blob: b19b551bb34bbf3da6b9ab9c4a0af69de71488b3 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03008 * Base on code in raid1.c. See raid1.c for further copyright information.
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Stephen Rothwell25570722008-10-15 09:09:21 +110022#include <linux/delay.h>
NeilBrownbff61972009-03-31 14:33:13 +110023#include <linux/blkdev.h>
Paul Gortmaker056075c2011-07-03 13:58:33 -040024#include <linux/module.h>
NeilBrownbff61972009-03-31 14:33:13 +110025#include <linux/seq_file.h>
Christian Dietrich8bda4702011-07-27 11:00:36 +100026#include <linux/ratelimit.h>
NeilBrown3ea7daa2012-05-22 13:53:47 +100027#include <linux/kthread.h>
NeilBrown43b2e5d2009-03-31 14:33:13 +110028#include "md.h"
Christoph Hellwigef740c32009-03-31 14:27:03 +110029#include "raid10.h"
Trela, Maciejdab8b292010-03-08 16:02:45 +110030#include "raid0.h"
Christoph Hellwigef740c32009-03-31 14:27:03 +110031#include "bitmap.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070032
33/*
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
36 * chunk_size
37 * raid_disks
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
NeilBrownc93983b2006-06-26 00:27:41 -070040 * far_offset (stored in bit 16 of layout )
Jonathan Brassow475901a2013-02-21 13:28:10 +110041 * use_far_sets (stored in bit 17 of layout )
NeilBrown8bce6d32015-10-22 13:20:15 +110042 * use_far_sets_bugfixed (stored in bit 18 of layout )
Linus Torvalds1da177e2005-04-16 15:20:36 -070043 *
Jonathan Brassow475901a2013-02-21 13:28:10 +110044 * The data to be stored is divided into chunks using chunksize. Each device
45 * is divided into far_copies sections. In each section, chunks are laid out
46 * in a style similar to raid0, but near_copies copies of each chunk is stored
47 * (each on a different drive). The starting device for each section is offset
48 * near_copies from the starting device of the previous section. Thus there
49 * are (near_copies * far_copies) of each chunk, and each is on a different
50 * drive. near_copies and far_copies must be at least one, and their product
51 * is at most raid_disks.
NeilBrownc93983b2006-06-26 00:27:41 -070052 *
53 * If far_offset is true, then the far_copies are handled a bit differently.
Jonathan Brassow475901a2013-02-21 13:28:10 +110054 * The copies are still in different stripes, but instead of being very far
55 * apart on disk, there are adjacent stripes.
56 *
57 * The far and offset algorithms are handled slightly differently if
58 * 'use_far_sets' is true. In this case, the array's devices are grouped into
59 * sets that are (near_copies * far_copies) in size. The far copied stripes
60 * are still shifted by 'near_copies' devices, but this shifting stays confined
61 * to the set rather than the entire array. This is done to improve the number
62 * of device combinations that can fail without causing the array to fail.
63 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
64 * on a device):
65 * A B C D A B C D E
66 * ... ...
67 * D A B C E A B C D
68 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
69 * [A B] [C D] [A B] [C D E]
70 * |...| |...| |...| | ... |
71 * [B A] [D C] [B A] [E C D]
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 */
73
74/*
75 * Number of guaranteed r10bios in case of extreme VM load:
76 */
77#define NR_RAID10_BIOS 256
78
Jonathan Brassow473e87c2012-07-31 10:03:52 +100079/* when we get a read error on a read-only array, we redirect to another
80 * device without failing the first device, or trying to over-write to
81 * correct the read error. To keep track of bad blocks on a per-bio
82 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
83 */
84#define IO_BLOCKED ((struct bio *)1)
85/* When we successfully write to a known bad-block, we need to remove the
86 * bad-block marking which must be done from process context. So we record
87 * the success by setting devs[n].bio to IO_MADE_GOOD
88 */
89#define IO_MADE_GOOD ((struct bio *)2)
90
91#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92
93/* When there are this many requests queued to be written by
NeilBrown34db0cd2011-10-11 16:50:01 +110094 * the raid10 thread, we become 'congested' to provide back-pressure
95 * for writeback.
96 */
97static int max_queued_requests = 1024;
98
NeilBrowne879a872011-10-11 16:49:02 +110099static void allow_barrier(struct r10conf *conf);
100static void lower_barrier(struct r10conf *conf);
NeilBrown635f6412013-06-11 14:57:09 +1000101static int _enough(struct r10conf *conf, int previous, int ignore);
NeilBrown3ea7daa2012-05-22 13:53:47 +1000102static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
103 int *skipped);
104static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200105static void end_reshape_write(struct bio *bio);
NeilBrown3ea7daa2012-05-22 13:53:47 +1000106static void end_reshape(struct r10conf *conf);
NeilBrown0a27ec92006-01-06 00:20:13 -0800107
Al Virodd0fc662005-10-07 07:46:04 +0100108static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109{
NeilBrowne879a872011-10-11 16:49:02 +1100110 struct r10conf *conf = data;
NeilBrown9f2c9d12011-10-11 16:48:43 +1100111 int size = offsetof(struct r10bio, devs[conf->copies]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112
NeilBrown69335ef2011-12-23 10:17:54 +1100113 /* allocate a r10bio with room for raid_disks entries in the
114 * bios array */
Jens Axboe7eaceac2011-03-10 08:52:07 +0100115 return kzalloc(size, gfp_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116}
117
118static void r10bio_pool_free(void *r10_bio, void *data)
119{
120 kfree(r10_bio);
121}
122
NeilBrown0310fa22008-08-05 15:54:14 +1000123/* Maximum size of each resync request */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124#define RESYNC_BLOCK_SIZE (64*1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
NeilBrown0310fa22008-08-05 15:54:14 +1000126/* amount of memory to reserve for resync requests */
127#define RESYNC_WINDOW (1024*1024)
128/* maximum number of concurrent requests, memory permitting */
129#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130
131/*
132 * When performing a resync, we need to read and compare, so
133 * we need as many pages are there are copies.
134 * When performing a recovery, we need 2 bios, one for read,
135 * one for write (we recover only one drive per r10buf)
136 *
137 */
Al Virodd0fc662005-10-07 07:46:04 +0100138static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139{
NeilBrowne879a872011-10-11 16:49:02 +1100140 struct r10conf *conf = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 struct page *page;
NeilBrown9f2c9d12011-10-11 16:48:43 +1100142 struct r10bio *r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 struct bio *bio;
144 int i, j;
145 int nalloc;
146
147 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
Jens Axboe7eaceac2011-03-10 08:52:07 +0100148 if (!r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150
NeilBrown3ea7daa2012-05-22 13:53:47 +1000151 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
152 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 nalloc = conf->copies; /* resync */
154 else
155 nalloc = 2; /* recovery */
156
157 /*
158 * Allocate bios.
159 */
160 for (j = nalloc ; j-- ; ) {
NeilBrown67465572010-10-26 17:33:54 +1100161 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162 if (!bio)
163 goto out_free_bio;
164 r10_bio->devs[j].bio = bio;
NeilBrown69335ef2011-12-23 10:17:54 +1100165 if (!conf->have_replacement)
166 continue;
167 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
168 if (!bio)
169 goto out_free_bio;
170 r10_bio->devs[j].repl_bio = bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 }
172 /*
173 * Allocate RESYNC_PAGES data pages and attach them
174 * where needed.
175 */
176 for (j = 0 ; j < nalloc; j++) {
NeilBrown69335ef2011-12-23 10:17:54 +1100177 struct bio *rbio = r10_bio->devs[j].repl_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 bio = r10_bio->devs[j].bio;
179 for (i = 0; i < RESYNC_PAGES; i++) {
NeilBrown3ea7daa2012-05-22 13:53:47 +1000180 if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
181 &conf->mddev->recovery)) {
182 /* we can share bv_page's during recovery
183 * and reshape */
Namhyung Kimc65060a2011-07-18 17:38:49 +1000184 struct bio *rbio = r10_bio->devs[0].bio;
185 page = rbio->bi_io_vec[i].bv_page;
186 get_page(page);
187 } else
188 page = alloc_page(gfp_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 if (unlikely(!page))
190 goto out_free_pages;
191
192 bio->bi_io_vec[i].bv_page = page;
NeilBrown69335ef2011-12-23 10:17:54 +1100193 if (rbio)
194 rbio->bi_io_vec[i].bv_page = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195 }
196 }
197
198 return r10_bio;
199
200out_free_pages:
201 for ( ; i > 0 ; i--)
NeilBrown1345b1d2006-01-06 00:20:40 -0800202 safe_put_page(bio->bi_io_vec[i-1].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203 while (j--)
204 for (i = 0; i < RESYNC_PAGES ; i++)
NeilBrown1345b1d2006-01-06 00:20:40 -0800205 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
majianpeng5fdd2cf2012-05-22 13:55:03 +1000206 j = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207out_free_bio:
majianpeng5fdd2cf2012-05-22 13:55:03 +1000208 for ( ; j < nalloc; j++) {
209 if (r10_bio->devs[j].bio)
210 bio_put(r10_bio->devs[j].bio);
NeilBrown69335ef2011-12-23 10:17:54 +1100211 if (r10_bio->devs[j].repl_bio)
212 bio_put(r10_bio->devs[j].repl_bio);
213 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214 r10bio_pool_free(r10_bio, conf);
215 return NULL;
216}
217
218static void r10buf_pool_free(void *__r10_bio, void *data)
219{
220 int i;
NeilBrowne879a872011-10-11 16:49:02 +1100221 struct r10conf *conf = data;
NeilBrown9f2c9d12011-10-11 16:48:43 +1100222 struct r10bio *r10bio = __r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 int j;
224
225 for (j=0; j < conf->copies; j++) {
226 struct bio *bio = r10bio->devs[j].bio;
227 if (bio) {
228 for (i = 0; i < RESYNC_PAGES; i++) {
NeilBrown1345b1d2006-01-06 00:20:40 -0800229 safe_put_page(bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230 bio->bi_io_vec[i].bv_page = NULL;
231 }
232 bio_put(bio);
233 }
NeilBrown69335ef2011-12-23 10:17:54 +1100234 bio = r10bio->devs[j].repl_bio;
235 if (bio)
236 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 }
238 r10bio_pool_free(r10bio, conf);
239}
240
NeilBrowne879a872011-10-11 16:49:02 +1100241static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242{
243 int i;
244
245 for (i = 0; i < conf->copies; i++) {
246 struct bio **bio = & r10_bio->devs[i].bio;
NeilBrown749c55e2011-07-28 11:39:24 +1000247 if (!BIO_SPECIAL(*bio))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248 bio_put(*bio);
249 *bio = NULL;
NeilBrown69335ef2011-12-23 10:17:54 +1100250 bio = &r10_bio->devs[i].repl_bio;
251 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
252 bio_put(*bio);
253 *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 }
255}
256
NeilBrown9f2c9d12011-10-11 16:48:43 +1100257static void free_r10bio(struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258{
NeilBrowne879a872011-10-11 16:49:02 +1100259 struct r10conf *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261 put_all_bios(conf, r10_bio);
262 mempool_free(r10_bio, conf->r10bio_pool);
263}
264
NeilBrown9f2c9d12011-10-11 16:48:43 +1100265static void put_buf(struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266{
NeilBrowne879a872011-10-11 16:49:02 +1100267 struct r10conf *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
269 mempool_free(r10_bio, conf->r10buf_pool);
270
NeilBrown0a27ec92006-01-06 00:20:13 -0800271 lower_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272}
273
NeilBrown9f2c9d12011-10-11 16:48:43 +1100274static void reschedule_retry(struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275{
276 unsigned long flags;
NeilBrownfd01b882011-10-11 16:47:53 +1100277 struct mddev *mddev = r10_bio->mddev;
NeilBrowne879a872011-10-11 16:49:02 +1100278 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
280 spin_lock_irqsave(&conf->device_lock, flags);
281 list_add(&r10_bio->retry_list, &conf->retry_list);
NeilBrown4443ae12006-01-06 00:20:28 -0800282 conf->nr_queued ++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283 spin_unlock_irqrestore(&conf->device_lock, flags);
284
Arthur Jones388667b2008-07-25 12:03:38 -0700285 /* wake up frozen array... */
286 wake_up(&conf->wait_barrier);
287
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288 md_wakeup_thread(mddev->thread);
289}
290
291/*
292 * raid_end_bio_io() is called when we have finished servicing a mirrored
293 * operation and are ready to return a success/failure code to the buffer
294 * cache layer.
295 */
NeilBrown9f2c9d12011-10-11 16:48:43 +1100296static void raid_end_bio_io(struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297{
298 struct bio *bio = r10_bio->master_bio;
NeilBrown856e08e2011-07-28 11:39:23 +1000299 int done;
NeilBrowne879a872011-10-11 16:49:02 +1100300 struct r10conf *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301
NeilBrown856e08e2011-07-28 11:39:23 +1000302 if (bio->bi_phys_segments) {
303 unsigned long flags;
304 spin_lock_irqsave(&conf->device_lock, flags);
305 bio->bi_phys_segments--;
306 done = (bio->bi_phys_segments == 0);
307 spin_unlock_irqrestore(&conf->device_lock, flags);
308 } else
309 done = 1;
310 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200311 bio->bi_error = -EIO;
NeilBrown856e08e2011-07-28 11:39:23 +1000312 if (done) {
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200313 bio_endio(bio);
NeilBrown856e08e2011-07-28 11:39:23 +1000314 /*
315 * Wake up any possible resync thread that waits for the device
316 * to go idle.
317 */
318 allow_barrier(conf);
319 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 free_r10bio(r10_bio);
321}
322
323/*
324 * Update disk head position estimator based on IRQ completion info.
325 */
NeilBrown9f2c9d12011-10-11 16:48:43 +1100326static inline void update_head_pos(int slot, struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327{
NeilBrowne879a872011-10-11 16:49:02 +1100328 struct r10conf *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329
330 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
331 r10_bio->devs[slot].addr + (r10_bio->sectors);
332}
333
Namhyung Kim778ca012011-07-18 17:38:47 +1000334/*
335 * Find the disk number which triggered given bio
336 */
NeilBrowne879a872011-10-11 16:49:02 +1100337static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
NeilBrown69335ef2011-12-23 10:17:54 +1100338 struct bio *bio, int *slotp, int *replp)
Namhyung Kim778ca012011-07-18 17:38:47 +1000339{
340 int slot;
NeilBrown69335ef2011-12-23 10:17:54 +1100341 int repl = 0;
Namhyung Kim778ca012011-07-18 17:38:47 +1000342
NeilBrown69335ef2011-12-23 10:17:54 +1100343 for (slot = 0; slot < conf->copies; slot++) {
Namhyung Kim778ca012011-07-18 17:38:47 +1000344 if (r10_bio->devs[slot].bio == bio)
345 break;
NeilBrown69335ef2011-12-23 10:17:54 +1100346 if (r10_bio->devs[slot].repl_bio == bio) {
347 repl = 1;
348 break;
349 }
350 }
Namhyung Kim778ca012011-07-18 17:38:47 +1000351
352 BUG_ON(slot == conf->copies);
353 update_head_pos(slot, r10_bio);
354
NeilBrown749c55e2011-07-28 11:39:24 +1000355 if (slotp)
356 *slotp = slot;
NeilBrown69335ef2011-12-23 10:17:54 +1100357 if (replp)
358 *replp = repl;
Namhyung Kim778ca012011-07-18 17:38:47 +1000359 return r10_bio->devs[slot].devnum;
360}
361
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200362static void raid10_end_read_request(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363{
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200364 int uptodate = !bio->bi_error;
NeilBrown9f2c9d12011-10-11 16:48:43 +1100365 struct r10bio *r10_bio = bio->bi_private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366 int slot, dev;
NeilBrownabbf0982011-12-23 10:17:54 +1100367 struct md_rdev *rdev;
NeilBrowne879a872011-10-11 16:49:02 +1100368 struct r10conf *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370 slot = r10_bio->read_slot;
371 dev = r10_bio->devs[slot].devnum;
NeilBrownabbf0982011-12-23 10:17:54 +1100372 rdev = r10_bio->devs[slot].rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373 /*
374 * this branch is our 'one mirror IO has finished' event handler:
375 */
NeilBrown4443ae12006-01-06 00:20:28 -0800376 update_head_pos(slot, r10_bio);
377
378 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 /*
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
383 *
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
387 */
388 set_bit(R10BIO_Uptodate, &r10_bio->state);
NeilBrownfae8cc52012-02-14 11:10:10 +1100389 } else {
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
394 */
NeilBrown635f6412013-06-11 14:57:09 +1000395 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
396 rdev->raid_disk))
NeilBrownfae8cc52012-02-14 11:10:10 +1100397 uptodate = 1;
NeilBrownfae8cc52012-02-14 11:10:10 +1100398 }
399 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 raid_end_bio_io(r10_bio);
NeilBrownabbf0982011-12-23 10:17:54 +1100401 rdev_dec_pending(rdev, conf->mddev);
NeilBrown4443ae12006-01-06 00:20:28 -0800402 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 /*
NeilBrown7c4e06f2011-05-11 14:53:17 +1000404 * oops, read error - keep the refcount on the rdev
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 */
406 char b[BDEVNAME_SIZE];
Christian Dietrich8bda4702011-07-27 11:00:36 +1000407 printk_ratelimited(KERN_ERR
408 "md/raid10:%s: %s: rescheduling sector %llu\n",
409 mdname(conf->mddev),
NeilBrownabbf0982011-12-23 10:17:54 +1100410 bdevname(rdev->bdev, b),
Christian Dietrich8bda4702011-07-27 11:00:36 +1000411 (unsigned long long)r10_bio->sector);
NeilBrown856e08e2011-07-28 11:39:23 +1000412 set_bit(R10BIO_ReadError, &r10_bio->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 reschedule_retry(r10_bio);
414 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415}
416
NeilBrown9f2c9d12011-10-11 16:48:43 +1100417static void close_write(struct r10bio *r10_bio)
NeilBrownbd870a12011-07-28 11:39:24 +1000418{
419 /* clear the bitmap if all writes complete successfully */
420 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
421 r10_bio->sectors,
422 !test_bit(R10BIO_Degraded, &r10_bio->state),
423 0);
424 md_write_end(r10_bio->mddev);
425}
426
NeilBrown9f2c9d12011-10-11 16:48:43 +1100427static void one_write_done(struct r10bio *r10_bio)
NeilBrown19d5f832011-09-10 17:21:17 +1000428{
429 if (atomic_dec_and_test(&r10_bio->remaining)) {
430 if (test_bit(R10BIO_WriteError, &r10_bio->state))
431 reschedule_retry(r10_bio);
432 else {
433 close_write(r10_bio);
434 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
435 reschedule_retry(r10_bio);
436 else
437 raid_end_bio_io(r10_bio);
438 }
439 }
440}
441
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200442static void raid10_end_write_request(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443{
NeilBrown9f2c9d12011-10-11 16:48:43 +1100444 struct r10bio *r10_bio = bio->bi_private;
Namhyung Kim778ca012011-07-18 17:38:47 +1000445 int dev;
NeilBrown749c55e2011-07-28 11:39:24 +1000446 int dec_rdev = 1;
NeilBrowne879a872011-10-11 16:49:02 +1100447 struct r10conf *conf = r10_bio->mddev->private;
NeilBrown475b0322011-12-23 10:17:55 +1100448 int slot, repl;
NeilBrown4ca40c22011-12-23 10:17:55 +1100449 struct md_rdev *rdev = NULL;
Shaohua Li579ed342016-10-06 14:13:52 -0700450 bool discard_error;
451
452 discard_error = bio->bi_error && bio_op(bio) == REQ_OP_DISCARD;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453
NeilBrown475b0322011-12-23 10:17:55 +1100454 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
NeilBrown475b0322011-12-23 10:17:55 +1100456 if (repl)
457 rdev = conf->mirrors[dev].replacement;
NeilBrown4ca40c22011-12-23 10:17:55 +1100458 if (!rdev) {
459 smp_rmb();
460 repl = 0;
NeilBrown475b0322011-12-23 10:17:55 +1100461 rdev = conf->mirrors[dev].rdev;
NeilBrown4ca40c22011-12-23 10:17:55 +1100462 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 /*
464 * this branch is our 'one mirror IO has finished' event handler:
465 */
Shaohua Li579ed342016-10-06 14:13:52 -0700466 if (bio->bi_error && !discard_error) {
NeilBrown475b0322011-12-23 10:17:55 +1100467 if (repl)
468 /* Never record new bad blocks to replacement,
469 * just fail it.
470 */
471 md_error(rdev->mddev, rdev);
472 else {
473 set_bit(WriteErrorSeen, &rdev->flags);
NeilBrownb7044d42011-12-23 10:17:56 +1100474 if (!test_and_set_bit(WantReplacement, &rdev->flags))
475 set_bit(MD_RECOVERY_NEEDED,
476 &rdev->mddev->recovery);
NeilBrown475b0322011-12-23 10:17:55 +1100477 set_bit(R10BIO_WriteError, &r10_bio->state);
478 dec_rdev = 0;
479 }
NeilBrown749c55e2011-07-28 11:39:24 +1000480 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 /*
482 * Set R10BIO_Uptodate in our master bio, so that
483 * we will return a good error code for to the higher
484 * levels even if IO on some other mirrored buffer fails.
485 *
486 * The 'master' represents the composite IO operation to
487 * user-side. So if something waits for IO, then it will
488 * wait for the 'master' bio.
489 */
NeilBrown749c55e2011-07-28 11:39:24 +1000490 sector_t first_bad;
491 int bad_sectors;
492
Alex Lyakas3056e3a2013-06-04 20:42:21 +0300493 /*
494 * Do not set R10BIO_Uptodate if the current device is
495 * rebuilding or Faulty. This is because we cannot use
496 * such device for properly reading the data back (we could
497 * potentially use it, if the current write would have felt
498 * before rdev->recovery_offset, but for simplicity we don't
499 * check this here.
500 */
501 if (test_bit(In_sync, &rdev->flags) &&
502 !test_bit(Faulty, &rdev->flags))
503 set_bit(R10BIO_Uptodate, &r10_bio->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504
NeilBrown749c55e2011-07-28 11:39:24 +1000505 /* Maybe we can clear some bad blocks. */
NeilBrown475b0322011-12-23 10:17:55 +1100506 if (is_badblock(rdev,
NeilBrown749c55e2011-07-28 11:39:24 +1000507 r10_bio->devs[slot].addr,
508 r10_bio->sectors,
Shaohua Li579ed342016-10-06 14:13:52 -0700509 &first_bad, &bad_sectors) && !discard_error) {
NeilBrown749c55e2011-07-28 11:39:24 +1000510 bio_put(bio);
NeilBrown475b0322011-12-23 10:17:55 +1100511 if (repl)
512 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
513 else
514 r10_bio->devs[slot].bio = IO_MADE_GOOD;
NeilBrown749c55e2011-07-28 11:39:24 +1000515 dec_rdev = 0;
516 set_bit(R10BIO_MadeGood, &r10_bio->state);
517 }
518 }
519
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 /*
521 *
522 * Let's see if all mirrored write operations have finished
523 * already.
524 */
NeilBrown19d5f832011-09-10 17:21:17 +1000525 one_write_done(r10_bio);
NeilBrown749c55e2011-07-28 11:39:24 +1000526 if (dec_rdev)
NeilBrown884162d2012-11-22 15:12:09 +1100527 rdev_dec_pending(rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528}
529
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530/*
531 * RAID10 layout manager
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300532 * As well as the chunksize and raid_disks count, there are two
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 * parameters: near_copies and far_copies.
534 * near_copies * far_copies must be <= raid_disks.
535 * Normally one of these will be 1.
536 * If both are 1, we get raid0.
537 * If near_copies == raid_disks, we get raid1.
538 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300539 * Chunks are laid out in raid0 style with near_copies copies of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 * first chunk, followed by near_copies copies of the next chunk and
541 * so on.
542 * If far_copies > 1, then after 1/far_copies of the array has been assigned
543 * as described above, we start again with a device offset of near_copies.
544 * So we effectively have another copy of the whole array further down all
545 * the drives, but with blocks on different drives.
546 * With this layout, and block is never stored twice on the one device.
547 *
548 * raid10_find_phys finds the sector offset of a given virtual sector
NeilBrownc93983b2006-06-26 00:27:41 -0700549 * on each device that it is on.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 *
551 * raid10_find_virt does the reverse mapping, from a device and a
552 * sector offset to a virtual address
553 */
554
NeilBrownf8c9e742012-05-21 09:28:33 +1000555static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556{
557 int n,f;
558 sector_t sector;
559 sector_t chunk;
560 sector_t stripe;
561 int dev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 int slot = 0;
Jonathan Brassow9a3152a2013-02-21 13:28:10 +1100563 int last_far_set_start, last_far_set_size;
564
565 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
566 last_far_set_start *= geo->far_set_size;
567
568 last_far_set_size = geo->far_set_size;
569 last_far_set_size += (geo->raid_disks % geo->far_set_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570
571 /* now calculate first sector/dev */
NeilBrown5cf00fc2012-05-21 09:28:20 +1000572 chunk = r10bio->sector >> geo->chunk_shift;
573 sector = r10bio->sector & geo->chunk_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574
NeilBrown5cf00fc2012-05-21 09:28:20 +1000575 chunk *= geo->near_copies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 stripe = chunk;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000577 dev = sector_div(stripe, geo->raid_disks);
578 if (geo->far_offset)
579 stripe *= geo->far_copies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580
NeilBrown5cf00fc2012-05-21 09:28:20 +1000581 sector += stripe << geo->chunk_shift;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582
583 /* and calculate all the others */
NeilBrown5cf00fc2012-05-21 09:28:20 +1000584 for (n = 0; n < geo->near_copies; n++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 int d = dev;
Jonathan Brassow475901a2013-02-21 13:28:10 +1100586 int set;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 sector_t s = sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 r10bio->devs[slot].devnum = d;
Jonathan Brassow4c0ca262013-02-21 13:28:09 +1100589 r10bio->devs[slot].addr = s;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 slot++;
591
NeilBrown5cf00fc2012-05-21 09:28:20 +1000592 for (f = 1; f < geo->far_copies; f++) {
Jonathan Brassow475901a2013-02-21 13:28:10 +1100593 set = d / geo->far_set_size;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000594 d += geo->near_copies;
Jonathan Brassow475901a2013-02-21 13:28:10 +1100595
Jonathan Brassow9a3152a2013-02-21 13:28:10 +1100596 if ((geo->raid_disks % geo->far_set_size) &&
597 (d > last_far_set_start)) {
598 d -= last_far_set_start;
599 d %= last_far_set_size;
600 d += last_far_set_start;
601 } else {
602 d %= geo->far_set_size;
603 d += geo->far_set_size * set;
604 }
NeilBrown5cf00fc2012-05-21 09:28:20 +1000605 s += geo->stride;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 r10bio->devs[slot].devnum = d;
607 r10bio->devs[slot].addr = s;
608 slot++;
609 }
610 dev++;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000611 if (dev >= geo->raid_disks) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 dev = 0;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000613 sector += (geo->chunk_mask + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 }
615 }
NeilBrownf8c9e742012-05-21 09:28:33 +1000616}
617
618static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
619{
620 struct geom *geo = &conf->geo;
621
622 if (conf->reshape_progress != MaxSector &&
623 ((r10bio->sector >= conf->reshape_progress) !=
624 conf->mddev->reshape_backwards)) {
625 set_bit(R10BIO_Previous, &r10bio->state);
626 geo = &conf->prev;
627 } else
628 clear_bit(R10BIO_Previous, &r10bio->state);
629
630 __raid10_find_phys(geo, r10bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631}
632
NeilBrowne879a872011-10-11 16:49:02 +1100633static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634{
635 sector_t offset, chunk, vchunk;
NeilBrownf8c9e742012-05-21 09:28:33 +1000636 /* Never use conf->prev as this is only called during resync
637 * or recovery, so reshape isn't happening
638 */
NeilBrown5cf00fc2012-05-21 09:28:20 +1000639 struct geom *geo = &conf->geo;
Jonathan Brassow475901a2013-02-21 13:28:10 +1100640 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
641 int far_set_size = geo->far_set_size;
Jonathan Brassow9a3152a2013-02-21 13:28:10 +1100642 int last_far_set_start;
643
644 if (geo->raid_disks % geo->far_set_size) {
645 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
646 last_far_set_start *= geo->far_set_size;
647
648 if (dev >= last_far_set_start) {
649 far_set_size = geo->far_set_size;
650 far_set_size += (geo->raid_disks % geo->far_set_size);
651 far_set_start = last_far_set_start;
652 }
653 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654
NeilBrown5cf00fc2012-05-21 09:28:20 +1000655 offset = sector & geo->chunk_mask;
656 if (geo->far_offset) {
NeilBrownc93983b2006-06-26 00:27:41 -0700657 int fc;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000658 chunk = sector >> geo->chunk_shift;
659 fc = sector_div(chunk, geo->far_copies);
660 dev -= fc * geo->near_copies;
Jonathan Brassow475901a2013-02-21 13:28:10 +1100661 if (dev < far_set_start)
662 dev += far_set_size;
NeilBrownc93983b2006-06-26 00:27:41 -0700663 } else {
NeilBrown5cf00fc2012-05-21 09:28:20 +1000664 while (sector >= geo->stride) {
665 sector -= geo->stride;
Jonathan Brassow475901a2013-02-21 13:28:10 +1100666 if (dev < (geo->near_copies + far_set_start))
667 dev += far_set_size - geo->near_copies;
NeilBrownc93983b2006-06-26 00:27:41 -0700668 else
NeilBrown5cf00fc2012-05-21 09:28:20 +1000669 dev -= geo->near_copies;
NeilBrownc93983b2006-06-26 00:27:41 -0700670 }
NeilBrown5cf00fc2012-05-21 09:28:20 +1000671 chunk = sector >> geo->chunk_shift;
NeilBrownc93983b2006-06-26 00:27:41 -0700672 }
NeilBrown5cf00fc2012-05-21 09:28:20 +1000673 vchunk = chunk * geo->raid_disks + dev;
674 sector_div(vchunk, geo->near_copies);
675 return (vchunk << geo->chunk_shift) + offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676}
677
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678/*
679 * This routine returns the disk from which the requested read should
680 * be done. There is a per-array 'next expected sequential IO' sector
681 * number - if this matches on the next IO then we use the last disk.
682 * There is also a per-disk 'last know head position' sector that is
683 * maintained from IRQ contexts, both the normal and the resync IO
684 * completion handlers update this position correctly. If there is no
685 * perfect sequential match then we pick the disk whose head is closest.
686 *
687 * If there are 2 mirrors in the same 2 devices, performance degrades
688 * because position is mirror, not device based.
689 *
690 * The rdev for the device selected will have nr_pending incremented.
691 */
692
693/*
694 * FIXME: possibly should rethink readbalancing and do it differently
695 * depending on near_copies / far_copies geometry.
696 */
NeilBrown96c3fd12011-12-23 10:17:54 +1100697static struct md_rdev *read_balance(struct r10conf *conf,
698 struct r10bio *r10_bio,
699 int *max_sectors)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700{
NeilBrownaf3a2cd2010-05-08 08:20:17 +1000701 const sector_t this_sector = r10_bio->sector;
NeilBrown56d99122011-05-11 14:27:03 +1000702 int disk, slot;
NeilBrown856e08e2011-07-28 11:39:23 +1000703 int sectors = r10_bio->sectors;
704 int best_good_sectors;
NeilBrown56d99122011-05-11 14:27:03 +1000705 sector_t new_distance, best_dist;
Jonathan Brassow3bbae042012-07-31 10:03:52 +1000706 struct md_rdev *best_rdev, *rdev = NULL;
NeilBrown56d99122011-05-11 14:27:03 +1000707 int do_balance;
708 int best_slot;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000709 struct geom *geo = &conf->geo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710
711 raid10_find_phys(conf, r10_bio);
712 rcu_read_lock();
NeilBrown856e08e2011-07-28 11:39:23 +1000713 sectors = r10_bio->sectors;
NeilBrown56d99122011-05-11 14:27:03 +1000714 best_slot = -1;
NeilBrownabbf0982011-12-23 10:17:54 +1100715 best_rdev = NULL;
NeilBrown56d99122011-05-11 14:27:03 +1000716 best_dist = MaxSector;
NeilBrown856e08e2011-07-28 11:39:23 +1000717 best_good_sectors = 0;
NeilBrown56d99122011-05-11 14:27:03 +1000718 do_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 /*
720 * Check if we can balance. We can balance on the whole
NeilBrown6cce3b232006-01-06 00:20:16 -0800721 * device if no resync is going on (recovery is ok), or below
722 * the resync window. We take the first readable disk when
723 * above the resync window.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724 */
725 if (conf->mddev->recovery_cp < MaxSector
NeilBrown56d99122011-05-11 14:27:03 +1000726 && (this_sector + sectors >= conf->next_resync))
727 do_balance = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728
NeilBrown56d99122011-05-11 14:27:03 +1000729 for (slot = 0; slot < conf->copies ; slot++) {
NeilBrown856e08e2011-07-28 11:39:23 +1000730 sector_t first_bad;
731 int bad_sectors;
732 sector_t dev_sector;
733
NeilBrown56d99122011-05-11 14:27:03 +1000734 if (r10_bio->devs[slot].bio == IO_BLOCKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735 continue;
NeilBrown56d99122011-05-11 14:27:03 +1000736 disk = r10_bio->devs[slot].devnum;
NeilBrownabbf0982011-12-23 10:17:54 +1100737 rdev = rcu_dereference(conf->mirrors[disk].replacement);
738 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
739 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
740 rdev = rcu_dereference(conf->mirrors[disk].rdev);
NeilBrown050b6612012-03-19 12:46:39 +1100741 if (rdev == NULL ||
Kent Overstreet8ae12662015-04-27 23:48:34 -0700742 test_bit(Faulty, &rdev->flags))
NeilBrownabbf0982011-12-23 10:17:54 +1100743 continue;
744 if (!test_bit(In_sync, &rdev->flags) &&
745 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
NeilBrown56d99122011-05-11 14:27:03 +1000746 continue;
747
NeilBrown856e08e2011-07-28 11:39:23 +1000748 dev_sector = r10_bio->devs[slot].addr;
749 if (is_badblock(rdev, dev_sector, sectors,
750 &first_bad, &bad_sectors)) {
751 if (best_dist < MaxSector)
752 /* Already have a better slot */
753 continue;
754 if (first_bad <= dev_sector) {
755 /* Cannot read here. If this is the
756 * 'primary' device, then we must not read
757 * beyond 'bad_sectors' from another device.
758 */
759 bad_sectors -= (dev_sector - first_bad);
760 if (!do_balance && sectors > bad_sectors)
761 sectors = bad_sectors;
762 if (best_good_sectors > sectors)
763 best_good_sectors = sectors;
764 } else {
765 sector_t good_sectors =
766 first_bad - dev_sector;
767 if (good_sectors > best_good_sectors) {
768 best_good_sectors = good_sectors;
769 best_slot = slot;
NeilBrownabbf0982011-12-23 10:17:54 +1100770 best_rdev = rdev;
NeilBrown856e08e2011-07-28 11:39:23 +1000771 }
772 if (!do_balance)
773 /* Must read from here */
774 break;
775 }
776 continue;
777 } else
778 best_good_sectors = sectors;
779
NeilBrown56d99122011-05-11 14:27:03 +1000780 if (!do_balance)
781 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782
NeilBrown22dfdf52005-11-28 13:44:09 -0800783 /* This optimisation is debatable, and completely destroys
784 * sequential read speed for 'far copies' arrays. So only
785 * keep it for 'near' arrays, and review those later.
786 */
NeilBrown5cf00fc2012-05-21 09:28:20 +1000787 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 break;
Keld Simonsen8ed3a192008-03-04 14:29:34 -0800789
790 /* for far > 1 always use the lowest address */
NeilBrown5cf00fc2012-05-21 09:28:20 +1000791 if (geo->far_copies > 1)
NeilBrown56d99122011-05-11 14:27:03 +1000792 new_distance = r10_bio->devs[slot].addr;
Keld Simonsen8ed3a192008-03-04 14:29:34 -0800793 else
NeilBrown56d99122011-05-11 14:27:03 +1000794 new_distance = abs(r10_bio->devs[slot].addr -
795 conf->mirrors[disk].head_position);
796 if (new_distance < best_dist) {
797 best_dist = new_distance;
798 best_slot = slot;
NeilBrownabbf0982011-12-23 10:17:54 +1100799 best_rdev = rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 }
801 }
NeilBrownabbf0982011-12-23 10:17:54 +1100802 if (slot >= conf->copies) {
NeilBrown56d99122011-05-11 14:27:03 +1000803 slot = best_slot;
NeilBrownabbf0982011-12-23 10:17:54 +1100804 rdev = best_rdev;
805 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806
NeilBrown56d99122011-05-11 14:27:03 +1000807 if (slot >= 0) {
NeilBrown56d99122011-05-11 14:27:03 +1000808 atomic_inc(&rdev->nr_pending);
NeilBrown56d99122011-05-11 14:27:03 +1000809 r10_bio->read_slot = slot;
810 } else
NeilBrown96c3fd12011-12-23 10:17:54 +1100811 rdev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812 rcu_read_unlock();
NeilBrown856e08e2011-07-28 11:39:23 +1000813 *max_sectors = best_good_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814
NeilBrown96c3fd12011-12-23 10:17:54 +1100815 return rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816}
817
NeilBrown5c675f82014-12-15 12:56:56 +1100818static int raid10_congested(struct mddev *mddev, int bits)
NeilBrown0d129222006-10-03 01:15:54 -0700819{
NeilBrowne879a872011-10-11 16:49:02 +1100820 struct r10conf *conf = mddev->private;
NeilBrown0d129222006-10-03 01:15:54 -0700821 int i, ret = 0;
822
Tejun Heo44522262015-05-22 17:13:26 -0400823 if ((bits & (1 << WB_async_congested)) &&
NeilBrown34db0cd2011-10-11 16:50:01 +1100824 conf->pending_count >= max_queued_requests)
825 return 1;
826
NeilBrown0d129222006-10-03 01:15:54 -0700827 rcu_read_lock();
NeilBrownf8c9e742012-05-21 09:28:33 +1000828 for (i = 0;
829 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
830 && ret == 0;
831 i++) {
NeilBrown3cb03002011-10-11 16:45:26 +1100832 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
NeilBrown0d129222006-10-03 01:15:54 -0700833 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200834 struct request_queue *q = bdev_get_queue(rdev->bdev);
NeilBrown0d129222006-10-03 01:15:54 -0700835
836 ret |= bdi_congested(&q->backing_dev_info, bits);
837 }
838 }
839 rcu_read_unlock();
840 return ret;
841}
842
NeilBrowne879a872011-10-11 16:49:02 +1100843static void flush_pending_writes(struct r10conf *conf)
NeilBrowna35e63e2008-03-04 14:29:29 -0800844{
845 /* Any writes that have been queued but are awaiting
846 * bitmap updates get flushed here.
NeilBrowna35e63e2008-03-04 14:29:29 -0800847 */
NeilBrowna35e63e2008-03-04 14:29:29 -0800848 spin_lock_irq(&conf->device_lock);
849
850 if (conf->pending_bio_list.head) {
851 struct bio *bio;
852 bio = bio_list_get(&conf->pending_bio_list);
NeilBrown34db0cd2011-10-11 16:50:01 +1100853 conf->pending_count = 0;
NeilBrowna35e63e2008-03-04 14:29:29 -0800854 spin_unlock_irq(&conf->device_lock);
855 /* flush any pending bitmap writes to disk
856 * before proceeding w/ I/O */
857 bitmap_unplug(conf->mddev->bitmap);
NeilBrown34db0cd2011-10-11 16:50:01 +1100858 wake_up(&conf->wait_barrier);
NeilBrowna35e63e2008-03-04 14:29:29 -0800859
860 while (bio) { /* submit pending writes */
861 struct bio *next = bio->bi_next;
862 bio->bi_next = NULL;
Mike Christie796a5cf2016-06-05 14:32:07 -0500863 if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
Shaohua Li532a2a32012-10-11 13:30:52 +1100864 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
865 /* Just ignore it */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200866 bio_endio(bio);
Shaohua Li532a2a32012-10-11 13:30:52 +1100867 else
868 generic_make_request(bio);
NeilBrowna35e63e2008-03-04 14:29:29 -0800869 bio = next;
870 }
NeilBrowna35e63e2008-03-04 14:29:29 -0800871 } else
872 spin_unlock_irq(&conf->device_lock);
NeilBrowna35e63e2008-03-04 14:29:29 -0800873}
Jens Axboe7eaceac2011-03-10 08:52:07 +0100874
NeilBrown0a27ec92006-01-06 00:20:13 -0800875/* Barriers....
876 * Sometimes we need to suspend IO while we do something else,
877 * either some resync/recovery, or reconfigure the array.
878 * To do this we raise a 'barrier'.
879 * The 'barrier' is a counter that can be raised multiple times
880 * to count how many activities are happening which preclude
881 * normal IO.
882 * We can only raise the barrier if there is no pending IO.
883 * i.e. if nr_pending == 0.
884 * We choose only to raise the barrier if no-one is waiting for the
885 * barrier to go down. This means that as soon as an IO request
886 * is ready, no other operations which require a barrier will start
887 * until the IO request has had a chance.
888 *
889 * So: regular IO calls 'wait_barrier'. When that returns there
890 * is no backgroup IO happening, It must arrange to call
891 * allow_barrier when it has finished its IO.
892 * backgroup IO calls must call raise_barrier. Once that returns
893 * there is no normal IO happeing. It must arrange to call
894 * lower_barrier when the particular background IO completes.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896
NeilBrowne879a872011-10-11 16:49:02 +1100897static void raise_barrier(struct r10conf *conf, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898{
NeilBrown6cce3b232006-01-06 00:20:16 -0800899 BUG_ON(force && !conf->barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 spin_lock_irq(&conf->resync_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901
NeilBrown6cce3b232006-01-06 00:20:16 -0800902 /* Wait until no block IO is waiting (unless 'force') */
903 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
Lukas Czernereed8c022012-11-30 11:42:40 +0100904 conf->resync_lock);
NeilBrown0a27ec92006-01-06 00:20:13 -0800905
906 /* block any new IO from starting */
907 conf->barrier++;
908
NeilBrownc3b328a2011-04-18 18:25:43 +1000909 /* Now wait for all pending IO to complete */
NeilBrown0a27ec92006-01-06 00:20:13 -0800910 wait_event_lock_irq(conf->wait_barrier,
Tomasz Majchrzak0e5313e2016-06-24 14:20:16 +0200911 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
Lukas Czernereed8c022012-11-30 11:42:40 +0100912 conf->resync_lock);
NeilBrown0a27ec92006-01-06 00:20:13 -0800913
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 spin_unlock_irq(&conf->resync_lock);
915}
916
NeilBrowne879a872011-10-11 16:49:02 +1100917static void lower_barrier(struct r10conf *conf)
NeilBrown0a27ec92006-01-06 00:20:13 -0800918{
919 unsigned long flags;
920 spin_lock_irqsave(&conf->resync_lock, flags);
921 conf->barrier--;
922 spin_unlock_irqrestore(&conf->resync_lock, flags);
923 wake_up(&conf->wait_barrier);
924}
925
NeilBrowne879a872011-10-11 16:49:02 +1100926static void wait_barrier(struct r10conf *conf)
NeilBrown0a27ec92006-01-06 00:20:13 -0800927{
928 spin_lock_irq(&conf->resync_lock);
929 if (conf->barrier) {
930 conf->nr_waiting++;
NeilBrownd6b42dc2012-03-19 12:46:38 +1100931 /* Wait for the barrier to drop.
932 * However if there are already pending
933 * requests (preventing the barrier from
934 * rising completely), and the
935 * pre-process bio queue isn't empty,
936 * then don't wait, as we need to empty
937 * that queue to get the nr_pending
938 * count down.
939 */
940 wait_event_lock_irq(conf->wait_barrier,
941 !conf->barrier ||
Tomasz Majchrzak0e5313e2016-06-24 14:20:16 +0200942 (atomic_read(&conf->nr_pending) &&
NeilBrownd6b42dc2012-03-19 12:46:38 +1100943 current->bio_list &&
NeilBrown5959cde2017-03-10 17:00:47 +1100944 (!bio_list_empty(&current->bio_list[0]) ||
945 !bio_list_empty(&current->bio_list[1]))),
Lukas Czernereed8c022012-11-30 11:42:40 +0100946 conf->resync_lock);
NeilBrown0a27ec92006-01-06 00:20:13 -0800947 conf->nr_waiting--;
Tomasz Majchrzak0e5313e2016-06-24 14:20:16 +0200948 if (!conf->nr_waiting)
949 wake_up(&conf->wait_barrier);
NeilBrown0a27ec92006-01-06 00:20:13 -0800950 }
Tomasz Majchrzak0e5313e2016-06-24 14:20:16 +0200951 atomic_inc(&conf->nr_pending);
NeilBrown0a27ec92006-01-06 00:20:13 -0800952 spin_unlock_irq(&conf->resync_lock);
953}
954
NeilBrowne879a872011-10-11 16:49:02 +1100955static void allow_barrier(struct r10conf *conf)
NeilBrown0a27ec92006-01-06 00:20:13 -0800956{
Tomasz Majchrzak0e5313e2016-06-24 14:20:16 +0200957 if ((atomic_dec_and_test(&conf->nr_pending)) ||
958 (conf->array_freeze_pending))
959 wake_up(&conf->wait_barrier);
NeilBrown0a27ec92006-01-06 00:20:13 -0800960}
961
NeilBrowne2d59922013-06-12 11:01:22 +1000962static void freeze_array(struct r10conf *conf, int extra)
NeilBrown4443ae12006-01-06 00:20:28 -0800963{
964 /* stop syncio and normal IO and wait for everything to
NeilBrownf1885932006-01-06 00:20:42 -0800965 * go quiet.
NeilBrown4443ae12006-01-06 00:20:28 -0800966 * We increment barrier and nr_waiting, and then
NeilBrowne2d59922013-06-12 11:01:22 +1000967 * wait until nr_pending match nr_queued+extra
NeilBrown1c830532008-03-04 14:29:35 -0800968 * This is called in the context of one normal IO request
969 * that has failed. Thus any sync request that might be pending
970 * will be blocked by nr_pending, and we need to wait for
971 * pending IO requests to complete or be queued for re-try.
NeilBrowne2d59922013-06-12 11:01:22 +1000972 * Thus the number queued (nr_queued) plus this request (extra)
NeilBrown1c830532008-03-04 14:29:35 -0800973 * must match the number of pending IOs (nr_pending) before
974 * we continue.
NeilBrown4443ae12006-01-06 00:20:28 -0800975 */
976 spin_lock_irq(&conf->resync_lock);
Tomasz Majchrzak0e5313e2016-06-24 14:20:16 +0200977 conf->array_freeze_pending++;
NeilBrown4443ae12006-01-06 00:20:28 -0800978 conf->barrier++;
979 conf->nr_waiting++;
Lukas Czernereed8c022012-11-30 11:42:40 +0100980 wait_event_lock_irq_cmd(conf->wait_barrier,
Tomasz Majchrzak0e5313e2016-06-24 14:20:16 +0200981 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
Lukas Czernereed8c022012-11-30 11:42:40 +0100982 conf->resync_lock,
983 flush_pending_writes(conf));
NeilBrownc3b328a2011-04-18 18:25:43 +1000984
Tomasz Majchrzak0e5313e2016-06-24 14:20:16 +0200985 conf->array_freeze_pending--;
NeilBrown4443ae12006-01-06 00:20:28 -0800986 spin_unlock_irq(&conf->resync_lock);
987}
988
NeilBrowne879a872011-10-11 16:49:02 +1100989static void unfreeze_array(struct r10conf *conf)
NeilBrown4443ae12006-01-06 00:20:28 -0800990{
991 /* reverse the effect of the freeze */
992 spin_lock_irq(&conf->resync_lock);
993 conf->barrier--;
994 conf->nr_waiting--;
995 wake_up(&conf->wait_barrier);
996 spin_unlock_irq(&conf->resync_lock);
997}
998
NeilBrownf8c9e742012-05-21 09:28:33 +1000999static sector_t choose_data_offset(struct r10bio *r10_bio,
1000 struct md_rdev *rdev)
1001{
1002 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1003 test_bit(R10BIO_Previous, &r10_bio->state))
1004 return rdev->data_offset;
1005 else
1006 return rdev->new_data_offset;
1007}
1008
NeilBrown57c67df2012-10-11 13:32:13 +11001009struct raid10_plug_cb {
1010 struct blk_plug_cb cb;
1011 struct bio_list pending;
1012 int pending_cnt;
1013};
1014
1015static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1016{
1017 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1018 cb);
1019 struct mddev *mddev = plug->cb.data;
1020 struct r10conf *conf = mddev->private;
1021 struct bio *bio;
1022
NeilBrown874807a2012-11-27 12:14:40 +11001023 if (from_schedule || current->bio_list) {
NeilBrown57c67df2012-10-11 13:32:13 +11001024 spin_lock_irq(&conf->device_lock);
1025 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1026 conf->pending_count += plug->pending_cnt;
1027 spin_unlock_irq(&conf->device_lock);
NeilBrownee0b0242013-02-25 12:38:29 +11001028 wake_up(&conf->wait_barrier);
NeilBrown57c67df2012-10-11 13:32:13 +11001029 md_wakeup_thread(mddev->thread);
1030 kfree(plug);
1031 return;
1032 }
1033
1034 /* we aren't scheduling, so we can do the write-out directly. */
1035 bio = bio_list_get(&plug->pending);
1036 bitmap_unplug(mddev->bitmap);
1037 wake_up(&conf->wait_barrier);
1038
1039 while (bio) { /* submit pending writes */
1040 struct bio *next = bio->bi_next;
1041 bio->bi_next = NULL;
Mike Christie796a5cf2016-06-05 14:32:07 -05001042 if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
Shaohua Li32f9f572013-04-28 18:26:38 +08001043 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1044 /* Just ignore it */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001045 bio_endio(bio);
Shaohua Li32f9f572013-04-28 18:26:38 +08001046 else
1047 generic_make_request(bio);
NeilBrown57c67df2012-10-11 13:32:13 +11001048 bio = next;
1049 }
1050 kfree(plug);
1051}
1052
Kent Overstreet20d01892013-11-23 18:21:01 -08001053static void __make_request(struct mddev *mddev, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054{
NeilBrowne879a872011-10-11 16:49:02 +11001055 struct r10conf *conf = mddev->private;
NeilBrown9f2c9d12011-10-11 16:48:43 +11001056 struct r10bio *r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057 struct bio *read_bio;
1058 int i;
Mike Christie796a5cf2016-06-05 14:32:07 -05001059 const int op = bio_op(bio);
Jens Axboea3623572005-11-01 09:26:16 +01001060 const int rw = bio_data_dir(bio);
Jens Axboe1eff9d32016-08-05 15:35:16 -06001061 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1062 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
NeilBrown6cce3b232006-01-06 00:20:16 -08001063 unsigned long flags;
NeilBrown3cb03002011-10-11 16:45:26 +11001064 struct md_rdev *blocked_rdev;
NeilBrown57c67df2012-10-11 13:32:13 +11001065 struct blk_plug_cb *cb;
1066 struct raid10_plug_cb *plug = NULL;
NeilBrownd4432c22011-07-28 11:39:24 +10001067 int sectors_handled;
1068 int max_sectors;
NeilBrown3ea7daa2012-05-22 13:53:47 +10001069 int sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070
Tomasz Majchrzak9b622e22016-07-28 10:28:25 +02001071 md_write_start(mddev, bio);
1072
NeilBrowncc13b1d2014-05-05 13:34:37 +10001073 /*
1074 * Register the new request and wait if the reconstruction
1075 * thread has put up a bar for new requests.
1076 * Continue immediately if no resync is active currently.
1077 */
1078 wait_barrier(conf);
1079
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08001080 sectors = bio_sectors(bio);
NeilBrown3ea7daa2012-05-22 13:53:47 +10001081 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
Kent Overstreet4f024f32013-10-11 15:44:27 -07001082 bio->bi_iter.bi_sector < conf->reshape_progress &&
1083 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
NeilBrown3ea7daa2012-05-22 13:53:47 +10001084 /* IO spans the reshape position. Need to wait for
1085 * reshape to pass
1086 */
1087 allow_barrier(conf);
1088 wait_event(conf->wait_barrier,
Kent Overstreet4f024f32013-10-11 15:44:27 -07001089 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1090 conf->reshape_progress >= bio->bi_iter.bi_sector +
1091 sectors);
NeilBrown3ea7daa2012-05-22 13:53:47 +10001092 wait_barrier(conf);
1093 }
1094 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1095 bio_data_dir(bio) == WRITE &&
1096 (mddev->reshape_backwards
Kent Overstreet4f024f32013-10-11 15:44:27 -07001097 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1098 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1099 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1100 bio->bi_iter.bi_sector < conf->reshape_progress))) {
NeilBrown3ea7daa2012-05-22 13:53:47 +10001101 /* Need to update reshape_position in metadata */
1102 mddev->reshape_position = conf->reshape_progress;
Guoqing Jiang85ad1d12016-05-03 22:22:13 -04001103 set_mask_bits(&mddev->flags, 0,
1104 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
NeilBrown3ea7daa2012-05-22 13:53:47 +10001105 md_wakeup_thread(mddev->thread);
1106 wait_event(mddev->sb_wait,
1107 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1108
1109 conf->reshape_safe = mddev->reshape_position;
1110 }
1111
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1113
1114 r10_bio->master_bio = bio;
NeilBrown3ea7daa2012-05-22 13:53:47 +10001115 r10_bio->sectors = sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116
1117 r10_bio->mddev = mddev;
Kent Overstreet4f024f32013-10-11 15:44:27 -07001118 r10_bio->sector = bio->bi_iter.bi_sector;
NeilBrown6cce3b232006-01-06 00:20:16 -08001119 r10_bio->state = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120
NeilBrown856e08e2011-07-28 11:39:23 +10001121 /* We might need to issue multiple reads to different
1122 * devices if there are bad blocks around, so we keep
1123 * track of the number of reads in bio->bi_phys_segments.
1124 * If this is 0, there is only one r10_bio and no locking
1125 * will be needed when the request completes. If it is
1126 * non-zero, then it is the number of not-completed requests.
1127 */
1128 bio->bi_phys_segments = 0;
Jens Axboeb7c44ed2015-07-24 12:37:59 -06001129 bio_clear_flag(bio, BIO_SEG_VALID);
NeilBrown856e08e2011-07-28 11:39:23 +10001130
Jens Axboea3623572005-11-01 09:26:16 +01001131 if (rw == READ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 /*
1133 * read balancing logic:
1134 */
NeilBrown96c3fd12011-12-23 10:17:54 +11001135 struct md_rdev *rdev;
NeilBrown856e08e2011-07-28 11:39:23 +10001136 int slot;
1137
1138read_again:
NeilBrown96c3fd12011-12-23 10:17:54 +11001139 rdev = read_balance(conf, r10_bio, &max_sectors);
1140 if (!rdev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 raid_end_bio_io(r10_bio);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02001142 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 }
NeilBrown96c3fd12011-12-23 10:17:54 +11001144 slot = r10_bio->read_slot;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145
NeilBrowna167f662010-10-26 18:31:13 +11001146 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
Kent Overstreet4f024f32013-10-11 15:44:27 -07001147 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
Kent Overstreet6678d832013-08-07 11:14:32 -07001148 max_sectors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149
1150 r10_bio->devs[slot].bio = read_bio;
NeilBrownabbf0982011-12-23 10:17:54 +11001151 r10_bio->devs[slot].rdev = rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152
Kent Overstreet4f024f32013-10-11 15:44:27 -07001153 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
NeilBrownf8c9e742012-05-21 09:28:33 +10001154 choose_data_offset(r10_bio, rdev);
NeilBrown96c3fd12011-12-23 10:17:54 +11001155 read_bio->bi_bdev = rdev->bdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 read_bio->bi_end_io = raid10_end_read_request;
Mike Christie796a5cf2016-06-05 14:32:07 -05001157 bio_set_op_attrs(read_bio, op, do_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 read_bio->bi_private = r10_bio;
1159
NeilBrown856e08e2011-07-28 11:39:23 +10001160 if (max_sectors < r10_bio->sectors) {
1161 /* Could not read all from this device, so we will
1162 * need another r10_bio.
1163 */
NeilBrownb50c2592014-01-14 10:38:09 +11001164 sectors_handled = (r10_bio->sector + max_sectors
Kent Overstreet4f024f32013-10-11 15:44:27 -07001165 - bio->bi_iter.bi_sector);
NeilBrown856e08e2011-07-28 11:39:23 +10001166 r10_bio->sectors = max_sectors;
1167 spin_lock_irq(&conf->device_lock);
1168 if (bio->bi_phys_segments == 0)
1169 bio->bi_phys_segments = 2;
1170 else
1171 bio->bi_phys_segments++;
NeilBrownb50c2592014-01-14 10:38:09 +11001172 spin_unlock_irq(&conf->device_lock);
NeilBrown856e08e2011-07-28 11:39:23 +10001173 /* Cannot call generic_make_request directly
1174 * as that will be queued in __generic_make_request
1175 * and subsequent mempool_alloc might block
1176 * waiting for it. so hand bio over to raid10d.
1177 */
1178 reschedule_retry(r10_bio);
1179
1180 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1181
1182 r10_bio->master_bio = bio;
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08001183 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
NeilBrown856e08e2011-07-28 11:39:23 +10001184 r10_bio->state = 0;
1185 r10_bio->mddev = mddev;
Kent Overstreet4f024f32013-10-11 15:44:27 -07001186 r10_bio->sector = bio->bi_iter.bi_sector +
1187 sectors_handled;
NeilBrown856e08e2011-07-28 11:39:23 +10001188 goto read_again;
1189 } else
1190 generic_make_request(read_bio);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02001191 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 }
1193
1194 /*
1195 * WRITE:
1196 */
NeilBrown34db0cd2011-10-11 16:50:01 +11001197 if (conf->pending_count >= max_queued_requests) {
1198 md_wakeup_thread(mddev->thread);
1199 wait_event(conf->wait_barrier,
1200 conf->pending_count < max_queued_requests);
1201 }
Dan Williams6bfe0b42008-04-30 00:52:32 -07001202 /* first select target devices under rcu_lock and
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203 * inc refcount on their rdev. Record them by setting
1204 * bios[x] to bio
NeilBrownd4432c22011-07-28 11:39:24 +10001205 * If there are known/acknowledged bad blocks on any device
1206 * on which we have seen a write error, we want to avoid
1207 * writing to those blocks. This potentially requires several
1208 * writes to write around the bad blocks. Each set of writes
1209 * gets its own r10_bio with a set of bios attached. The number
1210 * of r10_bios is recored in bio->bi_phys_segments just as with
1211 * the read case.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212 */
NeilBrownc3b328a2011-04-18 18:25:43 +10001213
NeilBrown69335ef2011-12-23 10:17:54 +11001214 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 raid10_find_phys(conf, r10_bio);
NeilBrownd4432c22011-07-28 11:39:24 +10001216retry_write:
Harvey Harrisoncb6969e2008-05-06 20:42:32 -07001217 blocked_rdev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 rcu_read_lock();
NeilBrownd4432c22011-07-28 11:39:24 +10001219 max_sectors = r10_bio->sectors;
1220
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 for (i = 0; i < conf->copies; i++) {
1222 int d = r10_bio->devs[i].devnum;
NeilBrown3cb03002011-10-11 16:45:26 +11001223 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
NeilBrown475b0322011-12-23 10:17:55 +11001224 struct md_rdev *rrdev = rcu_dereference(
1225 conf->mirrors[d].replacement);
NeilBrown4ca40c22011-12-23 10:17:55 +11001226 if (rdev == rrdev)
1227 rrdev = NULL;
Dan Williams6bfe0b42008-04-30 00:52:32 -07001228 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1229 atomic_inc(&rdev->nr_pending);
1230 blocked_rdev = rdev;
1231 break;
1232 }
NeilBrown475b0322011-12-23 10:17:55 +11001233 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1234 atomic_inc(&rrdev->nr_pending);
1235 blocked_rdev = rrdev;
1236 break;
1237 }
Kent Overstreet8ae12662015-04-27 23:48:34 -07001238 if (rdev && (test_bit(Faulty, &rdev->flags)))
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001239 rdev = NULL;
Kent Overstreet8ae12662015-04-27 23:48:34 -07001240 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
NeilBrown475b0322011-12-23 10:17:55 +11001241 rrdev = NULL;
1242
NeilBrownd4432c22011-07-28 11:39:24 +10001243 r10_bio->devs[i].bio = NULL;
NeilBrown475b0322011-12-23 10:17:55 +11001244 r10_bio->devs[i].repl_bio = NULL;
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001245
1246 if (!rdev && !rrdev) {
NeilBrown6cce3b232006-01-06 00:20:16 -08001247 set_bit(R10BIO_Degraded, &r10_bio->state);
NeilBrownd4432c22011-07-28 11:39:24 +10001248 continue;
NeilBrown6cce3b232006-01-06 00:20:16 -08001249 }
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001250 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
NeilBrownd4432c22011-07-28 11:39:24 +10001251 sector_t first_bad;
1252 sector_t dev_sector = r10_bio->devs[i].addr;
1253 int bad_sectors;
1254 int is_bad;
1255
1256 is_bad = is_badblock(rdev, dev_sector,
1257 max_sectors,
1258 &first_bad, &bad_sectors);
1259 if (is_bad < 0) {
1260 /* Mustn't write here until the bad block
1261 * is acknowledged
1262 */
1263 atomic_inc(&rdev->nr_pending);
1264 set_bit(BlockedBadBlocks, &rdev->flags);
1265 blocked_rdev = rdev;
1266 break;
1267 }
1268 if (is_bad && first_bad <= dev_sector) {
1269 /* Cannot write here at all */
1270 bad_sectors -= (dev_sector - first_bad);
1271 if (bad_sectors < max_sectors)
1272 /* Mustn't write more than bad_sectors
1273 * to other devices yet
1274 */
1275 max_sectors = bad_sectors;
1276 /* We don't set R10BIO_Degraded as that
1277 * only applies if the disk is missing,
1278 * so it might be re-added, and we want to
1279 * know to recover this chunk.
1280 * In this case the device is here, and the
1281 * fact that this chunk is not in-sync is
1282 * recorded in the bad block log.
1283 */
1284 continue;
1285 }
1286 if (is_bad) {
1287 int good_sectors = first_bad - dev_sector;
1288 if (good_sectors < max_sectors)
1289 max_sectors = good_sectors;
1290 }
1291 }
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001292 if (rdev) {
1293 r10_bio->devs[i].bio = bio;
1294 atomic_inc(&rdev->nr_pending);
1295 }
NeilBrown475b0322011-12-23 10:17:55 +11001296 if (rrdev) {
1297 r10_bio->devs[i].repl_bio = bio;
1298 atomic_inc(&rrdev->nr_pending);
1299 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 }
1301 rcu_read_unlock();
1302
Dan Williams6bfe0b42008-04-30 00:52:32 -07001303 if (unlikely(blocked_rdev)) {
1304 /* Have to wait for this device to get unblocked, then retry */
1305 int j;
1306 int d;
1307
NeilBrown475b0322011-12-23 10:17:55 +11001308 for (j = 0; j < i; j++) {
Dan Williams6bfe0b42008-04-30 00:52:32 -07001309 if (r10_bio->devs[j].bio) {
1310 d = r10_bio->devs[j].devnum;
1311 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1312 }
NeilBrown475b0322011-12-23 10:17:55 +11001313 if (r10_bio->devs[j].repl_bio) {
NeilBrown4ca40c22011-12-23 10:17:55 +11001314 struct md_rdev *rdev;
NeilBrown475b0322011-12-23 10:17:55 +11001315 d = r10_bio->devs[j].devnum;
NeilBrown4ca40c22011-12-23 10:17:55 +11001316 rdev = conf->mirrors[d].replacement;
1317 if (!rdev) {
1318 /* Race with remove_disk */
1319 smp_mb();
1320 rdev = conf->mirrors[d].rdev;
1321 }
1322 rdev_dec_pending(rdev, mddev);
NeilBrown475b0322011-12-23 10:17:55 +11001323 }
1324 }
Dan Williams6bfe0b42008-04-30 00:52:32 -07001325 allow_barrier(conf);
1326 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1327 wait_barrier(conf);
1328 goto retry_write;
1329 }
1330
NeilBrownd4432c22011-07-28 11:39:24 +10001331 if (max_sectors < r10_bio->sectors) {
1332 /* We are splitting this into multiple parts, so
1333 * we need to prepare for allocating another r10_bio.
1334 */
1335 r10_bio->sectors = max_sectors;
1336 spin_lock_irq(&conf->device_lock);
1337 if (bio->bi_phys_segments == 0)
1338 bio->bi_phys_segments = 2;
1339 else
1340 bio->bi_phys_segments++;
1341 spin_unlock_irq(&conf->device_lock);
1342 }
Kent Overstreet4f024f32013-10-11 15:44:27 -07001343 sectors_handled = r10_bio->sector + max_sectors -
1344 bio->bi_iter.bi_sector;
NeilBrownd4432c22011-07-28 11:39:24 +10001345
NeilBrown4e780642010-10-19 12:54:01 +11001346 atomic_set(&r10_bio->remaining, 1);
NeilBrownd4432c22011-07-28 11:39:24 +10001347 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
NeilBrown06d91a52005-06-21 17:17:12 -07001348
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 for (i = 0; i < conf->copies; i++) {
1350 struct bio *mbio;
1351 int d = r10_bio->devs[i].devnum;
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001352 if (r10_bio->devs[i].bio) {
1353 struct md_rdev *rdev = conf->mirrors[d].rdev;
1354 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
Kent Overstreet4f024f32013-10-11 15:44:27 -07001355 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
Kent Overstreet6678d832013-08-07 11:14:32 -07001356 max_sectors);
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001357 r10_bio->devs[i].bio = mbio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358
Kent Overstreet4f024f32013-10-11 15:44:27 -07001359 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001360 choose_data_offset(r10_bio,
1361 rdev));
1362 mbio->bi_bdev = rdev->bdev;
1363 mbio->bi_end_io = raid10_end_write_request;
Christoph Hellwig288dab82016-06-09 16:00:36 +02001364 bio_set_op_attrs(mbio, op, do_sync | do_fua);
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001365 mbio->bi_private = r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001367 atomic_inc(&r10_bio->remaining);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001369 cb = blk_check_plugged(raid10_unplug, mddev,
1370 sizeof(*plug));
1371 if (cb)
1372 plug = container_of(cb, struct raid10_plug_cb,
1373 cb);
1374 else
1375 plug = NULL;
1376 spin_lock_irqsave(&conf->device_lock, flags);
1377 if (plug) {
1378 bio_list_add(&plug->pending, mbio);
1379 plug->pending_cnt++;
1380 } else {
1381 bio_list_add(&conf->pending_bio_list, mbio);
1382 conf->pending_count++;
1383 }
1384 spin_unlock_irqrestore(&conf->device_lock, flags);
1385 if (!plug)
1386 md_wakeup_thread(mddev->thread);
1387 }
NeilBrown57c67df2012-10-11 13:32:13 +11001388
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001389 if (r10_bio->devs[i].repl_bio) {
1390 struct md_rdev *rdev = conf->mirrors[d].replacement;
1391 if (rdev == NULL) {
1392 /* Replacement just got moved to main 'rdev' */
1393 smp_mb();
1394 rdev = conf->mirrors[d].rdev;
1395 }
1396 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
Kent Overstreet4f024f32013-10-11 15:44:27 -07001397 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
Kent Overstreet6678d832013-08-07 11:14:32 -07001398 max_sectors);
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001399 r10_bio->devs[i].repl_bio = mbio;
1400
Kent Overstreet4f024f32013-10-11 15:44:27 -07001401 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001402 choose_data_offset(
1403 r10_bio, rdev));
1404 mbio->bi_bdev = rdev->bdev;
1405 mbio->bi_end_io = raid10_end_write_request;
Christoph Hellwig288dab82016-06-09 16:00:36 +02001406 bio_set_op_attrs(mbio, op, do_sync | do_fua);
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001407 mbio->bi_private = r10_bio;
1408
1409 atomic_inc(&r10_bio->remaining);
Shaohua Li4131c882017-02-23 12:26:41 -08001410
1411 cb = blk_check_plugged(raid10_unplug, mddev,
1412 sizeof(*plug));
1413 if (cb)
1414 plug = container_of(cb, struct raid10_plug_cb,
1415 cb);
1416 else
1417 plug = NULL;
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001418 spin_lock_irqsave(&conf->device_lock, flags);
Shaohua Li4131c882017-02-23 12:26:41 -08001419 if (plug) {
1420 bio_list_add(&plug->pending, mbio);
1421 plug->pending_cnt++;
1422 } else {
1423 bio_list_add(&conf->pending_bio_list, mbio);
1424 conf->pending_count++;
1425 }
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001426 spin_unlock_irqrestore(&conf->device_lock, flags);
Shaohua Li4131c882017-02-23 12:26:41 -08001427 if (!plug)
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001428 md_wakeup_thread(mddev->thread);
NeilBrown57c67df2012-10-11 13:32:13 +11001429 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 }
1431
NeilBrown079fa162011-09-10 17:21:23 +10001432 /* Don't remove the bias on 'remaining' (one_write_done) until
1433 * after checking if we need to go around again.
1434 */
NeilBrowna35e63e2008-03-04 14:29:29 -08001435
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08001436 if (sectors_handled < bio_sectors(bio)) {
NeilBrown079fa162011-09-10 17:21:23 +10001437 one_write_done(r10_bio);
NeilBrown5e570282011-07-28 11:39:25 +10001438 /* We need another r10_bio. It has already been counted
NeilBrownd4432c22011-07-28 11:39:24 +10001439 * in bio->bi_phys_segments.
1440 */
1441 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1442
1443 r10_bio->master_bio = bio;
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08001444 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
NeilBrownd4432c22011-07-28 11:39:24 +10001445
1446 r10_bio->mddev = mddev;
Kent Overstreet4f024f32013-10-11 15:44:27 -07001447 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
NeilBrownd4432c22011-07-28 11:39:24 +10001448 r10_bio->state = 0;
1449 goto retry_write;
1450 }
NeilBrown079fa162011-09-10 17:21:23 +10001451 one_write_done(r10_bio);
Kent Overstreet20d01892013-11-23 18:21:01 -08001452}
1453
Shaohua Li849674e2016-01-20 13:52:20 -08001454static void raid10_make_request(struct mddev *mddev, struct bio *bio)
Kent Overstreet20d01892013-11-23 18:21:01 -08001455{
1456 struct r10conf *conf = mddev->private;
1457 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1458 int chunk_sects = chunk_mask + 1;
1459
1460 struct bio *split;
1461
Jens Axboe1eff9d32016-08-05 15:35:16 -06001462 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
Kent Overstreet20d01892013-11-23 18:21:01 -08001463 md_flush_request(mddev, bio);
1464 return;
1465 }
1466
Kent Overstreet20d01892013-11-23 18:21:01 -08001467 do {
1468
1469 /*
1470 * If this request crosses a chunk boundary, we need to split
1471 * it.
1472 */
1473 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1474 bio_sectors(bio) > chunk_sects
1475 && (conf->geo.near_copies < conf->geo.raid_disks
1476 || conf->prev.near_copies <
1477 conf->prev.raid_disks))) {
1478 split = bio_split(bio, chunk_sects -
1479 (bio->bi_iter.bi_sector &
1480 (chunk_sects - 1)),
1481 GFP_NOIO, fs_bio_set);
1482 bio_chain(split, bio);
1483 } else {
1484 split = bio;
1485 }
1486
Shaohua Liad516642017-02-28 13:00:20 -08001487 /*
1488 * If a bio is splitted, the first part of bio will pass
1489 * barrier but the bio is queued in current->bio_list (see
1490 * generic_make_request). If there is a raise_barrier() called
1491 * here, the second part of bio can't pass barrier. But since
1492 * the first part bio isn't dispatched to underlaying disks
1493 * yet, the barrier is never released, hence raise_barrier will
1494 * alays wait. We have a deadlock.
1495 * Note, this only happens in read path. For write path, the
1496 * first part of bio is dispatched in a schedule() call
1497 * (because of blk plug) or offloaded to raid10d.
1498 * Quitting from the function immediately can change the bio
1499 * order queued in bio_list and avoid the deadlock.
1500 */
Kent Overstreet20d01892013-11-23 18:21:01 -08001501 __make_request(mddev, split);
Shaohua Liad516642017-02-28 13:00:20 -08001502 if (split != bio && bio_data_dir(bio) == READ) {
1503 generic_make_request(bio);
1504 break;
1505 }
Kent Overstreet20d01892013-11-23 18:21:01 -08001506 } while (split != bio);
NeilBrown079fa162011-09-10 17:21:23 +10001507
1508 /* In case raid10d snuck in to freeze_array */
1509 wake_up(&conf->wait_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510}
1511
Shaohua Li849674e2016-01-20 13:52:20 -08001512static void raid10_status(struct seq_file *seq, struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513{
NeilBrowne879a872011-10-11 16:49:02 +11001514 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515 int i;
1516
NeilBrown5cf00fc2012-05-21 09:28:20 +10001517 if (conf->geo.near_copies < conf->geo.raid_disks)
Andre Noll9d8f0362009-06-18 08:45:01 +10001518 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
NeilBrown5cf00fc2012-05-21 09:28:20 +10001519 if (conf->geo.near_copies > 1)
1520 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1521 if (conf->geo.far_copies > 1) {
1522 if (conf->geo.far_offset)
1523 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
NeilBrownc93983b2006-06-26 00:27:41 -07001524 else
NeilBrown5cf00fc2012-05-21 09:28:20 +10001525 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
NeilBrown8bce6d32015-10-22 13:20:15 +11001526 if (conf->geo.far_set_size != conf->geo.raid_disks)
1527 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
NeilBrownc93983b2006-06-26 00:27:41 -07001528 }
NeilBrown5cf00fc2012-05-21 09:28:20 +10001529 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1530 conf->geo.raid_disks - mddev->degraded);
NeilBrownd44b0a92016-06-02 16:19:52 +10001531 rcu_read_lock();
1532 for (i = 0; i < conf->geo.raid_disks; i++) {
1533 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1534 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1535 }
1536 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 seq_printf(seq, "]");
1538}
1539
NeilBrown700c7212011-07-27 11:00:36 +10001540/* check if there are enough drives for
1541 * every block to appear on atleast one.
1542 * Don't consider the device numbered 'ignore'
1543 * as we might be about to remove it.
1544 */
NeilBrown635f6412013-06-11 14:57:09 +10001545static int _enough(struct r10conf *conf, int previous, int ignore)
NeilBrown700c7212011-07-27 11:00:36 +10001546{
1547 int first = 0;
NeilBrown725d6e52013-06-11 15:08:03 +10001548 int has_enough = 0;
NeilBrown635f6412013-06-11 14:57:09 +10001549 int disks, ncopies;
1550 if (previous) {
1551 disks = conf->prev.raid_disks;
1552 ncopies = conf->prev.near_copies;
1553 } else {
1554 disks = conf->geo.raid_disks;
1555 ncopies = conf->geo.near_copies;
1556 }
NeilBrown700c7212011-07-27 11:00:36 +10001557
NeilBrown725d6e52013-06-11 15:08:03 +10001558 rcu_read_lock();
NeilBrown700c7212011-07-27 11:00:36 +10001559 do {
1560 int n = conf->copies;
1561 int cnt = 0;
NeilBrown80b48122012-09-27 12:35:21 +10001562 int this = first;
NeilBrown700c7212011-07-27 11:00:36 +10001563 while (n--) {
NeilBrown725d6e52013-06-11 15:08:03 +10001564 struct md_rdev *rdev;
1565 if (this != ignore &&
1566 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1567 test_bit(In_sync, &rdev->flags))
NeilBrown700c7212011-07-27 11:00:36 +10001568 cnt++;
NeilBrown635f6412013-06-11 14:57:09 +10001569 this = (this+1) % disks;
NeilBrown700c7212011-07-27 11:00:36 +10001570 }
1571 if (cnt == 0)
NeilBrown725d6e52013-06-11 15:08:03 +10001572 goto out;
NeilBrown635f6412013-06-11 14:57:09 +10001573 first = (first + ncopies) % disks;
NeilBrown700c7212011-07-27 11:00:36 +10001574 } while (first != 0);
NeilBrown725d6e52013-06-11 15:08:03 +10001575 has_enough = 1;
1576out:
1577 rcu_read_unlock();
1578 return has_enough;
NeilBrown700c7212011-07-27 11:00:36 +10001579}
1580
NeilBrownf8c9e742012-05-21 09:28:33 +10001581static int enough(struct r10conf *conf, int ignore)
1582{
NeilBrown635f6412013-06-11 14:57:09 +10001583 /* when calling 'enough', both 'prev' and 'geo' must
1584 * be stable.
1585 * This is ensured if ->reconfig_mutex or ->device_lock
1586 * is held.
1587 */
1588 return _enough(conf, 0, ignore) &&
1589 _enough(conf, 1, ignore);
NeilBrownf8c9e742012-05-21 09:28:33 +10001590}
1591
Shaohua Li849674e2016-01-20 13:52:20 -08001592static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593{
1594 char b[BDEVNAME_SIZE];
NeilBrowne879a872011-10-11 16:49:02 +11001595 struct r10conf *conf = mddev->private;
NeilBrown635f6412013-06-11 14:57:09 +10001596 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597
1598 /*
1599 * If it is not operational, then we have already marked it as dead
1600 * else if it is the last working disks, ignore the error, let the
1601 * next level up know.
1602 * else mark the drive as failed
1603 */
NeilBrown635f6412013-06-11 14:57:09 +10001604 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrownb2d444d2005-11-08 21:39:31 -08001605 if (test_bit(In_sync, &rdev->flags)
NeilBrown635f6412013-06-11 14:57:09 +10001606 && !enough(conf, rdev->raid_disk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607 /*
1608 * Don't fail the drive, just return an IO error.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609 */
NeilBrownc04be0a2006-10-03 01:15:53 -07001610 spin_unlock_irqrestore(&conf->device_lock, flags);
NeilBrown635f6412013-06-11 14:57:09 +10001611 return;
1612 }
NeilBrown2446dba2014-07-31 10:16:29 +10001613 if (test_and_clear_bit(In_sync, &rdev->flags))
NeilBrown635f6412013-06-11 14:57:09 +10001614 mddev->degraded++;
NeilBrown2446dba2014-07-31 10:16:29 +10001615 /*
1616 * If recovery is running, make sure it aborts.
1617 */
1618 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
NeilBrownde393cd2011-07-28 11:31:48 +10001619 set_bit(Blocked, &rdev->flags);
NeilBrownb2d444d2005-11-08 21:39:31 -08001620 set_bit(Faulty, &rdev->flags);
Guoqing Jiang85ad1d12016-05-03 22:22:13 -04001621 set_mask_bits(&mddev->flags, 0,
1622 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
NeilBrown635f6412013-06-11 14:57:09 +10001623 spin_unlock_irqrestore(&conf->device_lock, flags);
Joe Perches067032b2011-01-14 09:14:33 +11001624 printk(KERN_ALERT
1625 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1626 "md/raid10:%s: Operation continuing on %d devices.\n",
NeilBrown128595e2010-05-03 14:47:14 +10001627 mdname(mddev), bdevname(rdev->bdev, b),
NeilBrown5cf00fc2012-05-21 09:28:20 +10001628 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629}
1630
NeilBrowne879a872011-10-11 16:49:02 +11001631static void print_conf(struct r10conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632{
1633 int i;
NeilBrown4056ca52016-06-02 16:19:52 +10001634 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635
NeilBrown128595e2010-05-03 14:47:14 +10001636 printk(KERN_DEBUG "RAID10 conf printout:\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637 if (!conf) {
NeilBrown128595e2010-05-03 14:47:14 +10001638 printk(KERN_DEBUG "(!conf)\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 return;
1640 }
NeilBrown5cf00fc2012-05-21 09:28:20 +10001641 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1642 conf->geo.raid_disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643
NeilBrown4056ca52016-06-02 16:19:52 +10001644 /* This is only called with ->reconfix_mutex held, so
1645 * rcu protection of rdev is not needed */
NeilBrown5cf00fc2012-05-21 09:28:20 +10001646 for (i = 0; i < conf->geo.raid_disks; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647 char b[BDEVNAME_SIZE];
NeilBrown4056ca52016-06-02 16:19:52 +10001648 rdev = conf->mirrors[i].rdev;
1649 if (rdev)
NeilBrown128595e2010-05-03 14:47:14 +10001650 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
NeilBrown4056ca52016-06-02 16:19:52 +10001651 i, !test_bit(In_sync, &rdev->flags),
1652 !test_bit(Faulty, &rdev->flags),
1653 bdevname(rdev->bdev,b));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 }
1655}
1656
NeilBrowne879a872011-10-11 16:49:02 +11001657static void close_sync(struct r10conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658{
NeilBrown0a27ec92006-01-06 00:20:13 -08001659 wait_barrier(conf);
1660 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661
1662 mempool_destroy(conf->r10buf_pool);
1663 conf->r10buf_pool = NULL;
1664}
1665
NeilBrownfd01b882011-10-11 16:47:53 +11001666static int raid10_spare_active(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667{
1668 int i;
NeilBrowne879a872011-10-11 16:49:02 +11001669 struct r10conf *conf = mddev->private;
Jonathan Brassowdc280d982012-07-31 10:03:52 +10001670 struct raid10_info *tmp;
NeilBrown6b965622010-08-18 11:56:59 +10001671 int count = 0;
1672 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673
1674 /*
1675 * Find all non-in_sync disks within the RAID10 configuration
1676 * and mark them in_sync
1677 */
NeilBrown5cf00fc2012-05-21 09:28:20 +10001678 for (i = 0; i < conf->geo.raid_disks; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679 tmp = conf->mirrors + i;
NeilBrown4ca40c22011-12-23 10:17:55 +11001680 if (tmp->replacement
1681 && tmp->replacement->recovery_offset == MaxSector
1682 && !test_bit(Faulty, &tmp->replacement->flags)
1683 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1684 /* Replacement has just become active */
1685 if (!tmp->rdev
1686 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1687 count++;
1688 if (tmp->rdev) {
1689 /* Replaced device not technically faulty,
1690 * but we need to be sure it gets removed
1691 * and never re-added.
1692 */
1693 set_bit(Faulty, &tmp->rdev->flags);
1694 sysfs_notify_dirent_safe(
1695 tmp->rdev->sysfs_state);
1696 }
1697 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1698 } else if (tmp->rdev
Lukasz Dorau61e49472013-10-24 12:55:17 +11001699 && tmp->rdev->recovery_offset == MaxSector
NeilBrown4ca40c22011-12-23 10:17:55 +11001700 && !test_bit(Faulty, &tmp->rdev->flags)
1701 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
NeilBrown6b965622010-08-18 11:56:59 +10001702 count++;
Jonathan Brassow2863b9e2012-10-11 13:38:58 +11001703 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 }
1705 }
NeilBrown6b965622010-08-18 11:56:59 +10001706 spin_lock_irqsave(&conf->device_lock, flags);
1707 mddev->degraded -= count;
1708 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709
1710 print_conf(conf);
NeilBrown6b965622010-08-18 11:56:59 +10001711 return count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712}
1713
NeilBrownfd01b882011-10-11 16:47:53 +11001714static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715{
NeilBrowne879a872011-10-11 16:49:02 +11001716 struct r10conf *conf = mddev->private;
Neil Brown199050e2008-06-28 08:31:33 +10001717 int err = -EEXIST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 int mirror;
Neil Brown6c2fce22008-06-28 08:31:31 +10001719 int first = 0;
NeilBrown5cf00fc2012-05-21 09:28:20 +10001720 int last = conf->geo.raid_disks - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721
1722 if (mddev->recovery_cp < MaxSector)
1723 /* only hot-add to in-sync arrays, as recovery is
1724 * very different from resync
1725 */
Neil Brown199050e2008-06-28 08:31:33 +10001726 return -EBUSY;
NeilBrown635f6412013-06-11 14:57:09 +10001727 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
Neil Brown199050e2008-06-28 08:31:33 +10001728 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729
Dan Williams1501efa2016-01-13 16:00:07 -08001730 if (md_integrity_add_rdev(rdev, mddev))
1731 return -ENXIO;
1732
NeilBrowna53a6c82008-11-06 17:28:20 +11001733 if (rdev->raid_disk >= 0)
Neil Brown6c2fce22008-06-28 08:31:31 +10001734 first = last = rdev->raid_disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735
Namhyung Kim2c4193d2011-07-18 17:38:43 +10001736 if (rdev->saved_raid_disk >= first &&
NeilBrown6cce3b232006-01-06 00:20:16 -08001737 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1738 mirror = rdev->saved_raid_disk;
1739 else
Neil Brown6c2fce22008-06-28 08:31:31 +10001740 mirror = first;
NeilBrown2bb77732011-07-27 11:00:36 +10001741 for ( ; mirror <= last ; mirror++) {
Jonathan Brassowdc280d982012-07-31 10:03:52 +10001742 struct raid10_info *p = &conf->mirrors[mirror];
NeilBrown2bb77732011-07-27 11:00:36 +10001743 if (p->recovery_disabled == mddev->recovery_disabled)
1744 continue;
NeilBrownb7044d42011-12-23 10:17:56 +11001745 if (p->rdev) {
1746 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1747 p->replacement != NULL)
1748 continue;
1749 clear_bit(In_sync, &rdev->flags);
1750 set_bit(Replacement, &rdev->flags);
1751 rdev->raid_disk = mirror;
1752 err = 0;
Jonathan Brassow9092c022013-05-02 14:19:24 -05001753 if (mddev->gendisk)
1754 disk_stack_limits(mddev->gendisk, rdev->bdev,
1755 rdev->data_offset << 9);
NeilBrownb7044d42011-12-23 10:17:56 +11001756 conf->fullsync = 1;
1757 rcu_assign_pointer(p->replacement, rdev);
1758 break;
1759 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760
Jonathan Brassow9092c022013-05-02 14:19:24 -05001761 if (mddev->gendisk)
1762 disk_stack_limits(mddev->gendisk, rdev->bdev,
1763 rdev->data_offset << 9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764
NeilBrown2bb77732011-07-27 11:00:36 +10001765 p->head_position = 0;
NeilBrownd890fa22011-10-26 11:54:39 +11001766 p->recovery_disabled = mddev->recovery_disabled - 1;
NeilBrown2bb77732011-07-27 11:00:36 +10001767 rdev->raid_disk = mirror;
1768 err = 0;
1769 if (rdev->saved_raid_disk != mirror)
1770 conf->fullsync = 1;
1771 rcu_assign_pointer(p->rdev, rdev);
1772 break;
1773 }
Jonathan Brassowed30be02012-10-31 11:42:30 +11001774 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
Shaohua Li532a2a32012-10-11 13:30:52 +11001775 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1776
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 print_conf(conf);
Neil Brown199050e2008-06-28 08:31:33 +10001778 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779}
1780
NeilBrownb8321b62011-12-23 10:17:51 +11001781static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782{
NeilBrowne879a872011-10-11 16:49:02 +11001783 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 int err = 0;
NeilBrownb8321b62011-12-23 10:17:51 +11001785 int number = rdev->raid_disk;
NeilBrownc8ab9032011-12-23 10:17:54 +11001786 struct md_rdev **rdevp;
Jonathan Brassowdc280d982012-07-31 10:03:52 +10001787 struct raid10_info *p = conf->mirrors + number;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788
1789 print_conf(conf);
NeilBrownc8ab9032011-12-23 10:17:54 +11001790 if (rdev == p->rdev)
1791 rdevp = &p->rdev;
1792 else if (rdev == p->replacement)
1793 rdevp = &p->replacement;
1794 else
1795 return 0;
1796
1797 if (test_bit(In_sync, &rdev->flags) ||
1798 atomic_read(&rdev->nr_pending)) {
1799 err = -EBUSY;
1800 goto abort;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801 }
NeilBrownd787be42016-06-02 16:19:53 +10001802 /* Only remove non-faulty devices if recovery
NeilBrownc8ab9032011-12-23 10:17:54 +11001803 * is not possible.
1804 */
1805 if (!test_bit(Faulty, &rdev->flags) &&
1806 mddev->recovery_disabled != p->recovery_disabled &&
NeilBrown4ca40c22011-12-23 10:17:55 +11001807 (!p->replacement || p->replacement == rdev) &&
NeilBrown63aced62012-05-22 13:55:33 +10001808 number < conf->geo.raid_disks &&
NeilBrownc8ab9032011-12-23 10:17:54 +11001809 enough(conf, -1)) {
1810 err = -EBUSY;
1811 goto abort;
1812 }
1813 *rdevp = NULL;
NeilBrownd787be42016-06-02 16:19:53 +10001814 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1815 synchronize_rcu();
1816 if (atomic_read(&rdev->nr_pending)) {
1817 /* lost the race, try later */
1818 err = -EBUSY;
1819 *rdevp = rdev;
1820 goto abort;
1821 }
1822 }
1823 if (p->replacement) {
NeilBrown4ca40c22011-12-23 10:17:55 +11001824 /* We must have just cleared 'rdev' */
1825 p->rdev = p->replacement;
1826 clear_bit(Replacement, &p->replacement->flags);
1827 smp_mb(); /* Make sure other CPUs may see both as identical
1828 * but will never see neither -- if they are careful.
1829 */
1830 p->replacement = NULL;
1831 clear_bit(WantReplacement, &rdev->flags);
1832 } else
1833 /* We might have just remove the Replacement as faulty
1834 * Clear the flag just in case
1835 */
1836 clear_bit(WantReplacement, &rdev->flags);
1837
NeilBrownc8ab9032011-12-23 10:17:54 +11001838 err = md_integrity_register(mddev);
1839
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840abort:
1841
1842 print_conf(conf);
1843 return err;
1844}
1845
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001846static void end_sync_read(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847{
NeilBrown9f2c9d12011-10-11 16:48:43 +11001848 struct r10bio *r10_bio = bio->bi_private;
NeilBrowne879a872011-10-11 16:49:02 +11001849 struct r10conf *conf = r10_bio->mddev->private;
Namhyung Kim778ca012011-07-18 17:38:47 +10001850 int d;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851
NeilBrown3ea7daa2012-05-22 13:53:47 +10001852 if (bio == r10_bio->master_bio) {
1853 /* this is a reshape read */
1854 d = r10_bio->read_slot; /* really the read dev */
1855 } else
1856 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
NeilBrown0eb3ff12006-01-06 00:20:29 -08001857
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001858 if (!bio->bi_error)
NeilBrown0eb3ff12006-01-06 00:20:29 -08001859 set_bit(R10BIO_Uptodate, &r10_bio->state);
NeilBrowne684e412011-07-28 11:39:25 +10001860 else
1861 /* The write handler will notice the lack of
1862 * R10BIO_Uptodate and record any errors etc
1863 */
NeilBrown4dbcdc72006-01-06 00:20:52 -08001864 atomic_add(r10_bio->sectors,
1865 &conf->mirrors[d].rdev->corrected_errors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866
1867 /* for reconstruct, we always reschedule after a read.
1868 * for resync, only after all reads
1869 */
NeilBrown73d5c382009-02-25 13:18:47 +11001870 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1872 atomic_dec_and_test(&r10_bio->remaining)) {
1873 /* we have read all the blocks,
1874 * do the comparison in process context in raid10d
1875 */
1876 reschedule_retry(r10_bio);
1877 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878}
1879
NeilBrown9f2c9d12011-10-11 16:48:43 +11001880static void end_sync_request(struct r10bio *r10_bio)
NeilBrown5e570282011-07-28 11:39:25 +10001881{
NeilBrownfd01b882011-10-11 16:47:53 +11001882 struct mddev *mddev = r10_bio->mddev;
NeilBrown5e570282011-07-28 11:39:25 +10001883
1884 while (atomic_dec_and_test(&r10_bio->remaining)) {
1885 if (r10_bio->master_bio == NULL) {
1886 /* the primary of several recovery bios */
1887 sector_t s = r10_bio->sectors;
1888 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1889 test_bit(R10BIO_WriteError, &r10_bio->state))
1890 reschedule_retry(r10_bio);
1891 else
1892 put_buf(r10_bio);
1893 md_done_sync(mddev, s, 1);
1894 break;
1895 } else {
NeilBrown9f2c9d12011-10-11 16:48:43 +11001896 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
NeilBrown5e570282011-07-28 11:39:25 +10001897 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1898 test_bit(R10BIO_WriteError, &r10_bio->state))
1899 reschedule_retry(r10_bio);
1900 else
1901 put_buf(r10_bio);
1902 r10_bio = r10_bio2;
1903 }
1904 }
1905}
1906
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001907static void end_sync_write(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908{
NeilBrown9f2c9d12011-10-11 16:48:43 +11001909 struct r10bio *r10_bio = bio->bi_private;
NeilBrownfd01b882011-10-11 16:47:53 +11001910 struct mddev *mddev = r10_bio->mddev;
NeilBrowne879a872011-10-11 16:49:02 +11001911 struct r10conf *conf = mddev->private;
Namhyung Kim778ca012011-07-18 17:38:47 +10001912 int d;
NeilBrown749c55e2011-07-28 11:39:24 +10001913 sector_t first_bad;
1914 int bad_sectors;
1915 int slot;
NeilBrown9ad1aef2011-12-23 10:17:55 +11001916 int repl;
NeilBrown4ca40c22011-12-23 10:17:55 +11001917 struct md_rdev *rdev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918
NeilBrown9ad1aef2011-12-23 10:17:55 +11001919 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1920 if (repl)
1921 rdev = conf->mirrors[d].replacement;
NeilBrown547414d2012-03-13 11:21:20 +11001922 else
NeilBrown9ad1aef2011-12-23 10:17:55 +11001923 rdev = conf->mirrors[d].rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001925 if (bio->bi_error) {
NeilBrown9ad1aef2011-12-23 10:17:55 +11001926 if (repl)
1927 md_error(mddev, rdev);
1928 else {
1929 set_bit(WriteErrorSeen, &rdev->flags);
NeilBrownb7044d42011-12-23 10:17:56 +11001930 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1931 set_bit(MD_RECOVERY_NEEDED,
1932 &rdev->mddev->recovery);
NeilBrown9ad1aef2011-12-23 10:17:55 +11001933 set_bit(R10BIO_WriteError, &r10_bio->state);
1934 }
1935 } else if (is_badblock(rdev,
NeilBrown749c55e2011-07-28 11:39:24 +10001936 r10_bio->devs[slot].addr,
1937 r10_bio->sectors,
1938 &first_bad, &bad_sectors))
1939 set_bit(R10BIO_MadeGood, &r10_bio->state);
NeilBrowndfc70642008-05-23 13:04:39 -07001940
NeilBrown9ad1aef2011-12-23 10:17:55 +11001941 rdev_dec_pending(rdev, mddev);
NeilBrown5e570282011-07-28 11:39:25 +10001942
1943 end_sync_request(r10_bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944}
1945
1946/*
1947 * Note: sync and recover and handled very differently for raid10
1948 * This code is for resync.
1949 * For resync, we read through virtual addresses and read all blocks.
1950 * If there is any error, we schedule a write. The lowest numbered
1951 * drive is authoritative.
1952 * However requests come for physical address, so we need to map.
1953 * For every physical address there are raid_disks/copies virtual addresses,
1954 * which is always are least one, but is not necessarly an integer.
1955 * This means that a physical address can span multiple chunks, so we may
1956 * have to submit multiple io requests for a single sync request.
1957 */
1958/*
1959 * We check if all blocks are in-sync and only write to blocks that
1960 * aren't in sync
1961 */
NeilBrown9f2c9d12011-10-11 16:48:43 +11001962static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963{
NeilBrowne879a872011-10-11 16:49:02 +11001964 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 int i, first;
1966 struct bio *tbio, *fbio;
majianpengf4380a92012-04-12 16:04:47 +10001967 int vcnt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968
1969 atomic_set(&r10_bio->remaining, 1);
1970
1971 /* find the first device with a block */
1972 for (i=0; i<conf->copies; i++)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001973 if (!r10_bio->devs[i].bio->bi_error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974 break;
1975
1976 if (i == conf->copies)
1977 goto done;
1978
1979 first = i;
1980 fbio = r10_bio->devs[i].bio;
Artur Paszkiewiczcc578582015-12-18 15:19:16 +11001981 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
1982 fbio->bi_iter.bi_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983
majianpengf4380a92012-04-12 16:04:47 +10001984 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 /* now find blocks with errors */
NeilBrown0eb3ff12006-01-06 00:20:29 -08001986 for (i=0 ; i < conf->copies ; i++) {
1987 int j, d;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 tbio = r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001990
1991 if (tbio->bi_end_io != end_sync_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992 continue;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001993 if (i == first)
1994 continue;
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001995 if (!r10_bio->devs[i].bio->bi_error) {
NeilBrown0eb3ff12006-01-06 00:20:29 -08001996 /* We know that the bi_io_vec layout is the same for
1997 * both 'first' and 'i', so we just compare them.
1998 * All vec entries are PAGE_SIZE;
1999 */
NeilBrown7bb23c42013-07-16 16:50:47 +10002000 int sectors = r10_bio->sectors;
2001 for (j = 0; j < vcnt; j++) {
2002 int len = PAGE_SIZE;
2003 if (sectors < (len / 512))
2004 len = sectors * 512;
NeilBrown0eb3ff12006-01-06 00:20:29 -08002005 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2006 page_address(tbio->bi_io_vec[j].bv_page),
NeilBrown7bb23c42013-07-16 16:50:47 +10002007 len))
NeilBrown0eb3ff12006-01-06 00:20:29 -08002008 break;
NeilBrown7bb23c42013-07-16 16:50:47 +10002009 sectors -= len/512;
2010 }
NeilBrown0eb3ff12006-01-06 00:20:29 -08002011 if (j == vcnt)
2012 continue;
Jianpeng Ma7f7583d2012-10-11 14:17:59 +11002013 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
NeilBrownf84ee362011-07-28 11:39:25 +10002014 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2015 /* Don't fix anything. */
2016 continue;
NeilBrown0eb3ff12006-01-06 00:20:29 -08002017 }
NeilBrownf84ee362011-07-28 11:39:25 +10002018 /* Ok, we need to write this bio, either to correct an
2019 * inconsistency or to correct an unreadable block.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020 * First we need to fixup bv_offset, bv_len and
2021 * bi_vecs, as the read request might have corrupted these
2022 */
Kent Overstreet8be185f2012-09-06 14:14:43 -07002023 bio_reset(tbio);
2024
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025 tbio->bi_vcnt = vcnt;
Artur Paszkiewiczcc578582015-12-18 15:19:16 +11002026 tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 tbio->bi_private = r10_bio;
Kent Overstreet4f024f32013-10-11 15:44:27 -07002028 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029 tbio->bi_end_io = end_sync_write;
Mike Christie796a5cf2016-06-05 14:32:07 -05002030 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031
Kent Overstreetc31df252015-05-06 23:34:20 -07002032 bio_copy_data(tbio, fbio);
2033
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 d = r10_bio->devs[i].devnum;
2035 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2036 atomic_inc(&r10_bio->remaining);
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08002037 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038
Kent Overstreet4f024f32013-10-11 15:44:27 -07002039 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2041 generic_make_request(tbio);
2042 }
2043
NeilBrown9ad1aef2011-12-23 10:17:55 +11002044 /* Now write out to any replacement devices
2045 * that are active
2046 */
2047 for (i = 0; i < conf->copies; i++) {
Kent Overstreetc31df252015-05-06 23:34:20 -07002048 int d;
NeilBrown9ad1aef2011-12-23 10:17:55 +11002049
2050 tbio = r10_bio->devs[i].repl_bio;
2051 if (!tbio || !tbio->bi_end_io)
2052 continue;
2053 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2054 && r10_bio->devs[i].bio != fbio)
Kent Overstreetc31df252015-05-06 23:34:20 -07002055 bio_copy_data(tbio, fbio);
NeilBrown9ad1aef2011-12-23 10:17:55 +11002056 d = r10_bio->devs[i].devnum;
2057 atomic_inc(&r10_bio->remaining);
2058 md_sync_acct(conf->mirrors[d].replacement->bdev,
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08002059 bio_sectors(tbio));
NeilBrown9ad1aef2011-12-23 10:17:55 +11002060 generic_make_request(tbio);
2061 }
2062
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063done:
2064 if (atomic_dec_and_test(&r10_bio->remaining)) {
2065 md_done_sync(mddev, r10_bio->sectors, 1);
2066 put_buf(r10_bio);
2067 }
2068}
2069
2070/*
2071 * Now for the recovery code.
2072 * Recovery happens across physical sectors.
2073 * We recover all non-is_sync drives by finding the virtual address of
2074 * each, and then choose a working drive that also has that virt address.
2075 * There is a separate r10_bio for each non-in_sync drive.
2076 * Only the first two slots are in use. The first for reading,
2077 * The second for writing.
2078 *
2079 */
NeilBrown9f2c9d12011-10-11 16:48:43 +11002080static void fix_recovery_read_error(struct r10bio *r10_bio)
NeilBrown5e570282011-07-28 11:39:25 +10002081{
2082 /* We got a read error during recovery.
2083 * We repeat the read in smaller page-sized sections.
2084 * If a read succeeds, write it to the new device or record
2085 * a bad block if we cannot.
2086 * If a read fails, record a bad block on both old and
2087 * new devices.
2088 */
NeilBrownfd01b882011-10-11 16:47:53 +11002089 struct mddev *mddev = r10_bio->mddev;
NeilBrowne879a872011-10-11 16:49:02 +11002090 struct r10conf *conf = mddev->private;
NeilBrown5e570282011-07-28 11:39:25 +10002091 struct bio *bio = r10_bio->devs[0].bio;
2092 sector_t sect = 0;
2093 int sectors = r10_bio->sectors;
2094 int idx = 0;
2095 int dr = r10_bio->devs[0].devnum;
2096 int dw = r10_bio->devs[1].devnum;
2097
2098 while (sectors) {
2099 int s = sectors;
NeilBrown3cb03002011-10-11 16:45:26 +11002100 struct md_rdev *rdev;
NeilBrown5e570282011-07-28 11:39:25 +10002101 sector_t addr;
2102 int ok;
2103
2104 if (s > (PAGE_SIZE>>9))
2105 s = PAGE_SIZE >> 9;
2106
2107 rdev = conf->mirrors[dr].rdev;
2108 addr = r10_bio->devs[0].addr + sect,
2109 ok = sync_page_io(rdev,
2110 addr,
2111 s << 9,
2112 bio->bi_io_vec[idx].bv_page,
Mike Christie796a5cf2016-06-05 14:32:07 -05002113 REQ_OP_READ, 0, false);
NeilBrown5e570282011-07-28 11:39:25 +10002114 if (ok) {
2115 rdev = conf->mirrors[dw].rdev;
2116 addr = r10_bio->devs[1].addr + sect;
2117 ok = sync_page_io(rdev,
2118 addr,
2119 s << 9,
2120 bio->bi_io_vec[idx].bv_page,
Mike Christie796a5cf2016-06-05 14:32:07 -05002121 REQ_OP_WRITE, 0, false);
NeilBrownb7044d42011-12-23 10:17:56 +11002122 if (!ok) {
NeilBrown5e570282011-07-28 11:39:25 +10002123 set_bit(WriteErrorSeen, &rdev->flags);
NeilBrownb7044d42011-12-23 10:17:56 +11002124 if (!test_and_set_bit(WantReplacement,
2125 &rdev->flags))
2126 set_bit(MD_RECOVERY_NEEDED,
2127 &rdev->mddev->recovery);
2128 }
NeilBrown5e570282011-07-28 11:39:25 +10002129 }
2130 if (!ok) {
2131 /* We don't worry if we cannot set a bad block -
2132 * it really is bad so there is no loss in not
2133 * recording it yet
2134 */
2135 rdev_set_badblocks(rdev, addr, s, 0);
2136
2137 if (rdev != conf->mirrors[dw].rdev) {
2138 /* need bad block on destination too */
NeilBrown3cb03002011-10-11 16:45:26 +11002139 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
NeilBrown5e570282011-07-28 11:39:25 +10002140 addr = r10_bio->devs[1].addr + sect;
2141 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2142 if (!ok) {
2143 /* just abort the recovery */
2144 printk(KERN_NOTICE
2145 "md/raid10:%s: recovery aborted"
2146 " due to read error\n",
2147 mdname(mddev));
2148
2149 conf->mirrors[dw].recovery_disabled
2150 = mddev->recovery_disabled;
2151 set_bit(MD_RECOVERY_INTR,
2152 &mddev->recovery);
2153 break;
2154 }
2155 }
2156 }
2157
2158 sectors -= s;
2159 sect += s;
2160 idx++;
2161 }
2162}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163
NeilBrown9f2c9d12011-10-11 16:48:43 +11002164static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165{
NeilBrowne879a872011-10-11 16:49:02 +11002166 struct r10conf *conf = mddev->private;
Namhyung Kimc65060a2011-07-18 17:38:49 +10002167 int d;
NeilBrown24afd802011-12-23 10:17:55 +11002168 struct bio *wbio, *wbio2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169
NeilBrown5e570282011-07-28 11:39:25 +10002170 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2171 fix_recovery_read_error(r10_bio);
2172 end_sync_request(r10_bio);
2173 return;
2174 }
2175
Namhyung Kimc65060a2011-07-18 17:38:49 +10002176 /*
2177 * share the pages with the first bio
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 * and submit the write request
2179 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 d = r10_bio->devs[1].devnum;
NeilBrown24afd802011-12-23 10:17:55 +11002181 wbio = r10_bio->devs[1].bio;
2182 wbio2 = r10_bio->devs[1].repl_bio;
NeilBrown0eb25bb2013-07-24 15:37:42 +10002183 /* Need to test wbio2->bi_end_io before we call
2184 * generic_make_request as if the former is NULL,
2185 * the latter is free to free wbio2.
2186 */
2187 if (wbio2 && !wbio2->bi_end_io)
2188 wbio2 = NULL;
NeilBrown24afd802011-12-23 10:17:55 +11002189 if (wbio->bi_end_io) {
2190 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08002191 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
NeilBrown24afd802011-12-23 10:17:55 +11002192 generic_make_request(wbio);
2193 }
NeilBrown0eb25bb2013-07-24 15:37:42 +10002194 if (wbio2) {
NeilBrown24afd802011-12-23 10:17:55 +11002195 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2196 md_sync_acct(conf->mirrors[d].replacement->bdev,
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08002197 bio_sectors(wbio2));
NeilBrown24afd802011-12-23 10:17:55 +11002198 generic_make_request(wbio2);
2199 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200}
2201
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202/*
Robert Becker1e509152009-12-14 12:49:58 +11002203 * Used by fix_read_error() to decay the per rdev read_errors.
2204 * We halve the read error count for every hour that has elapsed
2205 * since the last recorded read error.
2206 *
2207 */
NeilBrownfd01b882011-10-11 16:47:53 +11002208static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
Robert Becker1e509152009-12-14 12:49:58 +11002209{
Arnd Bergmann0e3ef492016-06-17 17:33:10 +02002210 long cur_time_mon;
Robert Becker1e509152009-12-14 12:49:58 +11002211 unsigned long hours_since_last;
2212 unsigned int read_errors = atomic_read(&rdev->read_errors);
2213
Arnd Bergmann0e3ef492016-06-17 17:33:10 +02002214 cur_time_mon = ktime_get_seconds();
Robert Becker1e509152009-12-14 12:49:58 +11002215
Arnd Bergmann0e3ef492016-06-17 17:33:10 +02002216 if (rdev->last_read_error == 0) {
Robert Becker1e509152009-12-14 12:49:58 +11002217 /* first time we've seen a read error */
2218 rdev->last_read_error = cur_time_mon;
2219 return;
2220 }
2221
Arnd Bergmann0e3ef492016-06-17 17:33:10 +02002222 hours_since_last = (long)(cur_time_mon -
2223 rdev->last_read_error) / 3600;
Robert Becker1e509152009-12-14 12:49:58 +11002224
2225 rdev->last_read_error = cur_time_mon;
2226
2227 /*
2228 * if hours_since_last is > the number of bits in read_errors
2229 * just set read errors to 0. We do this to avoid
2230 * overflowing the shift of read_errors by hours_since_last.
2231 */
2232 if (hours_since_last >= 8 * sizeof(read_errors))
2233 atomic_set(&rdev->read_errors, 0);
2234 else
2235 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2236}
2237
NeilBrown3cb03002011-10-11 16:45:26 +11002238static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
NeilBrown58c54fc2011-07-28 11:39:25 +10002239 int sectors, struct page *page, int rw)
2240{
2241 sector_t first_bad;
2242 int bad_sectors;
2243
2244 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2245 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2246 return -1;
Mike Christie796a5cf2016-06-05 14:32:07 -05002247 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
NeilBrown58c54fc2011-07-28 11:39:25 +10002248 /* success */
2249 return 1;
NeilBrownb7044d42011-12-23 10:17:56 +11002250 if (rw == WRITE) {
NeilBrown58c54fc2011-07-28 11:39:25 +10002251 set_bit(WriteErrorSeen, &rdev->flags);
NeilBrownb7044d42011-12-23 10:17:56 +11002252 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2253 set_bit(MD_RECOVERY_NEEDED,
2254 &rdev->mddev->recovery);
2255 }
NeilBrown58c54fc2011-07-28 11:39:25 +10002256 /* need to record an error - either for the block or the device */
2257 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2258 md_error(rdev->mddev, rdev);
2259 return 0;
2260}
2261
Robert Becker1e509152009-12-14 12:49:58 +11002262/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 * This is a kernel thread which:
2264 *
2265 * 1. Retries failed read operations on working mirrors.
2266 * 2. Updates the raid superblock when problems encounter.
NeilBrown6814d532006-10-03 01:15:45 -07002267 * 3. Performs writes following reads for array synchronising.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 */
2269
NeilBrowne879a872011-10-11 16:49:02 +11002270static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
NeilBrown6814d532006-10-03 01:15:45 -07002271{
2272 int sect = 0; /* Offset from r10_bio->sector */
2273 int sectors = r10_bio->sectors;
NeilBrown3cb03002011-10-11 16:45:26 +11002274 struct md_rdev*rdev;
Robert Becker1e509152009-12-14 12:49:58 +11002275 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10002276 int d = r10_bio->devs[r10_bio->read_slot].devnum;
Robert Becker1e509152009-12-14 12:49:58 +11002277
NeilBrown7c4e06f2011-05-11 14:53:17 +10002278 /* still own a reference to this rdev, so it cannot
2279 * have been cleared recently.
2280 */
2281 rdev = conf->mirrors[d].rdev;
Robert Becker1e509152009-12-14 12:49:58 +11002282
NeilBrown7c4e06f2011-05-11 14:53:17 +10002283 if (test_bit(Faulty, &rdev->flags))
2284 /* drive has already been failed, just ignore any
2285 more fix_read_error() attempts */
2286 return;
2287
2288 check_decay_read_errors(mddev, rdev);
2289 atomic_inc(&rdev->read_errors);
2290 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2291 char b[BDEVNAME_SIZE];
Robert Becker1e509152009-12-14 12:49:58 +11002292 bdevname(rdev->bdev, b);
2293
NeilBrown7c4e06f2011-05-11 14:53:17 +10002294 printk(KERN_NOTICE
2295 "md/raid10:%s: %s: Raid device exceeded "
2296 "read_error threshold [cur %d:max %d]\n",
2297 mdname(mddev), b,
2298 atomic_read(&rdev->read_errors), max_read_errors);
2299 printk(KERN_NOTICE
2300 "md/raid10:%s: %s: Failing raid device\n",
2301 mdname(mddev), b);
NeilBrownd683c8e2016-06-02 16:19:52 +10002302 md_error(mddev, rdev);
NeilBrownfae8cc52012-02-14 11:10:10 +11002303 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
NeilBrown7c4e06f2011-05-11 14:53:17 +10002304 return;
Robert Becker1e509152009-12-14 12:49:58 +11002305 }
Robert Becker1e509152009-12-14 12:49:58 +11002306
NeilBrown6814d532006-10-03 01:15:45 -07002307 while(sectors) {
2308 int s = sectors;
2309 int sl = r10_bio->read_slot;
2310 int success = 0;
2311 int start;
2312
2313 if (s > (PAGE_SIZE>>9))
2314 s = PAGE_SIZE >> 9;
2315
2316 rcu_read_lock();
2317 do {
NeilBrown8dbed5c2011-07-28 11:39:24 +10002318 sector_t first_bad;
2319 int bad_sectors;
2320
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10002321 d = r10_bio->devs[sl].devnum;
NeilBrown6814d532006-10-03 01:15:45 -07002322 rdev = rcu_dereference(conf->mirrors[d].rdev);
2323 if (rdev &&
NeilBrown8dbed5c2011-07-28 11:39:24 +10002324 test_bit(In_sync, &rdev->flags) &&
NeilBrownf5b67ae2016-06-02 16:19:53 +10002325 !test_bit(Faulty, &rdev->flags) &&
NeilBrown8dbed5c2011-07-28 11:39:24 +10002326 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2327 &first_bad, &bad_sectors) == 0) {
NeilBrown6814d532006-10-03 01:15:45 -07002328 atomic_inc(&rdev->nr_pending);
2329 rcu_read_unlock();
NeilBrown2b193362010-10-27 15:16:40 +11002330 success = sync_page_io(rdev,
NeilBrown6814d532006-10-03 01:15:45 -07002331 r10_bio->devs[sl].addr +
Jonathan Brassowccebd4c2011-01-14 09:14:33 +11002332 sect,
NeilBrown6814d532006-10-03 01:15:45 -07002333 s<<9,
Mike Christie796a5cf2016-06-05 14:32:07 -05002334 conf->tmppage,
2335 REQ_OP_READ, 0, false);
NeilBrown6814d532006-10-03 01:15:45 -07002336 rdev_dec_pending(rdev, mddev);
2337 rcu_read_lock();
2338 if (success)
2339 break;
2340 }
2341 sl++;
2342 if (sl == conf->copies)
2343 sl = 0;
2344 } while (!success && sl != r10_bio->read_slot);
2345 rcu_read_unlock();
2346
2347 if (!success) {
NeilBrown58c54fc2011-07-28 11:39:25 +10002348 /* Cannot read from anywhere, just mark the block
2349 * as bad on the first device to discourage future
2350 * reads.
2351 */
NeilBrown6814d532006-10-03 01:15:45 -07002352 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
NeilBrown58c54fc2011-07-28 11:39:25 +10002353 rdev = conf->mirrors[dn].rdev;
2354
2355 if (!rdev_set_badblocks(
2356 rdev,
2357 r10_bio->devs[r10_bio->read_slot].addr
2358 + sect,
NeilBrownfae8cc52012-02-14 11:10:10 +11002359 s, 0)) {
NeilBrown58c54fc2011-07-28 11:39:25 +10002360 md_error(mddev, rdev);
NeilBrownfae8cc52012-02-14 11:10:10 +11002361 r10_bio->devs[r10_bio->read_slot].bio
2362 = IO_BLOCKED;
2363 }
NeilBrown6814d532006-10-03 01:15:45 -07002364 break;
2365 }
2366
2367 start = sl;
2368 /* write it back and re-read */
2369 rcu_read_lock();
2370 while (sl != r10_bio->read_slot) {
Robert Becker67b8dc42009-12-14 12:49:57 +11002371 char b[BDEVNAME_SIZE];
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10002372
NeilBrown6814d532006-10-03 01:15:45 -07002373 if (sl==0)
2374 sl = conf->copies;
2375 sl--;
2376 d = r10_bio->devs[sl].devnum;
2377 rdev = rcu_dereference(conf->mirrors[d].rdev);
NeilBrown1294b9c2011-07-28 11:39:23 +10002378 if (!rdev ||
NeilBrownf5b67ae2016-06-02 16:19:53 +10002379 test_bit(Faulty, &rdev->flags) ||
NeilBrown1294b9c2011-07-28 11:39:23 +10002380 !test_bit(In_sync, &rdev->flags))
2381 continue;
2382
2383 atomic_inc(&rdev->nr_pending);
2384 rcu_read_unlock();
NeilBrown58c54fc2011-07-28 11:39:25 +10002385 if (r10_sync_page_io(rdev,
2386 r10_bio->devs[sl].addr +
2387 sect,
NeilBrown055d3742012-07-03 15:55:33 +10002388 s, conf->tmppage, WRITE)
NeilBrown1294b9c2011-07-28 11:39:23 +10002389 == 0) {
2390 /* Well, this device is dead */
2391 printk(KERN_NOTICE
2392 "md/raid10:%s: read correction "
2393 "write failed"
2394 " (%d sectors at %llu on %s)\n",
2395 mdname(mddev), s,
2396 (unsigned long long)(
NeilBrownf8c9e742012-05-21 09:28:33 +10002397 sect +
2398 choose_data_offset(r10_bio,
2399 rdev)),
NeilBrown1294b9c2011-07-28 11:39:23 +10002400 bdevname(rdev->bdev, b));
2401 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2402 "drive\n",
2403 mdname(mddev),
2404 bdevname(rdev->bdev, b));
NeilBrown6814d532006-10-03 01:15:45 -07002405 }
NeilBrown1294b9c2011-07-28 11:39:23 +10002406 rdev_dec_pending(rdev, mddev);
2407 rcu_read_lock();
NeilBrown6814d532006-10-03 01:15:45 -07002408 }
2409 sl = start;
2410 while (sl != r10_bio->read_slot) {
NeilBrown1294b9c2011-07-28 11:39:23 +10002411 char b[BDEVNAME_SIZE];
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10002412
NeilBrown6814d532006-10-03 01:15:45 -07002413 if (sl==0)
2414 sl = conf->copies;
2415 sl--;
2416 d = r10_bio->devs[sl].devnum;
2417 rdev = rcu_dereference(conf->mirrors[d].rdev);
NeilBrown1294b9c2011-07-28 11:39:23 +10002418 if (!rdev ||
NeilBrownf5b67ae2016-06-02 16:19:53 +10002419 test_bit(Faulty, &rdev->flags) ||
NeilBrown1294b9c2011-07-28 11:39:23 +10002420 !test_bit(In_sync, &rdev->flags))
2421 continue;
Robert Becker67b8dc42009-12-14 12:49:57 +11002422
NeilBrown1294b9c2011-07-28 11:39:23 +10002423 atomic_inc(&rdev->nr_pending);
2424 rcu_read_unlock();
NeilBrown58c54fc2011-07-28 11:39:25 +10002425 switch (r10_sync_page_io(rdev,
2426 r10_bio->devs[sl].addr +
2427 sect,
NeilBrown055d3742012-07-03 15:55:33 +10002428 s, conf->tmppage,
NeilBrown58c54fc2011-07-28 11:39:25 +10002429 READ)) {
2430 case 0:
NeilBrown1294b9c2011-07-28 11:39:23 +10002431 /* Well, this device is dead */
2432 printk(KERN_NOTICE
2433 "md/raid10:%s: unable to read back "
2434 "corrected sectors"
2435 " (%d sectors at %llu on %s)\n",
2436 mdname(mddev), s,
2437 (unsigned long long)(
NeilBrownf8c9e742012-05-21 09:28:33 +10002438 sect +
2439 choose_data_offset(r10_bio, rdev)),
NeilBrown1294b9c2011-07-28 11:39:23 +10002440 bdevname(rdev->bdev, b));
2441 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2442 "drive\n",
2443 mdname(mddev),
2444 bdevname(rdev->bdev, b));
NeilBrown58c54fc2011-07-28 11:39:25 +10002445 break;
2446 case 1:
NeilBrown1294b9c2011-07-28 11:39:23 +10002447 printk(KERN_INFO
2448 "md/raid10:%s: read error corrected"
2449 " (%d sectors at %llu on %s)\n",
2450 mdname(mddev), s,
2451 (unsigned long long)(
NeilBrownf8c9e742012-05-21 09:28:33 +10002452 sect +
2453 choose_data_offset(r10_bio, rdev)),
NeilBrown1294b9c2011-07-28 11:39:23 +10002454 bdevname(rdev->bdev, b));
2455 atomic_add(s, &rdev->corrected_errors);
NeilBrown6814d532006-10-03 01:15:45 -07002456 }
NeilBrown1294b9c2011-07-28 11:39:23 +10002457
2458 rdev_dec_pending(rdev, mddev);
2459 rcu_read_lock();
NeilBrown6814d532006-10-03 01:15:45 -07002460 }
2461 rcu_read_unlock();
2462
2463 sectors -= s;
2464 sect += s;
2465 }
2466}
2467
NeilBrown9f2c9d12011-10-11 16:48:43 +11002468static int narrow_write_error(struct r10bio *r10_bio, int i)
NeilBrownbd870a12011-07-28 11:39:24 +10002469{
2470 struct bio *bio = r10_bio->master_bio;
NeilBrownfd01b882011-10-11 16:47:53 +11002471 struct mddev *mddev = r10_bio->mddev;
NeilBrowne879a872011-10-11 16:49:02 +11002472 struct r10conf *conf = mddev->private;
NeilBrown3cb03002011-10-11 16:45:26 +11002473 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
NeilBrownbd870a12011-07-28 11:39:24 +10002474 /* bio has the data to be written to slot 'i' where
2475 * we just recently had a write error.
2476 * We repeatedly clone the bio and trim down to one block,
2477 * then try the write. Where the write fails we record
2478 * a bad block.
2479 * It is conceivable that the bio doesn't exactly align with
2480 * blocks. We must handle this.
2481 *
2482 * We currently own a reference to the rdev.
2483 */
2484
2485 int block_sectors;
2486 sector_t sector;
2487 int sectors;
2488 int sect_to_write = r10_bio->sectors;
2489 int ok = 1;
2490
2491 if (rdev->badblocks.shift < 0)
2492 return 0;
2493
NeilBrownf04ebb02015-02-16 14:51:54 +11002494 block_sectors = roundup(1 << rdev->badblocks.shift,
2495 bdev_logical_block_size(rdev->bdev) >> 9);
NeilBrownbd870a12011-07-28 11:39:24 +10002496 sector = r10_bio->sector;
2497 sectors = ((r10_bio->sector + block_sectors)
2498 & ~(sector_t)(block_sectors - 1))
2499 - sector;
2500
2501 while (sect_to_write) {
2502 struct bio *wbio;
Tomasz Majchrzak27028622016-08-23 10:53:57 +02002503 sector_t wsector;
NeilBrownbd870a12011-07-28 11:39:24 +10002504 if (sectors > sect_to_write)
2505 sectors = sect_to_write;
2506 /* Write at 'sector' for 'sectors' */
2507 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
Kent Overstreet4f024f32013-10-11 15:44:27 -07002508 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
Tomasz Majchrzak27028622016-08-23 10:53:57 +02002509 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2510 wbio->bi_iter.bi_sector = wsector +
2511 choose_data_offset(r10_bio, rdev);
NeilBrownbd870a12011-07-28 11:39:24 +10002512 wbio->bi_bdev = rdev->bdev;
Mike Christie796a5cf2016-06-05 14:32:07 -05002513 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
Mike Christie4e49ea42016-06-05 14:31:41 -05002514
2515 if (submit_bio_wait(wbio) < 0)
NeilBrownbd870a12011-07-28 11:39:24 +10002516 /* Failure! */
Tomasz Majchrzak27028622016-08-23 10:53:57 +02002517 ok = rdev_set_badblocks(rdev, wsector,
NeilBrownbd870a12011-07-28 11:39:24 +10002518 sectors, 0)
2519 && ok;
2520
2521 bio_put(wbio);
2522 sect_to_write -= sectors;
2523 sector += sectors;
2524 sectors = block_sectors;
2525 }
2526 return ok;
2527}
2528
NeilBrown9f2c9d12011-10-11 16:48:43 +11002529static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
NeilBrown560f8e52011-07-28 11:39:23 +10002530{
2531 int slot = r10_bio->read_slot;
NeilBrown560f8e52011-07-28 11:39:23 +10002532 struct bio *bio;
NeilBrowne879a872011-10-11 16:49:02 +11002533 struct r10conf *conf = mddev->private;
NeilBrownabbf0982011-12-23 10:17:54 +11002534 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
NeilBrown560f8e52011-07-28 11:39:23 +10002535 char b[BDEVNAME_SIZE];
2536 unsigned long do_sync;
NeilBrown856e08e2011-07-28 11:39:23 +10002537 int max_sectors;
NeilBrown560f8e52011-07-28 11:39:23 +10002538
2539 /* we got a read error. Maybe the drive is bad. Maybe just
2540 * the block and we can fix it.
2541 * We freeze all other IO, and try reading the block from
2542 * other devices. When we find one, we re-write
2543 * and check it that fixes the read error.
2544 * This is all done synchronously while the array is
2545 * frozen.
2546 */
NeilBrownfae8cc52012-02-14 11:10:10 +11002547 bio = r10_bio->devs[slot].bio;
2548 bdevname(bio->bi_bdev, b);
2549 bio_put(bio);
2550 r10_bio->devs[slot].bio = NULL;
2551
NeilBrown560f8e52011-07-28 11:39:23 +10002552 if (mddev->ro == 0) {
NeilBrowne2d59922013-06-12 11:01:22 +10002553 freeze_array(conf, 1);
NeilBrown560f8e52011-07-28 11:39:23 +10002554 fix_read_error(conf, mddev, r10_bio);
2555 unfreeze_array(conf);
NeilBrownfae8cc52012-02-14 11:10:10 +11002556 } else
2557 r10_bio->devs[slot].bio = IO_BLOCKED;
2558
NeilBrownabbf0982011-12-23 10:17:54 +11002559 rdev_dec_pending(rdev, mddev);
NeilBrown560f8e52011-07-28 11:39:23 +10002560
NeilBrown7399c312011-07-28 11:39:23 +10002561read_more:
NeilBrown96c3fd12011-12-23 10:17:54 +11002562 rdev = read_balance(conf, r10_bio, &max_sectors);
2563 if (rdev == NULL) {
NeilBrown560f8e52011-07-28 11:39:23 +10002564 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2565 " read error for block %llu\n",
NeilBrown7399c312011-07-28 11:39:23 +10002566 mdname(mddev), b,
NeilBrown560f8e52011-07-28 11:39:23 +10002567 (unsigned long long)r10_bio->sector);
2568 raid_end_bio_io(r10_bio);
NeilBrown560f8e52011-07-28 11:39:23 +10002569 return;
2570 }
2571
Jens Axboe1eff9d32016-08-05 15:35:16 -06002572 do_sync = (r10_bio->master_bio->bi_opf & REQ_SYNC);
NeilBrown560f8e52011-07-28 11:39:23 +10002573 slot = r10_bio->read_slot;
NeilBrown560f8e52011-07-28 11:39:23 +10002574 printk_ratelimited(
2575 KERN_ERR
NeilBrown055d3742012-07-03 15:55:33 +10002576 "md/raid10:%s: %s: redirecting "
NeilBrown560f8e52011-07-28 11:39:23 +10002577 "sector %llu to another mirror\n",
2578 mdname(mddev),
2579 bdevname(rdev->bdev, b),
2580 (unsigned long long)r10_bio->sector);
2581 bio = bio_clone_mddev(r10_bio->master_bio,
2582 GFP_NOIO, mddev);
Kent Overstreet4f024f32013-10-11 15:44:27 -07002583 bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
NeilBrown560f8e52011-07-28 11:39:23 +10002584 r10_bio->devs[slot].bio = bio;
NeilBrownabbf0982011-12-23 10:17:54 +11002585 r10_bio->devs[slot].rdev = rdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -07002586 bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
NeilBrownf8c9e742012-05-21 09:28:33 +10002587 + choose_data_offset(r10_bio, rdev);
NeilBrown560f8e52011-07-28 11:39:23 +10002588 bio->bi_bdev = rdev->bdev;
Mike Christie796a5cf2016-06-05 14:32:07 -05002589 bio_set_op_attrs(bio, REQ_OP_READ, do_sync);
NeilBrown560f8e52011-07-28 11:39:23 +10002590 bio->bi_private = r10_bio;
2591 bio->bi_end_io = raid10_end_read_request;
NeilBrown7399c312011-07-28 11:39:23 +10002592 if (max_sectors < r10_bio->sectors) {
2593 /* Drat - have to split this up more */
2594 struct bio *mbio = r10_bio->master_bio;
2595 int sectors_handled =
2596 r10_bio->sector + max_sectors
Kent Overstreet4f024f32013-10-11 15:44:27 -07002597 - mbio->bi_iter.bi_sector;
NeilBrown7399c312011-07-28 11:39:23 +10002598 r10_bio->sectors = max_sectors;
2599 spin_lock_irq(&conf->device_lock);
2600 if (mbio->bi_phys_segments == 0)
2601 mbio->bi_phys_segments = 2;
2602 else
2603 mbio->bi_phys_segments++;
2604 spin_unlock_irq(&conf->device_lock);
2605 generic_make_request(bio);
NeilBrown7399c312011-07-28 11:39:23 +10002606
2607 r10_bio = mempool_alloc(conf->r10bio_pool,
2608 GFP_NOIO);
2609 r10_bio->master_bio = mbio;
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08002610 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
NeilBrown7399c312011-07-28 11:39:23 +10002611 r10_bio->state = 0;
2612 set_bit(R10BIO_ReadError,
2613 &r10_bio->state);
2614 r10_bio->mddev = mddev;
Kent Overstreet4f024f32013-10-11 15:44:27 -07002615 r10_bio->sector = mbio->bi_iter.bi_sector
NeilBrown7399c312011-07-28 11:39:23 +10002616 + sectors_handled;
2617
2618 goto read_more;
2619 } else
2620 generic_make_request(bio);
NeilBrown560f8e52011-07-28 11:39:23 +10002621}
2622
NeilBrowne879a872011-10-11 16:49:02 +11002623static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
NeilBrown749c55e2011-07-28 11:39:24 +10002624{
2625 /* Some sort of write request has finished and it
2626 * succeeded in writing where we thought there was a
2627 * bad block. So forget the bad block.
NeilBrown1a0b7cd2011-07-28 11:39:25 +10002628 * Or possibly if failed and we need to record
2629 * a bad block.
NeilBrown749c55e2011-07-28 11:39:24 +10002630 */
2631 int m;
NeilBrown3cb03002011-10-11 16:45:26 +11002632 struct md_rdev *rdev;
NeilBrown749c55e2011-07-28 11:39:24 +10002633
2634 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2635 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
NeilBrown1a0b7cd2011-07-28 11:39:25 +10002636 for (m = 0; m < conf->copies; m++) {
2637 int dev = r10_bio->devs[m].devnum;
2638 rdev = conf->mirrors[dev].rdev;
2639 if (r10_bio->devs[m].bio == NULL)
2640 continue;
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002641 if (!r10_bio->devs[m].bio->bi_error) {
NeilBrown749c55e2011-07-28 11:39:24 +10002642 rdev_clear_badblocks(
2643 rdev,
2644 r10_bio->devs[m].addr,
NeilBrownc6563a82012-05-21 09:27:00 +10002645 r10_bio->sectors, 0);
NeilBrown1a0b7cd2011-07-28 11:39:25 +10002646 } else {
2647 if (!rdev_set_badblocks(
2648 rdev,
2649 r10_bio->devs[m].addr,
2650 r10_bio->sectors, 0))
2651 md_error(conf->mddev, rdev);
NeilBrown749c55e2011-07-28 11:39:24 +10002652 }
NeilBrown9ad1aef2011-12-23 10:17:55 +11002653 rdev = conf->mirrors[dev].replacement;
2654 if (r10_bio->devs[m].repl_bio == NULL)
2655 continue;
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002656
2657 if (!r10_bio->devs[m].repl_bio->bi_error) {
NeilBrown9ad1aef2011-12-23 10:17:55 +11002658 rdev_clear_badblocks(
2659 rdev,
2660 r10_bio->devs[m].addr,
NeilBrownc6563a82012-05-21 09:27:00 +10002661 r10_bio->sectors, 0);
NeilBrown9ad1aef2011-12-23 10:17:55 +11002662 } else {
2663 if (!rdev_set_badblocks(
2664 rdev,
2665 r10_bio->devs[m].addr,
2666 r10_bio->sectors, 0))
2667 md_error(conf->mddev, rdev);
2668 }
NeilBrown1a0b7cd2011-07-28 11:39:25 +10002669 }
NeilBrown749c55e2011-07-28 11:39:24 +10002670 put_buf(r10_bio);
2671 } else {
NeilBrown95af5872015-08-14 11:26:17 +10002672 bool fail = false;
NeilBrownbd870a12011-07-28 11:39:24 +10002673 for (m = 0; m < conf->copies; m++) {
2674 int dev = r10_bio->devs[m].devnum;
2675 struct bio *bio = r10_bio->devs[m].bio;
2676 rdev = conf->mirrors[dev].rdev;
2677 if (bio == IO_MADE_GOOD) {
NeilBrown749c55e2011-07-28 11:39:24 +10002678 rdev_clear_badblocks(
2679 rdev,
2680 r10_bio->devs[m].addr,
NeilBrownc6563a82012-05-21 09:27:00 +10002681 r10_bio->sectors, 0);
NeilBrown749c55e2011-07-28 11:39:24 +10002682 rdev_dec_pending(rdev, conf->mddev);
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002683 } else if (bio != NULL && bio->bi_error) {
NeilBrown95af5872015-08-14 11:26:17 +10002684 fail = true;
NeilBrownbd870a12011-07-28 11:39:24 +10002685 if (!narrow_write_error(r10_bio, m)) {
2686 md_error(conf->mddev, rdev);
2687 set_bit(R10BIO_Degraded,
2688 &r10_bio->state);
2689 }
2690 rdev_dec_pending(rdev, conf->mddev);
NeilBrown749c55e2011-07-28 11:39:24 +10002691 }
NeilBrown475b0322011-12-23 10:17:55 +11002692 bio = r10_bio->devs[m].repl_bio;
2693 rdev = conf->mirrors[dev].replacement;
NeilBrown4ca40c22011-12-23 10:17:55 +11002694 if (rdev && bio == IO_MADE_GOOD) {
NeilBrown475b0322011-12-23 10:17:55 +11002695 rdev_clear_badblocks(
2696 rdev,
2697 r10_bio->devs[m].addr,
NeilBrownc6563a82012-05-21 09:27:00 +10002698 r10_bio->sectors, 0);
NeilBrown475b0322011-12-23 10:17:55 +11002699 rdev_dec_pending(rdev, conf->mddev);
2700 }
NeilBrownbd870a12011-07-28 11:39:24 +10002701 }
NeilBrown95af5872015-08-14 11:26:17 +10002702 if (fail) {
2703 spin_lock_irq(&conf->device_lock);
2704 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
Shaohua Li23ddba82016-03-14 11:49:32 -07002705 conf->nr_queued++;
NeilBrown95af5872015-08-14 11:26:17 +10002706 spin_unlock_irq(&conf->device_lock);
2707 md_wakeup_thread(conf->mddev->thread);
NeilBrownc3407022015-10-24 16:23:48 +11002708 } else {
2709 if (test_bit(R10BIO_WriteError,
2710 &r10_bio->state))
2711 close_write(r10_bio);
NeilBrown95af5872015-08-14 11:26:17 +10002712 raid_end_bio_io(r10_bio);
NeilBrownc3407022015-10-24 16:23:48 +11002713 }
NeilBrown749c55e2011-07-28 11:39:24 +10002714 }
2715}
2716
Shaohua Li4ed87312012-10-11 13:34:00 +11002717static void raid10d(struct md_thread *thread)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718{
Shaohua Li4ed87312012-10-11 13:34:00 +11002719 struct mddev *mddev = thread->mddev;
NeilBrown9f2c9d12011-10-11 16:48:43 +11002720 struct r10bio *r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721 unsigned long flags;
NeilBrowne879a872011-10-11 16:49:02 +11002722 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 struct list_head *head = &conf->retry_list;
NeilBrowne1dfa0a2011-04-18 18:25:41 +10002724 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725
2726 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727
NeilBrown95af5872015-08-14 11:26:17 +10002728 if (!list_empty_careful(&conf->bio_end_io_list) &&
2729 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2730 LIST_HEAD(tmp);
2731 spin_lock_irqsave(&conf->device_lock, flags);
2732 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
Shaohua Li23ddba82016-03-14 11:49:32 -07002733 while (!list_empty(&conf->bio_end_io_list)) {
2734 list_move(conf->bio_end_io_list.prev, &tmp);
2735 conf->nr_queued--;
2736 }
NeilBrown95af5872015-08-14 11:26:17 +10002737 }
2738 spin_unlock_irqrestore(&conf->device_lock, flags);
2739 while (!list_empty(&tmp)) {
Mikulas Patockaa4527442015-10-01 15:17:43 -04002740 r10_bio = list_first_entry(&tmp, struct r10bio,
2741 retry_list);
NeilBrown95af5872015-08-14 11:26:17 +10002742 list_del(&r10_bio->retry_list);
NeilBrownc3407022015-10-24 16:23:48 +11002743 if (mddev->degraded)
2744 set_bit(R10BIO_Degraded, &r10_bio->state);
2745
2746 if (test_bit(R10BIO_WriteError,
2747 &r10_bio->state))
2748 close_write(r10_bio);
NeilBrown95af5872015-08-14 11:26:17 +10002749 raid_end_bio_io(r10_bio);
2750 }
2751 }
2752
NeilBrowne1dfa0a2011-04-18 18:25:41 +10002753 blk_start_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754 for (;;) {
NeilBrowna35e63e2008-03-04 14:29:29 -08002755
NeilBrown0021b7b2012-07-31 09:08:14 +02002756 flush_pending_writes(conf);
NeilBrowna35e63e2008-03-04 14:29:29 -08002757
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrowna35e63e2008-03-04 14:29:29 -08002759 if (list_empty(head)) {
NeilBrown6cce3b232006-01-06 00:20:16 -08002760 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761 break;
NeilBrowna35e63e2008-03-04 14:29:29 -08002762 }
NeilBrown9f2c9d12011-10-11 16:48:43 +11002763 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 list_del(head->prev);
NeilBrown4443ae12006-01-06 00:20:28 -08002765 conf->nr_queued--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 spin_unlock_irqrestore(&conf->device_lock, flags);
2767
2768 mddev = r10_bio->mddev;
NeilBrown070ec552009-06-16 16:54:21 +10002769 conf = mddev->private;
NeilBrownbd870a12011-07-28 11:39:24 +10002770 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2771 test_bit(R10BIO_WriteError, &r10_bio->state))
NeilBrown749c55e2011-07-28 11:39:24 +10002772 handle_write_completed(conf, r10_bio);
NeilBrown3ea7daa2012-05-22 13:53:47 +10002773 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2774 reshape_request_write(mddev, r10_bio);
NeilBrown749c55e2011-07-28 11:39:24 +10002775 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776 sync_request_write(mddev, r10_bio);
Jens Axboe7eaceac2011-03-10 08:52:07 +01002777 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 recovery_request_write(mddev, r10_bio);
NeilBrown856e08e2011-07-28 11:39:23 +10002779 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
NeilBrown560f8e52011-07-28 11:39:23 +10002780 handle_read_error(mddev, r10_bio);
NeilBrown856e08e2011-07-28 11:39:23 +10002781 else {
2782 /* just a partial read to be scheduled from a
2783 * separate context
2784 */
2785 int slot = r10_bio->read_slot;
2786 generic_make_request(r10_bio->devs[slot].bio);
2787 }
NeilBrown4443ae12006-01-06 00:20:28 -08002788
NeilBrown1d9d5242009-10-16 15:55:32 +11002789 cond_resched();
NeilBrownde393cd2011-07-28 11:31:48 +10002790 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2791 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 }
NeilBrowne1dfa0a2011-04-18 18:25:41 +10002793 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794}
2795
NeilBrowne879a872011-10-11 16:49:02 +11002796static int init_resync(struct r10conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797{
2798 int buffs;
NeilBrown69335ef2011-12-23 10:17:54 +11002799 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800
2801 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
Eric Sesterhennb6385482006-04-02 13:34:29 +02002802 BUG_ON(conf->r10buf_pool);
NeilBrown69335ef2011-12-23 10:17:54 +11002803 conf->have_replacement = 0;
NeilBrown5cf00fc2012-05-21 09:28:20 +10002804 for (i = 0; i < conf->geo.raid_disks; i++)
NeilBrown69335ef2011-12-23 10:17:54 +11002805 if (conf->mirrors[i].replacement)
2806 conf->have_replacement = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2808 if (!conf->r10buf_pool)
2809 return -ENOMEM;
2810 conf->next_resync = 0;
2811 return 0;
2812}
2813
2814/*
2815 * perform a "sync" on one "block"
2816 *
2817 * We need to make sure that no normal I/O request - particularly write
2818 * requests - conflict with active sync requests.
2819 *
2820 * This is achieved by tracking pending requests and a 'barrier' concept
2821 * that can be installed to exclude normal IO requests.
2822 *
2823 * Resync and recovery are handled very differently.
2824 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2825 *
2826 * For resync, we iterate over virtual addresses, read all copies,
2827 * and update if there are differences. If only one copy is live,
2828 * skip it.
2829 * For recovery, we iterate over physical addresses, read a good
2830 * value for each non-in_sync drive, and over-write.
2831 *
2832 * So, for recovery we may have several outstanding complex requests for a
2833 * given address, one for each out-of-sync device. We model this by allocating
2834 * a number of r10_bio structures, one for each out-of-sync device.
2835 * As we setup these structures, we collect all bio's together into a list
2836 * which we then process collectively to add pages, and then process again
2837 * to pass to generic_make_request.
2838 *
2839 * The r10_bio structures are linked using a borrowed master_bio pointer.
2840 * This link is counted in ->remaining. When the r10_bio that points to NULL
2841 * has its remaining count decremented to 0, the whole complex operation
2842 * is complete.
2843 *
2844 */
2845
Shaohua Li849674e2016-01-20 13:52:20 -08002846static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
NeilBrown09314792015-02-19 16:04:40 +11002847 int *skipped)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848{
NeilBrowne879a872011-10-11 16:49:02 +11002849 struct r10conf *conf = mddev->private;
NeilBrown9f2c9d12011-10-11 16:48:43 +11002850 struct r10bio *r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 struct bio *biolist = NULL, *bio;
2852 sector_t max_sector, nr_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 int i;
NeilBrown6cce3b232006-01-06 00:20:16 -08002854 int max_sync;
NeilBrown57dab0b2010-10-19 10:03:39 +11002855 sector_t sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856 sector_t sectors_skipped = 0;
2857 int chunks_skipped = 0;
NeilBrown5cf00fc2012-05-21 09:28:20 +10002858 sector_t chunk_mask = conf->geo.chunk_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859
2860 if (!conf->r10buf_pool)
2861 if (init_resync(conf))
NeilBrown57afd892005-06-21 17:17:13 -07002862 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863
Martin Wilck7e83ccb2013-04-24 11:42:42 +10002864 /*
2865 * Allow skipping a full rebuild for incremental assembly
2866 * of a clean array, like RAID1 does.
2867 */
2868 if (mddev->bitmap == NULL &&
2869 mddev->recovery_cp == MaxSector &&
NeilBrown13765122013-07-04 16:41:53 +10002870 mddev->reshape_position == MaxSector &&
2871 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
Martin Wilck7e83ccb2013-04-24 11:42:42 +10002872 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
NeilBrown13765122013-07-04 16:41:53 +10002873 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
Martin Wilck7e83ccb2013-04-24 11:42:42 +10002874 conf->fullsync == 0) {
2875 *skipped = 1;
NeilBrown13765122013-07-04 16:41:53 +10002876 return mddev->dev_sectors - sector_nr;
Martin Wilck7e83ccb2013-04-24 11:42:42 +10002877 }
2878
Linus Torvalds1da177e2005-04-16 15:20:36 -07002879 skipped:
Andre Noll58c0fed2009-03-31 14:33:13 +11002880 max_sector = mddev->dev_sectors;
NeilBrown3ea7daa2012-05-22 13:53:47 +10002881 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2882 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883 max_sector = mddev->resync_max_sectors;
2884 if (sector_nr >= max_sector) {
NeilBrown6cce3b232006-01-06 00:20:16 -08002885 /* If we aborted, we need to abort the
2886 * sync on the 'current' bitmap chucks (there can
2887 * be several when recovering multiple devices).
2888 * as we may have started syncing it but not finished.
2889 * We can find the current address in
2890 * mddev->curr_resync, but for recovery,
2891 * we need to convert that to several
2892 * virtual addresses.
2893 */
NeilBrown3ea7daa2012-05-22 13:53:47 +10002894 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2895 end_reshape(conf);
NeilBrownb3968552014-08-18 13:59:50 +10002896 close_sync(conf);
NeilBrown3ea7daa2012-05-22 13:53:47 +10002897 return 0;
2898 }
2899
NeilBrown6cce3b232006-01-06 00:20:16 -08002900 if (mddev->curr_resync < max_sector) { /* aborted */
2901 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2902 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2903 &sync_blocks, 1);
NeilBrown5cf00fc2012-05-21 09:28:20 +10002904 else for (i = 0; i < conf->geo.raid_disks; i++) {
NeilBrown6cce3b232006-01-06 00:20:16 -08002905 sector_t sect =
2906 raid10_find_virt(conf, mddev->curr_resync, i);
2907 bitmap_end_sync(mddev->bitmap, sect,
2908 &sync_blocks, 1);
2909 }
NeilBrown9ad1aef2011-12-23 10:17:55 +11002910 } else {
2911 /* completed sync */
2912 if ((!mddev->bitmap || conf->fullsync)
2913 && conf->have_replacement
2914 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2915 /* Completed a full sync so the replacements
2916 * are now fully recovered.
2917 */
NeilBrownf90145f2016-06-02 16:19:52 +10002918 rcu_read_lock();
2919 for (i = 0; i < conf->geo.raid_disks; i++) {
2920 struct md_rdev *rdev =
2921 rcu_dereference(conf->mirrors[i].replacement);
2922 if (rdev)
2923 rdev->recovery_offset = MaxSector;
2924 }
2925 rcu_read_unlock();
NeilBrown9ad1aef2011-12-23 10:17:55 +11002926 }
NeilBrown6cce3b232006-01-06 00:20:16 -08002927 conf->fullsync = 0;
NeilBrown9ad1aef2011-12-23 10:17:55 +11002928 }
NeilBrown6cce3b232006-01-06 00:20:16 -08002929 bitmap_close_sync(mddev->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930 close_sync(conf);
NeilBrown57afd892005-06-21 17:17:13 -07002931 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002932 return sectors_skipped;
2933 }
NeilBrown3ea7daa2012-05-22 13:53:47 +10002934
2935 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2936 return reshape_request(mddev, sector_nr, skipped);
2937
NeilBrown5cf00fc2012-05-21 09:28:20 +10002938 if (chunks_skipped >= conf->geo.raid_disks) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939 /* if there has been nothing to do on any drive,
2940 * then there is nothing to do at all..
2941 */
NeilBrown57afd892005-06-21 17:17:13 -07002942 *skipped = 1;
2943 return (max_sector - sector_nr) + sectors_skipped;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944 }
2945
NeilBrownc6207272008-02-06 01:39:52 -08002946 if (max_sector > mddev->resync_max)
2947 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2948
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949 /* make sure whole request will fit in a chunk - if chunks
2950 * are meaningful
2951 */
NeilBrown5cf00fc2012-05-21 09:28:20 +10002952 if (conf->geo.near_copies < conf->geo.raid_disks &&
2953 max_sector > (sector_nr | chunk_mask))
2954 max_sector = (sector_nr | chunk_mask) + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955
Tomasz Majchrzak7ac50442016-06-13 15:51:19 +02002956 /*
2957 * If there is non-resync activity waiting for a turn, then let it
2958 * though before starting on this new sync request.
2959 */
2960 if (conf->nr_waiting)
2961 schedule_timeout_uninterruptible(1);
2962
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 /* Again, very different code for resync and recovery.
2964 * Both must result in an r10bio with a list of bios that
2965 * have bi_end_io, bi_sector, bi_bdev set,
2966 * and bi_private set to the r10bio.
2967 * For recovery, we may actually create several r10bios
2968 * with 2 bios in each, that correspond to the bios in the main one.
2969 * In this case, the subordinate r10bios link back through a
2970 * borrowed master_bio pointer, and the counter in the master
2971 * includes a ref from each subordinate.
2972 */
2973 /* First, we decide what to do and set ->bi_end_io
2974 * To end_sync_read if we want to read, and
2975 * end_sync_write if we will want to write.
2976 */
2977
NeilBrown6cce3b232006-01-06 00:20:16 -08002978 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2980 /* recovery... the complicated one */
NeilBrowne875ece2011-07-28 11:39:24 +10002981 int j;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982 r10_bio = NULL;
2983
NeilBrown5cf00fc2012-05-21 09:28:20 +10002984 for (i = 0 ; i < conf->geo.raid_disks; i++) {
NeilBrownab9d47e2011-05-11 14:54:41 +10002985 int still_degraded;
NeilBrown9f2c9d12011-10-11 16:48:43 +11002986 struct r10bio *rb2;
NeilBrownab9d47e2011-05-11 14:54:41 +10002987 sector_t sect;
2988 int must_sync;
NeilBrowne875ece2011-07-28 11:39:24 +10002989 int any_working;
Jonathan Brassowdc280d982012-07-31 10:03:52 +10002990 struct raid10_info *mirror = &conf->mirrors[i];
NeilBrownf90145f2016-06-02 16:19:52 +10002991 struct md_rdev *mrdev, *mreplace;
NeilBrownab9d47e2011-05-11 14:54:41 +10002992
NeilBrownf90145f2016-06-02 16:19:52 +10002993 rcu_read_lock();
2994 mrdev = rcu_dereference(mirror->rdev);
2995 mreplace = rcu_dereference(mirror->replacement);
2996
2997 if ((mrdev == NULL ||
NeilBrownf5b67ae2016-06-02 16:19:53 +10002998 test_bit(Faulty, &mrdev->flags) ||
NeilBrownf90145f2016-06-02 16:19:52 +10002999 test_bit(In_sync, &mrdev->flags)) &&
3000 (mreplace == NULL ||
3001 test_bit(Faulty, &mreplace->flags))) {
3002 rcu_read_unlock();
NeilBrownab9d47e2011-05-11 14:54:41 +10003003 continue;
NeilBrownf90145f2016-06-02 16:19:52 +10003004 }
NeilBrownab9d47e2011-05-11 14:54:41 +10003005
3006 still_degraded = 0;
3007 /* want to reconstruct this device */
3008 rb2 = r10_bio;
3009 sect = raid10_find_virt(conf, sector_nr, i);
NeilBrownfc448a12012-07-03 10:37:30 +10003010 if (sect >= mddev->resync_max_sectors) {
3011 /* last stripe is not complete - don't
3012 * try to recover this sector.
3013 */
NeilBrownf90145f2016-06-02 16:19:52 +10003014 rcu_read_unlock();
NeilBrownfc448a12012-07-03 10:37:30 +10003015 continue;
3016 }
NeilBrownf5b67ae2016-06-02 16:19:53 +10003017 if (mreplace && test_bit(Faulty, &mreplace->flags))
3018 mreplace = NULL;
NeilBrown24afd802011-12-23 10:17:55 +11003019 /* Unless we are doing a full sync, or a replacement
3020 * we only need to recover the block if it is set in
3021 * the bitmap
NeilBrownab9d47e2011-05-11 14:54:41 +10003022 */
3023 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3024 &sync_blocks, 1);
3025 if (sync_blocks < max_sync)
3026 max_sync = sync_blocks;
3027 if (!must_sync &&
NeilBrownf90145f2016-06-02 16:19:52 +10003028 mreplace == NULL &&
NeilBrownab9d47e2011-05-11 14:54:41 +10003029 !conf->fullsync) {
3030 /* yep, skip the sync_blocks here, but don't assume
3031 * that there will never be anything to do here
NeilBrown6cce3b232006-01-06 00:20:16 -08003032 */
NeilBrownab9d47e2011-05-11 14:54:41 +10003033 chunks_skipped = -1;
NeilBrownf90145f2016-06-02 16:19:52 +10003034 rcu_read_unlock();
NeilBrownab9d47e2011-05-11 14:54:41 +10003035 continue;
3036 }
NeilBrownf90145f2016-06-02 16:19:52 +10003037 atomic_inc(&mrdev->nr_pending);
3038 if (mreplace)
3039 atomic_inc(&mreplace->nr_pending);
3040 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041
NeilBrownab9d47e2011-05-11 14:54:41 +10003042 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
NeilBrowncb8b12b2014-08-18 14:38:45 +10003043 r10_bio->state = 0;
NeilBrownab9d47e2011-05-11 14:54:41 +10003044 raise_barrier(conf, rb2 != NULL);
3045 atomic_set(&r10_bio->remaining, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003046
NeilBrownab9d47e2011-05-11 14:54:41 +10003047 r10_bio->master_bio = (struct bio*)rb2;
3048 if (rb2)
3049 atomic_inc(&rb2->remaining);
3050 r10_bio->mddev = mddev;
3051 set_bit(R10BIO_IsRecover, &r10_bio->state);
3052 r10_bio->sector = sect;
NeilBrown6cce3b232006-01-06 00:20:16 -08003053
NeilBrownab9d47e2011-05-11 14:54:41 +10003054 raid10_find_phys(conf, r10_bio);
NeilBrown18055562009-05-07 12:48:10 +10003055
NeilBrownab9d47e2011-05-11 14:54:41 +10003056 /* Need to check if the array will still be
3057 * degraded
3058 */
NeilBrownf90145f2016-06-02 16:19:52 +10003059 rcu_read_lock();
3060 for (j = 0; j < conf->geo.raid_disks; j++) {
3061 struct md_rdev *rdev = rcu_dereference(
3062 conf->mirrors[j].rdev);
3063 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
NeilBrownab9d47e2011-05-11 14:54:41 +10003064 still_degraded = 1;
NeilBrown87fc7672005-09-09 16:24:04 -07003065 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066 }
NeilBrownf90145f2016-06-02 16:19:52 +10003067 }
NeilBrownab9d47e2011-05-11 14:54:41 +10003068
3069 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3070 &sync_blocks, still_degraded);
3071
NeilBrowne875ece2011-07-28 11:39:24 +10003072 any_working = 0;
NeilBrownab9d47e2011-05-11 14:54:41 +10003073 for (j=0; j<conf->copies;j++) {
NeilBrowne875ece2011-07-28 11:39:24 +10003074 int k;
NeilBrownab9d47e2011-05-11 14:54:41 +10003075 int d = r10_bio->devs[j].devnum;
NeilBrown5e570282011-07-28 11:39:25 +10003076 sector_t from_addr, to_addr;
NeilBrownf90145f2016-06-02 16:19:52 +10003077 struct md_rdev *rdev =
3078 rcu_dereference(conf->mirrors[d].rdev);
NeilBrown40c356c2011-07-28 11:39:24 +10003079 sector_t sector, first_bad;
3080 int bad_sectors;
NeilBrownf90145f2016-06-02 16:19:52 +10003081 if (!rdev ||
3082 !test_bit(In_sync, &rdev->flags))
NeilBrownab9d47e2011-05-11 14:54:41 +10003083 continue;
3084 /* This is where we read from */
NeilBrowne875ece2011-07-28 11:39:24 +10003085 any_working = 1;
NeilBrown40c356c2011-07-28 11:39:24 +10003086 sector = r10_bio->devs[j].addr;
3087
3088 if (is_badblock(rdev, sector, max_sync,
3089 &first_bad, &bad_sectors)) {
3090 if (first_bad > sector)
3091 max_sync = first_bad - sector;
3092 else {
3093 bad_sectors -= (sector
3094 - first_bad);
3095 if (max_sync > bad_sectors)
3096 max_sync = bad_sectors;
3097 continue;
3098 }
3099 }
NeilBrownab9d47e2011-05-11 14:54:41 +10003100 bio = r10_bio->devs[0].bio;
Kent Overstreet8be185f2012-09-06 14:14:43 -07003101 bio_reset(bio);
NeilBrownab9d47e2011-05-11 14:54:41 +10003102 bio->bi_next = biolist;
3103 biolist = bio;
3104 bio->bi_private = r10_bio;
3105 bio->bi_end_io = end_sync_read;
Mike Christie796a5cf2016-06-05 14:32:07 -05003106 bio_set_op_attrs(bio, REQ_OP_READ, 0);
NeilBrown5e570282011-07-28 11:39:25 +10003107 from_addr = r10_bio->devs[j].addr;
Kent Overstreet4f024f32013-10-11 15:44:27 -07003108 bio->bi_iter.bi_sector = from_addr +
3109 rdev->data_offset;
NeilBrown24afd802011-12-23 10:17:55 +11003110 bio->bi_bdev = rdev->bdev;
3111 atomic_inc(&rdev->nr_pending);
3112 /* and we write to 'i' (if not in_sync) */
NeilBrownab9d47e2011-05-11 14:54:41 +10003113
3114 for (k=0; k<conf->copies; k++)
3115 if (r10_bio->devs[k].devnum == i)
3116 break;
3117 BUG_ON(k == conf->copies);
NeilBrown5e570282011-07-28 11:39:25 +10003118 to_addr = r10_bio->devs[k].addr;
NeilBrownab9d47e2011-05-11 14:54:41 +10003119 r10_bio->devs[0].devnum = d;
NeilBrown5e570282011-07-28 11:39:25 +10003120 r10_bio->devs[0].addr = from_addr;
NeilBrownab9d47e2011-05-11 14:54:41 +10003121 r10_bio->devs[1].devnum = i;
NeilBrown5e570282011-07-28 11:39:25 +10003122 r10_bio->devs[1].addr = to_addr;
NeilBrownab9d47e2011-05-11 14:54:41 +10003123
NeilBrownf90145f2016-06-02 16:19:52 +10003124 if (!test_bit(In_sync, &mrdev->flags)) {
NeilBrown24afd802011-12-23 10:17:55 +11003125 bio = r10_bio->devs[1].bio;
Kent Overstreet8be185f2012-09-06 14:14:43 -07003126 bio_reset(bio);
NeilBrown24afd802011-12-23 10:17:55 +11003127 bio->bi_next = biolist;
3128 biolist = bio;
3129 bio->bi_private = r10_bio;
3130 bio->bi_end_io = end_sync_write;
Mike Christie796a5cf2016-06-05 14:32:07 -05003131 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
Kent Overstreet4f024f32013-10-11 15:44:27 -07003132 bio->bi_iter.bi_sector = to_addr
NeilBrownf90145f2016-06-02 16:19:52 +10003133 + mrdev->data_offset;
3134 bio->bi_bdev = mrdev->bdev;
NeilBrown24afd802011-12-23 10:17:55 +11003135 atomic_inc(&r10_bio->remaining);
3136 } else
3137 r10_bio->devs[1].bio->bi_end_io = NULL;
3138
3139 /* and maybe write to replacement */
3140 bio = r10_bio->devs[1].repl_bio;
3141 if (bio)
3142 bio->bi_end_io = NULL;
NeilBrownf90145f2016-06-02 16:19:52 +10003143 /* Note: if mreplace != NULL, then bio
NeilBrown24afd802011-12-23 10:17:55 +11003144 * cannot be NULL as r10buf_pool_alloc will
3145 * have allocated it.
3146 * So the second test here is pointless.
3147 * But it keeps semantic-checkers happy, and
3148 * this comment keeps human reviewers
3149 * happy.
3150 */
NeilBrownf90145f2016-06-02 16:19:52 +10003151 if (mreplace == NULL || bio == NULL ||
3152 test_bit(Faulty, &mreplace->flags))
NeilBrown24afd802011-12-23 10:17:55 +11003153 break;
Kent Overstreet8be185f2012-09-06 14:14:43 -07003154 bio_reset(bio);
NeilBrown24afd802011-12-23 10:17:55 +11003155 bio->bi_next = biolist;
3156 biolist = bio;
3157 bio->bi_private = r10_bio;
3158 bio->bi_end_io = end_sync_write;
Mike Christie796a5cf2016-06-05 14:32:07 -05003159 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
Kent Overstreet4f024f32013-10-11 15:44:27 -07003160 bio->bi_iter.bi_sector = to_addr +
NeilBrownf90145f2016-06-02 16:19:52 +10003161 mreplace->data_offset;
3162 bio->bi_bdev = mreplace->bdev;
NeilBrown24afd802011-12-23 10:17:55 +11003163 atomic_inc(&r10_bio->remaining);
NeilBrownab9d47e2011-05-11 14:54:41 +10003164 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165 }
NeilBrownf90145f2016-06-02 16:19:52 +10003166 rcu_read_unlock();
NeilBrownab9d47e2011-05-11 14:54:41 +10003167 if (j == conf->copies) {
NeilBrowne875ece2011-07-28 11:39:24 +10003168 /* Cannot recover, so abort the recovery or
3169 * record a bad block */
NeilBrowne875ece2011-07-28 11:39:24 +10003170 if (any_working) {
3171 /* problem is that there are bad blocks
3172 * on other device(s)
3173 */
3174 int k;
3175 for (k = 0; k < conf->copies; k++)
3176 if (r10_bio->devs[k].devnum == i)
3177 break;
NeilBrown24afd802011-12-23 10:17:55 +11003178 if (!test_bit(In_sync,
NeilBrownf90145f2016-06-02 16:19:52 +10003179 &mrdev->flags)
NeilBrown24afd802011-12-23 10:17:55 +11003180 && !rdev_set_badblocks(
NeilBrownf90145f2016-06-02 16:19:52 +10003181 mrdev,
NeilBrown24afd802011-12-23 10:17:55 +11003182 r10_bio->devs[k].addr,
3183 max_sync, 0))
3184 any_working = 0;
NeilBrownf90145f2016-06-02 16:19:52 +10003185 if (mreplace &&
NeilBrown24afd802011-12-23 10:17:55 +11003186 !rdev_set_badblocks(
NeilBrownf90145f2016-06-02 16:19:52 +10003187 mreplace,
NeilBrowne875ece2011-07-28 11:39:24 +10003188 r10_bio->devs[k].addr,
3189 max_sync, 0))
3190 any_working = 0;
3191 }
3192 if (!any_working) {
3193 if (!test_and_set_bit(MD_RECOVERY_INTR,
3194 &mddev->recovery))
3195 printk(KERN_INFO "md/raid10:%s: insufficient "
3196 "working devices for recovery.\n",
3197 mdname(mddev));
NeilBrown24afd802011-12-23 10:17:55 +11003198 mirror->recovery_disabled
NeilBrowne875ece2011-07-28 11:39:24 +10003199 = mddev->recovery_disabled;
3200 }
NeilBrowne8b84912014-01-06 10:35:34 +11003201 put_buf(r10_bio);
3202 if (rb2)
3203 atomic_dec(&rb2->remaining);
3204 r10_bio = rb2;
NeilBrownf90145f2016-06-02 16:19:52 +10003205 rdev_dec_pending(mrdev, mddev);
3206 if (mreplace)
3207 rdev_dec_pending(mreplace, mddev);
NeilBrownab9d47e2011-05-11 14:54:41 +10003208 break;
3209 }
NeilBrownf90145f2016-06-02 16:19:52 +10003210 rdev_dec_pending(mrdev, mddev);
3211 if (mreplace)
3212 rdev_dec_pending(mreplace, mddev);
NeilBrownab9d47e2011-05-11 14:54:41 +10003213 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214 if (biolist == NULL) {
3215 while (r10_bio) {
NeilBrown9f2c9d12011-10-11 16:48:43 +11003216 struct r10bio *rb2 = r10_bio;
3217 r10_bio = (struct r10bio*) rb2->master_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 rb2->master_bio = NULL;
3219 put_buf(rb2);
3220 }
3221 goto giveup;
3222 }
3223 } else {
3224 /* resync. Schedule a read for every block at this virt offset */
3225 int count = 0;
NeilBrown6cce3b232006-01-06 00:20:16 -08003226
Goldwyn Rodriguesc40f3412015-08-19 08:14:42 +10003227 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
NeilBrown78200d42009-02-25 13:18:47 +11003228
NeilBrown6cce3b232006-01-06 00:20:16 -08003229 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3230 &sync_blocks, mddev->degraded) &&
NeilBrownab9d47e2011-05-11 14:54:41 +10003231 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3232 &mddev->recovery)) {
NeilBrown6cce3b232006-01-06 00:20:16 -08003233 /* We can skip this block */
3234 *skipped = 1;
3235 return sync_blocks + sectors_skipped;
3236 }
3237 if (sync_blocks < max_sync)
3238 max_sync = sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003239 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
NeilBrowncb8b12b2014-08-18 14:38:45 +10003240 r10_bio->state = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242 r10_bio->mddev = mddev;
3243 atomic_set(&r10_bio->remaining, 0);
NeilBrown6cce3b232006-01-06 00:20:16 -08003244 raise_barrier(conf, 0);
3245 conf->next_resync = sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246
3247 r10_bio->master_bio = NULL;
3248 r10_bio->sector = sector_nr;
3249 set_bit(R10BIO_IsSync, &r10_bio->state);
3250 raid10_find_phys(conf, r10_bio);
NeilBrown5cf00fc2012-05-21 09:28:20 +10003251 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252
NeilBrown5cf00fc2012-05-21 09:28:20 +10003253 for (i = 0; i < conf->copies; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254 int d = r10_bio->devs[i].devnum;
NeilBrown40c356c2011-07-28 11:39:24 +10003255 sector_t first_bad, sector;
3256 int bad_sectors;
NeilBrownf90145f2016-06-02 16:19:52 +10003257 struct md_rdev *rdev;
NeilBrown40c356c2011-07-28 11:39:24 +10003258
NeilBrown9ad1aef2011-12-23 10:17:55 +11003259 if (r10_bio->devs[i].repl_bio)
3260 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3261
Linus Torvalds1da177e2005-04-16 15:20:36 -07003262 bio = r10_bio->devs[i].bio;
Kent Overstreet8be185f2012-09-06 14:14:43 -07003263 bio_reset(bio);
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02003264 bio->bi_error = -EIO;
NeilBrownf90145f2016-06-02 16:19:52 +10003265 rcu_read_lock();
3266 rdev = rcu_dereference(conf->mirrors[d].rdev);
3267 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3268 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269 continue;
NeilBrownf90145f2016-06-02 16:19:52 +10003270 }
NeilBrown40c356c2011-07-28 11:39:24 +10003271 sector = r10_bio->devs[i].addr;
NeilBrownf90145f2016-06-02 16:19:52 +10003272 if (is_badblock(rdev, sector, max_sync,
NeilBrown40c356c2011-07-28 11:39:24 +10003273 &first_bad, &bad_sectors)) {
3274 if (first_bad > sector)
3275 max_sync = first_bad - sector;
3276 else {
3277 bad_sectors -= (sector - first_bad);
3278 if (max_sync > bad_sectors)
Dan Carpenter91502f02012-10-11 14:20:58 +11003279 max_sync = bad_sectors;
NeilBrownf90145f2016-06-02 16:19:52 +10003280 rcu_read_unlock();
NeilBrown40c356c2011-07-28 11:39:24 +10003281 continue;
3282 }
3283 }
NeilBrownf90145f2016-06-02 16:19:52 +10003284 atomic_inc(&rdev->nr_pending);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285 atomic_inc(&r10_bio->remaining);
3286 bio->bi_next = biolist;
3287 biolist = bio;
3288 bio->bi_private = r10_bio;
3289 bio->bi_end_io = end_sync_read;
Mike Christie796a5cf2016-06-05 14:32:07 -05003290 bio_set_op_attrs(bio, REQ_OP_READ, 0);
NeilBrownf90145f2016-06-02 16:19:52 +10003291 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3292 bio->bi_bdev = rdev->bdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293 count++;
NeilBrown9ad1aef2011-12-23 10:17:55 +11003294
NeilBrownf90145f2016-06-02 16:19:52 +10003295 rdev = rcu_dereference(conf->mirrors[d].replacement);
3296 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3297 rcu_read_unlock();
NeilBrown9ad1aef2011-12-23 10:17:55 +11003298 continue;
NeilBrownf90145f2016-06-02 16:19:52 +10003299 }
3300 atomic_inc(&rdev->nr_pending);
3301 rcu_read_unlock();
NeilBrown9ad1aef2011-12-23 10:17:55 +11003302
3303 /* Need to set up for writing to the replacement */
3304 bio = r10_bio->devs[i].repl_bio;
Kent Overstreet8be185f2012-09-06 14:14:43 -07003305 bio_reset(bio);
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02003306 bio->bi_error = -EIO;
NeilBrown9ad1aef2011-12-23 10:17:55 +11003307
3308 sector = r10_bio->devs[i].addr;
NeilBrown9ad1aef2011-12-23 10:17:55 +11003309 bio->bi_next = biolist;
3310 biolist = bio;
3311 bio->bi_private = r10_bio;
3312 bio->bi_end_io = end_sync_write;
Mike Christie796a5cf2016-06-05 14:32:07 -05003313 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
NeilBrownf90145f2016-06-02 16:19:52 +10003314 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3315 bio->bi_bdev = rdev->bdev;
NeilBrown9ad1aef2011-12-23 10:17:55 +11003316 count++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317 }
3318
3319 if (count < 2) {
3320 for (i=0; i<conf->copies; i++) {
3321 int d = r10_bio->devs[i].devnum;
3322 if (r10_bio->devs[i].bio->bi_end_io)
NeilBrownab9d47e2011-05-11 14:54:41 +10003323 rdev_dec_pending(conf->mirrors[d].rdev,
3324 mddev);
NeilBrown9ad1aef2011-12-23 10:17:55 +11003325 if (r10_bio->devs[i].repl_bio &&
3326 r10_bio->devs[i].repl_bio->bi_end_io)
3327 rdev_dec_pending(
3328 conf->mirrors[d].replacement,
3329 mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330 }
3331 put_buf(r10_bio);
3332 biolist = NULL;
3333 goto giveup;
3334 }
3335 }
3336
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337 nr_sectors = 0;
NeilBrown6cce3b232006-01-06 00:20:16 -08003338 if (sector_nr + max_sync < max_sector)
3339 max_sector = sector_nr + max_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003340 do {
3341 struct page *page;
3342 int len = PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343 if (sector_nr + (len>>9) > max_sector)
3344 len = (max_sector - sector_nr) << 9;
3345 if (len == 0)
3346 break;
3347 for (bio= biolist ; bio ; bio=bio->bi_next) {
NeilBrownab9d47e2011-05-11 14:54:41 +10003348 struct bio *bio2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
NeilBrownab9d47e2011-05-11 14:54:41 +10003350 if (bio_add_page(bio, page, len, 0))
3351 continue;
3352
3353 /* stop here */
3354 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3355 for (bio2 = biolist;
3356 bio2 && bio2 != bio;
3357 bio2 = bio2->bi_next) {
3358 /* remove last page from this bio */
3359 bio2->bi_vcnt--;
Kent Overstreet4f024f32013-10-11 15:44:27 -07003360 bio2->bi_iter.bi_size -= len;
Jens Axboeb7c44ed2015-07-24 12:37:59 -06003361 bio_clear_flag(bio2, BIO_SEG_VALID);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362 }
NeilBrownab9d47e2011-05-11 14:54:41 +10003363 goto bio_full;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364 }
3365 nr_sectors += len>>9;
3366 sector_nr += len>>9;
3367 } while (biolist->bi_vcnt < RESYNC_PAGES);
3368 bio_full:
3369 r10_bio->sectors = nr_sectors;
3370
3371 while (biolist) {
3372 bio = biolist;
3373 biolist = biolist->bi_next;
3374
3375 bio->bi_next = NULL;
3376 r10_bio = bio->bi_private;
3377 r10_bio->sectors = nr_sectors;
3378
3379 if (bio->bi_end_io == end_sync_read) {
3380 md_sync_acct(bio->bi_bdev, nr_sectors);
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02003381 bio->bi_error = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382 generic_make_request(bio);
3383 }
3384 }
3385
NeilBrown57afd892005-06-21 17:17:13 -07003386 if (sectors_skipped)
3387 /* pretend they weren't skipped, it makes
3388 * no important difference in this case
3389 */
3390 md_done_sync(mddev, sectors_skipped, 1);
3391
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392 return sectors_skipped + nr_sectors;
3393 giveup:
3394 /* There is nowhere to write, so all non-sync
NeilBrowne875ece2011-07-28 11:39:24 +10003395 * drives must be failed or in resync, all drives
3396 * have a bad block, so try the next chunk...
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397 */
NeilBrown09b40682009-02-25 13:18:47 +11003398 if (sector_nr + max_sync < max_sector)
3399 max_sector = sector_nr + max_sync;
3400
3401 sectors_skipped += (max_sector - sector_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402 chunks_skipped ++;
3403 sector_nr = max_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404 goto skipped;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405}
3406
Dan Williams80c3a6c2009-03-17 18:10:40 -07003407static sector_t
NeilBrownfd01b882011-10-11 16:47:53 +11003408raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
Dan Williams80c3a6c2009-03-17 18:10:40 -07003409{
3410 sector_t size;
NeilBrowne879a872011-10-11 16:49:02 +11003411 struct r10conf *conf = mddev->private;
Dan Williams80c3a6c2009-03-17 18:10:40 -07003412
3413 if (!raid_disks)
NeilBrown3ea7daa2012-05-22 13:53:47 +10003414 raid_disks = min(conf->geo.raid_disks,
3415 conf->prev.raid_disks);
Dan Williams80c3a6c2009-03-17 18:10:40 -07003416 if (!sectors)
Trela, Maciejdab8b292010-03-08 16:02:45 +11003417 sectors = conf->dev_sectors;
Dan Williams80c3a6c2009-03-17 18:10:40 -07003418
NeilBrown5cf00fc2012-05-21 09:28:20 +10003419 size = sectors >> conf->geo.chunk_shift;
3420 sector_div(size, conf->geo.far_copies);
Dan Williams80c3a6c2009-03-17 18:10:40 -07003421 size = size * raid_disks;
NeilBrown5cf00fc2012-05-21 09:28:20 +10003422 sector_div(size, conf->geo.near_copies);
Dan Williams80c3a6c2009-03-17 18:10:40 -07003423
NeilBrown5cf00fc2012-05-21 09:28:20 +10003424 return size << conf->geo.chunk_shift;
Dan Williams80c3a6c2009-03-17 18:10:40 -07003425}
3426
NeilBrown6508fdb2012-05-17 10:08:45 +10003427static void calc_sectors(struct r10conf *conf, sector_t size)
3428{
3429 /* Calculate the number of sectors-per-device that will
3430 * actually be used, and set conf->dev_sectors and
3431 * conf->stride
3432 */
3433
NeilBrown5cf00fc2012-05-21 09:28:20 +10003434 size = size >> conf->geo.chunk_shift;
3435 sector_div(size, conf->geo.far_copies);
3436 size = size * conf->geo.raid_disks;
3437 sector_div(size, conf->geo.near_copies);
NeilBrown6508fdb2012-05-17 10:08:45 +10003438 /* 'size' is now the number of chunks in the array */
3439 /* calculate "used chunks per device" */
3440 size = size * conf->copies;
3441
3442 /* We need to round up when dividing by raid_disks to
3443 * get the stride size.
3444 */
NeilBrown5cf00fc2012-05-21 09:28:20 +10003445 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
NeilBrown6508fdb2012-05-17 10:08:45 +10003446
NeilBrown5cf00fc2012-05-21 09:28:20 +10003447 conf->dev_sectors = size << conf->geo.chunk_shift;
NeilBrown6508fdb2012-05-17 10:08:45 +10003448
NeilBrown5cf00fc2012-05-21 09:28:20 +10003449 if (conf->geo.far_offset)
3450 conf->geo.stride = 1 << conf->geo.chunk_shift;
NeilBrown6508fdb2012-05-17 10:08:45 +10003451 else {
NeilBrown5cf00fc2012-05-21 09:28:20 +10003452 sector_div(size, conf->geo.far_copies);
3453 conf->geo.stride = size << conf->geo.chunk_shift;
NeilBrown6508fdb2012-05-17 10:08:45 +10003454 }
3455}
Trela, Maciejdab8b292010-03-08 16:02:45 +11003456
NeilBrowndeb200d2012-05-21 09:28:33 +10003457enum geo_type {geo_new, geo_old, geo_start};
3458static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3459{
3460 int nc, fc, fo;
3461 int layout, chunk, disks;
3462 switch (new) {
3463 case geo_old:
3464 layout = mddev->layout;
3465 chunk = mddev->chunk_sectors;
3466 disks = mddev->raid_disks - mddev->delta_disks;
3467 break;
3468 case geo_new:
3469 layout = mddev->new_layout;
3470 chunk = mddev->new_chunk_sectors;
3471 disks = mddev->raid_disks;
3472 break;
3473 default: /* avoid 'may be unused' warnings */
3474 case geo_start: /* new when starting reshape - raid_disks not
3475 * updated yet. */
3476 layout = mddev->new_layout;
3477 chunk = mddev->new_chunk_sectors;
3478 disks = mddev->raid_disks + mddev->delta_disks;
3479 break;
3480 }
NeilBrown8bce6d32015-10-22 13:20:15 +11003481 if (layout >> 19)
NeilBrowndeb200d2012-05-21 09:28:33 +10003482 return -1;
3483 if (chunk < (PAGE_SIZE >> 9) ||
3484 !is_power_of_2(chunk))
3485 return -2;
3486 nc = layout & 255;
3487 fc = (layout >> 8) & 255;
3488 fo = layout & (1<<16);
3489 geo->raid_disks = disks;
3490 geo->near_copies = nc;
3491 geo->far_copies = fc;
3492 geo->far_offset = fo;
NeilBrown8bce6d32015-10-22 13:20:15 +11003493 switch (layout >> 17) {
3494 case 0: /* original layout. simple but not always optimal */
3495 geo->far_set_size = disks;
3496 break;
3497 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3498 * actually using this, but leave code here just in case.*/
3499 geo->far_set_size = disks/fc;
3500 WARN(geo->far_set_size < fc,
3501 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3502 break;
3503 case 2: /* "improved" layout fixed to match documentation */
3504 geo->far_set_size = fc * nc;
3505 break;
3506 default: /* Not a valid layout */
3507 return -1;
3508 }
NeilBrowndeb200d2012-05-21 09:28:33 +10003509 geo->chunk_mask = chunk - 1;
3510 geo->chunk_shift = ffz(~chunk);
3511 return nc*fc;
3512}
3513
NeilBrowne879a872011-10-11 16:49:02 +11003514static struct r10conf *setup_conf(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515{
NeilBrowne879a872011-10-11 16:49:02 +11003516 struct r10conf *conf = NULL;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003517 int err = -EINVAL;
NeilBrowndeb200d2012-05-21 09:28:33 +10003518 struct geom geo;
3519 int copies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520
NeilBrowndeb200d2012-05-21 09:28:33 +10003521 copies = setup_geo(&geo, mddev, geo_new);
3522
3523 if (copies == -2) {
NeilBrown128595e2010-05-03 14:47:14 +10003524 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3525 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3526 mdname(mddev), PAGE_SIZE);
Trela, Maciejdab8b292010-03-08 16:02:45 +11003527 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003528 }
NeilBrown2604b702006-01-06 00:20:36 -08003529
NeilBrowndeb200d2012-05-21 09:28:33 +10003530 if (copies < 2 || copies > mddev->raid_disks) {
NeilBrown128595e2010-05-03 14:47:14 +10003531 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
Maciej Trelaf73ea872010-06-16 11:46:29 +01003532 mdname(mddev), mddev->new_layout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533 goto out;
3534 }
Trela, Maciejdab8b292010-03-08 16:02:45 +11003535
3536 err = -ENOMEM;
NeilBrowne879a872011-10-11 16:49:02 +11003537 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
Trela, Maciejdab8b292010-03-08 16:02:45 +11003538 if (!conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 goto out;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003540
NeilBrown3ea7daa2012-05-22 13:53:47 +10003541 /* FIXME calc properly */
Jonathan Brassowdc280d982012-07-31 10:03:52 +10003542 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
NeilBrown78eaa0d2013-07-02 15:58:05 +10003543 max(0,-mddev->delta_disks)),
Trela, Maciejdab8b292010-03-08 16:02:45 +11003544 GFP_KERNEL);
3545 if (!conf->mirrors)
3546 goto out;
NeilBrown4443ae12006-01-06 00:20:28 -08003547
3548 conf->tmppage = alloc_page(GFP_KERNEL);
3549 if (!conf->tmppage)
Trela, Maciejdab8b292010-03-08 16:02:45 +11003550 goto out;
3551
NeilBrowndeb200d2012-05-21 09:28:33 +10003552 conf->geo = geo;
3553 conf->copies = copies;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003554 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3555 r10bio_pool_free, conf);
3556 if (!conf->r10bio_pool)
3557 goto out;
3558
NeilBrown6508fdb2012-05-17 10:08:45 +10003559 calc_sectors(conf, mddev->dev_sectors);
NeilBrown3ea7daa2012-05-22 13:53:47 +10003560 if (mddev->reshape_position == MaxSector) {
3561 conf->prev = conf->geo;
3562 conf->reshape_progress = MaxSector;
3563 } else {
3564 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3565 err = -EINVAL;
3566 goto out;
3567 }
3568 conf->reshape_progress = mddev->reshape_position;
3569 if (conf->prev.far_offset)
3570 conf->prev.stride = 1 << conf->prev.chunk_shift;
3571 else
3572 /* far_copies must be 1 */
3573 conf->prev.stride = conf->dev_sectors;
3574 }
NeilBrown299b0682015-07-06 17:37:49 +10003575 conf->reshape_safe = conf->reshape_progress;
Neil Browne7e72bf2008-05-14 16:05:54 -07003576 spin_lock_init(&conf->device_lock);
Trela, Maciejdab8b292010-03-08 16:02:45 +11003577 INIT_LIST_HEAD(&conf->retry_list);
NeilBrown95af5872015-08-14 11:26:17 +10003578 INIT_LIST_HEAD(&conf->bio_end_io_list);
Trela, Maciejdab8b292010-03-08 16:02:45 +11003579
3580 spin_lock_init(&conf->resync_lock);
3581 init_waitqueue_head(&conf->wait_barrier);
Tomasz Majchrzak0e5313e2016-06-24 14:20:16 +02003582 atomic_set(&conf->nr_pending, 0);
Trela, Maciejdab8b292010-03-08 16:02:45 +11003583
NeilBrown02326052012-07-03 15:56:52 +10003584 conf->thread = md_register_thread(raid10d, mddev, "raid10");
Trela, Maciejdab8b292010-03-08 16:02:45 +11003585 if (!conf->thread)
3586 goto out;
3587
Trela, Maciejdab8b292010-03-08 16:02:45 +11003588 conf->mddev = mddev;
3589 return conf;
3590
3591 out:
NeilBrown3ea7daa2012-05-22 13:53:47 +10003592 if (err == -ENOMEM)
3593 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3594 mdname(mddev));
Trela, Maciejdab8b292010-03-08 16:02:45 +11003595 if (conf) {
Julia Lawall644df1a2015-09-13 14:15:10 +02003596 mempool_destroy(conf->r10bio_pool);
Trela, Maciejdab8b292010-03-08 16:02:45 +11003597 kfree(conf->mirrors);
3598 safe_put_page(conf->tmppage);
3599 kfree(conf);
3600 }
3601 return ERR_PTR(err);
3602}
3603
Shaohua Li849674e2016-01-20 13:52:20 -08003604static int raid10_run(struct mddev *mddev)
Trela, Maciejdab8b292010-03-08 16:02:45 +11003605{
NeilBrowne879a872011-10-11 16:49:02 +11003606 struct r10conf *conf;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003607 int i, disk_idx, chunk_size;
Jonathan Brassowdc280d982012-07-31 10:03:52 +10003608 struct raid10_info *disk;
NeilBrown3cb03002011-10-11 16:45:26 +11003609 struct md_rdev *rdev;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003610 sector_t size;
NeilBrown3ea7daa2012-05-22 13:53:47 +10003611 sector_t min_offset_diff = 0;
3612 int first = 1;
Shaohua Li532a2a32012-10-11 13:30:52 +11003613 bool discard_supported = false;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003614
3615 if (mddev->private == NULL) {
3616 conf = setup_conf(mddev);
3617 if (IS_ERR(conf))
3618 return PTR_ERR(conf);
3619 mddev->private = conf;
3620 }
3621 conf = mddev->private;
3622 if (!conf)
3623 goto out;
3624
Trela, Maciejdab8b292010-03-08 16:02:45 +11003625 mddev->thread = conf->thread;
3626 conf->thread = NULL;
3627
Martin K. Petersen8f6c2e42009-07-01 11:13:45 +10003628 chunk_size = mddev->chunk_sectors << 9;
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003629 if (mddev->queue) {
Shaohua Li532a2a32012-10-11 13:30:52 +11003630 blk_queue_max_discard_sectors(mddev->queue,
3631 mddev->chunk_sectors);
H. Peter Anvin5026d7a2013-06-12 07:37:43 -07003632 blk_queue_max_write_same_sectors(mddev->queue, 0);
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003633 blk_queue_io_min(mddev->queue, chunk_size);
3634 if (conf->geo.raid_disks % conf->geo.near_copies)
3635 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3636 else
3637 blk_queue_io_opt(mddev->queue, chunk_size *
3638 (conf->geo.raid_disks / conf->geo.near_copies));
3639 }
Martin K. Petersen8f6c2e42009-07-01 11:13:45 +10003640
NeilBrowndafb20f2012-03-19 12:46:39 +11003641 rdev_for_each(rdev, mddev) {
NeilBrown3ea7daa2012-05-22 13:53:47 +10003642 long long diff;
NeilBrownaba336b2012-05-31 15:39:11 +10003643 struct request_queue *q;
NeilBrown34b343c2011-07-28 11:31:47 +10003644
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645 disk_idx = rdev->raid_disk;
NeilBrownf8c9e742012-05-21 09:28:33 +10003646 if (disk_idx < 0)
3647 continue;
3648 if (disk_idx >= conf->geo.raid_disks &&
3649 disk_idx >= conf->prev.raid_disks)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650 continue;
3651 disk = conf->mirrors + disk_idx;
3652
NeilBrown56a25592011-12-23 10:17:55 +11003653 if (test_bit(Replacement, &rdev->flags)) {
3654 if (disk->replacement)
3655 goto out_free_conf;
3656 disk->replacement = rdev;
3657 } else {
3658 if (disk->rdev)
3659 goto out_free_conf;
3660 disk->rdev = rdev;
3661 }
NeilBrownaba336b2012-05-31 15:39:11 +10003662 q = bdev_get_queue(rdev->bdev);
NeilBrown3ea7daa2012-05-22 13:53:47 +10003663 diff = (rdev->new_data_offset - rdev->data_offset);
3664 if (!mddev->reshape_backwards)
3665 diff = -diff;
3666 if (diff < 0)
3667 diff = 0;
3668 if (first || diff < min_offset_diff)
3669 min_offset_diff = diff;
NeilBrown56a25592011-12-23 10:17:55 +11003670
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003671 if (mddev->gendisk)
3672 disk_stack_limits(mddev->gendisk, rdev->bdev,
3673 rdev->data_offset << 9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003674
3675 disk->head_position = 0;
Shaohua Li532a2a32012-10-11 13:30:52 +11003676
3677 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3678 discard_supported = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003679 }
NeilBrown3ea7daa2012-05-22 13:53:47 +10003680
Jonathan Brassowed30be02012-10-31 11:42:30 +11003681 if (mddev->queue) {
3682 if (discard_supported)
3683 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3684 mddev->queue);
3685 else
3686 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3687 mddev->queue);
3688 }
NeilBrown6d508242005-09-09 16:24:03 -07003689 /* need to check that every block has at least one working mirror */
NeilBrown700c7212011-07-27 11:00:36 +10003690 if (!enough(conf, -1)) {
NeilBrown128595e2010-05-03 14:47:14 +10003691 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
NeilBrown6d508242005-09-09 16:24:03 -07003692 mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693 goto out_free_conf;
3694 }
3695
NeilBrown3ea7daa2012-05-22 13:53:47 +10003696 if (conf->reshape_progress != MaxSector) {
3697 /* must ensure that shape change is supported */
3698 if (conf->geo.far_copies != 1 &&
3699 conf->geo.far_offset == 0)
3700 goto out_free_conf;
3701 if (conf->prev.far_copies != 1 &&
NeilBrown78eaa0d2013-07-02 15:58:05 +10003702 conf->prev.far_offset == 0)
NeilBrown3ea7daa2012-05-22 13:53:47 +10003703 goto out_free_conf;
3704 }
3705
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706 mddev->degraded = 0;
NeilBrownf8c9e742012-05-21 09:28:33 +10003707 for (i = 0;
3708 i < conf->geo.raid_disks
3709 || i < conf->prev.raid_disks;
3710 i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711
3712 disk = conf->mirrors + i;
3713
NeilBrown56a25592011-12-23 10:17:55 +11003714 if (!disk->rdev && disk->replacement) {
3715 /* The replacement is all we have - use it */
3716 disk->rdev = disk->replacement;
3717 disk->replacement = NULL;
3718 clear_bit(Replacement, &disk->rdev->flags);
3719 }
3720
NeilBrown5fd6c1d2006-06-26 00:27:40 -07003721 if (!disk->rdev ||
NeilBrown2e333e82006-10-21 10:24:07 -07003722 !test_bit(In_sync, &disk->rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003723 disk->head_position = 0;
3724 mddev->degraded++;
NeilBrown0b59bb62014-01-14 16:30:10 +11003725 if (disk->rdev &&
3726 disk->rdev->saved_raid_disk < 0)
Neil Brown8c2e8702008-06-28 08:30:52 +10003727 conf->fullsync = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003728 }
NeilBrownd890fa22011-10-26 11:54:39 +11003729 disk->recovery_disabled = mddev->recovery_disabled - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003730 }
3731
Andre Noll8c6ac862009-06-18 08:48:06 +10003732 if (mddev->recovery_cp != MaxSector)
NeilBrown128595e2010-05-03 14:47:14 +10003733 printk(KERN_NOTICE "md/raid10:%s: not clean"
Andre Noll8c6ac862009-06-18 08:48:06 +10003734 " -- starting background reconstruction\n",
3735 mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003736 printk(KERN_INFO
NeilBrown128595e2010-05-03 14:47:14 +10003737 "md/raid10:%s: active with %d out of %d devices\n",
NeilBrown5cf00fc2012-05-21 09:28:20 +10003738 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3739 conf->geo.raid_disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740 /*
3741 * Ok, everything is just fine now
3742 */
Trela, Maciejdab8b292010-03-08 16:02:45 +11003743 mddev->dev_sectors = conf->dev_sectors;
3744 size = raid10_size(mddev, 0, 0);
3745 md_set_array_sectors(mddev, size);
3746 mddev->resync_max_sectors = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003747
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003748 if (mddev->queue) {
NeilBrown5cf00fc2012-05-21 09:28:20 +10003749 int stripe = conf->geo.raid_disks *
Andre Noll9d8f0362009-06-18 08:45:01 +10003750 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003751
3752 /* Calculate max read-ahead size.
3753 * We need to readahead at least twice a whole stripe....
3754 * maybe...
3755 */
NeilBrown5cf00fc2012-05-21 09:28:20 +10003756 stripe /= conf->geo.near_copies;
NeilBrown3ea7daa2012-05-22 13:53:47 +10003757 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3758 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759 }
3760
Martin K. Petersena91a2782011-03-17 11:11:05 +01003761 if (md_integrity_register(mddev))
3762 goto out_free_conf;
3763
NeilBrown3ea7daa2012-05-22 13:53:47 +10003764 if (conf->reshape_progress != MaxSector) {
3765 unsigned long before_length, after_length;
3766
3767 before_length = ((1 << conf->prev.chunk_shift) *
3768 conf->prev.far_copies);
3769 after_length = ((1 << conf->geo.chunk_shift) *
3770 conf->geo.far_copies);
3771
3772 if (max(before_length, after_length) > min_offset_diff) {
3773 /* This cannot work */
3774 printk("md/raid10: offset difference not enough to continue reshape\n");
3775 goto out_free_conf;
3776 }
3777 conf->offset_diff = min_offset_diff;
3778
NeilBrown3ea7daa2012-05-22 13:53:47 +10003779 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3780 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3781 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3782 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3783 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3784 "reshape");
3785 }
3786
Linus Torvalds1da177e2005-04-16 15:20:36 -07003787 return 0;
3788
3789out_free_conf:
NeilBrown01f96c02011-09-21 15:30:20 +10003790 md_unregister_thread(&mddev->thread);
Julia Lawall644df1a2015-09-13 14:15:10 +02003791 mempool_destroy(conf->r10bio_pool);
NeilBrown1345b1d2006-01-06 00:20:40 -08003792 safe_put_page(conf->tmppage);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07003793 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794 kfree(conf);
3795 mddev->private = NULL;
3796out:
3797 return -EIO;
3798}
3799
NeilBrownafa0f552014-12-15 12:56:58 +11003800static void raid10_free(struct mddev *mddev, void *priv)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801{
NeilBrownafa0f552014-12-15 12:56:58 +11003802 struct r10conf *conf = priv;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803
Julia Lawall644df1a2015-09-13 14:15:10 +02003804 mempool_destroy(conf->r10bio_pool);
Hirokazu Takahashi0fea7ed2013-04-24 11:42:44 +10003805 safe_put_page(conf->tmppage);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07003806 kfree(conf->mirrors);
NeilBrownc4796e22014-08-23 20:19:26 +10003807 kfree(conf->mirrors_old);
3808 kfree(conf->mirrors_new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809 kfree(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003810}
3811
NeilBrownfd01b882011-10-11 16:47:53 +11003812static void raid10_quiesce(struct mddev *mddev, int state)
NeilBrown6cce3b232006-01-06 00:20:16 -08003813{
NeilBrowne879a872011-10-11 16:49:02 +11003814 struct r10conf *conf = mddev->private;
NeilBrown6cce3b232006-01-06 00:20:16 -08003815
3816 switch(state) {
3817 case 1:
3818 raise_barrier(conf, 0);
3819 break;
3820 case 0:
3821 lower_barrier(conf);
3822 break;
3823 }
NeilBrown6cce3b232006-01-06 00:20:16 -08003824}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825
NeilBrown006a09a2012-03-19 12:46:40 +11003826static int raid10_resize(struct mddev *mddev, sector_t sectors)
3827{
3828 /* Resize of 'far' arrays is not supported.
3829 * For 'near' and 'offset' arrays we can set the
3830 * number of sectors used to be an appropriate multiple
3831 * of the chunk size.
3832 * For 'offset', this is far_copies*chunksize.
3833 * For 'near' the multiplier is the LCM of
3834 * near_copies and raid_disks.
3835 * So if far_copies > 1 && !far_offset, fail.
3836 * Else find LCM(raid_disks, near_copy)*far_copies and
3837 * multiply by chunk_size. Then round to this number.
3838 * This is mostly done by raid10_size()
3839 */
3840 struct r10conf *conf = mddev->private;
3841 sector_t oldsize, size;
3842
NeilBrownf8c9e742012-05-21 09:28:33 +10003843 if (mddev->reshape_position != MaxSector)
3844 return -EBUSY;
3845
NeilBrown5cf00fc2012-05-21 09:28:20 +10003846 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
NeilBrown006a09a2012-03-19 12:46:40 +11003847 return -EINVAL;
3848
3849 oldsize = raid10_size(mddev, 0, 0);
3850 size = raid10_size(mddev, sectors, 0);
NeilBrowna4a61252012-05-22 13:55:27 +10003851 if (mddev->external_size &&
3852 mddev->array_sectors > size)
NeilBrown006a09a2012-03-19 12:46:40 +11003853 return -EINVAL;
NeilBrowna4a61252012-05-22 13:55:27 +10003854 if (mddev->bitmap) {
3855 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3856 if (ret)
3857 return ret;
3858 }
3859 md_set_array_sectors(mddev, size);
Heinz Mauelshagen859644f2016-05-03 19:43:24 +02003860 if (mddev->queue) {
3861 set_capacity(mddev->gendisk, mddev->array_sectors);
3862 revalidate_disk(mddev->gendisk);
3863 }
NeilBrown006a09a2012-03-19 12:46:40 +11003864 if (sectors > mddev->dev_sectors &&
3865 mddev->recovery_cp > oldsize) {
3866 mddev->recovery_cp = oldsize;
3867 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3868 }
NeilBrown6508fdb2012-05-17 10:08:45 +10003869 calc_sectors(conf, sectors);
3870 mddev->dev_sectors = conf->dev_sectors;
NeilBrown006a09a2012-03-19 12:46:40 +11003871 mddev->resync_max_sectors = size;
3872 return 0;
3873}
3874
NeilBrown53a6ab42015-02-12 14:09:57 +11003875static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
Trela, Maciejdab8b292010-03-08 16:02:45 +11003876{
NeilBrown3cb03002011-10-11 16:45:26 +11003877 struct md_rdev *rdev;
NeilBrowne879a872011-10-11 16:49:02 +11003878 struct r10conf *conf;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003879
3880 if (mddev->degraded > 0) {
NeilBrown128595e2010-05-03 14:47:14 +10003881 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3882 mdname(mddev));
Trela, Maciejdab8b292010-03-08 16:02:45 +11003883 return ERR_PTR(-EINVAL);
3884 }
NeilBrown53a6ab42015-02-12 14:09:57 +11003885 sector_div(size, devs);
Trela, Maciejdab8b292010-03-08 16:02:45 +11003886
Trela, Maciejdab8b292010-03-08 16:02:45 +11003887 /* Set new parameters */
3888 mddev->new_level = 10;
3889 /* new layout: far_copies = 1, near_copies = 2 */
3890 mddev->new_layout = (1<<8) + 2;
3891 mddev->new_chunk_sectors = mddev->chunk_sectors;
3892 mddev->delta_disks = mddev->raid_disks;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003893 mddev->raid_disks *= 2;
3894 /* make sure it will be not marked as dirty */
3895 mddev->recovery_cp = MaxSector;
NeilBrown53a6ab42015-02-12 14:09:57 +11003896 mddev->dev_sectors = size;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003897
3898 conf = setup_conf(mddev);
Krzysztof Wojcik02214dc2011-02-04 14:18:26 +01003899 if (!IS_ERR(conf)) {
NeilBrowndafb20f2012-03-19 12:46:39 +11003900 rdev_for_each(rdev, mddev)
NeilBrown53a6ab42015-02-12 14:09:57 +11003901 if (rdev->raid_disk >= 0) {
NeilBrowne93f68a2010-06-15 09:36:03 +01003902 rdev->new_raid_disk = rdev->raid_disk * 2;
NeilBrown53a6ab42015-02-12 14:09:57 +11003903 rdev->sectors = size;
3904 }
Krzysztof Wojcik02214dc2011-02-04 14:18:26 +01003905 conf->barrier = 1;
3906 }
3907
Trela, Maciejdab8b292010-03-08 16:02:45 +11003908 return conf;
3909}
3910
NeilBrownfd01b882011-10-11 16:47:53 +11003911static void *raid10_takeover(struct mddev *mddev)
Trela, Maciejdab8b292010-03-08 16:02:45 +11003912{
NeilBrowne373ab12011-10-11 16:48:59 +11003913 struct r0conf *raid0_conf;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003914
3915 /* raid10 can take over:
3916 * raid0 - providing it has only two drives
3917 */
3918 if (mddev->level == 0) {
3919 /* for raid0 takeover only one zone is supported */
NeilBrowne373ab12011-10-11 16:48:59 +11003920 raid0_conf = mddev->private;
3921 if (raid0_conf->nr_strip_zones > 1) {
NeilBrown128595e2010-05-03 14:47:14 +10003922 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3923 " with more than one zone.\n",
3924 mdname(mddev));
Trela, Maciejdab8b292010-03-08 16:02:45 +11003925 return ERR_PTR(-EINVAL);
3926 }
NeilBrown53a6ab42015-02-12 14:09:57 +11003927 return raid10_takeover_raid0(mddev,
3928 raid0_conf->strip_zone->zone_end,
3929 raid0_conf->strip_zone->nb_dev);
Trela, Maciejdab8b292010-03-08 16:02:45 +11003930 }
3931 return ERR_PTR(-EINVAL);
3932}
3933
NeilBrown3ea7daa2012-05-22 13:53:47 +10003934static int raid10_check_reshape(struct mddev *mddev)
3935{
3936 /* Called when there is a request to change
3937 * - layout (to ->new_layout)
3938 * - chunk size (to ->new_chunk_sectors)
3939 * - raid_disks (by delta_disks)
3940 * or when trying to restart a reshape that was ongoing.
3941 *
3942 * We need to validate the request and possibly allocate
3943 * space if that might be an issue later.
3944 *
3945 * Currently we reject any reshape of a 'far' mode array,
3946 * allow chunk size to change if new is generally acceptable,
3947 * allow raid_disks to increase, and allow
3948 * a switch between 'near' mode and 'offset' mode.
3949 */
3950 struct r10conf *conf = mddev->private;
3951 struct geom geo;
3952
3953 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3954 return -EINVAL;
3955
3956 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3957 /* mustn't change number of copies */
3958 return -EINVAL;
3959 if (geo.far_copies > 1 && !geo.far_offset)
3960 /* Cannot switch to 'far' mode */
3961 return -EINVAL;
3962
3963 if (mddev->array_sectors & geo.chunk_mask)
3964 /* not factor of array size */
3965 return -EINVAL;
3966
NeilBrown3ea7daa2012-05-22 13:53:47 +10003967 if (!enough(conf, -1))
3968 return -EINVAL;
3969
3970 kfree(conf->mirrors_new);
3971 conf->mirrors_new = NULL;
3972 if (mddev->delta_disks > 0) {
3973 /* allocate new 'mirrors' list */
3974 conf->mirrors_new = kzalloc(
Jonathan Brassowdc280d982012-07-31 10:03:52 +10003975 sizeof(struct raid10_info)
NeilBrown3ea7daa2012-05-22 13:53:47 +10003976 *(mddev->raid_disks +
3977 mddev->delta_disks),
3978 GFP_KERNEL);
3979 if (!conf->mirrors_new)
3980 return -ENOMEM;
3981 }
3982 return 0;
3983}
3984
3985/*
3986 * Need to check if array has failed when deciding whether to:
3987 * - start an array
3988 * - remove non-faulty devices
3989 * - add a spare
3990 * - allow a reshape
3991 * This determination is simple when no reshape is happening.
3992 * However if there is a reshape, we need to carefully check
3993 * both the before and after sections.
3994 * This is because some failed devices may only affect one
3995 * of the two sections, and some non-in_sync devices may
3996 * be insync in the section most affected by failed devices.
3997 */
3998static int calc_degraded(struct r10conf *conf)
3999{
4000 int degraded, degraded2;
4001 int i;
4002
4003 rcu_read_lock();
4004 degraded = 0;
4005 /* 'prev' section first */
4006 for (i = 0; i < conf->prev.raid_disks; i++) {
4007 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4008 if (!rdev || test_bit(Faulty, &rdev->flags))
4009 degraded++;
4010 else if (!test_bit(In_sync, &rdev->flags))
4011 /* When we can reduce the number of devices in
4012 * an array, this might not contribute to
4013 * 'degraded'. It does now.
4014 */
4015 degraded++;
4016 }
4017 rcu_read_unlock();
4018 if (conf->geo.raid_disks == conf->prev.raid_disks)
4019 return degraded;
4020 rcu_read_lock();
4021 degraded2 = 0;
4022 for (i = 0; i < conf->geo.raid_disks; i++) {
4023 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4024 if (!rdev || test_bit(Faulty, &rdev->flags))
4025 degraded2++;
4026 else if (!test_bit(In_sync, &rdev->flags)) {
4027 /* If reshape is increasing the number of devices,
4028 * this section has already been recovered, so
4029 * it doesn't contribute to degraded.
4030 * else it does.
4031 */
4032 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4033 degraded2++;
4034 }
4035 }
4036 rcu_read_unlock();
4037 if (degraded2 > degraded)
4038 return degraded2;
4039 return degraded;
4040}
4041
4042static int raid10_start_reshape(struct mddev *mddev)
4043{
4044 /* A 'reshape' has been requested. This commits
4045 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4046 * This also checks if there are enough spares and adds them
4047 * to the array.
4048 * We currently require enough spares to make the final
4049 * array non-degraded. We also require that the difference
4050 * between old and new data_offset - on each device - is
4051 * enough that we never risk over-writing.
4052 */
4053
4054 unsigned long before_length, after_length;
4055 sector_t min_offset_diff = 0;
4056 int first = 1;
4057 struct geom new;
4058 struct r10conf *conf = mddev->private;
4059 struct md_rdev *rdev;
4060 int spares = 0;
NeilBrownbb63a702012-05-22 13:55:28 +10004061 int ret;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004062
4063 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4064 return -EBUSY;
4065
4066 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4067 return -EINVAL;
4068
4069 before_length = ((1 << conf->prev.chunk_shift) *
4070 conf->prev.far_copies);
4071 after_length = ((1 << conf->geo.chunk_shift) *
4072 conf->geo.far_copies);
4073
4074 rdev_for_each(rdev, mddev) {
4075 if (!test_bit(In_sync, &rdev->flags)
4076 && !test_bit(Faulty, &rdev->flags))
4077 spares++;
4078 if (rdev->raid_disk >= 0) {
4079 long long diff = (rdev->new_data_offset
4080 - rdev->data_offset);
4081 if (!mddev->reshape_backwards)
4082 diff = -diff;
4083 if (diff < 0)
4084 diff = 0;
4085 if (first || diff < min_offset_diff)
4086 min_offset_diff = diff;
4087 }
4088 }
4089
4090 if (max(before_length, after_length) > min_offset_diff)
4091 return -EINVAL;
4092
4093 if (spares < mddev->delta_disks)
4094 return -EINVAL;
4095
4096 conf->offset_diff = min_offset_diff;
4097 spin_lock_irq(&conf->device_lock);
4098 if (conf->mirrors_new) {
4099 memcpy(conf->mirrors_new, conf->mirrors,
Jonathan Brassowdc280d982012-07-31 10:03:52 +10004100 sizeof(struct raid10_info)*conf->prev.raid_disks);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004101 smp_mb();
NeilBrownc4796e22014-08-23 20:19:26 +10004102 kfree(conf->mirrors_old);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004103 conf->mirrors_old = conf->mirrors;
4104 conf->mirrors = conf->mirrors_new;
4105 conf->mirrors_new = NULL;
4106 }
4107 setup_geo(&conf->geo, mddev, geo_start);
4108 smp_mb();
4109 if (mddev->reshape_backwards) {
4110 sector_t size = raid10_size(mddev, 0, 0);
4111 if (size < mddev->array_sectors) {
4112 spin_unlock_irq(&conf->device_lock);
4113 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4114 mdname(mddev));
4115 return -EINVAL;
4116 }
4117 mddev->resync_max_sectors = size;
4118 conf->reshape_progress = size;
4119 } else
4120 conf->reshape_progress = 0;
NeilBrown299b0682015-07-06 17:37:49 +10004121 conf->reshape_safe = conf->reshape_progress;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004122 spin_unlock_irq(&conf->device_lock);
4123
NeilBrownbb63a702012-05-22 13:55:28 +10004124 if (mddev->delta_disks && mddev->bitmap) {
4125 ret = bitmap_resize(mddev->bitmap,
4126 raid10_size(mddev, 0,
4127 conf->geo.raid_disks),
4128 0, 0);
4129 if (ret)
4130 goto abort;
4131 }
NeilBrown3ea7daa2012-05-22 13:53:47 +10004132 if (mddev->delta_disks > 0) {
4133 rdev_for_each(rdev, mddev)
4134 if (rdev->raid_disk < 0 &&
4135 !test_bit(Faulty, &rdev->flags)) {
4136 if (raid10_add_disk(mddev, rdev) == 0) {
4137 if (rdev->raid_disk >=
4138 conf->prev.raid_disks)
4139 set_bit(In_sync, &rdev->flags);
4140 else
4141 rdev->recovery_offset = 0;
4142
4143 if (sysfs_link_rdev(mddev, rdev))
4144 /* Failure here is OK */;
4145 }
4146 } else if (rdev->raid_disk >= conf->prev.raid_disks
4147 && !test_bit(Faulty, &rdev->flags)) {
4148 /* This is a spare that was manually added */
4149 set_bit(In_sync, &rdev->flags);
4150 }
4151 }
4152 /* When a reshape changes the number of devices,
4153 * ->degraded is measured against the larger of the
4154 * pre and post numbers.
4155 */
4156 spin_lock_irq(&conf->device_lock);
4157 mddev->degraded = calc_degraded(conf);
4158 spin_unlock_irq(&conf->device_lock);
4159 mddev->raid_disks = conf->geo.raid_disks;
4160 mddev->reshape_position = conf->reshape_progress;
4161 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4162
4163 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4164 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
NeilBrownea358cd2015-06-12 20:05:04 +10004165 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004166 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4167 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4168
4169 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4170 "reshape");
4171 if (!mddev->sync_thread) {
NeilBrownbb63a702012-05-22 13:55:28 +10004172 ret = -EAGAIN;
4173 goto abort;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004174 }
4175 conf->reshape_checkpoint = jiffies;
4176 md_wakeup_thread(mddev->sync_thread);
4177 md_new_event(mddev);
4178 return 0;
NeilBrownbb63a702012-05-22 13:55:28 +10004179
4180abort:
4181 mddev->recovery = 0;
4182 spin_lock_irq(&conf->device_lock);
4183 conf->geo = conf->prev;
4184 mddev->raid_disks = conf->geo.raid_disks;
4185 rdev_for_each(rdev, mddev)
4186 rdev->new_data_offset = rdev->data_offset;
4187 smp_wmb();
4188 conf->reshape_progress = MaxSector;
NeilBrown299b0682015-07-06 17:37:49 +10004189 conf->reshape_safe = MaxSector;
NeilBrownbb63a702012-05-22 13:55:28 +10004190 mddev->reshape_position = MaxSector;
4191 spin_unlock_irq(&conf->device_lock);
4192 return ret;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004193}
4194
4195/* Calculate the last device-address that could contain
4196 * any block from the chunk that includes the array-address 's'
4197 * and report the next address.
4198 * i.e. the address returned will be chunk-aligned and after
4199 * any data that is in the chunk containing 's'.
4200 */
4201static sector_t last_dev_address(sector_t s, struct geom *geo)
4202{
4203 s = (s | geo->chunk_mask) + 1;
4204 s >>= geo->chunk_shift;
4205 s *= geo->near_copies;
4206 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4207 s *= geo->far_copies;
4208 s <<= geo->chunk_shift;
4209 return s;
4210}
4211
4212/* Calculate the first device-address that could contain
4213 * any block from the chunk that includes the array-address 's'.
4214 * This too will be the start of a chunk
4215 */
4216static sector_t first_dev_address(sector_t s, struct geom *geo)
4217{
4218 s >>= geo->chunk_shift;
4219 s *= geo->near_copies;
4220 sector_div(s, geo->raid_disks);
4221 s *= geo->far_copies;
4222 s <<= geo->chunk_shift;
4223 return s;
4224}
4225
4226static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4227 int *skipped)
4228{
4229 /* We simply copy at most one chunk (smallest of old and new)
4230 * at a time, possibly less if that exceeds RESYNC_PAGES,
4231 * or we hit a bad block or something.
4232 * This might mean we pause for normal IO in the middle of
NeilBrown02ec5022015-07-06 16:33:47 +10004233 * a chunk, but that is not a problem as mddev->reshape_position
NeilBrown3ea7daa2012-05-22 13:53:47 +10004234 * can record any location.
4235 *
4236 * If we will want to write to a location that isn't
4237 * yet recorded as 'safe' (i.e. in metadata on disk) then
4238 * we need to flush all reshape requests and update the metadata.
4239 *
4240 * When reshaping forwards (e.g. to more devices), we interpret
4241 * 'safe' as the earliest block which might not have been copied
4242 * down yet. We divide this by previous stripe size and multiply
4243 * by previous stripe length to get lowest device offset that we
4244 * cannot write to yet.
4245 * We interpret 'sector_nr' as an address that we want to write to.
4246 * From this we use last_device_address() to find where we might
4247 * write to, and first_device_address on the 'safe' position.
4248 * If this 'next' write position is after the 'safe' position,
4249 * we must update the metadata to increase the 'safe' position.
4250 *
4251 * When reshaping backwards, we round in the opposite direction
4252 * and perform the reverse test: next write position must not be
4253 * less than current safe position.
4254 *
4255 * In all this the minimum difference in data offsets
4256 * (conf->offset_diff - always positive) allows a bit of slack,
NeilBrown02ec5022015-07-06 16:33:47 +10004257 * so next can be after 'safe', but not by more than offset_diff
NeilBrown3ea7daa2012-05-22 13:53:47 +10004258 *
4259 * We need to prepare all the bios here before we start any IO
4260 * to ensure the size we choose is acceptable to all devices.
4261 * The means one for each copy for write-out and an extra one for
4262 * read-in.
4263 * We store the read-in bio in ->master_bio and the others in
4264 * ->devs[x].bio and ->devs[x].repl_bio.
4265 */
4266 struct r10conf *conf = mddev->private;
4267 struct r10bio *r10_bio;
4268 sector_t next, safe, last;
4269 int max_sectors;
4270 int nr_sectors;
4271 int s;
4272 struct md_rdev *rdev;
4273 int need_flush = 0;
4274 struct bio *blist;
4275 struct bio *bio, *read_bio;
4276 int sectors_done = 0;
4277
4278 if (sector_nr == 0) {
4279 /* If restarting in the middle, skip the initial sectors */
4280 if (mddev->reshape_backwards &&
4281 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4282 sector_nr = (raid10_size(mddev, 0, 0)
4283 - conf->reshape_progress);
4284 } else if (!mddev->reshape_backwards &&
4285 conf->reshape_progress > 0)
4286 sector_nr = conf->reshape_progress;
4287 if (sector_nr) {
4288 mddev->curr_resync_completed = sector_nr;
4289 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4290 *skipped = 1;
4291 return sector_nr;
4292 }
4293 }
4294
4295 /* We don't use sector_nr to track where we are up to
4296 * as that doesn't work well for ->reshape_backwards.
4297 * So just use ->reshape_progress.
4298 */
4299 if (mddev->reshape_backwards) {
4300 /* 'next' is the earliest device address that we might
4301 * write to for this chunk in the new layout
4302 */
4303 next = first_dev_address(conf->reshape_progress - 1,
4304 &conf->geo);
4305
4306 /* 'safe' is the last device address that we might read from
4307 * in the old layout after a restart
4308 */
4309 safe = last_dev_address(conf->reshape_safe - 1,
4310 &conf->prev);
4311
4312 if (next + conf->offset_diff < safe)
4313 need_flush = 1;
4314
4315 last = conf->reshape_progress - 1;
4316 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4317 & conf->prev.chunk_mask);
4318 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4319 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4320 } else {
4321 /* 'next' is after the last device address that we
4322 * might write to for this chunk in the new layout
4323 */
4324 next = last_dev_address(conf->reshape_progress, &conf->geo);
4325
4326 /* 'safe' is the earliest device address that we might
4327 * read from in the old layout after a restart
4328 */
4329 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4330
4331 /* Need to update metadata if 'next' might be beyond 'safe'
4332 * as that would possibly corrupt data
4333 */
4334 if (next > safe + conf->offset_diff)
4335 need_flush = 1;
4336
4337 sector_nr = conf->reshape_progress;
4338 last = sector_nr | (conf->geo.chunk_mask
4339 & conf->prev.chunk_mask);
4340
4341 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4342 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4343 }
4344
4345 if (need_flush ||
4346 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4347 /* Need to update reshape_position in metadata */
4348 wait_barrier(conf);
4349 mddev->reshape_position = conf->reshape_progress;
4350 if (mddev->reshape_backwards)
4351 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4352 - conf->reshape_progress;
4353 else
4354 mddev->curr_resync_completed = conf->reshape_progress;
4355 conf->reshape_checkpoint = jiffies;
4356 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4357 md_wakeup_thread(mddev->thread);
4358 wait_event(mddev->sb_wait, mddev->flags == 0 ||
NeilBrownc91abf52013-11-19 12:02:01 +11004359 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4360 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4361 allow_barrier(conf);
4362 return sectors_done;
4363 }
NeilBrown3ea7daa2012-05-22 13:53:47 +10004364 conf->reshape_safe = mddev->reshape_position;
4365 allow_barrier(conf);
4366 }
4367
4368read_more:
4369 /* Now schedule reads for blocks from sector_nr to last */
4370 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
NeilBrowncb8b12b2014-08-18 14:38:45 +10004371 r10_bio->state = 0;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004372 raise_barrier(conf, sectors_done != 0);
4373 atomic_set(&r10_bio->remaining, 0);
4374 r10_bio->mddev = mddev;
4375 r10_bio->sector = sector_nr;
4376 set_bit(R10BIO_IsReshape, &r10_bio->state);
4377 r10_bio->sectors = last - sector_nr + 1;
4378 rdev = read_balance(conf, r10_bio, &max_sectors);
4379 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4380
4381 if (!rdev) {
4382 /* Cannot read from here, so need to record bad blocks
4383 * on all the target devices.
4384 */
4385 // FIXME
NeilBrowne337aea2014-08-18 14:48:54 +10004386 mempool_free(r10_bio, conf->r10buf_pool);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004387 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4388 return sectors_done;
4389 }
4390
4391 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4392
4393 read_bio->bi_bdev = rdev->bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -07004394 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
NeilBrown3ea7daa2012-05-22 13:53:47 +10004395 + rdev->data_offset);
4396 read_bio->bi_private = r10_bio;
4397 read_bio->bi_end_io = end_sync_read;
Mike Christie796a5cf2016-06-05 14:32:07 -05004398 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
NeilBrownce0b0a42014-08-18 13:56:38 +10004399 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02004400 read_bio->bi_error = 0;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004401 read_bio->bi_vcnt = 0;
Kent Overstreet4f024f32013-10-11 15:44:27 -07004402 read_bio->bi_iter.bi_size = 0;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004403 r10_bio->master_bio = read_bio;
4404 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4405
4406 /* Now find the locations in the new layout */
4407 __raid10_find_phys(&conf->geo, r10_bio);
4408
4409 blist = read_bio;
4410 read_bio->bi_next = NULL;
4411
NeilBrownd094d682016-06-02 16:19:52 +10004412 rcu_read_lock();
NeilBrown3ea7daa2012-05-22 13:53:47 +10004413 for (s = 0; s < conf->copies*2; s++) {
4414 struct bio *b;
4415 int d = r10_bio->devs[s/2].devnum;
4416 struct md_rdev *rdev2;
4417 if (s&1) {
NeilBrownd094d682016-06-02 16:19:52 +10004418 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004419 b = r10_bio->devs[s/2].repl_bio;
4420 } else {
NeilBrownd094d682016-06-02 16:19:52 +10004421 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004422 b = r10_bio->devs[s/2].bio;
4423 }
4424 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4425 continue;
Kent Overstreet8be185f2012-09-06 14:14:43 -07004426
4427 bio_reset(b);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004428 b->bi_bdev = rdev2->bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -07004429 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4430 rdev2->new_data_offset;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004431 b->bi_private = r10_bio;
4432 b->bi_end_io = end_reshape_write;
Mike Christie796a5cf2016-06-05 14:32:07 -05004433 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004434 b->bi_next = blist;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004435 blist = b;
4436 }
4437
4438 /* Now add as many pages as possible to all of these bios. */
4439
4440 nr_sectors = 0;
4441 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4442 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4443 int len = (max_sectors - s) << 9;
4444 if (len > PAGE_SIZE)
4445 len = PAGE_SIZE;
4446 for (bio = blist; bio ; bio = bio->bi_next) {
4447 struct bio *bio2;
4448 if (bio_add_page(bio, page, len, 0))
4449 continue;
4450
4451 /* Didn't fit, must stop */
4452 for (bio2 = blist;
4453 bio2 && bio2 != bio;
4454 bio2 = bio2->bi_next) {
4455 /* Remove last page from this bio */
4456 bio2->bi_vcnt--;
Kent Overstreet4f024f32013-10-11 15:44:27 -07004457 bio2->bi_iter.bi_size -= len;
Jens Axboeb7c44ed2015-07-24 12:37:59 -06004458 bio_clear_flag(bio2, BIO_SEG_VALID);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004459 }
4460 goto bio_full;
4461 }
4462 sector_nr += len >> 9;
4463 nr_sectors += len >> 9;
4464 }
4465bio_full:
NeilBrownd094d682016-06-02 16:19:52 +10004466 rcu_read_unlock();
NeilBrown3ea7daa2012-05-22 13:53:47 +10004467 r10_bio->sectors = nr_sectors;
4468
4469 /* Now submit the read */
4470 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4471 atomic_inc(&r10_bio->remaining);
4472 read_bio->bi_next = NULL;
4473 generic_make_request(read_bio);
4474 sector_nr += nr_sectors;
4475 sectors_done += nr_sectors;
4476 if (sector_nr <= last)
4477 goto read_more;
4478
4479 /* Now that we have done the whole section we can
4480 * update reshape_progress
4481 */
4482 if (mddev->reshape_backwards)
4483 conf->reshape_progress -= sectors_done;
4484 else
4485 conf->reshape_progress += sectors_done;
4486
4487 return sectors_done;
4488}
4489
4490static void end_reshape_request(struct r10bio *r10_bio);
4491static int handle_reshape_read_error(struct mddev *mddev,
4492 struct r10bio *r10_bio);
4493static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4494{
4495 /* Reshape read completed. Hopefully we have a block
4496 * to write out.
4497 * If we got a read error then we do sync 1-page reads from
4498 * elsewhere until we find the data - or give up.
4499 */
4500 struct r10conf *conf = mddev->private;
4501 int s;
4502
4503 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4504 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4505 /* Reshape has been aborted */
4506 md_done_sync(mddev, r10_bio->sectors, 0);
4507 return;
4508 }
4509
4510 /* We definitely have the data in the pages, schedule the
4511 * writes.
4512 */
4513 atomic_set(&r10_bio->remaining, 1);
4514 for (s = 0; s < conf->copies*2; s++) {
4515 struct bio *b;
4516 int d = r10_bio->devs[s/2].devnum;
4517 struct md_rdev *rdev;
NeilBrownd094d682016-06-02 16:19:52 +10004518 rcu_read_lock();
NeilBrown3ea7daa2012-05-22 13:53:47 +10004519 if (s&1) {
NeilBrownd094d682016-06-02 16:19:52 +10004520 rdev = rcu_dereference(conf->mirrors[d].replacement);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004521 b = r10_bio->devs[s/2].repl_bio;
4522 } else {
NeilBrownd094d682016-06-02 16:19:52 +10004523 rdev = rcu_dereference(conf->mirrors[d].rdev);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004524 b = r10_bio->devs[s/2].bio;
4525 }
NeilBrownd094d682016-06-02 16:19:52 +10004526 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4527 rcu_read_unlock();
NeilBrown3ea7daa2012-05-22 13:53:47 +10004528 continue;
NeilBrownd094d682016-06-02 16:19:52 +10004529 }
NeilBrown3ea7daa2012-05-22 13:53:47 +10004530 atomic_inc(&rdev->nr_pending);
NeilBrownd094d682016-06-02 16:19:52 +10004531 rcu_read_unlock();
NeilBrown3ea7daa2012-05-22 13:53:47 +10004532 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4533 atomic_inc(&r10_bio->remaining);
4534 b->bi_next = NULL;
4535 generic_make_request(b);
4536 }
4537 end_reshape_request(r10_bio);
4538}
4539
4540static void end_reshape(struct r10conf *conf)
4541{
4542 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4543 return;
4544
4545 spin_lock_irq(&conf->device_lock);
4546 conf->prev = conf->geo;
4547 md_finish_reshape(conf->mddev);
4548 smp_wmb();
4549 conf->reshape_progress = MaxSector;
NeilBrown299b0682015-07-06 17:37:49 +10004550 conf->reshape_safe = MaxSector;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004551 spin_unlock_irq(&conf->device_lock);
4552
4553 /* read-ahead size must cover two whole stripes, which is
4554 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4555 */
4556 if (conf->mddev->queue) {
4557 int stripe = conf->geo.raid_disks *
4558 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4559 stripe /= conf->geo.near_copies;
4560 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4561 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4562 }
4563 conf->fullsync = 0;
4564}
4565
NeilBrown3ea7daa2012-05-22 13:53:47 +10004566static int handle_reshape_read_error(struct mddev *mddev,
4567 struct r10bio *r10_bio)
4568{
4569 /* Use sync reads to get the blocks from somewhere else */
4570 int sectors = r10_bio->sectors;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004571 struct r10conf *conf = mddev->private;
NeilBrowne0ee7782012-08-18 09:51:42 +10004572 struct {
4573 struct r10bio r10_bio;
4574 struct r10dev devs[conf->copies];
4575 } on_stack;
4576 struct r10bio *r10b = &on_stack.r10_bio;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004577 int slot = 0;
4578 int idx = 0;
4579 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4580
NeilBrowne0ee7782012-08-18 09:51:42 +10004581 r10b->sector = r10_bio->sector;
4582 __raid10_find_phys(&conf->prev, r10b);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004583
4584 while (sectors) {
4585 int s = sectors;
4586 int success = 0;
4587 int first_slot = slot;
4588
4589 if (s > (PAGE_SIZE >> 9))
4590 s = PAGE_SIZE >> 9;
4591
NeilBrownd094d682016-06-02 16:19:52 +10004592 rcu_read_lock();
NeilBrown3ea7daa2012-05-22 13:53:47 +10004593 while (!success) {
NeilBrowne0ee7782012-08-18 09:51:42 +10004594 int d = r10b->devs[slot].devnum;
NeilBrownd094d682016-06-02 16:19:52 +10004595 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004596 sector_t addr;
4597 if (rdev == NULL ||
4598 test_bit(Faulty, &rdev->flags) ||
4599 !test_bit(In_sync, &rdev->flags))
4600 goto failed;
4601
NeilBrowne0ee7782012-08-18 09:51:42 +10004602 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
NeilBrownd094d682016-06-02 16:19:52 +10004603 atomic_inc(&rdev->nr_pending);
4604 rcu_read_unlock();
NeilBrown3ea7daa2012-05-22 13:53:47 +10004605 success = sync_page_io(rdev,
4606 addr,
4607 s << 9,
4608 bvec[idx].bv_page,
Mike Christie796a5cf2016-06-05 14:32:07 -05004609 REQ_OP_READ, 0, false);
NeilBrownd094d682016-06-02 16:19:52 +10004610 rdev_dec_pending(rdev, mddev);
4611 rcu_read_lock();
NeilBrown3ea7daa2012-05-22 13:53:47 +10004612 if (success)
4613 break;
4614 failed:
4615 slot++;
4616 if (slot >= conf->copies)
4617 slot = 0;
4618 if (slot == first_slot)
4619 break;
4620 }
NeilBrownd094d682016-06-02 16:19:52 +10004621 rcu_read_unlock();
NeilBrown3ea7daa2012-05-22 13:53:47 +10004622 if (!success) {
4623 /* couldn't read this block, must give up */
4624 set_bit(MD_RECOVERY_INTR,
4625 &mddev->recovery);
4626 return -EIO;
4627 }
4628 sectors -= s;
4629 idx++;
4630 }
4631 return 0;
4632}
4633
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02004634static void end_reshape_write(struct bio *bio)
NeilBrown3ea7daa2012-05-22 13:53:47 +10004635{
NeilBrown3ea7daa2012-05-22 13:53:47 +10004636 struct r10bio *r10_bio = bio->bi_private;
4637 struct mddev *mddev = r10_bio->mddev;
4638 struct r10conf *conf = mddev->private;
4639 int d;
4640 int slot;
4641 int repl;
4642 struct md_rdev *rdev = NULL;
4643
4644 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4645 if (repl)
4646 rdev = conf->mirrors[d].replacement;
4647 if (!rdev) {
4648 smp_mb();
4649 rdev = conf->mirrors[d].rdev;
4650 }
4651
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02004652 if (bio->bi_error) {
NeilBrown3ea7daa2012-05-22 13:53:47 +10004653 /* FIXME should record badblock */
4654 md_error(mddev, rdev);
4655 }
4656
4657 rdev_dec_pending(rdev, mddev);
4658 end_reshape_request(r10_bio);
4659}
4660
4661static void end_reshape_request(struct r10bio *r10_bio)
4662{
4663 if (!atomic_dec_and_test(&r10_bio->remaining))
4664 return;
4665 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4666 bio_put(r10_bio->master_bio);
4667 put_buf(r10_bio);
4668}
4669
4670static void raid10_finish_reshape(struct mddev *mddev)
4671{
4672 struct r10conf *conf = mddev->private;
4673
4674 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4675 return;
4676
4677 if (mddev->delta_disks > 0) {
4678 sector_t size = raid10_size(mddev, 0, 0);
4679 md_set_array_sectors(mddev, size);
4680 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4681 mddev->recovery_cp = mddev->resync_max_sectors;
4682 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4683 }
4684 mddev->resync_max_sectors = size;
Heinz Mauelshagen859644f2016-05-03 19:43:24 +02004685 if (mddev->queue) {
4686 set_capacity(mddev->gendisk, mddev->array_sectors);
4687 revalidate_disk(mddev->gendisk);
4688 }
NeilBrown63aced62012-05-22 13:55:33 +10004689 } else {
4690 int d;
NeilBrownd094d682016-06-02 16:19:52 +10004691 rcu_read_lock();
NeilBrown63aced62012-05-22 13:55:33 +10004692 for (d = conf->geo.raid_disks ;
4693 d < conf->geo.raid_disks - mddev->delta_disks;
4694 d++) {
NeilBrownd094d682016-06-02 16:19:52 +10004695 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
NeilBrown63aced62012-05-22 13:55:33 +10004696 if (rdev)
4697 clear_bit(In_sync, &rdev->flags);
NeilBrownd094d682016-06-02 16:19:52 +10004698 rdev = rcu_dereference(conf->mirrors[d].replacement);
NeilBrown63aced62012-05-22 13:55:33 +10004699 if (rdev)
4700 clear_bit(In_sync, &rdev->flags);
4701 }
NeilBrownd094d682016-06-02 16:19:52 +10004702 rcu_read_unlock();
NeilBrown3ea7daa2012-05-22 13:53:47 +10004703 }
4704 mddev->layout = mddev->new_layout;
4705 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4706 mddev->reshape_position = MaxSector;
4707 mddev->delta_disks = 0;
4708 mddev->reshape_backwards = 0;
4709}
4710
NeilBrown84fc4b52011-10-11 16:49:58 +11004711static struct md_personality raid10_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07004712{
4713 .name = "raid10",
NeilBrown2604b702006-01-06 00:20:36 -08004714 .level = 10,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004715 .owner = THIS_MODULE,
Shaohua Li849674e2016-01-20 13:52:20 -08004716 .make_request = raid10_make_request,
4717 .run = raid10_run,
NeilBrownafa0f552014-12-15 12:56:58 +11004718 .free = raid10_free,
Shaohua Li849674e2016-01-20 13:52:20 -08004719 .status = raid10_status,
4720 .error_handler = raid10_error,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004721 .hot_add_disk = raid10_add_disk,
4722 .hot_remove_disk= raid10_remove_disk,
4723 .spare_active = raid10_spare_active,
Shaohua Li849674e2016-01-20 13:52:20 -08004724 .sync_request = raid10_sync_request,
NeilBrown6cce3b232006-01-06 00:20:16 -08004725 .quiesce = raid10_quiesce,
Dan Williams80c3a6c2009-03-17 18:10:40 -07004726 .size = raid10_size,
NeilBrown006a09a2012-03-19 12:46:40 +11004727 .resize = raid10_resize,
Trela, Maciejdab8b292010-03-08 16:02:45 +11004728 .takeover = raid10_takeover,
NeilBrown3ea7daa2012-05-22 13:53:47 +10004729 .check_reshape = raid10_check_reshape,
4730 .start_reshape = raid10_start_reshape,
4731 .finish_reshape = raid10_finish_reshape,
NeilBrown5c675f82014-12-15 12:56:56 +11004732 .congested = raid10_congested,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004733};
4734
4735static int __init raid_init(void)
4736{
NeilBrown2604b702006-01-06 00:20:36 -08004737 return register_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004738}
4739
4740static void raid_exit(void)
4741{
NeilBrown2604b702006-01-06 00:20:36 -08004742 unregister_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004743}
4744
4745module_init(raid_init);
4746module_exit(raid_exit);
4747MODULE_LICENSE("GPL");
NeilBrown0efb9e62009-12-14 12:49:58 +11004748MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004749MODULE_ALIAS("md-personality-9"); /* RAID10 */
NeilBrownd9d166c2006-01-06 00:20:51 -08004750MODULE_ALIAS("md-raid10");
NeilBrown2604b702006-01-06 00:20:36 -08004751MODULE_ALIAS("md-level-10");
NeilBrown34db0cd2011-10-11 16:50:01 +11004752
4753module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);