28 08:31:46 +10001922 /* minor versions 1 and 2; superblock before data */
Mike Snitzer77304d22010-11-08 14:39:12 +01001923 max_sectors = i_size_read(rdev->bdev->bd_inode) >> 9;
Andre Noll15f4a5f2008-07-21 14:42:12 +10001924 max_sectors -= rdev->data_offset;
1925 if (!num_sectors || num_sectors > max_sectors)
1926 num_sectors = max_sectors;
NeilBrownc3d97142009-12-14 12:49:52 +11001927 } else if (rdev->mddev->bitmap_info.offset) {
Chris Webb0cd17fe2008-06-28 08:31:46 +10001928 /* minor version 0 with bitmap we can't move */
1929 return 0;
1930 } else {
1931 /* minor version 0; superblock after data */
Andre Noll0f420352008-07-11 22:02:23 +10001932 sector_t sb_start;
Mike Snitzer77304d22010-11-08 14:39:12 +01001933 sb_start = (i_size_read(rdev->bdev->bd_inode) >> 9) - 8*2;
Andre Noll0f420352008-07-11 22:02:23 +10001934 sb_start &= ~(sector_t)(4*2 - 1);
Andre Nolldd8ac332009-03-31 14:33:13 +11001935 max_sectors = rdev->sectors + sb_start - rdev->sb_start;
Andre Noll15f4a5f2008-07-21 14:42:12 +10001936 if (!num_sectors || num_sectors > max_sectors)
1937 num_sectors = max_sectors;
Andre Noll0f420352008-07-11 22:02:23 +10001938 rdev->sb_start = sb_start;
Chris Webb0cd17fe2008-06-28 08:31:46 +10001939 }
Namhyung Kim65a06f062011-07-27 11:00:36 +10001940 sb = page_address(rdev->sb_page);
Andre Noll15f4a5f2008-07-21 14:42:12 +10001941 sb->data_size = cpu_to_le64(num_sectors);
Jason Yan3fb632e2017-03-10 11:27:23 +08001942 sb->super_offset = cpu_to_le64(rdev->sb_start);
Chris Webb0cd17fe2008-06-28 08:31:46 +10001943 sb->sb_csum = calc_sb_1_csum(sb);
NeilBrown46533ff2016-11-18 16:16:11 +11001944 do {
1945 md_super_write(rdev->mddev, rdev, rdev->sb_start, rdev->sb_size,
1946 rdev->sb_page);
1947 } while (md_super_wait(rdev->mddev) < 0);
Justin Maggardc26a44e2010-11-24 16:36:17 +11001948 return num_sectors;
NeilBrownc6563a82012-05-21 09:27:00 +10001949
1950}
1951
1952static int
1953super_1_allow_new_offset(struct md_rdev *rdev,
1954 unsigned long long new_offset)
1955{
1956 /* All necessary checks on new >= old have been done */
1957 struct bitmap *bitmap;
1958 if (new_offset >= rdev->data_offset)
1959 return 1;
1960
1961 /* with 1.0 metadata, there is no metadata to tread on
1962 * so we can always move back */
1963 if (rdev->mddev->minor_version == 0)
1964 return 1;
1965
1966 /* otherwise we must be sure not to step on
1967 * any metadata, so stay:
1968 * 36K beyond start of superblock
1969 * beyond end of badblocks
1970 * beyond write-intent bitmap
1971 */
1972 if (rdev->sb_start + (32+4)*2 > new_offset)
1973 return 0;
1974 bitmap = rdev->mddev->bitmap;
1975 if (bitmap && !rdev->mddev->bitmap_info.file &&
1976 rdev->sb_start + rdev->mddev->bitmap_info.offset +
NeilBrown1ec885c2012-05-22 13:55:10 +10001977 bitmap->storage.file_pages * (PAGE_SIZE>>9) > new_offset)
NeilBrownc6563a82012-05-21 09:27:00 +10001978 return 0;
1979 if (rdev->badblocks.sector + rdev->badblocks.size > new_offset)
1980 return 0;
1981
1982 return 1;
Chris Webb0cd17fe2008-06-28 08:31:46 +10001983}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984
Adrian Bunk75c96f82005-05-05 16:16:09 -07001985static struct super_type super_types[] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986 [0] = {
1987 .name = "0.90.0",
1988 .owner = THIS_MODULE,
Chris Webb0cd17fe2008-06-28 08:31:46 +10001989 .load_super = super_90_load,
1990 .validate_super = super_90_validate,
1991 .sync_super = super_90_sync,
1992 .rdev_size_change = super_90_rdev_size_change,
NeilBrownc6563a82012-05-21 09:27:00 +10001993 .allow_new_offset = super_90_allow_new_offset,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 },
1995 [1] = {
1996 .name = "md-1",
1997 .owner = THIS_MODULE,
Chris Webb0cd17fe2008-06-28 08:31:46 +10001998 .load_super = super_1_load,
1999 .validate_super = super_1_validate,
2000 .sync_super = super_1_sync,
2001 .rdev_size_change = super_1_rdev_size_change,
NeilBrownc6563a82012-05-21 09:27:00 +10002002 .allow_new_offset = super_1_allow_new_offset,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 },
2004};
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005
NeilBrownfd01b882011-10-11 16:47:53 +11002006static void sync_super(struct mddev *mddev, struct md_rdev *rdev)
Jonathan Brassow076f9682011-06-07 17:51:30 -05002007{
2008 if (mddev->sync_super) {
2009 mddev->sync_super(mddev, rdev);
2010 return;
2011 }
2012
2013 BUG_ON(mddev->major_version >= ARRAY_SIZE(super_types));
2014
2015 super_types[mddev->major_version].sync_super(mddev, rdev);
2016}
2017
NeilBrownfd01b882011-10-11 16:47:53 +11002018static int match_mddev_units(struct mddev *mddev1, struct mddev *mddev2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019{
NeilBrown3cb03002011-10-11 16:45:26 +11002020 struct md_rdev *rdev, *rdev2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021
NeilBrown4b809912008-07-21 17:05:25 +10002022 rcu_read_lock();
Song Liu0b020e82015-09-03 23:00:35 -07002023 rdev_for_each_rcu(rdev, mddev1) {
2024 if (test_bit(Faulty, &rdev->flags) ||
2025 test_bit(Journal, &rdev->flags) ||
2026 rdev->raid_disk == -1)
2027 continue;
2028 rdev_for_each_rcu(rdev2, mddev2) {
2029 if (test_bit(Faulty, &rdev2->flags) ||
2030 test_bit(Journal, &rdev2->flags) ||
2031 rdev2->raid_disk == -1)
2032 continue;
NeilBrown7dd5e7c32007-02-28 20:11:35 -08002033 if (rdev->bdev->bd_contains ==
NeilBrown4b809912008-07-21 17:05:25 +10002034 rdev2->bdev->bd_contains) {
2035 rcu_read_unlock();
NeilBrown7dd5e7c32007-02-28 20:11:35 -08002036 return 1;
NeilBrown4b809912008-07-21 17:05:25 +10002037 }
Song Liu0b020e82015-09-03 23:00:35 -07002038 }
2039 }
NeilBrown4b809912008-07-21 17:05:25 +10002040 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041 return 0;
2042}
2043
2044static LIST_HEAD(pending_raid_disks);
2045
Andre Nollac5e7112009-08-03 10:59:47 +10002046/*
2047 * Try to register data integrity profile for an mddev
2048 *
2049 * This is called when an array is started and after a disk has been kicked
2050 * from the array. It only succeeds if all working and active component devices
2051 * are integrity capable with matching profiles.
2052 */
NeilBrownfd01b882011-10-11 16:47:53 +11002053int md_integrity_register(struct mddev *mddev)
Martin K. Petersen3f9d99c2009-03-31 14:27:02 +11002054{
NeilBrown3cb03002011-10-11 16:45:26 +11002055 struct md_rdev *rdev, *reference = NULL;
Martin K. Petersen3f9d99c2009-03-31 14:27:02 +11002056
Andre Nollac5e7112009-08-03 10:59:47 +10002057 if (list_empty(&mddev->disks))
2058 return 0; /* nothing to do */
Jonathan Brassow629acb62011-06-08 15:10:08 +10002059 if (!mddev->gendisk || blk_get_integrity(mddev->gendisk))
2060 return 0; /* shouldn't register, or already is */
NeilBrowndafb20f2012-03-19 12:46:39 +11002061 rdev_for_each(rdev, mddev) {
Andre Nollac5e7112009-08-03 10:59:47 +10002062 /* skip spares and non-functional disks */
2063 if (test_bit(Faulty, &rdev->flags))
2064 continue;
2065 if (rdev->raid_disk < 0)
2066 continue;
Andre Nollac5e7112009-08-03 10:59:47 +10002067 if (!reference) {
2068 /* Use the first rdev as the reference */
2069 reference = rdev;
2070 continue;
2071 }
2072 /* does this rdev's profile match the reference profile? */
2073 if (blk_integrity_compare(reference->bdev->bd_disk,
2074 rdev->bdev->bd_disk) < 0)
2075 return -EINVAL;
Martin K. Petersen3f9d99c2009-03-31 14:27:02 +11002076 }
Martin K. Petersen89078d52011-03-28 20:09:12 -04002077 if (!reference || !bdev_get_integrity(reference->bdev))
2078 return 0;
Andre Nollac5e7112009-08-03 10:59:47 +10002079 /*
2080 * All component devices are integrity capable and have matching
2081 * profiles, register the common profile for the md device.
2082 */
Martin K. Petersen25520d52015-10-21 13:19:49 -04002083 blk_integrity_register(mddev->gendisk,
2084 bdev_get_integrity(reference->bdev));
2085
NeilBrown9d487392016-11-02 14:16:49 +11002086 pr_debug("md: data integrity enabled on %s\n", mdname(mddev));
Martin K. Petersena91a2782011-03-17 11:11:05 +01002087 if (bioset_integrity_create(mddev->bio_set, BIO_POOL_SIZE)) {
NeilBrown9d487392016-11-02 14:16:49 +11002088 pr_err("md: failed to create integrity pool for %s\n",
Martin K. Petersena91a2782011-03-17 11:11:05 +01002089 mdname(mddev));
2090 return -EINVAL;
2091 }
Andre Nollac5e7112009-08-03 10:59:47 +10002092 return 0;
Martin K. Petersen3f9d99c2009-03-31 14:27:02 +11002093}
Andre Nollac5e7112009-08-03 10:59:47 +10002094EXPORT_SYMBOL(md_integrity_register);
2095
Dan Williams1501efa2016-01-13 16:00:07 -08002096/*
2097 * Attempt to add an rdev, but only if it is consistent with the current
2098 * integrity profile
2099 */
2100int md_integrity_add_rdev(struct md_rdev *rdev, struct mddev *mddev)
Andre Nollac5e7112009-08-03 10:59:47 +10002101{
Jonathan Brassow2863b9e2012-10-11 13:38:58 +11002102 struct blk_integrity *bi_rdev;
2103 struct blk_integrity *bi_mddev;
Dan Williams1501efa2016-01-13 16:00:07 -08002104 char name[BDEVNAME_SIZE];
Jonathan Brassow2863b9e2012-10-11 13:38:58 +11002105
2106 if (!mddev->gendisk)
Dan Williams1501efa2016-01-13 16:00:07 -08002107 return 0;
Jonathan Brassow2863b9e2012-10-11 13:38:58 +11002108
2109 bi_rdev = bdev_get_integrity(rdev->bdev);
2110 bi_mddev = blk_get_integrity(mddev->gendisk);
Andre Nollac5e7112009-08-03 10:59:47 +10002111
2112 if (!bi_mddev) /* nothing to do */
Dan Williams1501efa2016-01-13 16:00:07 -08002113 return 0;
2114
2115 if (blk_integrity_compare(mddev->gendisk, rdev->bdev->bd_disk) != 0) {
NeilBrown9d487392016-11-02 14:16:49 +11002116 pr_err("%s: incompatible integrity profile for %s\n",
2117 mdname(mddev), bdevname(rdev->bdev, name));
Dan Williams1501efa2016-01-13 16:00:07 -08002118 return -ENXIO;
2119 }
2120
2121 return 0;
Andre Nollac5e7112009-08-03 10:59:47 +10002122}
2123EXPORT_SYMBOL(md_integrity_add_rdev);
Martin K. Petersen3f9d99c2009-03-31 14:27:02 +11002124
NeilBrownf72ffdd2014-09-30 14:23:59 +10002125static int bind_rdev_to_array(struct md_rdev *rdev, struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126{
NeilBrown7dd5e7c32007-02-28 20:11:35 -08002127 char b[BDEVNAME_SIZE];
NeilBrownf637b9f2005-11-08 21:39:37 -08002128 struct kobject *ko;
NeilBrown5e55e2f2007-03-26 21:32:14 -08002129 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130
Dan Williams11e2ede2008-04-30 00:52:32 -07002131 /* prevent duplicates */
2132 if (find_rdev(mddev, rdev->bdev->bd_dev))
2133 return -EEXIST;
2134
NeilBrown97b20ef2017-04-13 08:53:48 +10002135 if ((bdev_read_only(rdev->bdev) || bdev_read_only(rdev->meta_bdev)) &&
2136 mddev->pers)
2137 return -EROFS;
2138
Andre Nolldd8ac332009-03-31 14:33:13 +11002139 /* make sure rdev->sectors exceeds mddev->dev_sectors */
Shaohua Lif6b6ec52015-12-21 10:51:02 +11002140 if (!test_bit(Journal, &rdev->flags) &&
2141 rdev->sectors &&
2142 (mddev->dev_sectors == 0 || rdev->sectors < mddev->dev_sectors)) {
NeilBrowna778b732007-05-23 13:58:10 -07002143 if (mddev->pers) {
2144 /* Cannot change size, so fail
2145 * If mddev->level <= 0, then we don't care
2146 * about aligning sizes (e.g. linear)
2147 */
2148 if (mddev->level > 0)
2149 return -ENOSPC;
2150 } else
Andre Nolldd8ac332009-03-31 14:33:13 +11002151 mddev->dev_sectors = rdev->sectors;
NeilBrown2bf071b2006-01-06 00:20:55 -08002152 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153
2154 /* Verify rdev->desc_nr is unique.
2155 * If it is -1, assign a free number, else
2156 * check number is not in use
2157 */
NeilBrown4878e9e2014-09-25 17:00:11 +10002158 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 if (rdev->desc_nr < 0) {
2160 int choice = 0;
NeilBrown4878e9e2014-09-25 17:00:11 +10002161 if (mddev->pers)
2162 choice = mddev->raid_disks;
Goldwyn Rodrigues57d051d2015-04-14 10:43:55 -05002163 while (md_find_rdev_nr_rcu(mddev, choice))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 choice++;
2165 rdev->desc_nr = choice;
2166 } else {
Goldwyn Rodrigues57d051d2015-04-14 10:43:55 -05002167 if (md_find_rdev_nr_rcu(mddev, rdev->desc_nr)) {
NeilBrown4878e9e2014-09-25 17:00:11 +10002168 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169 return -EBUSY;
NeilBrown4878e9e2014-09-25 17:00:11 +10002170 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171 }
NeilBrown4878e9e2014-09-25 17:00:11 +10002172 rcu_read_unlock();
Shaohua Lif6b6ec52015-12-21 10:51:02 +11002173 if (!test_bit(Journal, &rdev->flags) &&
2174 mddev->max_disks && rdev->desc_nr >= mddev->max_disks) {
NeilBrown9d487392016-11-02 14:16:49 +11002175 pr_warn("md: %s: array is limited to %d devices\n",
2176 mdname(mddev), mddev->max_disks);
NeilBrownde01dfa2009-02-06 18:02:46 +11002177 return -EBUSY;
2178 }
NeilBrown19133a42005-11-08 21:39:35 -08002179 bdevname(rdev->bdev,b);
Rasmus Villemoes90a9bef2015-06-25 15:02:36 -07002180 strreplace(b, '/', '!');
Greg Kroah-Hartman649316b2007-12-17 23:05:35 -07002181
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182 rdev->mddev = mddev;
NeilBrown9d487392016-11-02 14:16:49 +11002183 pr_debug("md: bind<%s>\n", b);
NeilBrown86e6ffd2005-11-08 21:39:24 -08002184
Greg Kroah-Hartmanb2d6db52007-12-17 23:05:35 -07002185 if ((err = kobject_add(&rdev->kobj, &mddev->kobj, "dev-%s", b)))
NeilBrown5e55e2f2007-03-26 21:32:14 -08002186 goto fail;
NeilBrown86e6ffd2005-11-08 21:39:24 -08002187
Tejun Heo0762b8b2008-08-25 19:56:12 +09002188 ko = &part_to_dev(rdev->bdev->bd_part)->kobj;
NeilBrown00bcb4a2010-06-01 19:37:23 +10002189 if (sysfs_create_link(&rdev->kobj, ko, "block"))
2190 /* failure here is OK */;
2191 rdev->sysfs_state = sysfs_get_dirent_safe(rdev->kobj.sd, "state");
NeilBrown3c0ee632008-10-21 13:25:28 +11002192
NeilBrown4b809912008-07-21 17:05:25 +10002193 list_add_rcu(&rdev->same_set, &mddev->disks);
Tejun Heoe09b4572010-11-13 11:55:17 +01002194 bd_link_disk_holder(rdev->bdev, mddev->gendisk);
NeilBrown4044ba52009-01-09 08:31:11 +11002195
2196 /* May as well allow recovery to be retried once */
NeilBrown53890422011-07-27 11:00:36 +10002197 mddev->recovery_disabled++;
Martin K. Petersen3f9d99c2009-03-31 14:27:02 +11002198
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199 return 0;
NeilBrown5e55e2f2007-03-26 21:32:14 -08002200
2201 fail:
NeilBrown9d487392016-11-02 14:16:49 +11002202 pr_warn("md: failed to register dev-%s for %s\n",
2203 b, mdname(mddev));
NeilBrown5e55e2f2007-03-26 21:32:14 -08002204 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205}
2206
NeilBrown177a99b2008-02-06 01:39:56 -08002207static void md_delayed_delete(struct work_struct *ws)
NeilBrown5792a282007-04-04 19:08:18 -07002208{
NeilBrown3cb03002011-10-11 16:45:26 +11002209 struct md_rdev *rdev = container_of(ws, struct md_rdev, del_work);
NeilBrown5792a282007-04-04 19:08:18 -07002210 kobject_del(&rdev->kobj);
NeilBrown177a99b2008-02-06 01:39:56 -08002211 kobject_put(&rdev->kobj);
NeilBrown5792a282007-04-04 19:08:18 -07002212}
2213
NeilBrownf72ffdd2014-09-30 14:23:59 +10002214static void unbind_rdev_from_array(struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215{
2216 char b[BDEVNAME_SIZE];
NeilBrown403df472014-09-30 15:52:29 +10002217
Tejun Heo49731ba2011-01-14 18:43:57 +01002218 bd_unlink_disk_holder(rdev->bdev, rdev->mddev->gendisk);
NeilBrown4b809912008-07-21 17:05:25 +10002219 list_del_rcu(&rdev->same_set);
NeilBrown9d487392016-11-02 14:16:49 +11002220 pr_debug("md: unbind<%s>\n", bdevname(rdev->bdev,b));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 rdev->mddev = NULL;
NeilBrown86e6ffd2005-11-08 21:39:24 -08002222 sysfs_remove_link(&rdev->kobj, "block");
NeilBrown3c0ee632008-10-21 13:25:28 +11002223 sysfs_put(rdev->sysfs_state);
2224 rdev->sysfs_state = NULL;
NeilBrown2230dfe2011-07-28 11:31:46 +10002225 rdev->badblocks.count = 0;
NeilBrown5792a282007-04-04 19:08:18 -07002226 /* We need to delay this, otherwise we can deadlock when
NeilBrown4b809912008-07-21 17:05:25 +10002227 * writing to 'remove' to "dev/state". We also need
2228 * to delay it due to rcu usage.
NeilBrown5792a282007-04-04 19:08:18 -07002229 */
NeilBrown4b809912008-07-21 17:05:25 +10002230 synchronize_rcu();
NeilBrown177a99b2008-02-06 01:39:56 -08002231 INIT_WORK(&rdev->del_work, md_delayed_delete);
2232 kobject_get(&rdev->kobj);
Tejun Heoe804ac72010-10-15 15:36:08 +02002233 queue_work(md_misc_wq, &rdev->del_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234}
2235
2236/*
2237 * prevent the device from being mounted, repartitioned or
2238 * otherwise reused by a RAID array (or any other kernel
2239 * subsystem), by bd_claiming the device.
2240 */
NeilBrown3cb03002011-10-11 16:45:26 +11002241static int lock_rdev(struct md_rdev *rdev, dev_t dev, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242{
2243 int err = 0;
2244 struct block_device *bdev;
2245 char b[BDEVNAME_SIZE];
2246
Tejun Heod4d77622010-11-13 11:55:18 +01002247 bdev = blkdev_get_by_dev(dev, FMODE_READ|FMODE_WRITE|FMODE_EXCL,
NeilBrown3cb03002011-10-11 16:45:26 +11002248 shared ? (struct md_rdev *)lock_rdev : rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 if (IS_ERR(bdev)) {
NeilBrown9d487392016-11-02 14:16:49 +11002250 pr_warn("md: could not open %s.\n", __bdevname(dev, b));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251 return PTR_ERR(bdev);
2252 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253 rdev->bdev = bdev;
2254 return err;
2255}
2256
NeilBrown3cb03002011-10-11 16:45:26 +11002257static void unlock_rdev(struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258{
2259 struct block_device *bdev = rdev->bdev;
2260 rdev->bdev = NULL;
Tejun Heoe525fd82010-11-13 11:55:17 +01002261 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262}
2263
2264void md_autodetect_dev(dev_t dev);
2265
NeilBrownf72ffdd2014-09-30 14:23:59 +10002266static void export_rdev(struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267{
2268 char b[BDEVNAME_SIZE];
NeilBrown403df472014-09-30 15:52:29 +10002269
NeilBrown9d487392016-11-02 14:16:49 +11002270 pr_debug("md: export_rdev(%s)\n", bdevname(rdev->bdev,b));
NeilBrown545c8792012-05-22 13:54:30 +10002271 md_rdev_clear(rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272#ifndef MODULE
NeilBrownd0fae182008-03-04 14:29:31 -08002273 if (test_bit(AutoDetected, &rdev->flags))
2274 md_autodetect_dev(rdev->bdev->bd_dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275#endif
2276 unlock_rdev(rdev);
NeilBrown86e6ffd2005-11-08 21:39:24 -08002277 kobject_put(&rdev->kobj);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278}
2279
Goldwyn Rodriguesfb56dfe2015-04-14 10:43:24 -05002280void md_kick_rdev_from_array(struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281{
2282 unbind_rdev_from_array(rdev);
2283 export_rdev(rdev);
2284}
Goldwyn Rodriguesfb56dfe2015-04-14 10:43:24 -05002285EXPORT_SYMBOL_GPL(md_kick_rdev_from_array);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
NeilBrownfd01b882011-10-11 16:47:53 +11002287static void export_array(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288{
NeilBrown0638bb02014-09-25 17:43:47 +10002289 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290
NeilBrown0638bb02014-09-25 17:43:47 +10002291 while (!list_empty(&mddev->disks)) {
2292 rdev = list_first_entry(&mddev->disks, struct md_rdev,
2293 same_set);
Goldwyn Rodriguesfb56dfe2015-04-14 10:43:24 -05002294 md_kick_rdev_from_array(rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 mddev->raid_disks = 0;
2297 mddev->major_version = 0;
2298}
2299
NeilBrown6497709b2017-03-15 14:05:14 +11002300static bool set_in_sync(struct mddev *mddev)
2301{
Shaohua Lied9b66d2017-07-25 15:18:13 -07002302 WARN_ON_ONCE(NR_CPUS != 1 && !spin_is_locked(&mddev->lock));
NeilBrown4ad23a972017-03-15 14:05:14 +11002303 if (!mddev->in_sync) {
2304 mddev->sync_checkers++;
2305 spin_unlock(&mddev->lock);
2306 percpu_ref_switch_to_atomic_sync(&mddev->writes_pending);
2307 spin_lock(&mddev->lock);
2308 if (!mddev->in_sync &&
2309 percpu_ref_is_zero(&mddev->writes_pending)) {
NeilBrown6497709b2017-03-15 14:05:14 +11002310 mddev->in_sync = 1;
NeilBrown4ad23a972017-03-15 14:05:14 +11002311 /*
2312 * Ensure ->in_sync is visible before we clear
2313 * ->sync_checkers.
2314 */
NeilBrown55cc39f2017-03-15 14:05:14 +11002315 smp_mb();
NeilBrown6497709b2017-03-15 14:05:14 +11002316 set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
2317 sysfs_notify_dirent_safe(mddev->sysfs_state);
2318 }
NeilBrown4ad23a972017-03-15 14:05:14 +11002319 if (--mddev->sync_checkers == 0)
2320 percpu_ref_switch_to_percpu(&mddev->writes_pending);
NeilBrown6497709b2017-03-15 14:05:14 +11002321 }
2322 if (mddev->safemode == 1)
2323 mddev->safemode = 0;
2324 return mddev->in_sync;
2325}
2326
NeilBrownf72ffdd2014-09-30 14:23:59 +10002327static void sync_sbs(struct mddev *mddev, int nospares)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328{
NeilBrown42543762006-06-26 00:27:57 -07002329 /* Update each superblock (in-memory image), but
2330 * if we are allowed to, skip spares which already
2331 * have the right event counter, or have one earlier
2332 * (which would mean they aren't being marked as dirty
2333 * with the rest of the array)
2334 */
NeilBrown3cb03002011-10-11 16:45:26 +11002335 struct md_rdev *rdev;
NeilBrowndafb20f2012-03-19 12:46:39 +11002336 rdev_for_each(rdev, mddev) {
NeilBrown42543762006-06-26 00:27:57 -07002337 if (rdev->sb_events == mddev->events ||
2338 (nospares &&
2339 rdev->raid_disk < 0 &&
NeilBrown42543762006-06-26 00:27:57 -07002340 rdev->sb_events+1 == mddev->events)) {
2341 /* Don't update this superblock */
2342 rdev->sb_loaded = 2;
2343 } else {
Jonathan Brassow076f9682011-06-07 17:51:30 -05002344 sync_super(mddev, rdev);
NeilBrown42543762006-06-26 00:27:57 -07002345 rdev->sb_loaded = 1;
2346 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 }
2348}
2349
Goldwyn Rodrigues2aa82192015-09-28 19:21:35 -05002350static bool does_sb_need_changing(struct mddev *mddev)
2351{
2352 struct md_rdev *rdev;
2353 struct mdp_superblock_1 *sb;
2354 int role;
2355
2356 /* Find a good rdev */
2357 rdev_for_each(rdev, mddev)
2358 if ((rdev->raid_disk >= 0) && !test_bit(Faulty, &rdev->flags))
2359 break;
2360
2361 /* No good device found. */
2362 if (!rdev)
2363 return false;
2364
2365 sb = page_address(rdev->sb_page);
2366 /* Check if a device has become faulty or a spare become active */
2367 rdev_for_each(rdev, mddev) {
2368 role = le16_to_cpu(sb->dev_roles[rdev->desc_nr]);
2369 /* Device activated? */
2370 if (role == 0xffff && rdev->raid_disk >=0 &&
2371 !test_bit(Faulty, &rdev->flags))
2372 return true;
2373 /* Device turned faulty? */
2374 if (test_bit(Faulty, &rdev->flags) && (role < 0xfffd))
2375 return true;
2376 }
2377
2378 /* Check if any mddev parameters have changed */
2379 if ((mddev->dev_sectors != le64_to_cpu(sb->size)) ||
2380 (mddev->reshape_position != le64_to_cpu(sb->reshape_position)) ||
Jason Yan13459212017-03-10 11:49:12 +08002381 (mddev->layout != le32_to_cpu(sb->layout)) ||
Goldwyn Rodrigues2aa82192015-09-28 19:21:35 -05002382 (mddev->raid_disks != le32_to_cpu(sb->raid_disks)) ||
2383 (mddev->chunk_sectors != le32_to_cpu(sb->chunksize)))
2384 return true;
2385
2386 return false;
2387}
2388
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05002389void md_update_sb(struct mddev *mddev, int force_change)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390{
NeilBrown3cb03002011-10-11 16:45:26 +11002391 struct md_rdev *rdev;
NeilBrown06d91a52005-06-21 17:17:12 -07002392 int sync_req;
NeilBrown42543762006-06-26 00:27:57 -07002393 int nospares = 0;
NeilBrown2699b672011-07-28 11:31:47 +10002394 int any_badblocks_changed = 0;
Guoqing Jiang23b63f92015-10-12 17:21:30 +08002395 int ret = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396
NeilBrownd87f0642013-04-24 11:42:40 +10002397 if (mddev->ro) {
2398 if (force_change)
Shaohua Li29530792016-12-08 15:48:19 -08002399 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
NeilBrownd87f0642013-04-24 11:42:40 +10002400 return;
2401 }
Goldwyn Rodrigues2aa82192015-09-28 19:21:35 -05002402
Guoqing Jiang2c97cf12016-05-02 11:33:09 -04002403repeat:
Goldwyn Rodrigues2aa82192015-09-28 19:21:35 -05002404 if (mddev_is_clustered(mddev)) {
Shaohua Li29530792016-12-08 15:48:19 -08002405 if (test_and_clear_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags))
Goldwyn Rodrigues2aa82192015-09-28 19:21:35 -05002406 force_change = 1;
Shaohua Li29530792016-12-08 15:48:19 -08002407 if (test_and_clear_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags))
Guoqing Jiang85ad1d12016-05-03 22:22:13 -04002408 nospares = 1;
Guoqing Jiang23b63f92015-10-12 17:21:30 +08002409 ret = md_cluster_ops->metadata_update_start(mddev);
Goldwyn Rodrigues2aa82192015-09-28 19:21:35 -05002410 /* Has someone else has updated the sb */
2411 if (!does_sb_need_changing(mddev)) {
Guoqing Jiang23b63f92015-10-12 17:21:30 +08002412 if (ret == 0)
2413 md_cluster_ops->metadata_update_cancel(mddev);
Shaohua Li29530792016-12-08 15:48:19 -08002414 bit_clear_unless(&mddev->sb_flags, BIT(MD_SB_CHANGE_PENDING),
2415 BIT(MD_SB_CHANGE_DEVS) |
2416 BIT(MD_SB_CHANGE_CLEAN));
Goldwyn Rodrigues2aa82192015-09-28 19:21:35 -05002417 return;
2418 }
2419 }
Guoqing Jiang2c97cf12016-05-02 11:33:09 -04002420
NeilBrown3a3a5dd2010-08-16 18:09:31 +10002421 /* First make sure individual recovery_offsets are correct */
NeilBrowndafb20f2012-03-19 12:46:39 +11002422 rdev_for_each(rdev, mddev) {
NeilBrown3a3a5dd2010-08-16 18:09:31 +10002423 if (rdev->raid_disk >= 0 &&
2424 mddev->delta_disks >= 0 &&
Shaohua Lif2076e72015-10-08 21:54:12 -07002425 !test_bit(Journal, &rdev->flags) &&
NeilBrown3a3a5dd2010-08-16 18:09:31 +10002426 !test_bit(In_sync, &rdev->flags) &&
2427 mddev->curr_resync_completed > rdev->recovery_offset)
2428 rdev->recovery_offset = mddev->curr_resync_completed;
2429
NeilBrownf72ffdd2014-09-30 14:23:59 +10002430 }
Dan Williamsbd52b742010-08-30 17:33:33 +10002431 if (!mddev->persistent) {
Shaohua Li29530792016-12-08 15:48:19 -08002432 clear_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
2433 clear_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
NeilBrownde393cd2011-07-28 11:31:48 +10002434 if (!mddev->external) {
Shaohua Li29530792016-12-08 15:48:19 -08002435 clear_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
NeilBrowndafb20f2012-03-19 12:46:39 +11002436 rdev_for_each(rdev, mddev) {
NeilBrownde393cd2011-07-28 11:31:48 +10002437 if (rdev->badblocks.changed) {
NeilBrownd0962932012-03-19 12:46:41 +11002438 rdev->badblocks.changed = 0;
Vishal Vermafc974ee2015-12-24 19:20:34 -07002439 ack_all_badblocks(&rdev->badblocks);
NeilBrownde393cd2011-07-28 11:31:48 +10002440 md_error(mddev, rdev);
2441 }
2442 clear_bit(Blocked, &rdev->flags);
2443 clear_bit(BlockedBadBlocks, &rdev->flags);
2444 wake_up(&rdev->blocked_wait);
2445 }
2446 }
NeilBrown3a3a5dd2010-08-16 18:09:31 +10002447 wake_up(&mddev->sb_wait);
2448 return;
2449 }
2450
NeilBrown85572d72014-12-15 12:56:56 +11002451 spin_lock(&mddev->lock);
NeilBrown84692192006-08-27 01:23:49 -07002452
Deepa Dinamani9ebc6ef2015-12-21 10:51:01 +11002453 mddev->utime = ktime_get_real_seconds();
NeilBrown3a3a5dd2010-08-16 18:09:31 +10002454
Shaohua Li29530792016-12-08 15:48:19 -08002455 if (test_and_clear_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags))
NeilBrown850b2b42006-10-03 01:15:46 -07002456 force_change = 1;
Shaohua Li29530792016-12-08 15:48:19 -08002457 if (test_and_clear_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags))
NeilBrown850b2b42006-10-03 01:15:46 -07002458 /* just a clean<-> dirty transition, possibly leave spares alone,
2459 * though if events isn't the right even/odd, we will have to do
2460 * spares after all
2461 */
2462 nospares = 1;
2463 if (force_change)
2464 nospares = 0;
2465 if (mddev->degraded)
NeilBrown84692192006-08-27 01:23:49 -07002466 /* If the array is degraded, then skipping spares is both
2467 * dangerous and fairly pointless.
2468 * Dangerous because a device that was removed from the array
2469 * might have a event_count that still looks up-to-date,
2470 * so it can be re-added without a resync.
2471 * Pointless because if there are any spares to skip,
2472 * then a recovery will happen and soon that array won't
2473 * be degraded any more and the spare can go back to sleep then.
2474 */
NeilBrown850b2b42006-10-03 01:15:46 -07002475 nospares = 0;
NeilBrown84692192006-08-27 01:23:49 -07002476
NeilBrown06d91a52005-06-21 17:17:12 -07002477 sync_req = mddev->in_sync;
NeilBrown42543762006-06-26 00:27:57 -07002478
2479 /* If this is just a dirty<->clean transition, and the array is clean
2480 * and 'events' is odd, we can roll back to the previous clean state */
NeilBrown850b2b42006-10-03 01:15:46 -07002481 if (nospares
NeilBrown42543762006-06-26 00:27:57 -07002482 && (mddev->in_sync && mddev->recovery_cp == MaxSector)
NeilBrowna8707c02010-05-18 09:28:43 +10002483 && mddev->can_decrease_events
2484 && mddev->events != 1) {
NeilBrown42543762006-06-26 00:27:57 -07002485 mddev->events--;
NeilBrowna8707c02010-05-18 09:28:43 +10002486 mddev->can_decrease_events = 0;
2487 } else {
NeilBrown42543762006-06-26 00:27:57 -07002488 /* otherwise we have to go forward and ... */
2489 mddev->events ++;
NeilBrowna8707c02010-05-18 09:28:43 +10002490 mddev->can_decrease_events = nospares;
NeilBrown42543762006-06-26 00:27:57 -07002491 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492
NeilBrown403df472014-09-30 15:52:29 +10002493 /*
2494 * This 64-bit counter should never wrap.
2495 * Either we are in around ~1 trillion A.C., assuming
2496 * 1 reboot per second, or we have a bug...
2497 */
2498 WARN_ON(mddev->events == 0);
NeilBrown2699b672011-07-28 11:31:47 +10002499
NeilBrowndafb20f2012-03-19 12:46:39 +11002500 rdev_for_each(rdev, mddev) {
NeilBrown2699b672011-07-28 11:31:47 +10002501 if (rdev->badblocks.changed)
2502 any_badblocks_changed++;
NeilBrownde393cd2011-07-28 11:31:48 +10002503 if (test_bit(Faulty, &rdev->flags))
2504 set_bit(FaultRecorded, &rdev->flags);
2505 }
NeilBrown2699b672011-07-28 11:31:47 +10002506
NeilBrowne6910632008-02-06 01:39:51 -08002507 sync_sbs(mddev, nospares);
NeilBrown85572d72014-12-15 12:56:56 +11002508 spin_unlock(&mddev->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509
NeilBrown36a4e1f2011-10-07 14:23:17 +11002510 pr_debug("md: updating %s RAID superblock on device (in sync %d)\n",
2511 mdname(mddev), mddev->in_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512
Shaohua Li504634f2016-11-18 09:44:08 -08002513 if (mddev->queue)
2514 blk_add_trace_msg(mddev->queue, "md md_update_sb");
NeilBrown46533ff2016-11-18 16:16:11 +11002515rewrite:
NeilBrown4ad13662007-07-17 04:06:13 -07002516 bitmap_update_sb(mddev->bitmap);
NeilBrowndafb20f2012-03-19 12:46:39 +11002517 rdev_for_each(rdev, mddev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518 char b[BDEVNAME_SIZE];
NeilBrown36a4e1f2011-10-07 14:23:17 +11002519
NeilBrown42543762006-06-26 00:27:57 -07002520 if (rdev->sb_loaded != 1)
2521 continue; /* no noise on spare devices */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522
NeilBrownf4667222013-12-09 12:04:56 +11002523 if (!test_bit(Faulty, &rdev->flags)) {
NeilBrown7bfa19f2005-06-21 17:17:28 -07002524 md_super_write(mddev,rdev,
Andre Noll0f420352008-07-11 22:02:23 +10002525 rdev->sb_start, rdev->sb_size,
NeilBrown7bfa19f2005-06-21 17:17:28 -07002526 rdev->sb_page);
NeilBrown36a4e1f2011-10-07 14:23:17 +11002527 pr_debug("md: (write) %s's sb offset: %llu\n",
2528 bdevname(rdev->bdev, b),
2529 (unsigned long long)rdev->sb_start);
NeilBrown42543762006-06-26 00:27:57 -07002530 rdev->sb_events = mddev->events;
NeilBrown2699b672011-07-28 11:31:47 +10002531 if (rdev->badblocks.size) {
2532 md_super_write(mddev, rdev,
2533 rdev->badblocks.sector,
2534 rdev->badblocks.size << 9,
2535 rdev->bb_page);
2536 rdev->badblocks.size = 0;
2537 }
NeilBrown7bfa19f2005-06-21 17:17:28 -07002538
NeilBrownf4667222013-12-09 12:04:56 +11002539 } else
NeilBrown36a4e1f2011-10-07 14:23:17 +11002540 pr_debug("md: %s (skipping faulty)\n",
2541 bdevname(rdev->bdev, b));
Andrei Warkentind70ed2e2011-10-18 12:16:48 +11002542
NeilBrown7bfa19f2005-06-21 17:17:28 -07002543 if (mddev->level == LEVEL_MULTIPATH)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 /* only need to write one superblock... */
2545 break;
2546 }
NeilBrown46533ff2016-11-18 16:16:11 +11002547 if (md_super_wait(mddev) < 0)
2548 goto rewrite;
Shaohua Li29530792016-12-08 15:48:19 -08002549 /* if there was a failure, MD_SB_CHANGE_DEVS was set, and we re-write super */
NeilBrown7bfa19f2005-06-21 17:17:28 -07002550
Guoqing Jiang2c97cf12016-05-02 11:33:09 -04002551 if (mddev_is_clustered(mddev) && ret == 0)
2552 md_cluster_ops->metadata_update_finish(mddev);
2553
NeilBrown850b2b42006-10-03 01:15:46 -07002554 if (mddev->in_sync != sync_req ||
Shaohua Li29530792016-12-08 15:48:19 -08002555 !bit_clear_unless(&mddev->sb_flags, BIT(MD_SB_CHANGE_PENDING),
2556 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_CLEAN)))
NeilBrown06d91a52005-06-21 17:17:12 -07002557 /* have to write it out again */
NeilBrown06d91a52005-06-21 17:17:12 -07002558 goto repeat;
NeilBrown3d310eb2005-06-21 17:17:26 -07002559 wake_up(&mddev->sb_wait);
NeilBrownacb180b2009-04-14 16:28:34 +10002560 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
2561 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
NeilBrown06d91a52005-06-21 17:17:12 -07002562
NeilBrowndafb20f2012-03-19 12:46:39 +11002563 rdev_for_each(rdev, mddev) {
NeilBrownde393cd2011-07-28 11:31:48 +10002564 if (test_and_clear_bit(FaultRecorded, &rdev->flags))
2565 clear_bit(Blocked, &rdev->flags);
2566
2567 if (any_badblocks_changed)
Vishal Vermafc974ee2015-12-24 19:20:34 -07002568 ack_all_badblocks(&rdev->badblocks);
NeilBrownde393cd2011-07-28 11:31:48 +10002569 clear_bit(BlockedBadBlocks, &rdev->flags);
2570 wake_up(&rdev->blocked_wait);
2571 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572}
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05002573EXPORT_SYMBOL(md_update_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574
Goldwyn Rodriguesa6da4ef2015-04-14 10:45:22 -05002575static int add_bound_rdev(struct md_rdev *rdev)
2576{
2577 struct mddev *mddev = rdev->mddev;
2578 int err = 0;
Shaohua Li87d4d912016-01-06 14:37:14 -08002579 bool add_journal = test_bit(Journal, &rdev->flags);
Goldwyn Rodriguesa6da4ef2015-04-14 10:45:22 -05002580
Shaohua Li87d4d912016-01-06 14:37:14 -08002581 if (!mddev->pers->hot_remove_disk || add_journal) {
Goldwyn Rodriguesa6da4ef2015-04-14 10:45:22 -05002582 /* If there is hot_add_disk but no hot_remove_disk
2583 * then added disks for geometry changes,
2584 * and should be added immediately.
2585 */
2586 super_types[mddev->major_version].
2587 validate_super(mddev, rdev);
Shaohua Li87d4d912016-01-06 14:37:14 -08002588 if (add_journal)
2589 mddev_suspend(mddev);
Goldwyn Rodriguesa6da4ef2015-04-14 10:45:22 -05002590 err = mddev->pers->hot_add_disk(mddev, rdev);
Shaohua Li87d4d912016-01-06 14:37:14 -08002591 if (add_journal)
2592 mddev_resume(mddev);
Goldwyn Rodriguesa6da4ef2015-04-14 10:45:22 -05002593 if (err) {
Guoqing Jiangdb767672016-06-02 23:32:05 -04002594 md_kick_rdev_from_array(rdev);
Goldwyn Rodriguesa6da4ef2015-04-14 10:45:22 -05002595 return err;
2596 }
2597 }
2598 sysfs_notify_dirent_safe(rdev->sysfs_state);
2599
Shaohua Li29530792016-12-08 15:48:19 -08002600 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
Goldwyn Rodriguesa6da4ef2015-04-14 10:45:22 -05002601 if (mddev->degraded)
2602 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
2603 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2604 md_new_event(mddev);
2605 md_wakeup_thread(mddev->thread);
2606 return 0;
2607}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608
Andre Noll7f6ce762008-03-23 18:34:54 +01002609/* words written to sysfs files may, or may not, be \n terminated.
NeilBrownbce74da2006-01-06 00:20:41 -08002610 * We want to accept with case. For this we use cmd_match.
2611 */
2612static int cmd_match(const char *cmd, const char *str)
2613{
2614 /* See if cmd, written into a sysfs file, matches
2615 * str. They must either be the same, or cmd can
2616 * have a trailing newline
2617 */
2618 while (*cmd && *str && *cmd == *str) {
2619 cmd++;
2620 str++;
2621 }
2622 if (*cmd == '\n')
2623 cmd++;
2624 if (*str || *cmd)
2625 return 0;
2626 return 1;
2627}
2628
NeilBrown86e6ffd2005-11-08 21:39:24 -08002629struct rdev_sysfs_entry {
2630 struct attribute attr;
NeilBrown3cb03002011-10-11 16:45:26 +11002631 ssize_t (*show)(struct md_rdev *, char *);
2632 ssize_t (*store)(struct md_rdev *, const char *, size_t);
NeilBrown86e6ffd2005-11-08 21:39:24 -08002633};
2634
2635static ssize_t
NeilBrown3cb03002011-10-11 16:45:26 +11002636state_show(struct md_rdev *rdev, char *page)
NeilBrown86e6ffd2005-11-08 21:39:24 -08002637{
Tomasz Majchrzak35b785f2016-10-21 16:26:57 +02002638 char *sep = ",";
NeilBrown20a49ff2008-02-06 01:39:57 -08002639 size_t len = 0;
NeilBrown758bfc82014-12-15 12:56:59 +11002640 unsigned long flags = ACCESS_ONCE(rdev->flags);
NeilBrown86e6ffd2005-11-08 21:39:24 -08002641
NeilBrown758bfc82014-12-15 12:56:59 +11002642 if (test_bit(Faulty, &flags) ||
Tomasz Majchrzakdcbcb482016-10-21 16:27:08 +02002643 (!test_bit(ExternalBbl, &flags) &&
2644 rdev->badblocks.unacked_exist))
Tomasz Majchrzak35b785f2016-10-21 16:26:57 +02002645 len += sprintf(page+len, "faulty%s", sep);
2646 if (test_bit(In_sync, &flags))
2647 len += sprintf(page+len, "in_sync%s", sep);
2648 if (test_bit(Journal, &flags))
2649 len += sprintf(page+len, "journal%s", sep);
2650 if (test_bit(WriteMostly, &flags))
2651 len += sprintf(page+len, "write_mostly%s", sep);
NeilBrown758bfc82014-12-15 12:56:59 +11002652 if (test_bit(Blocked, &flags) ||
NeilBrown52c64152011-12-08 16:22:48 +11002653 (rdev->badblocks.unacked_exist
Tomasz Majchrzak35b785f2016-10-21 16:26:57 +02002654 && !test_bit(Faulty, &flags)))
2655 len += sprintf(page+len, "blocked%s", sep);
NeilBrown758bfc82014-12-15 12:56:59 +11002656 if (!test_bit(Faulty, &flags) &&
Shaohua Lif2076e72015-10-08 21:54:12 -07002657 !test_bit(Journal, &flags) &&
Tomasz Majchrzak35b785f2016-10-21 16:26:57 +02002658 !test_bit(In_sync, &flags))
2659 len += sprintf(page+len, "spare%s", sep);
2660 if (test_bit(WriteErrorSeen, &flags))
2661 len += sprintf(page+len, "write_error%s", sep);
2662 if (test_bit(WantReplacement, &flags))
2663 len += sprintf(page+len, "want_replacement%s", sep);
2664 if (test_bit(Replacement, &flags))
2665 len += sprintf(page+len, "replacement%s", sep);
2666 if (test_bit(ExternalBbl, &flags))
2667 len += sprintf(page+len, "external_bbl%s", sep);
NeilBrown688834e2016-11-18 16:16:11 +11002668 if (test_bit(FailFast, &flags))
2669 len += sprintf(page+len, "failfast%s", sep);
Tomasz Majchrzak35b785f2016-10-21 16:26:57 +02002670
2671 if (len)
2672 len -= strlen(sep);
NeilBrown2d78f8c2011-12-23 10:17:51 +11002673
NeilBrown86e6ffd2005-11-08 21:39:24 -08002674 return len+sprintf(page+len, "\n");
2675}
2676
NeilBrown45dc2de2006-06-26 00:27:58 -07002677static ssize_t
NeilBrown3cb03002011-10-11 16:45:26 +11002678state_store(struct md_rdev *rdev, const char *buf, size_t len)
NeilBrown45dc2de2006-06-26 00:27:58 -07002679{
2680 /* can write
NeilBrownde393cd2011-07-28 11:31:48 +10002681 * faulty - simulates an error
NeilBrown45dc2de2006-06-26 00:27:58 -07002682 * remove - disconnects the device
NeilBrownf6556752006-06-26 00:28:01 -07002683 * writemostly - sets write_mostly
2684 * -writemostly - clears write_mostly
NeilBrownde393cd2011-07-28 11:31:48 +10002685 * blocked - sets the Blocked flags
2686 * -blocked - clears the Blocked and possibly simulates an error
NeilBrown6d56e272009-04-14 12:01:57 +10002687 * insync - sets Insync providing device isn't active
NeilBrownf4667222013-12-09 12:04:56 +11002688 * -insync - clear Insync for a device with a slot assigned,
2689 * so that it gets rebuilt based on bitmap
NeilBrownd7a9d442011-07-28 11:31:48 +10002690 * write_error - sets WriteErrorSeen
2691 * -write_error - clears WriteErrorSeen
NeilBrown688834e2016-11-18 16:16:11 +11002692 * {,-}failfast - set/clear FailFast
NeilBrown45dc2de2006-06-26 00:27:58 -07002693 */
2694 int err = -EINVAL;
2695 if (cmd_match(buf, "faulty") && rdev->mddev->pers) {
2696 md_error(rdev->mddev, rdev);
NeilBrown5ef56c82011-08-25 14:42:51 +10002697 if (test_bit(Faulty, &rdev->flags))
2698 err = 0;
2699 else
2700 err = -EBUSY;
NeilBrown45dc2de2006-06-26 00:27:58 -07002701 } else if (cmd_match(buf, "remove")) {
Shaohua Li5d881782016-07-28 09:06:34 -07002702 if (rdev->mddev->pers) {
2703 clear_bit(Blocked, &rdev->flags);
2704 remove_and_add_spares(rdev->mddev, rdev);
2705 }
NeilBrown45dc2de2006-06-26 00:27:58 -07002706 if (rdev->raid_disk >= 0)
2707 err = -EBUSY;
2708 else {
NeilBrownfd01b882011-10-11 16:47:53 +11002709 struct mddev *mddev = rdev->mddev;
NeilBrown45dc2de2006-06-26 00:27:58 -07002710 err = 0;
Guoqing Jianga9720902015-10-12 17:21:27 +08002711 if (mddev_is_clustered(mddev))
2712 err = md_cluster_ops->remove_disk(mddev, rdev);
2713
2714 if (err == 0) {
2715 md_kick_rdev_from_array(rdev);
NeilBrown060b0682016-11-04 16:46:03 +11002716 if (mddev->pers) {
Shaohua Li29530792016-12-08 15:48:19 -08002717 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
NeilBrown060b0682016-11-04 16:46:03 +11002718 md_wakeup_thread(mddev->thread);
2719 }
Guoqing Jianga9720902015-10-12 17:21:27 +08002720 md_new_event(mddev);
2721 }
NeilBrown45dc2de2006-06-26 00:27:58 -07002722 }
NeilBrownf6556752006-06-26 00:28:01 -07002723 } else if (cmd_match(buf, "writemostly")) {
2724 set_bit(WriteMostly, &rdev->flags);
2725 err = 0;
2726 } else if (cmd_match(buf, "-writemostly")) {
2727 clear_bit(WriteMostly, &rdev->flags);
2728 err = 0;
Dan Williams6bfe0b42008-04-30 00:52:32 -07002729 } else if (cmd_match(buf, "blocked")) {
2730 set_bit(Blocked, &rdev->flags);
2731 err = 0;
2732 } else if (cmd_match(buf, "-blocked")) {
NeilBrownde393cd2011-07-28 11:31:48 +10002733 if (!test_bit(Faulty, &rdev->flags) &&
Tomasz Majchrzakdcbcb482016-10-21 16:27:08 +02002734 !test_bit(ExternalBbl, &rdev->flags) &&
NeilBrown7da64a02011-08-30 16:20:17 +10002735 rdev->badblocks.unacked_exist) {
NeilBrownde393cd2011-07-28 11:31:48 +10002736 /* metadata handler doesn't understand badblocks,
2737 * so we need to fail the device
2738 */
2739 md_error(rdev->mddev, rdev);
2740 }
Dan Williams6bfe0b42008-04-30 00:52:32 -07002741 clear_bit(Blocked, &rdev->flags);
NeilBrownde393cd2011-07-28 11:31:48 +10002742 clear_bit(BlockedBadBlocks, &rdev->flags);
Dan Williams6bfe0b42008-04-30 00:52:32 -07002743 wake_up(&rdev->blocked_wait);
2744 set_bit(MD_RECOVERY_NEEDED, &rdev->mddev->recovery);
2745 md_wakeup_thread(rdev->mddev->thread);
2746
2747 err = 0;
NeilBrown6d56e272009-04-14 12:01:57 +10002748 } else if (cmd_match(buf, "insync") && rdev->raid_disk == -1) {
2749 set_bit(In_sync, &rdev->flags);
2750 err = 0;
NeilBrown688834e2016-11-18 16:16:11 +11002751 } else if (cmd_match(buf, "failfast")) {
2752 set_bit(FailFast, &rdev->flags);
2753 err = 0;
2754 } else if (cmd_match(buf, "-failfast")) {
2755 clear_bit(FailFast, &rdev->flags);
2756 err = 0;
Shaohua Lif2076e72015-10-08 21:54:12 -07002757 } else if (cmd_match(buf, "-insync") && rdev->raid_disk >= 0 &&
2758 !test_bit(Journal, &rdev->flags)) {
NeilBrowne1960f82014-09-30 15:24:25 +10002759 if (rdev->mddev->pers == NULL) {
2760 clear_bit(In_sync, &rdev->flags);
2761 rdev->saved_raid_disk = rdev->raid_disk;
2762 rdev->raid_disk = -1;
2763 err = 0;
2764 }
NeilBrownd7a9d442011-07-28 11:31:48 +10002765 } else if (cmd_match(buf, "write_error")) {
2766 set_bit(WriteErrorSeen, &rdev->flags);
2767 err = 0;
2768 } else if (cmd_match(buf, "-write_error")) {
2769 clear_bit(WriteErrorSeen, &rdev->flags);
2770 err = 0;
NeilBrown2d78f8c2011-12-23 10:17:51 +11002771 } else if (cmd_match(buf, "want_replacement")) {
2772 /* Any non-spare device that is not a replacement can
2773 * become want_replacement at any time, but we then need to
2774 * check if recovery is needed.
2775 */
2776 if (rdev->raid_disk >= 0 &&
Shaohua Lif2076e72015-10-08 21:54:12 -07002777 !test_bit(Journal, &rdev->flags) &&
NeilBrown2d78f8c2011-12-23 10:17:51 +11002778 !test_bit(Replacement, &rdev->flags))
2779 set_bit(WantReplacement, &rdev->flags);
2780 set_bit(MD_RECOVERY_NEEDED, &rdev->mddev->recovery);
2781 md_wakeup_thread(rdev->mddev->thread);
2782 err = 0;
2783 } else if (cmd_match(buf, "-want_replacement")) {
2784 /* Clearing 'want_replacement' is always allowed.
2785 * Once replacements starts it is too late though.
2786 */
2787 err = 0;
2788 clear_bit(WantReplacement, &rdev->flags);
2789 } else if (cmd_match(buf, "replacement")) {
2790 /* Can only set a device as a replacement when array has not
2791 * yet been started. Once running, replacement is automatic
2792 * from spares, or by assigning 'slot'.
2793 */
2794 if (rdev->mddev->pers)
2795 err = -EBUSY;
2796 else {
2797 set_bit(Replacement, &rdev->flags);
2798 err = 0;
2799 }
2800 } else if (cmd_match(buf, "-replacement")) {
2801 /* Similarly, can only clear Replacement before start */
2802 if (rdev->mddev->pers)
2803 err = -EBUSY;
2804 else {
2805 clear_bit(Replacement, &rdev->flags);
2806 err = 0;
2807 }
Goldwyn Rodriguesa6da4ef2015-04-14 10:45:22 -05002808 } else if (cmd_match(buf, "re-add")) {
2809 if (test_bit(Faulty, &rdev->flags) && (rdev->raid_disk == -1)) {
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -05002810 /* clear_bit is performed _after_ all the devices
2811 * have their local Faulty bit cleared. If any writes
2812 * happen in the meantime in the local node, they
2813 * will land in the local bitmap, which will be synced
2814 * by this node eventually
2815 */
2816 if (!mddev_is_clustered(rdev->mddev) ||
2817 (err = md_cluster_ops->gather_bitmaps(rdev)) == 0) {
2818 clear_bit(Faulty, &rdev->flags);
2819 err = add_bound_rdev(rdev);
2820 }
Goldwyn Rodriguesa6da4ef2015-04-14 10:45:22 -05002821 } else
2822 err = -EBUSY;
Tomasz Majchrzak35b785f2016-10-21 16:26:57 +02002823 } else if (cmd_match(buf, "external_bbl") && (rdev->mddev->external)) {
2824 set_bit(ExternalBbl, &rdev->flags);
2825 rdev->badblocks.shift = 0;
2826 err = 0;
2827 } else if (cmd_match(buf, "-external_bbl") && (rdev->mddev->external)) {
2828 clear_bit(ExternalBbl, &rdev->flags);
2829 err = 0;
NeilBrown45dc2de2006-06-26 00:27:58 -07002830 }
NeilBrown00bcb4a2010-06-01 19:37:23 +10002831 if (!err)
2832 sysfs_notify_dirent_safe(rdev->sysfs_state);
NeilBrown45dc2de2006-06-26 00:27:58 -07002833 return err ? err : len;
2834}
NeilBrown80ca3a42006-07-10 04:44:18 -07002835static struct rdev_sysfs_entry rdev_state =
NeilBrown750f1992014-09-30 08:53:05 +10002836__ATTR_PREALLOC(state, S_IRUGO|S_IWUSR, state_show, state_store);
NeilBrown86e6ffd2005-11-08 21:39:24 -08002837
2838static ssize_t
NeilBrown3cb03002011-10-11 16:45:26 +11002839errors_show(struct md_rdev *rdev, char *page)
NeilBrown4dbcdc72006-01-06 00:20:52 -08002840{
2841 return sprintf(page, "%d\n", atomic_read(&rdev->corrected_errors));
2842}
2843
2844static ssize_t
NeilBrown3cb03002011-10-11 16:45:26 +11002845errors_store(struct md_rdev *rdev, const char *buf, size_t len)
NeilBrown4dbcdc72006-01-06 00:20:52 -08002846{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03002847 unsigned int n;
2848 int rv;
2849
2850 rv = kstrtouint(buf, 10, &n);
2851 if (rv < 0)
2852 return rv;
2853 atomic_set(&rdev->corrected_errors, n);
2854 return len;
NeilBrown4dbcdc72006-01-06 00:20:52 -08002855}
2856static struct rdev_sysfs_entry rdev_errors =
NeilBrown80ca3a42006-07-10 04:44:18 -07002857__ATTR(errors, S_IRUGO|S_IWUSR, errors_show, errors_store);
NeilBrown4dbcdc72006-01-06 00:20:52 -08002858
NeilBrown014236d2006-01-06 00:20:55 -08002859static ssize_t
NeilBrown3cb03002011-10-11 16:45:26 +11002860slot_show(struct md_rdev *rdev, char *page)
NeilBrown014236d2006-01-06 00:20:55 -08002861{
Shaohua Lif2076e72015-10-08 21:54:12 -07002862 if (test_bit(Journal, &rdev->flags))
2863 return sprintf(page, "journal\n");
2864 else if (rdev->raid_disk < 0)
NeilBrown014236d2006-01-06 00:20:55 -08002865 return sprintf(page, "none\n");
2866 else
2867 return sprintf(page, "%d\n", rdev->raid_disk);
2868}
2869
2870static ssize_t
NeilBrown3cb03002011-10-11 16:45:26 +11002871slot_store(struct md_rdev *rdev, const char *buf, size_t len)
NeilBrown014236d2006-01-06 00:20:55 -08002872{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03002873 int slot;
NeilBrownc303da62008-02-06 01:39:51 -08002874 int err;
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03002875
Shaohua Lif2076e72015-10-08 21:54:12 -07002876 if (test_bit(Journal, &rdev->flags))
2877 return -EBUSY;
NeilBrown014236d2006-01-06 00:20:55 -08002878 if (strncmp(buf, "none", 4)==0)
2879 slot = -1;
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03002880 else {
2881 err = kstrtouint(buf, 10, (unsigned int *)&slot);
2882 if (err < 0)
2883 return err;
2884 }
Neil Brown6c2fce22008-06-28 08:31:31 +10002885 if (rdev->mddev->pers && slot == -1) {
NeilBrownc303da62008-02-06 01:39:51 -08002886 /* Setting 'slot' on an active array requires also
2887 * updating the 'rd%d' link, and communicating
2888 * with the personality with ->hot_*_disk.
2889 * For now we only support removing
2890 * failed/spare devices. This normally happens automatically,
2891 * but not when the metadata is externally managed.
2892 */
NeilBrownc303da62008-02-06 01:39:51 -08002893 if (rdev->raid_disk == -1)
2894 return -EEXIST;
2895 /* personality does all needed checks */
Namhyung Kim01393f32011-06-09 11:42:54 +10002896 if (rdev->mddev->pers->hot_remove_disk == NULL)
NeilBrownc303da62008-02-06 01:39:51 -08002897 return -EINVAL;
NeilBrown746d3202013-04-24 11:42:41 +10002898 clear_bit(Blocked, &rdev->flags);
2899 remove_and_add_spares(rdev->mddev, rdev);
2900 if (rdev->raid_disk >= 0)
2901 return -EBUSY;
NeilBrownc303da62008-02-06 01:39:51 -08002902 set_bit(MD_RECOVERY_NEEDED, &rdev->mddev->recovery);
2903 md_wakeup_thread(rdev->mddev->thread);
Neil Brown6c2fce22008-06-28 08:31:31 +10002904 } else if (rdev->mddev->pers) {
Neil Brown6c2fce22008-06-28 08:31:31 +10002905 /* Activating a spare .. or possibly reactivating
NeilBrown6d56e272009-04-14 12:01:57 +10002906 * if we ever get bitmaps working here.
Neil Brown6c2fce22008-06-28 08:31:31 +10002907 */
Goldwyn Rodriguescb01c542015-12-18 15:19:16 +11002908 int err;
Neil Brown6c2fce22008-06-28 08:31:31 +10002909
2910 if (rdev->raid_disk != -1)
2911 return -EBUSY;
2912
NeilBrownc6751b22011-02-02 11:57:13 +11002913 if (test_bit(MD_RECOVERY_RUNNING, &rdev->mddev->recovery))
2914 return -EBUSY;
2915
Neil Brown6c2fce22008-06-28 08:31:31 +10002916 if (rdev->mddev->pers->hot_add_disk == NULL)
2917 return -EINVAL;
2918
NeilBrownba1b41b2011-01-14 09:14:34 +11002919 if (slot >= rdev->mddev->raid_disks &&
2920 slot >= rdev->mddev->raid_disks + rdev->mddev->delta_disks)
2921 return -ENOSPC;
2922
Neil Brown6c2fce22008-06-28 08:31:31 +10002923 rdev->raid_disk = slot;
2924 if (test_bit(In_sync, &rdev->flags))
2925 rdev->saved_raid_disk = slot;
2926 else
2927 rdev->saved_raid_disk = -1;
NeilBrownd30519f2011-10-18 12:13:47 +11002928 clear_bit(In_sync, &rdev->flags);
NeilBrown8313b8e2013-12-12 10:13:33 +11002929 clear_bit(Bitmap_sync, &rdev->flags);
Goldwyn Rodriguescb01c542015-12-18 15:19:16 +11002930 err = rdev->mddev->pers->
2931 hot_add_disk(rdev->mddev, rdev);
2932 if (err) {
2933 rdev->raid_disk = -1;
2934 return err;
2935 } else
2936 sysfs_notify_dirent_safe(rdev->sysfs_state);
2937 if (sysfs_link_rdev(rdev->mddev, rdev))
2938 /* failure here is OK */;
Neil Brown6c2fce22008-06-28 08:31:31 +10002939 /* don't wakeup anyone, leave that to userspace. */
NeilBrownc303da62008-02-06 01:39:51 -08002940 } else {
NeilBrownba1b41b2011-01-14 09:14:34 +11002941 if (slot >= rdev->mddev->raid_disks &&
2942 slot >= rdev->mddev->raid_disks + rdev->mddev->delta_disks)
NeilBrownc303da62008-02-06 01:39:51 -08002943 return -ENOSPC;
2944 rdev->raid_disk = slot;
2945 /* assume it is working */
NeilBrownc5d79ad2008-02-06 01:39:54 -08002946 clear_bit(Faulty, &rdev->flags);
2947 clear_bit(WriteMostly, &rdev->flags);
NeilBrownc303da62008-02-06 01:39:51 -08002948 set_bit(In_sync, &rdev->flags);
NeilBrown00bcb4a2010-06-01 19:37:23 +10002949 sysfs_notify_dirent_safe(rdev->sysfs_state);
NeilBrownc303da62008-02-06 01:39:51 -08002950 }
NeilBrown014236d2006-01-06 00:20:55 -08002951 return len;
2952}
2953
NeilBrown014236d2006-01-06 00:20:55 -08002954static struct rdev_sysfs_entry rdev_slot =
NeilBrown80ca3a42006-07-10 04:44:18 -07002955__ATTR(slot, S_IRUGO|S_IWUSR, slot_show, slot_store);
NeilBrown014236d2006-01-06 00:20:55 -08002956
NeilBrown93c8cad2006-01-06 00:20:56 -08002957static ssize_t
NeilBrown3cb03002011-10-11 16:45:26 +11002958offset_show(struct md_rdev *rdev, char *page)
NeilBrown93c8cad2006-01-06 00:20:56 -08002959{
Andrew Morton6961ece2006-01-06 00:20:59 -08002960 return sprintf(page, "%llu\n", (unsigned long long)rdev->data_offset);
NeilBrown93c8cad2006-01-06 00:20:56 -08002961}
2962
2963static ssize_t
NeilBrown3cb03002011-10-11 16:45:26 +11002964offset_store(struct md_rdev *rdev, const char *buf, size_t len)
NeilBrown93c8cad2006-01-06 00:20:56 -08002965{
NeilBrownc6563a82012-05-21 09:27:00 +10002966 unsigned long long offset;
Jingoo Hanb29bebd2013-06-01 16:15:16 +09002967 if (kstrtoull(buf, 10, &offset) < 0)
NeilBrown93c8cad2006-01-06 00:20:56 -08002968 return -EINVAL;
Neil Brown8ed0a522008-06-28 08:31:29 +10002969 if (rdev->mddev->pers && rdev->raid_disk >= 0)
NeilBrown93c8cad2006-01-06 00:20:56 -08002970 return -EBUSY;
Andre Nolldd8ac332009-03-31 14:33:13 +11002971 if (rdev->sectors && rdev->mddev->external)
NeilBrownc5d79ad2008-02-06 01:39:54 -08002972 /* Must set offset before size, so overlap checks
2973 * can be sane */
2974 return -EBUSY;
NeilBrown93c8cad2006-01-06 00:20:56 -08002975 rdev->data_offset = offset;
NeilBrown25f7fd42012-07-19 15:59:18 +10002976 rdev->new_data_offset = offset;
NeilBrown93c8cad2006-01-06 00:20:56 -08002977 return len;
2978}
2979
2980static struct rdev_sysfs_entry rdev_offset =
NeilBrown80ca3a42006-07-10 04:44:18 -07002981__ATTR(offset, S_IRUGO|S_IWUSR, offset_show, offset_store);
NeilBrown93c8cad2006-01-06 00:20:56 -08002982
NeilBrownc6563a82012-05-21 09:27:00 +10002983static ssize_t new_offset_show(struct md_rdev *rdev, char *page)
2984{
2985 return sprintf(page, "%llu\n",
2986 (unsigned long long)rdev->new_data_offset);
2987}
2988
2989static ssize_t new_offset_store(struct md_rdev *rdev,
2990 const char *buf, size_t len)
2991{
2992 unsigned long long new_offset;
2993 struct mddev *mddev = rdev->mddev;
2994
Jingoo Hanb29bebd2013-06-01 16:15:16 +09002995 if (kstrtoull(buf, 10, &new_offset) < 0)
NeilBrownc6563a82012-05-21 09:27:00 +10002996 return -EINVAL;
2997
NeilBrownf851b602014-12-11 10:02:10 +11002998 if (mddev->sync_thread ||
2999 test_bit(MD_RECOVERY_RUNNING,&mddev->recovery))
NeilBrownc6563a82012-05-21 09:27:00 +10003000 return -EBUSY;
3001 if (new_offset == rdev->data_offset)
3002 /* reset is always permitted */
3003 ;
3004 else if (new_offset > rdev->data_offset) {
3005 /* must not push array size beyond rdev_sectors */
3006 if (new_offset - rdev->data_offset
3007 + mddev->dev_sectors > rdev->sectors)
3008 return -E2BIG;
3009 }
3010 /* Metadata worries about other space details. */
3011
3012 /* decreasing the offset is inconsistent with a backwards
3013 * reshape.
3014 */
3015 if (new_offset < rdev->data_offset &&
3016 mddev->reshape_backwards)
3017 return -EINVAL;
3018 /* Increasing offset is inconsistent with forwards
3019 * reshape. reshape_direction should be set to
3020 * 'backwards' first.
3021 */
3022 if (new_offset > rdev->data_offset &&
3023 !mddev->reshape_backwards)
3024 return -EINVAL;
3025
3026 if (mddev->pers && mddev->persistent &&
3027 !super_types[mddev->major_version]
3028 .allow_new_offset(rdev, new_offset))
3029 return -E2BIG;
3030 rdev->new_data_offset = new_offset;
3031 if (new_offset > rdev->data_offset)
3032 mddev->reshape_backwards = 1;
3033 else if (new_offset < rdev->data_offset)
3034 mddev->reshape_backwards = 0;
3035
3036 return len;
3037}
3038static struct rdev_sysfs_entry rdev_new_offset =
3039__ATTR(new_offset, S_IRUGO|S_IWUSR, new_offset_show, new_offset_store);
3040
NeilBrown83303b62006-01-06 00:21:06 -08003041static ssize_t
NeilBrown3cb03002011-10-11 16:45:26 +11003042rdev_size_show(struct md_rdev *rdev, char *page)
NeilBrown83303b62006-01-06 00:21:06 -08003043{
Andre Nolldd8ac332009-03-31 14:33:13 +11003044 return sprintf(page, "%llu\n", (unsigned long long)rdev->sectors / 2);
NeilBrown83303b62006-01-06 00:21:06 -08003045}
3046
NeilBrownc5d79ad2008-02-06 01:39:54 -08003047static int overlaps(sector_t s1, sector_t l1, sector_t s2, sector_t l2)
3048{
3049 /* check if two start/length pairs overlap */
3050 if (s1+l1 <= s2)
3051 return 0;
3052 if (s2+l2 <= s1)
3053 return 0;
3054 return 1;
3055}
3056
Dan Williamsb522adc2009-03-31 15:00:31 +11003057static int strict_blocks_to_sectors(const char *buf, sector_t *sectors)
3058{
3059 unsigned long long blocks;
3060 sector_t new;
3061
Jingoo Hanb29bebd2013-06-01 16:15:16 +09003062 if (kstrtoull(buf, 10, &blocks) < 0)
Dan Williamsb522adc2009-03-31 15:00:31 +11003063 return -EINVAL;
3064
3065 if (blocks & 1ULL << (8 * sizeof(blocks) - 1))
3066 return -EINVAL; /* sector conversion overflow */
3067
3068 new = blocks * 2;
3069 if (new != blocks * 2)
3070 return -EINVAL; /* unsigned long long to sector_t overflow */
3071
3072 *sectors = new;
3073 return 0;
3074}
3075
NeilBrown83303b62006-01-06 00:21:06 -08003076static ssize_t
NeilBrown3cb03002011-10-11 16:45:26 +11003077rdev_size_store(struct md_rdev *rdev, const char *buf, size_t len)
NeilBrown83303b62006-01-06 00:21:06 -08003078{
NeilBrownfd01b882011-10-11 16:47:53 +11003079 struct mddev *my_mddev = rdev->mddev;
Andre Nolldd8ac332009-03-31 14:33:13 +11003080 sector_t oldsectors = rdev->sectors;
Dan Williamsb522adc2009-03-31 15:00:31 +11003081 sector_t sectors;
NeilBrown27c529b2008-03-04 14:29:33 -08003082
Shaohua Lif2076e72015-10-08 21:54:12 -07003083 if (test_bit(Journal, &rdev->flags))
3084 return -EBUSY;
Dan Williamsb522adc2009-03-31 15:00:31 +11003085 if (strict_blocks_to_sectors(buf, &sectors) < 0)
Neil Brownd7027452008-07-12 10:37:50 +10003086 return -EINVAL;
NeilBrownc6563a82012-05-21 09:27:00 +10003087 if (rdev->data_offset != rdev->new_data_offset)
3088 return -EINVAL; /* too confusing */
Chris Webb0cd17fe2008-06-28 08:31:46 +10003089 if (my_mddev->pers && rdev->raid_disk >= 0) {
Neil Brownd7027452008-07-12 10:37:50 +10003090 if (my_mddev->persistent) {
Andre Nolldd8ac332009-03-31 14:33:13 +11003091 sectors = super_types[my_mddev->major_version].
3092 rdev_size_change(rdev, sectors);
3093 if (!sectors)
Chris Webb0cd17fe2008-06-28 08:31:46 +10003094 return -EBUSY;
Andre Nolldd8ac332009-03-31 14:33:13 +11003095 } else if (!sectors)
Mike Snitzer77304d22010-11-08 14:39:12 +01003096 sectors = (i_size_read(rdev->bdev->bd_inode) >> 9) -
Andre Nolldd8ac332009-03-31 14:33:13 +11003097 rdev->data_offset;
NeilBrowna6468532013-02-21 14:33:17 +11003098 if (!my_mddev->pers->resize)
3099 /* Cannot change size for RAID0 or Linear etc */
3100 return -EINVAL;
Chris Webb0cd17fe2008-06-28 08:31:46 +10003101 }
Andre Nolldd8ac332009-03-31 14:33:13 +11003102 if (sectors < my_mddev->dev_sectors)
Chris Webb7d3c6f82008-10-13 11:55:11 +11003103 return -EINVAL; /* component must fit device */
Chris Webb0cd17fe2008-06-28 08:31:46 +10003104
Andre Nolldd8ac332009-03-31 14:33:13 +11003105 rdev->sectors = sectors;
3106 if (sectors > oldsectors && my_mddev->external) {
NeilBrown8b1afc32014-09-29 15:33:20 +10003107 /* Need to check that all other rdevs with the same
3108 * ->bdev do not overlap. 'rcu' is sufficient to walk
3109 * the rdev lists safely.
3110 * This check does not provide a hard guarantee, it
3111 * just helps avoid dangerous mistakes.
NeilBrownc5d79ad2008-02-06 01:39:54 -08003112 */
NeilBrownfd01b882011-10-11 16:47:53 +11003113 struct mddev *mddev;
NeilBrownc5d79ad2008-02-06 01:39:54 -08003114 int overlap = 0;
Cheng Renquan159ec1f2009-01-09 08:31:08 +11003115 struct list_head *tmp;
NeilBrownc5d79ad2008-02-06 01:39:54 -08003116
NeilBrown8b1afc32014-09-29 15:33:20 +10003117 rcu_read_lock();
NeilBrown29ac4aa2008-02-06 01:39:58 -08003118 for_each_mddev(mddev, tmp) {
NeilBrown3cb03002011-10-11 16:45:26 +11003119 struct md_rdev *rdev2;
NeilBrownc5d79ad2008-02-06 01:39:54 -08003120
NeilBrowndafb20f2012-03-19 12:46:39 +11003121 rdev_for_each(rdev2, mddev)
NeilBrownf21e9ff2011-01-31 12:10:09 +11003122 if (rdev->bdev == rdev2->bdev &&
3123 rdev != rdev2 &&
3124 overlaps(rdev->data_offset, rdev->sectors,
3125 rdev2->data_offset,
3126 rdev2->sectors)) {
NeilBrownc5d79ad2008-02-06 01:39:54 -08003127 overlap = 1;
3128 break;
3129 }
NeilBrownc5d79ad2008-02-06 01:39:54 -08003130 if (overlap) {
3131 mddev_put(mddev);
3132 break;
3133 }
3134 }
NeilBrown8b1afc32014-09-29 15:33:20 +10003135 rcu_read_unlock();
NeilBrownc5d79ad2008-02-06 01:39:54 -08003136 if (overlap) {
3137 /* Someone else could have slipped in a size
3138 * change here, but doing so is just silly.
Andre Nolldd8ac332009-03-31 14:33:13 +11003139 * We put oldsectors back because we *know* it is
NeilBrownc5d79ad2008-02-06 01:39:54 -08003140 * safe, and trust userspace not to race with
3141 * itself
3142 */
Andre Nolldd8ac332009-03-31 14:33:13 +11003143 rdev->sectors = oldsectors;
NeilBrownc5d79ad2008-02-06 01:39:54 -08003144 return -EBUSY;
3145 }
3146 }
NeilBrown83303b62006-01-06 00:21:06 -08003147 return len;
3148}
3149
3150static struct rdev_sysfs_entry rdev_size =
NeilBrown80ca3a42006-07-10 04:44:18 -07003151__ATTR(size, S_IRUGO|S_IWUSR, rdev_size_show, rdev_size_store);
NeilBrown83303b62006-01-06 00:21:06 -08003152
NeilBrown3cb03002011-10-11 16:45:26 +11003153static ssize_t recovery_start_show(struct md_rdev *rdev, char *page)
Dan Williams06e3c812009-12-12 21:17:12 -07003154{
3155 unsigned long long recovery_start = rdev->recovery_offset;
3156
3157 if (test_bit(In_sync, &rdev->flags) ||
3158 recovery_start == MaxSector)
3159 return sprintf(page, "none\n");
3160
3161 return sprintf(page, "%llu\n", recovery_start);
3162}
3163
NeilBrown3cb03002011-10-11 16:45:26 +11003164static ssize_t recovery_start_store(struct md_rdev *rdev, const char *buf, size_t len)
Dan Williams06e3c812009-12-12 21:17:12 -07003165{
3166 unsigned long long recovery_start;
3167
3168 if (cmd_match(buf, "none"))
3169 recovery_start = MaxSector;
Jingoo Hanb29bebd2013-06-01 16:15:16 +09003170 else if (kstrtoull(buf, 10, &recovery_start))
Dan Williams06e3c812009-12-12 21:17:12 -07003171 return -EINVAL;
3172
3173 if (rdev->mddev->pers &&
3174 rdev->raid_disk >= 0)
3175 return -EBUSY;
3176
3177 rdev->recovery_offset = recovery_start;
3178 if (recovery_start == MaxSector)
3179 set_bit(In_sync, &rdev->flags);
3180 else
3181 clear_bit(In_sync, &rdev->flags);
3182 return len;
3183}
3184
3185static struct rdev_sysfs_entry rdev_recovery_start =
3186__ATTR(recovery_start, S_IRUGO|S_IWUSR, recovery_start_show, recovery_start_store);
3187
Vishal Vermafc974ee2015-12-24 19:20:34 -07003188/* sysfs access to bad-blocks list.
3189 * We present two files.
3190 * 'bad-blocks' lists sector numbers and lengths of ranges that
3191 * are recorded as bad. The list is truncated to fit within
3192 * the one-page limit of sysfs.
3193 * Writing "sector length" to this file adds an acknowledged
3194 * bad block list.
3195 * 'unacknowledged-bad-blocks' lists bad blocks that have not yet
3196 * been acknowledged. Writing to this file adds bad blocks
3197 * without acknowledging them. This is largely for testing.
3198 */
NeilBrown3cb03002011-10-11 16:45:26 +11003199static ssize_t bb_show(struct md_rdev *rdev, char *page)
NeilBrown16c791a2011-07-28 11:31:47 +10003200{
3201 return badblocks_show(&rdev->badblocks, page, 0);
3202}
NeilBrown3cb03002011-10-11 16:45:26 +11003203static ssize_t bb_store(struct md_rdev *rdev, const char *page, size_t len)
NeilBrown16c791a2011-07-28 11:31:47 +10003204{
NeilBrownde393cd2011-07-28 11:31:48 +10003205 int rv = badblocks_store(&rdev->badblocks, page, len, 0);
3206 /* Maybe that ack was all we needed */
3207 if (test_and_clear_bit(BlockedBadBlocks, &rdev->flags))
3208 wake_up(&rdev->blocked_wait);
3209 return rv;
NeilBrown16c791a2011-07-28 11:31:47 +10003210}
3211static struct rdev_sysfs_entry rdev_bad_blocks =
3212__ATTR(bad_blocks, S_IRUGO|S_IWUSR, bb_show, bb_store);
3213
NeilBrown3cb03002011-10-11 16:45:26 +11003214static ssize_t ubb_show(struct md_rdev *rdev, char *page)
NeilBrown16c791a2011-07-28 11:31:47 +10003215{
3216 return badblocks_show(&rdev->badblocks, page, 1);
3217}
NeilBrown3cb03002011-10-11 16:45:26 +11003218static ssize_t ubb_store(struct md_rdev *rdev, const char *page, size_t len)
NeilBrown16c791a2011-07-28 11:31:47 +10003219{
3220 return badblocks_store(&rdev->badblocks, page, len, 1);
3221}
3222static struct rdev_sysfs_entry rdev_unack_bad_blocks =
3223__ATTR(unacknowledged_bad_blocks, S_IRUGO|S_IWUSR, ubb_show, ubb_store);
3224
Artur Paszkiewicz664aed02017-03-09 10:00:00 +01003225static ssize_t
3226ppl_sector_show(struct md_rdev *rdev, char *page)
3227{
3228 return sprintf(page, "%llu\n", (unsigned long long)rdev->ppl.sector);
3229}
3230
3231static ssize_t
3232ppl_sector_store(struct md_rdev *rdev, const char *buf, size_t len)
3233{
3234 unsigned long long sector;
3235
3236 if (kstrtoull(buf, 10, &sector) < 0)
3237 return -EINVAL;
3238 if (sector != (sector_t)sector)
3239 return -EINVAL;
3240
3241 if (rdev->mddev->pers && test_bit(MD_HAS_PPL, &rdev->mddev->flags) &&
3242 rdev->raid_disk >= 0)
3243 return -EBUSY;
3244
3245 if (rdev->mddev->persistent) {
3246 if (rdev->mddev->major_version == 0)
3247 return -EINVAL;
3248 if ((sector > rdev->sb_start &&
3249 sector - rdev->sb_start > S16_MAX) ||
3250 (sector < rdev->sb_start &&
3251 rdev->sb_start - sector > -S16_MIN))
3252 return -EINVAL;
3253 rdev->ppl.offset = sector - rdev->sb_start;
3254 } else if (!rdev->mddev->external) {
3255 return -EBUSY;
3256 }
3257 rdev->ppl.sector = sector;
3258 return len;
3259}
3260
3261static struct rdev_sysfs_entry rdev_ppl_sector =
3262__ATTR(ppl_sector, S_IRUGO|S_IWUSR, ppl_sector_show, ppl_sector_store);
3263
3264static ssize_t
3265ppl_size_show(struct md_rdev *rdev, char *page)
3266{
3267 return sprintf(page, "%u\n", rdev->ppl.size);
3268}
3269
3270static ssize_t
3271ppl_size_store(struct md_rdev *rdev, const char *buf, size_t len)
3272{
3273 unsigned int size;
3274
3275 if (kstrtouint(buf, 10, &size) < 0)
3276 return -EINVAL;
3277
3278 if (rdev->mddev->pers && test_bit(MD_HAS_PPL, &rdev->mddev->flags) &&
3279 rdev->raid_disk >= 0)
3280 return -EBUSY;
3281
3282 if (rdev->mddev->persistent) {
3283 if (rdev->mddev->major_version == 0)
3284 return -EINVAL;
3285 if (size > U16_MAX)
3286 return -EINVAL;
3287 } else if (!rdev->mddev->external) {
3288 return -EBUSY;
3289 }
3290 rdev->ppl.size = size;
3291 return len;
3292}
3293
3294static struct rdev_sysfs_entry rdev_ppl_size =
3295__ATTR(ppl_size, S_IRUGO|S_IWUSR, ppl_size_show, ppl_size_store);
3296
NeilBrown86e6ffd2005-11-08 21:39:24 -08003297static struct attribute *rdev_default_attrs[] = {
3298 &rdev_state.attr,
NeilBrown4dbcdc72006-01-06 00:20:52 -08003299 &rdev_errors.attr,
NeilBrown014236d2006-01-06 00:20:55 -08003300 &rdev_slot.attr,
NeilBrown93c8cad2006-01-06 00:20:56 -08003301 &rdev_offset.attr,
NeilBrownc6563a82012-05-21 09:27:00 +10003302 &rdev_new_offset.attr,
NeilBrown83303b62006-01-06 00:21:06 -08003303 &rdev_size.attr,
Dan Williams06e3c812009-12-12 21:17:12 -07003304 &rdev_recovery_start.attr,
NeilBrown16c791a2011-07-28 11:31:47 +10003305 &rdev_bad_blocks.attr,
3306 &rdev_unack_bad_blocks.attr,
Artur Paszkiewicz664aed02017-03-09 10:00:00 +01003307 &rdev_ppl_sector.attr,
3308 &rdev_ppl_size.attr,
NeilBrown86e6ffd2005-11-08 21:39:24 -08003309 NULL,
3310};
3311static ssize_t
3312rdev_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3313{
3314 struct rdev_sysfs_entry *entry = container_of(attr, struct rdev_sysfs_entry, attr);
NeilBrown3cb03002011-10-11 16:45:26 +11003315 struct md_rdev *rdev = container_of(kobj, struct md_rdev, kobj);
NeilBrown86e6ffd2005-11-08 21:39:24 -08003316
3317 if (!entry->show)
3318 return -EIO;
NeilBrown758bfc82014-12-15 12:56:59 +11003319 if (!rdev->mddev)
3320 return -EBUSY;
3321 return entry->show(rdev, page);
NeilBrown86e6ffd2005-11-08 21:39:24 -08003322}
3323
3324static ssize_t
3325rdev_attr_store(struct kobject *kobj, struct attribute *attr,
3326 const char *page, size_t length)
3327{
3328 struct rdev_sysfs_entry *entry = container_of(attr, struct rdev_sysfs_entry, attr);
NeilBrown3cb03002011-10-11 16:45:26 +11003329 struct md_rdev *rdev = container_of(kobj, struct md_rdev, kobj);
NeilBrown27c529b2008-03-04 14:29:33 -08003330 ssize_t rv;
NeilBrownfd01b882011-10-11 16:47:53 +11003331 struct mddev *mddev = rdev->mddev;
NeilBrown86e6ffd2005-11-08 21:39:24 -08003332
3333 if (!entry->store)
3334 return -EIO;
NeilBrown67463ac2006-07-10 04:44:19 -07003335 if (!capable(CAP_SYS_ADMIN))
3336 return -EACCES;
NeilBrown27c529b2008-03-04 14:29:33 -08003337 rv = mddev ? mddev_lock(mddev): -EBUSY;
NeilBrownca388052008-02-06 01:39:55 -08003338 if (!rv) {
NeilBrown27c529b2008-03-04 14:29:33 -08003339 if (rdev->mddev == NULL)
3340 rv = -EBUSY;
3341 else
3342 rv = entry->store(rdev, page, length);
Dan Williams6a518302008-04-30 00:52:28 -07003343 mddev_unlock(mddev);
NeilBrownca388052008-02-06 01:39:55 -08003344 }
3345 return rv;
NeilBrown86e6ffd2005-11-08 21:39:24 -08003346}
3347
3348static void rdev_free(struct kobject *ko)
3349{
NeilBrown3cb03002011-10-11 16:45:26 +11003350 struct md_rdev *rdev = container_of(ko, struct md_rdev, kobj);
NeilBrown86e6ffd2005-11-08 21:39:24 -08003351 kfree(rdev);
3352}
Emese Revfy52cf25d2010-01-19 02:58:23 +01003353static const struct sysfs_ops rdev_sysfs_ops = {
NeilBrown86e6ffd2005-11-08 21:39:24 -08003354 .show = rdev_attr_show,
3355 .store = rdev_attr_store,
3356};
3357static struct kobj_type rdev_ktype = {
3358 .release = rdev_free,
3359 .sysfs_ops = &rdev_sysfs_ops,
3360 .default_attrs = rdev_default_attrs,
3361};
3362
NeilBrown3cb03002011-10-11 16:45:26 +11003363int md_rdev_init(struct md_rdev *rdev)
NeilBrowne8bb9a82010-06-01 19:37:26 +10003364{
3365 rdev->desc_nr = -1;
3366 rdev->saved_raid_disk = -1;
3367 rdev->raid_disk = -1;
3368 rdev->flags = 0;
3369 rdev->data_offset = 0;
NeilBrownc6563a82012-05-21 09:27:00 +10003370 rdev->new_data_offset = 0;
NeilBrowne8bb9a82010-06-01 19:37:26 +10003371 rdev->sb_events = 0;
Arnd Bergmann0e3ef492016-06-17 17:33:10 +02003372 rdev->last_read_error = 0;
NeilBrown2699b672011-07-28 11:31:47 +10003373 rdev->sb_loaded = 0;
3374 rdev->bb_page = NULL;
NeilBrowne8bb9a82010-06-01 19:37:26 +10003375 atomic_set(&rdev->nr_pending, 0);
3376 atomic_set(&rdev->read_errors, 0);
3377 atomic_set(&rdev->corrected_errors, 0);
3378
3379 INIT_LIST_HEAD(&rdev->same_set);
3380 init_waitqueue_head(&rdev->blocked_wait);
NeilBrown2230dfe2011-07-28 11:31:46 +10003381
3382 /* Add space to store bad block list.
3383 * This reserves the space even on arrays where it cannot
3384 * be used - I wonder if that matters
3385 */
Vishal Vermafc974ee2015-12-24 19:20:34 -07003386 return badblocks_init(&rdev->badblocks, 0);
NeilBrowne8bb9a82010-06-01 19:37:26 +10003387}
3388EXPORT_SYMBOL_GPL(md_rdev_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389/*
3390 * Import a device. If 'super_format' >= 0, then sanity check the superblock
3391 *
3392 * mark the device faulty if:
3393 *
3394 * - the device is nonexistent (zero size)
3395 * - the device has no valid superblock
3396 *
3397 * a faulty rdev _never_ has rdev->sb set.
3398 */
NeilBrown3cb03002011-10-11 16:45:26 +11003399static struct md_rdev *md_import_device(dev_t newdev, int super_format, int super_minor)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400{
3401 char b[BDEVNAME_SIZE];
3402 int err;
NeilBrown3cb03002011-10-11 16:45:26 +11003403 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404 sector_t size;
3405
NeilBrown9ffae0c2006-01-06 00:20:32 -08003406 rdev = kzalloc(sizeof(*rdev), GFP_KERNEL);
NeilBrown9d487392016-11-02 14:16:49 +11003407 if (!rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003408 return ERR_PTR(-ENOMEM);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409
NeilBrown2230dfe2011-07-28 11:31:46 +10003410 err = md_rdev_init(rdev);
3411 if (err)
3412 goto abort_free;
3413 err = alloc_disk_sb(rdev);
3414 if (err)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415 goto abort_free;
3416
NeilBrownc5d79ad2008-02-06 01:39:54 -08003417 err = lock_rdev(rdev, newdev, super_format == -2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418 if (err)
3419 goto abort_free;
3420
Greg Kroah-Hartmanf9cb0742007-12-17 23:05:35 -07003421 kobject_init(&rdev->kobj, &rdev_ktype);
NeilBrown86e6ffd2005-11-08 21:39:24 -08003422
Mike Snitzer77304d22010-11-08 14:39:12 +01003423 size = i_size_read(rdev->bdev->bd_inode) >> BLOCK_SIZE_BITS;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424 if (!size) {
NeilBrown9d487392016-11-02 14:16:49 +11003425 pr_warn("md: %s has zero or unknown size, marking faulty!\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426 bdevname(rdev->bdev,b));
3427 err = -EINVAL;
3428 goto abort_free;
3429 }
3430
3431 if (super_format >= 0) {
3432 err = super_types[super_format].
3433 load_super(rdev, NULL, super_minor);
3434 if (err == -EINVAL) {
NeilBrown9d487392016-11-02 14:16:49 +11003435 pr_warn("md: %s does not have a valid v%d.%d superblock, not importing!\n",
NeilBrowndf968c42007-07-17 04:06:11 -07003436 bdevname(rdev->bdev,b),
NeilBrown9d487392016-11-02 14:16:49 +11003437 super_format, super_minor);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003438 goto abort_free;
3439 }
3440 if (err < 0) {
NeilBrown9d487392016-11-02 14:16:49 +11003441 pr_warn("md: could not read %s's sb, not importing!\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442 bdevname(rdev->bdev,b));
3443 goto abort_free;
3444 }
3445 }
Dan Williams6bfe0b42008-04-30 00:52:32 -07003446
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447 return rdev;
3448
3449abort_free:
NeilBrown2699b672011-07-28 11:31:47 +10003450 if (rdev->bdev)
3451 unlock_rdev(rdev);
NeilBrown545c8792012-05-22 13:54:30 +10003452 md_rdev_clear(rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453 kfree(rdev);
3454 return ERR_PTR(err);
3455}
3456
3457/*
3458 * Check a full RAID array for plausibility
3459 */
3460
NeilBrownf72ffdd2014-09-30 14:23:59 +10003461static void analyze_sbs(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462{
3463 int i;
NeilBrown3cb03002011-10-11 16:45:26 +11003464 struct md_rdev *rdev, *freshest, *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465 char b[BDEVNAME_SIZE];
3466
3467 freshest = NULL;
NeilBrowndafb20f2012-03-19 12:46:39 +11003468 rdev_for_each_safe(rdev, tmp, mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 switch (super_types[mddev->major_version].
3470 load_super(rdev, freshest, mddev->minor_version)) {
3471 case 1:
3472 freshest = rdev;
3473 break;
3474 case 0:
3475 break;
3476 default:
NeilBrown9d487392016-11-02 14:16:49 +11003477 pr_warn("md: fatal superblock inconsistency in %s -- removing from array\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478 bdevname(rdev->bdev,b));
Goldwyn Rodriguesfb56dfe2015-04-14 10:43:24 -05003479 md_kick_rdev_from_array(rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480 }
3481
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482 super_types[mddev->major_version].
3483 validate_super(mddev, freshest);
3484
3485 i = 0;
NeilBrowndafb20f2012-03-19 12:46:39 +11003486 rdev_for_each_safe(rdev, tmp, mddev) {
NeilBrown233fca32010-04-14 17:02:09 +10003487 if (mddev->max_disks &&
3488 (rdev->desc_nr >= mddev->max_disks ||
3489 i > mddev->max_disks)) {
NeilBrown9d487392016-11-02 14:16:49 +11003490 pr_warn("md: %s: %s: only %d devices permitted\n",
3491 mdname(mddev), bdevname(rdev->bdev, b),
3492 mddev->max_disks);
Goldwyn Rodriguesfb56dfe2015-04-14 10:43:24 -05003493 md_kick_rdev_from_array(rdev);
NeilBrownde01dfa2009-02-06 18:02:46 +11003494 continue;
3495 }
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05003496 if (rdev != freshest) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497 if (super_types[mddev->major_version].
3498 validate_super(mddev, rdev)) {
NeilBrown9d487392016-11-02 14:16:49 +11003499 pr_warn("md: kicking non-fresh %s from array!\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500 bdevname(rdev->bdev,b));
Goldwyn Rodriguesfb56dfe2015-04-14 10:43:24 -05003501 md_kick_rdev_from_array(rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502 continue;
3503 }
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05003504 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 if (mddev->level == LEVEL_MULTIPATH) {
3506 rdev->desc_nr = i++;
3507 rdev->raid_disk = rdev->desc_nr;
NeilBrownb2d444d2005-11-08 21:39:31 -08003508 set_bit(In_sync, &rdev->flags);
Shaohua Lif2076e72015-10-08 21:54:12 -07003509 } else if (rdev->raid_disk >=
3510 (mddev->raid_disks - min(0, mddev->delta_disks)) &&
3511 !test_bit(Journal, &rdev->flags)) {
NeilBrowna778b732007-05-23 13:58:10 -07003512 rdev->raid_disk = -1;
3513 clear_bit(In_sync, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 }
3515 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003516}
3517
NeilBrown72e02072009-12-14 12:49:55 +11003518/* Read a fixed-point number.
3519 * Numbers in sysfs attributes should be in "standard" units where
3520 * possible, so time should be in seconds.
NeilBrownf72ffdd2014-09-30 14:23:59 +10003521 * However we internally use a a much smaller unit such as
NeilBrown72e02072009-12-14 12:49:55 +11003522 * milliseconds or jiffies.
3523 * This function takes a decimal number with a possible fractional
3524 * component, and produces an integer which is the result of
3525 * multiplying that number by 10^'scale'.
3526 * all without any floating-point arithmetic.
3527 */
3528int strict_strtoul_scaled(const char *cp, unsigned long *res, int scale)
3529{
3530 unsigned long result = 0;
3531 long decimals = -1;
3532 while (isdigit(*cp) || (*cp == '.' && decimals < 0)) {
3533 if (*cp == '.')
3534 decimals = 0;
3535 else if (decimals < scale) {
3536 unsigned int value;
3537 value = *cp - '0';
3538 result = result * 10 + value;
3539 if (decimals >= 0)
3540 decimals++;
3541 }
3542 cp++;
3543 }
3544 if (*cp == '\n')
3545 cp++;
3546 if (*cp)
3547 return -EINVAL;
3548 if (decimals < 0)
3549 decimals = 0;
3550 while (decimals < scale) {
3551 result *= 10;
3552 decimals ++;
3553 }
3554 *res = result;
3555 return 0;
3556}
3557
NeilBrowneae17012005-11-08 21:39:23 -08003558static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003559safe_delay_show(struct mddev *mddev, char *page)
NeilBrown16f17b32006-06-26 00:27:37 -07003560{
3561 int msec = (mddev->safemode_delay*1000)/HZ;
3562 return sprintf(page, "%d.%03d\n", msec/1000, msec%1000);
3563}
3564static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003565safe_delay_store(struct mddev *mddev, const char *cbuf, size_t len)
NeilBrown16f17b32006-06-26 00:27:37 -07003566{
NeilBrown16f17b32006-06-26 00:27:37 -07003567 unsigned long msec;
Dan Williams97ce0a72008-09-24 22:48:19 -07003568
Goldwyn Rodrigues28c1b9f2015-10-22 16:01:25 +11003569 if (mddev_is_clustered(mddev)) {
NeilBrown9d487392016-11-02 14:16:49 +11003570 pr_warn("md: Safemode is disabled for clustered mode\n");
Goldwyn Rodrigues28c1b9f2015-10-22 16:01:25 +11003571 return -EINVAL;
3572 }
3573
NeilBrown72e02072009-12-14 12:49:55 +11003574 if (strict_strtoul_scaled(cbuf, &msec, 3) < 0)
NeilBrown16f17b32006-06-26 00:27:37 -07003575 return -EINVAL;
NeilBrown16f17b32006-06-26 00:27:37 -07003576 if (msec == 0)
3577 mddev->safemode_delay = 0;
3578 else {
NeilBrown19052c02008-08-05 15:54:13 +10003579 unsigned long old_delay = mddev->safemode_delay;
NeilBrown1b30e662014-12-15 12:57:00 +11003580 unsigned long new_delay = (msec*HZ)/1000;
3581
3582 if (new_delay == 0)
3583 new_delay = 1;
3584 mddev->safemode_delay = new_delay;
3585 if (new_delay < old_delay || old_delay == 0)
3586 mod_timer(&mddev->safemode_timer, jiffies+1);
NeilBrown16f17b32006-06-26 00:27:37 -07003587 }
3588 return len;
3589}
3590static struct md_sysfs_entry md_safe_delay =
NeilBrown80ca3a42006-07-10 04:44:18 -07003591__ATTR(safe_mode_delay, S_IRUGO|S_IWUSR,safe_delay_show, safe_delay_store);
NeilBrown16f17b32006-06-26 00:27:37 -07003592
3593static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003594level_show(struct mddev *mddev, char *page)
NeilBrowneae17012005-11-08 21:39:23 -08003595{
NeilBrown36d091f2014-12-15 12:56:58 +11003596 struct md_personality *p;
3597 int ret;
3598 spin_lock(&mddev->lock);
3599 p = mddev->pers;
NeilBrownd9d166c2006-01-06 00:20:51 -08003600 if (p)
NeilBrown36d091f2014-12-15 12:56:58 +11003601 ret = sprintf(page, "%s\n", p->name);
NeilBrownd9d166c2006-01-06 00:20:51 -08003602 else if (mddev->clevel[0])
NeilBrown36d091f2014-12-15 12:56:58 +11003603 ret = sprintf(page, "%s\n", mddev->clevel);
NeilBrownd9d166c2006-01-06 00:20:51 -08003604 else if (mddev->level != LEVEL_NONE)
NeilBrown36d091f2014-12-15 12:56:58 +11003605 ret = sprintf(page, "%d\n", mddev->level);
NeilBrownd9d166c2006-01-06 00:20:51 -08003606 else
NeilBrown36d091f2014-12-15 12:56:58 +11003607 ret = 0;
3608 spin_unlock(&mddev->lock);
3609 return ret;
NeilBrowneae17012005-11-08 21:39:23 -08003610}
3611
NeilBrownd9d166c2006-01-06 00:20:51 -08003612static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003613level_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrownd9d166c2006-01-06 00:20:51 -08003614{
Dan Williamsf2859af2010-05-02 10:04:16 -07003615 char clevel[16];
NeilBrown67918752014-12-15 12:57:01 +11003616 ssize_t rv;
3617 size_t slen = len;
NeilBrowndb721d32014-12-15 12:56:58 +11003618 struct md_personality *pers, *oldpers;
Dan Williamsf2859af2010-05-02 10:04:16 -07003619 long level;
NeilBrowndb721d32014-12-15 12:56:58 +11003620 void *priv, *oldpriv;
NeilBrown3cb03002011-10-11 16:45:26 +11003621 struct md_rdev *rdev;
NeilBrown245f46c2009-03-31 14:39:39 +11003622
NeilBrown67918752014-12-15 12:57:01 +11003623 if (slen == 0 || slen >= sizeof(clevel))
3624 return -EINVAL;
3625
3626 rv = mddev_lock(mddev);
3627 if (rv)
NeilBrown245f46c2009-03-31 14:39:39 +11003628 return rv;
NeilBrown67918752014-12-15 12:57:01 +11003629
3630 if (mddev->pers == NULL) {
3631 strncpy(mddev->clevel, buf, slen);
3632 if (mddev->clevel[slen-1] == '\n')
3633 slen--;
3634 mddev->clevel[slen] = 0;
3635 mddev->level = LEVEL_NONE;
3636 rv = len;
3637 goto out_unlock;
NeilBrown245f46c2009-03-31 14:39:39 +11003638 }
NeilBrown67918752014-12-15 12:57:01 +11003639 rv = -EROFS;
NeilBrownbd8839e2014-05-28 13:39:21 +10003640 if (mddev->ro)
NeilBrown67918752014-12-15 12:57:01 +11003641 goto out_unlock;
NeilBrown245f46c2009-03-31 14:39:39 +11003642
3643 /* request to change the personality. Need to ensure:
3644 * - array is not engaged in resync/recovery/reshape
3645 * - old personality can be suspended
3646 * - new personality will access other array.
3647 */
3648
NeilBrown67918752014-12-15 12:57:01 +11003649 rv = -EBUSY;
NeilBrownbb4f1e92010-08-08 21:18:03 +10003650 if (mddev->sync_thread ||
NeilBrownf851b602014-12-11 10:02:10 +11003651 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
NeilBrownbb4f1e92010-08-08 21:18:03 +10003652 mddev->reshape_position != MaxSector ||
3653 mddev->sysfs_active)
NeilBrown67918752014-12-15 12:57:01 +11003654 goto out_unlock;
NeilBrown245f46c2009-03-31 14:39:39 +11003655
NeilBrown67918752014-12-15 12:57:01 +11003656 rv = -EINVAL;
NeilBrown245f46c2009-03-31 14:39:39 +11003657 if (!mddev->pers->quiesce) {
NeilBrown9d487392016-11-02 14:16:49 +11003658 pr_warn("md: %s: %s does not support online personality change\n",
3659 mdname(mddev), mddev->pers->name);
NeilBrown67918752014-12-15 12:57:01 +11003660 goto out_unlock;
NeilBrown245f46c2009-03-31 14:39:39 +11003661 }
3662
3663 /* Now find the new personality */
NeilBrown67918752014-12-15 12:57:01 +11003664 strncpy(clevel, buf, slen);
3665 if (clevel[slen-1] == '\n')
3666 slen--;
3667 clevel[slen] = 0;
Jingoo Hanb29bebd2013-06-01 16:15:16 +09003668 if (kstrtol(clevel, 10, &level))
Dan Williamsf2859af2010-05-02 10:04:16 -07003669 level = LEVEL_NONE;
NeilBrown245f46c2009-03-31 14:39:39 +11003670
Dan Williamsf2859af2010-05-02 10:04:16 -07003671 if (request_module("md-%s", clevel) != 0)
3672 request_module("md-level-%s", clevel);
NeilBrown245f46c2009-03-31 14:39:39 +11003673 spin_lock(&pers_lock);
Dan Williamsf2859af2010-05-02 10:04:16 -07003674 pers = find_pers(level, clevel);
NeilBrown245f46c2009-03-31 14:39:39 +11003675 if (!pers || !try_module_get(pers->owner)) {
3676 spin_unlock(&pers_lock);
NeilBrown9d487392016-11-02 14:16:49 +11003677 pr_warn("md: personality %s not loaded\n", clevel);
NeilBrown67918752014-12-15 12:57:01 +11003678 rv = -EINVAL;
3679 goto out_unlock;
NeilBrown245f46c2009-03-31 14:39:39 +11003680 }
3681 spin_unlock(&pers_lock);
3682
3683 if (pers == mddev->pers) {
3684 /* Nothing to do! */
3685 module_put(pers->owner);
NeilBrown67918752014-12-15 12:57:01 +11003686 rv = len;
3687 goto out_unlock;
NeilBrown245f46c2009-03-31 14:39:39 +11003688 }
3689 if (!pers->takeover) {
3690 module_put(pers->owner);
NeilBrown9d487392016-11-02 14:16:49 +11003691 pr_warn("md: %s: %s does not support personality takeover\n",
3692 mdname(mddev), clevel);
NeilBrown67918752014-12-15 12:57:01 +11003693 rv = -EINVAL;
3694 goto out_unlock;
NeilBrown245f46c2009-03-31 14:39:39 +11003695 }
3696
NeilBrowndafb20f2012-03-19 12:46:39 +11003697 rdev_for_each(rdev, mddev)
NeilBrowne93f68a2010-06-15 09:36:03 +01003698 rdev->new_raid_disk = rdev->raid_disk;
3699
NeilBrown245f46c2009-03-31 14:39:39 +11003700 /* ->takeover must set new_* and/or delta_disks
3701 * if it succeeds, and may set them when it fails.
3702 */
3703 priv = pers->takeover(mddev);
3704 if (IS_ERR(priv)) {
3705 mddev->new_level = mddev->level;
3706 mddev->new_layout = mddev->layout;
Andre Noll664e7c42009-06-18 08:45:27 +10003707 mddev->new_chunk_sectors = mddev->chunk_sectors;
NeilBrown245f46c2009-03-31 14:39:39 +11003708 mddev->raid_disks -= mddev->delta_disks;
3709 mddev->delta_disks = 0;
NeilBrown2c810cd2012-05-21 09:27:00 +10003710 mddev->reshape_backwards = 0;
NeilBrown245f46c2009-03-31 14:39:39 +11003711 module_put(pers->owner);
NeilBrown9d487392016-11-02 14:16:49 +11003712 pr_warn("md: %s: %s would not accept array\n",
3713 mdname(mddev), clevel);
NeilBrown67918752014-12-15 12:57:01 +11003714 rv = PTR_ERR(priv);
3715 goto out_unlock;
NeilBrown245f46c2009-03-31 14:39:39 +11003716 }
3717
3718 /* Looks like we have a winner */
3719 mddev_suspend(mddev);
NeilBrown5aa61f42014-12-15 12:56:57 +11003720 mddev_detach(mddev);
NeilBrown36d091f2014-12-15 12:56:58 +11003721
3722 spin_lock(&mddev->lock);
NeilBrowndb721d32014-12-15 12:56:58 +11003723 oldpers = mddev->pers;
3724 oldpriv = mddev->private;
3725 mddev->pers = pers;
3726 mddev->private = priv;
3727 strlcpy(mddev->clevel, pers->name, sizeof(mddev->clevel));
3728 mddev->level = mddev->new_level;
3729 mddev->layout = mddev->new_layout;
3730 mddev->chunk_sectors = mddev->new_chunk_sectors;
3731 mddev->delta_disks = 0;
3732 mddev->reshape_backwards = 0;
3733 mddev->degraded = 0;
NeilBrown36d091f2014-12-15 12:56:58 +11003734 spin_unlock(&mddev->lock);
NeilBrownf72ffdd2014-09-30 14:23:59 +10003735
NeilBrowndb721d32014-12-15 12:56:58 +11003736 if (oldpers->sync_request == NULL &&
Trela Maciej54071b32010-03-08 16:02:42 +11003737 mddev->external) {
3738 /* We are converting from a no-redundancy array
3739 * to a redundancy array and metadata is managed
3740 * externally so we need to be sure that writes
3741 * won't block due to a need to transition
3742 * clean->dirty
3743 * until external management is started.
3744 */
3745 mddev->in_sync = 0;
3746 mddev->safemode_delay = 0;
3747 mddev->safemode = 0;
3748 }
3749
NeilBrowndb721d32014-12-15 12:56:58 +11003750 oldpers->free(mddev, oldpriv);
3751
3752 if (oldpers->sync_request == NULL &&
3753 pers->sync_request != NULL) {
3754 /* need to add the md_redundancy_group */
3755 if (sysfs_create_group(&mddev->kobj, &md_redundancy_group))
NeilBrown9d487392016-11-02 14:16:49 +11003756 pr_warn("md: cannot register extra attributes for %s\n",
3757 mdname(mddev));
NeilBrowndb721d32014-12-15 12:56:58 +11003758 mddev->sysfs_action = sysfs_get_dirent(mddev->kobj.sd, "sync_action");
3759 }
3760 if (oldpers->sync_request != NULL &&
3761 pers->sync_request == NULL) {
3762 /* need to remove the md_redundancy_group */
3763 if (mddev->to_remove == NULL)
3764 mddev->to_remove = &md_redundancy_group;
3765 }
3766
Alexey Obitotskiy4cb9da72016-06-23 12:11:01 +02003767 module_put(oldpers->owner);
3768
NeilBrowndafb20f2012-03-19 12:46:39 +11003769 rdev_for_each(rdev, mddev) {
NeilBrowne93f68a2010-06-15 09:36:03 +01003770 if (rdev->raid_disk < 0)
3771 continue;
NeilBrownbf2cb0d2011-01-14 09:14:34 +11003772 if (rdev->new_raid_disk >= mddev->raid_disks)
NeilBrowne93f68a2010-06-15 09:36:03 +01003773 rdev->new_raid_disk = -1;
3774 if (rdev->new_raid_disk == rdev->raid_disk)
3775 continue;
Namhyung Kim36fad852011-07-27 11:00:36 +10003776 sysfs_unlink_rdev(mddev, rdev);
NeilBrowne93f68a2010-06-15 09:36:03 +01003777 }
NeilBrowndafb20f2012-03-19 12:46:39 +11003778 rdev_for_each(rdev, mddev) {
NeilBrowne93f68a2010-06-15 09:36:03 +01003779 if (rdev->raid_disk < 0)
3780 continue;
3781 if (rdev->new_raid_disk == rdev->raid_disk)
3782 continue;
3783 rdev->raid_disk = rdev->new_raid_disk;
3784 if (rdev->raid_disk < 0)
NeilBrown3a981b02009-08-03 10:59:55 +10003785 clear_bit(In_sync, &rdev->flags);
NeilBrowne93f68a2010-06-15 09:36:03 +01003786 else {
Namhyung Kim36fad852011-07-27 11:00:36 +10003787 if (sysfs_link_rdev(mddev, rdev))
NeilBrown9d487392016-11-02 14:16:49 +11003788 pr_warn("md: cannot register rd%d for %s after level change\n",
3789 rdev->raid_disk, mdname(mddev));
NeilBrown3a981b02009-08-03 10:59:55 +10003790 }
NeilBrowne93f68a2010-06-15 09:36:03 +01003791 }
3792
NeilBrowndb721d32014-12-15 12:56:58 +11003793 if (pers->sync_request == NULL) {
Trela, Maciej9af204c2010-03-08 16:02:44 +11003794 /* this is now an array without redundancy, so
3795 * it must always be in_sync
3796 */
3797 mddev->in_sync = 1;
3798 del_timer_sync(&mddev->safemode_timer);
3799 }
NeilBrown02e5f5c2013-11-14 15:16:15 +11003800 blk_set_stacking_limits(&mddev->queue->limits);
NeilBrown245f46c2009-03-31 14:39:39 +11003801 pers->run(mddev);
Shaohua Li29530792016-12-08 15:48:19 -08003802 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
Jonathan Brassow47525e52012-05-22 13:55:29 +10003803 mddev_resume(mddev);
NeilBrown830778a2014-01-14 15:17:03 +11003804 if (!mddev->thread)
3805 md_update_sb(mddev, 1);
Maciej Trela5cac78612010-04-14 17:17:39 +10003806 sysfs_notify(&mddev->kobj, NULL, "level");
Dan Williamsbb7f8d22010-05-01 18:14:57 -07003807 md_new_event(mddev);
NeilBrown67918752014-12-15 12:57:01 +11003808 rv = len;
3809out_unlock:
3810 mddev_unlock(mddev);
NeilBrownd9d166c2006-01-06 00:20:51 -08003811 return rv;
3812}
3813
3814static struct md_sysfs_entry md_level =
NeilBrown80ca3a42006-07-10 04:44:18 -07003815__ATTR(level, S_IRUGO|S_IWUSR, level_show, level_store);
NeilBrowneae17012005-11-08 21:39:23 -08003816
NeilBrownd4dbd022006-06-26 00:27:59 -07003817static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003818layout_show(struct mddev *mddev, char *page)
NeilBrownd4dbd022006-06-26 00:27:59 -07003819{
3820 /* just a number, not meaningful for all levels */
NeilBrown08a02ec2007-05-09 02:35:38 -07003821 if (mddev->reshape_position != MaxSector &&
3822 mddev->layout != mddev->new_layout)
3823 return sprintf(page, "%d (%d)\n",
3824 mddev->new_layout, mddev->layout);
NeilBrownd4dbd022006-06-26 00:27:59 -07003825 return sprintf(page, "%d\n", mddev->layout);
3826}
3827
3828static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003829layout_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrownd4dbd022006-06-26 00:27:59 -07003830{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03003831 unsigned int n;
NeilBrown67918752014-12-15 12:57:01 +11003832 int err;
NeilBrownd4dbd022006-06-26 00:27:59 -07003833
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03003834 err = kstrtouint(buf, 10, &n);
3835 if (err < 0)
3836 return err;
NeilBrown67918752014-12-15 12:57:01 +11003837 err = mddev_lock(mddev);
3838 if (err)
3839 return err;
NeilBrownd4dbd022006-06-26 00:27:59 -07003840
NeilBrownb3546032009-03-31 14:56:41 +11003841 if (mddev->pers) {
NeilBrown50ac1682009-06-18 08:47:55 +10003842 if (mddev->pers->check_reshape == NULL)
NeilBrown67918752014-12-15 12:57:01 +11003843 err = -EBUSY;
3844 else if (mddev->ro)
3845 err = -EROFS;
3846 else {
3847 mddev->new_layout = n;
3848 err = mddev->pers->check_reshape(mddev);
3849 if (err)
3850 mddev->new_layout = mddev->layout;
NeilBrown597a7112009-06-18 08:47:42 +10003851 }
NeilBrownb3546032009-03-31 14:56:41 +11003852 } else {
NeilBrown08a02ec2007-05-09 02:35:38 -07003853 mddev->new_layout = n;
NeilBrownb3546032009-03-31 14:56:41 +11003854 if (mddev->reshape_position == MaxSector)
3855 mddev->layout = n;
3856 }
NeilBrown67918752014-12-15 12:57:01 +11003857 mddev_unlock(mddev);
3858 return err ?: len;
NeilBrownd4dbd022006-06-26 00:27:59 -07003859}
3860static struct md_sysfs_entry md_layout =
NeilBrown80ca3a42006-07-10 04:44:18 -07003861__ATTR(layout, S_IRUGO|S_IWUSR, layout_show, layout_store);
NeilBrownd4dbd022006-06-26 00:27:59 -07003862
NeilBrowneae17012005-11-08 21:39:23 -08003863static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003864raid_disks_show(struct mddev *mddev, char *page)
NeilBrowneae17012005-11-08 21:39:23 -08003865{
NeilBrownbb636542005-11-08 21:39:45 -08003866 if (mddev->raid_disks == 0)
3867 return 0;
NeilBrown08a02ec2007-05-09 02:35:38 -07003868 if (mddev->reshape_position != MaxSector &&
3869 mddev->delta_disks != 0)
3870 return sprintf(page, "%d (%d)\n", mddev->raid_disks,
3871 mddev->raid_disks - mddev->delta_disks);
NeilBrowneae17012005-11-08 21:39:23 -08003872 return sprintf(page, "%d\n", mddev->raid_disks);
3873}
3874
NeilBrownfd01b882011-10-11 16:47:53 +11003875static int update_raid_disks(struct mddev *mddev, int raid_disks);
NeilBrownda943b992006-01-06 00:20:54 -08003876
3877static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003878raid_disks_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrownda943b992006-01-06 00:20:54 -08003879{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03003880 unsigned int n;
NeilBrown67918752014-12-15 12:57:01 +11003881 int err;
NeilBrownda943b992006-01-06 00:20:54 -08003882
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03003883 err = kstrtouint(buf, 10, &n);
3884 if (err < 0)
3885 return err;
NeilBrownda943b992006-01-06 00:20:54 -08003886
NeilBrown67918752014-12-15 12:57:01 +11003887 err = mddev_lock(mddev);
3888 if (err)
3889 return err;
NeilBrownda943b992006-01-06 00:20:54 -08003890 if (mddev->pers)
NeilBrown67918752014-12-15 12:57:01 +11003891 err = update_raid_disks(mddev, n);
NeilBrown08a02ec2007-05-09 02:35:38 -07003892 else if (mddev->reshape_position != MaxSector) {
NeilBrownc6563a82012-05-21 09:27:00 +10003893 struct md_rdev *rdev;
NeilBrown08a02ec2007-05-09 02:35:38 -07003894 int olddisks = mddev->raid_disks - mddev->delta_disks;
NeilBrownc6563a82012-05-21 09:27:00 +10003895
NeilBrown67918752014-12-15 12:57:01 +11003896 err = -EINVAL;
NeilBrownc6563a82012-05-21 09:27:00 +10003897 rdev_for_each(rdev, mddev) {
3898 if (olddisks < n &&
3899 rdev->data_offset < rdev->new_data_offset)
NeilBrown67918752014-12-15 12:57:01 +11003900 goto out_unlock;
NeilBrownc6563a82012-05-21 09:27:00 +10003901 if (olddisks > n &&
3902 rdev->data_offset > rdev->new_data_offset)
NeilBrown67918752014-12-15 12:57:01 +11003903 goto out_unlock;
NeilBrownc6563a82012-05-21 09:27:00 +10003904 }
NeilBrown67918752014-12-15 12:57:01 +11003905 err = 0;
NeilBrown08a02ec2007-05-09 02:35:38 -07003906 mddev->delta_disks = n - olddisks;
3907 mddev->raid_disks = n;
NeilBrown2c810cd2012-05-21 09:27:00 +10003908 mddev->reshape_backwards = (mddev->delta_disks < 0);
NeilBrown08a02ec2007-05-09 02:35:38 -07003909 } else
NeilBrownda943b992006-01-06 00:20:54 -08003910 mddev->raid_disks = n;
NeilBrown67918752014-12-15 12:57:01 +11003911out_unlock:
3912 mddev_unlock(mddev);
3913 return err ? err : len;
NeilBrownda943b992006-01-06 00:20:54 -08003914}
3915static struct md_sysfs_entry md_raid_disks =
NeilBrown80ca3a42006-07-10 04:44:18 -07003916__ATTR(raid_disks, S_IRUGO|S_IWUSR, raid_disks_show, raid_disks_store);
NeilBrowneae17012005-11-08 21:39:23 -08003917
NeilBrown24dd4692005-11-08 21:39:26 -08003918static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003919chunk_size_show(struct mddev *mddev, char *page)
NeilBrown3b343802006-01-06 00:20:47 -08003920{
NeilBrown08a02ec2007-05-09 02:35:38 -07003921 if (mddev->reshape_position != MaxSector &&
Andre Noll664e7c42009-06-18 08:45:27 +10003922 mddev->chunk_sectors != mddev->new_chunk_sectors)
3923 return sprintf(page, "%d (%d)\n",
3924 mddev->new_chunk_sectors << 9,
Andre Noll9d8f0362009-06-18 08:45:01 +10003925 mddev->chunk_sectors << 9);
3926 return sprintf(page, "%d\n", mddev->chunk_sectors << 9);
NeilBrown3b343802006-01-06 00:20:47 -08003927}
3928
3929static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003930chunk_size_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrown3b343802006-01-06 00:20:47 -08003931{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03003932 unsigned long n;
NeilBrown67918752014-12-15 12:57:01 +11003933 int err;
NeilBrown3b343802006-01-06 00:20:47 -08003934
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03003935 err = kstrtoul(buf, 10, &n);
3936 if (err < 0)
3937 return err;
NeilBrown3b343802006-01-06 00:20:47 -08003938
NeilBrown67918752014-12-15 12:57:01 +11003939 err = mddev_lock(mddev);
3940 if (err)
3941 return err;
NeilBrownb3546032009-03-31 14:56:41 +11003942 if (mddev->pers) {
NeilBrown50ac1682009-06-18 08:47:55 +10003943 if (mddev->pers->check_reshape == NULL)
NeilBrown67918752014-12-15 12:57:01 +11003944 err = -EBUSY;
3945 else if (mddev->ro)
3946 err = -EROFS;
3947 else {
3948 mddev->new_chunk_sectors = n >> 9;
3949 err = mddev->pers->check_reshape(mddev);
3950 if (err)
3951 mddev->new_chunk_sectors = mddev->chunk_sectors;
NeilBrown597a7112009-06-18 08:47:42 +10003952 }
NeilBrownb3546032009-03-31 14:56:41 +11003953 } else {
Andre Noll664e7c42009-06-18 08:45:27 +10003954 mddev->new_chunk_sectors = n >> 9;
NeilBrownb3546032009-03-31 14:56:41 +11003955 if (mddev->reshape_position == MaxSector)
Andre Noll9d8f0362009-06-18 08:45:01 +10003956 mddev->chunk_sectors = n >> 9;
NeilBrownb3546032009-03-31 14:56:41 +11003957 }
NeilBrown67918752014-12-15 12:57:01 +11003958 mddev_unlock(mddev);
3959 return err ?: len;
NeilBrown3b343802006-01-06 00:20:47 -08003960}
3961static struct md_sysfs_entry md_chunk_size =
NeilBrown80ca3a42006-07-10 04:44:18 -07003962__ATTR(chunk_size, S_IRUGO|S_IWUSR, chunk_size_show, chunk_size_store);
NeilBrown3b343802006-01-06 00:20:47 -08003963
NeilBrowna94213b2006-06-26 00:28:00 -07003964static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003965resync_start_show(struct mddev *mddev, char *page)
NeilBrowna94213b2006-06-26 00:28:00 -07003966{
NeilBrownd1a7c502009-03-31 15:24:32 +11003967 if (mddev->recovery_cp == MaxSector)
3968 return sprintf(page, "none\n");
NeilBrowna94213b2006-06-26 00:28:00 -07003969 return sprintf(page, "%llu\n", (unsigned long long)mddev->recovery_cp);
3970}
3971
3972static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11003973resync_start_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrowna94213b2006-06-26 00:28:00 -07003974{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03003975 unsigned long long n;
NeilBrown67918752014-12-15 12:57:01 +11003976 int err;
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03003977
3978 if (cmd_match(buf, "none"))
3979 n = MaxSector;
3980 else {
3981 err = kstrtoull(buf, 10, &n);
3982 if (err < 0)
3983 return err;
3984 if (n != (sector_t)n)
3985 return -EINVAL;
3986 }
NeilBrowna94213b2006-06-26 00:28:00 -07003987
NeilBrown67918752014-12-15 12:57:01 +11003988 err = mddev_lock(mddev);
3989 if (err)
3990 return err;
NeilBrownb0986362011-05-11 15:52:21 +10003991 if (mddev->pers && !test_bit(MD_RECOVERY_FROZEN, &mddev->recovery))
NeilBrown67918752014-12-15 12:57:01 +11003992 err = -EBUSY;
NeilBrowna94213b2006-06-26 00:28:00 -07003993
NeilBrown67918752014-12-15 12:57:01 +11003994 if (!err) {
3995 mddev->recovery_cp = n;
3996 if (mddev->pers)
Shaohua Li29530792016-12-08 15:48:19 -08003997 set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
NeilBrown67918752014-12-15 12:57:01 +11003998 }
3999 mddev_unlock(mddev);
4000 return err ?: len;
NeilBrowna94213b2006-06-26 00:28:00 -07004001}
4002static struct md_sysfs_entry md_resync_start =
NeilBrown750f1992014-09-30 08:53:05 +10004003__ATTR_PREALLOC(resync_start, S_IRUGO|S_IWUSR,
4004 resync_start_show, resync_start_store);
NeilBrowna94213b2006-06-26 00:28:00 -07004005
NeilBrown9e653b62006-06-26 00:27:58 -07004006/*
4007 * The array state can be:
4008 *
4009 * clear
4010 * No devices, no size, no level
4011 * Equivalent to STOP_ARRAY ioctl
4012 * inactive
4013 * May have some settings, but array is not active
4014 * all IO results in error
4015 * When written, doesn't tear down array, but just stops it
4016 * suspended (not supported yet)
4017 * All IO requests will block. The array can be reconfigured.
Andre Noll910d8cb2008-03-25 21:00:53 +01004018 * Writing this, if accepted, will block until array is quiescent
NeilBrown9e653b62006-06-26 00:27:58 -07004019 * readonly
4020 * no resync can happen. no superblocks get written.
4021 * write requests fail
4022 * read-auto
4023 * like readonly, but behaves like 'clean' on a write request.
4024 *
4025 * clean - no pending writes, but otherwise active.
4026 * When written to inactive array, starts without resync
4027 * If a write request arrives then
4028 * if metadata is known, mark 'dirty' and switch to 'active'.
4029 * if not known, block and switch to write-pending
4030 * If written to an active array that has pending writes, then fails.
4031 * active
4032 * fully active: IO and resync can be happening.
4033 * When written to inactive array, starts with resync
4034 *
4035 * write-pending
4036 * clean, but writes are blocked waiting for 'active' to be written.
4037 *
4038 * active-idle
4039 * like active, but no writes have been seen for a while (100msec).
4040 *
4041 */
4042enum array_state { clear, inactive, suspended, readonly, read_auto, clean, active,
4043 write_pending, active_idle, bad_word};
Adrian Bunk05381952006-06-26 00:28:01 -07004044static char *array_states[] = {
NeilBrown9e653b62006-06-26 00:27:58 -07004045 "clear", "inactive", "suspended", "readonly", "read-auto", "clean", "active",
4046 "write-pending", "active-idle", NULL };
4047
4048static int match_word(const char *word, char **list)
4049{
4050 int n;
4051 for (n=0; list[n]; n++)
4052 if (cmd_match(word, list[n]))
4053 break;
4054 return n;
4055}
4056
4057static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004058array_state_show(struct mddev *mddev, char *page)
NeilBrown9e653b62006-06-26 00:27:58 -07004059{
4060 enum array_state st = inactive;
4061
4062 if (mddev->pers)
4063 switch(mddev->ro) {
4064 case 1:
4065 st = readonly;
4066 break;
4067 case 2:
4068 st = read_auto;
4069 break;
4070 case 0:
NeilBrown55cc39f2017-03-15 14:05:14 +11004071 spin_lock(&mddev->lock);
Shaohua Li29530792016-12-08 15:48:19 -08004072 if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
NeilBrowne6910632008-02-06 01:39:51 -08004073 st = write_pending;
Tomasz Majchrzak16f88942016-10-24 12:47:28 +02004074 else if (mddev->in_sync)
4075 st = clean;
NeilBrown9e653b62006-06-26 00:27:58 -07004076 else if (mddev->safemode)
4077 st = active_idle;
4078 else
4079 st = active;
NeilBrown55cc39f2017-03-15 14:05:14 +11004080 spin_unlock(&mddev->lock);
NeilBrown9e653b62006-06-26 00:27:58 -07004081 }
4082 else {
4083 if (list_empty(&mddev->disks) &&
4084 mddev->raid_disks == 0 &&
Andre Noll58c0fed2009-03-31 14:33:13 +11004085 mddev->dev_sectors == 0)
NeilBrown9e653b62006-06-26 00:27:58 -07004086 st = clear;
4087 else
4088 st = inactive;
4089 }
4090 return sprintf(page, "%s\n", array_states[st]);
4091}
4092
NeilBrownf72ffdd2014-09-30 14:23:59 +10004093static int do_md_stop(struct mddev *mddev, int ro, struct block_device *bdev);
4094static int md_set_readonly(struct mddev *mddev, struct block_device *bdev);
4095static int do_md_run(struct mddev *mddev);
NeilBrownfd01b882011-10-11 16:47:53 +11004096static int restart_array(struct mddev *mddev);
NeilBrown9e653b62006-06-26 00:27:58 -07004097
4098static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004099array_state_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrown9e653b62006-06-26 00:27:58 -07004100{
NeilBrown6497709b2017-03-15 14:05:14 +11004101 int err = 0;
NeilBrown9e653b62006-06-26 00:27:58 -07004102 enum array_state st = match_word(buf, array_states);
NeilBrown67918752014-12-15 12:57:01 +11004103
4104 if (mddev->pers && (st == active || st == clean) && mddev->ro != 1) {
4105 /* don't take reconfig_mutex when toggling between
4106 * clean and active
4107 */
4108 spin_lock(&mddev->lock);
4109 if (st == active) {
4110 restart_array(mddev);
Shaohua Li29530792016-12-08 15:48:19 -08004111 clear_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
Tomasz Majchrzak91a6c4a2016-10-25 17:07:08 +02004112 md_wakeup_thread(mddev->thread);
NeilBrown67918752014-12-15 12:57:01 +11004113 wake_up(&mddev->sb_wait);
NeilBrown67918752014-12-15 12:57:01 +11004114 } else /* st == clean */ {
4115 restart_array(mddev);
NeilBrown6497709b2017-03-15 14:05:14 +11004116 if (!set_in_sync(mddev))
NeilBrown67918752014-12-15 12:57:01 +11004117 err = -EBUSY;
4118 }
Tomasz Majchrzak573275b2016-06-30 10:47:09 +02004119 if (!err)
4120 sysfs_notify_dirent_safe(mddev->sysfs_state);
NeilBrown67918752014-12-15 12:57:01 +11004121 spin_unlock(&mddev->lock);
NeilBrownc008f1d2015-06-12 19:46:44 +10004122 return err ?: len;
NeilBrown67918752014-12-15 12:57:01 +11004123 }
4124 err = mddev_lock(mddev);
4125 if (err)
4126 return err;
4127 err = -EINVAL;
NeilBrown9e653b62006-06-26 00:27:58 -07004128 switch(st) {
4129 case bad_word:
4130 break;
4131 case clear:
4132 /* stopping an active array */
NeilBrowna05b7ea2012-07-19 15:59:18 +10004133 err = do_md_stop(mddev, 0, NULL);
NeilBrown9e653b62006-06-26 00:27:58 -07004134 break;
4135 case inactive:
4136 /* stopping an active array */
NeilBrown90cf1952012-07-31 10:04:55 +10004137 if (mddev->pers)
NeilBrowna05b7ea2012-07-19 15:59:18 +10004138 err = do_md_stop(mddev, 2, NULL);
NeilBrown90cf1952012-07-31 10:04:55 +10004139 else
NeilBrowne6910632008-02-06 01:39:51 -08004140 err = 0; /* already inactive */
NeilBrown9e653b62006-06-26 00:27:58 -07004141 break;
4142 case suspended:
4143 break; /* not supported yet */
4144 case readonly:
4145 if (mddev->pers)
NeilBrowna05b7ea2012-07-19 15:59:18 +10004146 err = md_set_readonly(mddev, NULL);
NeilBrown9e653b62006-06-26 00:27:58 -07004147 else {
4148 mddev->ro = 1;
NeilBrown648b6292008-04-30 00:52:30 -07004149 set_disk_ro(mddev->gendisk, 1);
NeilBrown9e653b62006-06-26 00:27:58 -07004150 err = do_md_run(mddev);
4151 }
4152 break;
4153 case read_auto:
NeilBrown9e653b62006-06-26 00:27:58 -07004154 if (mddev->pers) {
NeilBrown80268ee2008-10-13 11:55:12 +11004155 if (mddev->ro == 0)
NeilBrowna05b7ea2012-07-19 15:59:18 +10004156 err = md_set_readonly(mddev, NULL);
NeilBrown80268ee2008-10-13 11:55:12 +11004157 else if (mddev->ro == 1)
NeilBrown648b6292008-04-30 00:52:30 -07004158 err = restart_array(mddev);
4159 if (err == 0) {
4160 mddev->ro = 2;
4161 set_disk_ro(mddev->gendisk, 0);
4162 }
NeilBrown9e653b62006-06-26 00:27:58 -07004163 } else {
4164 mddev->ro = 2;
4165 err = do_md_run(mddev);
4166 }
4167 break;
4168 case clean:
4169 if (mddev->pers) {
Song Liu339421d2015-10-08 21:54:13 -07004170 err = restart_array(mddev);
4171 if (err)
4172 break;
NeilBrown85572d72014-12-15 12:56:56 +11004173 spin_lock(&mddev->lock);
NeilBrown6497709b2017-03-15 14:05:14 +11004174 if (!set_in_sync(mddev))
NeilBrowne6910632008-02-06 01:39:51 -08004175 err = -EBUSY;
NeilBrown85572d72014-12-15 12:56:56 +11004176 spin_unlock(&mddev->lock);
NeilBrown5bf29592009-05-07 12:50:57 +10004177 } else
4178 err = -EINVAL;
NeilBrown9e653b62006-06-26 00:27:58 -07004179 break;
4180 case active:
4181 if (mddev->pers) {
Song Liu339421d2015-10-08 21:54:13 -07004182 err = restart_array(mddev);
4183 if (err)
4184 break;
Shaohua Li29530792016-12-08 15:48:19 -08004185 clear_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
NeilBrown9e653b62006-06-26 00:27:58 -07004186 wake_up(&mddev->sb_wait);
4187 err = 0;
4188 } else {
4189 mddev->ro = 0;
NeilBrown648b6292008-04-30 00:52:30 -07004190 set_disk_ro(mddev->gendisk, 0);
NeilBrown9e653b62006-06-26 00:27:58 -07004191 err = do_md_run(mddev);
4192 }
4193 break;
4194 case write_pending:
4195 case active_idle:
4196 /* these cannot be set */
4197 break;
4198 }
NeilBrown67918752014-12-15 12:57:01 +11004199
4200 if (!err) {
NeilBrown1d23f172011-12-08 15:49:12 +11004201 if (mddev->hold_active == UNTIL_IOCTL)
4202 mddev->hold_active = 0;
NeilBrown00bcb4a2010-06-01 19:37:23 +10004203 sysfs_notify_dirent_safe(mddev->sysfs_state);
Neil Brown0fd62b82008-06-28 08:31:36 +10004204 }
NeilBrown67918752014-12-15 12:57:01 +11004205 mddev_unlock(mddev);
4206 return err ?: len;
NeilBrown9e653b62006-06-26 00:27:58 -07004207}
NeilBrown80ca3a42006-07-10 04:44:18 -07004208static struct md_sysfs_entry md_array_state =
NeilBrown750f1992014-09-30 08:53:05 +10004209__ATTR_PREALLOC(array_state, S_IRUGO|S_IWUSR, array_state_show, array_state_store);
NeilBrown9e653b62006-06-26 00:27:58 -07004210
NeilBrown6d7ff732006-01-06 00:21:16 -08004211static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004212max_corrected_read_errors_show(struct mddev *mddev, char *page) {
Robert Becker1e509152009-12-14 12:49:58 +11004213 return sprintf(page, "%d\n",
4214 atomic_read(&mddev->max_corr_read_errors));
4215}
4216
4217static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004218max_corrected_read_errors_store(struct mddev *mddev, const char *buf, size_t len)
Robert Becker1e509152009-12-14 12:49:58 +11004219{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004220 unsigned int n;
4221 int rv;
Robert Becker1e509152009-12-14 12:49:58 +11004222
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004223 rv = kstrtouint(buf, 10, &n);
4224 if (rv < 0)
4225 return rv;
4226 atomic_set(&mddev->max_corr_read_errors, n);
4227 return len;
Robert Becker1e509152009-12-14 12:49:58 +11004228}
4229
4230static struct md_sysfs_entry max_corr_read_errors =
4231__ATTR(max_read_errors, S_IRUGO|S_IWUSR, max_corrected_read_errors_show,
4232 max_corrected_read_errors_store);
4233
4234static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004235null_show(struct mddev *mddev, char *page)
NeilBrown6d7ff732006-01-06 00:21:16 -08004236{
4237 return -EINVAL;
4238}
4239
4240static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004241new_dev_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrown6d7ff732006-01-06 00:21:16 -08004242{
4243 /* buf must be %d:%d\n? giving major and minor numbers */
4244 /* The new device is added to the array.
4245 * If the array has a persistent superblock, we read the
4246 * superblock to initialise info and check validity.
4247 * Otherwise, only checking done is that in bind_rdev_to_array,
4248 * which mainly checks size.
4249 */
4250 char *e;
4251 int major = simple_strtoul(buf, &e, 10);
4252 int minor;
4253 dev_t dev;
NeilBrown3cb03002011-10-11 16:45:26 +11004254 struct md_rdev *rdev;
NeilBrown6d7ff732006-01-06 00:21:16 -08004255 int err;
4256
4257 if (!*buf || *e != ':' || !e[1] || e[1] == '\n')
4258 return -EINVAL;
4259 minor = simple_strtoul(e+1, &e, 10);
4260 if (*e && *e != '\n')
4261 return -EINVAL;
4262 dev = MKDEV(major, minor);
4263 if (major != MAJOR(dev) ||
4264 minor != MINOR(dev))
4265 return -EOVERFLOW;
4266
NeilBrown67918752014-12-15 12:57:01 +11004267 flush_workqueue(md_misc_wq);
4268
4269 err = mddev_lock(mddev);
4270 if (err)
4271 return err;
NeilBrown6d7ff732006-01-06 00:21:16 -08004272 if (mddev->persistent) {
4273 rdev = md_import_device(dev, mddev->major_version,
4274 mddev->minor_version);
4275 if (!IS_ERR(rdev) && !list_empty(&mddev->disks)) {
NeilBrown3cb03002011-10-11 16:45:26 +11004276 struct md_rdev *rdev0
4277 = list_entry(mddev->disks.next,
4278 struct md_rdev, same_set);
NeilBrown6d7ff732006-01-06 00:21:16 -08004279 err = super_types[mddev->major_version]
4280 .load_super(rdev, rdev0, mddev->minor_version);
4281 if (err < 0)
4282 goto out;
4283 }
NeilBrownc5d79ad2008-02-06 01:39:54 -08004284 } else if (mddev->external)
4285 rdev = md_import_device(dev, -2, -1);
4286 else
NeilBrown6d7ff732006-01-06 00:21:16 -08004287 rdev = md_import_device(dev, -1, -1);
4288
NeilBrown9a8c0fa2015-06-25 17:06:40 +10004289 if (IS_ERR(rdev)) {
4290 mddev_unlock(mddev);
NeilBrown6d7ff732006-01-06 00:21:16 -08004291 return PTR_ERR(rdev);
NeilBrown9a8c0fa2015-06-25 17:06:40 +10004292 }
NeilBrown6d7ff732006-01-06 00:21:16 -08004293 err = bind_rdev_to_array(rdev, mddev);
4294 out:
4295 if (err)
4296 export_rdev(rdev);
NeilBrown67918752014-12-15 12:57:01 +11004297 mddev_unlock(mddev);
Alexey Obitotskiy5492c462017-07-28 15:49:25 +02004298 if (!err)
4299 md_new_event(mddev);
NeilBrown6d7ff732006-01-06 00:21:16 -08004300 return err ? err : len;
4301}
4302
4303static struct md_sysfs_entry md_new_device =
NeilBrown80ca3a42006-07-10 04:44:18 -07004304__ATTR(new_dev, S_IWUSR, null_show, new_dev_store);
NeilBrown3b343802006-01-06 00:20:47 -08004305
4306static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004307bitmap_store(struct mddev *mddev, const char *buf, size_t len)
Paul Clements9b1d1da2006-10-03 01:15:49 -07004308{
4309 char *end;
4310 unsigned long chunk, end_chunk;
NeilBrown67918752014-12-15 12:57:01 +11004311 int err;
Paul Clements9b1d1da2006-10-03 01:15:49 -07004312
NeilBrown67918752014-12-15 12:57:01 +11004313 err = mddev_lock(mddev);
4314 if (err)
4315 return err;
Paul Clements9b1d1da2006-10-03 01:15:49 -07004316 if (!mddev->bitmap)
4317 goto out;
4318 /* buf should be <chunk> <chunk> ... or <chunk>-<chunk> ... (range) */
4319 while (*buf) {
4320 chunk = end_chunk = simple_strtoul(buf, &end, 0);
4321 if (buf == end) break;
4322 if (*end == '-') { /* range */
4323 buf = end + 1;
4324 end_chunk = simple_strtoul(buf, &end, 0);
4325 if (buf == end) break;
4326 }
4327 if (*end && !isspace(*end)) break;
4328 bitmap_dirty_bits(mddev->bitmap, chunk, end_chunk);
André Goddard Rosae7d28602009-12-14 18:01:06 -08004329 buf = skip_spaces(end);
Paul Clements9b1d1da2006-10-03 01:15:49 -07004330 }
4331 bitmap_unplug(mddev->bitmap); /* flush the bits to disk */
4332out:
NeilBrown67918752014-12-15 12:57:01 +11004333 mddev_unlock(mddev);
Paul Clements9b1d1da2006-10-03 01:15:49 -07004334 return len;
4335}
4336
4337static struct md_sysfs_entry md_bitmap =
4338__ATTR(bitmap_set_bits, S_IWUSR, null_show, bitmap_store);
4339
4340static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004341size_show(struct mddev *mddev, char *page)
NeilBrowna35b0d62006-01-06 00:20:49 -08004342{
Andre Noll58c0fed2009-03-31 14:33:13 +11004343 return sprintf(page, "%llu\n",
4344 (unsigned long long)mddev->dev_sectors / 2);
NeilBrowna35b0d62006-01-06 00:20:49 -08004345}
4346
NeilBrownfd01b882011-10-11 16:47:53 +11004347static int update_size(struct mddev *mddev, sector_t num_sectors);
NeilBrowna35b0d62006-01-06 00:20:49 -08004348
4349static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004350size_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrowna35b0d62006-01-06 00:20:49 -08004351{
4352 /* If array is inactive, we can reduce the component size, but
4353 * not increase it (except from 0).
4354 * If array is active, we can try an on-line resize
4355 */
Dan Williamsb522adc2009-03-31 15:00:31 +11004356 sector_t sectors;
4357 int err = strict_blocks_to_sectors(buf, &sectors);
NeilBrowna35b0d62006-01-06 00:20:49 -08004358
Andre Noll58c0fed2009-03-31 14:33:13 +11004359 if (err < 0)
4360 return err;
NeilBrown67918752014-12-15 12:57:01 +11004361 err = mddev_lock(mddev);
4362 if (err)
4363 return err;
NeilBrowna35b0d62006-01-06 00:20:49 -08004364 if (mddev->pers) {
Andre Noll58c0fed2009-03-31 14:33:13 +11004365 err = update_size(mddev, sectors);
Xiao Ni4ba1e782016-06-12 17:18:00 +08004366 if (err == 0)
4367 md_update_sb(mddev, 1);
NeilBrowna35b0d62006-01-06 00:20:49 -08004368 } else {
Andre Noll58c0fed2009-03-31 14:33:13 +11004369 if (mddev->dev_sectors == 0 ||
4370 mddev->dev_sectors > sectors)
4371 mddev->dev_sectors = sectors;
NeilBrowna35b0d62006-01-06 00:20:49 -08004372 else
4373 err = -ENOSPC;
4374 }
NeilBrown67918752014-12-15 12:57:01 +11004375 mddev_unlock(mddev);
NeilBrowna35b0d62006-01-06 00:20:49 -08004376 return err ? err : len;
4377}
4378
4379static struct md_sysfs_entry md_size =
NeilBrown80ca3a42006-07-10 04:44:18 -07004380__ATTR(component_size, S_IRUGO|S_IWUSR, size_show, size_store);
NeilBrowna35b0d62006-01-06 00:20:49 -08004381
Masanari Iida83f0d772012-10-30 00:18:08 +09004382/* Metadata version.
NeilBrowne6910632008-02-06 01:39:51 -08004383 * This is one of
4384 * 'none' for arrays with no metadata (good luck...)
4385 * 'external' for arrays with externally managed metadata,
NeilBrown8bb93aa2006-01-06 00:20:50 -08004386 * or N.M for internally known formats
4387 */
4388static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004389metadata_show(struct mddev *mddev, char *page)
NeilBrown8bb93aa2006-01-06 00:20:50 -08004390{
4391 if (mddev->persistent)
4392 return sprintf(page, "%d.%d\n",
4393 mddev->major_version, mddev->minor_version);
NeilBrowne6910632008-02-06 01:39:51 -08004394 else if (mddev->external)
4395 return sprintf(page, "external:%s\n", mddev->metadata_type);
NeilBrown8bb93aa2006-01-06 00:20:50 -08004396 else
4397 return sprintf(page, "none\n");
4398}
4399
4400static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004401metadata_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrown8bb93aa2006-01-06 00:20:50 -08004402{
4403 int major, minor;
4404 char *e;
NeilBrown67918752014-12-15 12:57:01 +11004405 int err;
NeilBrownea43ddd2008-10-13 11:55:11 +11004406 /* Changing the details of 'external' metadata is
4407 * always permitted. Otherwise there must be
4408 * no devices attached to the array.
4409 */
NeilBrown67918752014-12-15 12:57:01 +11004410
4411 err = mddev_lock(mddev);
4412 if (err)
4413 return err;
4414 err = -EBUSY;
NeilBrownea43ddd2008-10-13 11:55:11 +11004415 if (mddev->external && strncmp(buf, "external:", 9) == 0)
4416 ;
4417 else if (!list_empty(&mddev->disks))
NeilBrown67918752014-12-15 12:57:01 +11004418 goto out_unlock;
NeilBrown8bb93aa2006-01-06 00:20:50 -08004419
NeilBrown67918752014-12-15 12:57:01 +11004420 err = 0;
NeilBrown8bb93aa2006-01-06 00:20:50 -08004421 if (cmd_match(buf, "none")) {
4422 mddev->persistent = 0;
NeilBrowne6910632008-02-06 01:39:51 -08004423 mddev->external = 0;
4424 mddev->major_version = 0;
4425 mddev->minor_version = 90;
NeilBrown67918752014-12-15 12:57:01 +11004426 goto out_unlock;
NeilBrowne6910632008-02-06 01:39:51 -08004427 }
4428 if (strncmp(buf, "external:", 9) == 0) {
NeilBrown20a49ff2008-02-06 01:39:57 -08004429 size_t namelen = len-9;
NeilBrowne6910632008-02-06 01:39:51 -08004430 if (namelen >= sizeof(mddev->metadata_type))
4431 namelen = sizeof(mddev->metadata_type)-1;
4432 strncpy(mddev->metadata_type, buf+9, namelen);
4433 mddev->metadata_type[namelen] = 0;
4434 if (namelen && mddev->metadata_type[namelen-1] == '\n')
4435 mddev->metadata_type[--namelen] = 0;
4436 mddev->persistent = 0;
4437 mddev->external = 1;
NeilBrown8bb93aa2006-01-06 00:20:50 -08004438 mddev->major_version = 0;
4439 mddev->minor_version = 90;
NeilBrown67918752014-12-15 12:57:01 +11004440 goto out_unlock;
NeilBrown8bb93aa2006-01-06 00:20:50 -08004441 }
4442 major = simple_strtoul(buf, &e, 10);
NeilBrown67918752014-12-15 12:57:01 +11004443 err = -EINVAL;
NeilBrown8bb93aa2006-01-06 00:20:50 -08004444 if (e==buf || *e != '.')
NeilBrown67918752014-12-15 12:57:01 +11004445 goto out_unlock;
NeilBrown8bb93aa2006-01-06 00:20:50 -08004446 buf = e+1;
4447 minor = simple_strtoul(buf, &e, 10);
NeilBrown3f9d7b02006-12-22 01:11:41 -08004448 if (e==buf || (*e && *e != '\n') )
NeilBrown67918752014-12-15 12:57:01 +11004449 goto out_unlock;
4450 err = -ENOENT;
Ahmed S. Darwish50511da2007-05-09 02:35:34 -07004451 if (major >= ARRAY_SIZE(super_types) || super_types[major].name == NULL)
NeilBrown67918752014-12-15 12:57:01 +11004452 goto out_unlock;
NeilBrown8bb93aa2006-01-06 00:20:50 -08004453 mddev->major_version = major;
4454 mddev->minor_version = minor;
4455 mddev->persistent = 1;
NeilBrowne6910632008-02-06 01:39:51 -08004456 mddev->external = 0;
NeilBrown67918752014-12-15 12:57:01 +11004457 err = 0;
4458out_unlock:
4459 mddev_unlock(mddev);
4460 return err ?: len;
NeilBrown8bb93aa2006-01-06 00:20:50 -08004461}
4462
4463static struct md_sysfs_entry md_metadata =
NeilBrown750f1992014-09-30 08:53:05 +10004464__ATTR_PREALLOC(metadata_version, S_IRUGO|S_IWUSR, metadata_show, metadata_store);
NeilBrown8bb93aa2006-01-06 00:20:50 -08004465
NeilBrowna35b0d62006-01-06 00:20:49 -08004466static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004467action_show(struct mddev *mddev, char *page)
NeilBrown24dd4692005-11-08 21:39:26 -08004468{
NeilBrown7eec3142005-11-08 21:39:44 -08004469 char *type = "idle";
NeilBrownb7b17c92014-12-15 12:56:59 +11004470 unsigned long recovery = mddev->recovery;
4471 if (test_bit(MD_RECOVERY_FROZEN, &recovery))
NeilBrownb6a9ce62009-05-26 09:41:17 +10004472 type = "frozen";
NeilBrownb7b17c92014-12-15 12:56:59 +11004473 else if (test_bit(MD_RECOVERY_RUNNING, &recovery) ||
4474 (!mddev->ro && test_bit(MD_RECOVERY_NEEDED, &recovery))) {
4475 if (test_bit(MD_RECOVERY_RESHAPE, &recovery))
NeilBrownccfcc3c2006-03-27 01:18:09 -08004476 type = "reshape";
NeilBrownb7b17c92014-12-15 12:56:59 +11004477 else if (test_bit(MD_RECOVERY_SYNC, &recovery)) {
4478 if (!test_bit(MD_RECOVERY_REQUESTED, &recovery))
NeilBrown24dd4692005-11-08 21:39:26 -08004479 type = "resync";
NeilBrownb7b17c92014-12-15 12:56:59 +11004480 else if (test_bit(MD_RECOVERY_CHECK, &recovery))
NeilBrown24dd4692005-11-08 21:39:26 -08004481 type = "check";
4482 else
4483 type = "repair";
NeilBrownb7b17c92014-12-15 12:56:59 +11004484 } else if (test_bit(MD_RECOVERY_RECOVER, &recovery))
NeilBrown24dd4692005-11-08 21:39:26 -08004485 type = "recover";
NeilBrown985ca972015-07-06 12:26:57 +10004486 else if (mddev->reshape_position != MaxSector)
4487 type = "reshape";
NeilBrown24dd4692005-11-08 21:39:26 -08004488 }
4489 return sprintf(page, "%s\n", type);
4490}
4491
4492static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004493action_store(struct mddev *mddev, const char *page, size_t len)
NeilBrown24dd4692005-11-08 21:39:26 -08004494{
NeilBrown7eec3142005-11-08 21:39:44 -08004495 if (!mddev->pers || !mddev->pers->sync_request)
4496 return -EINVAL;
4497
NeilBrownb6a9ce62009-05-26 09:41:17 +10004498
4499 if (cmd_match(page, "idle") || cmd_match(page, "frozen")) {
NeilBrown56ccc112015-05-28 17:53:29 +10004500 if (cmd_match(page, "frozen"))
4501 set_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
4502 else
4503 clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
NeilBrown8e8e2512015-06-12 19:51:27 +10004504 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) &&
4505 mddev_lock(mddev) == 0) {
4506 flush_workqueue(md_misc_wq);
4507 if (mddev->sync_thread) {
4508 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
NeilBrown67918752014-12-15 12:57:01 +11004509 md_reap_sync_thread(mddev);
NeilBrown67918752014-12-15 12:57:01 +11004510 }
NeilBrown8e8e2512015-06-12 19:51:27 +10004511 mddev_unlock(mddev);
NeilBrown7eec3142005-11-08 21:39:44 -08004512 }
NeilBrown312045e2015-12-21 11:01:21 +11004513 } else if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
NeilBrown24dd4692005-11-08 21:39:26 -08004514 return -EBUSY;
Neil Brown72a23c22008-06-28 08:31:41 +10004515 else if (cmd_match(page, "resync"))
NeilBrown56ccc112015-05-28 17:53:29 +10004516 clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
Neil Brown72a23c22008-06-28 08:31:41 +10004517 else if (cmd_match(page, "recover")) {
NeilBrown56ccc112015-05-28 17:53:29 +10004518 clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
Neil Brown72a23c22008-06-28 08:31:41 +10004519 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
Neil Brown72a23c22008-06-28 08:31:41 +10004520 } else if (cmd_match(page, "reshape")) {
NeilBrown16484bf2006-03-27 01:18:13 -08004521 int err;
4522 if (mddev->pers->start_reshape == NULL)
4523 return -EINVAL;
NeilBrown67918752014-12-15 12:57:01 +11004524 err = mddev_lock(mddev);
4525 if (!err) {
NeilBrown312045e2015-12-21 11:01:21 +11004526 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4527 err = -EBUSY;
4528 else {
4529 clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
4530 err = mddev->pers->start_reshape(mddev);
4531 }
NeilBrown67918752014-12-15 12:57:01 +11004532 mddev_unlock(mddev);
4533 }
NeilBrown16484bf2006-03-27 01:18:13 -08004534 if (err)
4535 return err;
Neil Browna99ac972008-06-28 08:31:43 +10004536 sysfs_notify(&mddev->kobj, NULL, "degraded");
NeilBrown16484bf2006-03-27 01:18:13 -08004537 } else {
NeilBrownbce74da2006-01-06 00:20:41 -08004538 if (cmd_match(page, "check"))
NeilBrown7eec3142005-11-08 21:39:44 -08004539 set_bit(MD_RECOVERY_CHECK, &mddev->recovery);
NeilBrown2adc7d42006-05-20 14:59:57 -07004540 else if (!cmd_match(page, "repair"))
NeilBrown7eec3142005-11-08 21:39:44 -08004541 return -EINVAL;
NeilBrown56ccc112015-05-28 17:53:29 +10004542 clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
NeilBrown7eec3142005-11-08 21:39:44 -08004543 set_bit(MD_RECOVERY_REQUESTED, &mddev->recovery);
4544 set_bit(MD_RECOVERY_SYNC, &mddev->recovery);
NeilBrown7eec3142005-11-08 21:39:44 -08004545 }
NeilBrown48c26dd2012-10-11 14:19:39 +11004546 if (mddev->ro == 2) {
4547 /* A write to sync_action is enough to justify
4548 * canceling read-auto mode
4549 */
4550 mddev->ro = 0;
4551 md_wakeup_thread(mddev->sync_thread);
4552 }
NeilBrown03c902e2006-01-06 00:20:46 -08004553 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
NeilBrown24dd4692005-11-08 21:39:26 -08004554 md_wakeup_thread(mddev->thread);
NeilBrown00bcb4a2010-06-01 19:37:23 +10004555 sysfs_notify_dirent_safe(mddev->sysfs_action);
NeilBrown24dd4692005-11-08 21:39:26 -08004556 return len;
4557}
4558
Jonathan Brassowc4a39552013-06-25 01:23:59 -05004559static struct md_sysfs_entry md_scan_mode =
NeilBrown750f1992014-09-30 08:53:05 +10004560__ATTR_PREALLOC(sync_action, S_IRUGO|S_IWUSR, action_show, action_store);
Jonathan Brassowc4a39552013-06-25 01:23:59 -05004561
4562static ssize_t
4563last_sync_action_show(struct mddev *mddev, char *page)
4564{
4565 return sprintf(page, "%s\n", mddev->last_sync_action);
4566}
4567
4568static struct md_sysfs_entry md_last_scan_mode = __ATTR_RO(last_sync_action);
4569
NeilBrown9d888832005-11-08 21:39:26 -08004570static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004571mismatch_cnt_show(struct mddev *mddev, char *page)
NeilBrown9d888832005-11-08 21:39:26 -08004572{
4573 return sprintf(page, "%llu\n",
Jianpeng Ma7f7583d2012-10-11 14:17:59 +11004574 (unsigned long long)
4575 atomic64_read(&mddev->resync_mismatches));
NeilBrown9d888832005-11-08 21:39:26 -08004576}
4577
NeilBrown80ca3a42006-07-10 04:44:18 -07004578static struct md_sysfs_entry md_mismatches = __ATTR_RO(mismatch_cnt);
NeilBrown9d888832005-11-08 21:39:26 -08004579
NeilBrown88202a02006-01-06 00:21:36 -08004580static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004581sync_min_show(struct mddev *mddev, char *page)
NeilBrown88202a02006-01-06 00:21:36 -08004582{
4583 return sprintf(page, "%d (%s)\n", speed_min(mddev),
4584 mddev->sync_speed_min ? "local": "system");
4585}
4586
4587static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004588sync_min_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrown88202a02006-01-06 00:21:36 -08004589{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004590 unsigned int min;
4591 int rv;
4592
NeilBrown88202a02006-01-06 00:21:36 -08004593 if (strncmp(buf, "system", 6)==0) {
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004594 min = 0;
4595 } else {
4596 rv = kstrtouint(buf, 10, &min);
4597 if (rv < 0)
4598 return rv;
4599 if (min == 0)
4600 return -EINVAL;
NeilBrown88202a02006-01-06 00:21:36 -08004601 }
NeilBrown88202a02006-01-06 00:21:36 -08004602 mddev->sync_speed_min = min;
4603 return len;
4604}
4605
4606static struct md_sysfs_entry md_sync_min =
4607__ATTR(sync_speed_min, S_IRUGO|S_IWUSR, sync_min_show, sync_min_store);
4608
4609static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004610sync_max_show(struct mddev *mddev, char *page)
NeilBrown88202a02006-01-06 00:21:36 -08004611{
4612 return sprintf(page, "%d (%s)\n", speed_max(mddev),
4613 mddev->sync_speed_max ? "local": "system");
4614}
4615
4616static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004617sync_max_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrown88202a02006-01-06 00:21:36 -08004618{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004619 unsigned int max;
4620 int rv;
4621
NeilBrown88202a02006-01-06 00:21:36 -08004622 if (strncmp(buf, "system", 6)==0) {
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004623 max = 0;
4624 } else {
4625 rv = kstrtouint(buf, 10, &max);
4626 if (rv < 0)
4627 return rv;
4628 if (max == 0)
4629 return -EINVAL;
NeilBrown88202a02006-01-06 00:21:36 -08004630 }
NeilBrown88202a02006-01-06 00:21:36 -08004631 mddev->sync_speed_max = max;
4632 return len;
4633}
4634
4635static struct md_sysfs_entry md_sync_max =
4636__ATTR(sync_speed_max, S_IRUGO|S_IWUSR, sync_max_show, sync_max_store);
4637
Iustin Popd7f3d292007-10-16 23:30:54 -07004638static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004639degraded_show(struct mddev *mddev, char *page)
Iustin Popd7f3d292007-10-16 23:30:54 -07004640{
4641 return sprintf(page, "%d\n", mddev->degraded);
4642}
4643static struct md_sysfs_entry md_degraded = __ATTR_RO(degraded);
NeilBrown88202a02006-01-06 00:21:36 -08004644
4645static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004646sync_force_parallel_show(struct mddev *mddev, char *page)
Bernd Schubert90b08712008-05-23 13:04:38 -07004647{
4648 return sprintf(page, "%d\n", mddev->parallel_resync);
4649}
4650
4651static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004652sync_force_parallel_store(struct mddev *mddev, const char *buf, size_t len)
Bernd Schubert90b08712008-05-23 13:04:38 -07004653{
4654 long n;
4655
Jingoo Hanb29bebd2013-06-01 16:15:16 +09004656 if (kstrtol(buf, 10, &n))
Bernd Schubert90b08712008-05-23 13:04:38 -07004657 return -EINVAL;
4658
4659 if (n != 0 && n != 1)
4660 return -EINVAL;
4661
4662 mddev->parallel_resync = n;
4663
4664 if (mddev->sync_thread)
4665 wake_up(&resync_wait);
4666
4667 return len;
4668}
4669
4670/* force parallel resync, even with shared block devices */
4671static struct md_sysfs_entry md_sync_force_parallel =
4672__ATTR(sync_force_parallel, S_IRUGO|S_IWUSR,
4673 sync_force_parallel_show, sync_force_parallel_store);
4674
4675static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004676sync_speed_show(struct mddev *mddev, char *page)
NeilBrown88202a02006-01-06 00:21:36 -08004677{
4678 unsigned long resync, dt, db;
NeilBrownd1a7c502009-03-31 15:24:32 +11004679 if (mddev->curr_resync == 0)
4680 return sprintf(page, "none\n");
Andre Noll9687a602008-03-25 22:24:09 +01004681 resync = mddev->curr_mark_cnt - atomic_read(&mddev->recovery_active);
4682 dt = (jiffies - mddev->resync_mark) / HZ;
NeilBrown88202a02006-01-06 00:21:36 -08004683 if (!dt) dt++;
Andre Noll9687a602008-03-25 22:24:09 +01004684 db = resync - mddev->resync_mark_cnt;
4685 return sprintf(page, "%lu\n", db/dt/2); /* K/sec */
NeilBrown88202a02006-01-06 00:21:36 -08004686}
4687
NeilBrown80ca3a42006-07-10 04:44:18 -07004688static struct md_sysfs_entry md_sync_speed = __ATTR_RO(sync_speed);
NeilBrown88202a02006-01-06 00:21:36 -08004689
4690static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004691sync_completed_show(struct mddev *mddev, char *page)
NeilBrown88202a02006-01-06 00:21:36 -08004692{
RĂ©mi RĂ©rolle13ae8642011-01-14 09:14:34 +11004693 unsigned long long max_sectors, resync;
NeilBrown88202a02006-01-06 00:21:36 -08004694
NeilBrownacb180b2009-04-14 16:28:34 +10004695 if (!test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4696 return sprintf(page, "none\n");
4697
NeilBrown72f36d52012-10-11 14:25:57 +11004698 if (mddev->curr_resync == 1 ||
4699 mddev->curr_resync == 2)
4700 return sprintf(page, "delayed\n");
4701
NeilBrownc804cde2012-05-21 09:28:33 +10004702 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
4703 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
Andre Noll58c0fed2009-03-31 14:33:13 +11004704 max_sectors = mddev->resync_max_sectors;
NeilBrown88202a02006-01-06 00:21:36 -08004705 else
Andre Noll58c0fed2009-03-31 14:33:13 +11004706 max_sectors = mddev->dev_sectors;
NeilBrown88202a02006-01-06 00:21:36 -08004707
NeilBrownacb180b2009-04-14 16:28:34 +10004708 resync = mddev->curr_resync_completed;
RĂ©mi RĂ©rolle13ae8642011-01-14 09:14:34 +11004709 return sprintf(page, "%llu / %llu\n", resync, max_sectors);
NeilBrown88202a02006-01-06 00:21:36 -08004710}
4711
NeilBrown750f1992014-09-30 08:53:05 +10004712static struct md_sysfs_entry md_sync_completed =
4713 __ATTR_PREALLOC(sync_completed, S_IRUGO, sync_completed_show, NULL);
NeilBrown88202a02006-01-06 00:21:36 -08004714
NeilBrowne464eaf2006-03-27 01:18:14 -08004715static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004716min_sync_show(struct mddev *mddev, char *page)
Neil Brown5e96ee62008-06-28 08:31:24 +10004717{
4718 return sprintf(page, "%llu\n",
4719 (unsigned long long)mddev->resync_min);
4720}
4721static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004722min_sync_store(struct mddev *mddev, const char *buf, size_t len)
Neil Brown5e96ee62008-06-28 08:31:24 +10004723{
4724 unsigned long long min;
NeilBrown23da4222014-12-15 12:57:01 +11004725 int err;
NeilBrown23da4222014-12-15 12:57:01 +11004726
Jingoo Hanb29bebd2013-06-01 16:15:16 +09004727 if (kstrtoull(buf, 10, &min))
Neil Brown5e96ee62008-06-28 08:31:24 +10004728 return -EINVAL;
NeilBrown23da4222014-12-15 12:57:01 +11004729
4730 spin_lock(&mddev->lock);
4731 err = -EINVAL;
Neil Brown5e96ee62008-06-28 08:31:24 +10004732 if (min > mddev->resync_max)
NeilBrown23da4222014-12-15 12:57:01 +11004733 goto out_unlock;
4734
4735 err = -EBUSY;
Neil Brown5e96ee62008-06-28 08:31:24 +10004736 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
NeilBrown23da4222014-12-15 12:57:01 +11004737 goto out_unlock;
Neil Brown5e96ee62008-06-28 08:31:24 +10004738
NeilBrown50c37b12015-03-23 17:36:38 +11004739 /* Round down to multiple of 4K for safety */
4740 mddev->resync_min = round_down(min, 8);
NeilBrown23da4222014-12-15 12:57:01 +11004741 err = 0;
Neil Brown5e96ee62008-06-28 08:31:24 +10004742
NeilBrown23da4222014-12-15 12:57:01 +11004743out_unlock:
4744 spin_unlock(&mddev->lock);
4745 return err ?: len;
Neil Brown5e96ee62008-06-28 08:31:24 +10004746}
4747
4748static struct md_sysfs_entry md_min_sync =
4749__ATTR(sync_min, S_IRUGO|S_IWUSR, min_sync_show, min_sync_store);
4750
4751static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004752max_sync_show(struct mddev *mddev, char *page)
NeilBrownc6207272008-02-06 01:39:52 -08004753{
4754 if (mddev->resync_max == MaxSector)
4755 return sprintf(page, "max\n");
4756 else
4757 return sprintf(page, "%llu\n",
4758 (unsigned long long)mddev->resync_max);
4759}
4760static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004761max_sync_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrownc6207272008-02-06 01:39:52 -08004762{
NeilBrown23da4222014-12-15 12:57:01 +11004763 int err;
4764 spin_lock(&mddev->lock);
NeilBrownc6207272008-02-06 01:39:52 -08004765 if (strncmp(buf, "max", 3) == 0)
4766 mddev->resync_max = MaxSector;
4767 else {
Neil Brown5e96ee62008-06-28 08:31:24 +10004768 unsigned long long max;
NeilBrown23da4222014-12-15 12:57:01 +11004769 int chunk;
4770
4771 err = -EINVAL;
Jingoo Hanb29bebd2013-06-01 16:15:16 +09004772 if (kstrtoull(buf, 10, &max))
NeilBrown23da4222014-12-15 12:57:01 +11004773 goto out_unlock;
Neil Brown5e96ee62008-06-28 08:31:24 +10004774 if (max < mddev->resync_min)
NeilBrown23da4222014-12-15 12:57:01 +11004775 goto out_unlock;
4776
4777 err = -EBUSY;
NeilBrownc6207272008-02-06 01:39:52 -08004778 if (max < mddev->resync_max &&
NeilBrown4d484a42009-08-13 10:41:50 +10004779 mddev->ro == 0 &&
NeilBrownc6207272008-02-06 01:39:52 -08004780 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
NeilBrown23da4222014-12-15 12:57:01 +11004781 goto out_unlock;
NeilBrownc6207272008-02-06 01:39:52 -08004782
4783 /* Must be a multiple of chunk_size */
NeilBrown23da4222014-12-15 12:57:01 +11004784 chunk = mddev->chunk_sectors;
4785 if (chunk) {
raz ben yehuda2ac06c32009-06-16 17:01:42 +10004786 sector_t temp = max;
NeilBrown23da4222014-12-15 12:57:01 +11004787
4788 err = -EINVAL;
4789 if (sector_div(temp, chunk))
4790 goto out_unlock;
NeilBrownc6207272008-02-06 01:39:52 -08004791 }
4792 mddev->resync_max = max;
4793 }
4794 wake_up(&mddev->recovery_wait);
NeilBrown23da4222014-12-15 12:57:01 +11004795 err = 0;
4796out_unlock:
4797 spin_unlock(&mddev->lock);
4798 return err ?: len;
NeilBrownc6207272008-02-06 01:39:52 -08004799}
4800
4801static struct md_sysfs_entry md_max_sync =
4802__ATTR(sync_max, S_IRUGO|S_IWUSR, max_sync_show, max_sync_store);
4803
4804static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004805suspend_lo_show(struct mddev *mddev, char *page)
NeilBrowne464eaf2006-03-27 01:18:14 -08004806{
4807 return sprintf(page, "%llu\n", (unsigned long long)mddev->suspend_lo);
4808}
4809
4810static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004811suspend_lo_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrowne464eaf2006-03-27 01:18:14 -08004812{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004813 unsigned long long old, new;
NeilBrown67918752014-12-15 12:57:01 +11004814 int err;
NeilBrowne464eaf2006-03-27 01:18:14 -08004815
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004816 err = kstrtoull(buf, 10, &new);
4817 if (err < 0)
4818 return err;
4819 if (new != (sector_t)new)
NeilBrowne464eaf2006-03-27 01:18:14 -08004820 return -EINVAL;
NeilBrown23ddff32011-01-14 09:14:34 +11004821
NeilBrown67918752014-12-15 12:57:01 +11004822 err = mddev_lock(mddev);
4823 if (err)
4824 return err;
4825 err = -EINVAL;
4826 if (mddev->pers == NULL ||
4827 mddev->pers->quiesce == NULL)
4828 goto unlock;
4829 old = mddev->suspend_lo;
NeilBrown23ddff32011-01-14 09:14:34 +11004830 mddev->suspend_lo = new;
4831 if (new >= old)
4832 /* Shrinking suspended region */
NeilBrowne464eaf2006-03-27 01:18:14 -08004833 mddev->pers->quiesce(mddev, 2);
NeilBrown23ddff32011-01-14 09:14:34 +11004834 else {
4835 /* Expanding suspended region - need to wait */
4836 mddev->pers->quiesce(mddev, 1);
4837 mddev->pers->quiesce(mddev, 0);
4838 }
NeilBrown67918752014-12-15 12:57:01 +11004839 err = 0;
4840unlock:
4841 mddev_unlock(mddev);
4842 return err ?: len;
NeilBrowne464eaf2006-03-27 01:18:14 -08004843}
4844static struct md_sysfs_entry md_suspend_lo =
4845__ATTR(suspend_lo, S_IRUGO|S_IWUSR, suspend_lo_show, suspend_lo_store);
4846
NeilBrowne464eaf2006-03-27 01:18:14 -08004847static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004848suspend_hi_show(struct mddev *mddev, char *page)
NeilBrowne464eaf2006-03-27 01:18:14 -08004849{
4850 return sprintf(page, "%llu\n", (unsigned long long)mddev->suspend_hi);
4851}
4852
4853static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004854suspend_hi_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrowne464eaf2006-03-27 01:18:14 -08004855{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004856 unsigned long long old, new;
NeilBrown67918752014-12-15 12:57:01 +11004857 int err;
NeilBrowne464eaf2006-03-27 01:18:14 -08004858
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004859 err = kstrtoull(buf, 10, &new);
4860 if (err < 0)
4861 return err;
4862 if (new != (sector_t)new)
NeilBrowne464eaf2006-03-27 01:18:14 -08004863 return -EINVAL;
NeilBrown23ddff32011-01-14 09:14:34 +11004864
NeilBrown67918752014-12-15 12:57:01 +11004865 err = mddev_lock(mddev);
4866 if (err)
4867 return err;
4868 err = -EINVAL;
4869 if (mddev->pers == NULL ||
4870 mddev->pers->quiesce == NULL)
4871 goto unlock;
4872 old = mddev->suspend_hi;
NeilBrown23ddff32011-01-14 09:14:34 +11004873 mddev->suspend_hi = new;
4874 if (new <= old)
4875 /* Shrinking suspended region */
4876 mddev->pers->quiesce(mddev, 2);
4877 else {
4878 /* Expanding suspended region - need to wait */
NeilBrowne464eaf2006-03-27 01:18:14 -08004879 mddev->pers->quiesce(mddev, 1);
4880 mddev->pers->quiesce(mddev, 0);
NeilBrown23ddff32011-01-14 09:14:34 +11004881 }
NeilBrown67918752014-12-15 12:57:01 +11004882 err = 0;
4883unlock:
4884 mddev_unlock(mddev);
4885 return err ?: len;
NeilBrowne464eaf2006-03-27 01:18:14 -08004886}
4887static struct md_sysfs_entry md_suspend_hi =
4888__ATTR(suspend_hi, S_IRUGO|S_IWUSR, suspend_hi_show, suspend_hi_store);
4889
NeilBrown08a02ec2007-05-09 02:35:38 -07004890static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004891reshape_position_show(struct mddev *mddev, char *page)
NeilBrown08a02ec2007-05-09 02:35:38 -07004892{
4893 if (mddev->reshape_position != MaxSector)
4894 return sprintf(page, "%llu\n",
4895 (unsigned long long)mddev->reshape_position);
4896 strcpy(page, "none\n");
4897 return 5;
4898}
4899
4900static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004901reshape_position_store(struct mddev *mddev, const char *buf, size_t len)
NeilBrown08a02ec2007-05-09 02:35:38 -07004902{
NeilBrownc6563a82012-05-21 09:27:00 +10004903 struct md_rdev *rdev;
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004904 unsigned long long new;
NeilBrown67918752014-12-15 12:57:01 +11004905 int err;
NeilBrown67918752014-12-15 12:57:01 +11004906
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03004907 err = kstrtoull(buf, 10, &new);
4908 if (err < 0)
4909 return err;
4910 if (new != (sector_t)new)
NeilBrown08a02ec2007-05-09 02:35:38 -07004911 return -EINVAL;
NeilBrown67918752014-12-15 12:57:01 +11004912 err = mddev_lock(mddev);
4913 if (err)
4914 return err;
4915 err = -EBUSY;
4916 if (mddev->pers)
4917 goto unlock;
NeilBrown08a02ec2007-05-09 02:35:38 -07004918 mddev->reshape_position = new;
4919 mddev->delta_disks = 0;
NeilBrown2c810cd2012-05-21 09:27:00 +10004920 mddev->reshape_backwards = 0;
NeilBrown08a02ec2007-05-09 02:35:38 -07004921 mddev->new_level = mddev->level;
4922 mddev->new_layout = mddev->layout;
Andre Noll664e7c42009-06-18 08:45:27 +10004923 mddev->new_chunk_sectors = mddev->chunk_sectors;
NeilBrownc6563a82012-05-21 09:27:00 +10004924 rdev_for_each(rdev, mddev)
4925 rdev->new_data_offset = rdev->data_offset;
NeilBrown67918752014-12-15 12:57:01 +11004926 err = 0;
4927unlock:
4928 mddev_unlock(mddev);
4929 return err ?: len;
NeilBrown08a02ec2007-05-09 02:35:38 -07004930}
4931
4932static struct md_sysfs_entry md_reshape_position =
4933__ATTR(reshape_position, S_IRUGO|S_IWUSR, reshape_position_show,
4934 reshape_position_store);
4935
Dan Williamsb522adc2009-03-31 15:00:31 +11004936static ssize_t
NeilBrown2c810cd2012-05-21 09:27:00 +10004937reshape_direction_show(struct mddev *mddev, char *page)
4938{
4939 return sprintf(page, "%s\n",
4940 mddev->reshape_backwards ? "backwards" : "forwards");
4941}
4942
4943static ssize_t
4944reshape_direction_store(struct mddev *mddev, const char *buf, size_t len)
4945{
4946 int backwards = 0;
NeilBrown67918752014-12-15 12:57:01 +11004947 int err;
4948
NeilBrown2c810cd2012-05-21 09:27:00 +10004949 if (cmd_match(buf, "forwards"))
4950 backwards = 0;
4951 else if (cmd_match(buf, "backwards"))
4952 backwards = 1;
4953 else
4954 return -EINVAL;
4955 if (mddev->reshape_backwards == backwards)
4956 return len;
4957
NeilBrown67918752014-12-15 12:57:01 +11004958 err = mddev_lock(mddev);
4959 if (err)
4960 return err;
NeilBrown2c810cd2012-05-21 09:27:00 +10004961 /* check if we are allowed to change */
4962 if (mddev->delta_disks)
NeilBrown67918752014-12-15 12:57:01 +11004963 err = -EBUSY;
4964 else if (mddev->persistent &&
NeilBrown2c810cd2012-05-21 09:27:00 +10004965 mddev->major_version == 0)
NeilBrown67918752014-12-15 12:57:01 +11004966 err = -EINVAL;
4967 else
4968 mddev->reshape_backwards = backwards;
4969 mddev_unlock(mddev);
4970 return err ?: len;
NeilBrown2c810cd2012-05-21 09:27:00 +10004971}
4972
4973static struct md_sysfs_entry md_reshape_direction =
4974__ATTR(reshape_direction, S_IRUGO|S_IWUSR, reshape_direction_show,
4975 reshape_direction_store);
4976
4977static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004978array_size_show(struct mddev *mddev, char *page)
Dan Williamsb522adc2009-03-31 15:00:31 +11004979{
4980 if (mddev->external_size)
4981 return sprintf(page, "%llu\n",
4982 (unsigned long long)mddev->array_sectors/2);
4983 else
4984 return sprintf(page, "default\n");
4985}
4986
4987static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004988array_size_store(struct mddev *mddev, const char *buf, size_t len)
Dan Williamsb522adc2009-03-31 15:00:31 +11004989{
4990 sector_t sectors;
NeilBrown67918752014-12-15 12:57:01 +11004991 int err;
4992
4993 err = mddev_lock(mddev);
4994 if (err)
4995 return err;
Dan Williamsb522adc2009-03-31 15:00:31 +11004996
Guoqing Jiangab5a98b2016-05-02 11:33:13 -04004997 /* cluster raid doesn't support change array_sectors */
Zhilong Liub6708832017-04-10 14:15:55 +08004998 if (mddev_is_clustered(mddev)) {
4999 mddev_unlock(mddev);
Guoqing Jiangab5a98b2016-05-02 11:33:13 -04005000 return -EINVAL;
Zhilong Liub6708832017-04-10 14:15:55 +08005001 }
Guoqing Jiangab5a98b2016-05-02 11:33:13 -04005002
Dan Williamsb522adc2009-03-31 15:00:31 +11005003 if (strncmp(buf, "default", 7) == 0) {
5004 if (mddev->pers)
5005 sectors = mddev->pers->size(mddev, 0, 0);
5006 else
5007 sectors = mddev->array_sectors;
5008
5009 mddev->external_size = 0;
5010 } else {
5011 if (strict_blocks_to_sectors(buf, &sectors) < 0)
NeilBrown67918752014-12-15 12:57:01 +11005012 err = -EINVAL;
5013 else if (mddev->pers && mddev->pers->size(mddev, 0, 0) < sectors)
5014 err = -E2BIG;
5015 else
5016 mddev->external_size = 1;
Dan Williamsb522adc2009-03-31 15:00:31 +11005017 }
5018
NeilBrown67918752014-12-15 12:57:01 +11005019 if (!err) {
5020 mddev->array_sectors = sectors;
5021 if (mddev->pers) {
5022 set_capacity(mddev->gendisk, mddev->array_sectors);
5023 revalidate_disk(mddev->gendisk);
5024 }
NeilBrowncbe6ef12011-02-16 13:58:38 +11005025 }
NeilBrown67918752014-12-15 12:57:01 +11005026 mddev_unlock(mddev);
5027 return err ?: len;
Dan Williamsb522adc2009-03-31 15:00:31 +11005028}
5029
5030static struct md_sysfs_entry md_array_size =
5031__ATTR(array_size, S_IRUGO|S_IWUSR, array_size_show,
5032 array_size_store);
NeilBrowne464eaf2006-03-27 01:18:14 -08005033
Artur Paszkiewicz664aed02017-03-09 10:00:00 +01005034static ssize_t
5035consistency_policy_show(struct mddev *mddev, char *page)
5036{
5037 int ret;
5038
5039 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
5040 ret = sprintf(page, "journal\n");
5041 } else if (test_bit(MD_HAS_PPL, &mddev->flags)) {
5042 ret = sprintf(page, "ppl\n");
5043 } else if (mddev->bitmap) {
5044 ret = sprintf(page, "bitmap\n");
5045 } else if (mddev->pers) {
5046 if (mddev->pers->sync_request)
5047 ret = sprintf(page, "resync\n");
5048 else
5049 ret = sprintf(page, "none\n");
5050 } else {
5051 ret = sprintf(page, "unknown\n");
5052 }
5053
5054 return ret;
5055}
5056
5057static ssize_t
5058consistency_policy_store(struct mddev *mddev, const char *buf, size_t len)
5059{
Artur Paszkiewiczba903a32017-03-09 10:00:03 +01005060 int err = 0;
5061
Artur Paszkiewicz664aed02017-03-09 10:00:00 +01005062 if (mddev->pers) {
Artur Paszkiewiczba903a32017-03-09 10:00:03 +01005063 if (mddev->pers->change_consistency_policy)
5064 err = mddev->pers->change_consistency_policy(mddev, buf);
5065 else
5066 err = -EBUSY;
Artur Paszkiewicz664aed02017-03-09 10:00:00 +01005067 } else if (mddev->external && strncmp(buf, "ppl", 3) == 0) {
5068 set_bit(MD_HAS_PPL, &mddev->flags);
Artur Paszkiewicz664aed02017-03-09 10:00:00 +01005069 } else {
Artur Paszkiewiczba903a32017-03-09 10:00:03 +01005070 err = -EINVAL;
Artur Paszkiewicz664aed02017-03-09 10:00:00 +01005071 }
Artur Paszkiewiczba903a32017-03-09 10:00:03 +01005072
5073 return err ? err : len;
Artur Paszkiewicz664aed02017-03-09 10:00:00 +01005074}
5075
5076static struct md_sysfs_entry md_consistency_policy =
5077__ATTR(consistency_policy, S_IRUGO | S_IWUSR, consistency_policy_show,
5078 consistency_policy_store);
5079
NeilBrowneae17012005-11-08 21:39:23 -08005080static struct attribute *md_default_attrs[] = {
5081 &md_level.attr,
NeilBrownd4dbd022006-06-26 00:27:59 -07005082 &md_layout.attr,
NeilBrowneae17012005-11-08 21:39:23 -08005083 &md_raid_disks.attr,
NeilBrown3b343802006-01-06 00:20:47 -08005084 &md_chunk_size.attr,
NeilBrowna35b0d62006-01-06 00:20:49 -08005085 &md_size.attr,
NeilBrowna94213b2006-06-26 00:28:00 -07005086 &md_resync_start.attr,
NeilBrown8bb93aa2006-01-06 00:20:50 -08005087 &md_metadata.attr,
NeilBrown6d7ff732006-01-06 00:21:16 -08005088 &md_new_device.attr,
NeilBrown16f17b32006-06-26 00:27:37 -07005089 &md_safe_delay.attr,
NeilBrown9e653b62006-06-26 00:27:58 -07005090 &md_array_state.attr,
NeilBrown08a02ec2007-05-09 02:35:38 -07005091 &md_reshape_position.attr,
NeilBrown2c810cd2012-05-21 09:27:00 +10005092 &md_reshape_direction.attr,
Dan Williamsb522adc2009-03-31 15:00:31 +11005093 &md_array_size.attr,
Robert Becker1e509152009-12-14 12:49:58 +11005094 &max_corr_read_errors.attr,
Artur Paszkiewicz664aed02017-03-09 10:00:00 +01005095 &md_consistency_policy.attr,
NeilBrown411036f2005-11-08 21:39:40 -08005096 NULL,
5097};
5098
5099static struct attribute *md_redundancy_attrs[] = {
NeilBrown24dd4692005-11-08 21:39:26 -08005100 &md_scan_mode.attr,
Jonathan Brassowc4a39552013-06-25 01:23:59 -05005101 &md_last_scan_mode.attr,
NeilBrown9d888832005-11-08 21:39:26 -08005102 &md_mismatches.attr,
NeilBrown88202a02006-01-06 00:21:36 -08005103 &md_sync_min.attr,
5104 &md_sync_max.attr,
5105 &md_sync_speed.attr,
Bernd Schubert90b08712008-05-23 13:04:38 -07005106 &md_sync_force_parallel.attr,
NeilBrown88202a02006-01-06 00:21:36 -08005107 &md_sync_completed.attr,
Neil Brown5e96ee62008-06-28 08:31:24 +10005108 &md_min_sync.attr,
NeilBrownc6207272008-02-06 01:39:52 -08005109 &md_max_sync.attr,
NeilBrowne464eaf2006-03-27 01:18:14 -08005110 &md_suspend_lo.attr,
5111 &md_suspend_hi.attr,
Paul Clements9b1d1da2006-10-03 01:15:49 -07005112 &md_bitmap.attr,
Iustin Popd7f3d292007-10-16 23:30:54 -07005113 &md_degraded.attr,
NeilBrowneae17012005-11-08 21:39:23 -08005114 NULL,
5115};
NeilBrown411036f2005-11-08 21:39:40 -08005116static struct attribute_group md_redundancy_group = {
5117 .name = NULL,
5118 .attrs = md_redundancy_attrs,
5119};
5120
NeilBrowneae17012005-11-08 21:39:23 -08005121static ssize_t
5122md_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
5123{
5124 struct md_sysfs_entry *entry = container_of(attr, struct md_sysfs_entry, attr);
NeilBrownfd01b882011-10-11 16:47:53 +11005125 struct mddev *mddev = container_of(kobj, struct mddev, kobj);
NeilBrown96de1e62005-11-08 21:39:39 -08005126 ssize_t rv;
NeilBrowneae17012005-11-08 21:39:23 -08005127
5128 if (!entry->show)
5129 return -EIO;
NeilBrownaf8a2432011-12-08 15:49:46 +11005130 spin_lock(&all_mddevs_lock);
5131 if (list_empty(&mddev->all_mddevs)) {
5132 spin_unlock(&all_mddevs_lock);
5133 return -EBUSY;
5134 }
5135 mddev_get(mddev);
5136 spin_unlock(&all_mddevs_lock);
5137
NeilBrownb7b17c92014-12-15 12:56:59 +11005138 rv = entry->show(mddev, page);
NeilBrownaf8a2432011-12-08 15:49:46 +11005139 mddev_put(mddev);
NeilBrown96de1e62005-11-08 21:39:39 -08005140 return rv;
NeilBrowneae17012005-11-08 21:39:23 -08005141}
5142
5143static ssize_t
5144md_attr_store(struct kobject *kobj, struct attribute *attr,
5145 const char *page, size_t length)
5146{
5147 struct md_sysfs_entry *entry = container_of(attr, struct md_sysfs_entry, attr);
NeilBrownfd01b882011-10-11 16:47:53 +11005148 struct mddev *mddev = container_of(kobj, struct mddev, kobj);
NeilBrown96de1e62005-11-08 21:39:39 -08005149 ssize_t rv;
NeilBrowneae17012005-11-08 21:39:23 -08005150
5151 if (!entry->store)
5152 return -EIO;
NeilBrown67463ac2006-07-10 04:44:19 -07005153 if (!capable(CAP_SYS_ADMIN))
5154 return -EACCES;
NeilBrownaf8a2432011-12-08 15:49:46 +11005155 spin_lock(&all_mddevs_lock);
5156 if (list_empty(&mddev->all_mddevs)) {
5157 spin_unlock(&all_mddevs_lock);
5158 return -EBUSY;
5159 }
5160 mddev_get(mddev);
5161 spin_unlock(&all_mddevs_lock);
NeilBrown67918752014-12-15 12:57:01 +11005162 rv = entry->store(mddev, page, length);
NeilBrownaf8a2432011-12-08 15:49:46 +11005163 mddev_put(mddev);
NeilBrown96de1e62005-11-08 21:39:39 -08005164 return rv;
NeilBrowneae17012005-11-08 21:39:23 -08005165}
5166
5167static void md_free(struct kobject *ko)
5168{
NeilBrownfd01b882011-10-11 16:47:53 +11005169 struct mddev *mddev = container_of(ko, struct mddev, kobj);
NeilBrowna21d1502009-01-09 08:31:09 +11005170
5171 if (mddev->sysfs_state)
5172 sysfs_put(mddev->sysfs_state);
5173
NeilBrown6cd18e72015-04-27 14:12:22 +10005174 if (mddev->queue)
5175 blk_cleanup_queue(mddev->queue);
NeilBrowna21d1502009-01-09 08:31:09 +11005176 if (mddev->gendisk) {
5177 del_gendisk(mddev->gendisk);
5178 put_disk(mddev->gendisk);
5179 }
NeilBrown4ad23a972017-03-15 14:05:14 +11005180 percpu_ref_exit(&mddev->writes_pending);
NeilBrowna21d1502009-01-09 08:31:09 +11005181
NeilBrowneae17012005-11-08 21:39:23 -08005182 kfree(mddev);
5183}
5184
Emese Revfy52cf25d2010-01-19 02:58:23 +01005185static const struct sysfs_ops md_sysfs_ops = {
NeilBrowneae17012005-11-08 21:39:23 -08005186 .show = md_attr_show,
5187 .store = md_attr_store,
5188};
5189static struct kobj_type md_ktype = {
5190 .release = md_free,
5191 .sysfs_ops = &md_sysfs_ops,
5192 .default_attrs = md_default_attrs,
5193};
5194
Linus Torvalds1da177e2005-04-16 15:20:36 -07005195int mdp_major = 0;
5196
Dan Williams5fd3a172009-03-04 00:57:25 -07005197static void mddev_delayed_delete(struct work_struct *ws)
5198{
NeilBrownfd01b882011-10-11 16:47:53 +11005199 struct mddev *mddev = container_of(ws, struct mddev, del_work);
Dan Williams5fd3a172009-03-04 00:57:25 -07005200
NeilBrown43a70502009-12-14 12:49:55 +11005201 sysfs_remove_group(&mddev->kobj, &md_bitmap_group);
Dan Williams5fd3a172009-03-04 00:57:25 -07005202 kobject_del(&mddev->kobj);
5203 kobject_put(&mddev->kobj);
5204}
5205
NeilBrown4ad23a972017-03-15 14:05:14 +11005206static void no_op(struct percpu_ref *r) {}
5207
NeilBrowna415c0f2017-06-05 16:05:13 +10005208int mddev_init_writes_pending(struct mddev *mddev)
5209{
5210 if (mddev->writes_pending.percpu_count_ptr)
5211 return 0;
5212 if (percpu_ref_init(&mddev->writes_pending, no_op, 0, GFP_KERNEL) < 0)
5213 return -ENOMEM;
5214 /* We want to start with the refcount at zero */
5215 percpu_ref_put(&mddev->writes_pending);
5216 return 0;
5217}
5218EXPORT_SYMBOL_GPL(mddev_init_writes_pending);
5219
NeilBrownefeb53c2009-01-09 08:31:10 +11005220static int md_alloc(dev_t dev, char *name)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005221{
NeilBrown039b7222017-04-12 16:26:13 +10005222 /*
5223 * If dev is zero, name is the name of a device to allocate with
5224 * an arbitrary minor number. It will be "md_???"
5225 * If dev is non-zero it must be a device number with a MAJOR of
5226 * MD_MAJOR or mdp_major. In this case, if "name" is NULL, then
5227 * the device is being created by opening a node in /dev.
5228 * If "name" is not NULL, the device is being created by
5229 * writing to /sys/module/md_mod/parameters/new_array.
5230 */
Arjan van de Ven48c9c272006-03-27 01:18:20 -08005231 static DEFINE_MUTEX(disks_mutex);
NeilBrownfd01b882011-10-11 16:47:53 +11005232 struct mddev *mddev = mddev_find(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005233 struct gendisk *disk;
NeilBrownefeb53c2009-01-09 08:31:10 +11005234 int partitioned;
5235 int shift;
5236 int unit;
Greg Kroah-Hartman3830c622007-12-17 15:54:39 -04005237 int error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005238
5239 if (!mddev)
NeilBrownefeb53c2009-01-09 08:31:10 +11005240 return -ENODEV;
5241
5242 partitioned = (MAJOR(mddev->unit) != MD_MAJOR);
5243 shift = partitioned ? MdpMinorShift : 0;
5244 unit = MINOR(mddev->unit) >> shift;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005245
Tejun Heoe804ac72010-10-15 15:36:08 +02005246 /* wait for any previous instance of this device to be
5247 * completely removed (mddev_delayed_delete).
NeilBrownd3374822009-01-09 08:31:10 +11005248 */
Tejun Heoe804ac72010-10-15 15:36:08 +02005249 flush_workqueue(md_misc_wq);
NeilBrownd3374822009-01-09 08:31:10 +11005250
Arjan van de Ven48c9c272006-03-27 01:18:20 -08005251 mutex_lock(&disks_mutex);
NeilBrown0909dc42009-07-01 12:27:21 +10005252 error = -EEXIST;
5253 if (mddev->gendisk)
5254 goto abort;
NeilBrownefeb53c2009-01-09 08:31:10 +11005255
NeilBrown039b7222017-04-12 16:26:13 +10005256 if (name && !dev) {
NeilBrownefeb53c2009-01-09 08:31:10 +11005257 /* Need to ensure that 'name' is not a duplicate.
5258 */
NeilBrownfd01b882011-10-11 16:47:53 +11005259 struct mddev *mddev2;
NeilBrownefeb53c2009-01-09 08:31:10 +11005260 spin_lock(&all_mddevs_lock);
5261
5262 list_for_each_entry(mddev2, &all_mddevs, all_mddevs)
5263 if (mddev2->gendisk &&
5264 strcmp(mddev2->gendisk->disk_name, name) == 0) {
5265 spin_unlock(&all_mddevs_lock);
NeilBrown0909dc42009-07-01 12:27:21 +10005266 goto abort;
NeilBrownefeb53c2009-01-09 08:31:10 +11005267 }
5268 spin_unlock(&all_mddevs_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005269 }
NeilBrown039b7222017-04-12 16:26:13 +10005270 if (name && dev)
5271 /*
5272 * Creating /dev/mdNNN via "newarray", so adjust hold_active.
5273 */
5274 mddev->hold_active = UNTIL_STOP;
NeilBrown8b765392009-01-09 08:31:08 +11005275
NeilBrown0909dc42009-07-01 12:27:21 +10005276 error = -ENOMEM;
NeilBrown8b765392009-01-09 08:31:08 +11005277 mddev->queue = blk_alloc_queue(GFP_KERNEL);
NeilBrown0909dc42009-07-01 12:27:21 +10005278 if (!mddev->queue)
5279 goto abort;
NeilBrown409c57f2009-03-31 14:39:39 +11005280 mddev->queue->queuedata = mddev;
5281
NeilBrown409c57f2009-03-31 14:39:39 +11005282 blk_queue_make_request(mddev->queue, md_make_request);
Martin K. Petersenb1bd0552012-01-11 16:27:11 +01005283 blk_set_stacking_limits(&mddev->queue->limits);
NeilBrown8b765392009-01-09 08:31:08 +11005284
Linus Torvalds1da177e2005-04-16 15:20:36 -07005285 disk = alloc_disk(1 << shift);
5286 if (!disk) {
NeilBrown8b765392009-01-09 08:31:08 +11005287 blk_cleanup_queue(mddev->queue);
5288 mddev->queue = NULL;
NeilBrown0909dc42009-07-01 12:27:21 +10005289 goto abort;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005290 }
NeilBrownefeb53c2009-01-09 08:31:10 +11005291 disk->major = MAJOR(mddev->unit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005292 disk->first_minor = unit << shift;
NeilBrownefeb53c2009-01-09 08:31:10 +11005293 if (name)
5294 strcpy(disk->disk_name, name);
5295 else if (partitioned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005296 sprintf(disk->disk_name, "md_d%d", unit);
Greg Kroah-Hartmance7b0f42005-06-20 21:15:16 -07005297 else
Linus Torvalds1da177e2005-04-16 15:20:36 -07005298 sprintf(disk->disk_name, "md%d", unit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005299 disk->fops = &md_fops;
5300 disk->private_data = mddev;
5301 disk->queue = mddev->queue;
Jens Axboe56883a72016-03-30 10:16:53 -06005302 blk_queue_write_cache(mddev->queue, true, true);
NeilBrown92850bb2008-10-21 13:25:32 +11005303 /* Allow extended partitions. This makes the
NeilBrownd3374822009-01-09 08:31:10 +11005304 * 'mdp' device redundant, but we can't really
NeilBrown92850bb2008-10-21 13:25:32 +11005305 * remove it now.
5306 */
5307 disk->flags |= GENHD_FL_EXT_DEVT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005308 mddev->gendisk = disk;
NeilBrownb0140892011-05-10 17:49:01 +10005309 /* As soon as we call add_disk(), another thread could get
5310 * through to md_open, so make sure it doesn't get too far
5311 */
5312 mutex_lock(&mddev->open_mutex);
5313 add_disk(disk);
5314
Tejun Heoed9e1982008-08-25 19:56:05 +09005315 error = kobject_init_and_add(&mddev->kobj, &md_ktype,
5316 &disk_to_dev(disk)->kobj, "%s", "md");
NeilBrown0909dc42009-07-01 12:27:21 +10005317 if (error) {
5318 /* This isn't possible, but as kobject_init_and_add is marked
5319 * __must_check, we must do something with the result
5320 */
NeilBrown9d487392016-11-02 14:16:49 +11005321 pr_debug("md: cannot register %s/md - name in use\n",
5322 disk->disk_name);
NeilBrown0909dc42009-07-01 12:27:21 +10005323 error = 0;
5324 }
NeilBrown00bcb4a2010-06-01 19:37:23 +10005325 if (mddev->kobj.sd &&
5326 sysfs_create_group(&mddev->kobj, &md_bitmap_group))
NeilBrown9d487392016-11-02 14:16:49 +11005327 pr_debug("pointless warning\n");
NeilBrownb0140892011-05-10 17:49:01 +10005328 mutex_unlock(&mddev->open_mutex);
NeilBrown0909dc42009-07-01 12:27:21 +10005329 abort:
5330 mutex_unlock(&disks_mutex);
NeilBrown00bcb4a2010-06-01 19:37:23 +10005331 if (!error && mddev->kobj.sd) {
Greg Kroah-Hartman3830c622007-12-17 15:54:39 -04005332 kobject_uevent(&mddev->kobj, KOBJ_ADD);
NeilBrown00bcb4a2010-06-01 19:37:23 +10005333 mddev->sysfs_state = sysfs_get_dirent_safe(mddev->kobj.sd, "array_state");
NeilBrownb62b7592008-10-21 13:25:21 +11005334 }
NeilBrownd3374822009-01-09 08:31:10 +11005335 mddev_put(mddev);
NeilBrown0909dc42009-07-01 12:27:21 +10005336 return error;
NeilBrownefeb53c2009-01-09 08:31:10 +11005337}
5338
5339static struct kobject *md_probe(dev_t dev, int *part, void *data)
5340{
NeilBrown78b63502017-04-12 16:26:13 +10005341 if (create_on_open)
5342 md_alloc(dev, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005343 return NULL;
5344}
5345
NeilBrownefeb53c2009-01-09 08:31:10 +11005346static int add_named_array(const char *val, struct kernel_param *kp)
5347{
NeilBrown039b7222017-04-12 16:26:13 +10005348 /*
5349 * val must be "md_*" or "mdNNN".
5350 * For "md_*" we allocate an array with a large free minor number, and
NeilBrownefeb53c2009-01-09 08:31:10 +11005351 * set the name to val. val must not already be an active name.
NeilBrown039b7222017-04-12 16:26:13 +10005352 * For "mdNNN" we allocate an array with the minor number NNN
5353 * which must not already be in use.
NeilBrownefeb53c2009-01-09 08:31:10 +11005354 */
5355 int len = strlen(val);
5356 char buf[DISK_NAME_LEN];
NeilBrown039b7222017-04-12 16:26:13 +10005357 unsigned long devnum;
NeilBrownefeb53c2009-01-09 08:31:10 +11005358
5359 while (len && val[len-1] == '\n')
5360 len--;
5361 if (len >= DISK_NAME_LEN)
5362 return -E2BIG;
5363 strlcpy(buf, val, len+1);
NeilBrown039b7222017-04-12 16:26:13 +10005364 if (strncmp(buf, "md_", 3) == 0)
5365 return md_alloc(0, buf);
5366 if (strncmp(buf, "md", 2) == 0 &&
5367 isdigit(buf[2]) &&
5368 kstrtoul(buf+2, 10, &devnum) == 0 &&
5369 devnum <= MINORMASK)
5370 return md_alloc(MKDEV(MD_MAJOR, devnum), NULL);
5371
5372 return -EINVAL;
NeilBrownefeb53c2009-01-09 08:31:10 +11005373}
5374
Linus Torvalds1da177e2005-04-16 15:20:36 -07005375static void md_safemode_timeout(unsigned long data)
5376{
NeilBrownfd01b882011-10-11 16:47:53 +11005377 struct mddev *mddev = (struct mddev *) data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005378
NeilBrown4ad23a972017-03-15 14:05:14 +11005379 mddev->safemode = 1;
5380 if (mddev->external)
5381 sysfs_notify_dirent_safe(mddev->sysfs_state);
5382
Linus Torvalds1da177e2005-04-16 15:20:36 -07005383 md_wakeup_thread(mddev->thread);
5384}
5385
NeilBrown6ff8d8ec2006-01-06 00:20:15 -08005386static int start_dirty_degraded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005387
NeilBrownfd01b882011-10-11 16:47:53 +11005388int md_run(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005389{
NeilBrown2604b702006-01-06 00:20:36 -08005390 int err;
NeilBrown3cb03002011-10-11 16:45:26 +11005391 struct md_rdev *rdev;
NeilBrown84fc4b52011-10-11 16:49:58 +11005392 struct md_personality *pers;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005393
NeilBrowna757e642005-04-16 15:26:42 -07005394 if (list_empty(&mddev->disks))
5395 /* cannot run an array with no devices.. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005396 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005397
5398 if (mddev->pers)
5399 return -EBUSY;
NeilBrownbb4f1e92010-08-08 21:18:03 +10005400 /* Cannot run until previous stop completes properly */
5401 if (mddev->sysfs_active)
5402 return -EBUSY;
NeilBrownb6eb1272010-04-15 10:13:47 +10005403
Linus Torvalds1da177e2005-04-16 15:20:36 -07005404 /*
5405 * Analyze all RAID superblock(s)
5406 */
NeilBrown1ec4a932008-02-06 01:39:53 -08005407 if (!mddev->raid_disks) {
5408 if (!mddev->persistent)
5409 return -EINVAL;
NeilBrowna757e642005-04-16 15:26:42 -07005410 analyze_sbs(mddev);
NeilBrown1ec4a932008-02-06 01:39:53 -08005411 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005412
NeilBrownd9d166c2006-01-06 00:20:51 -08005413 if (mddev->level != LEVEL_NONE)
5414 request_module("md-level-%d", mddev->level);
5415 else if (mddev->clevel[0])
5416 request_module("md-%s", mddev->clevel);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005417
5418 /*
5419 * Drop all container device buffers, from now on
5420 * the only valid external interface is through the md
5421 * device.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005422 */
NeilBrowndafb20f2012-03-19 12:46:39 +11005423 rdev_for_each(rdev, mddev) {
NeilBrownb2d444d2005-11-08 21:39:31 -08005424 if (test_bit(Faulty, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005425 continue;
5426 sync_blockdev(rdev->bdev);
Peter Zijlstraf98393a2007-05-06 14:49:54 -07005427 invalidate_bdev(rdev->bdev);
NeilBrown97b20ef2017-04-13 08:53:48 +10005428 if (mddev->ro != 1 &&
5429 (bdev_read_only(rdev->bdev) ||
5430 bdev_read_only(rdev->meta_bdev))) {
5431 mddev->ro = 1;
5432 if (mddev->gendisk)
5433 set_disk_ro(mddev->gendisk, 1);
5434 }
NeilBrownf0d76d72007-07-17 04:06:12 -07005435
5436 /* perform some consistency tests on the device.
5437 * We don't want the data to overlap the metadata,
Andre Noll58c0fed2009-03-31 14:33:13 +11005438 * Internal Bitmap issues have been handled elsewhere.
NeilBrownf0d76d72007-07-17 04:06:12 -07005439 */
Jonathan Brassowa6ff7e02011-01-14 09:14:34 +11005440 if (rdev->meta_bdev) {
5441 /* Nothing to check */;
5442 } else if (rdev->data_offset < rdev->sb_start) {
Andre Noll58c0fed2009-03-31 14:33:13 +11005443 if (mddev->dev_sectors &&
5444 rdev->data_offset + mddev->dev_sectors
Andre Noll0f420352008-07-11 22:02:23 +10005445 > rdev->sb_start) {
NeilBrown9d487392016-11-02 14:16:49 +11005446 pr_warn("md: %s: data overlaps metadata\n",
5447 mdname(mddev));
NeilBrownf0d76d72007-07-17 04:06:12 -07005448 return -EINVAL;
5449 }
5450 } else {
Andre Noll0f420352008-07-11 22:02:23 +10005451 if (rdev->sb_start + rdev->sb_size/512
NeilBrownf0d76d72007-07-17 04:06:12 -07005452 > rdev->data_offset) {
NeilBrown9d487392016-11-02 14:16:49 +11005453 pr_warn("md: %s: metadata overlaps data\n",
5454 mdname(mddev));
NeilBrownf0d76d72007-07-17 04:06:12 -07005455 return -EINVAL;
5456 }
5457 }
NeilBrown00bcb4a2010-06-01 19:37:23 +10005458 sysfs_notify_dirent_safe(rdev->sysfs_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005459 }
5460
Ming Lei10273172017-02-14 23:29:00 +08005461 if (mddev->bio_set == NULL) {
NeilBrown011067b2017-06-18 14:38:57 +10005462 mddev->bio_set = bioset_create(BIO_POOL_SIZE, 0, BIOSET_NEED_BVECS);
Ming Lei10273172017-02-14 23:29:00 +08005463 if (!mddev->bio_set)
5464 return -ENOMEM;
5465 }
NeilBrown5a850712017-06-21 09:12:21 +10005466 if (mddev->sync_set == NULL) {
Linus Torvalds026d15f2017-07-08 12:50:18 -07005467 mddev->sync_set = bioset_create(BIO_POOL_SIZE, 0, BIOSET_NEED_BVECS);
NeilBrown5a850712017-06-21 09:12:21 +10005468 if (!mddev->sync_set)
5469 return -ENOMEM;
5470 }
NeilBrowna167f662010-10-26 18:31:13 +11005471
Linus Torvalds1da177e2005-04-16 15:20:36 -07005472 spin_lock(&pers_lock);
NeilBrownd9d166c2006-01-06 00:20:51 -08005473 pers = find_pers(mddev->level, mddev->clevel);
NeilBrown2604b702006-01-06 00:20:36 -08005474 if (!pers || !try_module_get(pers->owner)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005475 spin_unlock(&pers_lock);
NeilBrownd9d166c2006-01-06 00:20:51 -08005476 if (mddev->level != LEVEL_NONE)
NeilBrown9d487392016-11-02 14:16:49 +11005477 pr_warn("md: personality for level %d is not loaded!\n",
5478 mddev->level);
NeilBrownd9d166c2006-01-06 00:20:51 -08005479 else
NeilBrown9d487392016-11-02 14:16:49 +11005480 pr_warn("md: personality for level %s is not loaded!\n",
5481 mddev->clevel);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005482 return -EINVAL;
5483 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005484 spin_unlock(&pers_lock);
NeilBrown34817e82009-03-31 14:39:38 +11005485 if (mddev->level != pers->level) {
5486 mddev->level = pers->level;
5487 mddev->new_level = pers->level;
5488 }
NeilBrownd9d166c2006-01-06 00:20:51 -08005489 strlcpy(mddev->clevel, pers->name, sizeof(mddev->clevel));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005490
NeilBrownf6705572006-03-27 01:18:11 -08005491 if (mddev->reshape_position != MaxSector &&
NeilBrown63c70c42006-03-27 01:18:13 -08005492 pers->start_reshape == NULL) {
NeilBrownf6705572006-03-27 01:18:11 -08005493 /* This personality cannot handle reshaping... */
NeilBrownf6705572006-03-27 01:18:11 -08005494 module_put(pers->owner);
5495 return -EINVAL;
5496 }
5497
NeilBrown7dd5e7c32007-02-28 20:11:35 -08005498 if (pers->sync_request) {
5499 /* Warn if this is a potentially silly
5500 * configuration.
5501 */
5502 char b[BDEVNAME_SIZE], b2[BDEVNAME_SIZE];
NeilBrown3cb03002011-10-11 16:45:26 +11005503 struct md_rdev *rdev2;
NeilBrown7dd5e7c32007-02-28 20:11:35 -08005504 int warned = 0;
Cheng Renquan159ec1f2009-01-09 08:31:08 +11005505
NeilBrowndafb20f2012-03-19 12:46:39 +11005506 rdev_for_each(rdev, mddev)
5507 rdev_for_each(rdev2, mddev) {
NeilBrown7dd5e7c32007-02-28 20:11:35 -08005508 if (rdev < rdev2 &&
5509 rdev->bdev->bd_contains ==
5510 rdev2->bdev->bd_contains) {
NeilBrown9d487392016-11-02 14:16:49 +11005511 pr_warn("%s: WARNING: %s appears to be on the same physical disk as %s.\n",
5512 mdname(mddev),
5513 bdevname(rdev->bdev,b),
5514 bdevname(rdev2->bdev,b2));
NeilBrown7dd5e7c32007-02-28 20:11:35 -08005515 warned = 1;
5516 }
5517 }
Cheng Renquan159ec1f2009-01-09 08:31:08 +11005518
NeilBrown7dd5e7c32007-02-28 20:11:35 -08005519 if (warned)
NeilBrown9d487392016-11-02 14:16:49 +11005520 pr_warn("True protection against single-disk failure might be compromised.\n");
NeilBrown7dd5e7c32007-02-28 20:11:35 -08005521 }
5522
NeilBrown657390d2005-08-26 18:34:16 -07005523 mddev->recovery = 0;
Andre Noll58c0fed2009-03-31 14:33:13 +11005524 /* may be over-ridden by personality */
5525 mddev->resync_max_sectors = mddev->dev_sectors;
5526
NeilBrown6ff8d8ec2006-01-06 00:20:15 -08005527 mddev->ok_start_degraded = start_dirty_degraded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005528
NeilBrown0f9552b2009-12-30 12:08:50 +11005529 if (start_readonly && mddev->ro == 0)
NeilBrownf91de922005-11-08 21:39:36 -08005530 mddev->ro = 2; /* read-only, but switch on first write */
5531
Song Liua85dd7b2017-01-23 17:12:57 -08005532 /*
5533 * NOTE: some pers->run(), for example r5l_recovery_log(), wakes
5534 * up mddev->thread. It is important to initialize critical
5535 * resources for mddev->thread BEFORE calling pers->run().
5536 */
NeilBrown36d091f2014-12-15 12:56:58 +11005537 err = pers->run(mddev);
Andre Noll13e53df2008-03-26 00:07:03 +01005538 if (err)
NeilBrown9d487392016-11-02 14:16:49 +11005539 pr_warn("md: pers->run() failed ...\n");
NeilBrown36d091f2014-12-15 12:56:58 +11005540 else if (pers->size(mddev, 0, 0) < mddev->array_sectors) {
NeilBrown9d487392016-11-02 14:16:49 +11005541 WARN_ONCE(!mddev->external_size,
5542 "%s: default size too small, but 'external_size' not in effect?\n",
5543 __func__);
5544 pr_warn("md: invalid array_size %llu > default size %llu\n",
5545 (unsigned long long)mddev->array_sectors / 2,
5546 (unsigned long long)pers->size(mddev, 0, 0) / 2);
Dan Williamsb522adc2009-03-31 15:00:31 +11005547 err = -EINVAL;
Dan Williamsb522adc2009-03-31 15:00:31 +11005548 }
NeilBrown36d091f2014-12-15 12:56:58 +11005549 if (err == 0 && pers->sync_request &&
NeilBrownef99bf42012-05-22 13:55:08 +10005550 (mddev->bitmap_info.file || mddev->bitmap_info.offset)) {
Goldwyn Rodriguesf9209a32014-06-06 12:43:49 -05005551 struct bitmap *bitmap;
5552
5553 bitmap = bitmap_create(mddev, -1);
5554 if (IS_ERR(bitmap)) {
5555 err = PTR_ERR(bitmap);
NeilBrown9d487392016-11-02 14:16:49 +11005556 pr_warn("%s: failed to create bitmap (%d)\n",
5557 mdname(mddev), err);
Goldwyn Rodriguesf9209a32014-06-06 12:43:49 -05005558 } else
5559 mddev->bitmap = bitmap;
5560
NeilBrownb15c2e52006-01-06 00:20:16 -08005561 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005562 if (err) {
NeilBrown5aa61f42014-12-15 12:56:57 +11005563 mddev_detach(mddev);
NeilBrown0c35bd42015-03-13 11:51:18 +11005564 if (mddev->private)
5565 pers->free(mddev, mddev->private);
NeilBrownbd691922015-06-25 17:01:40 +10005566 mddev->private = NULL;
NeilBrown36d091f2014-12-15 12:56:58 +11005567 module_put(pers->owner);
NeilBrown32a76272005-06-21 17:17:14 -07005568 bitmap_destroy(mddev);
5569 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005570 }
NeilBrown5c675f82014-12-15 12:56:56 +11005571 if (mddev->queue) {
Shaohua Libb086a82016-09-30 09:45:40 -07005572 bool nonrot = true;
5573
5574 rdev_for_each(rdev, mddev) {
5575 if (rdev->raid_disk >= 0 &&
5576 !blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
5577 nonrot = false;
5578 break;
5579 }
5580 }
5581 if (mddev->degraded)
5582 nonrot = false;
5583 if (nonrot)
5584 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, mddev->queue);
5585 else
5586 queue_flag_clear_unlocked(QUEUE_FLAG_NONROT, mddev->queue);
Jan Karadc3b17c2017-02-02 15:56:50 +01005587 mddev->queue->backing_dev_info->congested_data = mddev;
5588 mddev->queue->backing_dev_info->congested_fn = md_congested;
NeilBrown5c675f82014-12-15 12:56:56 +11005589 }
NeilBrown36d091f2014-12-15 12:56:58 +11005590 if (pers->sync_request) {
NeilBrown00bcb4a2010-06-01 19:37:23 +10005591 if (mddev->kobj.sd &&
5592 sysfs_create_group(&mddev->kobj, &md_redundancy_group))
NeilBrown9d487392016-11-02 14:16:49 +11005593 pr_warn("md: cannot register extra attributes for %s\n",
5594 mdname(mddev));
NeilBrown00bcb4a2010-06-01 19:37:23 +10005595 mddev->sysfs_action = sysfs_get_dirent_safe(mddev->kobj.sd, "sync_action");
NeilBrown5e55e2f2007-03-26 21:32:14 -08005596 } else if (mddev->ro == 2) /* auto-readonly not meaningful */
NeilBrownfd9d49c2005-11-08 21:39:42 -08005597 mddev->ro = 0;
5598
Robert Becker1e509152009-12-14 12:49:58 +11005599 atomic_set(&mddev->max_corr_read_errors,
5600 MD_DEFAULT_MAX_CORRECTED_READ_ERRORS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005601 mddev->safemode = 0;
Goldwyn Rodrigues28c1b9f2015-10-22 16:01:25 +11005602 if (mddev_is_clustered(mddev))
5603 mddev->safemode_delay = 0;
5604 else
5605 mddev->safemode_delay = (200 * HZ)/1000 +1; /* 200 msec delay */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005606 mddev->in_sync = 1;
NeilBrown0ca69882011-01-14 09:14:33 +11005607 smp_wmb();
NeilBrown36d091f2014-12-15 12:56:58 +11005608 spin_lock(&mddev->lock);
5609 mddev->pers = pers;
NeilBrown36d091f2014-12-15 12:56:58 +11005610 spin_unlock(&mddev->lock);
NeilBrowndafb20f2012-03-19 12:46:39 +11005611 rdev_for_each(rdev, mddev)
Namhyung Kim36fad852011-07-27 11:00:36 +10005612 if (rdev->raid_disk >= 0)
5613 if (sysfs_link_rdev(mddev, rdev))
NeilBrown00bcb4a2010-06-01 19:37:23 +10005614 /* failure here is OK */;
NeilBrownf72ffdd2014-09-30 14:23:59 +10005615
NeilBrowna4a3d262015-07-17 11:57:30 +10005616 if (mddev->degraded && !mddev->ro)
5617 /* This ensures that recovering status is reported immediately
5618 * via sysfs - until a lack of spares is confirmed.
5619 */
5620 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005621 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
NeilBrownf72ffdd2014-09-30 14:23:59 +10005622
Shaohua Li29530792016-12-08 15:48:19 -08005623 if (mddev->sb_flags)
NeilBrown850b2b42006-10-03 01:15:46 -07005624 md_update_sb(mddev, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005625
NeilBrownd7603b72006-01-06 00:20:30 -08005626 md_new_event(mddev);
NeilBrown00bcb4a2010-06-01 19:37:23 +10005627 sysfs_notify_dirent_safe(mddev->sysfs_state);
5628 sysfs_notify_dirent_safe(mddev->sysfs_action);
Neil Browna99ac972008-06-28 08:31:43 +10005629 sysfs_notify(&mddev->kobj, NULL, "degraded");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005630 return 0;
5631}
NeilBrown390ee602010-06-01 19:37:27 +10005632EXPORT_SYMBOL_GPL(md_run);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005633
NeilBrownfd01b882011-10-11 16:47:53 +11005634static int do_md_run(struct mddev *mddev)
NeilBrownfe60b012010-03-29 11:10:42 +11005635{
5636 int err;
5637
5638 err = md_run(mddev);
5639 if (err)
5640 goto out;
NeilBrown69e51b42010-06-01 19:37:35 +10005641 err = bitmap_load(mddev);
5642 if (err) {
5643 bitmap_destroy(mddev);
5644 goto out;
5645 }
Jonathan Brassow0fd018a2011-06-07 17:49:36 -05005646
Goldwyn Rodrigues28c1b9f2015-10-22 16:01:25 +11005647 if (mddev_is_clustered(mddev))
5648 md_allow_write(mddev);
5649
Jonathan Brassow0fd018a2011-06-07 17:49:36 -05005650 md_wakeup_thread(mddev->thread);
5651 md_wakeup_thread(mddev->sync_thread); /* possibly kick off a reshape */
5652
NeilBrownfe60b012010-03-29 11:10:42 +11005653 set_capacity(mddev->gendisk, mddev->array_sectors);
5654 revalidate_disk(mddev->gendisk);
NeilBrownf0b4f7e2011-02-24 17:26:41 +11005655 mddev->changed = 1;
NeilBrownfe60b012010-03-29 11:10:42 +11005656 kobject_uevent(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE);
5657out:
5658 return err;
5659}
5660
NeilBrownfd01b882011-10-11 16:47:53 +11005661static int restart_array(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005662{
5663 struct gendisk *disk = mddev->gendisk;
NeilBrown97b20ef2017-04-13 08:53:48 +10005664 struct md_rdev *rdev;
5665 bool has_journal = false;
5666 bool has_readonly = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005667
Andre Noll80fab1d2008-07-11 22:02:21 +10005668 /* Complain if it has no devices */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005669 if (list_empty(&mddev->disks))
Andre Noll80fab1d2008-07-11 22:02:21 +10005670 return -ENXIO;
5671 if (!mddev->pers)
5672 return -EINVAL;
5673 if (!mddev->ro)
5674 return -EBUSY;
Song Liu339421d2015-10-08 21:54:13 -07005675
NeilBrown97b20ef2017-04-13 08:53:48 +10005676 rcu_read_lock();
5677 rdev_for_each_rcu(rdev, mddev) {
5678 if (test_bit(Journal, &rdev->flags) &&
5679 !test_bit(Faulty, &rdev->flags))
5680 has_journal = true;
5681 if (bdev_read_only(rdev->bdev))
5682 has_readonly = true;
Song Liu339421d2015-10-08 21:54:13 -07005683 }
NeilBrown97b20ef2017-04-13 08:53:48 +10005684 rcu_read_unlock();
5685 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) && !has_journal)
5686 /* Don't restart rw with journal missing/faulty */
5687 return -EINVAL;
5688 if (has_readonly)
5689 return -EROFS;
Song Liu339421d2015-10-08 21:54:13 -07005690
Andre Noll80fab1d2008-07-11 22:02:21 +10005691 mddev->safemode = 0;
5692 mddev->ro = 0;
5693 set_disk_ro(disk, 0);
NeilBrown9d487392016-11-02 14:16:49 +11005694 pr_debug("md: %s switched to read-write mode.\n", mdname(mddev));
Andre Noll80fab1d2008-07-11 22:02:21 +10005695 /* Kick recovery or resync if necessary */
5696 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5697 md_wakeup_thread(mddev->thread);
5698 md_wakeup_thread(mddev->sync_thread);
NeilBrown00bcb4a2010-06-01 19:37:23 +10005699 sysfs_notify_dirent_safe(mddev->sysfs_state);
Andre Noll80fab1d2008-07-11 22:02:21 +10005700 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005701}
5702
NeilBrownfd01b882011-10-11 16:47:53 +11005703static void md_clean(struct mddev *mddev)
NeilBrown6177b472010-03-29 11:37:13 +11005704{
5705 mddev->array_sectors = 0;
5706 mddev->external_size = 0;
5707 mddev->dev_sectors = 0;
5708 mddev->raid_disks = 0;
5709 mddev->recovery_cp = 0;
5710 mddev->resync_min = 0;
5711 mddev->resync_max = MaxSector;
5712 mddev->reshape_position = MaxSector;
5713 mddev->external = 0;
5714 mddev->persistent = 0;
5715 mddev->level = LEVEL_NONE;
5716 mddev->clevel[0] = 0;
5717 mddev->flags = 0;
Shaohua Li29530792016-12-08 15:48:19 -08005718 mddev->sb_flags = 0;
NeilBrown6177b472010-03-29 11:37:13 +11005719 mddev->ro = 0;
5720 mddev->metadata_type[0] = 0;
5721 mddev->chunk_sectors = 0;
5722 mddev->ctime = mddev->utime = 0;
5723 mddev->layout = 0;
5724 mddev->max_disks = 0;
5725 mddev->events = 0;
NeilBrowna8707c02010-05-18 09:28:43 +10005726 mddev->can_decrease_events = 0;
NeilBrown6177b472010-03-29 11:37:13 +11005727 mddev->delta_disks = 0;
NeilBrown2c810cd2012-05-21 09:27:00 +10005728 mddev->reshape_backwards = 0;
NeilBrown6177b472010-03-29 11:37:13 +11005729 mddev->new_level = LEVEL_NONE;
5730 mddev->new_layout = 0;
5731 mddev->new_chunk_sectors = 0;
5732 mddev->curr_resync = 0;
Jianpeng Ma7f7583d2012-10-11 14:17:59 +11005733 atomic64_set(&mddev->resync_mismatches, 0);
NeilBrown6177b472010-03-29 11:37:13 +11005734 mddev->suspend_lo = mddev->suspend_hi = 0;
5735 mddev->sync_speed_min = mddev->sync_speed_max = 0;
5736 mddev->recovery = 0;
5737 mddev->in_sync = 0;
NeilBrownf0b4f7e2011-02-24 17:26:41 +11005738 mddev->changed = 0;
NeilBrown6177b472010-03-29 11:37:13 +11005739 mddev->degraded = 0;
NeilBrown6177b472010-03-29 11:37:13 +11005740 mddev->safemode = 0;
NeilBrownbd691922015-06-25 17:01:40 +10005741 mddev->private = NULL;
Guoqing Jiangc20c33f2016-08-12 13:42:38 +08005742 mddev->cluster_info = NULL;
NeilBrown6177b472010-03-29 11:37:13 +11005743 mddev->bitmap_info.offset = 0;
5744 mddev->bitmap_info.default_offset = 0;
NeilBrown6409bb02012-05-22 13:55:07 +10005745 mddev->bitmap_info.default_space = 0;
NeilBrown6177b472010-03-29 11:37:13 +11005746 mddev->bitmap_info.chunksize = 0;
5747 mddev->bitmap_info.daemon_sleep = 0;
5748 mddev->bitmap_info.max_write_behind = 0;
Guoqing Jiangc20c33f2016-08-12 13:42:38 +08005749 mddev->bitmap_info.nodes = 0;
NeilBrown6177b472010-03-29 11:37:13 +11005750}
5751
NeilBrownfd01b882011-10-11 16:47:53 +11005752static void __md_stop_writes(struct mddev *mddev)
NeilBrowna047e122010-03-29 12:07:53 +11005753{
NeilBrown6b6204e2013-05-09 09:48:30 +10005754 set_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
NeilBrownf851b602014-12-11 10:02:10 +11005755 flush_workqueue(md_misc_wq);
NeilBrowna047e122010-03-29 12:07:53 +11005756 if (mddev->sync_thread) {
NeilBrowna047e122010-03-29 12:07:53 +11005757 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
Jonathan Brassowa91d5ac2013-04-24 11:42:43 +10005758 md_reap_sync_thread(mddev);
NeilBrowna047e122010-03-29 12:07:53 +11005759 }
5760
5761 del_timer_sync(&mddev->safemode_timer);
5762
Shaohua Li034e33f2016-11-21 10:29:19 -08005763 if (mddev->pers && mddev->pers->quiesce) {
5764 mddev->pers->quiesce(mddev, 1);
5765 mddev->pers->quiesce(mddev, 0);
5766 }
NeilBrowna047e122010-03-29 12:07:53 +11005767 bitmap_flush(mddev);
NeilBrowna047e122010-03-29 12:07:53 +11005768
NeilBrownb6d428c2013-04-24 11:42:42 +10005769 if (mddev->ro == 0 &&
Goldwyn Rodrigues28c1b9f2015-10-22 16:01:25 +11005770 ((!mddev->in_sync && !mddev_is_clustered(mddev)) ||
Shaohua Li29530792016-12-08 15:48:19 -08005771 mddev->sb_flags)) {
NeilBrowna047e122010-03-29 12:07:53 +11005772 /* mark array as shutdown cleanly */
Goldwyn Rodrigues28c1b9f2015-10-22 16:01:25 +11005773 if (!mddev_is_clustered(mddev))
5774 mddev->in_sync = 1;
NeilBrowna047e122010-03-29 12:07:53 +11005775 md_update_sb(mddev, 1);
5776 }
5777}
NeilBrowndefad612011-01-14 09:14:33 +11005778
NeilBrownfd01b882011-10-11 16:47:53 +11005779void md_stop_writes(struct mddev *mddev)
NeilBrowndefad612011-01-14 09:14:33 +11005780{
NeilBrown29f097c2013-11-14 17:54:51 +11005781 mddev_lock_nointr(mddev);
NeilBrowndefad612011-01-14 09:14:33 +11005782 __md_stop_writes(mddev);
5783 mddev_unlock(mddev);
5784}
NeilBrown390ee602010-06-01 19:37:27 +10005785EXPORT_SYMBOL_GPL(md_stop_writes);
NeilBrowna047e122010-03-29 12:07:53 +11005786
NeilBrown5aa61f42014-12-15 12:56:57 +11005787static void mddev_detach(struct mddev *mddev)
5788{
Guoqing Jiang48df4982017-03-14 09:40:20 +08005789 bitmap_wait_behind_writes(mddev);
NeilBrown36d091f2014-12-15 12:56:58 +11005790 if (mddev->pers && mddev->pers->quiesce) {
NeilBrown5aa61f42014-12-15 12:56:57 +11005791 mddev->pers->quiesce(mddev, 1);
5792 mddev->pers->quiesce(mddev, 0);
5793 }
5794 md_unregister_thread(&mddev->thread);
5795 if (mddev->queue)
5796 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
5797}
5798
NeilBrown5eff3c42012-11-19 10:47:48 +11005799static void __md_stop(struct mddev *mddev)
NeilBrown6177b472010-03-29 11:37:13 +11005800{
NeilBrown36d091f2014-12-15 12:56:58 +11005801 struct md_personality *pers = mddev->pers;
Guoqing Jiang48df4982017-03-14 09:40:20 +08005802 bitmap_destroy(mddev);
NeilBrown5aa61f42014-12-15 12:56:57 +11005803 mddev_detach(mddev);
NeilBrownee5d0042015-07-22 10:20:07 +10005804 /* Ensure ->event_work is done */
5805 flush_workqueue(md_misc_wq);
NeilBrown36d091f2014-12-15 12:56:58 +11005806 spin_lock(&mddev->lock);
NeilBrown6177b472010-03-29 11:37:13 +11005807 mddev->pers = NULL;
NeilBrown36d091f2014-12-15 12:56:58 +11005808 spin_unlock(&mddev->lock);
5809 pers->free(mddev, mddev->private);
NeilBrownbd691922015-06-25 17:01:40 +10005810 mddev->private = NULL;
NeilBrown36d091f2014-12-15 12:56:58 +11005811 if (pers->sync_request && mddev->to_remove == NULL)
5812 mddev->to_remove = &md_redundancy_group;
5813 module_put(pers->owner);
NeilBrowncca9cf92010-04-01 12:08:16 +11005814 clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
NeilBrown6177b472010-03-29 11:37:13 +11005815}
NeilBrown5eff3c42012-11-19 10:47:48 +11005816
5817void md_stop(struct mddev *mddev)
5818{
5819 /* stop the array and free an attached data structures.
5820 * This is called from dm-raid
5821 */
5822 __md_stop(mddev);
NeilBrown5eff3c42012-11-19 10:47:48 +11005823 if (mddev->bio_set)
5824 bioset_free(mddev->bio_set);
5825}
5826
NeilBrown390ee602010-06-01 19:37:27 +10005827EXPORT_SYMBOL_GPL(md_stop);
NeilBrown6177b472010-03-29 11:37:13 +11005828
NeilBrowna05b7ea2012-07-19 15:59:18 +10005829static int md_set_readonly(struct mddev *mddev, struct block_device *bdev)
NeilBrowna4bd82d2010-03-29 13:23:10 +11005830{
5831 int err = 0;
NeilBrown30b8feb2013-11-14 15:16:17 +11005832 int did_freeze = 0;
5833
5834 if (!test_bit(MD_RECOVERY_FROZEN, &mddev->recovery)) {
5835 did_freeze = 1;
5836 set_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
5837 md_wakeup_thread(mddev->thread);
5838 }
NeilBrownf851b602014-12-11 10:02:10 +11005839 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
NeilBrown30b8feb2013-11-14 15:16:17 +11005840 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
NeilBrownf851b602014-12-11 10:02:10 +11005841 if (mddev->sync_thread)
NeilBrown30b8feb2013-11-14 15:16:17 +11005842 /* Thread might be blocked waiting for metadata update
5843 * which will now never happen */
5844 wake_up_process(mddev->sync_thread->tsk);
NeilBrownf851b602014-12-11 10:02:10 +11005845
Shaohua Li29530792016-12-08 15:48:19 -08005846 if (mddev->external && test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
NeilBrown88724bf2015-09-24 14:00:51 +10005847 return -EBUSY;
NeilBrown30b8feb2013-11-14 15:16:17 +11005848 mddev_unlock(mddev);
NeilBrownf851b602014-12-11 10:02:10 +11005849 wait_event(resync_wait, !test_bit(MD_RECOVERY_RUNNING,
5850 &mddev->recovery));
NeilBrown88724bf2015-09-24 14:00:51 +10005851 wait_event(mddev->sb_wait,
Shaohua Li29530792016-12-08 15:48:19 -08005852 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
NeilBrown30b8feb2013-11-14 15:16:17 +11005853 mddev_lock_nointr(mddev);
5854
NeilBrowna4bd82d2010-03-29 13:23:10 +11005855 mutex_lock(&mddev->open_mutex);
NeilBrown9ba3b7f2014-09-09 14:00:15 +10005856 if ((mddev->pers && atomic_read(&mddev->openers) > !!bdev) ||
NeilBrown30b8feb2013-11-14 15:16:17 +11005857 mddev->sync_thread ||
Guoqing Jiangaf8d8e62016-08-12 13:42:37 +08005858 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery)) {
NeilBrown9d487392016-11-02 14:16:49 +11005859 pr_warn("md: %s still in use.\n",mdname(mddev));
NeilBrown30b8feb2013-11-14 15:16:17 +11005860 if (did_freeze) {
5861 clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
NeilBrown45eaf452014-10-29 08:49:50 +11005862 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
NeilBrown30b8feb2013-11-14 15:16:17 +11005863 md_wakeup_thread(mddev->thread);
5864 }
NeilBrowna4bd82d2010-03-29 13:23:10 +11005865 err = -EBUSY;
5866 goto out;
5867 }
5868 if (mddev->pers) {
NeilBrowndefad612011-01-14 09:14:33 +11005869 __md_stop_writes(mddev);
NeilBrowna4bd82d2010-03-29 13:23:10 +11005870
5871 err = -ENXIO;
5872 if (mddev->ro==1)
5873 goto out;
5874 mddev->ro = 1;
5875 set_disk_ro(mddev->gendisk, 1);
5876 clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
NeilBrown45eaf452014-10-29 08:49:50 +11005877 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5878 md_wakeup_thread(mddev->thread);
NeilBrown00bcb4a2010-06-01 19:37:23 +10005879 sysfs_notify_dirent_safe(mddev->sysfs_state);
NeilBrown30b8feb2013-11-14 15:16:17 +11005880 err = 0;
NeilBrowna4bd82d2010-03-29 13:23:10 +11005881 }
5882out:
5883 mutex_unlock(&mddev->open_mutex);
5884 return err;
5885}
5886
NeilBrown9e653b62006-06-26 00:27:58 -07005887/* mode:
5888 * 0 - completely stop and dis-assemble array
NeilBrown9e653b62006-06-26 00:27:58 -07005889 * 2 - stop but do not disassemble array
5890 */
NeilBrownf72ffdd2014-09-30 14:23:59 +10005891static int do_md_stop(struct mddev *mddev, int mode,
NeilBrowna05b7ea2012-07-19 15:59:18 +10005892 struct block_device *bdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005893{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005894 struct gendisk *disk = mddev->gendisk;
NeilBrown3cb03002011-10-11 16:45:26 +11005895 struct md_rdev *rdev;
NeilBrown30b8feb2013-11-14 15:16:17 +11005896 int did_freeze = 0;
5897
5898 if (!test_bit(MD_RECOVERY_FROZEN, &mddev->recovery)) {
5899 did_freeze = 1;
5900 set_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
5901 md_wakeup_thread(mddev->thread);
5902 }
NeilBrownf851b602014-12-11 10:02:10 +11005903 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
NeilBrown30b8feb2013-11-14 15:16:17 +11005904 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
NeilBrownf851b602014-12-11 10:02:10 +11005905 if (mddev->sync_thread)
NeilBrown30b8feb2013-11-14 15:16:17 +11005906 /* Thread might be blocked waiting for metadata update
5907 * which will now never happen */
5908 wake_up_process(mddev->sync_thread->tsk);
NeilBrownf851b602014-12-11 10:02:10 +11005909
NeilBrown30b8feb2013-11-14 15:16:17 +11005910 mddev_unlock(mddev);
NeilBrownf851b602014-12-11 10:02:10 +11005911 wait_event(resync_wait, (mddev->sync_thread == NULL &&
5912 !test_bit(MD_RECOVERY_RUNNING,
5913 &mddev->recovery)));
NeilBrown30b8feb2013-11-14 15:16:17 +11005914 mddev_lock_nointr(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005915
NeilBrownc8c00a62009-08-10 12:50:52 +10005916 mutex_lock(&mddev->open_mutex);
NeilBrown9ba3b7f2014-09-09 14:00:15 +10005917 if ((mddev->pers && atomic_read(&mddev->openers) > !!bdev) ||
NeilBrown30b8feb2013-11-14 15:16:17 +11005918 mddev->sysfs_active ||
5919 mddev->sync_thread ||
Guoqing Jiangaf8d8e62016-08-12 13:42:37 +08005920 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery)) {
NeilBrown9d487392016-11-02 14:16:49 +11005921 pr_warn("md: %s still in use.\n",mdname(mddev));
NeilBrown6e17b022010-08-07 21:41:19 +10005922 mutex_unlock(&mddev->open_mutex);
NeilBrown30b8feb2013-11-14 15:16:17 +11005923 if (did_freeze) {
5924 clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
NeilBrown45eaf452014-10-29 08:49:50 +11005925 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
NeilBrown30b8feb2013-11-14 15:16:17 +11005926 md_wakeup_thread(mddev->thread);
5927 }
NeilBrown260fa032013-08-27 16:44:13 +10005928 return -EBUSY;
5929 }
NeilBrown6e17b022010-08-07 21:41:19 +10005930 if (mddev->pers) {
NeilBrowna4bd82d2010-03-29 13:23:10 +11005931 if (mddev->ro)
5932 set_disk_ro(disk, 0);
NeilBrown409c57f2009-03-31 14:39:39 +11005933
NeilBrowndefad612011-01-14 09:14:33 +11005934 __md_stop_writes(mddev);
NeilBrown5eff3c42012-11-19 10:47:48 +11005935 __md_stop(mddev);
Jan Karadc3b17c2017-02-02 15:56:50 +01005936 mddev->queue->backing_dev_info->congested_fn = NULL;
NeilBrown6177b472010-03-29 11:37:13 +11005937
NeilBrowna4bd82d2010-03-29 13:23:10 +11005938 /* tell userspace to handle 'inactive' */
NeilBrown00bcb4a2010-06-01 19:37:23 +10005939 sysfs_notify_dirent_safe(mddev->sysfs_state);
NeilBrown0d4ca602006-12-10 02:20:44 -08005940
NeilBrowndafb20f2012-03-19 12:46:39 +11005941 rdev_for_each(rdev, mddev)
Namhyung Kim36fad852011-07-27 11:00:36 +10005942 if (rdev->raid_disk >= 0)
5943 sysfs_unlink_rdev(mddev, rdev);
NeilBrownc4647292009-05-07 12:51:06 +10005944
NeilBrowna4bd82d2010-03-29 13:23:10 +11005945 set_capacity(disk, 0);
NeilBrown6e17b022010-08-07 21:41:19 +10005946 mutex_unlock(&mddev->open_mutex);
NeilBrownf0b4f7e2011-02-24 17:26:41 +11005947 mddev->changed = 1;
NeilBrowna4bd82d2010-03-29 13:23:10 +11005948 revalidate_disk(disk);
NeilBrown0d4ca602006-12-10 02:20:44 -08005949
NeilBrowna4bd82d2010-03-29 13:23:10 +11005950 if (mddev->ro)
5951 mddev->ro = 0;
NeilBrown6e17b022010-08-07 21:41:19 +10005952 } else
5953 mutex_unlock(&mddev->open_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005954 /*
5955 * Free resources if final stop
5956 */
NeilBrown9e653b62006-06-26 00:27:58 -07005957 if (mode == 0) {
NeilBrown9d487392016-11-02 14:16:49 +11005958 pr_info("md: %s stopped.\n", mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005959
NeilBrownc3d97142009-12-14 12:49:52 +11005960 if (mddev->bitmap_info.file) {
NeilBrown4af1a042014-12-15 12:57:00 +11005961 struct file *f = mddev->bitmap_info.file;
5962 spin_lock(&mddev->lock);
NeilBrownc3d97142009-12-14 12:49:52 +11005963 mddev->bitmap_info.file = NULL;
NeilBrown4af1a042014-12-15 12:57:00 +11005964 spin_unlock(&mddev->lock);
5965 fput(f);
NeilBrown978f9462006-02-02 14:28:05 -08005966 }
NeilBrownc3d97142009-12-14 12:49:52 +11005967 mddev->bitmap_info.offset = 0;
NeilBrown978f9462006-02-02 14:28:05 -08005968
Linus Torvalds1da177e2005-04-16 15:20:36 -07005969 export_array(mddev);
5970
NeilBrown6177b472010-03-29 11:37:13 +11005971 md_clean(mddev);
NeilBrownefeb53c2009-01-09 08:31:10 +11005972 if (mddev->hold_active == UNTIL_STOP)
5973 mddev->hold_active = 0;
NeilBrowna4bd82d2010-03-29 13:23:10 +11005974 }
NeilBrownd7603b72006-01-06 00:20:30 -08005975 md_new_event(mddev);
NeilBrown00bcb4a2010-06-01 19:37:23 +10005976 sysfs_notify_dirent_safe(mddev->sysfs_state);
NeilBrown6e17b022010-08-07 21:41:19 +10005977 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005978}
5979
Jeff Garzikfdee8ae2006-12-10 02:20:50 -08005980#ifndef MODULE
NeilBrownfd01b882011-10-11 16:47:53 +11005981static void autorun_array(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005982{
NeilBrown3cb03002011-10-11 16:45:26 +11005983 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005984 int err;
5985
NeilBrowna757e642005-04-16 15:26:42 -07005986 if (list_empty(&mddev->disks))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005987 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005988
NeilBrown9d487392016-11-02 14:16:49 +11005989 pr_info("md: running: ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005990
NeilBrowndafb20f2012-03-19 12:46:39 +11005991 rdev_for_each(rdev, mddev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005992 char b[BDEVNAME_SIZE];
NeilBrown9d487392016-11-02 14:16:49 +11005993 pr_cont("<%s>", bdevname(rdev->bdev,b));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005994 }
NeilBrown9d487392016-11-02 14:16:49 +11005995 pr_cont("\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005996
NeilBrownd710e132008-10-13 11:55:12 +11005997 err = do_md_run(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005998 if (err) {
NeilBrown9d487392016-11-02 14:16:49 +11005999 pr_warn("md: do_md_run() returned %d\n", err);
NeilBrowna05b7ea2012-07-19 15:59:18 +10006000 do_md_stop(mddev, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006001 }
6002}
6003
6004/*
6005 * lets try to run arrays based on all disks that have arrived
6006 * until now. (those are in pending_raid_disks)
6007 *
6008 * the method: pick the first pending disk, collect all disks with
6009 * the same UUID, remove all from the pending list and put them into
6010 * the 'same_array' list. Then order this list based on superblock
6011 * update time (freshest comes first), kick out 'old' disks and
6012 * compare superblocks. If everything's fine then run it.
6013 *
6014 * If "unit" is allocated, then bump its reference count
6015 */
6016static void autorun_devices(int part)
6017{
NeilBrown3cb03002011-10-11 16:45:26 +11006018 struct md_rdev *rdev0, *rdev, *tmp;
NeilBrownfd01b882011-10-11 16:47:53 +11006019 struct mddev *mddev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006020 char b[BDEVNAME_SIZE];
6021
NeilBrown9d487392016-11-02 14:16:49 +11006022 pr_info("md: autorun ...\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006023 while (!list_empty(&pending_raid_disks)) {
NeilBrowne8703fe2006-10-03 01:15:59 -07006024 int unit;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006025 dev_t dev;
NeilBrownad01c9e2006-03-27 01:18:07 -08006026 LIST_HEAD(candidates);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006027 rdev0 = list_entry(pending_raid_disks.next,
NeilBrown3cb03002011-10-11 16:45:26 +11006028 struct md_rdev, same_set);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006029
NeilBrown9d487392016-11-02 14:16:49 +11006030 pr_debug("md: considering %s ...\n", bdevname(rdev0->bdev,b));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006031 INIT_LIST_HEAD(&candidates);
Cheng Renquan159ec1f2009-01-09 08:31:08 +11006032 rdev_for_each_list(rdev, tmp, &pending_raid_disks)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006033 if (super_90_load(rdev, rdev0, 0) >= 0) {
NeilBrown9d487392016-11-02 14:16:49 +11006034 pr_debug("md: adding %s ...\n",
6035 bdevname(rdev->bdev,b));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006036 list_move(&rdev->same_set, &candidates);
6037 }
6038 /*
6039 * now we have a set of devices, with all of them having
6040 * mostly sane superblocks. It's time to allocate the
6041 * mddev.
6042 */
NeilBrowne8703fe2006-10-03 01:15:59 -07006043 if (part) {
6044 dev = MKDEV(mdp_major,
6045 rdev0->preferred_minor << MdpMinorShift);
6046 unit = MINOR(dev) >> MdpMinorShift;
6047 } else {
6048 dev = MKDEV(MD_MAJOR, rdev0->preferred_minor);
6049 unit = MINOR(dev);
6050 }
6051 if (rdev0->preferred_minor != unit) {
NeilBrown9d487392016-11-02 14:16:49 +11006052 pr_warn("md: unit number in %s is bad: %d\n",
6053 bdevname(rdev0->bdev, b), rdev0->preferred_minor);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006054 break;
6055 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006056
6057 md_probe(dev, NULL, NULL);
6058 mddev = mddev_find(dev);
Neil Brown9bbbca32008-06-28 08:31:17 +10006059 if (!mddev || !mddev->gendisk) {
6060 if (mddev)
6061 mddev_put(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006062 break;
6063 }
NeilBrownf72ffdd2014-09-30 14:23:59 +10006064 if (mddev_lock(mddev))
NeilBrown9d487392016-11-02 14:16:49 +11006065 pr_warn("md: %s locked, cannot run\n", mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006066 else if (mddev->raid_disks || mddev->major_version
6067 || !list_empty(&mddev->disks)) {
NeilBrown9d487392016-11-02 14:16:49 +11006068 pr_warn("md: %s already running, cannot run %s\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07006069 mdname(mddev), bdevname(rdev0->bdev,b));
6070 mddev_unlock(mddev);
6071 } else {
NeilBrown9d487392016-11-02 14:16:49 +11006072 pr_debug("md: created %s\n", mdname(mddev));
NeilBrown1ec4a932008-02-06 01:39:53 -08006073 mddev->persistent = 1;
Cheng Renquan159ec1f2009-01-09 08:31:08 +11006074 rdev_for_each_list(rdev, tmp, &candidates) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006075 list_del_init(&rdev->same_set);
6076 if (bind_rdev_to_array(rdev, mddev))
6077 export_rdev(rdev);
6078 }
6079 autorun_array(mddev);
6080 mddev_unlock(mddev);
6081 }
6082 /* on success, candidates will be empty, on error
6083 * it won't...
6084 */
Cheng Renquan159ec1f2009-01-09 08:31:08 +11006085 rdev_for_each_list(rdev, tmp, &candidates) {
NeilBrown4b809912008-07-21 17:05:25 +10006086 list_del_init(&rdev->same_set);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006087 export_rdev(rdev);
NeilBrown4b809912008-07-21 17:05:25 +10006088 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006089 mddev_put(mddev);
6090 }
NeilBrown9d487392016-11-02 14:16:49 +11006091 pr_info("md: ... autorun DONE.\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006092}
Jeff Garzikfdee8ae2006-12-10 02:20:50 -08006093#endif /* !MODULE */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006094
NeilBrownf72ffdd2014-09-30 14:23:59 +10006095static int get_version(void __user *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006096{
6097 mdu_version_t ver;
6098
6099 ver.major = MD_MAJOR_VERSION;
6100 ver.minor = MD_MINOR_VERSION;
6101 ver.patchlevel = MD_PATCHLEVEL_VERSION;
6102
6103 if (copy_to_user(arg, &ver, sizeof(ver)))
6104 return -EFAULT;
6105
6106 return 0;
6107}
6108
NeilBrownf72ffdd2014-09-30 14:23:59 +10006109static int get_array_info(struct mddev *mddev, void __user *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006110{
6111 mdu_array_info_t info;
NeilBrowna9f326e2009-09-23 18:06:41 +10006112 int nr,working,insync,failed,spare;
NeilBrown3cb03002011-10-11 16:45:26 +11006113 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006114
NeilBrown1ca69c42012-10-11 13:37:33 +11006115 nr = working = insync = failed = spare = 0;
6116 rcu_read_lock();
6117 rdev_for_each_rcu(rdev, mddev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006118 nr++;
NeilBrownb2d444d2005-11-08 21:39:31 -08006119 if (test_bit(Faulty, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006120 failed++;
6121 else {
6122 working++;
NeilBrownb2d444d2005-11-08 21:39:31 -08006123 if (test_bit(In_sync, &rdev->flags))
NeilBrownf72ffdd2014-09-30 14:23:59 +10006124 insync++;
Song Liub347af82016-08-11 17:14:45 -07006125 else if (test_bit(Journal, &rdev->flags))
6126 /* TODO: add journal count to md_u.h */
6127 ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006128 else
6129 spare++;
6130 }
6131 }
NeilBrown1ca69c42012-10-11 13:37:33 +11006132 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006133
6134 info.major_version = mddev->major_version;
6135 info.minor_version = mddev->minor_version;
6136 info.patch_version = MD_PATCHLEVEL_VERSION;
Deepa Dinamani9ebc6ef2015-12-21 10:51:01 +11006137 info.ctime = clamp_t(time64_t, mddev->ctime, 0, U32_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006138 info.level = mddev->level;
Andre Noll58c0fed2009-03-31 14:33:13 +11006139 info.size = mddev->dev_sectors / 2;
6140 if (info.size != mddev->dev_sectors / 2) /* overflow */
NeilBrown284ae7c2006-02-03 03:03:40 -08006141 info.size = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006142 info.nr_disks = nr;
6143 info.raid_disks = mddev->raid_disks;
6144 info.md_minor = mddev->md_minor;
6145 info.not_persistent= !mddev->persistent;
6146
Deepa Dinamani9ebc6ef2015-12-21 10:51:01 +11006147 info.utime = clamp_t(time64_t, mddev->utime, 0, U32_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006148 info.state = 0;
6149 if (mddev->in_sync)
6150 info.state = (1<<MD_SB_CLEAN);
NeilBrownc3d97142009-12-14 12:49:52 +11006151 if (mddev->bitmap && mddev->bitmap_info.offset)
NeilBrown9bd35922014-07-02 11:35:06 +10006152 info.state |= (1<<MD_SB_BITMAP_PRESENT);
Goldwyn Rodriguesca8895d2014-11-26 12:22:03 -06006153 if (mddev_is_clustered(mddev))
6154 info.state |= (1<<MD_SB_CLUSTERED);
NeilBrowna9f326e2009-09-23 18:06:41 +10006155 info.active_disks = insync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006156 info.working_disks = working;
6157 info.failed_disks = failed;
6158 info.spare_disks = spare;
6159
6160 info.layout = mddev->layout;
Andre Noll9d8f0362009-06-18 08:45:01 +10006161 info.chunk_size = mddev->chunk_sectors << 9;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006162
6163 if (copy_to_user(arg, &info, sizeof(info)))
6164 return -EFAULT;
6165
6166 return 0;
6167}
6168
NeilBrownf72ffdd2014-09-30 14:23:59 +10006169static int get_bitmap_file(struct mddev *mddev, void __user * arg)
NeilBrown32a76272005-06-21 17:17:14 -07006170{
6171 mdu_bitmap_file_t *file = NULL; /* too big for stack allocation */
NeilBrownf4ad3d32014-12-15 12:57:00 +11006172 char *ptr;
NeilBrown4af1a042014-12-15 12:57:00 +11006173 int err;
NeilBrown32a76272005-06-21 17:17:14 -07006174
Benjamin Randazzob6878d92015-07-25 16:36:50 +02006175 file = kzalloc(sizeof(*file), GFP_NOIO);
NeilBrown32a76272005-06-21 17:17:14 -07006176 if (!file)
NeilBrown4af1a042014-12-15 12:57:00 +11006177 return -ENOMEM;
NeilBrown32a76272005-06-21 17:17:14 -07006178
NeilBrown32a76272005-06-21 17:17:14 -07006179 err = 0;
NeilBrown4af1a042014-12-15 12:57:00 +11006180 spin_lock(&mddev->lock);
Benjamin Randazzo25eafe12015-07-25 16:36:50 +02006181 /* bitmap enabled */
6182 if (mddev->bitmap_info.file) {
6183 ptr = file_path(mddev->bitmap_info.file, file->pathname,
6184 sizeof(file->pathname));
6185 if (IS_ERR(ptr))
6186 err = PTR_ERR(ptr);
6187 else
6188 memmove(file->pathname, ptr,
6189 sizeof(file->pathname)-(ptr-file->pathname));
6190 }
NeilBrown4af1a042014-12-15 12:57:00 +11006191 spin_unlock(&mddev->lock);
6192
6193 if (err == 0 &&
6194 copy_to_user(arg, file, sizeof(*file)))
NeilBrown32a76272005-06-21 17:17:14 -07006195 err = -EFAULT;
NeilBrown4af1a042014-12-15 12:57:00 +11006196
NeilBrown32a76272005-06-21 17:17:14 -07006197 kfree(file);
6198 return err;
6199}
6200
NeilBrownf72ffdd2014-09-30 14:23:59 +10006201static int get_disk_info(struct mddev *mddev, void __user * arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006202{
6203 mdu_disk_info_t info;
NeilBrown3cb03002011-10-11 16:45:26 +11006204 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006205
6206 if (copy_from_user(&info, arg, sizeof(info)))
6207 return -EFAULT;
6208
NeilBrown1ca69c42012-10-11 13:37:33 +11006209 rcu_read_lock();
Goldwyn Rodrigues57d051d2015-04-14 10:43:55 -05006210 rdev = md_find_rdev_nr_rcu(mddev, info.number);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006211 if (rdev) {
6212 info.major = MAJOR(rdev->bdev->bd_dev);
6213 info.minor = MINOR(rdev->bdev->bd_dev);
6214 info.raid_disk = rdev->raid_disk;
6215 info.state = 0;
NeilBrownb2d444d2005-11-08 21:39:31 -08006216 if (test_bit(Faulty, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006217 info.state |= (1<<MD_DISK_FAULTY);
NeilBrownb2d444d2005-11-08 21:39:31 -08006218 else if (test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006219 info.state |= (1<<MD_DISK_ACTIVE);
6220 info.state |= (1<<MD_DISK_SYNC);
6221 }
Shaohua Li9efdca12015-10-12 16:59:50 -07006222 if (test_bit(Journal, &rdev->flags))
Song Liubac624f2015-08-13 14:31:55 -07006223 info.state |= (1<<MD_DISK_JOURNAL);
NeilBrown8ddf9ef2005-09-09 16:23:45 -07006224 if (test_bit(WriteMostly, &rdev->flags))
6225 info.state |= (1<<MD_DISK_WRITEMOSTLY);
NeilBrown688834e2016-11-18 16:16:11 +11006226 if (test_bit(FailFast, &rdev->flags))
6227 info.state |= (1<<MD_DISK_FAILFAST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006228 } else {
6229 info.major = info.minor = 0;
6230 info.raid_disk = -1;
6231 info.state = (1<<MD_DISK_REMOVED);
6232 }
NeilBrown1ca69c42012-10-11 13:37:33 +11006233 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006234
6235 if (copy_to_user(arg, &info, sizeof(info)))
6236 return -EFAULT;
6237
6238 return 0;
6239}
6240
NeilBrownf72ffdd2014-09-30 14:23:59 +10006241static int add_new_disk(struct mddev *mddev, mdu_disk_info_t *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006242{
6243 char b[BDEVNAME_SIZE], b2[BDEVNAME_SIZE];
NeilBrown3cb03002011-10-11 16:45:26 +11006244 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006245 dev_t dev = MKDEV(info->major,info->minor);
6246
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05006247 if (mddev_is_clustered(mddev) &&
6248 !(info->state & ((1 << MD_DISK_CLUSTER_ADD) | (1 << MD_DISK_CANDIDATE)))) {
NeilBrown9d487392016-11-02 14:16:49 +11006249 pr_warn("%s: Cannot add to clustered mddev.\n",
6250 mdname(mddev));
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05006251 return -EINVAL;
6252 }
6253
Linus Torvalds1da177e2005-04-16 15:20:36 -07006254 if (info->major != MAJOR(dev) || info->minor != MINOR(dev))
6255 return -EOVERFLOW;
6256
6257 if (!mddev->raid_disks) {
6258 int err;
6259 /* expecting a device which has a superblock */
6260 rdev = md_import_device(dev, mddev->major_version, mddev->minor_version);
6261 if (IS_ERR(rdev)) {
NeilBrown9d487392016-11-02 14:16:49 +11006262 pr_warn("md: md_import_device returned %ld\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07006263 PTR_ERR(rdev));
6264 return PTR_ERR(rdev);
6265 }
6266 if (!list_empty(&mddev->disks)) {
NeilBrown3cb03002011-10-11 16:45:26 +11006267 struct md_rdev *rdev0
6268 = list_entry(mddev->disks.next,
6269 struct md_rdev, same_set);
NeilBrowna9f326e2009-09-23 18:06:41 +10006270 err = super_types[mddev->major_version]
Linus Torvalds1da177e2005-04-16 15:20:36 -07006271 .load_super(rdev, rdev0, mddev->minor_version);
6272 if (err < 0) {
NeilBrown9d487392016-11-02 14:16:49 +11006273 pr_warn("md: %s has different UUID to %s\n",
NeilBrownf72ffdd2014-09-30 14:23:59 +10006274 bdevname(rdev->bdev,b),
Linus Torvalds1da177e2005-04-16 15:20:36 -07006275 bdevname(rdev0->bdev,b2));
6276 export_rdev(rdev);
6277 return -EINVAL;
6278 }
6279 }
6280 err = bind_rdev_to_array(rdev, mddev);
6281 if (err)
6282 export_rdev(rdev);
6283 return err;
6284 }
6285
6286 /*
6287 * add_new_disk can be used once the array is assembled
6288 * to add "hot spares". They must already have a superblock
6289 * written
6290 */
6291 if (mddev->pers) {
6292 int err;
6293 if (!mddev->pers->hot_add_disk) {
NeilBrown9d487392016-11-02 14:16:49 +11006294 pr_warn("%s: personality does not support diskops!\n",
6295 mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006296 return -EINVAL;
6297 }
NeilBrown7b1e35f2005-09-09 16:23:50 -07006298 if (mddev->persistent)
6299 rdev = md_import_device(dev, mddev->major_version,
6300 mddev->minor_version);
6301 else
6302 rdev = md_import_device(dev, -1, -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006303 if (IS_ERR(rdev)) {
NeilBrown9d487392016-11-02 14:16:49 +11006304 pr_warn("md: md_import_device returned %ld\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07006305 PTR_ERR(rdev));
6306 return PTR_ERR(rdev);
6307 }
NeilBrown1a855a02010-12-09 16:36:28 +11006308 /* set saved_raid_disk if appropriate */
NeilBrown41158c72005-06-21 17:17:25 -07006309 if (!mddev->persistent) {
6310 if (info->state & (1<<MD_DISK_SYNC) &&
NeilBrownbf572542011-01-12 09:03:35 +11006311 info->raid_disk < mddev->raid_disks) {
NeilBrown41158c72005-06-21 17:17:25 -07006312 rdev->raid_disk = info->raid_disk;
NeilBrownbf572542011-01-12 09:03:35 +11006313 set_bit(In_sync, &rdev->flags);
NeilBrown8313b8e2013-12-12 10:13:33 +11006314 clear_bit(Bitmap_sync, &rdev->flags);
NeilBrownbf572542011-01-12 09:03:35 +11006315 } else
NeilBrown41158c72005-06-21 17:17:25 -07006316 rdev->raid_disk = -1;
NeilBrownf4667222013-12-09 12:04:56 +11006317 rdev->saved_raid_disk = rdev->raid_disk;
NeilBrown41158c72005-06-21 17:17:25 -07006318 } else
6319 super_types[mddev->major_version].
6320 validate_super(mddev, rdev);
NeilBrownbedd86b2011-05-11 14:26:20 +10006321 if ((info->state & (1<<MD_DISK_SYNC)) &&
NeilBrownf4563092012-07-03 15:59:06 +10006322 rdev->raid_disk != info->raid_disk) {
NeilBrownbedd86b2011-05-11 14:26:20 +10006323 /* This was a hot-add request, but events doesn't
6324 * match, so reject it.
6325 */
6326 export_rdev(rdev);
6327 return -EINVAL;
6328 }
6329
NeilBrownb2d444d2005-11-08 21:39:31 -08006330 clear_bit(In_sync, &rdev->flags); /* just to be sure */
NeilBrown8ddf9ef2005-09-09 16:23:45 -07006331 if (info->state & (1<<MD_DISK_WRITEMOSTLY))
6332 set_bit(WriteMostly, &rdev->flags);
NeilBrown575a80f2009-03-31 14:33:13 +11006333 else
6334 clear_bit(WriteMostly, &rdev->flags);
NeilBrown688834e2016-11-18 16:16:11 +11006335 if (info->state & (1<<MD_DISK_FAILFAST))
6336 set_bit(FailFast, &rdev->flags);
6337 else
6338 clear_bit(FailFast, &rdev->flags);
NeilBrown8ddf9ef2005-09-09 16:23:45 -07006339
Shaohua Lif6b6ec52015-12-21 10:51:02 +11006340 if (info->state & (1<<MD_DISK_JOURNAL)) {
6341 struct md_rdev *rdev2;
6342 bool has_journal = false;
6343
6344 /* make sure no existing journal disk */
6345 rdev_for_each(rdev2, mddev) {
6346 if (test_bit(Journal, &rdev2->flags)) {
6347 has_journal = true;
6348 break;
6349 }
6350 }
6351 if (has_journal) {
6352 export_rdev(rdev);
6353 return -EBUSY;
6354 }
Song Liubac624f2015-08-13 14:31:55 -07006355 set_bit(Journal, &rdev->flags);
Shaohua Lif6b6ec52015-12-21 10:51:02 +11006356 }
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05006357 /*
6358 * check whether the device shows up in other nodes
6359 */
6360 if (mddev_is_clustered(mddev)) {
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05006361 if (info->state & (1 << MD_DISK_CANDIDATE))
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05006362 set_bit(Candidate, &rdev->flags);
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05006363 else if (info->state & (1 << MD_DISK_CLUSTER_ADD)) {
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05006364 /* --add initiated by this node */
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05006365 err = md_cluster_ops->add_new_disk(mddev, rdev);
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05006366 if (err) {
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05006367 export_rdev(rdev);
6368 return err;
6369 }
6370 }
6371 }
6372
Linus Torvalds1da177e2005-04-16 15:20:36 -07006373 rdev->raid_disk = -1;
6374 err = bind_rdev_to_array(rdev, mddev);
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05006375
Linus Torvalds1da177e2005-04-16 15:20:36 -07006376 if (err)
6377 export_rdev(rdev);
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05006378
6379 if (mddev_is_clustered(mddev)) {
Guoqing Jiange566aef2016-08-12 13:42:34 +08006380 if (info->state & (1 << MD_DISK_CANDIDATE)) {
6381 if (!err) {
6382 err = md_cluster_ops->new_disk_ack(mddev,
6383 err == 0);
6384 if (err)
6385 md_kick_rdev_from_array(rdev);
6386 }
6387 } else {
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05006388 if (err)
6389 md_cluster_ops->add_new_disk_cancel(mddev);
6390 else
6391 err = add_bound_rdev(rdev);
6392 }
6393
6394 } else if (!err)
Goldwyn Rodriguesa6da4ef2015-04-14 10:45:22 -05006395 err = add_bound_rdev(rdev);
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05006396
Linus Torvalds1da177e2005-04-16 15:20:36 -07006397 return err;
6398 }
6399
6400 /* otherwise, add_new_disk is only allowed
6401 * for major_version==0 superblocks
6402 */
6403 if (mddev->major_version != 0) {
NeilBrown9d487392016-11-02 14:16:49 +11006404 pr_warn("%s: ADD_NEW_DISK not supported\n", mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006405 return -EINVAL;
6406 }
6407
6408 if (!(info->state & (1<<MD_DISK_FAULTY))) {
6409 int err;
NeilBrownd710e132008-10-13 11:55:12 +11006410 rdev = md_import_device(dev, -1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006411 if (IS_ERR(rdev)) {
NeilBrown9d487392016-11-02 14:16:49 +11006412 pr_warn("md: error, md_import_device() returned %ld\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07006413 PTR_ERR(rdev));
6414 return PTR_ERR(rdev);
6415 }
6416 rdev->desc_nr = info->number;
6417 if (info->raid_disk < mddev->raid_disks)
6418 rdev->raid_disk = info->raid_disk;
6419 else
6420 rdev->raid_disk = -1;
6421
Linus Torvalds1da177e2005-04-16 15:20:36 -07006422 if (rdev->raid_disk < mddev->raid_disks)
NeilBrownb2d444d2005-11-08 21:39:31 -08006423 if (info->state & (1<<MD_DISK_SYNC))
6424 set_bit(In_sync, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006425
NeilBrown8ddf9ef2005-09-09 16:23:45 -07006426 if (info->state & (1<<MD_DISK_WRITEMOSTLY))
6427 set_bit(WriteMostly, &rdev->flags);
NeilBrown688834e2016-11-18 16:16:11 +11006428 if (info->state & (1<<MD_DISK_FAILFAST))
6429 set_bit(FailFast, &rdev->flags);
NeilBrown8ddf9ef2005-09-09 16:23:45 -07006430
Linus Torvalds1da177e2005-04-16 15:20:36 -07006431 if (!mddev->persistent) {
NeilBrown9d487392016-11-02 14:16:49 +11006432 pr_debug("md: nonpersistent superblock ...\n");
Mike Snitzer77304d22010-11-08 14:39:12 +01006433 rdev->sb_start = i_size_read(rdev->bdev->bd_inode) / 512;
6434 } else
Jonathan Brassow57b2caa2011-01-14 09:14:33 +11006435 rdev->sb_start = calc_dev_sboffset(rdev);
NeilBrown8190e752009-06-18 08:48:58 +10006436 rdev->sectors = rdev->sb_start;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006437
NeilBrown2bf071b2006-01-06 00:20:55 -08006438 err = bind_rdev_to_array(rdev, mddev);
6439 if (err) {
6440 export_rdev(rdev);
6441 return err;
6442 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006443 }
6444
6445 return 0;
6446}
6447
NeilBrownf72ffdd2014-09-30 14:23:59 +10006448static int hot_remove_disk(struct mddev *mddev, dev_t dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006449{
6450 char b[BDEVNAME_SIZE];
NeilBrown3cb03002011-10-11 16:45:26 +11006451 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006452
Linus Torvalds1da177e2005-04-16 15:20:36 -07006453 rdev = find_rdev(mddev, dev);
6454 if (!rdev)
6455 return -ENXIO;
6456
Goldwyn Rodrigues2910ff12015-09-28 10:27:26 -05006457 if (rdev->raid_disk < 0)
6458 goto kick_rdev;
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05006459
NeilBrown3ea8929d2013-04-24 11:42:41 +10006460 clear_bit(Blocked, &rdev->flags);
6461 remove_and_add_spares(mddev, rdev);
6462
Linus Torvalds1da177e2005-04-16 15:20:36 -07006463 if (rdev->raid_disk >= 0)
6464 goto busy;
6465
Goldwyn Rodrigues2910ff12015-09-28 10:27:26 -05006466kick_rdev:
Goldwyn Rodrigues54a88392015-12-21 10:51:00 +11006467 if (mddev_is_clustered(mddev))
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -05006468 md_cluster_ops->remove_disk(mddev, rdev);
6469
Goldwyn Rodriguesfb56dfe2015-04-14 10:43:24 -05006470 md_kick_rdev_from_array(rdev);
Shaohua Li29530792016-12-08 15:48:19 -08006471 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
NeilBrown060b0682016-11-04 16:46:03 +11006472 if (mddev->thread)
6473 md_wakeup_thread(mddev->thread);
6474 else
6475 md_update_sb(mddev, 1);
NeilBrownd7603b72006-01-06 00:20:30 -08006476 md_new_event(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006477
6478 return 0;
6479busy:
NeilBrown9d487392016-11-02 14:16:49 +11006480 pr_debug("md: cannot remove active disk %s from %s ...\n",
6481 bdevname(rdev->bdev,b), mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006482 return -EBUSY;
6483}
6484
NeilBrownf72ffdd2014-09-30 14:23:59 +10006485static int hot_add_disk(struct mddev *mddev, dev_t dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006486{
6487 char b[BDEVNAME_SIZE];
6488 int err;
NeilBrown3cb03002011-10-11 16:45:26 +11006489 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006490
6491 if (!mddev->pers)
6492 return -ENODEV;
6493
6494 if (mddev->major_version != 0) {
NeilBrown9d487392016-11-02 14:16:49 +11006495 pr_warn("%s: HOT_ADD may only be used with version-0 superblocks.\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07006496 mdname(mddev));
6497 return -EINVAL;
6498 }
6499 if (!mddev->pers->hot_add_disk) {
NeilBrown9d487392016-11-02 14:16:49 +11006500 pr_warn("%s: personality does not support diskops!\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07006501 mdname(mddev));
6502 return -EINVAL;
6503 }
6504
NeilBrownd710e132008-10-13 11:55:12 +11006505 rdev = md_import_device(dev, -1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006506 if (IS_ERR(rdev)) {
NeilBrown9d487392016-11-02 14:16:49 +11006507 pr_warn("md: error, md_import_device() returned %ld\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07006508 PTR_ERR(rdev));
6509 return -EINVAL;
6510 }
6511
6512 if (mddev->persistent)
Jonathan Brassow57b2caa2011-01-14 09:14:33 +11006513 rdev->sb_start = calc_dev_sboffset(rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006514 else
Mike Snitzer77304d22010-11-08 14:39:12 +01006515 rdev->sb_start = i_size_read(rdev->bdev->bd_inode) / 512;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006516
NeilBrown8190e752009-06-18 08:48:58 +10006517 rdev->sectors = rdev->sb_start;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006518
NeilBrownb2d444d2005-11-08 21:39:31 -08006519 if (test_bit(Faulty, &rdev->flags)) {
NeilBrown9d487392016-11-02 14:16:49 +11006520 pr_warn("md: can not hot-add faulty %s disk to %s!\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07006521 bdevname(rdev->bdev,b), mdname(mddev));
6522 err = -EINVAL;
6523 goto abort_export;
6524 }
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05006525
NeilBrownb2d444d2005-11-08 21:39:31 -08006526 clear_bit(In_sync, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006527 rdev->desc_nr = -1;
NeilBrown58427302006-10-06 00:44:04 -07006528 rdev->saved_raid_disk = -1;
NeilBrown2bf071b2006-01-06 00:20:55 -08006529 err = bind_rdev_to_array(rdev, mddev);
6530 if (err)
Goldwyn Rodrigues2aa82192015-09-28 19:21:35 -05006531 goto abort_export;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006532
6533 /*
6534 * The rest should better be atomic, we can have disk failures
6535 * noticed in interrupt contexts ...
6536 */
6537
Linus Torvalds1da177e2005-04-16 15:20:36 -07006538 rdev->raid_disk = -1;
6539
Shaohua Li29530792016-12-08 15:48:19 -08006540 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
NeilBrown060b0682016-11-04 16:46:03 +11006541 if (!mddev->thread)
6542 md_update_sb(mddev, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006543 /*
6544 * Kick recovery, maybe this spare has to be added to the
6545 * array immediately.
6546 */
6547 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
6548 md_wakeup_thread(mddev->thread);
NeilBrownd7603b72006-01-06 00:20:30 -08006549 md_new_event(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006550 return 0;
6551
Linus Torvalds1da177e2005-04-16 15:20:36 -07006552abort_export:
6553 export_rdev(rdev);
6554 return err;
6555}
6556
NeilBrownfd01b882011-10-11 16:47:53 +11006557static int set_bitmap_file(struct mddev *mddev, int fd)
NeilBrown32a76272005-06-21 17:17:14 -07006558{
NeilBrown035328c2014-04-09 12:25:40 +10006559 int err = 0;
NeilBrown32a76272005-06-21 17:17:14 -07006560
NeilBrown36fa3062005-09-09 16:23:45 -07006561 if (mddev->pers) {
NeilBrownd66b1b32014-08-08 15:40:24 +10006562 if (!mddev->pers->quiesce || !mddev->thread)
NeilBrown36fa3062005-09-09 16:23:45 -07006563 return -EBUSY;
6564 if (mddev->recovery || mddev->sync_thread)
6565 return -EBUSY;
6566 /* we should be able to change the bitmap.. */
NeilBrown32a76272005-06-21 17:17:14 -07006567 }
6568
NeilBrown36fa3062005-09-09 16:23:45 -07006569 if (fd >= 0) {
NeilBrown035328c2014-04-09 12:25:40 +10006570 struct inode *inode;
NeilBrown1e594bb2014-12-15 12:57:00 +11006571 struct file *f;
NeilBrown36fa3062005-09-09 16:23:45 -07006572
NeilBrown1e594bb2014-12-15 12:57:00 +11006573 if (mddev->bitmap || mddev->bitmap_info.file)
6574 return -EEXIST; /* cannot add when bitmap is present */
6575 f = fget(fd);
6576
6577 if (f == NULL) {
NeilBrown9d487392016-11-02 14:16:49 +11006578 pr_warn("%s: error: failed to get bitmap file\n",
6579 mdname(mddev));
NeilBrown36fa3062005-09-09 16:23:45 -07006580 return -EBADF;
6581 }
6582
NeilBrown1e594bb2014-12-15 12:57:00 +11006583 inode = f->f_mapping->host;
NeilBrown035328c2014-04-09 12:25:40 +10006584 if (!S_ISREG(inode->i_mode)) {
NeilBrown9d487392016-11-02 14:16:49 +11006585 pr_warn("%s: error: bitmap file must be a regular file\n",
6586 mdname(mddev));
NeilBrown035328c2014-04-09 12:25:40 +10006587 err = -EBADF;
NeilBrown1e594bb2014-12-15 12:57:00 +11006588 } else if (!(f->f_mode & FMODE_WRITE)) {
NeilBrown9d487392016-11-02 14:16:49 +11006589 pr_warn("%s: error: bitmap file must open for write\n",
6590 mdname(mddev));
NeilBrown035328c2014-04-09 12:25:40 +10006591 err = -EBADF;
6592 } else if (atomic_read(&inode->i_writecount) != 1) {
NeilBrown9d487392016-11-02 14:16:49 +11006593 pr_warn("%s: error: bitmap file is already in use\n",
6594 mdname(mddev));
NeilBrown035328c2014-04-09 12:25:40 +10006595 err = -EBUSY;
6596 }
6597 if (err) {
NeilBrown1e594bb2014-12-15 12:57:00 +11006598 fput(f);
NeilBrown36fa3062005-09-09 16:23:45 -07006599 return err;
6600 }
NeilBrown1e594bb2014-12-15 12:57:00 +11006601 mddev->bitmap_info.file = f;
NeilBrownc3d97142009-12-14 12:49:52 +11006602 mddev->bitmap_info.offset = 0; /* file overrides offset */
NeilBrown36fa3062005-09-09 16:23:45 -07006603 } else if (mddev->bitmap == NULL)
6604 return -ENOENT; /* cannot remove what isn't there */
6605 err = 0;
6606 if (mddev->pers) {
6607 mddev->pers->quiesce(mddev, 1);
NeilBrown69e51b42010-06-01 19:37:35 +10006608 if (fd >= 0) {
Goldwyn Rodriguesf9209a32014-06-06 12:43:49 -05006609 struct bitmap *bitmap;
6610
6611 bitmap = bitmap_create(mddev, -1);
6612 if (!IS_ERR(bitmap)) {
6613 mddev->bitmap = bitmap;
NeilBrown69e51b42010-06-01 19:37:35 +10006614 err = bitmap_load(mddev);
NeilBrownba599ac2015-02-25 11:44:11 +11006615 } else
6616 err = PTR_ERR(bitmap);
NeilBrown69e51b42010-06-01 19:37:35 +10006617 }
NeilBrownd7375ab322006-06-26 00:27:43 -07006618 if (fd < 0 || err) {
NeilBrown36fa3062005-09-09 16:23:45 -07006619 bitmap_destroy(mddev);
NeilBrownd7375ab322006-06-26 00:27:43 -07006620 fd = -1; /* make sure to put the file */
6621 }
NeilBrown36fa3062005-09-09 16:23:45 -07006622 mddev->pers->quiesce(mddev, 0);
NeilBrownd7375ab322006-06-26 00:27:43 -07006623 }
6624 if (fd < 0) {
NeilBrown4af1a042014-12-15 12:57:00 +11006625 struct file *f = mddev->bitmap_info.file;
6626 if (f) {
6627 spin_lock(&mddev->lock);
6628 mddev->bitmap_info.file = NULL;
6629 spin_unlock(&mddev->lock);
6630 fput(f);
6631 }
NeilBrown36fa3062005-09-09 16:23:45 -07006632 }
6633
NeilBrown32a76272005-06-21 17:17:14 -07006634 return err;
6635}
6636
Linus Torvalds1da177e2005-04-16 15:20:36 -07006637/*
6638 * set_array_info is used two different ways
6639 * The original usage is when creating a new array.
6640 * In this usage, raid_disks is > 0 and it together with
6641 * level, size, not_persistent,layout,chunksize determine the
6642 * shape of the array.
6643 * This will always create an array with a type-0.90.0 superblock.
6644 * The newer usage is when assembling an array.
6645 * In this case raid_disks will be 0, and the major_version field is
6646 * use to determine which style super-blocks are to be found on the devices.
6647 * The minor and patch _version numbers are also kept incase the
6648 * super_block handler wishes to interpret them.
6649 */
NeilBrownf72ffdd2014-09-30 14:23:59 +10006650static int set_array_info(struct mddev *mddev, mdu_array_info_t *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006651{
6652
6653 if (info->raid_disks == 0) {
6654 /* just setting version number for superblock loading */
6655 if (info->major_version < 0 ||
Ahmed S. Darwish50511da2007-05-09 02:35:34 -07006656 info->major_version >= ARRAY_SIZE(super_types) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07006657 super_types[info->major_version].name == NULL) {
6658 /* maybe try to auto-load a module? */
NeilBrown9d487392016-11-02 14:16:49 +11006659 pr_warn("md: superblock version %d not known\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07006660 info->major_version);
6661 return -EINVAL;
6662 }
6663 mddev->major_version = info->major_version;
6664 mddev->minor_version = info->minor_version;
6665 mddev->patch_version = info->patch_version;
NeilBrown3f9d7b02006-12-22 01:11:41 -08006666 mddev->persistent = !info->not_persistent;
NeilBrowncbd19982009-12-30 12:08:49 +11006667 /* ensure mddev_put doesn't delete this now that there
6668 * is some minimal configuration.
6669 */
Deepa Dinamani9ebc6ef2015-12-21 10:51:01 +11006670 mddev->ctime = ktime_get_real_seconds();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006671 return 0;
6672 }
6673 mddev->major_version = MD_MAJOR_VERSION;
6674 mddev->minor_version = MD_MINOR_VERSION;
6675 mddev->patch_version = MD_PATCHLEVEL_VERSION;
Deepa Dinamani9ebc6ef2015-12-21 10:51:01 +11006676 mddev->ctime = ktime_get_real_seconds();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006677
6678 mddev->level = info->level;
NeilBrown17115e02006-01-16 22:14:57 -08006679 mddev->clevel[0] = 0;
Andre Noll58c0fed2009-03-31 14:33:13 +11006680 mddev->dev_sectors = 2 * (sector_t)info->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006681 mddev->raid_disks = info->raid_disks;
6682 /* don't set md_minor, it is determined by which /dev/md* was
6683 * openned
6684 */
6685 if (info->state & (1<<MD_SB_CLEAN))
6686 mddev->recovery_cp = MaxSector;
6687 else
6688 mddev->recovery_cp = 0;
6689 mddev->persistent = ! info->not_persistent;
NeilBrowne6910632008-02-06 01:39:51 -08006690 mddev->external = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006691
6692 mddev->layout = info->layout;
Andre Noll9d8f0362009-06-18 08:45:01 +10006693 mddev->chunk_sectors = info->chunk_size >> 9;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006694
Shaohua Li29530792016-12-08 15:48:19 -08006695 if (mddev->persistent) {
NeilBrown1b3bae42017-03-01 07:31:28 +11006696 mddev->max_disks = MD_SB_DISKS;
6697 mddev->flags = 0;
6698 mddev->sb_flags = 0;
Shaohua Li29530792016-12-08 15:48:19 -08006699 }
6700 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006701
NeilBrownc3d97142009-12-14 12:49:52 +11006702 mddev->bitmap_info.default_offset = MD_SB_BYTES >> 9;
NeilBrown6409bb02012-05-22 13:55:07 +10006703 mddev->bitmap_info.default_space = 64*2 - (MD_SB_BYTES >> 9);
NeilBrownc3d97142009-12-14 12:49:52 +11006704 mddev->bitmap_info.offset = 0;
NeilBrownb2a27032005-11-28 13:44:12 -08006705
NeilBrownf6705572006-03-27 01:18:11 -08006706 mddev->reshape_position = MaxSector;
6707
Linus Torvalds1da177e2005-04-16 15:20:36 -07006708 /*
6709 * Generate a 128 bit UUID
6710 */
6711 get_random_bytes(mddev->uuid, 16);
6712
NeilBrownf6705572006-03-27 01:18:11 -08006713 mddev->new_level = mddev->level;
Andre Noll664e7c42009-06-18 08:45:27 +10006714 mddev->new_chunk_sectors = mddev->chunk_sectors;
NeilBrownf6705572006-03-27 01:18:11 -08006715 mddev->new_layout = mddev->layout;
6716 mddev->delta_disks = 0;
NeilBrown2c810cd2012-05-21 09:27:00 +10006717 mddev->reshape_backwards = 0;
NeilBrownf6705572006-03-27 01:18:11 -08006718
Linus Torvalds1da177e2005-04-16 15:20:36 -07006719 return 0;
6720}
6721
NeilBrownfd01b882011-10-11 16:47:53 +11006722void md_set_array_sectors(struct mddev *mddev, sector_t array_sectors)
Dan Williams1f403622009-03-31 14:59:03 +11006723{
Dan Williamsb522adc2009-03-31 15:00:31 +11006724 WARN(!mddev_is_locked(mddev), "%s: unlocked mddev!\n", __func__);
6725
6726 if (mddev->external_size)
6727 return;
6728
Dan Williams1f403622009-03-31 14:59:03 +11006729 mddev->array_sectors = array_sectors;
6730}
6731EXPORT_SYMBOL(md_set_array_sectors);
6732
NeilBrownfd01b882011-10-11 16:47:53 +11006733static int update_size(struct mddev *mddev, sector_t num_sectors)
NeilBrowna35b0d62006-01-06 00:20:49 -08006734{
NeilBrown3cb03002011-10-11 16:45:26 +11006735 struct md_rdev *rdev;
NeilBrowna35b0d62006-01-06 00:20:49 -08006736 int rv;
Andre Nolld71f9f82008-07-11 22:02:22 +10006737 int fit = (num_sectors == 0);
Guoqing Jiang818da592017-03-01 16:42:40 +08006738 sector_t old_dev_sectors = mddev->dev_sectors;
Guoqing Jiangab5a98b2016-05-02 11:33:13 -04006739
NeilBrowna35b0d62006-01-06 00:20:49 -08006740 if (mddev->pers->resize == NULL)
6741 return -EINVAL;
Andre Nolld71f9f82008-07-11 22:02:22 +10006742 /* The "num_sectors" is the number of sectors of each device that
6743 * is used. This can only make sense for arrays with redundancy.
6744 * linear and raid0 always use whatever space is available. We can only
6745 * consider changing this number if no resync or reconstruction is
6746 * happening, and if the new size is acceptable. It must fit before the
Andre Noll0f420352008-07-11 22:02:23 +10006747 * sb_start or, if that is <data_offset, it must fit before the size
Andre Nolld71f9f82008-07-11 22:02:22 +10006748 * of each device. If num_sectors is zero, we find the largest size
6749 * that fits.
NeilBrowna35b0d62006-01-06 00:20:49 -08006750 */
NeilBrownf851b602014-12-11 10:02:10 +11006751 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
6752 mddev->sync_thread)
NeilBrowna35b0d62006-01-06 00:20:49 -08006753 return -EBUSY;
NeilBrownbd8839e2014-05-28 13:39:21 +10006754 if (mddev->ro)
6755 return -EROFS;
NeilBrowna4a61252012-05-22 13:55:27 +10006756
NeilBrowndafb20f2012-03-19 12:46:39 +11006757 rdev_for_each(rdev, mddev) {
Andre Nolldd8ac332009-03-31 14:33:13 +11006758 sector_t avail = rdev->sectors;
NeilBrown01ab5662006-10-28 10:38:30 -07006759
Andre Nolld71f9f82008-07-11 22:02:22 +10006760 if (fit && (num_sectors == 0 || num_sectors > avail))
6761 num_sectors = avail;
6762 if (avail < num_sectors)
NeilBrowna35b0d62006-01-06 00:20:49 -08006763 return -ENOSPC;
6764 }
Andre Nolld71f9f82008-07-11 22:02:22 +10006765 rv = mddev->pers->resize(mddev, num_sectors);
Guoqing Jiangc9483632017-02-24 11:15:23 +08006766 if (!rv) {
Guoqing Jiang818da592017-03-01 16:42:40 +08006767 if (mddev_is_clustered(mddev))
6768 md_cluster_ops->update_size(mddev, old_dev_sectors);
6769 else if (mddev->queue) {
Guoqing Jiangc9483632017-02-24 11:15:23 +08006770 set_capacity(mddev->gendisk, mddev->array_sectors);
6771 revalidate_disk(mddev->gendisk);
6772 }
6773 }
NeilBrowna35b0d62006-01-06 00:20:49 -08006774 return rv;
6775}
6776
NeilBrownfd01b882011-10-11 16:47:53 +11006777static int update_raid_disks(struct mddev *mddev, int raid_disks)
NeilBrownda943b992006-01-06 00:20:54 -08006778{
6779 int rv;
NeilBrownc6563a82012-05-21 09:27:00 +10006780 struct md_rdev *rdev;
NeilBrownda943b992006-01-06 00:20:54 -08006781 /* change the number of raid disks */
NeilBrown63c70c42006-03-27 01:18:13 -08006782 if (mddev->pers->check_reshape == NULL)
NeilBrownda943b992006-01-06 00:20:54 -08006783 return -EINVAL;
NeilBrownbd8839e2014-05-28 13:39:21 +10006784 if (mddev->ro)
6785 return -EROFS;
NeilBrownda943b992006-01-06 00:20:54 -08006786 if (raid_disks <= 0 ||
NeilBrown233fca32010-04-14 17:02:09 +10006787 (mddev->max_disks && raid_disks >= mddev->max_disks))
NeilBrownda943b992006-01-06 00:20:54 -08006788 return -EINVAL;
NeilBrownf851b602014-12-11 10:02:10 +11006789 if (mddev->sync_thread ||
6790 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
6791 mddev->reshape_position != MaxSector)
NeilBrownda943b992006-01-06 00:20:54 -08006792 return -EBUSY;
NeilBrownc6563a82012-05-21 09:27:00 +10006793
6794 rdev_for_each(rdev, mddev) {
6795 if (mddev->raid_disks < raid_disks &&
6796 rdev->data_offset < rdev->new_data_offset)
6797 return -EINVAL;
6798 if (mddev->raid_disks > raid_disks &&
6799 rdev->data_offset > rdev->new_data_offset)
6800 return -EINVAL;
6801 }
6802
NeilBrown63c70c42006-03-27 01:18:13 -08006803 mddev->delta_disks = raid_disks - mddev->raid_disks;
NeilBrown2c810cd2012-05-21 09:27:00 +10006804 if (mddev->delta_disks < 0)
6805 mddev->reshape_backwards = 1;
6806 else if (mddev->delta_disks > 0)
6807 mddev->reshape_backwards = 0;
NeilBrown63c70c42006-03-27 01:18:13 -08006808
6809 rv = mddev->pers->check_reshape(mddev);
NeilBrown2c810cd2012-05-21 09:27:00 +10006810 if (rv < 0) {
NeilBrownde171cb2011-01-31 11:57:42 +11006811 mddev->delta_disks = 0;
NeilBrown2c810cd2012-05-21 09:27:00 +10006812 mddev->reshape_backwards = 0;
6813 }
NeilBrownda943b992006-01-06 00:20:54 -08006814 return rv;
6815}
6816
Linus Torvalds1da177e2005-04-16 15:20:36 -07006817/*
6818 * update_array_info is used to change the configuration of an
6819 * on-line array.
6820 * The version, ctime,level,size,raid_disks,not_persistent, layout,chunk_size
6821 * fields in the info are checked against the array.
6822 * Any differences that cannot be handled will cause an error.
6823 * Normally, only one change can be managed at a time.
6824 */
NeilBrownfd01b882011-10-11 16:47:53 +11006825static int update_array_info(struct mddev *mddev, mdu_array_info_t *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006826{
6827 int rv = 0;
6828 int cnt = 0;
NeilBrown36fa3062005-09-09 16:23:45 -07006829 int state = 0;
6830
6831 /* calculate expected state,ignoring low bits */
NeilBrownc3d97142009-12-14 12:49:52 +11006832 if (mddev->bitmap && mddev->bitmap_info.offset)
NeilBrown36fa3062005-09-09 16:23:45 -07006833 state |= (1 << MD_SB_BITMAP_PRESENT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006834
6835 if (mddev->major_version != info->major_version ||
6836 mddev->minor_version != info->minor_version ||
6837/* mddev->patch_version != info->patch_version || */
6838 mddev->ctime != info->ctime ||
6839 mddev->level != info->level ||
6840/* mddev->layout != info->layout || */
Firo Yang4e023612015-06-11 09:41:10 +08006841 mddev->persistent != !info->not_persistent ||
Andre Noll9d8f0362009-06-18 08:45:01 +10006842 mddev->chunk_sectors != info->chunk_size >> 9 ||
NeilBrown36fa3062005-09-09 16:23:45 -07006843 /* ignore bottom 8 bits of state, and allow SB_BITMAP_PRESENT to change */
6844 ((state^info->state) & 0xfffffe00)
6845 )
Linus Torvalds1da177e2005-04-16 15:20:36 -07006846 return -EINVAL;
6847 /* Check there is only one change */
Andre Noll58c0fed2009-03-31 14:33:13 +11006848 if (info->size >= 0 && mddev->dev_sectors / 2 != info->size)
6849 cnt++;
6850 if (mddev->raid_disks != info->raid_disks)
6851 cnt++;
6852 if (mddev->layout != info->layout)
6853 cnt++;
6854 if ((state ^ info->state) & (1<<MD_SB_BITMAP_PRESENT))
6855 cnt++;
6856 if (cnt == 0)
6857 return 0;
6858 if (cnt > 1)
6859 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006860
6861 if (mddev->layout != info->layout) {
6862 /* Change layout
6863 * we don't need to do anything at the md level, the
6864 * personality will take care of it all.
6865 */
NeilBrown50ac1682009-06-18 08:47:55 +10006866 if (mddev->pers->check_reshape == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006867 return -EINVAL;
NeilBrown597a7112009-06-18 08:47:42 +10006868 else {
6869 mddev->new_layout = info->layout;
NeilBrown50ac1682009-06-18 08:47:55 +10006870 rv = mddev->pers->check_reshape(mddev);
NeilBrown597a7112009-06-18 08:47:42 +10006871 if (rv)
6872 mddev->new_layout = mddev->layout;
6873 return rv;
6874 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006875 }
Andre Noll58c0fed2009-03-31 14:33:13 +11006876 if (info->size >= 0 && mddev->dev_sectors / 2 != info->size)
Andre Nolld71f9f82008-07-11 22:02:22 +10006877 rv = update_size(mddev, (sector_t)info->size * 2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006878
NeilBrownda943b992006-01-06 00:20:54 -08006879 if (mddev->raid_disks != info->raid_disks)
6880 rv = update_raid_disks(mddev, info->raid_disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006881
NeilBrown36fa3062005-09-09 16:23:45 -07006882 if ((state ^ info->state) & (1<<MD_SB_BITMAP_PRESENT)) {
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05006883 if (mddev->pers->quiesce == NULL || mddev->thread == NULL) {
6884 rv = -EINVAL;
6885 goto err;
6886 }
6887 if (mddev->recovery || mddev->sync_thread) {
6888 rv = -EBUSY;
6889 goto err;
6890 }
NeilBrown36fa3062005-09-09 16:23:45 -07006891 if (info->state & (1<<MD_SB_BITMAP_PRESENT)) {
Goldwyn Rodriguesf9209a32014-06-06 12:43:49 -05006892 struct bitmap *bitmap;
NeilBrown36fa3062005-09-09 16:23:45 -07006893 /* add the bitmap */
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05006894 if (mddev->bitmap) {
6895 rv = -EEXIST;
6896 goto err;
6897 }
6898 if (mddev->bitmap_info.default_offset == 0) {
6899 rv = -EINVAL;
6900 goto err;
6901 }
NeilBrownc3d97142009-12-14 12:49:52 +11006902 mddev->bitmap_info.offset =
6903 mddev->bitmap_info.default_offset;
NeilBrown6409bb02012-05-22 13:55:07 +10006904 mddev->bitmap_info.space =
6905 mddev->bitmap_info.default_space;
NeilBrown36fa3062005-09-09 16:23:45 -07006906 mddev->pers->quiesce(mddev, 1);
Goldwyn Rodriguesf9209a32014-06-06 12:43:49 -05006907 bitmap = bitmap_create(mddev, -1);
6908 if (!IS_ERR(bitmap)) {
6909 mddev->bitmap = bitmap;
NeilBrown69e51b42010-06-01 19:37:35 +10006910 rv = bitmap_load(mddev);
NeilBrownba599ac2015-02-25 11:44:11 +11006911 } else
6912 rv = PTR_ERR(bitmap);
NeilBrown36fa3062005-09-09 16:23:45 -07006913 if (rv)
6914 bitmap_destroy(mddev);
6915 mddev->pers->quiesce(mddev, 0);
6916 } else {
6917 /* remove the bitmap */
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05006918 if (!mddev->bitmap) {
6919 rv = -ENOENT;
6920 goto err;
6921 }
6922 if (mddev->bitmap->storage.file) {
6923 rv = -EINVAL;
6924 goto err;
6925 }
Guoqing Jiangf6a2dc62015-12-21 10:51:00 +11006926 if (mddev->bitmap_info.nodes) {
6927 /* hold PW on all the bitmap lock */
6928 if (md_cluster_ops->lock_all_bitmaps(mddev) <= 0) {
NeilBrown9d487392016-11-02 14:16:49 +11006929 pr_warn("md: can't change bitmap to none since the array is in use by more than one node\n");
Guoqing Jiangf6a2dc62015-12-21 10:51:00 +11006930 rv = -EPERM;
6931 md_cluster_ops->unlock_all_bitmaps(mddev);
6932 goto err;
6933 }
6934
6935 mddev->bitmap_info.nodes = 0;
6936 md_cluster_ops->leave(mddev);
6937 }
NeilBrown36fa3062005-09-09 16:23:45 -07006938 mddev->pers->quiesce(mddev, 1);
6939 bitmap_destroy(mddev);
6940 mddev->pers->quiesce(mddev, 0);
NeilBrownc3d97142009-12-14 12:49:52 +11006941 mddev->bitmap_info.offset = 0;
NeilBrown36fa3062005-09-09 16:23:45 -07006942 }
6943 }
NeilBrown850b2b42006-10-03 01:15:46 -07006944 md_update_sb(mddev, 1);
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05006945 return rv;
6946err:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006947 return rv;
6948}
6949
NeilBrownfd01b882011-10-11 16:47:53 +11006950static int set_disk_faulty(struct mddev *mddev, dev_t dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006951{
NeilBrown3cb03002011-10-11 16:45:26 +11006952 struct md_rdev *rdev;
NeilBrown1ca69c42012-10-11 13:37:33 +11006953 int err = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006954
6955 if (mddev->pers == NULL)
6956 return -ENODEV;
6957
NeilBrown1ca69c42012-10-11 13:37:33 +11006958 rcu_read_lock();
6959 rdev = find_rdev_rcu(mddev, dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006960 if (!rdev)
NeilBrown1ca69c42012-10-11 13:37:33 +11006961 err = -ENODEV;
6962 else {
6963 md_error(mddev, rdev);
6964 if (!test_bit(Faulty, &rdev->flags))
6965 err = -EBUSY;
6966 }
6967 rcu_read_unlock();
6968 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006969}
6970
Andre Noll2f9618c2008-04-25 18:57:58 +02006971/*
6972 * We have a problem here : there is no easy way to give a CHS
6973 * virtual geometry. We currently pretend that we have a 2 heads
6974 * 4 sectors (with a BIG number of cylinders...). This drives
6975 * dosfs just mad... ;-)
6976 */
Christoph Hellwiga885c8c2006-01-08 01:02:50 -08006977static int md_getgeo(struct block_device *bdev, struct hd_geometry *geo)
6978{
NeilBrownfd01b882011-10-11 16:47:53 +11006979 struct mddev *mddev = bdev->bd_disk->private_data;
Christoph Hellwiga885c8c2006-01-08 01:02:50 -08006980
6981 geo->heads = 2;
6982 geo->sectors = 4;
NeilBrown49ce6ce2010-03-29 10:51:42 +11006983 geo->cylinders = mddev->array_sectors / 8;
Christoph Hellwiga885c8c2006-01-08 01:02:50 -08006984 return 0;
6985}
6986
Nicolas Schichancb335f82014-01-15 16:58:52 +01006987static inline bool md_ioctl_valid(unsigned int cmd)
6988{
6989 switch (cmd) {
6990 case ADD_NEW_DISK:
6991 case BLKROSET:
6992 case GET_ARRAY_INFO:
6993 case GET_BITMAP_FILE:
6994 case GET_DISK_INFO:
6995 case HOT_ADD_DISK:
6996 case HOT_REMOVE_DISK:
Nicolas Schichancb335f82014-01-15 16:58:52 +01006997 case RAID_AUTORUN:
6998 case RAID_VERSION:
6999 case RESTART_ARRAY_RW:
7000 case RUN_ARRAY:
7001 case SET_ARRAY_INFO:
7002 case SET_BITMAP_FILE:
7003 case SET_DISK_FAULTY:
7004 case STOP_ARRAY:
7005 case STOP_ARRAY_RO:
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05007006 case CLUSTERED_DISK_NACK:
Nicolas Schichancb335f82014-01-15 16:58:52 +01007007 return true;
7008 default:
7009 return false;
7010 }
7011}
7012
Al Viroa39907f2008-03-02 10:31:15 -05007013static int md_ioctl(struct block_device *bdev, fmode_t mode,
Linus Torvalds1da177e2005-04-16 15:20:36 -07007014 unsigned int cmd, unsigned long arg)
7015{
7016 int err = 0;
7017 void __user *argp = (void __user *)arg;
NeilBrownfd01b882011-10-11 16:47:53 +11007018 struct mddev *mddev = NULL;
Dan Williamse2218352010-05-12 08:25:37 +10007019 int ro;
NeilBrown065e5192017-04-06 11:16:33 +08007020 bool did_set_md_closing = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007021
Nicolas Schichancb335f82014-01-15 16:58:52 +01007022 if (!md_ioctl_valid(cmd))
7023 return -ENOTTY;
7024
NeilBrown506c9e42011-12-23 10:17:26 +11007025 switch (cmd) {
7026 case RAID_VERSION:
7027 case GET_ARRAY_INFO:
7028 case GET_DISK_INFO:
7029 break;
7030 default:
7031 if (!capable(CAP_SYS_ADMIN))
7032 return -EACCES;
7033 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007034
7035 /*
7036 * Commands dealing with the RAID driver but not any
7037 * particular array:
7038 */
NeilBrownc02c0ae2012-12-11 13:39:21 +11007039 switch (cmd) {
7040 case RAID_VERSION:
7041 err = get_version(argp);
NeilBrown3adc28d2014-09-30 15:46:41 +10007042 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007043
Linus Torvalds1da177e2005-04-16 15:20:36 -07007044#ifndef MODULE
NeilBrownc02c0ae2012-12-11 13:39:21 +11007045 case RAID_AUTORUN:
7046 err = 0;
7047 autostart_arrays(arg);
NeilBrown3adc28d2014-09-30 15:46:41 +10007048 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007049#endif
NeilBrownc02c0ae2012-12-11 13:39:21 +11007050 default:;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007051 }
7052
7053 /*
7054 * Commands creating/starting a new array:
7055 */
7056
Al Viroa39907f2008-03-02 10:31:15 -05007057 mddev = bdev->bd_disk->private_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007058
7059 if (!mddev) {
7060 BUG();
NeilBrown3adc28d2014-09-30 15:46:41 +10007061 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007062 }
7063
NeilBrown1ca69c42012-10-11 13:37:33 +11007064 /* Some actions do not requires the mutex */
7065 switch (cmd) {
7066 case GET_ARRAY_INFO:
7067 if (!mddev->raid_disks && !mddev->external)
7068 err = -ENODEV;
7069 else
7070 err = get_array_info(mddev, argp);
NeilBrown3adc28d2014-09-30 15:46:41 +10007071 goto out;
NeilBrown1ca69c42012-10-11 13:37:33 +11007072
7073 case GET_DISK_INFO:
7074 if (!mddev->raid_disks && !mddev->external)
7075 err = -ENODEV;
7076 else
7077 err = get_disk_info(mddev, argp);
NeilBrown3adc28d2014-09-30 15:46:41 +10007078 goto out;
NeilBrown1ca69c42012-10-11 13:37:33 +11007079
7080 case SET_DISK_FAULTY:
7081 err = set_disk_faulty(mddev, new_decode_dev(arg));
NeilBrown3adc28d2014-09-30 15:46:41 +10007082 goto out;
NeilBrown4af1a042014-12-15 12:57:00 +11007083
7084 case GET_BITMAP_FILE:
7085 err = get_bitmap_file(mddev, argp);
7086 goto out;
7087
NeilBrown1ca69c42012-10-11 13:37:33 +11007088 }
7089
NeilBrowna7a3f082012-12-11 13:35:54 +11007090 if (cmd == ADD_NEW_DISK)
7091 /* need to ensure md_delayed_delete() has completed */
7092 flush_workqueue(md_misc_wq);
7093
Hannes Reinecke90f5f7a2013-04-02 08:38:55 +02007094 if (cmd == HOT_REMOVE_DISK)
7095 /* need to ensure recovery thread has run */
7096 wait_event_interruptible_timeout(mddev->sb_wait,
7097 !test_bit(MD_RECOVERY_NEEDED,
Shaohua Li82a301c2016-12-08 15:48:18 -08007098 &mddev->recovery),
Hannes Reinecke90f5f7a2013-04-02 08:38:55 +02007099 msecs_to_jiffies(5000));
NeilBrown260fa032013-08-27 16:44:13 +10007100 if (cmd == STOP_ARRAY || cmd == STOP_ARRAY_RO) {
7101 /* Need to flush page cache, and ensure no-one else opens
7102 * and writes
7103 */
7104 mutex_lock(&mddev->open_mutex);
NeilBrown9ba3b7f2014-09-09 14:00:15 +10007105 if (mddev->pers && atomic_read(&mddev->openers) > 1) {
NeilBrown260fa032013-08-27 16:44:13 +10007106 mutex_unlock(&mddev->open_mutex);
7107 err = -EBUSY;
NeilBrown3adc28d2014-09-30 15:46:41 +10007108 goto out;
NeilBrown260fa032013-08-27 16:44:13 +10007109 }
NeilBrown065e5192017-04-06 11:16:33 +08007110 WARN_ON_ONCE(test_bit(MD_CLOSING, &mddev->flags));
Guoqing Jiangaf8d8e62016-08-12 13:42:37 +08007111 set_bit(MD_CLOSING, &mddev->flags);
NeilBrown065e5192017-04-06 11:16:33 +08007112 did_set_md_closing = true;
NeilBrown260fa032013-08-27 16:44:13 +10007113 mutex_unlock(&mddev->open_mutex);
7114 sync_blockdev(bdev);
7115 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007116 err = mddev_lock(mddev);
7117 if (err) {
NeilBrown9d487392016-11-02 14:16:49 +11007118 pr_debug("md: ioctl lock interrupted, reason %d, cmd %d\n",
7119 err, cmd);
NeilBrown3adc28d2014-09-30 15:46:41 +10007120 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007121 }
7122
NeilBrownc02c0ae2012-12-11 13:39:21 +11007123 if (cmd == SET_ARRAY_INFO) {
7124 mdu_array_info_t info;
7125 if (!arg)
7126 memset(&info, 0, sizeof(info));
7127 else if (copy_from_user(&info, argp, sizeof(info))) {
7128 err = -EFAULT;
NeilBrown3adc28d2014-09-30 15:46:41 +10007129 goto unlock;
NeilBrownc02c0ae2012-12-11 13:39:21 +11007130 }
7131 if (mddev->pers) {
7132 err = update_array_info(mddev, &info);
7133 if (err) {
NeilBrown9d487392016-11-02 14:16:49 +11007134 pr_warn("md: couldn't update array info. %d\n", err);
NeilBrown3adc28d2014-09-30 15:46:41 +10007135 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007136 }
NeilBrown3adc28d2014-09-30 15:46:41 +10007137 goto unlock;
NeilBrownc02c0ae2012-12-11 13:39:21 +11007138 }
7139 if (!list_empty(&mddev->disks)) {
NeilBrown9d487392016-11-02 14:16:49 +11007140 pr_warn("md: array %s already has disks!\n", mdname(mddev));
NeilBrownc02c0ae2012-12-11 13:39:21 +11007141 err = -EBUSY;
NeilBrown3adc28d2014-09-30 15:46:41 +10007142 goto unlock;
NeilBrownc02c0ae2012-12-11 13:39:21 +11007143 }
7144 if (mddev->raid_disks) {
NeilBrown9d487392016-11-02 14:16:49 +11007145 pr_warn("md: array %s already initialised!\n", mdname(mddev));
NeilBrownc02c0ae2012-12-11 13:39:21 +11007146 err = -EBUSY;
NeilBrown3adc28d2014-09-30 15:46:41 +10007147 goto unlock;
NeilBrownc02c0ae2012-12-11 13:39:21 +11007148 }
7149 err = set_array_info(mddev, &info);
7150 if (err) {
NeilBrown9d487392016-11-02 14:16:49 +11007151 pr_warn("md: couldn't set array info. %d\n", err);
NeilBrown3adc28d2014-09-30 15:46:41 +10007152 goto unlock;
NeilBrownc02c0ae2012-12-11 13:39:21 +11007153 }
NeilBrown3adc28d2014-09-30 15:46:41 +10007154 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007155 }
7156
7157 /*
7158 * Commands querying/configuring an existing array:
7159 */
NeilBrown32a76272005-06-21 17:17:14 -07007160 /* if we are not initialised yet, only ADD_NEW_DISK, STOP_ARRAY,
NeilBrown3f9d7b02006-12-22 01:11:41 -08007161 * RUN_ARRAY, and GET_ and SET_BITMAP_FILE are allowed */
NeilBrowna17184a2008-02-06 01:39:55 -08007162 if ((!mddev->raid_disks && !mddev->external)
7163 && cmd != ADD_NEW_DISK && cmd != STOP_ARRAY
7164 && cmd != RUN_ARRAY && cmd != SET_BITMAP_FILE
7165 && cmd != GET_BITMAP_FILE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007166 err = -ENODEV;
NeilBrown3adc28d2014-09-30 15:46:41 +10007167 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007168 }
7169
7170 /*
7171 * Commands even a read-only array can execute:
7172 */
NeilBrownc02c0ae2012-12-11 13:39:21 +11007173 switch (cmd) {
NeilBrownc02c0ae2012-12-11 13:39:21 +11007174 case RESTART_ARRAY_RW:
7175 err = restart_array(mddev);
NeilBrown3adc28d2014-09-30 15:46:41 +10007176 goto unlock;
NeilBrownc02c0ae2012-12-11 13:39:21 +11007177
7178 case STOP_ARRAY:
7179 err = do_md_stop(mddev, 0, bdev);
NeilBrown3adc28d2014-09-30 15:46:41 +10007180 goto unlock;
NeilBrownc02c0ae2012-12-11 13:39:21 +11007181
7182 case STOP_ARRAY_RO:
7183 err = md_set_readonly(mddev, bdev);
NeilBrown3adc28d2014-09-30 15:46:41 +10007184 goto unlock;
NeilBrownc02c0ae2012-12-11 13:39:21 +11007185
NeilBrown3ea8929d2013-04-24 11:42:41 +10007186 case HOT_REMOVE_DISK:
7187 err = hot_remove_disk(mddev, new_decode_dev(arg));
NeilBrown3adc28d2014-09-30 15:46:41 +10007188 goto unlock;
NeilBrown3ea8929d2013-04-24 11:42:41 +10007189
NeilBrown7ceb17e2013-04-24 11:42:42 +10007190 case ADD_NEW_DISK:
7191 /* We can support ADD_NEW_DISK on read-only arrays
Wei Fang466ad292016-03-21 19:19:30 +08007192 * only if we are re-adding a preexisting device.
NeilBrown7ceb17e2013-04-24 11:42:42 +10007193 * So require mddev->pers and MD_DISK_SYNC.
7194 */
7195 if (mddev->pers) {
7196 mdu_disk_info_t info;
7197 if (copy_from_user(&info, argp, sizeof(info)))
7198 err = -EFAULT;
7199 else if (!(info.state & (1<<MD_DISK_SYNC)))
7200 /* Need to clear read-only for this */
7201 break;
7202 else
7203 err = add_new_disk(mddev, &info);
NeilBrown3adc28d2014-09-30 15:46:41 +10007204 goto unlock;
NeilBrown7ceb17e2013-04-24 11:42:42 +10007205 }
7206 break;
7207
NeilBrownc02c0ae2012-12-11 13:39:21 +11007208 case BLKROSET:
7209 if (get_user(ro, (int __user *)(arg))) {
7210 err = -EFAULT;
NeilBrown3adc28d2014-09-30 15:46:41 +10007211 goto unlock;
NeilBrownc02c0ae2012-12-11 13:39:21 +11007212 }
7213 err = -EINVAL;
7214
7215 /* if the bdev is going readonly the value of mddev->ro
7216 * does not matter, no writes are coming
7217 */
7218 if (ro)
NeilBrown3adc28d2014-09-30 15:46:41 +10007219 goto unlock;
NeilBrown32a76272005-06-21 17:17:14 -07007220
NeilBrownc02c0ae2012-12-11 13:39:21 +11007221 /* are we are already prepared for writes? */
7222 if (mddev->ro != 1)
NeilBrown3adc28d2014-09-30 15:46:41 +10007223 goto unlock;
NeilBrownc02c0ae2012-12-11 13:39:21 +11007224
7225 /* transitioning to readauto need only happen for
7226 * arrays that call md_write_start
7227 */
7228 if (mddev->pers) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007229 err = restart_array(mddev);
NeilBrownc02c0ae2012-12-11 13:39:21 +11007230 if (err == 0) {
7231 mddev->ro = 2;
7232 set_disk_ro(mddev->gendisk, 0);
Dan Williamse2218352010-05-12 08:25:37 +10007233 }
NeilBrownc02c0ae2012-12-11 13:39:21 +11007234 }
NeilBrown3adc28d2014-09-30 15:46:41 +10007235 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007236 }
7237
7238 /*
7239 * The remaining ioctls are changing the state of the
NeilBrownf91de922005-11-08 21:39:36 -08007240 * superblock, so we do not allow them on read-only arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007241 */
NeilBrown326eb172014-09-30 15:36:28 +10007242 if (mddev->ro && mddev->pers) {
NeilBrownf91de922005-11-08 21:39:36 -08007243 if (mddev->ro == 2) {
7244 mddev->ro = 0;
NeilBrown00bcb4a2010-06-01 19:37:23 +10007245 sysfs_notify_dirent_safe(mddev->sysfs_state);
Neil Brown0fd62b82008-06-28 08:31:36 +10007246 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
NeilBrownf3378b42013-02-28 11:59:03 +11007247 /* mddev_unlock will wake thread */
7248 /* If a device failed while we were read-only, we
7249 * need to make sure the metadata is updated now.
7250 */
Shaohua Li29530792016-12-08 15:48:19 -08007251 if (test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)) {
NeilBrownf3378b42013-02-28 11:59:03 +11007252 mddev_unlock(mddev);
7253 wait_event(mddev->sb_wait,
Shaohua Li29530792016-12-08 15:48:19 -08007254 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags) &&
7255 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
NeilBrown29f097c2013-11-14 17:54:51 +11007256 mddev_lock_nointr(mddev);
NeilBrownf3378b42013-02-28 11:59:03 +11007257 }
NeilBrownf91de922005-11-08 21:39:36 -08007258 } else {
7259 err = -EROFS;
NeilBrown3adc28d2014-09-30 15:46:41 +10007260 goto unlock;
NeilBrownf91de922005-11-08 21:39:36 -08007261 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007262 }
7263
NeilBrownc02c0ae2012-12-11 13:39:21 +11007264 switch (cmd) {
7265 case ADD_NEW_DISK:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007266 {
NeilBrownc02c0ae2012-12-11 13:39:21 +11007267 mdu_disk_info_t info;
7268 if (copy_from_user(&info, argp, sizeof(info)))
7269 err = -EFAULT;
7270 else
7271 err = add_new_disk(mddev, &info);
NeilBrown3adc28d2014-09-30 15:46:41 +10007272 goto unlock;
NeilBrownc02c0ae2012-12-11 13:39:21 +11007273 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007274
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05007275 case CLUSTERED_DISK_NACK:
7276 if (mddev_is_clustered(mddev))
7277 md_cluster_ops->new_disk_ack(mddev, false);
7278 else
7279 err = -EINVAL;
7280 goto unlock;
7281
NeilBrownc02c0ae2012-12-11 13:39:21 +11007282 case HOT_ADD_DISK:
7283 err = hot_add_disk(mddev, new_decode_dev(arg));
NeilBrown3adc28d2014-09-30 15:46:41 +10007284 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007285
NeilBrownc02c0ae2012-12-11 13:39:21 +11007286 case RUN_ARRAY:
7287 err = do_md_run(mddev);
NeilBrown3adc28d2014-09-30 15:46:41 +10007288 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007289
NeilBrownc02c0ae2012-12-11 13:39:21 +11007290 case SET_BITMAP_FILE:
7291 err = set_bitmap_file(mddev, (int)arg);
NeilBrown3adc28d2014-09-30 15:46:41 +10007292 goto unlock;
NeilBrown32a76272005-06-21 17:17:14 -07007293
NeilBrownc02c0ae2012-12-11 13:39:21 +11007294 default:
7295 err = -EINVAL;
NeilBrown3adc28d2014-09-30 15:46:41 +10007296 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007297 }
7298
NeilBrown3adc28d2014-09-30 15:46:41 +10007299unlock:
NeilBrownd3374822009-01-09 08:31:10 +11007300 if (mddev->hold_active == UNTIL_IOCTL &&
7301 err != -EINVAL)
7302 mddev->hold_active = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007303 mddev_unlock(mddev);
NeilBrown3adc28d2014-09-30 15:46:41 +10007304out:
NeilBrown065e5192017-04-06 11:16:33 +08007305 if(did_set_md_closing)
7306 clear_bit(MD_CLOSING, &mddev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007307 return err;
7308}
Arnd Bergmannaa98aa32009-12-14 12:50:05 +11007309#ifdef CONFIG_COMPAT
7310static int md_compat_ioctl(struct block_device *bdev, fmode_t mode,
7311 unsigned int cmd, unsigned long arg)
7312{
7313 switch (cmd) {
7314 case HOT_REMOVE_DISK:
7315 case HOT_ADD_DISK:
7316 case SET_DISK_FAULTY:
7317 case SET_BITMAP_FILE:
7318 /* These take in integer arg, do not convert */
7319 break;
7320 default:
7321 arg = (unsigned long)compat_ptr(arg);
7322 break;
7323 }
7324
7325 return md_ioctl(bdev, mode, cmd, arg);
7326}
7327#endif /* CONFIG_COMPAT */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007328
Al Viroa39907f2008-03-02 10:31:15 -05007329static int md_open(struct block_device *bdev, fmode_t mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007330{
7331 /*
7332 * Succeed if we can lock the mddev, which confirms that
7333 * it isn't being stopped right now.
7334 */
NeilBrownfd01b882011-10-11 16:47:53 +11007335 struct mddev *mddev = mddev_find(bdev->bd_dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007336 int err;
7337
Yuanhan Liu0c098222012-05-22 13:55:32 +10007338 if (!mddev)
7339 return -ENODEV;
7340
NeilBrownd3374822009-01-09 08:31:10 +11007341 if (mddev->gendisk != bdev->bd_disk) {
7342 /* we are racing with mddev_put which is discarding this
7343 * bd_disk.
7344 */
7345 mddev_put(mddev);
7346 /* Wait until bdev->bd_disk is definitely gone */
Tejun Heoe804ac72010-10-15 15:36:08 +02007347 flush_workqueue(md_misc_wq);
NeilBrownd3374822009-01-09 08:31:10 +11007348 /* Then retry the open from the top */
7349 return -ERESTARTSYS;
7350 }
7351 BUG_ON(mddev != bdev->bd_disk->private_data);
7352
NeilBrownc8c00a62009-08-10 12:50:52 +10007353 if ((err = mutex_lock_interruptible(&mddev->open_mutex)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007354 goto out;
7355
Guoqing Jiangaf8d8e62016-08-12 13:42:37 +08007356 if (test_bit(MD_CLOSING, &mddev->flags)) {
7357 mutex_unlock(&mddev->open_mutex);
NeilBrowne2342ca2016-12-05 16:40:50 +11007358 err = -ENODEV;
7359 goto out;
Guoqing Jiangaf8d8e62016-08-12 13:42:37 +08007360 }
7361
Linus Torvalds1da177e2005-04-16 15:20:36 -07007362 err = 0;
NeilBrownf2ea68c2008-07-21 17:05:25 +10007363 atomic_inc(&mddev->openers);
NeilBrownc8c00a62009-08-10 12:50:52 +10007364 mutex_unlock(&mddev->open_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007365
NeilBrownf0b4f7e2011-02-24 17:26:41 +11007366 check_disk_change(bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007367 out:
NeilBrowne2342ca2016-12-05 16:40:50 +11007368 if (err)
7369 mddev_put(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007370 return err;
7371}
7372
Al Virodb2a1442013-05-05 21:52:57 -04007373static void md_release(struct gendisk *disk, fmode_t mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007374{
NeilBrownf72ffdd2014-09-30 14:23:59 +10007375 struct mddev *mddev = disk->private_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007376
Eric Sesterhenn52e5f9d2006-10-03 23:33:23 +02007377 BUG_ON(!mddev);
NeilBrownf2ea68c2008-07-21 17:05:25 +10007378 atomic_dec(&mddev->openers);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007379 mddev_put(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007380}
NeilBrownf0b4f7e2011-02-24 17:26:41 +11007381
7382static int md_media_changed(struct gendisk *disk)
7383{
NeilBrownfd01b882011-10-11 16:47:53 +11007384 struct mddev *mddev = disk->private_data;
NeilBrownf0b4f7e2011-02-24 17:26:41 +11007385
7386 return mddev->changed;
7387}
7388
7389static int md_revalidate(struct gendisk *disk)
7390{
NeilBrownfd01b882011-10-11 16:47:53 +11007391 struct mddev *mddev = disk->private_data;
NeilBrownf0b4f7e2011-02-24 17:26:41 +11007392
7393 mddev->changed = 0;
7394 return 0;
7395}
Alexey Dobriyan83d5cde2009-09-21 17:01:13 -07007396static const struct block_device_operations md_fops =
Linus Torvalds1da177e2005-04-16 15:20:36 -07007397{
7398 .owner = THIS_MODULE,
Al Viroa39907f2008-03-02 10:31:15 -05007399 .open = md_open,
7400 .release = md_release,
NeilBrownb492b852009-05-26 12:57:36 +10007401 .ioctl = md_ioctl,
Arnd Bergmannaa98aa32009-12-14 12:50:05 +11007402#ifdef CONFIG_COMPAT
7403 .compat_ioctl = md_compat_ioctl,
7404#endif
Christoph Hellwiga885c8c2006-01-08 01:02:50 -08007405 .getgeo = md_getgeo,
NeilBrownf0b4f7e2011-02-24 17:26:41 +11007406 .media_changed = md_media_changed,
7407 .revalidate_disk= md_revalidate,
Linus Torvalds1da177e2005-04-16 15:20:36 -07007408};
7409
NeilBrownf72ffdd2014-09-30 14:23:59 +10007410static int md_thread(void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007411{
NeilBrown2b8bf342011-10-11 16:48:23 +11007412 struct md_thread *thread = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007413
Linus Torvalds1da177e2005-04-16 15:20:36 -07007414 /*
7415 * md_thread is a 'system-thread', it's priority should be very
7416 * high. We avoid resource deadlocks individually in each
7417 * raid personality. (RAID5 does preallocation) We also use RR and
7418 * the very same RT priority as kswapd, thus we will never get
7419 * into a priority inversion deadlock.
7420 *
7421 * we definitely have to have equal or higher priority than
7422 * bdflush, otherwise bdflush will deadlock if there are too
7423 * many dirty RAID5 blocks.
7424 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007425
NeilBrown6985c432005-10-19 21:23:47 -07007426 allow_signal(SIGKILL);
NeilBrowna6fb0932005-09-09 16:23:56 -07007427 while (!kthread_should_stop()) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007428
NeilBrown93588e22005-11-15 00:09:12 -08007429 /* We need to wait INTERRUPTIBLE so that
7430 * we don't add to the load-average.
7431 * That means we need to be sure no signals are
7432 * pending
7433 */
7434 if (signal_pending(current))
7435 flush_signals(current);
7436
7437 wait_event_interruptible_timeout
7438 (thread->wqueue,
7439 test_bit(THREAD_WAKEUP, &thread->flags)
Shaohua Lice1ccd02016-11-21 10:29:18 -08007440 || kthread_should_stop() || kthread_should_park(),
NeilBrown93588e22005-11-15 00:09:12 -08007441 thread->timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007442
NeilBrown6c987912011-01-14 09:13:53 +11007443 clear_bit(THREAD_WAKEUP, &thread->flags);
Shaohua Lice1ccd02016-11-21 10:29:18 -08007444 if (kthread_should_park())
7445 kthread_parkme();
NeilBrown6c987912011-01-14 09:13:53 +11007446 if (!kthread_should_stop())
Shaohua Li4ed87312012-10-11 13:34:00 +11007447 thread->run(thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007448 }
NeilBrowna6fb0932005-09-09 16:23:56 -07007449
Linus Torvalds1da177e2005-04-16 15:20:36 -07007450 return 0;
7451}
7452
NeilBrown2b8bf342011-10-11 16:48:23 +11007453void md_wakeup_thread(struct md_thread *thread)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007454{
7455 if (thread) {
NeilBrown36a4e1f2011-10-07 14:23:17 +11007456 pr_debug("md: waking up MD thread %s.\n", thread->tsk->comm);
NeilBrown4ad23a972017-03-15 14:05:14 +11007457 if (!test_and_set_bit(THREAD_WAKEUP, &thread->flags))
7458 wake_up(&thread->wqueue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007459 }
7460}
NeilBrown6c144d32014-09-30 16:15:38 +10007461EXPORT_SYMBOL(md_wakeup_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007462
Shaohua Li4ed87312012-10-11 13:34:00 +11007463struct md_thread *md_register_thread(void (*run) (struct md_thread *),
7464 struct mddev *mddev, const char *name)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007465{
NeilBrown2b8bf342011-10-11 16:48:23 +11007466 struct md_thread *thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007467
NeilBrown2b8bf342011-10-11 16:48:23 +11007468 thread = kzalloc(sizeof(struct md_thread), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007469 if (!thread)
7470 return NULL;
7471
Linus Torvalds1da177e2005-04-16 15:20:36 -07007472 init_waitqueue_head(&thread->wqueue);
7473
Linus Torvalds1da177e2005-04-16 15:20:36 -07007474 thread->run = run;
7475 thread->mddev = mddev;
NeilBrown32a76272005-06-21 17:17:14 -07007476 thread->timeout = MAX_SCHEDULE_TIMEOUT;
NeilBrown0da3c612009-09-23 18:09:45 +10007477 thread->tsk = kthread_run(md_thread, thread,
7478 "%s_%s",
7479 mdname(thread->mddev),
NeilBrown02326052012-07-03 15:56:52 +10007480 name);
NeilBrowna6fb0932005-09-09 16:23:56 -07007481 if (IS_ERR(thread->tsk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007482 kfree(thread);
7483 return NULL;
7484 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007485 return thread;
7486}
NeilBrown6c144d32014-09-30 16:15:38 +10007487EXPORT_SYMBOL(md_register_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007488
NeilBrown2b8bf342011-10-11 16:48:23 +11007489void md_unregister_thread(struct md_thread **threadp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007490{
NeilBrown2b8bf342011-10-11 16:48:23 +11007491 struct md_thread *thread = *threadp;
NeilBrowne0cf8f02009-03-31 14:39:39 +11007492 if (!thread)
7493 return;
NeilBrown36a4e1f2011-10-07 14:23:17 +11007494 pr_debug("interrupting MD-thread pid %d\n", task_pid_nr(thread->tsk));
NeilBrown01f96c02011-09-21 15:30:20 +10007495 /* Locking ensures that mddev_unlock does not wake_up a
7496 * non-existent thread
7497 */
7498 spin_lock(&pers_lock);
7499 *threadp = NULL;
7500 spin_unlock(&pers_lock);
NeilBrowna6fb0932005-09-09 16:23:56 -07007501
7502 kthread_stop(thread->tsk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007503 kfree(thread);
7504}
NeilBrown6c144d32014-09-30 16:15:38 +10007505EXPORT_SYMBOL(md_unregister_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007506
NeilBrownfd01b882011-10-11 16:47:53 +11007507void md_error(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007508{
NeilBrownb2d444d2005-11-08 21:39:31 -08007509 if (!rdev || test_bit(Faulty, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007510 return;
Dan Williams6bfe0b42008-04-30 00:52:32 -07007511
NeilBrownde393cd2011-07-28 11:31:48 +10007512 if (!mddev->pers || !mddev->pers->error_handler)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007513 return;
7514 mddev->pers->error_handler(mddev,rdev);
Neil Brown72a23c22008-06-28 08:31:41 +10007515 if (mddev->degraded)
7516 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
NeilBrown00bcb4a2010-06-01 19:37:23 +10007517 sysfs_notify_dirent_safe(rdev->sysfs_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007518 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
7519 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7520 md_wakeup_thread(mddev->thread);
NeilBrown768a4182010-07-26 11:49:55 +10007521 if (mddev->event_work.func)
Tejun Heoe804ac72010-10-15 15:36:08 +02007522 queue_work(md_misc_wq, &mddev->event_work);
Guoqing Jiangbb9ef712015-12-28 10:46:38 +08007523 md_new_event(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007524}
NeilBrown6c144d32014-09-30 16:15:38 +10007525EXPORT_SYMBOL(md_error);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007526
7527/* seq_file implementation /proc/mdstat */
7528
7529static void status_unused(struct seq_file *seq)
7530{
7531 int i = 0;
NeilBrown3cb03002011-10-11 16:45:26 +11007532 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007533
7534 seq_printf(seq, "unused devices: ");
7535
Cheng Renquan159ec1f2009-01-09 08:31:08 +11007536 list_for_each_entry(rdev, &pending_raid_disks, same_set) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007537 char b[BDEVNAME_SIZE];
7538 i++;
7539 seq_printf(seq, "%s ",
7540 bdevname(rdev->bdev,b));
7541 }
7542 if (!i)
7543 seq_printf(seq, "<none>");
7544
7545 seq_printf(seq, "\n");
7546}
7547
NeilBrownf7851be2015-07-02 17:12:58 +10007548static int status_resync(struct seq_file *seq, struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007549{
NeilBrowndd71cf62009-05-07 12:49:35 +10007550 sector_t max_sectors, resync, res;
7551 unsigned long dt, db;
7552 sector_t rt;
NeilBrown4588b422006-03-27 01:18:04 -08007553 int scale;
7554 unsigned int per_milli;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007555
NeilBrownc804cde2012-05-21 09:28:33 +10007556 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
7557 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
NeilBrowndd71cf62009-05-07 12:49:35 +10007558 max_sectors = mddev->resync_max_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007559 else
NeilBrowndd71cf62009-05-07 12:49:35 +10007560 max_sectors = mddev->dev_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007561
NeilBrownf7851be2015-07-02 17:12:58 +10007562 resync = mddev->curr_resync;
7563 if (resync <= 3) {
7564 if (test_bit(MD_RECOVERY_DONE, &mddev->recovery))
7565 /* Still cleaning up */
7566 resync = max_sectors;
7567 } else
7568 resync -= atomic_read(&mddev->recovery_active);
7569
7570 if (resync == 0) {
7571 if (mddev->recovery_cp < MaxSector) {
7572 seq_printf(seq, "\tresync=PENDING");
7573 return 1;
7574 }
7575 return 0;
7576 }
7577 if (resync < 3) {
7578 seq_printf(seq, "\tresync=DELAYED");
7579 return 1;
7580 }
7581
NeilBrown403df472014-09-30 15:52:29 +10007582 WARN_ON(max_sectors == 0);
NeilBrown4588b422006-03-27 01:18:04 -08007583 /* Pick 'scale' such that (resync>>scale)*1000 will fit
NeilBrowndd71cf62009-05-07 12:49:35 +10007584 * in a sector_t, and (max_sectors>>scale) will fit in a
NeilBrown4588b422006-03-27 01:18:04 -08007585 * u32, as those are the requirements for sector_div.
7586 * Thus 'scale' must be at least 10
7587 */
7588 scale = 10;
7589 if (sizeof(sector_t) > sizeof(unsigned long)) {
NeilBrowndd71cf62009-05-07 12:49:35 +10007590 while ( max_sectors/2 > (1ULL<<(scale+32)))
NeilBrown4588b422006-03-27 01:18:04 -08007591 scale++;
7592 }
7593 res = (resync>>scale)*1000;
NeilBrowndd71cf62009-05-07 12:49:35 +10007594 sector_div(res, (u32)((max_sectors>>scale)+1));
NeilBrown4588b422006-03-27 01:18:04 -08007595
7596 per_milli = res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007597 {
NeilBrown4588b422006-03-27 01:18:04 -08007598 int i, x = per_milli/50, y = 20-x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007599 seq_printf(seq, "[");
7600 for (i = 0; i < x; i++)
7601 seq_printf(seq, "=");
7602 seq_printf(seq, ">");
7603 for (i = 0; i < y; i++)
7604 seq_printf(seq, ".");
7605 seq_printf(seq, "] ");
7606 }
NeilBrown4588b422006-03-27 01:18:04 -08007607 seq_printf(seq, " %s =%3u.%u%% (%llu/%llu)",
NeilBrownccfcc3c2006-03-27 01:18:09 -08007608 (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)?
7609 "reshape" :
NeilBrown61df9d92006-10-03 01:15:57 -07007610 (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)?
7611 "check" :
7612 (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ?
7613 "resync" : "recovery"))),
7614 per_milli/10, per_milli % 10,
NeilBrowndd71cf62009-05-07 12:49:35 +10007615 (unsigned long long) resync/2,
7616 (unsigned long long) max_sectors/2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007617
7618 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07007619 * dt: time from mark until now
7620 * db: blocks written from mark until now
7621 * rt: remaining time
NeilBrowndd71cf62009-05-07 12:49:35 +10007622 *
7623 * rt is a sector_t, so could be 32bit or 64bit.
7624 * So we divide before multiply in case it is 32bit and close
7625 * to the limit.
Lucas De Marchi25985ed2011-03-30 22:57:33 -03007626 * We scale the divisor (db) by 32 to avoid losing precision
NeilBrowndd71cf62009-05-07 12:49:35 +10007627 * near the end of resync when the number of remaining sectors
7628 * is close to 'db'.
7629 * We then divide rt by 32 after multiplying by db to compensate.
7630 * The '+1' avoids division by zero if db is very small.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007631 */
7632 dt = ((jiffies - mddev->resync_mark) / HZ);
7633 if (!dt) dt++;
NeilBrownff4e8d92006-07-10 04:44:16 -07007634 db = (mddev->curr_mark_cnt - atomic_read(&mddev->recovery_active))
7635 - mddev->resync_mark_cnt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007636
NeilBrowndd71cf62009-05-07 12:49:35 +10007637 rt = max_sectors - resync; /* number of remaining sectors */
7638 sector_div(rt, db/32+1);
7639 rt *= dt;
7640 rt >>= 5;
7641
7642 seq_printf(seq, " finish=%lu.%lumin", (unsigned long)rt / 60,
7643 ((unsigned long)rt % 60)/6);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007644
NeilBrownff4e8d92006-07-10 04:44:16 -07007645 seq_printf(seq, " speed=%ldK/sec", db/2/dt);
NeilBrownf7851be2015-07-02 17:12:58 +10007646 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007647}
7648
7649static void *md_seq_start(struct seq_file *seq, loff_t *pos)
7650{
7651 struct list_head *tmp;
7652 loff_t l = *pos;
NeilBrownfd01b882011-10-11 16:47:53 +11007653 struct mddev *mddev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007654
7655 if (l >= 0x10000)
7656 return NULL;
7657 if (!l--)
7658 /* header */
7659 return (void*)1;
7660
7661 spin_lock(&all_mddevs_lock);
7662 list_for_each(tmp,&all_mddevs)
7663 if (!l--) {
NeilBrownfd01b882011-10-11 16:47:53 +11007664 mddev = list_entry(tmp, struct mddev, all_mddevs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007665 mddev_get(mddev);
7666 spin_unlock(&all_mddevs_lock);
7667 return mddev;
7668 }
7669 spin_unlock(&all_mddevs_lock);
7670 if (!l--)
7671 return (void*)2;/* tail */
7672 return NULL;
7673}
7674
7675static void *md_seq_next(struct seq_file *seq, void *v, loff_t *pos)
7676{
7677 struct list_head *tmp;
NeilBrownfd01b882011-10-11 16:47:53 +11007678 struct mddev *next_mddev, *mddev = v;
NeilBrownf72ffdd2014-09-30 14:23:59 +10007679
Linus Torvalds1da177e2005-04-16 15:20:36 -07007680 ++*pos;
7681 if (v == (void*)2)
7682 return NULL;
7683
7684 spin_lock(&all_mddevs_lock);
7685 if (v == (void*)1)
7686 tmp = all_mddevs.next;
7687 else
7688 tmp = mddev->all_mddevs.next;
7689 if (tmp != &all_mddevs)
NeilBrownfd01b882011-10-11 16:47:53 +11007690 next_mddev = mddev_get(list_entry(tmp,struct mddev,all_mddevs));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007691 else {
7692 next_mddev = (void*)2;
7693 *pos = 0x10000;
NeilBrownf72ffdd2014-09-30 14:23:59 +10007694 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007695 spin_unlock(&all_mddevs_lock);
7696
7697 if (v != (void*)1)
7698 mddev_put(mddev);
7699 return next_mddev;
7700
7701}
7702
7703static void md_seq_stop(struct seq_file *seq, void *v)
7704{
NeilBrownfd01b882011-10-11 16:47:53 +11007705 struct mddev *mddev = v;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007706
7707 if (mddev && v != (void*)1 && v != (void*)2)
7708 mddev_put(mddev);
7709}
7710
7711static int md_seq_show(struct seq_file *seq, void *v)
7712{
NeilBrownfd01b882011-10-11 16:47:53 +11007713 struct mddev *mddev = v;
Andre Nolldd8ac332009-03-31 14:33:13 +11007714 sector_t sectors;
NeilBrown3cb03002011-10-11 16:45:26 +11007715 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007716
7717 if (v == (void*)1) {
NeilBrown84fc4b52011-10-11 16:49:58 +11007718 struct md_personality *pers;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007719 seq_printf(seq, "Personalities : ");
7720 spin_lock(&pers_lock);
NeilBrown2604b702006-01-06 00:20:36 -08007721 list_for_each_entry(pers, &pers_list, list)
7722 seq_printf(seq, "[%s] ", pers->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007723
7724 spin_unlock(&pers_lock);
7725 seq_printf(seq, "\n");
Kay Sieversf1514632011-07-12 20:48:39 +02007726 seq->poll_event = atomic_read(&md_event_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007727 return 0;
7728 }
7729 if (v == (void*)2) {
7730 status_unused(seq);
7731 return 0;
7732 }
7733
NeilBrown36d091f2014-12-15 12:56:58 +11007734 spin_lock(&mddev->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007735 if (mddev->pers || mddev->raid_disks || !list_empty(&mddev->disks)) {
7736 seq_printf(seq, "%s : %sactive", mdname(mddev),
7737 mddev->pers ? "" : "in");
7738 if (mddev->pers) {
NeilBrownf91de922005-11-08 21:39:36 -08007739 if (mddev->ro==1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007740 seq_printf(seq, " (read-only)");
NeilBrownf91de922005-11-08 21:39:36 -08007741 if (mddev->ro==2)
NeilBrown52720ae2008-03-10 11:43:47 -07007742 seq_printf(seq, " (auto-read-only)");
Linus Torvalds1da177e2005-04-16 15:20:36 -07007743 seq_printf(seq, " %s", mddev->pers->name);
7744 }
7745
Andre Nolldd8ac332009-03-31 14:33:13 +11007746 sectors = 0;
NeilBrownf97fcad2014-12-15 12:56:59 +11007747 rcu_read_lock();
7748 rdev_for_each_rcu(rdev, mddev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007749 char b[BDEVNAME_SIZE];
7750 seq_printf(seq, " %s[%d]",
7751 bdevname(rdev->bdev,b), rdev->desc_nr);
NeilBrown8ddf9ef2005-09-09 16:23:45 -07007752 if (test_bit(WriteMostly, &rdev->flags))
7753 seq_printf(seq, "(W)");
Shaohua Li9efdca12015-10-12 16:59:50 -07007754 if (test_bit(Journal, &rdev->flags))
7755 seq_printf(seq, "(J)");
NeilBrownb2d444d2005-11-08 21:39:31 -08007756 if (test_bit(Faulty, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007757 seq_printf(seq, "(F)");
7758 continue;
NeilBrown2d78f8c2011-12-23 10:17:51 +11007759 }
7760 if (rdev->raid_disk < 0)
NeilBrownb325a322005-09-09 16:24:00 -07007761 seq_printf(seq, "(S)"); /* spare */
NeilBrown2d78f8c2011-12-23 10:17:51 +11007762 if (test_bit(Replacement, &rdev->flags))
7763 seq_printf(seq, "(R)");
Andre Nolldd8ac332009-03-31 14:33:13 +11007764 sectors += rdev->sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007765 }
NeilBrownf97fcad2014-12-15 12:56:59 +11007766 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007767
7768 if (!list_empty(&mddev->disks)) {
7769 if (mddev->pers)
7770 seq_printf(seq, "\n %llu blocks",
Andre Nollf233ea52008-07-21 17:05:22 +10007771 (unsigned long long)
7772 mddev->array_sectors / 2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007773 else
7774 seq_printf(seq, "\n %llu blocks",
Andre Nolldd8ac332009-03-31 14:33:13 +11007775 (unsigned long long)sectors / 2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007776 }
NeilBrown1cd6bf12005-09-09 16:24:00 -07007777 if (mddev->persistent) {
7778 if (mddev->major_version != 0 ||
7779 mddev->minor_version != 90) {
7780 seq_printf(seq," super %d.%d",
7781 mddev->major_version,
7782 mddev->minor_version);
7783 }
NeilBrowne6910632008-02-06 01:39:51 -08007784 } else if (mddev->external)
7785 seq_printf(seq, " super external:%s",
7786 mddev->metadata_type);
7787 else
NeilBrown1cd6bf12005-09-09 16:24:00 -07007788 seq_printf(seq, " super non-persistent");
Linus Torvalds1da177e2005-04-16 15:20:36 -07007789
7790 if (mddev->pers) {
NeilBrownd710e132008-10-13 11:55:12 +11007791 mddev->pers->status(seq, mddev);
NeilBrownf72ffdd2014-09-30 14:23:59 +10007792 seq_printf(seq, "\n ");
NeilBrown8e1b39d2005-11-08 21:39:41 -08007793 if (mddev->pers->sync_request) {
NeilBrownf7851be2015-07-02 17:12:58 +10007794 if (status_resync(seq, mddev))
NeilBrown8e1b39d2005-11-08 21:39:41 -08007795 seq_printf(seq, "\n ");
NeilBrown8e1b39d2005-11-08 21:39:41 -08007796 }
NeilBrown32a76272005-06-21 17:17:14 -07007797 } else
7798 seq_printf(seq, "\n ");
7799
NeilBrown57148962012-03-19 12:46:40 +11007800 bitmap_status(seq, mddev->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007801
7802 seq_printf(seq, "\n");
7803 }
NeilBrown36d091f2014-12-15 12:56:58 +11007804 spin_unlock(&mddev->lock);
NeilBrownf72ffdd2014-09-30 14:23:59 +10007805
Linus Torvalds1da177e2005-04-16 15:20:36 -07007806 return 0;
7807}
7808
Jan Engelhardt110518b2009-05-07 12:49:37 +10007809static const struct seq_operations md_seq_ops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007810 .start = md_seq_start,
7811 .next = md_seq_next,
7812 .stop = md_seq_stop,
7813 .show = md_seq_show,
7814};
7815
7816static int md_seq_open(struct inode *inode, struct file *file)
7817{
Kay Sieversf1514632011-07-12 20:48:39 +02007818 struct seq_file *seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007819 int error;
7820
7821 error = seq_open(file, &md_seq_ops);
NeilBrownd7603b72006-01-06 00:20:30 -08007822 if (error)
Kay Sieversf1514632011-07-12 20:48:39 +02007823 return error;
7824
7825 seq = file->private_data;
7826 seq->poll_event = atomic_read(&md_event_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007827 return error;
7828}
7829
NeilBrowne2f23b62014-04-09 14:33:51 +10007830static int md_unloading;
NeilBrownd7603b72006-01-06 00:20:30 -08007831static unsigned int mdstat_poll(struct file *filp, poll_table *wait)
7832{
Kay Sieversf1514632011-07-12 20:48:39 +02007833 struct seq_file *seq = filp->private_data;
NeilBrownd7603b72006-01-06 00:20:30 -08007834 int mask;
7835
NeilBrowne2f23b62014-04-09 14:33:51 +10007836 if (md_unloading)
kbuild test robot7d7e64f2014-12-03 16:07:59 +11007837 return POLLIN|POLLRDNORM|POLLERR|POLLPRI;
NeilBrownd7603b72006-01-06 00:20:30 -08007838 poll_wait(filp, &md_event_waiters, wait);
7839
7840 /* always allow read */
7841 mask = POLLIN | POLLRDNORM;
7842
Kay Sieversf1514632011-07-12 20:48:39 +02007843 if (seq->poll_event != atomic_read(&md_event_count))
NeilBrownd7603b72006-01-06 00:20:30 -08007844 mask |= POLLERR | POLLPRI;
7845 return mask;
7846}
7847
Arjan van de Venfa027c22007-02-12 00:55:33 -08007848static const struct file_operations md_seq_fops = {
Akinobu Mitae24650c2006-10-17 00:09:38 -07007849 .owner = THIS_MODULE,
Linus Torvalds1da177e2005-04-16 15:20:36 -07007850 .open = md_seq_open,
7851 .read = seq_read,
7852 .llseek = seq_lseek,
Cihangir Akturk26e13042017-07-29 19:52:45 +03007853 .release = seq_release,
NeilBrownd7603b72006-01-06 00:20:30 -08007854 .poll = mdstat_poll,
Linus Torvalds1da177e2005-04-16 15:20:36 -07007855};
7856
NeilBrown84fc4b52011-10-11 16:49:58 +11007857int register_md_personality(struct md_personality *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007858{
NeilBrown9d487392016-11-02 14:16:49 +11007859 pr_debug("md: %s personality registered for level %d\n",
7860 p->name, p->level);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007861 spin_lock(&pers_lock);
NeilBrown2604b702006-01-06 00:20:36 -08007862 list_add_tail(&p->list, &pers_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007863 spin_unlock(&pers_lock);
7864 return 0;
7865}
NeilBrown6c144d32014-09-30 16:15:38 +10007866EXPORT_SYMBOL(register_md_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007867
NeilBrown84fc4b52011-10-11 16:49:58 +11007868int unregister_md_personality(struct md_personality *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007869{
NeilBrown9d487392016-11-02 14:16:49 +11007870 pr_debug("md: %s personality unregistered\n", p->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007871 spin_lock(&pers_lock);
NeilBrown2604b702006-01-06 00:20:36 -08007872 list_del_init(&p->list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007873 spin_unlock(&pers_lock);
7874 return 0;
7875}
NeilBrown6c144d32014-09-30 16:15:38 +10007876EXPORT_SYMBOL(unregister_md_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007877
NeilBrown6022e752015-08-13 12:32:55 +10007878int register_md_cluster_operations(struct md_cluster_operations *ops,
7879 struct module *module)
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05007880{
NeilBrown6022e752015-08-13 12:32:55 +10007881 int ret = 0;
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05007882 spin_lock(&pers_lock);
NeilBrown6022e752015-08-13 12:32:55 +10007883 if (md_cluster_ops != NULL)
7884 ret = -EALREADY;
7885 else {
7886 md_cluster_ops = ops;
7887 md_cluster_mod = module;
7888 }
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05007889 spin_unlock(&pers_lock);
NeilBrown6022e752015-08-13 12:32:55 +10007890 return ret;
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05007891}
7892EXPORT_SYMBOL(register_md_cluster_operations);
7893
7894int unregister_md_cluster_operations(void)
7895{
7896 spin_lock(&pers_lock);
7897 md_cluster_ops = NULL;
7898 spin_unlock(&pers_lock);
7899 return 0;
7900}
7901EXPORT_SYMBOL(unregister_md_cluster_operations);
7902
7903int md_setup_cluster(struct mddev *mddev, int nodes)
7904{
Guoqing Jiang47a7b0d2016-09-04 22:17:28 -04007905 if (!md_cluster_ops)
7906 request_module("md-cluster");
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05007907 spin_lock(&pers_lock);
Guoqing Jiang47a7b0d2016-09-04 22:17:28 -04007908 /* ensure module won't be unloaded */
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05007909 if (!md_cluster_ops || !try_module_get(md_cluster_mod)) {
NeilBrown9d487392016-11-02 14:16:49 +11007910 pr_warn("can't find md-cluster module or get it's reference.\n");
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05007911 spin_unlock(&pers_lock);
7912 return -ENOENT;
7913 }
7914 spin_unlock(&pers_lock);
7915
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -05007916 return md_cluster_ops->join(mddev, nodes);
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05007917}
7918
7919void md_cluster_stop(struct mddev *mddev)
7920{
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -05007921 if (!md_cluster_ops)
7922 return;
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05007923 md_cluster_ops->leave(mddev);
7924 module_put(md_cluster_mod);
7925}
7926
NeilBrownfd01b882011-10-11 16:47:53 +11007927static int is_mddev_idle(struct mddev *mddev, int init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007928{
NeilBrownf72ffdd2014-09-30 14:23:59 +10007929 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007930 int idle;
NeilBrowneea1bf32009-03-31 14:27:02 +11007931 int curr_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007932
7933 idle = 1;
NeilBrown4b809912008-07-21 17:05:25 +10007934 rcu_read_lock();
7935 rdev_for_each_rcu(rdev, mddev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007936 struct gendisk *disk = rdev->bdev->bd_contains->bd_disk;
NeilBrowneea1bf32009-03-31 14:27:02 +11007937 curr_events = (int)part_stat_read(&disk->part0, sectors[0]) +
7938 (int)part_stat_read(&disk->part0, sectors[1]) -
7939 atomic_read(&disk->sync_io);
NeilBrown713f6ab2007-07-17 04:06:12 -07007940 /* sync IO will cause sync_io to increase before the disk_stats
7941 * as sync_io is counted when a request starts, and
7942 * disk_stats is counted when it completes.
7943 * So resync activity will cause curr_events to be smaller than
7944 * when there was no such activity.
7945 * non-sync IO will cause disk_stat to increase without
7946 * increasing sync_io so curr_events will (eventually)
7947 * be larger than it was before. Once it becomes
7948 * substantially larger, the test below will cause
7949 * the array to appear non-idle, and resync will slow
7950 * down.
7951 * If there is a lot of outstanding resync activity when
7952 * we set last_event to curr_events, then all that activity
7953 * completing might cause the array to appear non-idle
7954 * and resync will be slowed down even though there might
7955 * not have been non-resync activity. This will only
7956 * happen once though. 'last_events' will soon reflect
7957 * the state where there is little or no outstanding
7958 * resync requests, and further resync activity will
7959 * always make curr_events less than last_events.
NeilBrownc0e48522005-11-18 01:11:01 -08007960 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07007961 */
NeilBrowneea1bf32009-03-31 14:27:02 +11007962 if (init || curr_events - rdev->last_events > 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007963 rdev->last_events = curr_events;
7964 idle = 0;
7965 }
7966 }
NeilBrown4b809912008-07-21 17:05:25 +10007967 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007968 return idle;
7969}
7970
NeilBrownfd01b882011-10-11 16:47:53 +11007971void md_done_sync(struct mddev *mddev, int blocks, int ok)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007972{
7973 /* another "blocks" (512byte) blocks have been synced */
7974 atomic_sub(blocks, &mddev->recovery_active);
7975 wake_up(&mddev->recovery_wait);
7976 if (!ok) {
NeilBrowndfc70642008-05-23 13:04:39 -07007977 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
majianpeng0a19caa2012-11-19 19:57:34 +08007978 set_bit(MD_RECOVERY_ERROR, &mddev->recovery);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007979 md_wakeup_thread(mddev->thread);
7980 // stop recovery, signal do_sync ....
7981 }
7982}
NeilBrown6c144d32014-09-30 16:15:38 +10007983EXPORT_SYMBOL(md_done_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007984
NeilBrown06d91a52005-06-21 17:17:12 -07007985/* md_write_start(mddev, bi)
7986 * If we need to update some array metadata (e.g. 'active' flag
NeilBrown3d310eb2005-06-21 17:17:26 -07007987 * in superblock) before writing, schedule a superblock update
7988 * and wait for it to complete.
NeilBrowncc27b0c2017-06-05 16:49:39 +10007989 * A return value of 'false' means that the write wasn't recorded
7990 * and cannot proceed as the array is being suspend.
NeilBrown06d91a52005-06-21 17:17:12 -07007991 */
NeilBrowncc27b0c2017-06-05 16:49:39 +10007992bool md_write_start(struct mddev *mddev, struct bio *bi)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007993{
Neil Brown0fd62b82008-06-28 08:31:36 +10007994 int did_change = 0;
NeilBrown06d91a52005-06-21 17:17:12 -07007995 if (bio_data_dir(bi) != WRITE)
NeilBrowncc27b0c2017-06-05 16:49:39 +10007996 return true;
NeilBrown06d91a52005-06-21 17:17:12 -07007997
NeilBrownf91de922005-11-08 21:39:36 -08007998 BUG_ON(mddev->ro == 1);
7999 if (mddev->ro == 2) {
8000 /* need to switch to read/write */
8001 mddev->ro = 0;
8002 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8003 md_wakeup_thread(mddev->thread);
NeilBrown25156192008-03-04 14:29:32 -08008004 md_wakeup_thread(mddev->sync_thread);
Neil Brown0fd62b82008-06-28 08:31:36 +10008005 did_change = 1;
NeilBrownf91de922005-11-08 21:39:36 -08008006 }
NeilBrown4ad23a972017-03-15 14:05:14 +11008007 rcu_read_lock();
8008 percpu_ref_get(&mddev->writes_pending);
NeilBrown55cc39f2017-03-15 14:05:14 +11008009 smp_mb(); /* Match smp_mb in set_in_sync() */
NeilBrown31a59e32008-04-30 00:52:30 -07008010 if (mddev->safemode == 1)
8011 mddev->safemode = 0;
NeilBrown4ad23a972017-03-15 14:05:14 +11008012 /* sync_checkers is always 0 when writes_pending is in per-cpu mode */
NeilBrown81fe48e2017-08-08 16:56:36 +10008013 if (mddev->in_sync || mddev->sync_checkers) {
NeilBrown85572d72014-12-15 12:56:56 +11008014 spin_lock(&mddev->lock);
NeilBrown3d310eb2005-06-21 17:17:26 -07008015 if (mddev->in_sync) {
8016 mddev->in_sync = 0;
Shaohua Li29530792016-12-08 15:48:19 -08008017 set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
8018 set_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
NeilBrown3d310eb2005-06-21 17:17:26 -07008019 md_wakeup_thread(mddev->thread);
Neil Brown0fd62b82008-06-28 08:31:36 +10008020 did_change = 1;
NeilBrown3d310eb2005-06-21 17:17:26 -07008021 }
NeilBrown85572d72014-12-15 12:56:56 +11008022 spin_unlock(&mddev->lock);
NeilBrown06d91a52005-06-21 17:17:12 -07008023 }
NeilBrown4ad23a972017-03-15 14:05:14 +11008024 rcu_read_unlock();
Neil Brown0fd62b82008-06-28 08:31:36 +10008025 if (did_change)
NeilBrown00bcb4a2010-06-01 19:37:23 +10008026 sysfs_notify_dirent_safe(mddev->sysfs_state);
NeilBrown09a44cc2008-05-23 13:04:36 -07008027 wait_event(mddev->sb_wait,
NeilBrowncc27b0c2017-06-05 16:49:39 +10008028 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags) && !mddev->suspended);
8029 if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
8030 percpu_ref_put(&mddev->writes_pending);
8031 return false;
8032 }
8033 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008034}
NeilBrown6c144d32014-09-30 16:15:38 +10008035EXPORT_SYMBOL(md_write_start);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008036
NeilBrown49728052017-03-15 14:05:12 +11008037/* md_write_inc can only be called when md_write_start() has
8038 * already been called at least once of the current request.
8039 * It increments the counter and is useful when a single request
8040 * is split into several parts. Each part causes an increment and
8041 * so needs a matching md_write_end().
8042 * Unlike md_write_start(), it is safe to call md_write_inc() inside
8043 * a spinlocked region.
8044 */
8045void md_write_inc(struct mddev *mddev, struct bio *bi)
8046{
8047 if (bio_data_dir(bi) != WRITE)
8048 return;
8049 WARN_ON_ONCE(mddev->in_sync || mddev->ro);
NeilBrown4ad23a972017-03-15 14:05:14 +11008050 percpu_ref_get(&mddev->writes_pending);
NeilBrown49728052017-03-15 14:05:12 +11008051}
8052EXPORT_SYMBOL(md_write_inc);
8053
NeilBrownfd01b882011-10-11 16:47:53 +11008054void md_write_end(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008055{
NeilBrown4ad23a972017-03-15 14:05:14 +11008056 percpu_ref_put(&mddev->writes_pending);
8057
8058 if (mddev->safemode == 2)
8059 md_wakeup_thread(mddev->thread);
8060 else if (mddev->safemode_delay)
8061 /* The roundup() ensures this only performs locking once
8062 * every ->safemode_delay jiffies
8063 */
8064 mod_timer(&mddev->safemode_timer,
8065 roundup(jiffies, mddev->safemode_delay) +
8066 mddev->safemode_delay);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008067}
NeilBrown4ad23a972017-03-15 14:05:14 +11008068
NeilBrown6c144d32014-09-30 16:15:38 +10008069EXPORT_SYMBOL(md_write_end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008070
NeilBrown2a2275d2007-01-26 00:57:11 -08008071/* md_allow_write(mddev)
8072 * Calling this ensures that the array is marked 'active' so that writes
8073 * may proceed without blocking. It is important to call this before
8074 * attempting a GFP_KERNEL allocation while holding the mddev lock.
8075 * Must be called with mddev_lock held.
8076 */
Artur Paszkiewicz2214c262017-05-08 11:56:55 +02008077void md_allow_write(struct mddev *mddev)
NeilBrown2a2275d2007-01-26 00:57:11 -08008078{
8079 if (!mddev->pers)
Artur Paszkiewicz2214c262017-05-08 11:56:55 +02008080 return;
NeilBrown2a2275d2007-01-26 00:57:11 -08008081 if (mddev->ro)
Artur Paszkiewicz2214c262017-05-08 11:56:55 +02008082 return;
Neil Brown1a0fd492008-06-28 08:31:27 +10008083 if (!mddev->pers->sync_request)
Artur Paszkiewicz2214c262017-05-08 11:56:55 +02008084 return;
NeilBrown2a2275d2007-01-26 00:57:11 -08008085
NeilBrown85572d72014-12-15 12:56:56 +11008086 spin_lock(&mddev->lock);
NeilBrown2a2275d2007-01-26 00:57:11 -08008087 if (mddev->in_sync) {
8088 mddev->in_sync = 0;
Shaohua Li29530792016-12-08 15:48:19 -08008089 set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
8090 set_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
NeilBrown2a2275d2007-01-26 00:57:11 -08008091 if (mddev->safemode_delay &&
8092 mddev->safemode == 0)
8093 mddev->safemode = 1;
NeilBrown85572d72014-12-15 12:56:56 +11008094 spin_unlock(&mddev->lock);
NeilBrown2a2275d2007-01-26 00:57:11 -08008095 md_update_sb(mddev, 0);
NeilBrown00bcb4a2010-06-01 19:37:23 +10008096 sysfs_notify_dirent_safe(mddev->sysfs_state);
Artur Paszkiewicz2214c262017-05-08 11:56:55 +02008097 /* wait for the dirty state to be recorded in the metadata */
8098 wait_event(mddev->sb_wait,
8099 !test_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags) &&
8100 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
NeilBrown2a2275d2007-01-26 00:57:11 -08008101 } else
NeilBrown85572d72014-12-15 12:56:56 +11008102 spin_unlock(&mddev->lock);
NeilBrown2a2275d2007-01-26 00:57:11 -08008103}
8104EXPORT_SYMBOL_GPL(md_allow_write);
8105
Linus Torvalds1da177e2005-04-16 15:20:36 -07008106#define SYNC_MARKS 10
8107#define SYNC_MARK_STEP (3*HZ)
majianpeng54f89342012-10-31 11:59:10 +11008108#define UPDATE_FREQUENCY (5*60*HZ)
Shaohua Li4ed87312012-10-11 13:34:00 +11008109void md_do_sync(struct md_thread *thread)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008110{
Shaohua Li4ed87312012-10-11 13:34:00 +11008111 struct mddev *mddev = thread->mddev;
NeilBrownfd01b882011-10-11 16:47:53 +11008112 struct mddev *mddev2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008113 unsigned int currspeed = 0,
8114 window;
Xiao Niac7e50a2014-08-07 09:37:41 -04008115 sector_t max_sectors,j, io_sectors, recovery_done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008116 unsigned long mark[SYNC_MARKS];
majianpeng54f89342012-10-31 11:59:10 +11008117 unsigned long update_time;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008118 sector_t mark_cnt[SYNC_MARKS];
8119 int last_mark,m;
8120 struct list_head *tmp;
8121 sector_t last_check;
NeilBrown57afd892005-06-21 17:17:13 -07008122 int skipped = 0;
NeilBrown3cb03002011-10-11 16:45:26 +11008123 struct md_rdev *rdev;
Jonathan Brassowc4a39552013-06-25 01:23:59 -05008124 char *desc, *action = NULL;
majianpeng7c2c57c2012-07-03 12:12:26 +10008125 struct blk_plug plug;
Guoqing Jiang41a9a0d2016-05-02 11:33:08 -04008126 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008127
8128 /* just incase thread restarts... */
8129 if (test_bit(MD_RECOVERY_DONE, &mddev->recovery))
8130 return;
NeilBrown3991b312014-05-28 13:39:23 +10008131 if (mddev->ro) {/* never try to sync a read-only array */
8132 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008133 return;
NeilBrown3991b312014-05-28 13:39:23 +10008134 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008135
Guoqing Jiang41a9a0d2016-05-02 11:33:08 -04008136 if (mddev_is_clustered(mddev)) {
8137 ret = md_cluster_ops->resync_start(mddev);
8138 if (ret)
8139 goto skip;
8140
Guoqing Jiangbb8bf152016-06-02 23:32:04 -04008141 set_bit(MD_CLUSTER_RESYNC_LOCKED, &mddev->flags);
Guoqing Jiang41a9a0d2016-05-02 11:33:08 -04008142 if (!(test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
8143 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) ||
8144 test_bit(MD_RECOVERY_RECOVER, &mddev->recovery))
8145 && ((unsigned long long)mddev->curr_resync_completed
8146 < (unsigned long long)mddev->resync_max_sectors))
8147 goto skip;
8148 }
8149
NeilBrown61df9d92006-10-03 01:15:57 -07008150 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
Jonathan Brassowc4a39552013-06-25 01:23:59 -05008151 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
NeilBrown61df9d92006-10-03 01:15:57 -07008152 desc = "data-check";
Jonathan Brassowc4a39552013-06-25 01:23:59 -05008153 action = "check";
8154 } else if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
NeilBrown61df9d92006-10-03 01:15:57 -07008155 desc = "requested-resync";
Jonathan Brassowc4a39552013-06-25 01:23:59 -05008156 action = "repair";
8157 } else
NeilBrown61df9d92006-10-03 01:15:57 -07008158 desc = "resync";
8159 } else if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
8160 desc = "reshape";
8161 else
8162 desc = "recovery";
8163
Jonathan Brassowc4a39552013-06-25 01:23:59 -05008164 mddev->last_sync_action = action ?: desc;
8165
Linus Torvalds1da177e2005-04-16 15:20:36 -07008166 /* we overload curr_resync somewhat here.
8167 * 0 == not engaged in resync at all
8168 * 2 == checking that there is no conflict with another sync
8169 * 1 == like 2, but have yielded to allow conflicting resync to
8170 * commense
8171 * other == active in resync - this many blocks
8172 *
8173 * Before starting a resync we must have set curr_resync to
8174 * 2, and then checked that every "conflicting" array has curr_resync
8175 * less than ours. When we find one that is the same or higher
8176 * we wait on resync_wait. To avoid deadlock, we reduce curr_resync
8177 * to 1 if we choose to yield (based arbitrarily on address of mddev structure).
8178 * This will mean we have to start checking from the beginning again.
8179 *
8180 */
8181
8182 do {
Artur Paszkiewiczc622ca52016-08-16 14:26:08 +02008183 int mddev2_minor = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008184 mddev->curr_resync = 2;
8185
8186 try_again:
NeilBrown404e4b42009-12-30 15:25:23 +11008187 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
Linus Torvalds1da177e2005-04-16 15:20:36 -07008188 goto skip;
NeilBrown29ac4aa2008-02-06 01:39:58 -08008189 for_each_mddev(mddev2, tmp) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008190 if (mddev2 == mddev)
8191 continue;
Bernd Schubert90b08712008-05-23 13:04:38 -07008192 if (!mddev->parallel_resync
8193 && mddev2->curr_resync
8194 && match_mddev_units(mddev, mddev2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008195 DEFINE_WAIT(wq);
8196 if (mddev < mddev2 && mddev->curr_resync == 2) {
8197 /* arbitrarily yield */
8198 mddev->curr_resync = 1;
8199 wake_up(&resync_wait);
8200 }
8201 if (mddev > mddev2 && mddev->curr_resync == 1)
8202 /* no need to wait here, we can wait the next
8203 * time 'round when curr_resync == 2
8204 */
8205 continue;
NeilBrown97441972008-09-19 11:49:54 +10008206 /* We need to wait 'interruptible' so as not to
8207 * contribute to the load average, and not to
8208 * be caught by 'softlockup'
8209 */
8210 prepare_to_wait(&resync_wait, &wq, TASK_INTERRUPTIBLE);
NeilBrownc91abf52013-11-19 12:02:01 +11008211 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery) &&
NeilBrown8712e552005-10-26 01:58:58 -07008212 mddev2->curr_resync >= mddev->curr_resync) {
Artur Paszkiewiczc622ca52016-08-16 14:26:08 +02008213 if (mddev2_minor != mddev2->md_minor) {
8214 mddev2_minor = mddev2->md_minor;
NeilBrown9d487392016-11-02 14:16:49 +11008215 pr_info("md: delaying %s of %s until %s has finished (they share one or more physical units)\n",
8216 desc, mdname(mddev),
8217 mdname(mddev2));
Artur Paszkiewiczc622ca52016-08-16 14:26:08 +02008218 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008219 mddev_put(mddev2);
NeilBrown97441972008-09-19 11:49:54 +10008220 if (signal_pending(current))
8221 flush_signals(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008222 schedule();
8223 finish_wait(&resync_wait, &wq);
8224 goto try_again;
8225 }
8226 finish_wait(&resync_wait, &wq);
8227 }
8228 }
8229 } while (mddev->curr_resync < 2);
8230
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008231 j = 0;
NeilBrown9d888832005-11-08 21:39:26 -08008232 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008233 /* resync follows the size requested by the personality,
NeilBrown57afd892005-06-21 17:17:13 -07008234 * which defaults to physical size, but can be virtual size
Linus Torvalds1da177e2005-04-16 15:20:36 -07008235 */
8236 max_sectors = mddev->resync_max_sectors;
Jianpeng Ma7f7583d2012-10-11 14:17:59 +11008237 atomic64_set(&mddev->resync_mismatches, 0);
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008238 /* we don't use the checkpoint if there's a bitmap */
Neil Brown5e96ee62008-06-28 08:31:24 +10008239 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
8240 j = mddev->resync_min;
8241 else if (!mddev->bitmap)
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008242 j = mddev->recovery_cp;
Neil Brown5e96ee62008-06-28 08:31:24 +10008243
NeilBrownccfcc3c2006-03-27 01:18:09 -08008244 } else if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
NeilBrownc804cde2012-05-21 09:28:33 +10008245 max_sectors = mddev->resync_max_sectors;
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008246 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008247 /* recovery follows the physical size of devices */
Andre Noll58c0fed2009-03-31 14:33:13 +11008248 max_sectors = mddev->dev_sectors;
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008249 j = MaxSector;
Dan Williams4e59ca72009-12-12 21:17:06 -07008250 rcu_read_lock();
NeilBrowndafb20f2012-03-19 12:46:39 +11008251 rdev_for_each_rcu(rdev, mddev)
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008252 if (rdev->raid_disk >= 0 &&
Shaohua Lif2076e72015-10-08 21:54:12 -07008253 !test_bit(Journal, &rdev->flags) &&
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008254 !test_bit(Faulty, &rdev->flags) &&
8255 !test_bit(In_sync, &rdev->flags) &&
8256 rdev->recovery_offset < j)
8257 j = rdev->recovery_offset;
Dan Williams4e59ca72009-12-12 21:17:06 -07008258 rcu_read_unlock();
NeilBrown133d4522014-07-02 12:04:14 +10008259
8260 /* If there is a bitmap, we need to make sure all
8261 * writes that started before we added a spare
8262 * complete before we start doing a recovery.
8263 * Otherwise the write might complete and (via
8264 * bitmap_endwrite) set a bit in the bitmap after the
8265 * recovery has checked that bit and skipped that
8266 * region.
8267 */
8268 if (mddev->bitmap) {
8269 mddev->pers->quiesce(mddev, 1);
8270 mddev->pers->quiesce(mddev, 0);
8271 }
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008272 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008273
NeilBrown9d487392016-11-02 14:16:49 +11008274 pr_info("md: %s of RAID array %s\n", desc, mdname(mddev));
8275 pr_debug("md: minimum _guaranteed_ speed: %d KB/sec/disk.\n", speed_min(mddev));
8276 pr_debug("md: using maximum available idle IO bandwidth (but not more than %d KB/sec) for %s.\n",
8277 speed_max(mddev), desc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008278
NeilBrowneea1bf32009-03-31 14:27:02 +11008279 is_mddev_idle(mddev, 1); /* this initializes IO event counters */
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008280
NeilBrown57afd892005-06-21 17:17:13 -07008281 io_sectors = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008282 for (m = 0; m < SYNC_MARKS; m++) {
8283 mark[m] = jiffies;
NeilBrown57afd892005-06-21 17:17:13 -07008284 mark_cnt[m] = io_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008285 }
8286 last_mark = 0;
8287 mddev->resync_mark = mark[last_mark];
8288 mddev->resync_mark_cnt = mark_cnt[last_mark];
8289
8290 /*
8291 * Tune reconstruction:
8292 */
8293 window = 32*(PAGE_SIZE/512);
NeilBrown9d487392016-11-02 14:16:49 +11008294 pr_debug("md: using %dk window, over a total of %lluk.\n",
8295 window/2, (unsigned long long)max_sectors/2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008296
8297 atomic_set(&mddev->recovery_active, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008298 last_check = 0;
8299
8300 if (j>2) {
NeilBrown9d487392016-11-02 14:16:49 +11008301 pr_debug("md: resuming %s of %s from checkpoint.\n",
8302 desc, mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07008303 mddev->curr_resync = j;
NeilBrown72f36d52012-10-11 14:25:57 +11008304 } else
8305 mddev->curr_resync = 3; /* no longer delayed */
NeilBrown75d3da42011-01-14 09:14:34 +11008306 mddev->curr_resync_completed = j;
NeilBrown72f36d52012-10-11 14:25:57 +11008307 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
8308 md_new_event(mddev);
majianpeng54f89342012-10-31 11:59:10 +11008309 update_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008310
majianpeng7c2c57c2012-07-03 12:12:26 +10008311 blk_start_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008312 while (j < max_sectors) {
NeilBrown57afd892005-06-21 17:17:13 -07008313 sector_t sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008314
NeilBrown57afd892005-06-21 17:17:13 -07008315 skipped = 0;
NeilBrown97e4f422009-03-31 14:33:13 +11008316
NeilBrown7a91ee12009-05-26 12:57:21 +10008317 if (!test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
8318 ((mddev->curr_resync > mddev->curr_resync_completed &&
8319 (mddev->curr_resync - mddev->curr_resync_completed)
8320 > (max_sectors >> 4)) ||
majianpeng54f89342012-10-31 11:59:10 +11008321 time_after_eq(jiffies, update_time + UPDATE_FREQUENCY) ||
NeilBrown7a91ee12009-05-26 12:57:21 +10008322 (j - mddev->curr_resync_completed)*2
NeilBrownc5e19d92015-07-17 12:06:02 +10008323 >= mddev->resync_max - mddev->curr_resync_completed ||
8324 mddev->curr_resync_completed > mddev->resync_max
NeilBrown7a91ee12009-05-26 12:57:21 +10008325 )) {
NeilBrown97e4f422009-03-31 14:33:13 +11008326 /* time to update curr_resync_completed */
NeilBrown97e4f422009-03-31 14:33:13 +11008327 wait_event(mddev->recovery_wait,
8328 atomic_read(&mddev->recovery_active) == 0);
NeilBrown75d3da42011-01-14 09:14:34 +11008329 mddev->curr_resync_completed = j;
kernelmail35d78c62012-10-31 11:59:10 +11008330 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
8331 j > mddev->recovery_cp)
8332 mddev->recovery_cp = j;
majianpeng54f89342012-10-31 11:59:10 +11008333 update_time = jiffies;
Shaohua Li29530792016-12-08 15:48:19 -08008334 set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
NeilBrownacb180b2009-04-14 16:28:34 +10008335 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
NeilBrown97e4f422009-03-31 14:33:13 +11008336 }
NeilBrownacb180b2009-04-14 16:28:34 +10008337
NeilBrownc91abf52013-11-19 12:02:01 +11008338 while (j >= mddev->resync_max &&
8339 !test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
NeilBrowne62e58a2009-07-01 13:15:35 +10008340 /* As this condition is controlled by user-space,
8341 * we can block indefinitely, so use '_interruptible'
8342 * to avoid triggering warnings.
8343 */
8344 flush_signals(current); /* just in case */
8345 wait_event_interruptible(mddev->recovery_wait,
8346 mddev->resync_max > j
NeilBrownc91abf52013-11-19 12:02:01 +11008347 || test_bit(MD_RECOVERY_INTR,
8348 &mddev->recovery));
NeilBrowne62e58a2009-07-01 13:15:35 +10008349 }
NeilBrownacb180b2009-04-14 16:28:34 +10008350
NeilBrownc91abf52013-11-19 12:02:01 +11008351 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
8352 break;
NeilBrownacb180b2009-04-14 16:28:34 +10008353
NeilBrown09314792015-02-19 16:04:40 +11008354 sectors = mddev->pers->sync_request(mddev, j, &skipped);
NeilBrown57afd892005-06-21 17:17:13 -07008355 if (sectors == 0) {
NeilBrowndfc70642008-05-23 13:04:39 -07008356 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
NeilBrownc91abf52013-11-19 12:02:01 +11008357 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008358 }
NeilBrown57afd892005-06-21 17:17:13 -07008359
8360 if (!skipped) { /* actual IO requested */
8361 io_sectors += sectors;
8362 atomic_add(sectors, &mddev->recovery_active);
8363 }
8364
NeilBrowne875ece2011-07-28 11:39:24 +10008365 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
8366 break;
8367
Linus Torvalds1da177e2005-04-16 15:20:36 -07008368 j += sectors;
NeilBrown5ed1df22015-07-24 13:27:08 +10008369 if (j > max_sectors)
8370 /* when skipping, extra large numbers can be returned. */
8371 j = max_sectors;
NeilBrown72f36d52012-10-11 14:25:57 +11008372 if (j > 2)
8373 mddev->curr_resync = j;
NeilBrownff4e8d92006-07-10 04:44:16 -07008374 mddev->curr_mark_cnt = io_sectors;
NeilBrownd7603b72006-01-06 00:20:30 -08008375 if (last_check == 0)
NeilBrowne875ece2011-07-28 11:39:24 +10008376 /* this is the earliest that rebuild will be
NeilBrownd7603b72006-01-06 00:20:30 -08008377 * visible in /proc/mdstat
8378 */
8379 md_new_event(mddev);
NeilBrown57afd892005-06-21 17:17:13 -07008380
8381 if (last_check + window > io_sectors || j == max_sectors)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008382 continue;
8383
NeilBrown57afd892005-06-21 17:17:13 -07008384 last_check = io_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008385 repeat:
8386 if (time_after_eq(jiffies, mark[last_mark] + SYNC_MARK_STEP )) {
8387 /* step marks */
8388 int next = (last_mark+1) % SYNC_MARKS;
8389
8390 mddev->resync_mark = mark[next];
8391 mddev->resync_mark_cnt = mark_cnt[next];
8392 mark[next] = jiffies;
NeilBrown57afd892005-06-21 17:17:13 -07008393 mark_cnt[next] = io_sectors - atomic_read(&mddev->recovery_active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008394 last_mark = next;
8395 }
8396
NeilBrownc91abf52013-11-19 12:02:01 +11008397 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
8398 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008399
8400 /*
8401 * this loop exits only if either when we are slower than
8402 * the 'hard' speed limit, or the system was IO-idle for
8403 * a jiffy.
8404 * the system might be non-idle CPU-wise, but we only care
8405 * about not overloading the IO subsystem. (things like an
8406 * e2fsck being done on the RAID array should execute fast)
8407 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008408 cond_resched();
8409
Xiao Niac7e50a2014-08-07 09:37:41 -04008410 recovery_done = io_sectors - atomic_read(&mddev->recovery_active);
8411 currspeed = ((unsigned long)(recovery_done - mddev->resync_mark_cnt))/2
NeilBrown57afd892005-06-21 17:17:13 -07008412 /((jiffies-mddev->resync_mark)/HZ +1) +1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008413
NeilBrown88202a02006-01-06 00:21:36 -08008414 if (currspeed > speed_min(mddev)) {
NeilBrownac8fa412015-02-19 16:55:00 +11008415 if (currspeed > speed_max(mddev)) {
NeilBrownc0e48522005-11-18 01:11:01 -08008416 msleep(500);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008417 goto repeat;
8418 }
NeilBrownac8fa412015-02-19 16:55:00 +11008419 if (!is_mddev_idle(mddev, 0)) {
8420 /*
8421 * Give other IO more of a chance.
8422 * The faster the devices, the less we wait.
8423 */
8424 wait_event(mddev->recovery_wait,
8425 !atomic_read(&mddev->recovery_active));
8426 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008427 }
8428 }
NeilBrown9d487392016-11-02 14:16:49 +11008429 pr_info("md: %s: %s %s.\n",mdname(mddev), desc,
8430 test_bit(MD_RECOVERY_INTR, &mddev->recovery)
8431 ? "interrupted" : "done");
Linus Torvalds1da177e2005-04-16 15:20:36 -07008432 /*
8433 * this also signals 'finished resyncing' to md_stop
8434 */
majianpeng7c2c57c2012-07-03 12:12:26 +10008435 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008436 wait_event(mddev->recovery_wait, !atomic_read(&mddev->recovery_active));
8437
NeilBrown5ed1df22015-07-24 13:27:08 +10008438 if (!test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
8439 !test_bit(MD_RECOVERY_INTR, &mddev->recovery) &&
NeilBrown1217e1d2016-10-28 15:59:41 +11008440 mddev->curr_resync > 3) {
NeilBrown5ed1df22015-07-24 13:27:08 +10008441 mddev->curr_resync_completed = mddev->curr_resync;
8442 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
8443 }
NeilBrown09314792015-02-19 16:04:40 +11008444 mddev->pers->sync_request(mddev, max_sectors, &skipped);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008445
NeilBrowndfc70642008-05-23 13:04:39 -07008446 if (!test_bit(MD_RECOVERY_CHECK, &mddev->recovery) &&
NeilBrown1217e1d2016-10-28 15:59:41 +11008447 mddev->curr_resync > 3) {
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008448 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
8449 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8450 if (mddev->curr_resync >= mddev->recovery_cp) {
NeilBrown9d487392016-11-02 14:16:49 +11008451 pr_debug("md: checkpointing %s of %s.\n",
8452 desc, mdname(mddev));
majianpeng0a19caa2012-11-19 19:57:34 +08008453 if (test_bit(MD_RECOVERY_ERROR,
8454 &mddev->recovery))
8455 mddev->recovery_cp =
8456 mddev->curr_resync_completed;
8457 else
8458 mddev->recovery_cp =
8459 mddev->curr_resync;
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008460 }
8461 } else
8462 mddev->recovery_cp = MaxSector;
8463 } else {
8464 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery))
8465 mddev->curr_resync = MaxSector;
Dan Williams4e59ca72009-12-12 21:17:06 -07008466 rcu_read_lock();
NeilBrowndafb20f2012-03-19 12:46:39 +11008467 rdev_for_each_rcu(rdev, mddev)
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008468 if (rdev->raid_disk >= 0 &&
NeilBrown70fffd02010-06-16 17:01:25 +10008469 mddev->delta_disks >= 0 &&
Shaohua Lif2076e72015-10-08 21:54:12 -07008470 !test_bit(Journal, &rdev->flags) &&
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008471 !test_bit(Faulty, &rdev->flags) &&
8472 !test_bit(In_sync, &rdev->flags) &&
8473 rdev->recovery_offset < mddev->curr_resync)
8474 rdev->recovery_offset = mddev->curr_resync;
Dan Williams4e59ca72009-12-12 21:17:06 -07008475 rcu_read_unlock();
NeilBrown5fd6c1d2006-06-26 00:27:40 -07008476 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008477 }
NeilBrowndb91ff52012-02-07 12:01:51 +11008478 skip:
Guoqing Jiangbb8bf152016-06-02 23:32:04 -04008479 /* set CHANGE_PENDING here since maybe another update is needed,
8480 * so other nodes are informed. It should be harmless for normal
8481 * raid */
Shaohua Li29530792016-12-08 15:48:19 -08008482 set_mask_bits(&mddev->sb_flags, 0,
8483 BIT(MD_SB_CHANGE_PENDING) | BIT(MD_SB_CHANGE_DEVS));
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -05008484
NeilBrown23da4222014-12-15 12:57:01 +11008485 spin_lock(&mddev->lock);
NeilBrownc07b70a2009-12-14 12:49:48 +11008486 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8487 /* We completed so min/max setting can be forgotten if used. */
8488 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
8489 mddev->resync_min = 0;
8490 mddev->resync_max = MaxSector;
8491 } else if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
8492 mddev->resync_min = mddev->curr_resync_completed;
NeilBrownf7851be2015-07-02 17:12:58 +10008493 set_bit(MD_RECOVERY_DONE, &mddev->recovery);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008494 mddev->curr_resync = 0;
NeilBrown23da4222014-12-15 12:57:01 +11008495 spin_unlock(&mddev->lock);
8496
Linus Torvalds1da177e2005-04-16 15:20:36 -07008497 wake_up(&resync_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008498 md_wakeup_thread(mddev->thread);
NeilBrownc6207272008-02-06 01:39:52 -08008499 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008500}
NeilBrown29269552006-03-27 01:18:10 -08008501EXPORT_SYMBOL_GPL(md_do_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008502
NeilBrown746d3202013-04-24 11:42:41 +10008503static int remove_and_add_spares(struct mddev *mddev,
8504 struct md_rdev *this)
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008505{
NeilBrown3cb03002011-10-11 16:45:26 +11008506 struct md_rdev *rdev;
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008507 int spares = 0;
NeilBrownf2a371c2012-01-09 00:46:41 +11008508 int removed = 0;
NeilBrownd787be42016-06-02 16:19:53 +10008509 bool remove_some = false;
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008510
NeilBrownd787be42016-06-02 16:19:53 +10008511 rdev_for_each(rdev, mddev) {
NeilBrown746d3202013-04-24 11:42:41 +10008512 if ((this == NULL || rdev == this) &&
8513 rdev->raid_disk >= 0 &&
Dan Williams6bfe0b42008-04-30 00:52:32 -07008514 !test_bit(Blocked, &rdev->flags) &&
NeilBrownd787be42016-06-02 16:19:53 +10008515 test_bit(Faulty, &rdev->flags) &&
8516 atomic_read(&rdev->nr_pending)==0) {
8517 /* Faulty non-Blocked devices with nr_pending == 0
8518 * never get nr_pending incremented,
8519 * never get Faulty cleared, and never get Blocked set.
8520 * So we can synchronize_rcu now rather than once per device
8521 */
8522 remove_some = true;
8523 set_bit(RemoveSynchronized, &rdev->flags);
8524 }
8525 }
8526
8527 if (remove_some)
8528 synchronize_rcu();
8529 rdev_for_each(rdev, mddev) {
8530 if ((this == NULL || rdev == this) &&
8531 rdev->raid_disk >= 0 &&
8532 !test_bit(Blocked, &rdev->flags) &&
8533 ((test_bit(RemoveSynchronized, &rdev->flags) ||
Shaohua Lif2076e72015-10-08 21:54:12 -07008534 (!test_bit(In_sync, &rdev->flags) &&
8535 !test_bit(Journal, &rdev->flags))) &&
NeilBrownd787be42016-06-02 16:19:53 +10008536 atomic_read(&rdev->nr_pending)==0)) {
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008537 if (mddev->pers->hot_remove_disk(
NeilBrownb8321b62011-12-23 10:17:51 +11008538 mddev, rdev) == 0) {
Namhyung Kim36fad852011-07-27 11:00:36 +10008539 sysfs_unlink_rdev(mddev, rdev);
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008540 rdev->raid_disk = -1;
NeilBrownf2a371c2012-01-09 00:46:41 +11008541 removed++;
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008542 }
8543 }
NeilBrownd787be42016-06-02 16:19:53 +10008544 if (remove_some && test_bit(RemoveSynchronized, &rdev->flags))
8545 clear_bit(RemoveSynchronized, &rdev->flags);
8546 }
8547
Jonathan Brassow90584fc2013-03-07 16:24:26 -06008548 if (removed && mddev->kobj.sd)
8549 sysfs_notify(&mddev->kobj, NULL, "degraded");
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008550
Goldwyn Rodrigues2910ff12015-09-28 10:27:26 -05008551 if (this && removed)
NeilBrown746d3202013-04-24 11:42:41 +10008552 goto no_add;
8553
NeilBrowndafb20f2012-03-19 12:46:39 +11008554 rdev_for_each(rdev, mddev) {
Goldwyn Rodrigues2910ff12015-09-28 10:27:26 -05008555 if (this && this != rdev)
8556 continue;
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05008557 if (test_bit(Candidate, &rdev->flags))
8558 continue;
NeilBrown7bfec5f2011-12-23 10:17:53 +11008559 if (rdev->raid_disk >= 0 &&
8560 !test_bit(In_sync, &rdev->flags) &&
Shaohua Lif2076e72015-10-08 21:54:12 -07008561 !test_bit(Journal, &rdev->flags) &&
NeilBrown7bfec5f2011-12-23 10:17:53 +11008562 !test_bit(Faulty, &rdev->flags))
8563 spares++;
NeilBrown7ceb17e2013-04-24 11:42:42 +10008564 if (rdev->raid_disk >= 0)
8565 continue;
8566 if (test_bit(Faulty, &rdev->flags))
8567 continue;
Shaohua Lif6b6ec52015-12-21 10:51:02 +11008568 if (!test_bit(Journal, &rdev->flags)) {
8569 if (mddev->ro &&
8570 ! (rdev->saved_raid_disk >= 0 &&
8571 !test_bit(Bitmap_sync, &rdev->flags)))
8572 continue;
NeilBrown7ceb17e2013-04-24 11:42:42 +10008573
Shaohua Lif6b6ec52015-12-21 10:51:02 +11008574 rdev->recovery_offset = 0;
8575 }
NeilBrown7ceb17e2013-04-24 11:42:42 +10008576 if (mddev->pers->
8577 hot_add_disk(mddev, rdev) == 0) {
8578 if (sysfs_link_rdev(mddev, rdev))
8579 /* failure here is OK */;
Shaohua Lif6b6ec52015-12-21 10:51:02 +11008580 if (!test_bit(Journal, &rdev->flags))
8581 spares++;
NeilBrown7ceb17e2013-04-24 11:42:42 +10008582 md_new_event(mddev);
Shaohua Li29530792016-12-08 15:48:19 -08008583 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
NeilBrowndfc70642008-05-23 13:04:39 -07008584 }
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008585 }
NeilBrown746d3202013-04-24 11:42:41 +10008586no_add:
NeilBrown6dafab62012-09-19 12:54:22 +10008587 if (removed)
Shaohua Li29530792016-12-08 15:48:19 -08008588 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008589 return spares;
8590}
NeilBrown7ebc0be2011-01-14 09:14:33 +11008591
NeilBrownac05f252014-09-30 08:10:42 +10008592static void md_start_sync(struct work_struct *ws)
8593{
8594 struct mddev *mddev = container_of(ws, struct mddev, del_work);
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -05008595
NeilBrownac05f252014-09-30 08:10:42 +10008596 mddev->sync_thread = md_register_thread(md_do_sync,
8597 mddev,
8598 "resync");
8599 if (!mddev->sync_thread) {
NeilBrown9d487392016-11-02 14:16:49 +11008600 pr_warn("%s: could not start resync thread...\n",
8601 mdname(mddev));
NeilBrownac05f252014-09-30 08:10:42 +10008602 /* leave the spares where they are, it shouldn't hurt */
8603 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8604 clear_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8605 clear_bit(MD_RECOVERY_REQUESTED, &mddev->recovery);
8606 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8607 clear_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
NeilBrownf851b602014-12-11 10:02:10 +11008608 wake_up(&resync_wait);
NeilBrownac05f252014-09-30 08:10:42 +10008609 if (test_and_clear_bit(MD_RECOVERY_RECOVER,
8610 &mddev->recovery))
8611 if (mddev->sysfs_action)
8612 sysfs_notify_dirent_safe(mddev->sysfs_action);
8613 } else
8614 md_wakeup_thread(mddev->sync_thread);
8615 sysfs_notify_dirent_safe(mddev->sysfs_action);
8616 md_new_event(mddev);
8617}
8618
Linus Torvalds1da177e2005-04-16 15:20:36 -07008619/*
8620 * This routine is regularly called by all per-raid-array threads to
8621 * deal with generic issues like resync and super-block update.
8622 * Raid personalities that don't have a thread (linear/raid0) do not
8623 * need this as they never do any recovery or update the superblock.
8624 *
8625 * It does not do any resync itself, but rather "forks" off other threads
8626 * to do that as needed.
8627 * When it is determined that resync is needed, we set MD_RECOVERY_RUNNING in
8628 * "->recovery" and create a thread at ->sync_thread.
NeilBrowndfc70642008-05-23 13:04:39 -07008629 * When the thread finishes it sets MD_RECOVERY_DONE
Linus Torvalds1da177e2005-04-16 15:20:36 -07008630 * and wakeups up this thread which will reap the thread and finish up.
8631 * This thread also removes any faulty devices (with nr_pending == 0).
8632 *
8633 * The overall approach is:
8634 * 1/ if the superblock needs updating, update it.
8635 * 2/ If a recovery thread is running, don't do anything else.
8636 * 3/ If recovery has finished, clean up, possibly marking spares active.
8637 * 4/ If there are any faulty devices, remove them.
8638 * 5/ If array is degraded, try to add spares devices
8639 * 6/ If array has spares or is not in-sync, start a resync thread.
8640 */
NeilBrownfd01b882011-10-11 16:47:53 +11008641void md_check_recovery(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008642{
Jonathan Brassow68866e42011-06-08 15:10:08 +10008643 if (mddev->suspended)
8644 return;
8645
NeilBrown5f404022005-06-21 17:17:16 -07008646 if (mddev->bitmap)
NeilBrownaa5cbd12009-12-14 12:49:46 +11008647 bitmap_daemon_work(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008648
NeilBrownfca4d842005-06-21 17:17:11 -07008649 if (signal_pending(current)) {
NeilBrown31a59e32008-04-30 00:52:30 -07008650 if (mddev->pers->sync_request && !mddev->external) {
NeilBrown9d487392016-11-02 14:16:49 +11008651 pr_debug("md: %s in immediate safe mode\n",
8652 mdname(mddev));
NeilBrownfca4d842005-06-21 17:17:11 -07008653 mddev->safemode = 2;
8654 }
8655 flush_signals(current);
8656 }
8657
NeilBrownc89a8ee2008-08-05 15:54:13 +10008658 if (mddev->ro && !test_bit(MD_RECOVERY_NEEDED, &mddev->recovery))
8659 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008660 if ( ! (
Shaohua Li29530792016-12-08 15:48:19 -08008661 (mddev->sb_flags & ~ (1<<MD_SB_CHANGE_PENDING)) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07008662 test_bit(MD_RECOVERY_NEEDED, &mddev->recovery) ||
NeilBrownfca4d842005-06-21 17:17:11 -07008663 test_bit(MD_RECOVERY_DONE, &mddev->recovery) ||
NeilBrown31a59e32008-04-30 00:52:30 -07008664 (mddev->external == 0 && mddev->safemode == 1) ||
NeilBrown4ad23a972017-03-15 14:05:14 +11008665 (mddev->safemode == 2
NeilBrownfca4d842005-06-21 17:17:11 -07008666 && !mddev->in_sync && mddev->recovery_cp == MaxSector)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008667 ))
8668 return;
NeilBrownfca4d842005-06-21 17:17:11 -07008669
NeilBrowndf5b89b2006-03-27 01:18:20 -08008670 if (mddev_trylock(mddev)) {
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008671 int spares = 0;
NeilBrownfca4d842005-06-21 17:17:11 -07008672
Shaohua Liafc1f552017-08-11 20:34:45 -07008673 if (!mddev->external && mddev->safemode == 1)
NeilBrown33182d12017-08-08 16:56:36 +10008674 mddev->safemode = 0;
8675
NeilBrownc89a8ee2008-08-05 15:54:13 +10008676 if (mddev->ro) {
Neil Brownab16bfc2015-06-17 12:31:46 +10008677 struct md_rdev *rdev;
8678 if (!mddev->external && mddev->in_sync)
8679 /* 'Blocked' flag not needed as failed devices
8680 * will be recorded if array switched to read/write.
8681 * Leaving it set will prevent the device
8682 * from being removed.
8683 */
8684 rdev_for_each(rdev, mddev)
8685 clear_bit(Blocked, &rdev->flags);
NeilBrown7ceb17e2013-04-24 11:42:42 +10008686 /* On a read-only array we can:
8687 * - remove failed devices
8688 * - add already-in_sync devices if the array itself
8689 * is in-sync.
8690 * As we only add devices that are already in-sync,
8691 * we can activate the spares immediately.
NeilBrownc89a8ee2008-08-05 15:54:13 +10008692 */
NeilBrown7ceb17e2013-04-24 11:42:42 +10008693 remove_and_add_spares(mddev, NULL);
NeilBrown8313b8e2013-12-12 10:13:33 +11008694 /* There is no thread, but we need to call
8695 * ->spare_active and clear saved_raid_disk
8696 */
NeilBrown2ac295a2014-05-29 11:40:03 +10008697 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
NeilBrown8313b8e2013-12-12 10:13:33 +11008698 md_reap_sync_thread(mddev);
NeilBrowna4a3d262015-07-17 11:57:30 +10008699 clear_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
NeilBrown8313b8e2013-12-12 10:13:33 +11008700 clear_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
Shaohua Li29530792016-12-08 15:48:19 -08008701 clear_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
NeilBrownc89a8ee2008-08-05 15:54:13 +10008702 goto unlock;
8703 }
8704
Guoqing Jiang659b2542015-12-21 10:50:59 +11008705 if (mddev_is_clustered(mddev)) {
8706 struct md_rdev *rdev;
8707 /* kick the device if another node issued a
8708 * remove disk.
8709 */
8710 rdev_for_each(rdev, mddev) {
8711 if (test_and_clear_bit(ClusterRemove, &rdev->flags) &&
8712 rdev->raid_disk < 0)
8713 md_kick_rdev_from_array(rdev);
8714 }
8715 }
8716
NeilBrown6497709b2017-03-15 14:05:14 +11008717 if (!mddev->external && !mddev->in_sync) {
NeilBrown85572d72014-12-15 12:56:56 +11008718 spin_lock(&mddev->lock);
NeilBrown6497709b2017-03-15 14:05:14 +11008719 set_in_sync(mddev);
NeilBrown85572d72014-12-15 12:56:56 +11008720 spin_unlock(&mddev->lock);
NeilBrownfca4d842005-06-21 17:17:11 -07008721 }
NeilBrownfca4d842005-06-21 17:17:11 -07008722
Shaohua Li29530792016-12-08 15:48:19 -08008723 if (mddev->sb_flags)
NeilBrown850b2b42006-10-03 01:15:46 -07008724 md_update_sb(mddev, 0);
NeilBrown06d91a52005-06-21 17:17:12 -07008725
Linus Torvalds1da177e2005-04-16 15:20:36 -07008726 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) &&
8727 !test_bit(MD_RECOVERY_DONE, &mddev->recovery)) {
8728 /* resync/recovery still happening */
8729 clear_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8730 goto unlock;
8731 }
8732 if (mddev->sync_thread) {
Jonathan Brassowa91d5ac2013-04-24 11:42:43 +10008733 md_reap_sync_thread(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008734 goto unlock;
8735 }
Neil Brown72a23c22008-06-28 08:31:41 +10008736 /* Set RUNNING before clearing NEEDED to avoid
8737 * any transients in the value of "sync_action".
8738 */
NeilBrown72f36d52012-10-11 14:25:57 +11008739 mddev->curr_resync_completed = 0;
NeilBrown23da4222014-12-15 12:57:01 +11008740 spin_lock(&mddev->lock);
Neil Brown72a23c22008-06-28 08:31:41 +10008741 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
NeilBrown23da4222014-12-15 12:57:01 +11008742 spin_unlock(&mddev->lock);
NeilBrown24dd4692005-11-08 21:39:26 -08008743 /* Clear some bits that don't mean anything, but
8744 * might be left set
8745 */
NeilBrown24dd4692005-11-08 21:39:26 -08008746 clear_bit(MD_RECOVERY_INTR, &mddev->recovery);
8747 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008748
NeilBrowned209582012-04-24 10:23:14 +10008749 if (!test_and_clear_bit(MD_RECOVERY_NEEDED, &mddev->recovery) ||
8750 test_bit(MD_RECOVERY_FROZEN, &mddev->recovery))
NeilBrownac05f252014-09-30 08:10:42 +10008751 goto not_running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008752 /* no recovery is running.
8753 * remove any failed drives, then
8754 * add spares if possible.
NeilBrown72f36d52012-10-11 14:25:57 +11008755 * Spares are also removed and re-added, to allow
Linus Torvalds1da177e2005-04-16 15:20:36 -07008756 * the personality to fail the re-add.
8757 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008758
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008759 if (mddev->reshape_position != MaxSector) {
NeilBrown50ac1682009-06-18 08:47:55 +10008760 if (mddev->pers->check_reshape == NULL ||
8761 mddev->pers->check_reshape(mddev) != 0)
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008762 /* Cannot proceed */
NeilBrownac05f252014-09-30 08:10:42 +10008763 goto not_running;
NeilBrownb4c4c7b2007-02-28 20:11:48 -08008764 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
Neil Brown72a23c22008-06-28 08:31:41 +10008765 clear_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
NeilBrown746d3202013-04-24 11:42:41 +10008766 } else if ((spares = remove_and_add_spares(mddev, NULL))) {
NeilBrown24dd4692005-11-08 21:39:26 -08008767 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8768 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
Dan Williams56ac36d2008-08-07 10:02:47 -07008769 clear_bit(MD_RECOVERY_REQUESTED, &mddev->recovery);
Neil Brown72a23c22008-06-28 08:31:41 +10008770 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
NeilBrown24dd4692005-11-08 21:39:26 -08008771 } else if (mddev->recovery_cp < MaxSector) {
8772 set_bit(MD_RECOVERY_SYNC, &mddev->recovery);
Neil Brown72a23c22008-06-28 08:31:41 +10008773 clear_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
NeilBrown24dd4692005-11-08 21:39:26 -08008774 } else if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
8775 /* nothing to be done ... */
NeilBrownac05f252014-09-30 08:10:42 +10008776 goto not_running;
NeilBrown24dd4692005-11-08 21:39:26 -08008777
Linus Torvalds1da177e2005-04-16 15:20:36 -07008778 if (mddev->pers->sync_request) {
NeilBrownef99bf42012-05-22 13:55:08 +10008779 if (spares) {
NeilBrowna654b9d82005-06-21 17:17:27 -07008780 /* We are adding a device or devices to an array
8781 * which has the bitmap stored on all devices.
8782 * So make sure all bitmap pages get written
8783 */
8784 bitmap_write_all(mddev->bitmap);
8785 }
NeilBrownac05f252014-09-30 08:10:42 +10008786 INIT_WORK(&mddev->del_work, md_start_sync);
8787 queue_work(md_misc_wq, &mddev->del_work);
8788 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008789 }
NeilBrownac05f252014-09-30 08:10:42 +10008790 not_running:
Neil Brown72a23c22008-06-28 08:31:41 +10008791 if (!mddev->sync_thread) {
8792 clear_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
NeilBrownf851b602014-12-11 10:02:10 +11008793 wake_up(&resync_wait);
Neil Brown72a23c22008-06-28 08:31:41 +10008794 if (test_and_clear_bit(MD_RECOVERY_RECOVER,
8795 &mddev->recovery))
NeilBrown0c3573f2009-01-09 08:31:05 +11008796 if (mddev->sysfs_action)
NeilBrown00bcb4a2010-06-01 19:37:23 +10008797 sysfs_notify_dirent_safe(mddev->sysfs_action);
Neil Brown72a23c22008-06-28 08:31:41 +10008798 }
NeilBrownac05f252014-09-30 08:10:42 +10008799 unlock:
8800 wake_up(&mddev->sb_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008801 mddev_unlock(mddev);
8802 }
8803}
NeilBrown6c144d32014-09-30 16:15:38 +10008804EXPORT_SYMBOL(md_check_recovery);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008805
Jonathan Brassowa91d5ac2013-04-24 11:42:43 +10008806void md_reap_sync_thread(struct mddev *mddev)
8807{
8808 struct md_rdev *rdev;
8809
8810 /* resync has finished, collect result */
8811 md_unregister_thread(&mddev->sync_thread);
8812 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery) &&
8813 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
8814 /* success...*/
8815 /* activate any spares */
8816 if (mddev->pers->spare_active(mddev)) {
8817 sysfs_notify(&mddev->kobj, NULL,
8818 "degraded");
Shaohua Li29530792016-12-08 15:48:19 -08008819 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
Jonathan Brassowa91d5ac2013-04-24 11:42:43 +10008820 }
8821 }
8822 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
8823 mddev->pers->finish_reshape)
8824 mddev->pers->finish_reshape(mddev);
8825
8826 /* If array is no-longer degraded, then any saved_raid_disk
NeilBrownf4667222013-12-09 12:04:56 +11008827 * information must be scrapped.
Jonathan Brassowa91d5ac2013-04-24 11:42:43 +10008828 */
NeilBrownf4667222013-12-09 12:04:56 +11008829 if (!mddev->degraded)
8830 rdev_for_each(rdev, mddev)
Jonathan Brassowa91d5ac2013-04-24 11:42:43 +10008831 rdev->saved_raid_disk = -1;
8832
8833 md_update_sb(mddev, 1);
Shaohua Li29530792016-12-08 15:48:19 -08008834 /* MD_SB_CHANGE_PENDING should be cleared by md_update_sb, so we can
Guoqing Jiangbb8bf152016-06-02 23:32:04 -04008835 * call resync_finish here if MD_CLUSTER_RESYNC_LOCKED is set by
8836 * clustered raid */
8837 if (test_and_clear_bit(MD_CLUSTER_RESYNC_LOCKED, &mddev->flags))
8838 md_cluster_ops->resync_finish(mddev);
Jonathan Brassowa91d5ac2013-04-24 11:42:43 +10008839 clear_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
NeilBrownea358cd2015-06-12 20:05:04 +10008840 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
Jonathan Brassowa91d5ac2013-04-24 11:42:43 +10008841 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8842 clear_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8843 clear_bit(MD_RECOVERY_REQUESTED, &mddev->recovery);
8844 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
NeilBrownf851b602014-12-11 10:02:10 +11008845 wake_up(&resync_wait);
Jonathan Brassowa91d5ac2013-04-24 11:42:43 +10008846 /* flag recovery needed just to double check */
8847 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8848 sysfs_notify_dirent_safe(mddev->sysfs_action);
8849 md_new_event(mddev);
8850 if (mddev->event_work.func)
8851 queue_work(md_misc_wq, &mddev->event_work);
8852}
NeilBrown6c144d32014-09-30 16:15:38 +10008853EXPORT_SYMBOL(md_reap_sync_thread);
Jonathan Brassowa91d5ac2013-04-24 11:42:43 +10008854
NeilBrownfd01b882011-10-11 16:47:53 +11008855void md_wait_for_blocked_rdev(struct md_rdev *rdev, struct mddev *mddev)
Dan Williams6bfe0b42008-04-30 00:52:32 -07008856{
NeilBrown00bcb4a2010-06-01 19:37:23 +10008857 sysfs_notify_dirent_safe(rdev->sysfs_state);
Dan Williams6bfe0b42008-04-30 00:52:32 -07008858 wait_event_timeout(rdev->blocked_wait,
NeilBrownde393cd2011-07-28 11:31:48 +10008859 !test_bit(Blocked, &rdev->flags) &&
8860 !test_bit(BlockedBadBlocks, &rdev->flags),
Dan Williams6bfe0b42008-04-30 00:52:32 -07008861 msecs_to_jiffies(5000));
8862 rdev_dec_pending(rdev, mddev);
8863}
8864EXPORT_SYMBOL(md_wait_for_blocked_rdev);
8865
NeilBrownc6563a82012-05-21 09:27:00 +10008866void md_finish_reshape(struct mddev *mddev)
8867{
8868 /* called be personality module when reshape completes. */
8869 struct md_rdev *rdev;
8870
8871 rdev_for_each(rdev, mddev) {
8872 if (rdev->data_offset > rdev->new_data_offset)
8873 rdev->sectors += rdev->data_offset - rdev->new_data_offset;
8874 else
8875 rdev->sectors -= rdev->new_data_offset - rdev->data_offset;
8876 rdev->data_offset = rdev->new_data_offset;
8877 }
8878}
8879EXPORT_SYMBOL(md_finish_reshape);
NeilBrown2230dfe2011-07-28 11:31:46 +10008880
Vishal Vermafc974ee2015-12-24 19:20:34 -07008881/* Bad block management */
NeilBrown2230dfe2011-07-28 11:31:46 +10008882
Vishal Vermafc974ee2015-12-24 19:20:34 -07008883/* Returns 1 on success, 0 on failure */
NeilBrown3cb03002011-10-11 16:45:26 +11008884int rdev_set_badblocks(struct md_rdev *rdev, sector_t s, int sectors,
NeilBrownc6563a82012-05-21 09:27:00 +10008885 int is_new)
NeilBrown2230dfe2011-07-28 11:31:46 +10008886{
Guoqing Jiang85ad1d12016-05-03 22:22:13 -04008887 struct mddev *mddev = rdev->mddev;
NeilBrownc6563a82012-05-21 09:27:00 +10008888 int rv;
8889 if (is_new)
8890 s += rdev->new_data_offset;
8891 else
8892 s += rdev->data_offset;
Vishal Vermafc974ee2015-12-24 19:20:34 -07008893 rv = badblocks_set(&rdev->badblocks, s, sectors, 0);
8894 if (rv == 0) {
NeilBrown2230dfe2011-07-28 11:31:46 +10008895 /* Make sure they get written out promptly */
Tomasz Majchrzak35b785f2016-10-21 16:26:57 +02008896 if (test_bit(ExternalBbl, &rdev->flags))
8897 sysfs_notify(&rdev->kobj, NULL,
8898 "unacknowledged_bad_blocks");
NeilBrown8bd2f0a2011-12-08 16:26:08 +11008899 sysfs_notify_dirent_safe(rdev->sysfs_state);
Shaohua Li29530792016-12-08 15:48:19 -08008900 set_mask_bits(&mddev->sb_flags, 0,
8901 BIT(MD_SB_CHANGE_CLEAN) | BIT(MD_SB_CHANGE_PENDING));
NeilBrown2230dfe2011-07-28 11:31:46 +10008902 md_wakeup_thread(rdev->mddev->thread);
Vishal Vermafc974ee2015-12-24 19:20:34 -07008903 return 1;
8904 } else
8905 return 0;
NeilBrown2230dfe2011-07-28 11:31:46 +10008906}
8907EXPORT_SYMBOL_GPL(rdev_set_badblocks);
8908
NeilBrownc6563a82012-05-21 09:27:00 +10008909int rdev_clear_badblocks(struct md_rdev *rdev, sector_t s, int sectors,
8910 int is_new)
NeilBrown2230dfe2011-07-28 11:31:46 +10008911{
Tomasz Majchrzak35b785f2016-10-21 16:26:57 +02008912 int rv;
NeilBrownc6563a82012-05-21 09:27:00 +10008913 if (is_new)
8914 s += rdev->new_data_offset;
8915 else
8916 s += rdev->data_offset;
Tomasz Majchrzak35b785f2016-10-21 16:26:57 +02008917 rv = badblocks_clear(&rdev->badblocks, s, sectors);
8918 if ((rv == 0) && test_bit(ExternalBbl, &rdev->flags))
8919 sysfs_notify(&rdev->kobj, NULL, "bad_blocks");
8920 return rv;
NeilBrown2230dfe2011-07-28 11:31:46 +10008921}
8922EXPORT_SYMBOL_GPL(rdev_clear_badblocks);
8923
Adrian Bunk75c96f82005-05-05 16:16:09 -07008924static int md_notify_reboot(struct notifier_block *this,
8925 unsigned long code, void *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008926{
8927 struct list_head *tmp;
NeilBrownfd01b882011-10-11 16:47:53 +11008928 struct mddev *mddev;
Daniel P. Berrange2dba6a92011-09-23 10:40:45 +01008929 int need_delay = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008930
NeilBrownc744a652012-03-19 12:46:37 +11008931 for_each_mddev(mddev, tmp) {
8932 if (mddev_trylock(mddev)) {
NeilBrown30b8aa92012-04-24 10:23:16 +10008933 if (mddev->pers)
8934 __md_stop_writes(mddev);
NeilBrown0f62fb22014-05-06 09:36:08 +10008935 if (mddev->persistent)
8936 mddev->safemode = 2;
NeilBrownc744a652012-03-19 12:46:37 +11008937 mddev_unlock(mddev);
Daniel P. Berrange2dba6a92011-09-23 10:40:45 +01008938 }
NeilBrownc744a652012-03-19 12:46:37 +11008939 need_delay = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008940 }
NeilBrownc744a652012-03-19 12:46:37 +11008941 /*
8942 * certain more exotic SCSI devices are known to be
8943 * volatile wrt too early system reboots. While the
8944 * right place to handle this issue is the given
8945 * driver, we do want to have a safe RAID driver ...
8946 */
8947 if (need_delay)
8948 mdelay(1000*1);
8949
Linus Torvalds1da177e2005-04-16 15:20:36 -07008950 return NOTIFY_DONE;
8951}
8952
Adrian Bunk75c96f82005-05-05 16:16:09 -07008953static struct notifier_block md_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008954 .notifier_call = md_notify_reboot,
8955 .next = NULL,
8956 .priority = INT_MAX, /* before any real devices */
8957};
8958
8959static void md_geninit(void)
8960{
NeilBrown36a4e1f2011-10-07 14:23:17 +11008961 pr_debug("md: sizeof(mdp_super_t) = %d\n", (int)sizeof(mdp_super_t));
Linus Torvalds1da177e2005-04-16 15:20:36 -07008962
Denis V. Lunevc7705f3442008-04-29 01:02:35 -07008963 proc_create("mdstat", S_IRUGO, NULL, &md_seq_fops);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008964}
8965
Adrian Bunk75c96f82005-05-05 16:16:09 -07008966static int __init md_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008967{
Tejun Heoe804ac72010-10-15 15:36:08 +02008968 int ret = -ENOMEM;
8969
Tejun Heoada609e2011-01-25 14:35:54 +01008970 md_wq = alloc_workqueue("md", WQ_MEM_RECLAIM, 0);
Tejun Heoe804ac72010-10-15 15:36:08 +02008971 if (!md_wq)
8972 goto err_wq;
8973
8974 md_misc_wq = alloc_workqueue("md_misc", 0, 0);
8975 if (!md_misc_wq)
8976 goto err_misc_wq;
8977
8978 if ((ret = register_blkdev(MD_MAJOR, "md")) < 0)
8979 goto err_md;
8980
8981 if ((ret = register_blkdev(0, "mdp")) < 0)
8982 goto err_mdp;
8983 mdp_major = ret;
8984
NeilBrownaf5628f2014-07-31 13:54:54 +10008985 blk_register_region(MKDEV(MD_MAJOR, 0), 512, THIS_MODULE,
NeilBrowne8703fe2006-10-03 01:15:59 -07008986 md_probe, NULL, NULL);
8987 blk_register_region(MKDEV(mdp_major, 0), 1UL<<MINORBITS, THIS_MODULE,
Linus Torvalds1da177e2005-04-16 15:20:36 -07008988 md_probe, NULL, NULL);
8989
Linus Torvalds1da177e2005-04-16 15:20:36 -07008990 register_reboot_notifier(&md_notifier);
Eric W. Biederman0b4d4142007-02-14 00:34:09 -08008991 raid_table_header = register_sysctl_table(raid_root_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008992
8993 md_geninit();
NeilBrownd710e132008-10-13 11:55:12 +11008994 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008995
Tejun Heoe804ac72010-10-15 15:36:08 +02008996err_mdp:
8997 unregister_blkdev(MD_MAJOR, "md");
8998err_md:
8999 destroy_workqueue(md_misc_wq);
9000err_misc_wq:
9001 destroy_workqueue(md_wq);
9002err_wq:
9003 return ret;
9004}
Linus Torvalds1da177e2005-04-16 15:20:36 -07009005
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05009006static void check_sb_changes(struct mddev *mddev, struct md_rdev *rdev)
Goldwyn Rodrigues1d7e3e92014-06-07 01:53:00 -05009007{
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05009008 struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
9009 struct md_rdev *rdev2;
9010 int role, ret;
9011 char b[BDEVNAME_SIZE];
Goldwyn Rodrigues1d7e3e92014-06-07 01:53:00 -05009012
Guoqing Jiang818da592017-03-01 16:42:40 +08009013 /*
9014 * If size is changed in another node then we need to
9015 * do resize as well.
9016 */
9017 if (mddev->dev_sectors != le64_to_cpu(sb->size)) {
9018 ret = mddev->pers->resize(mddev, le64_to_cpu(sb->size));
9019 if (ret)
9020 pr_info("md-cluster: resize failed\n");
9021 else
9022 bitmap_update_sb(mddev->bitmap);
9023 }
9024
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05009025 /* Check for change of roles in the active devices */
9026 rdev_for_each(rdev2, mddev) {
9027 if (test_bit(Faulty, &rdev2->flags))
9028 continue;
9029
9030 /* Check if the roles changed */
9031 role = le16_to_cpu(sb->dev_roles[rdev2->desc_nr]);
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05009032
9033 if (test_bit(Candidate, &rdev2->flags)) {
9034 if (role == 0xfffe) {
9035 pr_info("md: Removing Candidate device %s because add failed\n", bdevname(rdev2->bdev,b));
9036 md_kick_rdev_from_array(rdev2);
9037 continue;
9038 }
9039 else
9040 clear_bit(Candidate, &rdev2->flags);
9041 }
9042
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05009043 if (role != rdev2->raid_disk) {
9044 /* got activated */
9045 if (rdev2->raid_disk == -1 && role != 0xffff) {
9046 rdev2->saved_raid_disk = role;
9047 ret = remove_and_add_spares(mddev, rdev2);
9048 pr_info("Activated spare: %s\n",
NeilBrown9d487392016-11-02 14:16:49 +11009049 bdevname(rdev2->bdev,b));
Guoqing Jianga5781832016-05-02 11:33:14 -04009050 /* wakeup mddev->thread here, so array could
9051 * perform resync with the new activated disk */
9052 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
9053 md_wakeup_thread(mddev->thread);
9054
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05009055 }
9056 /* device faulty
9057 * We just want to do the minimum to mark the disk
9058 * as faulty. The recovery is performed by the
9059 * one who initiated the error.
9060 */
9061 if ((role == 0xfffe) || (role == 0xfffd)) {
9062 md_error(mddev, rdev2);
9063 clear_bit(Blocked, &rdev2->flags);
9064 }
9065 }
Goldwyn Rodrigues1d7e3e92014-06-07 01:53:00 -05009066 }
9067
Goldwyn Rodrigues28c1b9f2015-10-22 16:01:25 +11009068 if (mddev->raid_disks != le32_to_cpu(sb->raid_disks))
9069 update_raid_disks(mddev, le32_to_cpu(sb->raid_disks));
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05009070
9071 /* Finally set the event to be up to date */
9072 mddev->events = le64_to_cpu(sb->events);
9073}
9074
9075static int read_rdev(struct mddev *mddev, struct md_rdev *rdev)
9076{
9077 int err;
9078 struct page *swapout = rdev->sb_page;
9079 struct mdp_superblock_1 *sb;
9080
9081 /* Store the sb page of the rdev in the swapout temporary
9082 * variable in case we err in the future
9083 */
9084 rdev->sb_page = NULL;
NeilBrown7f0f0d82016-11-02 14:16:49 +11009085 err = alloc_disk_sb(rdev);
9086 if (err == 0) {
9087 ClearPageUptodate(rdev->sb_page);
9088 rdev->sb_loaded = 0;
9089 err = super_types[mddev->major_version].
9090 load_super(rdev, NULL, mddev->minor_version);
9091 }
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05009092 if (err < 0) {
9093 pr_warn("%s: %d Could not reload rdev(%d) err: %d. Restoring old values\n",
9094 __func__, __LINE__, rdev->desc_nr, err);
NeilBrown7f0f0d82016-11-02 14:16:49 +11009095 if (rdev->sb_page)
9096 put_page(rdev->sb_page);
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05009097 rdev->sb_page = swapout;
9098 rdev->sb_loaded = 1;
9099 return err;
9100 }
9101
9102 sb = page_address(rdev->sb_page);
9103 /* Read the offset unconditionally, even if MD_FEATURE_RECOVERY_OFFSET
9104 * is not set
9105 */
9106
9107 if ((le32_to_cpu(sb->feature_map) & MD_FEATURE_RECOVERY_OFFSET))
9108 rdev->recovery_offset = le64_to_cpu(sb->recovery_offset);
9109
9110 /* The other node finished recovery, call spare_active to set
9111 * device In_sync and mddev->degraded
9112 */
9113 if (rdev->recovery_offset == MaxSector &&
9114 !test_bit(In_sync, &rdev->flags) &&
9115 mddev->pers->spare_active(mddev))
9116 sysfs_notify(&mddev->kobj, NULL, "degraded");
9117
9118 put_page(swapout);
9119 return 0;
9120}
9121
9122void md_reload_sb(struct mddev *mddev, int nr)
9123{
9124 struct md_rdev *rdev;
9125 int err;
9126
9127 /* Find the rdev */
9128 rdev_for_each_rcu(rdev, mddev) {
9129 if (rdev->desc_nr == nr)
9130 break;
9131 }
9132
9133 if (!rdev || rdev->desc_nr != nr) {
9134 pr_warn("%s: %d Could not find rdev with nr %d\n", __func__, __LINE__, nr);
9135 return;
9136 }
9137
9138 err = read_rdev(mddev, rdev);
9139 if (err < 0)
9140 return;
9141
9142 check_sb_changes(mddev, rdev);
9143
9144 /* Read all rdev's to update recovery_offset */
9145 rdev_for_each_rcu(rdev, mddev)
9146 read_rdev(mddev, rdev);
Goldwyn Rodrigues1d7e3e92014-06-07 01:53:00 -05009147}
9148EXPORT_SYMBOL(md_reload_sb);
9149
Linus Torvalds1da177e2005-04-16 15:20:36 -07009150#ifndef MODULE
9151
9152/*
9153 * Searches all registered partitions for autorun RAID arrays
9154 * at boot time.
9155 */
Michael J. Evans4d936ec2007-10-16 23:30:52 -07009156
Cong Wang5b1f5bc32016-06-08 09:20:16 -07009157static DEFINE_MUTEX(detected_devices_mutex);
Michael J. Evans4d936ec2007-10-16 23:30:52 -07009158static LIST_HEAD(all_detected_devices);
9159struct detected_devices_node {
9160 struct list_head list;
9161 dev_t dev;
9162};
Linus Torvalds1da177e2005-04-16 15:20:36 -07009163
9164void md_autodetect_dev(dev_t dev)
9165{
Michael J. Evans4d936ec2007-10-16 23:30:52 -07009166 struct detected_devices_node *node_detected_dev;
9167
9168 node_detected_dev = kzalloc(sizeof(*node_detected_dev), GFP_KERNEL);
9169 if (node_detected_dev) {
9170 node_detected_dev->dev = dev;
Cong Wang5b1f5bc32016-06-08 09:20:16 -07009171 mutex_lock(&detected_devices_mutex);
Michael J. Evans4d936ec2007-10-16 23:30:52 -07009172 list_add_tail(&node_detected_dev->list, &all_detected_devices);
Cong Wang5b1f5bc32016-06-08 09:20:16 -07009173 mutex_unlock(&detected_devices_mutex);
Michael J. Evans4d936ec2007-10-16 23:30:52 -07009174 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07009175}
9176
Linus Torvalds1da177e2005-04-16 15:20:36 -07009177static void autostart_arrays(int part)
9178{
NeilBrown3cb03002011-10-11 16:45:26 +11009179 struct md_rdev *rdev;
Michael J. Evans4d936ec2007-10-16 23:30:52 -07009180 struct detected_devices_node *node_detected_dev;
9181 dev_t dev;
9182 int i_scanned, i_passed;
9183
9184 i_scanned = 0;
9185 i_passed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009186
NeilBrown9d487392016-11-02 14:16:49 +11009187 pr_info("md: Autodetecting RAID arrays.\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07009188
Cong Wang5b1f5bc32016-06-08 09:20:16 -07009189 mutex_lock(&detected_devices_mutex);
Michael J. Evans4d936ec2007-10-16 23:30:52 -07009190 while (!list_empty(&all_detected_devices) && i_scanned < INT_MAX) {
9191 i_scanned++;
9192 node_detected_dev = list_entry(all_detected_devices.next,
9193 struct detected_devices_node, list);
9194 list_del(&node_detected_dev->list);
9195 dev = node_detected_dev->dev;
9196 kfree(node_detected_dev);
Shaohua Li90bcf1332016-09-14 14:26:54 -07009197 mutex_unlock(&detected_devices_mutex);
NeilBrowndf968c42007-07-17 04:06:11 -07009198 rdev = md_import_device(dev,0, 90);
Shaohua Li90bcf1332016-09-14 14:26:54 -07009199 mutex_lock(&detected_devices_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009200 if (IS_ERR(rdev))
9201 continue;
9202
NeilBrown403df472014-09-30 15:52:29 +10009203 if (test_bit(Faulty, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07009204 continue;
NeilBrown403df472014-09-30 15:52:29 +10009205
NeilBrownd0fae182008-03-04 14:29:31 -08009206 set_bit(AutoDetected, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009207 list_add(&rdev->same_set, &pending_raid_disks);
Michael J. Evans4d936ec2007-10-16 23:30:52 -07009208 i_passed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009209 }
Cong Wang5b1f5bc32016-06-08 09:20:16 -07009210 mutex_unlock(&detected_devices_mutex);
Michael J. Evans4d936ec2007-10-16 23:30:52 -07009211
NeilBrown9d487392016-11-02 14:16:49 +11009212 pr_debug("md: Scanned %d and added %d devices.\n", i_scanned, i_passed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009213
9214 autorun_devices(part);
9215}
9216
Jeff Garzikfdee8ae2006-12-10 02:20:50 -08009217#endif /* !MODULE */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009218
9219static __exit void md_exit(void)
9220{
NeilBrownfd01b882011-10-11 16:47:53 +11009221 struct mddev *mddev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009222 struct list_head *tmp;
NeilBrowne2f23b62014-04-09 14:33:51 +10009223 int delay = 1;
Greg Kroah-Hartman8ab5e4c2005-06-20 21:15:16 -07009224
NeilBrownaf5628f2014-07-31 13:54:54 +10009225 blk_unregister_region(MKDEV(MD_MAJOR,0), 512);
NeilBrowne8703fe2006-10-03 01:15:59 -07009226 blk_unregister_region(MKDEV(mdp_major,0), 1U << MINORBITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009227
Christoph Hellwig3dbd8c22009-03-31 14:27:02 +11009228 unregister_blkdev(MD_MAJOR,"md");
Linus Torvalds1da177e2005-04-16 15:20:36 -07009229 unregister_blkdev(mdp_major, "mdp");
9230 unregister_reboot_notifier(&md_notifier);
9231 unregister_sysctl_table(raid_table_header);
NeilBrowne2f23b62014-04-09 14:33:51 +10009232
9233 /* We cannot unload the modules while some process is
9234 * waiting for us in select() or poll() - wake them up
9235 */
9236 md_unloading = 1;
9237 while (waitqueue_active(&md_event_waiters)) {
9238 /* not safe to leave yet */
9239 wake_up(&md_event_waiters);
9240 msleep(delay);
9241 delay += delay;
9242 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07009243 remove_proc_entry("mdstat", NULL);
NeilBrowne2f23b62014-04-09 14:33:51 +10009244
NeilBrown29ac4aa2008-02-06 01:39:58 -08009245 for_each_mddev(mddev, tmp) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07009246 export_array(mddev);
NeilBrown93568632017-02-06 13:41:39 +11009247 mddev->ctime = 0;
NeilBrownd3374822009-01-09 08:31:10 +11009248 mddev->hold_active = 0;
NeilBrown93568632017-02-06 13:41:39 +11009249 /*
9250 * for_each_mddev() will call mddev_put() at the end of each
9251 * iteration. As the mddev is now fully clear, this will
9252 * schedule the mddev for destruction by a workqueue, and the
9253 * destroy_workqueue() below will wait for that to complete.
9254 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009255 }
Tejun Heoe804ac72010-10-15 15:36:08 +02009256 destroy_workqueue(md_misc_wq);
9257 destroy_workqueue(md_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009258}
9259
Dan Williams685784a2007-07-09 11:56:42 -07009260subsys_initcall(md_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009261module_exit(md_exit)
9262
NeilBrownf91de922005-11-08 21:39:36 -08009263static int get_ro(char *buffer, struct kernel_param *kp)
9264{
9265 return sprintf(buffer, "%d", start_readonly);
9266}
9267static int set_ro(const char *val, struct kernel_param *kp)
9268{
Alexey Dobriyan4c9309c2015-05-16 14:02:38 +03009269 return kstrtouint(val, 10, (unsigned int *)&start_readonly);
NeilBrownf91de922005-11-08 21:39:36 -08009270}
9271
NeilBrown80ca3a42006-07-10 04:44:18 -07009272module_param_call(start_ro, set_ro, get_ro, NULL, S_IRUSR|S_IWUSR);
9273module_param(start_dirty_degraded, int, S_IRUGO|S_IWUSR);
NeilBrownefeb53c2009-01-09 08:31:10 +11009274module_param_call(new_array, add_named_array, NULL, NULL, S_IWUSR);
NeilBrown78b63502017-04-12 16:26:13 +10009275module_param(create_on_open, bool, S_IRUSR|S_IWUSR);
NeilBrownf91de922005-11-08 21:39:36 -08009276
Linus Torvalds1da177e2005-04-16 15:20:36 -07009277MODULE_LICENSE("GPL");
NeilBrown0efb9e62009-12-14 12:49:58 +11009278MODULE_DESCRIPTION("MD RAID framework");
NeilBrownaa1595e2005-08-04 12:53:32 -07009279MODULE_ALIAS("md");
NeilBrown72008652005-08-26 18:34:15 -07009280MODULE_ALIAS_BLOCKDEV_MAJOR(MD_MAJOR);