blob: e75d48c0421a41788c9159ef7e74d22ad93d9695 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
NeilBrown16a53ec2006-06-26 00:27:38 -07005 * Copyright (C) 2002, 2003 H. Peter Anvin
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
NeilBrown16a53ec2006-06-26 00:27:38 -07007 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
Linus Torvalds1da177e2005-04-16 15:20:36 -070010 *
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
NeilBrownae3c20c2006-07-10 04:44:17 -070021/*
22 * BITMAP UNPLUGGING:
23 *
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
26 * explanation.
27 *
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
NeilBrown7c13edc2011-04-18 18:25:43 +100030 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
NeilBrownae3c20c2006-07-10 04:44:17 -070032 * new additions.
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
NeilBrown7c13edc2011-04-18 18:25:43 +100035 * the number of the batch it will be in. This is seq_flush+1.
NeilBrownae3c20c2006-07-10 04:44:17 -070036 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
39 * batch.
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
43 * miss any bits.
44 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
NeilBrownbff61972009-03-31 14:33:13 +110046#include <linux/blkdev.h>
NeilBrownf6705572006-03-27 01:18:11 -080047#include <linux/kthread.h>
Dan Williamsf701d582009-03-31 15:09:39 +110048#include <linux/raid/pq.h>
Dan Williams91c00922007-01-02 13:52:30 -070049#include <linux/async_tx.h>
Paul Gortmaker056075c2011-07-03 13:58:33 -040050#include <linux/module.h>
Dan Williams07a3b412009-08-29 19:13:13 -070051#include <linux/async.h>
NeilBrownbff61972009-03-31 14:33:13 +110052#include <linux/seq_file.h>
Dan Williams36d1c642009-07-14 11:48:22 -070053#include <linux/cpu.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090054#include <linux/slab.h>
Christian Dietrich8bda4702011-07-27 11:00:36 +100055#include <linux/ratelimit.h>
Shaohua Li851c30c2013-08-28 14:30:16 +080056#include <linux/nodemask.h>
NeilBrowna9add5d2012-10-31 11:59:09 +110057#include <trace/events/block.h>
58
NeilBrown43b2e5d2009-03-31 14:33:13 +110059#include "md.h"
NeilBrownbff61972009-03-31 14:33:13 +110060#include "raid5.h"
Trela Maciej54071b32010-03-08 16:02:42 +110061#include "raid0.h"
Christoph Hellwigef740c32009-03-31 14:27:03 +110062#include "bitmap.h"
NeilBrown72626682005-09-09 16:23:54 -070063
Shaohua Li851c30c2013-08-28 14:30:16 +080064#define cpu_to_group(cpu) cpu_to_node(cpu)
65#define ANY_GROUP NUMA_NO_NODE
66
NeilBrown8e0e99b2014-10-02 13:45:00 +100067static bool devices_handle_discard_safely = false;
68module_param(devices_handle_discard_safely, bool, 0644);
69MODULE_PARM_DESC(devices_handle_discard_safely,
70 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
Shaohua Li851c30c2013-08-28 14:30:16 +080071static struct workqueue_struct *raid5_wq;
Linus Torvalds1da177e2005-04-16 15:20:36 -070072/*
73 * Stripe cache
74 */
75
76#define NR_STRIPES 256
77#define STRIPE_SIZE PAGE_SIZE
78#define STRIPE_SHIFT (PAGE_SHIFT - 9)
79#define STRIPE_SECTORS (STRIPE_SIZE>>9)
80#define IO_THRESHOLD 1
Dan Williams8b3e6cd2008-04-28 02:15:53 -070081#define BYPASS_THRESHOLD 1
NeilBrownfccddba2006-01-06 00:20:33 -080082#define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
Linus Torvalds1da177e2005-04-16 15:20:36 -070083#define HASH_MASK (NR_HASH - 1)
Shaohua Libfc90cb2013-08-29 15:40:32 +080084#define MAX_STRIPE_BATCH 8
Linus Torvalds1da177e2005-04-16 15:20:36 -070085
NeilBrownd1688a62011-10-11 16:49:52 +110086static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
NeilBrowndb298e12011-10-07 14:23:00 +110087{
88 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
89 return &conf->stripe_hashtbl[hash];
90}
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
Shaohua Li566c09c2013-11-14 15:16:17 +110092static inline int stripe_hash_locks_hash(sector_t sect)
93{
94 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
95}
96
97static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
98{
99 spin_lock_irq(conf->hash_locks + hash);
100 spin_lock(&conf->device_lock);
101}
102
103static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
104{
105 spin_unlock(&conf->device_lock);
106 spin_unlock_irq(conf->hash_locks + hash);
107}
108
109static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
110{
111 int i;
112 local_irq_disable();
113 spin_lock(conf->hash_locks);
114 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
115 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
116 spin_lock(&conf->device_lock);
117}
118
119static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
120{
121 int i;
122 spin_unlock(&conf->device_lock);
123 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
124 spin_unlock(conf->hash_locks + i - 1);
125 local_irq_enable();
126}
127
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128/* bio's attached to a stripe+device for I/O are linked together in bi_sector
129 * order without overlap. There may be several bio's per stripe+device, and
130 * a bio could span several devices.
131 * When walking this list for a particular stripe+device, we must never proceed
132 * beyond a bio that extends past this device, as the next bio might no longer
133 * be valid.
NeilBrowndb298e12011-10-07 14:23:00 +1100134 * This function is used to determine the 'next' bio in the list, given the sector
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 * of the current stripe+device
136 */
NeilBrowndb298e12011-10-07 14:23:00 +1100137static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
138{
Kent Overstreetaa8b57a2013-02-05 15:19:29 -0800139 int sectors = bio_sectors(bio);
Kent Overstreet4f024f32013-10-11 15:44:27 -0700140 if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
NeilBrowndb298e12011-10-07 14:23:00 +1100141 return bio->bi_next;
142 else
143 return NULL;
144}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
Jens Axboe960e7392008-08-15 10:41:18 +0200146/*
Jens Axboe5b99c2f2008-08-15 10:56:11 +0200147 * We maintain a biased count of active stripes in the bottom 16 bits of
148 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
Jens Axboe960e7392008-08-15 10:41:18 +0200149 */
Shaohua Lie7836bd62012-07-19 16:01:31 +1000150static inline int raid5_bi_processed_stripes(struct bio *bio)
Jens Axboe960e7392008-08-15 10:41:18 +0200151{
Shaohua Lie7836bd62012-07-19 16:01:31 +1000152 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
153 return (atomic_read(segments) >> 16) & 0xffff;
Jens Axboe960e7392008-08-15 10:41:18 +0200154}
155
Shaohua Lie7836bd62012-07-19 16:01:31 +1000156static inline int raid5_dec_bi_active_stripes(struct bio *bio)
Jens Axboe960e7392008-08-15 10:41:18 +0200157{
Shaohua Lie7836bd62012-07-19 16:01:31 +1000158 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
159 return atomic_sub_return(1, segments) & 0xffff;
Jens Axboe960e7392008-08-15 10:41:18 +0200160}
161
Shaohua Lie7836bd62012-07-19 16:01:31 +1000162static inline void raid5_inc_bi_active_stripes(struct bio *bio)
Jens Axboe960e7392008-08-15 10:41:18 +0200163{
Shaohua Lie7836bd62012-07-19 16:01:31 +1000164 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
165 atomic_inc(segments);
Jens Axboe960e7392008-08-15 10:41:18 +0200166}
167
Shaohua Lie7836bd62012-07-19 16:01:31 +1000168static inline void raid5_set_bi_processed_stripes(struct bio *bio,
169 unsigned int cnt)
Jens Axboe960e7392008-08-15 10:41:18 +0200170{
Shaohua Lie7836bd62012-07-19 16:01:31 +1000171 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
172 int old, new;
Jens Axboe960e7392008-08-15 10:41:18 +0200173
Shaohua Lie7836bd62012-07-19 16:01:31 +1000174 do {
175 old = atomic_read(segments);
176 new = (old & 0xffff) | (cnt << 16);
177 } while (atomic_cmpxchg(segments, old, new) != old);
Jens Axboe960e7392008-08-15 10:41:18 +0200178}
179
Shaohua Lie7836bd62012-07-19 16:01:31 +1000180static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
Jens Axboe960e7392008-08-15 10:41:18 +0200181{
Shaohua Lie7836bd62012-07-19 16:01:31 +1000182 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
183 atomic_set(segments, cnt);
Jens Axboe960e7392008-08-15 10:41:18 +0200184}
185
NeilBrownd0dabf72009-03-31 14:39:38 +1100186/* Find first data disk in a raid6 stripe */
187static inline int raid6_d0(struct stripe_head *sh)
188{
NeilBrown67cc2b82009-03-31 14:39:38 +1100189 if (sh->ddf_layout)
190 /* ddf always start from first device */
191 return 0;
192 /* md starts just after Q block */
NeilBrownd0dabf72009-03-31 14:39:38 +1100193 if (sh->qd_idx == sh->disks - 1)
194 return 0;
195 else
196 return sh->qd_idx + 1;
197}
NeilBrown16a53ec2006-06-26 00:27:38 -0700198static inline int raid6_next_disk(int disk, int raid_disks)
199{
200 disk++;
201 return (disk < raid_disks) ? disk : 0;
202}
Dan Williamsa4456852007-07-09 11:56:43 -0700203
NeilBrownd0dabf72009-03-31 14:39:38 +1100204/* When walking through the disks in a raid5, starting at raid6_d0,
205 * We need to map each disk to a 'slot', where the data disks are slot
206 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
207 * is raid_disks-1. This help does that mapping.
208 */
NeilBrown67cc2b82009-03-31 14:39:38 +1100209static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
210 int *count, int syndrome_disks)
NeilBrownd0dabf72009-03-31 14:39:38 +1100211{
Dan Williams66295422009-10-19 18:09:32 -0700212 int slot = *count;
NeilBrown67cc2b82009-03-31 14:39:38 +1100213
NeilBrowne4424fe2009-10-16 16:27:34 +1100214 if (sh->ddf_layout)
Dan Williams66295422009-10-19 18:09:32 -0700215 (*count)++;
NeilBrownd0dabf72009-03-31 14:39:38 +1100216 if (idx == sh->pd_idx)
NeilBrown67cc2b82009-03-31 14:39:38 +1100217 return syndrome_disks;
NeilBrownd0dabf72009-03-31 14:39:38 +1100218 if (idx == sh->qd_idx)
NeilBrown67cc2b82009-03-31 14:39:38 +1100219 return syndrome_disks + 1;
NeilBrowne4424fe2009-10-16 16:27:34 +1100220 if (!sh->ddf_layout)
Dan Williams66295422009-10-19 18:09:32 -0700221 (*count)++;
NeilBrownd0dabf72009-03-31 14:39:38 +1100222 return slot;
223}
224
Dan Williamsa4456852007-07-09 11:56:43 -0700225static void return_io(struct bio *return_bi)
226{
227 struct bio *bi = return_bi;
228 while (bi) {
Dan Williamsa4456852007-07-09 11:56:43 -0700229
230 return_bi = bi->bi_next;
231 bi->bi_next = NULL;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700232 bi->bi_iter.bi_size = 0;
Linus Torvalds0a82a8d2013-04-18 09:00:26 -0700233 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
234 bi, 0);
Neil Brown0e13fe232008-06-28 08:31:20 +1000235 bio_endio(bi, 0);
Dan Williamsa4456852007-07-09 11:56:43 -0700236 bi = return_bi;
237 }
238}
239
NeilBrownd1688a62011-10-11 16:49:52 +1100240static void print_raid5_conf (struct r5conf *conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241
Dan Williams600aa102008-06-28 08:32:05 +1000242static int stripe_operations_active(struct stripe_head *sh)
243{
244 return sh->check_state || sh->reconstruct_state ||
245 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
246 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
247}
248
Shaohua Li851c30c2013-08-28 14:30:16 +0800249static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
250{
251 struct r5conf *conf = sh->raid_conf;
252 struct r5worker_group *group;
Shaohua Libfc90cb2013-08-29 15:40:32 +0800253 int thread_cnt;
Shaohua Li851c30c2013-08-28 14:30:16 +0800254 int i, cpu = sh->cpu;
255
256 if (!cpu_online(cpu)) {
257 cpu = cpumask_any(cpu_online_mask);
258 sh->cpu = cpu;
259 }
260
261 if (list_empty(&sh->lru)) {
262 struct r5worker_group *group;
263 group = conf->worker_groups + cpu_to_group(cpu);
264 list_add_tail(&sh->lru, &group->handle_list);
Shaohua Libfc90cb2013-08-29 15:40:32 +0800265 group->stripes_cnt++;
266 sh->group = group;
Shaohua Li851c30c2013-08-28 14:30:16 +0800267 }
268
269 if (conf->worker_cnt_per_group == 0) {
270 md_wakeup_thread(conf->mddev->thread);
271 return;
272 }
273
274 group = conf->worker_groups + cpu_to_group(sh->cpu);
275
Shaohua Libfc90cb2013-08-29 15:40:32 +0800276 group->workers[0].working = true;
277 /* at least one worker should run to avoid race */
278 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
279
280 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
281 /* wakeup more workers */
282 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
283 if (group->workers[i].working == false) {
284 group->workers[i].working = true;
285 queue_work_on(sh->cpu, raid5_wq,
286 &group->workers[i].work);
287 thread_cnt--;
288 }
289 }
Shaohua Li851c30c2013-08-28 14:30:16 +0800290}
291
Shaohua Li566c09c2013-11-14 15:16:17 +1100292static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
293 struct list_head *temp_inactive_list)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294{
Shaohua Li4eb788d2012-07-19 16:01:31 +1000295 BUG_ON(!list_empty(&sh->lru));
296 BUG_ON(atomic_read(&conf->active_stripes)==0);
297 if (test_bit(STRIPE_HANDLE, &sh->state)) {
298 if (test_bit(STRIPE_DELAYED, &sh->state) &&
Jes Sorensenad3ab8b2015-01-29 12:38:29 -0500299 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
Shaohua Li4eb788d2012-07-19 16:01:31 +1000300 list_add_tail(&sh->lru, &conf->delayed_list);
Jes Sorensenad3ab8b2015-01-29 12:38:29 -0500301 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
Shaohua Li4eb788d2012-07-19 16:01:31 +1000302 sh->bm_seq - conf->seq_write > 0)
303 list_add_tail(&sh->lru, &conf->bitmap_list);
304 else {
305 clear_bit(STRIPE_DELAYED, &sh->state);
306 clear_bit(STRIPE_BIT_DELAY, &sh->state);
Shaohua Li851c30c2013-08-28 14:30:16 +0800307 if (conf->worker_cnt_per_group == 0) {
308 list_add_tail(&sh->lru, &conf->handle_list);
309 } else {
310 raid5_wakeup_stripe_thread(sh);
311 return;
312 }
Shaohua Li4eb788d2012-07-19 16:01:31 +1000313 }
314 md_wakeup_thread(conf->mddev->thread);
315 } else {
316 BUG_ON(stripe_operations_active(sh));
317 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
318 if (atomic_dec_return(&conf->preread_active_stripes)
319 < IO_THRESHOLD)
320 md_wakeup_thread(conf->mddev->thread);
321 atomic_dec(&conf->active_stripes);
Shaohua Li566c09c2013-11-14 15:16:17 +1100322 if (!test_bit(STRIPE_EXPANDING, &sh->state))
323 list_add_tail(&sh->lru, temp_inactive_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324 }
325}
NeilBrownd0dabf72009-03-31 14:39:38 +1100326
Shaohua Li566c09c2013-11-14 15:16:17 +1100327static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
328 struct list_head *temp_inactive_list)
Shaohua Li4eb788d2012-07-19 16:01:31 +1000329{
330 if (atomic_dec_and_test(&sh->count))
Shaohua Li566c09c2013-11-14 15:16:17 +1100331 do_release_stripe(conf, sh, temp_inactive_list);
332}
333
334/*
335 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
336 *
337 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
338 * given time. Adding stripes only takes device lock, while deleting stripes
339 * only takes hash lock.
340 */
341static void release_inactive_stripe_list(struct r5conf *conf,
342 struct list_head *temp_inactive_list,
343 int hash)
344{
345 int size;
346 bool do_wakeup = false;
347 unsigned long flags;
348
349 if (hash == NR_STRIPE_HASH_LOCKS) {
350 size = NR_STRIPE_HASH_LOCKS;
351 hash = NR_STRIPE_HASH_LOCKS - 1;
352 } else
353 size = 1;
354 while (size) {
355 struct list_head *list = &temp_inactive_list[size - 1];
356
357 /*
358 * We don't hold any lock here yet, get_active_stripe() might
359 * remove stripes from the list
360 */
361 if (!list_empty_careful(list)) {
362 spin_lock_irqsave(conf->hash_locks + hash, flags);
Shaohua Li4bda5562013-11-14 15:16:17 +1100363 if (list_empty(conf->inactive_list + hash) &&
364 !list_empty(list))
365 atomic_dec(&conf->empty_inactive_list_nr);
Shaohua Li566c09c2013-11-14 15:16:17 +1100366 list_splice_tail_init(list, conf->inactive_list + hash);
367 do_wakeup = true;
368 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
369 }
370 size--;
371 hash--;
372 }
373
374 if (do_wakeup) {
375 wake_up(&conf->wait_for_stripe);
376 if (conf->retry_read_aligned)
377 md_wakeup_thread(conf->mddev->thread);
378 }
Shaohua Li4eb788d2012-07-19 16:01:31 +1000379}
380
Shaohua Li773ca822013-08-27 17:50:39 +0800381/* should hold conf->device_lock already */
Shaohua Li566c09c2013-11-14 15:16:17 +1100382static int release_stripe_list(struct r5conf *conf,
383 struct list_head *temp_inactive_list)
Shaohua Li773ca822013-08-27 17:50:39 +0800384{
385 struct stripe_head *sh;
386 int count = 0;
387 struct llist_node *head;
388
389 head = llist_del_all(&conf->released_stripes);
Shaohua Lid265d9d2013-08-28 14:29:05 +0800390 head = llist_reverse_order(head);
Shaohua Li773ca822013-08-27 17:50:39 +0800391 while (head) {
Shaohua Li566c09c2013-11-14 15:16:17 +1100392 int hash;
393
Shaohua Li773ca822013-08-27 17:50:39 +0800394 sh = llist_entry(head, struct stripe_head, release_list);
395 head = llist_next(head);
396 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
397 smp_mb();
398 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
399 /*
400 * Don't worry the bit is set here, because if the bit is set
401 * again, the count is always > 1. This is true for
402 * STRIPE_ON_UNPLUG_LIST bit too.
403 */
Shaohua Li566c09c2013-11-14 15:16:17 +1100404 hash = sh->hash_lock_index;
405 __release_stripe(conf, sh, &temp_inactive_list[hash]);
Shaohua Li773ca822013-08-27 17:50:39 +0800406 count++;
407 }
408
409 return count;
410}
411
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412static void release_stripe(struct stripe_head *sh)
413{
NeilBrownd1688a62011-10-11 16:49:52 +1100414 struct r5conf *conf = sh->raid_conf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 unsigned long flags;
Shaohua Li566c09c2013-11-14 15:16:17 +1100416 struct list_head list;
417 int hash;
Shaohua Li773ca822013-08-27 17:50:39 +0800418 bool wakeup;
NeilBrown16a53ec2006-06-26 00:27:38 -0700419
Eivind Sartocf170f32014-05-28 13:39:23 +1000420 /* Avoid release_list until the last reference.
421 */
422 if (atomic_add_unless(&sh->count, -1, 1))
423 return;
424
majianpengad4068d2013-11-14 15:16:15 +1100425 if (unlikely(!conf->mddev->thread) ||
426 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
Shaohua Li773ca822013-08-27 17:50:39 +0800427 goto slow_path;
428 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
429 if (wakeup)
430 md_wakeup_thread(conf->mddev->thread);
431 return;
432slow_path:
Shaohua Li4eb788d2012-07-19 16:01:31 +1000433 local_irq_save(flags);
Shaohua Li773ca822013-08-27 17:50:39 +0800434 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
Shaohua Li4eb788d2012-07-19 16:01:31 +1000435 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
Shaohua Li566c09c2013-11-14 15:16:17 +1100436 INIT_LIST_HEAD(&list);
437 hash = sh->hash_lock_index;
438 do_release_stripe(conf, sh, &list);
Shaohua Li4eb788d2012-07-19 16:01:31 +1000439 spin_unlock(&conf->device_lock);
Shaohua Li566c09c2013-11-14 15:16:17 +1100440 release_inactive_stripe_list(conf, &list, hash);
Shaohua Li4eb788d2012-07-19 16:01:31 +1000441 }
442 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443}
444
NeilBrownfccddba2006-01-06 00:20:33 -0800445static inline void remove_hash(struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446{
Dan Williams45b42332007-07-09 11:56:43 -0700447 pr_debug("remove_hash(), stripe %llu\n",
448 (unsigned long long)sh->sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449
NeilBrownfccddba2006-01-06 00:20:33 -0800450 hlist_del_init(&sh->hash);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451}
452
NeilBrownd1688a62011-10-11 16:49:52 +1100453static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454{
NeilBrownfccddba2006-01-06 00:20:33 -0800455 struct hlist_head *hp = stripe_hash(conf, sh->sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456
Dan Williams45b42332007-07-09 11:56:43 -0700457 pr_debug("insert_hash(), stripe %llu\n",
458 (unsigned long long)sh->sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459
NeilBrownfccddba2006-01-06 00:20:33 -0800460 hlist_add_head(&sh->hash, hp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461}
462
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463/* find an idle stripe, make sure it is unhashed, and return it. */
Shaohua Li566c09c2013-11-14 15:16:17 +1100464static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465{
466 struct stripe_head *sh = NULL;
467 struct list_head *first;
468
Shaohua Li566c09c2013-11-14 15:16:17 +1100469 if (list_empty(conf->inactive_list + hash))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 goto out;
Shaohua Li566c09c2013-11-14 15:16:17 +1100471 first = (conf->inactive_list + hash)->next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 sh = list_entry(first, struct stripe_head, lru);
473 list_del_init(first);
474 remove_hash(sh);
475 atomic_inc(&conf->active_stripes);
Shaohua Li566c09c2013-11-14 15:16:17 +1100476 BUG_ON(hash != sh->hash_lock_index);
Shaohua Li4bda5562013-11-14 15:16:17 +1100477 if (list_empty(conf->inactive_list + hash))
478 atomic_inc(&conf->empty_inactive_list_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479out:
480 return sh;
481}
482
NeilBrowne4e11e32010-06-16 16:45:16 +1000483static void shrink_buffers(struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484{
485 struct page *p;
486 int i;
NeilBrowne4e11e32010-06-16 16:45:16 +1000487 int num = sh->raid_conf->pool_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488
NeilBrowne4e11e32010-06-16 16:45:16 +1000489 for (i = 0; i < num ; i++) {
Shaohua Lid592a992014-05-21 17:57:44 +0800490 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 p = sh->dev[i].page;
492 if (!p)
493 continue;
494 sh->dev[i].page = NULL;
NeilBrown2d1f3b52006-01-06 00:20:31 -0800495 put_page(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496 }
497}
498
NeilBrowne4e11e32010-06-16 16:45:16 +1000499static int grow_buffers(struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500{
501 int i;
NeilBrowne4e11e32010-06-16 16:45:16 +1000502 int num = sh->raid_conf->pool_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503
NeilBrowne4e11e32010-06-16 16:45:16 +1000504 for (i = 0; i < num; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 struct page *page;
506
507 if (!(page = alloc_page(GFP_KERNEL))) {
508 return 1;
509 }
510 sh->dev[i].page = page;
Shaohua Lid592a992014-05-21 17:57:44 +0800511 sh->dev[i].orig_page = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 }
513 return 0;
514}
515
NeilBrown784052e2009-03-31 15:19:07 +1100516static void raid5_build_block(struct stripe_head *sh, int i, int previous);
NeilBrownd1688a62011-10-11 16:49:52 +1100517static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
NeilBrown911d4ee2009-03-31 14:39:38 +1100518 struct stripe_head *sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519
NeilBrownb5663ba2009-03-31 14:39:38 +1100520static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521{
NeilBrownd1688a62011-10-11 16:49:52 +1100522 struct r5conf *conf = sh->raid_conf;
Shaohua Li566c09c2013-11-14 15:16:17 +1100523 int i, seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +0200525 BUG_ON(atomic_read(&sh->count) != 0);
526 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
Dan Williams600aa102008-06-28 08:32:05 +1000527 BUG_ON(stripe_operations_active(sh));
Dan Williamsd84e0f12007-01-02 13:52:30 -0700528
Dan Williams45b42332007-07-09 11:56:43 -0700529 pr_debug("init_stripe called, stripe %llu\n",
Markus Stockhausenb8e6a152014-08-23 20:19:27 +1000530 (unsigned long long)sector);
Shaohua Li566c09c2013-11-14 15:16:17 +1100531retry:
532 seq = read_seqcount_begin(&conf->gen_lock);
NeilBrown86b42c72009-03-31 15:19:03 +1100533 sh->generation = conf->generation - previous;
NeilBrownb5663ba2009-03-31 14:39:38 +1100534 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 sh->sector = sector;
NeilBrown911d4ee2009-03-31 14:39:38 +1100536 stripe_set_idx(sector, conf, previous, sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 sh->state = 0;
538
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800539 for (i = sh->disks; i--; ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 struct r5dev *dev = &sh->dev[i];
541
Dan Williamsd84e0f12007-01-02 13:52:30 -0700542 if (dev->toread || dev->read || dev->towrite || dev->written ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 test_bit(R5_LOCKED, &dev->flags)) {
Dan Williamsd84e0f12007-01-02 13:52:30 -0700544 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 (unsigned long long)sh->sector, i, dev->toread,
Dan Williamsd84e0f12007-01-02 13:52:30 -0700546 dev->read, dev->towrite, dev->written,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 test_bit(R5_LOCKED, &dev->flags));
NeilBrown8cfa7b02011-07-27 11:00:36 +1000548 WARN_ON(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 }
550 dev->flags = 0;
NeilBrown784052e2009-03-31 15:19:07 +1100551 raid5_build_block(sh, i, previous);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 }
Shaohua Li566c09c2013-11-14 15:16:17 +1100553 if (read_seqcount_retry(&conf->gen_lock, seq))
554 goto retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555 insert_hash(conf, sh);
Shaohua Li851c30c2013-08-28 14:30:16 +0800556 sh->cpu = smp_processor_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557}
558
NeilBrownd1688a62011-10-11 16:49:52 +1100559static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
NeilBrown86b42c72009-03-31 15:19:03 +1100560 short generation)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561{
562 struct stripe_head *sh;
563
Dan Williams45b42332007-07-09 11:56:43 -0700564 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
Sasha Levinb67bfe02013-02-27 17:06:00 -0800565 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
NeilBrown86b42c72009-03-31 15:19:03 +1100566 if (sh->sector == sector && sh->generation == generation)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 return sh;
Dan Williams45b42332007-07-09 11:56:43 -0700568 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 return NULL;
570}
571
NeilBrown674806d2010-06-16 17:17:53 +1000572/*
573 * Need to check if array has failed when deciding whether to:
574 * - start an array
575 * - remove non-faulty devices
576 * - add a spare
577 * - allow a reshape
578 * This determination is simple when no reshape is happening.
579 * However if there is a reshape, we need to carefully check
580 * both the before and after sections.
581 * This is because some failed devices may only affect one
582 * of the two sections, and some non-in_sync devices may
583 * be insync in the section most affected by failed devices.
584 */
NeilBrown908f4fb2011-12-23 10:17:50 +1100585static int calc_degraded(struct r5conf *conf)
NeilBrown674806d2010-06-16 17:17:53 +1000586{
NeilBrown908f4fb2011-12-23 10:17:50 +1100587 int degraded, degraded2;
NeilBrown674806d2010-06-16 17:17:53 +1000588 int i;
NeilBrown674806d2010-06-16 17:17:53 +1000589
590 rcu_read_lock();
591 degraded = 0;
592 for (i = 0; i < conf->previous_raid_disks; i++) {
NeilBrown3cb03002011-10-11 16:45:26 +1100593 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
NeilBrowne5c86472012-09-19 12:52:30 +1000594 if (rdev && test_bit(Faulty, &rdev->flags))
595 rdev = rcu_dereference(conf->disks[i].replacement);
NeilBrown674806d2010-06-16 17:17:53 +1000596 if (!rdev || test_bit(Faulty, &rdev->flags))
597 degraded++;
598 else if (test_bit(In_sync, &rdev->flags))
599 ;
600 else
601 /* not in-sync or faulty.
602 * If the reshape increases the number of devices,
603 * this is being recovered by the reshape, so
604 * this 'previous' section is not in_sync.
605 * If the number of devices is being reduced however,
606 * the device can only be part of the array if
607 * we are reverting a reshape, so this section will
608 * be in-sync.
609 */
610 if (conf->raid_disks >= conf->previous_raid_disks)
611 degraded++;
612 }
613 rcu_read_unlock();
NeilBrown908f4fb2011-12-23 10:17:50 +1100614 if (conf->raid_disks == conf->previous_raid_disks)
615 return degraded;
NeilBrown674806d2010-06-16 17:17:53 +1000616 rcu_read_lock();
NeilBrown908f4fb2011-12-23 10:17:50 +1100617 degraded2 = 0;
NeilBrown674806d2010-06-16 17:17:53 +1000618 for (i = 0; i < conf->raid_disks; i++) {
NeilBrown3cb03002011-10-11 16:45:26 +1100619 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
NeilBrowne5c86472012-09-19 12:52:30 +1000620 if (rdev && test_bit(Faulty, &rdev->flags))
621 rdev = rcu_dereference(conf->disks[i].replacement);
NeilBrown674806d2010-06-16 17:17:53 +1000622 if (!rdev || test_bit(Faulty, &rdev->flags))
NeilBrown908f4fb2011-12-23 10:17:50 +1100623 degraded2++;
NeilBrown674806d2010-06-16 17:17:53 +1000624 else if (test_bit(In_sync, &rdev->flags))
625 ;
626 else
627 /* not in-sync or faulty.
628 * If reshape increases the number of devices, this
629 * section has already been recovered, else it
630 * almost certainly hasn't.
631 */
632 if (conf->raid_disks <= conf->previous_raid_disks)
NeilBrown908f4fb2011-12-23 10:17:50 +1100633 degraded2++;
NeilBrown674806d2010-06-16 17:17:53 +1000634 }
635 rcu_read_unlock();
NeilBrown908f4fb2011-12-23 10:17:50 +1100636 if (degraded2 > degraded)
637 return degraded2;
638 return degraded;
639}
640
641static int has_failed(struct r5conf *conf)
642{
643 int degraded;
644
645 if (conf->mddev->reshape_position == MaxSector)
646 return conf->mddev->degraded > conf->max_degraded;
647
648 degraded = calc_degraded(conf);
NeilBrown674806d2010-06-16 17:17:53 +1000649 if (degraded > conf->max_degraded)
650 return 1;
651 return 0;
652}
653
NeilBrownb5663ba2009-03-31 14:39:38 +1100654static struct stripe_head *
NeilBrownd1688a62011-10-11 16:49:52 +1100655get_active_stripe(struct r5conf *conf, sector_t sector,
NeilBrowna8c906c2009-06-09 14:39:59 +1000656 int previous, int noblock, int noquiesce)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657{
658 struct stripe_head *sh;
Shaohua Li566c09c2013-11-14 15:16:17 +1100659 int hash = stripe_hash_locks_hash(sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660
Dan Williams45b42332007-07-09 11:56:43 -0700661 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
Shaohua Li566c09c2013-11-14 15:16:17 +1100663 spin_lock_irq(conf->hash_locks + hash);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
665 do {
NeilBrown72626682005-09-09 16:23:54 -0700666 wait_event_lock_irq(conf->wait_for_stripe,
NeilBrowna8c906c2009-06-09 14:39:59 +1000667 conf->quiesce == 0 || noquiesce,
Shaohua Li566c09c2013-11-14 15:16:17 +1100668 *(conf->hash_locks + hash));
NeilBrown86b42c72009-03-31 15:19:03 +1100669 sh = __find_stripe(conf, sector, conf->generation - previous);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670 if (!sh) {
671 if (!conf->inactive_blocked)
Shaohua Li566c09c2013-11-14 15:16:17 +1100672 sh = get_free_stripe(conf, hash);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 if (noblock && sh == NULL)
674 break;
675 if (!sh) {
676 conf->inactive_blocked = 1;
Shaohua Li566c09c2013-11-14 15:16:17 +1100677 wait_event_lock_irq(
678 conf->wait_for_stripe,
679 !list_empty(conf->inactive_list + hash) &&
680 (atomic_read(&conf->active_stripes)
681 < (conf->max_nr_stripes * 3 / 4)
682 || !conf->inactive_blocked),
683 *(conf->hash_locks + hash));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 conf->inactive_blocked = 0;
NeilBrown7da9d452014-01-22 11:45:03 +1100685 } else {
NeilBrownb5663ba2009-03-31 14:39:38 +1100686 init_stripe(sh, sector, previous);
NeilBrown7da9d452014-01-22 11:45:03 +1100687 atomic_inc(&sh->count);
688 }
Shaohua Lie240c182014-04-09 11:27:42 +0800689 } else if (!atomic_inc_not_zero(&sh->count)) {
NeilBrown6d183de2013-11-28 10:55:27 +1100690 spin_lock(&conf->device_lock);
Shaohua Lie240c182014-04-09 11:27:42 +0800691 if (!atomic_read(&sh->count)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 if (!test_bit(STRIPE_HANDLE, &sh->state))
693 atomic_inc(&conf->active_stripes);
NeilBrown5af9bef2014-01-14 15:16:10 +1100694 BUG_ON(list_empty(&sh->lru) &&
695 !test_bit(STRIPE_EXPANDING, &sh->state));
NeilBrown16a53ec2006-06-26 00:27:38 -0700696 list_del_init(&sh->lru);
Shaohua Libfc90cb2013-08-29 15:40:32 +0800697 if (sh->group) {
698 sh->group->stripes_cnt--;
699 sh->group = NULL;
700 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701 }
NeilBrown7da9d452014-01-22 11:45:03 +1100702 atomic_inc(&sh->count);
NeilBrown6d183de2013-11-28 10:55:27 +1100703 spin_unlock(&conf->device_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 }
705 } while (sh == NULL);
706
Shaohua Li566c09c2013-11-14 15:16:17 +1100707 spin_unlock_irq(conf->hash_locks + hash);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708 return sh;
709}
710
NeilBrown05616be2012-05-21 09:27:00 +1000711/* Determine if 'data_offset' or 'new_data_offset' should be used
712 * in this stripe_head.
713 */
714static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
715{
716 sector_t progress = conf->reshape_progress;
717 /* Need a memory barrier to make sure we see the value
718 * of conf->generation, or ->data_offset that was set before
719 * reshape_progress was updated.
720 */
721 smp_rmb();
722 if (progress == MaxSector)
723 return 0;
724 if (sh->generation == conf->generation - 1)
725 return 0;
726 /* We are in a reshape, and this is a new-generation stripe,
727 * so use new_data_offset.
728 */
729 return 1;
730}
731
NeilBrown6712ecf2007-09-27 12:47:43 +0200732static void
733raid5_end_read_request(struct bio *bi, int error);
734static void
735raid5_end_write_request(struct bio *bi, int error);
Dan Williams91c00922007-01-02 13:52:30 -0700736
Dan Williamsc4e5ac02008-06-28 08:31:53 +1000737static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
Dan Williams91c00922007-01-02 13:52:30 -0700738{
NeilBrownd1688a62011-10-11 16:49:52 +1100739 struct r5conf *conf = sh->raid_conf;
Dan Williams91c00922007-01-02 13:52:30 -0700740 int i, disks = sh->disks;
741
742 might_sleep();
743
744 for (i = disks; i--; ) {
745 int rw;
NeilBrown9a3e1102011-12-23 10:17:53 +1100746 int replace_only = 0;
NeilBrown977df362011-12-23 10:17:53 +1100747 struct bio *bi, *rbi;
748 struct md_rdev *rdev, *rrdev = NULL;
Tejun Heoe9c74692010-09-03 11:56:18 +0200749 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
750 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
751 rw = WRITE_FUA;
752 else
753 rw = WRITE;
Shaohua Li9e4447682012-10-11 13:49:49 +1100754 if (test_bit(R5_Discard, &sh->dev[i].flags))
Shaohua Li620125f2012-10-11 13:49:05 +1100755 rw |= REQ_DISCARD;
Tejun Heoe9c74692010-09-03 11:56:18 +0200756 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
Dan Williams91c00922007-01-02 13:52:30 -0700757 rw = READ;
NeilBrown9a3e1102011-12-23 10:17:53 +1100758 else if (test_and_clear_bit(R5_WantReplace,
759 &sh->dev[i].flags)) {
760 rw = WRITE;
761 replace_only = 1;
762 } else
Dan Williams91c00922007-01-02 13:52:30 -0700763 continue;
Shaohua Libc0934f2012-05-22 13:55:05 +1000764 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
765 rw |= REQ_SYNC;
Dan Williams91c00922007-01-02 13:52:30 -0700766
767 bi = &sh->dev[i].req;
NeilBrown977df362011-12-23 10:17:53 +1100768 rbi = &sh->dev[i].rreq; /* For writing to replacement */
Dan Williams91c00922007-01-02 13:52:30 -0700769
Dan Williams91c00922007-01-02 13:52:30 -0700770 rcu_read_lock();
NeilBrown9a3e1102011-12-23 10:17:53 +1100771 rrdev = rcu_dereference(conf->disks[i].replacement);
NeilBrowndd054fc2011-12-23 10:17:53 +1100772 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
773 rdev = rcu_dereference(conf->disks[i].rdev);
774 if (!rdev) {
775 rdev = rrdev;
776 rrdev = NULL;
777 }
NeilBrown9a3e1102011-12-23 10:17:53 +1100778 if (rw & WRITE) {
779 if (replace_only)
780 rdev = NULL;
NeilBrowndd054fc2011-12-23 10:17:53 +1100781 if (rdev == rrdev)
782 /* We raced and saw duplicates */
783 rrdev = NULL;
NeilBrown9a3e1102011-12-23 10:17:53 +1100784 } else {
NeilBrowndd054fc2011-12-23 10:17:53 +1100785 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
NeilBrown9a3e1102011-12-23 10:17:53 +1100786 rdev = rrdev;
787 rrdev = NULL;
788 }
NeilBrown977df362011-12-23 10:17:53 +1100789
Dan Williams91c00922007-01-02 13:52:30 -0700790 if (rdev && test_bit(Faulty, &rdev->flags))
791 rdev = NULL;
792 if (rdev)
793 atomic_inc(&rdev->nr_pending);
NeilBrown977df362011-12-23 10:17:53 +1100794 if (rrdev && test_bit(Faulty, &rrdev->flags))
795 rrdev = NULL;
796 if (rrdev)
797 atomic_inc(&rrdev->nr_pending);
Dan Williams91c00922007-01-02 13:52:30 -0700798 rcu_read_unlock();
799
NeilBrown73e92e52011-07-28 11:39:22 +1000800 /* We have already checked bad blocks for reads. Now
NeilBrown977df362011-12-23 10:17:53 +1100801 * need to check for writes. We never accept write errors
802 * on the replacement, so we don't to check rrdev.
NeilBrown73e92e52011-07-28 11:39:22 +1000803 */
804 while ((rw & WRITE) && rdev &&
805 test_bit(WriteErrorSeen, &rdev->flags)) {
806 sector_t first_bad;
807 int bad_sectors;
808 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
809 &first_bad, &bad_sectors);
810 if (!bad)
811 break;
812
813 if (bad < 0) {
814 set_bit(BlockedBadBlocks, &rdev->flags);
815 if (!conf->mddev->external &&
816 conf->mddev->flags) {
817 /* It is very unlikely, but we might
818 * still need to write out the
819 * bad block log - better give it
820 * a chance*/
821 md_check_recovery(conf->mddev);
822 }
majianpeng18507532012-07-03 12:11:54 +1000823 /*
824 * Because md_wait_for_blocked_rdev
825 * will dec nr_pending, we must
826 * increment it first.
827 */
828 atomic_inc(&rdev->nr_pending);
NeilBrown73e92e52011-07-28 11:39:22 +1000829 md_wait_for_blocked_rdev(rdev, conf->mddev);
830 } else {
831 /* Acknowledged bad block - skip the write */
832 rdev_dec_pending(rdev, conf->mddev);
833 rdev = NULL;
834 }
835 }
836
Dan Williams91c00922007-01-02 13:52:30 -0700837 if (rdev) {
NeilBrown9a3e1102011-12-23 10:17:53 +1100838 if (s->syncing || s->expanding || s->expanded
839 || s->replacing)
Dan Williams91c00922007-01-02 13:52:30 -0700840 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
841
Dan Williams2b7497f2008-06-28 08:31:52 +1000842 set_bit(STRIPE_IO_STARTED, &sh->state);
843
Kent Overstreet2f6db2a2012-09-11 12:26:38 -0700844 bio_reset(bi);
Dan Williams91c00922007-01-02 13:52:30 -0700845 bi->bi_bdev = rdev->bdev;
Kent Overstreet2f6db2a2012-09-11 12:26:38 -0700846 bi->bi_rw = rw;
847 bi->bi_end_io = (rw & WRITE)
848 ? raid5_end_write_request
849 : raid5_end_read_request;
850 bi->bi_private = sh;
851
Dan Williams91c00922007-01-02 13:52:30 -0700852 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
Harvey Harrisone46b2722008-04-28 02:15:50 -0700853 __func__, (unsigned long long)sh->sector,
Dan Williams91c00922007-01-02 13:52:30 -0700854 bi->bi_rw, i);
855 atomic_inc(&sh->count);
NeilBrown05616be2012-05-21 09:27:00 +1000856 if (use_new_offset(conf, sh))
Kent Overstreet4f024f32013-10-11 15:44:27 -0700857 bi->bi_iter.bi_sector = (sh->sector
NeilBrown05616be2012-05-21 09:27:00 +1000858 + rdev->new_data_offset);
859 else
Kent Overstreet4f024f32013-10-11 15:44:27 -0700860 bi->bi_iter.bi_sector = (sh->sector
NeilBrown05616be2012-05-21 09:27:00 +1000861 + rdev->data_offset);
majianpeng3f9e7c12012-07-31 10:04:21 +1000862 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
majianpenge59aa232013-11-14 15:16:19 +1100863 bi->bi_rw |= REQ_NOMERGE;
majianpeng3f9e7c12012-07-31 10:04:21 +1000864
Shaohua Lid592a992014-05-21 17:57:44 +0800865 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
866 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
867 sh->dev[i].vec.bv_page = sh->dev[i].page;
Kent Overstreet4997b722013-05-30 08:44:39 +0200868 bi->bi_vcnt = 1;
Dan Williams91c00922007-01-02 13:52:30 -0700869 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
870 bi->bi_io_vec[0].bv_offset = 0;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700871 bi->bi_iter.bi_size = STRIPE_SIZE;
Shaohua Li37c61ff2013-10-19 14:50:28 +0800872 /*
873 * If this is discard request, set bi_vcnt 0. We don't
874 * want to confuse SCSI because SCSI will replace payload
875 */
876 if (rw & REQ_DISCARD)
877 bi->bi_vcnt = 0;
NeilBrown977df362011-12-23 10:17:53 +1100878 if (rrdev)
879 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
Jonathan Brassowe3620a32013-03-07 16:22:01 -0600880
881 if (conf->mddev->gendisk)
882 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
883 bi, disk_devt(conf->mddev->gendisk),
884 sh->dev[i].sector);
Dan Williams91c00922007-01-02 13:52:30 -0700885 generic_make_request(bi);
NeilBrown977df362011-12-23 10:17:53 +1100886 }
887 if (rrdev) {
NeilBrown9a3e1102011-12-23 10:17:53 +1100888 if (s->syncing || s->expanding || s->expanded
889 || s->replacing)
NeilBrown977df362011-12-23 10:17:53 +1100890 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
891
892 set_bit(STRIPE_IO_STARTED, &sh->state);
893
Kent Overstreet2f6db2a2012-09-11 12:26:38 -0700894 bio_reset(rbi);
NeilBrown977df362011-12-23 10:17:53 +1100895 rbi->bi_bdev = rrdev->bdev;
Kent Overstreet2f6db2a2012-09-11 12:26:38 -0700896 rbi->bi_rw = rw;
897 BUG_ON(!(rw & WRITE));
898 rbi->bi_end_io = raid5_end_write_request;
899 rbi->bi_private = sh;
900
NeilBrown977df362011-12-23 10:17:53 +1100901 pr_debug("%s: for %llu schedule op %ld on "
902 "replacement disc %d\n",
903 __func__, (unsigned long long)sh->sector,
904 rbi->bi_rw, i);
905 atomic_inc(&sh->count);
NeilBrown05616be2012-05-21 09:27:00 +1000906 if (use_new_offset(conf, sh))
Kent Overstreet4f024f32013-10-11 15:44:27 -0700907 rbi->bi_iter.bi_sector = (sh->sector
NeilBrown05616be2012-05-21 09:27:00 +1000908 + rrdev->new_data_offset);
909 else
Kent Overstreet4f024f32013-10-11 15:44:27 -0700910 rbi->bi_iter.bi_sector = (sh->sector
NeilBrown05616be2012-05-21 09:27:00 +1000911 + rrdev->data_offset);
Shaohua Lid592a992014-05-21 17:57:44 +0800912 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
913 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
914 sh->dev[i].rvec.bv_page = sh->dev[i].page;
Kent Overstreet4997b722013-05-30 08:44:39 +0200915 rbi->bi_vcnt = 1;
NeilBrown977df362011-12-23 10:17:53 +1100916 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
917 rbi->bi_io_vec[0].bv_offset = 0;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700918 rbi->bi_iter.bi_size = STRIPE_SIZE;
Shaohua Li37c61ff2013-10-19 14:50:28 +0800919 /*
920 * If this is discard request, set bi_vcnt 0. We don't
921 * want to confuse SCSI because SCSI will replace payload
922 */
923 if (rw & REQ_DISCARD)
924 rbi->bi_vcnt = 0;
Jonathan Brassowe3620a32013-03-07 16:22:01 -0600925 if (conf->mddev->gendisk)
926 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
927 rbi, disk_devt(conf->mddev->gendisk),
928 sh->dev[i].sector);
NeilBrown977df362011-12-23 10:17:53 +1100929 generic_make_request(rbi);
930 }
931 if (!rdev && !rrdev) {
Namhyung Kimb0629622011-06-14 14:20:19 +1000932 if (rw & WRITE)
Dan Williams91c00922007-01-02 13:52:30 -0700933 set_bit(STRIPE_DEGRADED, &sh->state);
934 pr_debug("skip op %ld on disc %d for sector %llu\n",
935 bi->bi_rw, i, (unsigned long long)sh->sector);
936 clear_bit(R5_LOCKED, &sh->dev[i].flags);
937 set_bit(STRIPE_HANDLE, &sh->state);
938 }
939 }
940}
941
942static struct dma_async_tx_descriptor *
Shaohua Lid592a992014-05-21 17:57:44 +0800943async_copy_data(int frombio, struct bio *bio, struct page **page,
944 sector_t sector, struct dma_async_tx_descriptor *tx,
945 struct stripe_head *sh)
Dan Williams91c00922007-01-02 13:52:30 -0700946{
Kent Overstreet79886132013-11-23 17:19:00 -0800947 struct bio_vec bvl;
948 struct bvec_iter iter;
Dan Williams91c00922007-01-02 13:52:30 -0700949 struct page *bio_page;
Dan Williams91c00922007-01-02 13:52:30 -0700950 int page_offset;
Dan Williamsa08abd82009-06-03 11:43:59 -0700951 struct async_submit_ctl submit;
Dan Williams0403e382009-09-08 17:42:50 -0700952 enum async_tx_flags flags = 0;
Dan Williams91c00922007-01-02 13:52:30 -0700953
Kent Overstreet4f024f32013-10-11 15:44:27 -0700954 if (bio->bi_iter.bi_sector >= sector)
955 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
Dan Williams91c00922007-01-02 13:52:30 -0700956 else
Kent Overstreet4f024f32013-10-11 15:44:27 -0700957 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
Dan Williamsa08abd82009-06-03 11:43:59 -0700958
Dan Williams0403e382009-09-08 17:42:50 -0700959 if (frombio)
960 flags |= ASYNC_TX_FENCE;
961 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
962
Kent Overstreet79886132013-11-23 17:19:00 -0800963 bio_for_each_segment(bvl, bio, iter) {
964 int len = bvl.bv_len;
Dan Williams91c00922007-01-02 13:52:30 -0700965 int clen;
966 int b_offset = 0;
967
968 if (page_offset < 0) {
969 b_offset = -page_offset;
970 page_offset += b_offset;
971 len -= b_offset;
972 }
973
974 if (len > 0 && page_offset + len > STRIPE_SIZE)
975 clen = STRIPE_SIZE - page_offset;
976 else
977 clen = len;
978
979 if (clen > 0) {
Kent Overstreet79886132013-11-23 17:19:00 -0800980 b_offset += bvl.bv_offset;
981 bio_page = bvl.bv_page;
Shaohua Lid592a992014-05-21 17:57:44 +0800982 if (frombio) {
983 if (sh->raid_conf->skip_copy &&
984 b_offset == 0 && page_offset == 0 &&
985 clen == STRIPE_SIZE)
986 *page = bio_page;
987 else
988 tx = async_memcpy(*page, bio_page, page_offset,
Dan Williamsa08abd82009-06-03 11:43:59 -0700989 b_offset, clen, &submit);
Shaohua Lid592a992014-05-21 17:57:44 +0800990 } else
991 tx = async_memcpy(bio_page, *page, b_offset,
Dan Williamsa08abd82009-06-03 11:43:59 -0700992 page_offset, clen, &submit);
Dan Williams91c00922007-01-02 13:52:30 -0700993 }
Dan Williamsa08abd82009-06-03 11:43:59 -0700994 /* chain the operations */
995 submit.depend_tx = tx;
996
Dan Williams91c00922007-01-02 13:52:30 -0700997 if (clen < len) /* hit end of page */
998 break;
999 page_offset += len;
1000 }
1001
1002 return tx;
1003}
1004
1005static void ops_complete_biofill(void *stripe_head_ref)
1006{
1007 struct stripe_head *sh = stripe_head_ref;
1008 struct bio *return_bi = NULL;
Dan Williamse4d84902007-09-24 10:06:13 -07001009 int i;
Dan Williams91c00922007-01-02 13:52:30 -07001010
Harvey Harrisone46b2722008-04-28 02:15:50 -07001011 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001012 (unsigned long long)sh->sector);
1013
1014 /* clear completed biofills */
1015 for (i = sh->disks; i--; ) {
1016 struct r5dev *dev = &sh->dev[i];
Dan Williams91c00922007-01-02 13:52:30 -07001017
1018 /* acknowledge completion of a biofill operation */
Dan Williamse4d84902007-09-24 10:06:13 -07001019 /* and check if we need to reply to a read request,
1020 * new R5_Wantfill requests are held off until
Dan Williams83de75c2008-06-28 08:31:58 +10001021 * !STRIPE_BIOFILL_RUN
Dan Williamse4d84902007-09-24 10:06:13 -07001022 */
1023 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
Dan Williams91c00922007-01-02 13:52:30 -07001024 struct bio *rbi, *rbi2;
Dan Williams91c00922007-01-02 13:52:30 -07001025
Dan Williams91c00922007-01-02 13:52:30 -07001026 BUG_ON(!dev->read);
1027 rbi = dev->read;
1028 dev->read = NULL;
Kent Overstreet4f024f32013-10-11 15:44:27 -07001029 while (rbi && rbi->bi_iter.bi_sector <
Dan Williams91c00922007-01-02 13:52:30 -07001030 dev->sector + STRIPE_SECTORS) {
1031 rbi2 = r5_next_bio(rbi, dev->sector);
Shaohua Lie7836bd62012-07-19 16:01:31 +10001032 if (!raid5_dec_bi_active_stripes(rbi)) {
Dan Williams91c00922007-01-02 13:52:30 -07001033 rbi->bi_next = return_bi;
1034 return_bi = rbi;
1035 }
Dan Williams91c00922007-01-02 13:52:30 -07001036 rbi = rbi2;
1037 }
1038 }
1039 }
Dan Williams83de75c2008-06-28 08:31:58 +10001040 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
Dan Williams91c00922007-01-02 13:52:30 -07001041
1042 return_io(return_bi);
1043
Dan Williamse4d84902007-09-24 10:06:13 -07001044 set_bit(STRIPE_HANDLE, &sh->state);
Dan Williams91c00922007-01-02 13:52:30 -07001045 release_stripe(sh);
1046}
1047
1048static void ops_run_biofill(struct stripe_head *sh)
1049{
1050 struct dma_async_tx_descriptor *tx = NULL;
Dan Williamsa08abd82009-06-03 11:43:59 -07001051 struct async_submit_ctl submit;
Dan Williams91c00922007-01-02 13:52:30 -07001052 int i;
1053
Harvey Harrisone46b2722008-04-28 02:15:50 -07001054 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001055 (unsigned long long)sh->sector);
1056
1057 for (i = sh->disks; i--; ) {
1058 struct r5dev *dev = &sh->dev[i];
1059 if (test_bit(R5_Wantfill, &dev->flags)) {
1060 struct bio *rbi;
Shaohua Lib17459c2012-07-19 16:01:31 +10001061 spin_lock_irq(&sh->stripe_lock);
Dan Williams91c00922007-01-02 13:52:30 -07001062 dev->read = rbi = dev->toread;
1063 dev->toread = NULL;
Shaohua Lib17459c2012-07-19 16:01:31 +10001064 spin_unlock_irq(&sh->stripe_lock);
Kent Overstreet4f024f32013-10-11 15:44:27 -07001065 while (rbi && rbi->bi_iter.bi_sector <
Dan Williams91c00922007-01-02 13:52:30 -07001066 dev->sector + STRIPE_SECTORS) {
Shaohua Lid592a992014-05-21 17:57:44 +08001067 tx = async_copy_data(0, rbi, &dev->page,
1068 dev->sector, tx, sh);
Dan Williams91c00922007-01-02 13:52:30 -07001069 rbi = r5_next_bio(rbi, dev->sector);
1070 }
1071 }
1072 }
1073
1074 atomic_inc(&sh->count);
Dan Williamsa08abd82009-06-03 11:43:59 -07001075 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1076 async_trigger_callback(&submit);
Dan Williams91c00922007-01-02 13:52:30 -07001077}
1078
Dan Williams4e7d2c02009-08-29 19:13:11 -07001079static void mark_target_uptodate(struct stripe_head *sh, int target)
1080{
1081 struct r5dev *tgt;
1082
1083 if (target < 0)
1084 return;
1085
1086 tgt = &sh->dev[target];
1087 set_bit(R5_UPTODATE, &tgt->flags);
1088 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1089 clear_bit(R5_Wantcompute, &tgt->flags);
1090}
1091
Dan Williamsac6b53b2009-07-14 13:40:19 -07001092static void ops_complete_compute(void *stripe_head_ref)
Dan Williams91c00922007-01-02 13:52:30 -07001093{
1094 struct stripe_head *sh = stripe_head_ref;
Dan Williams91c00922007-01-02 13:52:30 -07001095
Harvey Harrisone46b2722008-04-28 02:15:50 -07001096 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001097 (unsigned long long)sh->sector);
1098
Dan Williamsac6b53b2009-07-14 13:40:19 -07001099 /* mark the computed target(s) as uptodate */
Dan Williams4e7d2c02009-08-29 19:13:11 -07001100 mark_target_uptodate(sh, sh->ops.target);
Dan Williamsac6b53b2009-07-14 13:40:19 -07001101 mark_target_uptodate(sh, sh->ops.target2);
Dan Williams4e7d2c02009-08-29 19:13:11 -07001102
Dan Williamsecc65c92008-06-28 08:31:57 +10001103 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1104 if (sh->check_state == check_state_compute_run)
1105 sh->check_state = check_state_compute_result;
Dan Williams91c00922007-01-02 13:52:30 -07001106 set_bit(STRIPE_HANDLE, &sh->state);
1107 release_stripe(sh);
1108}
1109
Dan Williamsd6f38f32009-07-14 11:50:52 -07001110/* return a pointer to the address conversion region of the scribble buffer */
1111static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1112 struct raid5_percpu *percpu)
Dan Williams91c00922007-01-02 13:52:30 -07001113{
Dan Williamsd6f38f32009-07-14 11:50:52 -07001114 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
1115}
1116
1117static struct dma_async_tx_descriptor *
1118ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1119{
Dan Williams91c00922007-01-02 13:52:30 -07001120 int disks = sh->disks;
Dan Williamsd6f38f32009-07-14 11:50:52 -07001121 struct page **xor_srcs = percpu->scribble;
Dan Williams91c00922007-01-02 13:52:30 -07001122 int target = sh->ops.target;
1123 struct r5dev *tgt = &sh->dev[target];
1124 struct page *xor_dest = tgt->page;
1125 int count = 0;
1126 struct dma_async_tx_descriptor *tx;
Dan Williamsa08abd82009-06-03 11:43:59 -07001127 struct async_submit_ctl submit;
Dan Williams91c00922007-01-02 13:52:30 -07001128 int i;
1129
1130 pr_debug("%s: stripe %llu block: %d\n",
Harvey Harrisone46b2722008-04-28 02:15:50 -07001131 __func__, (unsigned long long)sh->sector, target);
Dan Williams91c00922007-01-02 13:52:30 -07001132 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1133
1134 for (i = disks; i--; )
1135 if (i != target)
1136 xor_srcs[count++] = sh->dev[i].page;
1137
1138 atomic_inc(&sh->count);
1139
Dan Williams0403e382009-09-08 17:42:50 -07001140 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
Dan Williamsac6b53b2009-07-14 13:40:19 -07001141 ops_complete_compute, sh, to_addr_conv(sh, percpu));
Dan Williams91c00922007-01-02 13:52:30 -07001142 if (unlikely(count == 1))
Dan Williamsa08abd82009-06-03 11:43:59 -07001143 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
Dan Williams91c00922007-01-02 13:52:30 -07001144 else
Dan Williamsa08abd82009-06-03 11:43:59 -07001145 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
Dan Williams91c00922007-01-02 13:52:30 -07001146
Dan Williams91c00922007-01-02 13:52:30 -07001147 return tx;
1148}
1149
Dan Williamsac6b53b2009-07-14 13:40:19 -07001150/* set_syndrome_sources - populate source buffers for gen_syndrome
1151 * @srcs - (struct page *) array of size sh->disks
1152 * @sh - stripe_head to parse
1153 *
1154 * Populates srcs in proper layout order for the stripe and returns the
1155 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1156 * destination buffer is recorded in srcs[count] and the Q destination
1157 * is recorded in srcs[count+1]].
1158 */
1159static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
1160{
1161 int disks = sh->disks;
1162 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1163 int d0_idx = raid6_d0(sh);
1164 int count;
1165 int i;
1166
1167 for (i = 0; i < disks; i++)
NeilBrown5dd33c92009-10-16 16:40:25 +11001168 srcs[i] = NULL;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001169
1170 count = 0;
1171 i = d0_idx;
1172 do {
1173 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1174
1175 srcs[slot] = sh->dev[i].page;
1176 i = raid6_next_disk(i, disks);
1177 } while (i != d0_idx);
Dan Williamsac6b53b2009-07-14 13:40:19 -07001178
NeilBrowne4424fe2009-10-16 16:27:34 +11001179 return syndrome_disks;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001180}
1181
1182static struct dma_async_tx_descriptor *
1183ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1184{
1185 int disks = sh->disks;
1186 struct page **blocks = percpu->scribble;
1187 int target;
1188 int qd_idx = sh->qd_idx;
1189 struct dma_async_tx_descriptor *tx;
1190 struct async_submit_ctl submit;
1191 struct r5dev *tgt;
1192 struct page *dest;
1193 int i;
1194 int count;
1195
1196 if (sh->ops.target < 0)
1197 target = sh->ops.target2;
1198 else if (sh->ops.target2 < 0)
1199 target = sh->ops.target;
1200 else
1201 /* we should only have one valid target */
1202 BUG();
1203 BUG_ON(target < 0);
1204 pr_debug("%s: stripe %llu block: %d\n",
1205 __func__, (unsigned long long)sh->sector, target);
1206
1207 tgt = &sh->dev[target];
1208 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1209 dest = tgt->page;
1210
1211 atomic_inc(&sh->count);
1212
1213 if (target == qd_idx) {
1214 count = set_syndrome_sources(blocks, sh);
1215 blocks[count] = NULL; /* regenerating p is not necessary */
1216 BUG_ON(blocks[count+1] != dest); /* q should already be set */
Dan Williams0403e382009-09-08 17:42:50 -07001217 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1218 ops_complete_compute, sh,
Dan Williamsac6b53b2009-07-14 13:40:19 -07001219 to_addr_conv(sh, percpu));
1220 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1221 } else {
1222 /* Compute any data- or p-drive using XOR */
1223 count = 0;
1224 for (i = disks; i-- ; ) {
1225 if (i == target || i == qd_idx)
1226 continue;
1227 blocks[count++] = sh->dev[i].page;
1228 }
1229
Dan Williams0403e382009-09-08 17:42:50 -07001230 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1231 NULL, ops_complete_compute, sh,
Dan Williamsac6b53b2009-07-14 13:40:19 -07001232 to_addr_conv(sh, percpu));
1233 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1234 }
1235
1236 return tx;
1237}
1238
1239static struct dma_async_tx_descriptor *
1240ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1241{
1242 int i, count, disks = sh->disks;
1243 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1244 int d0_idx = raid6_d0(sh);
1245 int faila = -1, failb = -1;
1246 int target = sh->ops.target;
1247 int target2 = sh->ops.target2;
1248 struct r5dev *tgt = &sh->dev[target];
1249 struct r5dev *tgt2 = &sh->dev[target2];
1250 struct dma_async_tx_descriptor *tx;
1251 struct page **blocks = percpu->scribble;
1252 struct async_submit_ctl submit;
1253
1254 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1255 __func__, (unsigned long long)sh->sector, target, target2);
1256 BUG_ON(target < 0 || target2 < 0);
1257 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1258 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1259
Dan Williams6c910a72009-09-16 12:24:54 -07001260 /* we need to open-code set_syndrome_sources to handle the
Dan Williamsac6b53b2009-07-14 13:40:19 -07001261 * slot number conversion for 'faila' and 'failb'
1262 */
1263 for (i = 0; i < disks ; i++)
NeilBrown5dd33c92009-10-16 16:40:25 +11001264 blocks[i] = NULL;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001265 count = 0;
1266 i = d0_idx;
1267 do {
1268 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1269
1270 blocks[slot] = sh->dev[i].page;
1271
1272 if (i == target)
1273 faila = slot;
1274 if (i == target2)
1275 failb = slot;
1276 i = raid6_next_disk(i, disks);
1277 } while (i != d0_idx);
Dan Williamsac6b53b2009-07-14 13:40:19 -07001278
1279 BUG_ON(faila == failb);
1280 if (failb < faila)
1281 swap(faila, failb);
1282 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1283 __func__, (unsigned long long)sh->sector, faila, failb);
1284
1285 atomic_inc(&sh->count);
1286
1287 if (failb == syndrome_disks+1) {
1288 /* Q disk is one of the missing disks */
1289 if (faila == syndrome_disks) {
1290 /* Missing P+Q, just recompute */
Dan Williams0403e382009-09-08 17:42:50 -07001291 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1292 ops_complete_compute, sh,
1293 to_addr_conv(sh, percpu));
NeilBrowne4424fe2009-10-16 16:27:34 +11001294 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
Dan Williamsac6b53b2009-07-14 13:40:19 -07001295 STRIPE_SIZE, &submit);
1296 } else {
1297 struct page *dest;
1298 int data_target;
1299 int qd_idx = sh->qd_idx;
1300
1301 /* Missing D+Q: recompute D from P, then recompute Q */
1302 if (target == qd_idx)
1303 data_target = target2;
1304 else
1305 data_target = target;
1306
1307 count = 0;
1308 for (i = disks; i-- ; ) {
1309 if (i == data_target || i == qd_idx)
1310 continue;
1311 blocks[count++] = sh->dev[i].page;
1312 }
1313 dest = sh->dev[data_target].page;
Dan Williams0403e382009-09-08 17:42:50 -07001314 init_async_submit(&submit,
1315 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1316 NULL, NULL, NULL,
1317 to_addr_conv(sh, percpu));
Dan Williamsac6b53b2009-07-14 13:40:19 -07001318 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1319 &submit);
1320
1321 count = set_syndrome_sources(blocks, sh);
Dan Williams0403e382009-09-08 17:42:50 -07001322 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1323 ops_complete_compute, sh,
1324 to_addr_conv(sh, percpu));
Dan Williamsac6b53b2009-07-14 13:40:19 -07001325 return async_gen_syndrome(blocks, 0, count+2,
1326 STRIPE_SIZE, &submit);
1327 }
Dan Williamsac6b53b2009-07-14 13:40:19 -07001328 } else {
Dan Williams6c910a72009-09-16 12:24:54 -07001329 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1330 ops_complete_compute, sh,
1331 to_addr_conv(sh, percpu));
1332 if (failb == syndrome_disks) {
1333 /* We're missing D+P. */
1334 return async_raid6_datap_recov(syndrome_disks+2,
1335 STRIPE_SIZE, faila,
1336 blocks, &submit);
1337 } else {
1338 /* We're missing D+D. */
1339 return async_raid6_2data_recov(syndrome_disks+2,
1340 STRIPE_SIZE, faila, failb,
1341 blocks, &submit);
1342 }
Dan Williamsac6b53b2009-07-14 13:40:19 -07001343 }
1344}
1345
Dan Williams91c00922007-01-02 13:52:30 -07001346static void ops_complete_prexor(void *stripe_head_ref)
1347{
1348 struct stripe_head *sh = stripe_head_ref;
1349
Harvey Harrisone46b2722008-04-28 02:15:50 -07001350 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001351 (unsigned long long)sh->sector);
Dan Williams91c00922007-01-02 13:52:30 -07001352}
1353
1354static struct dma_async_tx_descriptor *
Dan Williamsd6f38f32009-07-14 11:50:52 -07001355ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1356 struct dma_async_tx_descriptor *tx)
Dan Williams91c00922007-01-02 13:52:30 -07001357{
Dan Williams91c00922007-01-02 13:52:30 -07001358 int disks = sh->disks;
Dan Williamsd6f38f32009-07-14 11:50:52 -07001359 struct page **xor_srcs = percpu->scribble;
Dan Williams91c00922007-01-02 13:52:30 -07001360 int count = 0, pd_idx = sh->pd_idx, i;
Dan Williamsa08abd82009-06-03 11:43:59 -07001361 struct async_submit_ctl submit;
Dan Williams91c00922007-01-02 13:52:30 -07001362
1363 /* existing parity data subtracted */
1364 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1365
Harvey Harrisone46b2722008-04-28 02:15:50 -07001366 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001367 (unsigned long long)sh->sector);
1368
1369 for (i = disks; i--; ) {
1370 struct r5dev *dev = &sh->dev[i];
1371 /* Only process blocks that are known to be uptodate */
Dan Williamsd8ee0722008-06-28 08:32:06 +10001372 if (test_bit(R5_Wantdrain, &dev->flags))
Dan Williams91c00922007-01-02 13:52:30 -07001373 xor_srcs[count++] = dev->page;
1374 }
1375
Dan Williams0403e382009-09-08 17:42:50 -07001376 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
Dan Williamsd6f38f32009-07-14 11:50:52 -07001377 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
Dan Williamsa08abd82009-06-03 11:43:59 -07001378 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
Dan Williams91c00922007-01-02 13:52:30 -07001379
1380 return tx;
1381}
1382
1383static struct dma_async_tx_descriptor *
Dan Williamsd8ee0722008-06-28 08:32:06 +10001384ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
Dan Williams91c00922007-01-02 13:52:30 -07001385{
1386 int disks = sh->disks;
Dan Williamsd8ee0722008-06-28 08:32:06 +10001387 int i;
Dan Williams91c00922007-01-02 13:52:30 -07001388
Harvey Harrisone46b2722008-04-28 02:15:50 -07001389 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001390 (unsigned long long)sh->sector);
1391
1392 for (i = disks; i--; ) {
1393 struct r5dev *dev = &sh->dev[i];
1394 struct bio *chosen;
Dan Williams91c00922007-01-02 13:52:30 -07001395
Dan Williamsd8ee0722008-06-28 08:32:06 +10001396 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
Dan Williams91c00922007-01-02 13:52:30 -07001397 struct bio *wbi;
1398
Shaohua Lib17459c2012-07-19 16:01:31 +10001399 spin_lock_irq(&sh->stripe_lock);
Dan Williams91c00922007-01-02 13:52:30 -07001400 chosen = dev->towrite;
1401 dev->towrite = NULL;
1402 BUG_ON(dev->written);
1403 wbi = dev->written = chosen;
Shaohua Lib17459c2012-07-19 16:01:31 +10001404 spin_unlock_irq(&sh->stripe_lock);
Shaohua Lid592a992014-05-21 17:57:44 +08001405 WARN_ON(dev->page != dev->orig_page);
Dan Williams91c00922007-01-02 13:52:30 -07001406
Kent Overstreet4f024f32013-10-11 15:44:27 -07001407 while (wbi && wbi->bi_iter.bi_sector <
Dan Williams91c00922007-01-02 13:52:30 -07001408 dev->sector + STRIPE_SECTORS) {
Tejun Heoe9c74692010-09-03 11:56:18 +02001409 if (wbi->bi_rw & REQ_FUA)
1410 set_bit(R5_WantFUA, &dev->flags);
Shaohua Libc0934f2012-05-22 13:55:05 +10001411 if (wbi->bi_rw & REQ_SYNC)
1412 set_bit(R5_SyncIO, &dev->flags);
Shaohua Li9e4447682012-10-11 13:49:49 +11001413 if (wbi->bi_rw & REQ_DISCARD)
Shaohua Li620125f2012-10-11 13:49:05 +11001414 set_bit(R5_Discard, &dev->flags);
Shaohua Lid592a992014-05-21 17:57:44 +08001415 else {
1416 tx = async_copy_data(1, wbi, &dev->page,
1417 dev->sector, tx, sh);
1418 if (dev->page != dev->orig_page) {
1419 set_bit(R5_SkipCopy, &dev->flags);
1420 clear_bit(R5_UPTODATE, &dev->flags);
1421 clear_bit(R5_OVERWRITE, &dev->flags);
1422 }
1423 }
Dan Williams91c00922007-01-02 13:52:30 -07001424 wbi = r5_next_bio(wbi, dev->sector);
1425 }
1426 }
1427 }
1428
1429 return tx;
1430}
1431
Dan Williamsac6b53b2009-07-14 13:40:19 -07001432static void ops_complete_reconstruct(void *stripe_head_ref)
Dan Williams91c00922007-01-02 13:52:30 -07001433{
1434 struct stripe_head *sh = stripe_head_ref;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001435 int disks = sh->disks;
1436 int pd_idx = sh->pd_idx;
1437 int qd_idx = sh->qd_idx;
1438 int i;
Shaohua Li9e4447682012-10-11 13:49:49 +11001439 bool fua = false, sync = false, discard = false;
Dan Williams91c00922007-01-02 13:52:30 -07001440
Harvey Harrisone46b2722008-04-28 02:15:50 -07001441 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001442 (unsigned long long)sh->sector);
1443
Shaohua Libc0934f2012-05-22 13:55:05 +10001444 for (i = disks; i--; ) {
Tejun Heoe9c74692010-09-03 11:56:18 +02001445 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
Shaohua Libc0934f2012-05-22 13:55:05 +10001446 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
Shaohua Li9e4447682012-10-11 13:49:49 +11001447 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
Shaohua Libc0934f2012-05-22 13:55:05 +10001448 }
Tejun Heoe9c74692010-09-03 11:56:18 +02001449
Dan Williams91c00922007-01-02 13:52:30 -07001450 for (i = disks; i--; ) {
1451 struct r5dev *dev = &sh->dev[i];
Dan Williamsac6b53b2009-07-14 13:40:19 -07001452
Tejun Heoe9c74692010-09-03 11:56:18 +02001453 if (dev->written || i == pd_idx || i == qd_idx) {
Shaohua Lid592a992014-05-21 17:57:44 +08001454 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
Shaohua Li9e4447682012-10-11 13:49:49 +11001455 set_bit(R5_UPTODATE, &dev->flags);
Tejun Heoe9c74692010-09-03 11:56:18 +02001456 if (fua)
1457 set_bit(R5_WantFUA, &dev->flags);
Shaohua Libc0934f2012-05-22 13:55:05 +10001458 if (sync)
1459 set_bit(R5_SyncIO, &dev->flags);
Tejun Heoe9c74692010-09-03 11:56:18 +02001460 }
Dan Williams91c00922007-01-02 13:52:30 -07001461 }
1462
Dan Williamsd8ee0722008-06-28 08:32:06 +10001463 if (sh->reconstruct_state == reconstruct_state_drain_run)
1464 sh->reconstruct_state = reconstruct_state_drain_result;
1465 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1466 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1467 else {
1468 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1469 sh->reconstruct_state = reconstruct_state_result;
1470 }
Dan Williams91c00922007-01-02 13:52:30 -07001471
1472 set_bit(STRIPE_HANDLE, &sh->state);
1473 release_stripe(sh);
1474}
1475
1476static void
Dan Williamsac6b53b2009-07-14 13:40:19 -07001477ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1478 struct dma_async_tx_descriptor *tx)
Dan Williams91c00922007-01-02 13:52:30 -07001479{
Dan Williams91c00922007-01-02 13:52:30 -07001480 int disks = sh->disks;
Dan Williamsd6f38f32009-07-14 11:50:52 -07001481 struct page **xor_srcs = percpu->scribble;
Dan Williamsa08abd82009-06-03 11:43:59 -07001482 struct async_submit_ctl submit;
Dan Williams91c00922007-01-02 13:52:30 -07001483 int count = 0, pd_idx = sh->pd_idx, i;
1484 struct page *xor_dest;
Dan Williamsd8ee0722008-06-28 08:32:06 +10001485 int prexor = 0;
Dan Williams91c00922007-01-02 13:52:30 -07001486 unsigned long flags;
Dan Williams91c00922007-01-02 13:52:30 -07001487
Harvey Harrisone46b2722008-04-28 02:15:50 -07001488 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001489 (unsigned long long)sh->sector);
1490
Shaohua Li620125f2012-10-11 13:49:05 +11001491 for (i = 0; i < sh->disks; i++) {
1492 if (pd_idx == i)
1493 continue;
1494 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1495 break;
1496 }
1497 if (i >= sh->disks) {
1498 atomic_inc(&sh->count);
Shaohua Li620125f2012-10-11 13:49:05 +11001499 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1500 ops_complete_reconstruct(sh);
1501 return;
1502 }
Dan Williams91c00922007-01-02 13:52:30 -07001503 /* check if prexor is active which means only process blocks
1504 * that are part of a read-modify-write (written)
1505 */
Dan Williamsd8ee0722008-06-28 08:32:06 +10001506 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1507 prexor = 1;
Dan Williams91c00922007-01-02 13:52:30 -07001508 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1509 for (i = disks; i--; ) {
1510 struct r5dev *dev = &sh->dev[i];
1511 if (dev->written)
1512 xor_srcs[count++] = dev->page;
1513 }
1514 } else {
1515 xor_dest = sh->dev[pd_idx].page;
1516 for (i = disks; i--; ) {
1517 struct r5dev *dev = &sh->dev[i];
1518 if (i != pd_idx)
1519 xor_srcs[count++] = dev->page;
1520 }
1521 }
1522
Dan Williams91c00922007-01-02 13:52:30 -07001523 /* 1/ if we prexor'd then the dest is reused as a source
1524 * 2/ if we did not prexor then we are redoing the parity
1525 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1526 * for the synchronous xor case
1527 */
Dan Williams88ba2aa2009-04-09 16:16:18 -07001528 flags = ASYNC_TX_ACK |
Dan Williams91c00922007-01-02 13:52:30 -07001529 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1530
1531 atomic_inc(&sh->count);
1532
Dan Williamsac6b53b2009-07-14 13:40:19 -07001533 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
Dan Williamsd6f38f32009-07-14 11:50:52 -07001534 to_addr_conv(sh, percpu));
Dan Williamsa08abd82009-06-03 11:43:59 -07001535 if (unlikely(count == 1))
1536 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1537 else
1538 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
Dan Williams91c00922007-01-02 13:52:30 -07001539}
1540
Dan Williamsac6b53b2009-07-14 13:40:19 -07001541static void
1542ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1543 struct dma_async_tx_descriptor *tx)
1544{
1545 struct async_submit_ctl submit;
1546 struct page **blocks = percpu->scribble;
Shaohua Li620125f2012-10-11 13:49:05 +11001547 int count, i;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001548
1549 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1550
Shaohua Li620125f2012-10-11 13:49:05 +11001551 for (i = 0; i < sh->disks; i++) {
1552 if (sh->pd_idx == i || sh->qd_idx == i)
1553 continue;
1554 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1555 break;
1556 }
1557 if (i >= sh->disks) {
1558 atomic_inc(&sh->count);
Shaohua Li620125f2012-10-11 13:49:05 +11001559 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1560 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1561 ops_complete_reconstruct(sh);
1562 return;
1563 }
1564
Dan Williamsac6b53b2009-07-14 13:40:19 -07001565 count = set_syndrome_sources(blocks, sh);
1566
1567 atomic_inc(&sh->count);
1568
1569 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1570 sh, to_addr_conv(sh, percpu));
1571 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
Dan Williams91c00922007-01-02 13:52:30 -07001572}
1573
1574static void ops_complete_check(void *stripe_head_ref)
1575{
1576 struct stripe_head *sh = stripe_head_ref;
Dan Williams91c00922007-01-02 13:52:30 -07001577
Harvey Harrisone46b2722008-04-28 02:15:50 -07001578 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001579 (unsigned long long)sh->sector);
1580
Dan Williamsecc65c92008-06-28 08:31:57 +10001581 sh->check_state = check_state_check_result;
Dan Williams91c00922007-01-02 13:52:30 -07001582 set_bit(STRIPE_HANDLE, &sh->state);
1583 release_stripe(sh);
1584}
1585
Dan Williamsac6b53b2009-07-14 13:40:19 -07001586static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
Dan Williams91c00922007-01-02 13:52:30 -07001587{
Dan Williams91c00922007-01-02 13:52:30 -07001588 int disks = sh->disks;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001589 int pd_idx = sh->pd_idx;
1590 int qd_idx = sh->qd_idx;
1591 struct page *xor_dest;
Dan Williamsd6f38f32009-07-14 11:50:52 -07001592 struct page **xor_srcs = percpu->scribble;
Dan Williams91c00922007-01-02 13:52:30 -07001593 struct dma_async_tx_descriptor *tx;
Dan Williamsa08abd82009-06-03 11:43:59 -07001594 struct async_submit_ctl submit;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001595 int count;
1596 int i;
Dan Williams91c00922007-01-02 13:52:30 -07001597
Harvey Harrisone46b2722008-04-28 02:15:50 -07001598 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001599 (unsigned long long)sh->sector);
1600
Dan Williamsac6b53b2009-07-14 13:40:19 -07001601 count = 0;
1602 xor_dest = sh->dev[pd_idx].page;
1603 xor_srcs[count++] = xor_dest;
Dan Williams91c00922007-01-02 13:52:30 -07001604 for (i = disks; i--; ) {
Dan Williamsac6b53b2009-07-14 13:40:19 -07001605 if (i == pd_idx || i == qd_idx)
1606 continue;
1607 xor_srcs[count++] = sh->dev[i].page;
Dan Williams91c00922007-01-02 13:52:30 -07001608 }
1609
Dan Williamsd6f38f32009-07-14 11:50:52 -07001610 init_async_submit(&submit, 0, NULL, NULL, NULL,
1611 to_addr_conv(sh, percpu));
Dan Williams099f53c2009-04-08 14:28:37 -07001612 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
Dan Williamsa08abd82009-06-03 11:43:59 -07001613 &sh->ops.zero_sum_result, &submit);
Dan Williams91c00922007-01-02 13:52:30 -07001614
Dan Williams91c00922007-01-02 13:52:30 -07001615 atomic_inc(&sh->count);
Dan Williamsa08abd82009-06-03 11:43:59 -07001616 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1617 tx = async_trigger_callback(&submit);
Dan Williams91c00922007-01-02 13:52:30 -07001618}
1619
Dan Williamsac6b53b2009-07-14 13:40:19 -07001620static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1621{
1622 struct page **srcs = percpu->scribble;
1623 struct async_submit_ctl submit;
1624 int count;
1625
1626 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1627 (unsigned long long)sh->sector, checkp);
1628
1629 count = set_syndrome_sources(srcs, sh);
1630 if (!checkp)
1631 srcs[count] = NULL;
1632
1633 atomic_inc(&sh->count);
1634 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1635 sh, to_addr_conv(sh, percpu));
1636 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1637 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1638}
1639
NeilBrown51acbce2013-02-28 09:08:34 +11001640static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
Dan Williams91c00922007-01-02 13:52:30 -07001641{
1642 int overlap_clear = 0, i, disks = sh->disks;
1643 struct dma_async_tx_descriptor *tx = NULL;
NeilBrownd1688a62011-10-11 16:49:52 +11001644 struct r5conf *conf = sh->raid_conf;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001645 int level = conf->level;
Dan Williamsd6f38f32009-07-14 11:50:52 -07001646 struct raid5_percpu *percpu;
1647 unsigned long cpu;
Dan Williams91c00922007-01-02 13:52:30 -07001648
Dan Williamsd6f38f32009-07-14 11:50:52 -07001649 cpu = get_cpu();
1650 percpu = per_cpu_ptr(conf->percpu, cpu);
Dan Williams83de75c2008-06-28 08:31:58 +10001651 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
Dan Williams91c00922007-01-02 13:52:30 -07001652 ops_run_biofill(sh);
1653 overlap_clear++;
1654 }
1655
Dan Williams7b3a8712008-06-28 08:32:09 +10001656 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
Dan Williamsac6b53b2009-07-14 13:40:19 -07001657 if (level < 6)
1658 tx = ops_run_compute5(sh, percpu);
1659 else {
1660 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1661 tx = ops_run_compute6_1(sh, percpu);
1662 else
1663 tx = ops_run_compute6_2(sh, percpu);
1664 }
1665 /* terminate the chain if reconstruct is not set to be run */
1666 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
Dan Williams7b3a8712008-06-28 08:32:09 +10001667 async_tx_ack(tx);
1668 }
Dan Williams91c00922007-01-02 13:52:30 -07001669
Dan Williams600aa102008-06-28 08:32:05 +10001670 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
Dan Williamsd6f38f32009-07-14 11:50:52 -07001671 tx = ops_run_prexor(sh, percpu, tx);
Dan Williams91c00922007-01-02 13:52:30 -07001672
Dan Williams600aa102008-06-28 08:32:05 +10001673 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
Dan Williamsd8ee0722008-06-28 08:32:06 +10001674 tx = ops_run_biodrain(sh, tx);
Dan Williams91c00922007-01-02 13:52:30 -07001675 overlap_clear++;
1676 }
1677
Dan Williamsac6b53b2009-07-14 13:40:19 -07001678 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1679 if (level < 6)
1680 ops_run_reconstruct5(sh, percpu, tx);
1681 else
1682 ops_run_reconstruct6(sh, percpu, tx);
1683 }
Dan Williams91c00922007-01-02 13:52:30 -07001684
Dan Williamsac6b53b2009-07-14 13:40:19 -07001685 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1686 if (sh->check_state == check_state_run)
1687 ops_run_check_p(sh, percpu);
1688 else if (sh->check_state == check_state_run_q)
1689 ops_run_check_pq(sh, percpu, 0);
1690 else if (sh->check_state == check_state_run_pq)
1691 ops_run_check_pq(sh, percpu, 1);
1692 else
1693 BUG();
1694 }
Dan Williams91c00922007-01-02 13:52:30 -07001695
Dan Williams91c00922007-01-02 13:52:30 -07001696 if (overlap_clear)
1697 for (i = disks; i--; ) {
1698 struct r5dev *dev = &sh->dev[i];
1699 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1700 wake_up(&sh->raid_conf->wait_for_overlap);
1701 }
Dan Williamsd6f38f32009-07-14 11:50:52 -07001702 put_cpu();
Dan Williams91c00922007-01-02 13:52:30 -07001703}
1704
Shaohua Li566c09c2013-11-14 15:16:17 +11001705static int grow_one_stripe(struct r5conf *conf, int hash)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706{
1707 struct stripe_head *sh;
Namhyung Kim6ce32842011-07-18 17:38:50 +10001708 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
NeilBrown3f294f42005-11-08 21:39:25 -08001709 if (!sh)
1710 return 0;
Namhyung Kim6ce32842011-07-18 17:38:50 +10001711
NeilBrown3f294f42005-11-08 21:39:25 -08001712 sh->raid_conf = conf;
NeilBrown3f294f42005-11-08 21:39:25 -08001713
Shaohua Lib17459c2012-07-19 16:01:31 +10001714 spin_lock_init(&sh->stripe_lock);
1715
NeilBrowne4e11e32010-06-16 16:45:16 +10001716 if (grow_buffers(sh)) {
1717 shrink_buffers(sh);
NeilBrown3f294f42005-11-08 21:39:25 -08001718 kmem_cache_free(conf->slab_cache, sh);
1719 return 0;
1720 }
Shaohua Li566c09c2013-11-14 15:16:17 +11001721 sh->hash_lock_index = hash;
NeilBrown3f294f42005-11-08 21:39:25 -08001722 /* we just created an active stripe so... */
1723 atomic_set(&sh->count, 1);
1724 atomic_inc(&conf->active_stripes);
1725 INIT_LIST_HEAD(&sh->lru);
1726 release_stripe(sh);
1727 return 1;
1728}
1729
NeilBrownd1688a62011-10-11 16:49:52 +11001730static int grow_stripes(struct r5conf *conf, int num)
NeilBrown3f294f42005-11-08 21:39:25 -08001731{
Christoph Lametere18b8902006-12-06 20:33:20 -08001732 struct kmem_cache *sc;
NeilBrown5e5e3e72009-10-16 16:35:30 +11001733 int devs = max(conf->raid_disks, conf->previous_raid_disks);
Shaohua Li566c09c2013-11-14 15:16:17 +11001734 int hash;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735
NeilBrownf4be6b42010-06-01 19:37:25 +10001736 if (conf->mddev->gendisk)
1737 sprintf(conf->cache_name[0],
1738 "raid%d-%s", conf->level, mdname(conf->mddev));
1739 else
1740 sprintf(conf->cache_name[0],
1741 "raid%d-%p", conf->level, conf->mddev);
1742 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1743
NeilBrownad01c9e2006-03-27 01:18:07 -08001744 conf->active_name = 0;
1745 sc = kmem_cache_create(conf->cache_name[conf->active_name],
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
Paul Mundt20c2df82007-07-20 10:11:58 +09001747 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 if (!sc)
1749 return 1;
1750 conf->slab_cache = sc;
NeilBrownad01c9e2006-03-27 01:18:07 -08001751 conf->pool_size = devs;
Shaohua Li566c09c2013-11-14 15:16:17 +11001752 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
1753 while (num--) {
1754 if (!grow_one_stripe(conf, hash))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755 return 1;
Shaohua Li566c09c2013-11-14 15:16:17 +11001756 conf->max_nr_stripes++;
1757 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
1758 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759 return 0;
1760}
NeilBrown29269552006-03-27 01:18:10 -08001761
Dan Williamsd6f38f32009-07-14 11:50:52 -07001762/**
1763 * scribble_len - return the required size of the scribble region
1764 * @num - total number of disks in the array
1765 *
1766 * The size must be enough to contain:
1767 * 1/ a struct page pointer for each device in the array +2
1768 * 2/ room to convert each entry in (1) to its corresponding dma
1769 * (dma_map_page()) or page (page_address()) address.
1770 *
1771 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1772 * calculate over all devices (not just the data blocks), using zeros in place
1773 * of the P and Q blocks.
1774 */
1775static size_t scribble_len(int num)
1776{
1777 size_t len;
1778
1779 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1780
1781 return len;
1782}
1783
NeilBrownd1688a62011-10-11 16:49:52 +11001784static int resize_stripes(struct r5conf *conf, int newsize)
NeilBrownad01c9e2006-03-27 01:18:07 -08001785{
1786 /* Make all the stripes able to hold 'newsize' devices.
1787 * New slots in each stripe get 'page' set to a new page.
1788 *
1789 * This happens in stages:
1790 * 1/ create a new kmem_cache and allocate the required number of
1791 * stripe_heads.
Masanari Iida83f0d772012-10-30 00:18:08 +09001792 * 2/ gather all the old stripe_heads and transfer the pages across
NeilBrownad01c9e2006-03-27 01:18:07 -08001793 * to the new stripe_heads. This will have the side effect of
1794 * freezing the array as once all stripe_heads have been collected,
1795 * no IO will be possible. Old stripe heads are freed once their
1796 * pages have been transferred over, and the old kmem_cache is
1797 * freed when all stripes are done.
1798 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1799 * we simple return a failre status - no need to clean anything up.
1800 * 4/ allocate new pages for the new slots in the new stripe_heads.
1801 * If this fails, we don't bother trying the shrink the
1802 * stripe_heads down again, we just leave them as they are.
1803 * As each stripe_head is processed the new one is released into
1804 * active service.
1805 *
1806 * Once step2 is started, we cannot afford to wait for a write,
1807 * so we use GFP_NOIO allocations.
1808 */
1809 struct stripe_head *osh, *nsh;
1810 LIST_HEAD(newstripes);
1811 struct disk_info *ndisks;
Dan Williamsd6f38f32009-07-14 11:50:52 -07001812 unsigned long cpu;
Dan Williamsb5470dc2008-06-27 21:44:04 -07001813 int err;
Christoph Lametere18b8902006-12-06 20:33:20 -08001814 struct kmem_cache *sc;
NeilBrownad01c9e2006-03-27 01:18:07 -08001815 int i;
Shaohua Li566c09c2013-11-14 15:16:17 +11001816 int hash, cnt;
NeilBrownad01c9e2006-03-27 01:18:07 -08001817
1818 if (newsize <= conf->pool_size)
1819 return 0; /* never bother to shrink */
1820
Dan Williamsb5470dc2008-06-27 21:44:04 -07001821 err = md_allow_write(conf->mddev);
1822 if (err)
1823 return err;
NeilBrown2a2275d2007-01-26 00:57:11 -08001824
NeilBrownad01c9e2006-03-27 01:18:07 -08001825 /* Step 1 */
1826 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1827 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
Paul Mundt20c2df82007-07-20 10:11:58 +09001828 0, 0, NULL);
NeilBrownad01c9e2006-03-27 01:18:07 -08001829 if (!sc)
1830 return -ENOMEM;
1831
1832 for (i = conf->max_nr_stripes; i; i--) {
Namhyung Kim6ce32842011-07-18 17:38:50 +10001833 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
NeilBrownad01c9e2006-03-27 01:18:07 -08001834 if (!nsh)
1835 break;
1836
NeilBrownad01c9e2006-03-27 01:18:07 -08001837 nsh->raid_conf = conf;
NeilBrowncb13ff62012-09-24 16:27:20 +10001838 spin_lock_init(&nsh->stripe_lock);
NeilBrownad01c9e2006-03-27 01:18:07 -08001839
1840 list_add(&nsh->lru, &newstripes);
1841 }
1842 if (i) {
1843 /* didn't get enough, give up */
1844 while (!list_empty(&newstripes)) {
1845 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1846 list_del(&nsh->lru);
1847 kmem_cache_free(sc, nsh);
1848 }
1849 kmem_cache_destroy(sc);
1850 return -ENOMEM;
1851 }
1852 /* Step 2 - Must use GFP_NOIO now.
1853 * OK, we have enough stripes, start collecting inactive
1854 * stripes and copying them over
1855 */
Shaohua Li566c09c2013-11-14 15:16:17 +11001856 hash = 0;
1857 cnt = 0;
NeilBrownad01c9e2006-03-27 01:18:07 -08001858 list_for_each_entry(nsh, &newstripes, lru) {
Shaohua Li566c09c2013-11-14 15:16:17 +11001859 lock_device_hash_lock(conf, hash);
1860 wait_event_cmd(conf->wait_for_stripe,
1861 !list_empty(conf->inactive_list + hash),
1862 unlock_device_hash_lock(conf, hash),
1863 lock_device_hash_lock(conf, hash));
1864 osh = get_free_stripe(conf, hash);
1865 unlock_device_hash_lock(conf, hash);
NeilBrownad01c9e2006-03-27 01:18:07 -08001866 atomic_set(&nsh->count, 1);
Shaohua Lid592a992014-05-21 17:57:44 +08001867 for(i=0; i<conf->pool_size; i++) {
NeilBrownad01c9e2006-03-27 01:18:07 -08001868 nsh->dev[i].page = osh->dev[i].page;
Shaohua Lid592a992014-05-21 17:57:44 +08001869 nsh->dev[i].orig_page = osh->dev[i].page;
1870 }
NeilBrownad01c9e2006-03-27 01:18:07 -08001871 for( ; i<newsize; i++)
1872 nsh->dev[i].page = NULL;
Shaohua Li566c09c2013-11-14 15:16:17 +11001873 nsh->hash_lock_index = hash;
NeilBrownad01c9e2006-03-27 01:18:07 -08001874 kmem_cache_free(conf->slab_cache, osh);
Shaohua Li566c09c2013-11-14 15:16:17 +11001875 cnt++;
1876 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
1877 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
1878 hash++;
1879 cnt = 0;
1880 }
NeilBrownad01c9e2006-03-27 01:18:07 -08001881 }
1882 kmem_cache_destroy(conf->slab_cache);
1883
1884 /* Step 3.
1885 * At this point, we are holding all the stripes so the array
1886 * is completely stalled, so now is a good time to resize
Dan Williamsd6f38f32009-07-14 11:50:52 -07001887 * conf->disks and the scribble region
NeilBrownad01c9e2006-03-27 01:18:07 -08001888 */
1889 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1890 if (ndisks) {
1891 for (i=0; i<conf->raid_disks; i++)
1892 ndisks[i] = conf->disks[i];
1893 kfree(conf->disks);
1894 conf->disks = ndisks;
1895 } else
1896 err = -ENOMEM;
1897
Dan Williamsd6f38f32009-07-14 11:50:52 -07001898 get_online_cpus();
1899 conf->scribble_len = scribble_len(newsize);
1900 for_each_present_cpu(cpu) {
1901 struct raid5_percpu *percpu;
1902 void *scribble;
1903
1904 percpu = per_cpu_ptr(conf->percpu, cpu);
1905 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1906
1907 if (scribble) {
1908 kfree(percpu->scribble);
1909 percpu->scribble = scribble;
1910 } else {
1911 err = -ENOMEM;
1912 break;
1913 }
1914 }
1915 put_online_cpus();
1916
NeilBrownad01c9e2006-03-27 01:18:07 -08001917 /* Step 4, return new stripes to service */
1918 while(!list_empty(&newstripes)) {
1919 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1920 list_del_init(&nsh->lru);
Dan Williamsd6f38f32009-07-14 11:50:52 -07001921
NeilBrownad01c9e2006-03-27 01:18:07 -08001922 for (i=conf->raid_disks; i < newsize; i++)
1923 if (nsh->dev[i].page == NULL) {
1924 struct page *p = alloc_page(GFP_NOIO);
1925 nsh->dev[i].page = p;
Shaohua Lid592a992014-05-21 17:57:44 +08001926 nsh->dev[i].orig_page = p;
NeilBrownad01c9e2006-03-27 01:18:07 -08001927 if (!p)
1928 err = -ENOMEM;
1929 }
1930 release_stripe(nsh);
1931 }
1932 /* critical section pass, GFP_NOIO no longer needed */
1933
1934 conf->slab_cache = sc;
1935 conf->active_name = 1-conf->active_name;
1936 conf->pool_size = newsize;
1937 return err;
1938}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939
Shaohua Li566c09c2013-11-14 15:16:17 +11001940static int drop_one_stripe(struct r5conf *conf, int hash)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941{
1942 struct stripe_head *sh;
1943
Shaohua Li566c09c2013-11-14 15:16:17 +11001944 spin_lock_irq(conf->hash_locks + hash);
1945 sh = get_free_stripe(conf, hash);
1946 spin_unlock_irq(conf->hash_locks + hash);
NeilBrown3f294f42005-11-08 21:39:25 -08001947 if (!sh)
1948 return 0;
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +02001949 BUG_ON(atomic_read(&sh->count));
NeilBrowne4e11e32010-06-16 16:45:16 +10001950 shrink_buffers(sh);
NeilBrown3f294f42005-11-08 21:39:25 -08001951 kmem_cache_free(conf->slab_cache, sh);
1952 atomic_dec(&conf->active_stripes);
1953 return 1;
1954}
1955
NeilBrownd1688a62011-10-11 16:49:52 +11001956static void shrink_stripes(struct r5conf *conf)
NeilBrown3f294f42005-11-08 21:39:25 -08001957{
Shaohua Li566c09c2013-11-14 15:16:17 +11001958 int hash;
1959 for (hash = 0; hash < NR_STRIPE_HASH_LOCKS; hash++)
1960 while (drop_one_stripe(conf, hash))
1961 ;
NeilBrown3f294f42005-11-08 21:39:25 -08001962
NeilBrown29fc7e32006-02-03 03:03:41 -08001963 if (conf->slab_cache)
1964 kmem_cache_destroy(conf->slab_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 conf->slab_cache = NULL;
1966}
1967
NeilBrown6712ecf2007-09-27 12:47:43 +02001968static void raid5_end_read_request(struct bio * bi, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969{
NeilBrown99c0fb52009-03-31 14:39:38 +11001970 struct stripe_head *sh = bi->bi_private;
NeilBrownd1688a62011-10-11 16:49:52 +11001971 struct r5conf *conf = sh->raid_conf;
NeilBrown7ecaa1e2006-03-27 01:18:08 -08001972 int disks = sh->disks, i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
NeilBrownd6950432006-07-10 04:44:20 -07001974 char b[BDEVNAME_SIZE];
NeilBrowndd054fc2011-12-23 10:17:53 +11001975 struct md_rdev *rdev = NULL;
NeilBrown05616be2012-05-21 09:27:00 +10001976 sector_t s;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977
1978 for (i=0 ; i<disks; i++)
1979 if (bi == &sh->dev[i].req)
1980 break;
1981
Dan Williams45b42332007-07-09 11:56:43 -07001982 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1983 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984 uptodate);
1985 if (i == disks) {
1986 BUG();
NeilBrown6712ecf2007-09-27 12:47:43 +02001987 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 }
NeilBrown14a75d32011-12-23 10:17:52 +11001989 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
NeilBrowndd054fc2011-12-23 10:17:53 +11001990 /* If replacement finished while this request was outstanding,
1991 * 'replacement' might be NULL already.
1992 * In that case it moved down to 'rdev'.
1993 * rdev is not removed until all requests are finished.
1994 */
NeilBrown14a75d32011-12-23 10:17:52 +11001995 rdev = conf->disks[i].replacement;
NeilBrowndd054fc2011-12-23 10:17:53 +11001996 if (!rdev)
NeilBrown14a75d32011-12-23 10:17:52 +11001997 rdev = conf->disks[i].rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998
NeilBrown05616be2012-05-21 09:27:00 +10001999 if (use_new_offset(conf, sh))
2000 s = sh->sector + rdev->new_data_offset;
2001 else
2002 s = sh->sector + rdev->data_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004 set_bit(R5_UPTODATE, &sh->dev[i].flags);
NeilBrown4e5314b2005-11-08 21:39:22 -08002005 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
NeilBrown14a75d32011-12-23 10:17:52 +11002006 /* Note that this cannot happen on a
2007 * replacement device. We just fail those on
2008 * any error
2009 */
Christian Dietrich8bda4702011-07-27 11:00:36 +10002010 printk_ratelimited(
2011 KERN_INFO
2012 "md/raid:%s: read error corrected"
2013 " (%lu sectors at %llu on %s)\n",
2014 mdname(conf->mddev), STRIPE_SECTORS,
NeilBrown05616be2012-05-21 09:27:00 +10002015 (unsigned long long)s,
Christian Dietrich8bda4702011-07-27 11:00:36 +10002016 bdevname(rdev->bdev, b));
Namhyung Kimddd51152011-07-27 11:00:36 +10002017 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
NeilBrown4e5314b2005-11-08 21:39:22 -08002018 clear_bit(R5_ReadError, &sh->dev[i].flags);
2019 clear_bit(R5_ReWrite, &sh->dev[i].flags);
majianpeng3f9e7c12012-07-31 10:04:21 +10002020 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2021 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2022
NeilBrown14a75d32011-12-23 10:17:52 +11002023 if (atomic_read(&rdev->read_errors))
2024 atomic_set(&rdev->read_errors, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025 } else {
NeilBrown14a75d32011-12-23 10:17:52 +11002026 const char *bdn = bdevname(rdev->bdev, b);
NeilBrownba22dcb2005-11-08 21:39:31 -08002027 int retry = 0;
majianpeng2e8ac3032012-07-03 15:57:02 +10002028 int set_bad = 0;
NeilBrownd6950432006-07-10 04:44:20 -07002029
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
NeilBrownd6950432006-07-10 04:44:20 -07002031 atomic_inc(&rdev->read_errors);
NeilBrown14a75d32011-12-23 10:17:52 +11002032 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2033 printk_ratelimited(
2034 KERN_WARNING
2035 "md/raid:%s: read error on replacement device "
2036 "(sector %llu on %s).\n",
2037 mdname(conf->mddev),
NeilBrown05616be2012-05-21 09:27:00 +10002038 (unsigned long long)s,
NeilBrown14a75d32011-12-23 10:17:52 +11002039 bdn);
majianpeng2e8ac3032012-07-03 15:57:02 +10002040 else if (conf->mddev->degraded >= conf->max_degraded) {
2041 set_bad = 1;
Christian Dietrich8bda4702011-07-27 11:00:36 +10002042 printk_ratelimited(
2043 KERN_WARNING
2044 "md/raid:%s: read error not correctable "
2045 "(sector %llu on %s).\n",
2046 mdname(conf->mddev),
NeilBrown05616be2012-05-21 09:27:00 +10002047 (unsigned long long)s,
Christian Dietrich8bda4702011-07-27 11:00:36 +10002048 bdn);
majianpeng2e8ac3032012-07-03 15:57:02 +10002049 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
NeilBrown4e5314b2005-11-08 21:39:22 -08002050 /* Oh, no!!! */
majianpeng2e8ac3032012-07-03 15:57:02 +10002051 set_bad = 1;
Christian Dietrich8bda4702011-07-27 11:00:36 +10002052 printk_ratelimited(
2053 KERN_WARNING
2054 "md/raid:%s: read error NOT corrected!! "
2055 "(sector %llu on %s).\n",
2056 mdname(conf->mddev),
NeilBrown05616be2012-05-21 09:27:00 +10002057 (unsigned long long)s,
Christian Dietrich8bda4702011-07-27 11:00:36 +10002058 bdn);
majianpeng2e8ac3032012-07-03 15:57:02 +10002059 } else if (atomic_read(&rdev->read_errors)
NeilBrownba22dcb2005-11-08 21:39:31 -08002060 > conf->max_nr_stripes)
NeilBrown14f8d262006-01-06 00:20:14 -08002061 printk(KERN_WARNING
NeilBrown0c55e022010-05-03 14:09:02 +10002062 "md/raid:%s: Too many read errors, failing device %s.\n",
NeilBrownd6950432006-07-10 04:44:20 -07002063 mdname(conf->mddev), bdn);
NeilBrownba22dcb2005-11-08 21:39:31 -08002064 else
2065 retry = 1;
Bian Yuedfa1f62013-11-14 15:16:17 +11002066 if (set_bad && test_bit(In_sync, &rdev->flags)
2067 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2068 retry = 1;
NeilBrownba22dcb2005-11-08 21:39:31 -08002069 if (retry)
majianpeng3f9e7c12012-07-31 10:04:21 +10002070 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2071 set_bit(R5_ReadError, &sh->dev[i].flags);
2072 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2073 } else
2074 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
NeilBrownba22dcb2005-11-08 21:39:31 -08002075 else {
NeilBrown4e5314b2005-11-08 21:39:22 -08002076 clear_bit(R5_ReadError, &sh->dev[i].flags);
2077 clear_bit(R5_ReWrite, &sh->dev[i].flags);
majianpeng2e8ac3032012-07-03 15:57:02 +10002078 if (!(set_bad
2079 && test_bit(In_sync, &rdev->flags)
2080 && rdev_set_badblocks(
2081 rdev, sh->sector, STRIPE_SECTORS, 0)))
2082 md_error(conf->mddev, rdev);
NeilBrownba22dcb2005-11-08 21:39:31 -08002083 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084 }
NeilBrown14a75d32011-12-23 10:17:52 +11002085 rdev_dec_pending(rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2087 set_bit(STRIPE_HANDLE, &sh->state);
2088 release_stripe(sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089}
2090
NeilBrownd710e132008-10-13 11:55:12 +11002091static void raid5_end_write_request(struct bio *bi, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092{
NeilBrown99c0fb52009-03-31 14:39:38 +11002093 struct stripe_head *sh = bi->bi_private;
NeilBrownd1688a62011-10-11 16:49:52 +11002094 struct r5conf *conf = sh->raid_conf;
NeilBrown7ecaa1e2006-03-27 01:18:08 -08002095 int disks = sh->disks, i;
NeilBrown977df362011-12-23 10:17:53 +11002096 struct md_rdev *uninitialized_var(rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
NeilBrownb84db562011-07-28 11:39:23 +10002098 sector_t first_bad;
2099 int bad_sectors;
NeilBrown977df362011-12-23 10:17:53 +11002100 int replacement = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101
NeilBrown977df362011-12-23 10:17:53 +11002102 for (i = 0 ; i < disks; i++) {
2103 if (bi == &sh->dev[i].req) {
2104 rdev = conf->disks[i].rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105 break;
NeilBrown977df362011-12-23 10:17:53 +11002106 }
2107 if (bi == &sh->dev[i].rreq) {
2108 rdev = conf->disks[i].replacement;
NeilBrowndd054fc2011-12-23 10:17:53 +11002109 if (rdev)
2110 replacement = 1;
2111 else
2112 /* rdev was removed and 'replacement'
2113 * replaced it. rdev is not removed
2114 * until all requests are finished.
2115 */
2116 rdev = conf->disks[i].rdev;
NeilBrown977df362011-12-23 10:17:53 +11002117 break;
2118 }
2119 }
Dan Williams45b42332007-07-09 11:56:43 -07002120 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2122 uptodate);
2123 if (i == disks) {
2124 BUG();
NeilBrown6712ecf2007-09-27 12:47:43 +02002125 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 }
2127
NeilBrown977df362011-12-23 10:17:53 +11002128 if (replacement) {
2129 if (!uptodate)
2130 md_error(conf->mddev, rdev);
2131 else if (is_badblock(rdev, sh->sector,
2132 STRIPE_SECTORS,
2133 &first_bad, &bad_sectors))
2134 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2135 } else {
2136 if (!uptodate) {
NeilBrown9f97e4b2014-01-16 09:35:38 +11002137 set_bit(STRIPE_DEGRADED, &sh->state);
NeilBrown977df362011-12-23 10:17:53 +11002138 set_bit(WriteErrorSeen, &rdev->flags);
2139 set_bit(R5_WriteError, &sh->dev[i].flags);
NeilBrown3a6de292011-12-23 10:17:54 +11002140 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2141 set_bit(MD_RECOVERY_NEEDED,
2142 &rdev->mddev->recovery);
NeilBrown977df362011-12-23 10:17:53 +11002143 } else if (is_badblock(rdev, sh->sector,
2144 STRIPE_SECTORS,
NeilBrownc0b32972013-04-24 11:42:42 +10002145 &first_bad, &bad_sectors)) {
NeilBrown977df362011-12-23 10:17:53 +11002146 set_bit(R5_MadeGood, &sh->dev[i].flags);
NeilBrownc0b32972013-04-24 11:42:42 +10002147 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2148 /* That was a successful write so make
2149 * sure it looks like we already did
2150 * a re-write.
2151 */
2152 set_bit(R5_ReWrite, &sh->dev[i].flags);
2153 }
NeilBrown977df362011-12-23 10:17:53 +11002154 }
2155 rdev_dec_pending(rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156
NeilBrown977df362011-12-23 10:17:53 +11002157 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2158 clear_bit(R5_LOCKED, &sh->dev[i].flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 set_bit(STRIPE_HANDLE, &sh->state);
NeilBrownc04be0a2006-10-03 01:15:53 -07002160 release_stripe(sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161}
2162
NeilBrown784052e2009-03-31 15:19:07 +11002163static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
Shaohua Lid592a992014-05-21 17:57:44 +08002164
NeilBrown784052e2009-03-31 15:19:07 +11002165static void raid5_build_block(struct stripe_head *sh, int i, int previous)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166{
2167 struct r5dev *dev = &sh->dev[i];
2168
2169 bio_init(&dev->req);
2170 dev->req.bi_io_vec = &dev->vec;
Shaohua Lid592a992014-05-21 17:57:44 +08002171 dev->req.bi_max_vecs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172 dev->req.bi_private = sh;
2173
NeilBrown977df362011-12-23 10:17:53 +11002174 bio_init(&dev->rreq);
2175 dev->rreq.bi_io_vec = &dev->rvec;
Shaohua Lid592a992014-05-21 17:57:44 +08002176 dev->rreq.bi_max_vecs = 1;
NeilBrown977df362011-12-23 10:17:53 +11002177 dev->rreq.bi_private = sh;
NeilBrown977df362011-12-23 10:17:53 +11002178
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 dev->flags = 0;
NeilBrown784052e2009-03-31 15:19:07 +11002180 dev->sector = compute_blocknr(sh, i, previous);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181}
2182
NeilBrownfd01b882011-10-11 16:47:53 +11002183static void error(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184{
2185 char b[BDEVNAME_SIZE];
NeilBrownd1688a62011-10-11 16:49:52 +11002186 struct r5conf *conf = mddev->private;
NeilBrown908f4fb2011-12-23 10:17:50 +11002187 unsigned long flags;
NeilBrown0c55e022010-05-03 14:09:02 +10002188 pr_debug("raid456: error called\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189
NeilBrown908f4fb2011-12-23 10:17:50 +11002190 spin_lock_irqsave(&conf->device_lock, flags);
2191 clear_bit(In_sync, &rdev->flags);
2192 mddev->degraded = calc_degraded(conf);
2193 spin_unlock_irqrestore(&conf->device_lock, flags);
2194 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2195
NeilBrownde393cd2011-07-28 11:31:48 +10002196 set_bit(Blocked, &rdev->flags);
NeilBrown6f8d0c72011-05-11 14:38:44 +10002197 set_bit(Faulty, &rdev->flags);
2198 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2199 printk(KERN_ALERT
2200 "md/raid:%s: Disk failure on %s, disabling device.\n"
2201 "md/raid:%s: Operation continuing on %d devices.\n",
2202 mdname(mddev),
2203 bdevname(rdev->bdev, b),
2204 mdname(mddev),
2205 conf->raid_disks - mddev->degraded);
NeilBrown16a53ec2006-06-26 00:27:38 -07002206}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207
2208/*
2209 * Input: a 'big' sector number,
2210 * Output: index of the data and parity disk, and the sector # in them.
2211 */
NeilBrownd1688a62011-10-11 16:49:52 +11002212static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
NeilBrown911d4ee2009-03-31 14:39:38 +11002213 int previous, int *dd_idx,
2214 struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215{
NeilBrown6e3b96e2010-04-23 07:08:28 +10002216 sector_t stripe, stripe2;
NeilBrown35f2a592010-04-20 14:13:34 +10002217 sector_t chunk_number;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 unsigned int chunk_offset;
NeilBrown911d4ee2009-03-31 14:39:38 +11002219 int pd_idx, qd_idx;
NeilBrown67cc2b82009-03-31 14:39:38 +11002220 int ddf_layout = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 sector_t new_sector;
NeilBrowne183eae2009-03-31 15:20:22 +11002222 int algorithm = previous ? conf->prev_algo
2223 : conf->algorithm;
Andre Noll09c9e5f2009-06-18 08:45:55 +10002224 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2225 : conf->chunk_sectors;
NeilBrown112bf892009-03-31 14:39:38 +11002226 int raid_disks = previous ? conf->previous_raid_disks
2227 : conf->raid_disks;
2228 int data_disks = raid_disks - conf->max_degraded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229
2230 /* First compute the information on this sector */
2231
2232 /*
2233 * Compute the chunk number and the sector offset inside the chunk
2234 */
2235 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2236 chunk_number = r_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237
2238 /*
2239 * Compute the stripe number
2240 */
NeilBrown35f2a592010-04-20 14:13:34 +10002241 stripe = chunk_number;
2242 *dd_idx = sector_div(stripe, data_disks);
NeilBrown6e3b96e2010-04-23 07:08:28 +10002243 stripe2 = stripe;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 /*
2245 * Select the parity disk based on the user selected algorithm.
2246 */
NeilBrown84789552011-07-27 11:00:36 +10002247 pd_idx = qd_idx = -1;
NeilBrown16a53ec2006-06-26 00:27:38 -07002248 switch(conf->level) {
2249 case 4:
NeilBrown911d4ee2009-03-31 14:39:38 +11002250 pd_idx = data_disks;
NeilBrown16a53ec2006-06-26 00:27:38 -07002251 break;
2252 case 5:
NeilBrowne183eae2009-03-31 15:20:22 +11002253 switch (algorithm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254 case ALGORITHM_LEFT_ASYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002255 pd_idx = data_disks - sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002256 if (*dd_idx >= pd_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 (*dd_idx)++;
2258 break;
2259 case ALGORITHM_RIGHT_ASYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002260 pd_idx = sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002261 if (*dd_idx >= pd_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 (*dd_idx)++;
2263 break;
2264 case ALGORITHM_LEFT_SYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002265 pd_idx = data_disks - sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002266 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 break;
2268 case ALGORITHM_RIGHT_SYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002269 pd_idx = sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002270 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 break;
NeilBrown99c0fb52009-03-31 14:39:38 +11002272 case ALGORITHM_PARITY_0:
2273 pd_idx = 0;
2274 (*dd_idx)++;
2275 break;
2276 case ALGORITHM_PARITY_N:
2277 pd_idx = data_disks;
2278 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279 default:
NeilBrown99c0fb52009-03-31 14:39:38 +11002280 BUG();
NeilBrown16a53ec2006-06-26 00:27:38 -07002281 }
2282 break;
2283 case 6:
2284
NeilBrowne183eae2009-03-31 15:20:22 +11002285 switch (algorithm) {
NeilBrown16a53ec2006-06-26 00:27:38 -07002286 case ALGORITHM_LEFT_ASYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002287 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002288 qd_idx = pd_idx + 1;
2289 if (pd_idx == raid_disks-1) {
NeilBrown99c0fb52009-03-31 14:39:38 +11002290 (*dd_idx)++; /* Q D D D P */
NeilBrown911d4ee2009-03-31 14:39:38 +11002291 qd_idx = 0;
2292 } else if (*dd_idx >= pd_idx)
NeilBrown16a53ec2006-06-26 00:27:38 -07002293 (*dd_idx) += 2; /* D D P Q D */
2294 break;
2295 case ALGORITHM_RIGHT_ASYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002296 pd_idx = sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002297 qd_idx = pd_idx + 1;
2298 if (pd_idx == raid_disks-1) {
NeilBrown99c0fb52009-03-31 14:39:38 +11002299 (*dd_idx)++; /* Q D D D P */
NeilBrown911d4ee2009-03-31 14:39:38 +11002300 qd_idx = 0;
2301 } else if (*dd_idx >= pd_idx)
NeilBrown16a53ec2006-06-26 00:27:38 -07002302 (*dd_idx) += 2; /* D D P Q D */
2303 break;
2304 case ALGORITHM_LEFT_SYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002305 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002306 qd_idx = (pd_idx + 1) % raid_disks;
2307 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
NeilBrown16a53ec2006-06-26 00:27:38 -07002308 break;
2309 case ALGORITHM_RIGHT_SYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002310 pd_idx = sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002311 qd_idx = (pd_idx + 1) % raid_disks;
2312 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
NeilBrown16a53ec2006-06-26 00:27:38 -07002313 break;
NeilBrown99c0fb52009-03-31 14:39:38 +11002314
2315 case ALGORITHM_PARITY_0:
2316 pd_idx = 0;
2317 qd_idx = 1;
2318 (*dd_idx) += 2;
2319 break;
2320 case ALGORITHM_PARITY_N:
2321 pd_idx = data_disks;
2322 qd_idx = data_disks + 1;
2323 break;
2324
2325 case ALGORITHM_ROTATING_ZERO_RESTART:
2326 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2327 * of blocks for computing Q is different.
2328 */
NeilBrown6e3b96e2010-04-23 07:08:28 +10002329 pd_idx = sector_div(stripe2, raid_disks);
NeilBrown99c0fb52009-03-31 14:39:38 +11002330 qd_idx = pd_idx + 1;
2331 if (pd_idx == raid_disks-1) {
2332 (*dd_idx)++; /* Q D D D P */
2333 qd_idx = 0;
2334 } else if (*dd_idx >= pd_idx)
2335 (*dd_idx) += 2; /* D D P Q D */
NeilBrown67cc2b82009-03-31 14:39:38 +11002336 ddf_layout = 1;
NeilBrown99c0fb52009-03-31 14:39:38 +11002337 break;
2338
2339 case ALGORITHM_ROTATING_N_RESTART:
2340 /* Same a left_asymmetric, by first stripe is
2341 * D D D P Q rather than
2342 * Q D D D P
2343 */
NeilBrown6e3b96e2010-04-23 07:08:28 +10002344 stripe2 += 1;
2345 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
NeilBrown99c0fb52009-03-31 14:39:38 +11002346 qd_idx = pd_idx + 1;
2347 if (pd_idx == raid_disks-1) {
2348 (*dd_idx)++; /* Q D D D P */
2349 qd_idx = 0;
2350 } else if (*dd_idx >= pd_idx)
2351 (*dd_idx) += 2; /* D D P Q D */
NeilBrown67cc2b82009-03-31 14:39:38 +11002352 ddf_layout = 1;
NeilBrown99c0fb52009-03-31 14:39:38 +11002353 break;
2354
2355 case ALGORITHM_ROTATING_N_CONTINUE:
2356 /* Same as left_symmetric but Q is before P */
NeilBrown6e3b96e2010-04-23 07:08:28 +10002357 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
NeilBrown99c0fb52009-03-31 14:39:38 +11002358 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2359 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
NeilBrown67cc2b82009-03-31 14:39:38 +11002360 ddf_layout = 1;
NeilBrown99c0fb52009-03-31 14:39:38 +11002361 break;
2362
2363 case ALGORITHM_LEFT_ASYMMETRIC_6:
2364 /* RAID5 left_asymmetric, with Q on last device */
NeilBrown6e3b96e2010-04-23 07:08:28 +10002365 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
NeilBrown99c0fb52009-03-31 14:39:38 +11002366 if (*dd_idx >= pd_idx)
2367 (*dd_idx)++;
2368 qd_idx = raid_disks - 1;
2369 break;
2370
2371 case ALGORITHM_RIGHT_ASYMMETRIC_6:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002372 pd_idx = sector_div(stripe2, raid_disks-1);
NeilBrown99c0fb52009-03-31 14:39:38 +11002373 if (*dd_idx >= pd_idx)
2374 (*dd_idx)++;
2375 qd_idx = raid_disks - 1;
2376 break;
2377
2378 case ALGORITHM_LEFT_SYMMETRIC_6:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002379 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
NeilBrown99c0fb52009-03-31 14:39:38 +11002380 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2381 qd_idx = raid_disks - 1;
2382 break;
2383
2384 case ALGORITHM_RIGHT_SYMMETRIC_6:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002385 pd_idx = sector_div(stripe2, raid_disks-1);
NeilBrown99c0fb52009-03-31 14:39:38 +11002386 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2387 qd_idx = raid_disks - 1;
2388 break;
2389
2390 case ALGORITHM_PARITY_0_6:
2391 pd_idx = 0;
2392 (*dd_idx)++;
2393 qd_idx = raid_disks - 1;
2394 break;
2395
NeilBrown16a53ec2006-06-26 00:27:38 -07002396 default:
NeilBrown99c0fb52009-03-31 14:39:38 +11002397 BUG();
NeilBrown16a53ec2006-06-26 00:27:38 -07002398 }
2399 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400 }
2401
NeilBrown911d4ee2009-03-31 14:39:38 +11002402 if (sh) {
2403 sh->pd_idx = pd_idx;
2404 sh->qd_idx = qd_idx;
NeilBrown67cc2b82009-03-31 14:39:38 +11002405 sh->ddf_layout = ddf_layout;
NeilBrown911d4ee2009-03-31 14:39:38 +11002406 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407 /*
2408 * Finally, compute the new sector number
2409 */
2410 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2411 return new_sector;
2412}
2413
NeilBrown784052e2009-03-31 15:19:07 +11002414static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415{
NeilBrownd1688a62011-10-11 16:49:52 +11002416 struct r5conf *conf = sh->raid_conf;
NeilBrownb875e532006-12-10 02:20:49 -08002417 int raid_disks = sh->disks;
2418 int data_disks = raid_disks - conf->max_degraded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 sector_t new_sector = sh->sector, check;
Andre Noll09c9e5f2009-06-18 08:45:55 +10002420 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2421 : conf->chunk_sectors;
NeilBrowne183eae2009-03-31 15:20:22 +11002422 int algorithm = previous ? conf->prev_algo
2423 : conf->algorithm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 sector_t stripe;
2425 int chunk_offset;
NeilBrown35f2a592010-04-20 14:13:34 +10002426 sector_t chunk_number;
2427 int dummy1, dd_idx = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 sector_t r_sector;
NeilBrown911d4ee2009-03-31 14:39:38 +11002429 struct stripe_head sh2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430
2431 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2432 stripe = new_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433
NeilBrown16a53ec2006-06-26 00:27:38 -07002434 if (i == sh->pd_idx)
2435 return 0;
2436 switch(conf->level) {
2437 case 4: break;
2438 case 5:
NeilBrowne183eae2009-03-31 15:20:22 +11002439 switch (algorithm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440 case ALGORITHM_LEFT_ASYMMETRIC:
2441 case ALGORITHM_RIGHT_ASYMMETRIC:
2442 if (i > sh->pd_idx)
2443 i--;
2444 break;
2445 case ALGORITHM_LEFT_SYMMETRIC:
2446 case ALGORITHM_RIGHT_SYMMETRIC:
2447 if (i < sh->pd_idx)
2448 i += raid_disks;
2449 i -= (sh->pd_idx + 1);
2450 break;
NeilBrown99c0fb52009-03-31 14:39:38 +11002451 case ALGORITHM_PARITY_0:
2452 i -= 1;
2453 break;
2454 case ALGORITHM_PARITY_N:
2455 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456 default:
NeilBrown99c0fb52009-03-31 14:39:38 +11002457 BUG();
NeilBrown16a53ec2006-06-26 00:27:38 -07002458 }
2459 break;
2460 case 6:
NeilBrownd0dabf72009-03-31 14:39:38 +11002461 if (i == sh->qd_idx)
NeilBrown16a53ec2006-06-26 00:27:38 -07002462 return 0; /* It is the Q disk */
NeilBrowne183eae2009-03-31 15:20:22 +11002463 switch (algorithm) {
NeilBrown16a53ec2006-06-26 00:27:38 -07002464 case ALGORITHM_LEFT_ASYMMETRIC:
2465 case ALGORITHM_RIGHT_ASYMMETRIC:
NeilBrown99c0fb52009-03-31 14:39:38 +11002466 case ALGORITHM_ROTATING_ZERO_RESTART:
2467 case ALGORITHM_ROTATING_N_RESTART:
2468 if (sh->pd_idx == raid_disks-1)
2469 i--; /* Q D D D P */
NeilBrown16a53ec2006-06-26 00:27:38 -07002470 else if (i > sh->pd_idx)
2471 i -= 2; /* D D P Q D */
2472 break;
2473 case ALGORITHM_LEFT_SYMMETRIC:
2474 case ALGORITHM_RIGHT_SYMMETRIC:
2475 if (sh->pd_idx == raid_disks-1)
2476 i--; /* Q D D D P */
2477 else {
2478 /* D D P Q D */
2479 if (i < sh->pd_idx)
2480 i += raid_disks;
2481 i -= (sh->pd_idx + 2);
2482 }
2483 break;
NeilBrown99c0fb52009-03-31 14:39:38 +11002484 case ALGORITHM_PARITY_0:
2485 i -= 2;
2486 break;
2487 case ALGORITHM_PARITY_N:
2488 break;
2489 case ALGORITHM_ROTATING_N_CONTINUE:
NeilBrowne4424fe2009-10-16 16:27:34 +11002490 /* Like left_symmetric, but P is before Q */
NeilBrown99c0fb52009-03-31 14:39:38 +11002491 if (sh->pd_idx == 0)
2492 i--; /* P D D D Q */
NeilBrowne4424fe2009-10-16 16:27:34 +11002493 else {
2494 /* D D Q P D */
2495 if (i < sh->pd_idx)
2496 i += raid_disks;
2497 i -= (sh->pd_idx + 1);
2498 }
NeilBrown99c0fb52009-03-31 14:39:38 +11002499 break;
2500 case ALGORITHM_LEFT_ASYMMETRIC_6:
2501 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2502 if (i > sh->pd_idx)
2503 i--;
2504 break;
2505 case ALGORITHM_LEFT_SYMMETRIC_6:
2506 case ALGORITHM_RIGHT_SYMMETRIC_6:
2507 if (i < sh->pd_idx)
2508 i += data_disks + 1;
2509 i -= (sh->pd_idx + 1);
2510 break;
2511 case ALGORITHM_PARITY_0_6:
2512 i -= 1;
2513 break;
NeilBrown16a53ec2006-06-26 00:27:38 -07002514 default:
NeilBrown99c0fb52009-03-31 14:39:38 +11002515 BUG();
NeilBrown16a53ec2006-06-26 00:27:38 -07002516 }
2517 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518 }
2519
2520 chunk_number = stripe * data_disks + i;
NeilBrown35f2a592010-04-20 14:13:34 +10002521 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522
NeilBrown112bf892009-03-31 14:39:38 +11002523 check = raid5_compute_sector(conf, r_sector,
NeilBrown784052e2009-03-31 15:19:07 +11002524 previous, &dummy1, &sh2);
NeilBrown911d4ee2009-03-31 14:39:38 +11002525 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2526 || sh2.qd_idx != sh->qd_idx) {
NeilBrown0c55e022010-05-03 14:09:02 +10002527 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2528 mdname(conf->mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 return 0;
2530 }
2531 return r_sector;
2532}
2533
Dan Williams600aa102008-06-28 08:32:05 +10002534static void
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002535schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
Dan Williams600aa102008-06-28 08:32:05 +10002536 int rcw, int expand)
Dan Williamse33129d2007-01-02 13:52:30 -07002537{
2538 int i, pd_idx = sh->pd_idx, disks = sh->disks;
NeilBrownd1688a62011-10-11 16:49:52 +11002539 struct r5conf *conf = sh->raid_conf;
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002540 int level = conf->level;
NeilBrown16a53ec2006-06-26 00:27:38 -07002541
Dan Williamse33129d2007-01-02 13:52:30 -07002542 if (rcw) {
Dan Williamse33129d2007-01-02 13:52:30 -07002543
2544 for (i = disks; i--; ) {
2545 struct r5dev *dev = &sh->dev[i];
2546
2547 if (dev->towrite) {
2548 set_bit(R5_LOCKED, &dev->flags);
Dan Williamsd8ee0722008-06-28 08:32:06 +10002549 set_bit(R5_Wantdrain, &dev->flags);
Dan Williamse33129d2007-01-02 13:52:30 -07002550 if (!expand)
2551 clear_bit(R5_UPTODATE, &dev->flags);
Dan Williams600aa102008-06-28 08:32:05 +10002552 s->locked++;
Dan Williamse33129d2007-01-02 13:52:30 -07002553 }
2554 }
NeilBrownce7d3632013-03-04 12:37:14 +11002555 /* if we are not expanding this is a proper write request, and
2556 * there will be bios with new data to be drained into the
2557 * stripe cache
2558 */
2559 if (!expand) {
2560 if (!s->locked)
2561 /* False alarm, nothing to do */
2562 return;
2563 sh->reconstruct_state = reconstruct_state_drain_run;
2564 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2565 } else
2566 sh->reconstruct_state = reconstruct_state_run;
2567
2568 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2569
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002570 if (s->locked + conf->max_degraded == disks)
Dan Williams8b3e6cd2008-04-28 02:15:53 -07002571 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002572 atomic_inc(&conf->pending_full_writes);
Dan Williamse33129d2007-01-02 13:52:30 -07002573 } else {
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002574 BUG_ON(level == 6);
Dan Williamse33129d2007-01-02 13:52:30 -07002575 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2576 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2577
Dan Williamse33129d2007-01-02 13:52:30 -07002578 for (i = disks; i--; ) {
2579 struct r5dev *dev = &sh->dev[i];
2580 if (i == pd_idx)
2581 continue;
2582
Dan Williamse33129d2007-01-02 13:52:30 -07002583 if (dev->towrite &&
2584 (test_bit(R5_UPTODATE, &dev->flags) ||
Dan Williamsd8ee0722008-06-28 08:32:06 +10002585 test_bit(R5_Wantcompute, &dev->flags))) {
2586 set_bit(R5_Wantdrain, &dev->flags);
Dan Williamse33129d2007-01-02 13:52:30 -07002587 set_bit(R5_LOCKED, &dev->flags);
2588 clear_bit(R5_UPTODATE, &dev->flags);
Dan Williams600aa102008-06-28 08:32:05 +10002589 s->locked++;
Dan Williamse33129d2007-01-02 13:52:30 -07002590 }
2591 }
NeilBrownce7d3632013-03-04 12:37:14 +11002592 if (!s->locked)
2593 /* False alarm - nothing to do */
2594 return;
2595 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2596 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2597 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2598 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
Dan Williamse33129d2007-01-02 13:52:30 -07002599 }
2600
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002601 /* keep the parity disk(s) locked while asynchronous operations
Dan Williamse33129d2007-01-02 13:52:30 -07002602 * are in flight
2603 */
2604 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2605 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
Dan Williams600aa102008-06-28 08:32:05 +10002606 s->locked++;
Dan Williamse33129d2007-01-02 13:52:30 -07002607
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002608 if (level == 6) {
2609 int qd_idx = sh->qd_idx;
2610 struct r5dev *dev = &sh->dev[qd_idx];
2611
2612 set_bit(R5_LOCKED, &dev->flags);
2613 clear_bit(R5_UPTODATE, &dev->flags);
2614 s->locked++;
2615 }
2616
Dan Williams600aa102008-06-28 08:32:05 +10002617 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
Harvey Harrisone46b2722008-04-28 02:15:50 -07002618 __func__, (unsigned long long)sh->sector,
Dan Williams600aa102008-06-28 08:32:05 +10002619 s->locked, s->ops_request);
Dan Williamse33129d2007-01-02 13:52:30 -07002620}
NeilBrown16a53ec2006-06-26 00:27:38 -07002621
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622/*
2623 * Each stripe/dev can have one or more bion attached.
NeilBrown16a53ec2006-06-26 00:27:38 -07002624 * toread/towrite point to the first in a chain.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625 * The bi_next chain must be in order.
2626 */
2627static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2628{
2629 struct bio **bip;
NeilBrownd1688a62011-10-11 16:49:52 +11002630 struct r5conf *conf = sh->raid_conf;
NeilBrown72626682005-09-09 16:23:54 -07002631 int firstwrite=0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632
NeilBrowncbe47ec2011-07-26 11:20:35 +10002633 pr_debug("adding bi b#%llu to stripe s#%llu\n",
Kent Overstreet4f024f32013-10-11 15:44:27 -07002634 (unsigned long long)bi->bi_iter.bi_sector,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635 (unsigned long long)sh->sector);
2636
Shaohua Lib17459c2012-07-19 16:01:31 +10002637 /*
2638 * If several bio share a stripe. The bio bi_phys_segments acts as a
2639 * reference count to avoid race. The reference count should already be
2640 * increased before this function is called (for example, in
2641 * make_request()), so other bio sharing this stripe will not free the
2642 * stripe. If a stripe is owned by one stripe, the stripe lock will
2643 * protect it.
2644 */
2645 spin_lock_irq(&sh->stripe_lock);
NeilBrown72626682005-09-09 16:23:54 -07002646 if (forwrite) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647 bip = &sh->dev[dd_idx].towrite;
Shaohua Li7eaf7e82012-07-19 16:01:31 +10002648 if (*bip == NULL)
NeilBrown72626682005-09-09 16:23:54 -07002649 firstwrite = 1;
2650 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651 bip = &sh->dev[dd_idx].toread;
Kent Overstreet4f024f32013-10-11 15:44:27 -07002652 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2653 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 goto overlap;
2655 bip = & (*bip)->bi_next;
2656 }
Kent Overstreet4f024f32013-10-11 15:44:27 -07002657 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658 goto overlap;
2659
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +02002660 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661 if (*bip)
2662 bi->bi_next = *bip;
2663 *bip = bi;
Shaohua Lie7836bd62012-07-19 16:01:31 +10002664 raid5_inc_bi_active_stripes(bi);
NeilBrown72626682005-09-09 16:23:54 -07002665
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666 if (forwrite) {
2667 /* check if page is covered */
2668 sector_t sector = sh->dev[dd_idx].sector;
2669 for (bi=sh->dev[dd_idx].towrite;
2670 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
Kent Overstreet4f024f32013-10-11 15:44:27 -07002671 bi && bi->bi_iter.bi_sector <= sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
Kent Overstreetf73a1c72012-09-25 15:05:12 -07002673 if (bio_end_sector(bi) >= sector)
2674 sector = bio_end_sector(bi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 }
2676 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2677 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2678 }
NeilBrowncbe47ec2011-07-26 11:20:35 +10002679
2680 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
Kent Overstreet4f024f32013-10-11 15:44:27 -07002681 (unsigned long long)(*bip)->bi_iter.bi_sector,
NeilBrowncbe47ec2011-07-26 11:20:35 +10002682 (unsigned long long)sh->sector, dd_idx);
NeilBrownb97390a2012-10-11 13:50:12 +11002683 spin_unlock_irq(&sh->stripe_lock);
NeilBrowncbe47ec2011-07-26 11:20:35 +10002684
2685 if (conf->mddev->bitmap && firstwrite) {
2686 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2687 STRIPE_SECTORS, 0);
2688 sh->bm_seq = conf->seq_flush+1;
2689 set_bit(STRIPE_BIT_DELAY, &sh->state);
2690 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691 return 1;
2692
2693 overlap:
2694 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
Shaohua Lib17459c2012-07-19 16:01:31 +10002695 spin_unlock_irq(&sh->stripe_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696 return 0;
2697}
2698
NeilBrownd1688a62011-10-11 16:49:52 +11002699static void end_reshape(struct r5conf *conf);
NeilBrown29269552006-03-27 01:18:10 -08002700
NeilBrownd1688a62011-10-11 16:49:52 +11002701static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
NeilBrown911d4ee2009-03-31 14:39:38 +11002702 struct stripe_head *sh)
NeilBrownccfcc3c2006-03-27 01:18:09 -08002703{
NeilBrown784052e2009-03-31 15:19:07 +11002704 int sectors_per_chunk =
Andre Noll09c9e5f2009-06-18 08:45:55 +10002705 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
NeilBrown911d4ee2009-03-31 14:39:38 +11002706 int dd_idx;
Coywolf Qi Hunt2d2063c2006-10-03 01:15:50 -07002707 int chunk_offset = sector_div(stripe, sectors_per_chunk);
NeilBrown112bf892009-03-31 14:39:38 +11002708 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
Coywolf Qi Hunt2d2063c2006-10-03 01:15:50 -07002709
NeilBrown112bf892009-03-31 14:39:38 +11002710 raid5_compute_sector(conf,
2711 stripe * (disks - conf->max_degraded)
NeilBrownb875e532006-12-10 02:20:49 -08002712 *sectors_per_chunk + chunk_offset,
NeilBrown112bf892009-03-31 14:39:38 +11002713 previous,
NeilBrown911d4ee2009-03-31 14:39:38 +11002714 &dd_idx, sh);
NeilBrownccfcc3c2006-03-27 01:18:09 -08002715}
2716
Dan Williamsa4456852007-07-09 11:56:43 -07002717static void
NeilBrownd1688a62011-10-11 16:49:52 +11002718handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
Dan Williamsa4456852007-07-09 11:56:43 -07002719 struct stripe_head_state *s, int disks,
2720 struct bio **return_bi)
2721{
2722 int i;
2723 for (i = disks; i--; ) {
2724 struct bio *bi;
2725 int bitmap_end = 0;
2726
2727 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
NeilBrown3cb03002011-10-11 16:45:26 +11002728 struct md_rdev *rdev;
Dan Williamsa4456852007-07-09 11:56:43 -07002729 rcu_read_lock();
2730 rdev = rcu_dereference(conf->disks[i].rdev);
2731 if (rdev && test_bit(In_sync, &rdev->flags))
NeilBrown7f0da592011-07-28 11:39:22 +10002732 atomic_inc(&rdev->nr_pending);
2733 else
2734 rdev = NULL;
Dan Williamsa4456852007-07-09 11:56:43 -07002735 rcu_read_unlock();
NeilBrown7f0da592011-07-28 11:39:22 +10002736 if (rdev) {
2737 if (!rdev_set_badblocks(
2738 rdev,
2739 sh->sector,
2740 STRIPE_SECTORS, 0))
2741 md_error(conf->mddev, rdev);
2742 rdev_dec_pending(rdev, conf->mddev);
2743 }
Dan Williamsa4456852007-07-09 11:56:43 -07002744 }
Shaohua Lib17459c2012-07-19 16:01:31 +10002745 spin_lock_irq(&sh->stripe_lock);
Dan Williamsa4456852007-07-09 11:56:43 -07002746 /* fail all writes first */
2747 bi = sh->dev[i].towrite;
2748 sh->dev[i].towrite = NULL;
Shaohua Lib17459c2012-07-19 16:01:31 +10002749 spin_unlock_irq(&sh->stripe_lock);
NeilBrown1ed850f2012-10-11 13:50:13 +11002750 if (bi)
Dan Williamsa4456852007-07-09 11:56:43 -07002751 bitmap_end = 1;
Dan Williamsa4456852007-07-09 11:56:43 -07002752
2753 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2754 wake_up(&conf->wait_for_overlap);
2755
Kent Overstreet4f024f32013-10-11 15:44:27 -07002756 while (bi && bi->bi_iter.bi_sector <
Dan Williamsa4456852007-07-09 11:56:43 -07002757 sh->dev[i].sector + STRIPE_SECTORS) {
2758 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2759 clear_bit(BIO_UPTODATE, &bi->bi_flags);
Shaohua Lie7836bd62012-07-19 16:01:31 +10002760 if (!raid5_dec_bi_active_stripes(bi)) {
Dan Williamsa4456852007-07-09 11:56:43 -07002761 md_write_end(conf->mddev);
2762 bi->bi_next = *return_bi;
2763 *return_bi = bi;
2764 }
2765 bi = nextbi;
2766 }
Shaohua Li7eaf7e82012-07-19 16:01:31 +10002767 if (bitmap_end)
2768 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2769 STRIPE_SECTORS, 0, 0);
2770 bitmap_end = 0;
Dan Williamsa4456852007-07-09 11:56:43 -07002771 /* and fail all 'written' */
2772 bi = sh->dev[i].written;
2773 sh->dev[i].written = NULL;
Shaohua Lid592a992014-05-21 17:57:44 +08002774 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
2775 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
2776 sh->dev[i].page = sh->dev[i].orig_page;
2777 }
2778
Dan Williamsa4456852007-07-09 11:56:43 -07002779 if (bi) bitmap_end = 1;
Kent Overstreet4f024f32013-10-11 15:44:27 -07002780 while (bi && bi->bi_iter.bi_sector <
Dan Williamsa4456852007-07-09 11:56:43 -07002781 sh->dev[i].sector + STRIPE_SECTORS) {
2782 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2783 clear_bit(BIO_UPTODATE, &bi->bi_flags);
Shaohua Lie7836bd62012-07-19 16:01:31 +10002784 if (!raid5_dec_bi_active_stripes(bi)) {
Dan Williamsa4456852007-07-09 11:56:43 -07002785 md_write_end(conf->mddev);
2786 bi->bi_next = *return_bi;
2787 *return_bi = bi;
2788 }
2789 bi = bi2;
2790 }
2791
Dan Williamsb5e98d62007-01-02 13:52:31 -07002792 /* fail any reads if this device is non-operational and
2793 * the data has not reached the cache yet.
2794 */
2795 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2796 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2797 test_bit(R5_ReadError, &sh->dev[i].flags))) {
NeilBrown143c4d02012-10-11 13:50:12 +11002798 spin_lock_irq(&sh->stripe_lock);
Dan Williamsa4456852007-07-09 11:56:43 -07002799 bi = sh->dev[i].toread;
2800 sh->dev[i].toread = NULL;
NeilBrown143c4d02012-10-11 13:50:12 +11002801 spin_unlock_irq(&sh->stripe_lock);
Dan Williamsa4456852007-07-09 11:56:43 -07002802 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2803 wake_up(&conf->wait_for_overlap);
Kent Overstreet4f024f32013-10-11 15:44:27 -07002804 while (bi && bi->bi_iter.bi_sector <
Dan Williamsa4456852007-07-09 11:56:43 -07002805 sh->dev[i].sector + STRIPE_SECTORS) {
2806 struct bio *nextbi =
2807 r5_next_bio(bi, sh->dev[i].sector);
2808 clear_bit(BIO_UPTODATE, &bi->bi_flags);
Shaohua Lie7836bd62012-07-19 16:01:31 +10002809 if (!raid5_dec_bi_active_stripes(bi)) {
Dan Williamsa4456852007-07-09 11:56:43 -07002810 bi->bi_next = *return_bi;
2811 *return_bi = bi;
2812 }
2813 bi = nextbi;
2814 }
2815 }
Dan Williamsa4456852007-07-09 11:56:43 -07002816 if (bitmap_end)
2817 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2818 STRIPE_SECTORS, 0, 0);
NeilBrown8cfa7b02011-07-27 11:00:36 +10002819 /* If we were in the middle of a write the parity block might
2820 * still be locked - so just clear all R5_LOCKED flags
2821 */
2822 clear_bit(R5_LOCKED, &sh->dev[i].flags);
Dan Williamsa4456852007-07-09 11:56:43 -07002823 }
2824
Dan Williams8b3e6cd2008-04-28 02:15:53 -07002825 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2826 if (atomic_dec_and_test(&conf->pending_full_writes))
2827 md_wakeup_thread(conf->mddev->thread);
Dan Williamsa4456852007-07-09 11:56:43 -07002828}
2829
NeilBrown7f0da592011-07-28 11:39:22 +10002830static void
NeilBrownd1688a62011-10-11 16:49:52 +11002831handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
NeilBrown7f0da592011-07-28 11:39:22 +10002832 struct stripe_head_state *s)
2833{
2834 int abort = 0;
2835 int i;
2836
NeilBrown7f0da592011-07-28 11:39:22 +10002837 clear_bit(STRIPE_SYNCING, &sh->state);
NeilBrownf8dfcff2013-03-12 12:18:06 +11002838 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2839 wake_up(&conf->wait_for_overlap);
NeilBrown7f0da592011-07-28 11:39:22 +10002840 s->syncing = 0;
NeilBrown9a3e1102011-12-23 10:17:53 +11002841 s->replacing = 0;
NeilBrown7f0da592011-07-28 11:39:22 +10002842 /* There is nothing more to do for sync/check/repair.
NeilBrown18b98372012-04-01 23:48:38 +10002843 * Don't even need to abort as that is handled elsewhere
2844 * if needed, and not always wanted e.g. if there is a known
2845 * bad block here.
NeilBrown9a3e1102011-12-23 10:17:53 +11002846 * For recover/replace we need to record a bad block on all
NeilBrown7f0da592011-07-28 11:39:22 +10002847 * non-sync devices, or abort the recovery
2848 */
NeilBrown18b98372012-04-01 23:48:38 +10002849 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2850 /* During recovery devices cannot be removed, so
2851 * locking and refcounting of rdevs is not needed
2852 */
2853 for (i = 0; i < conf->raid_disks; i++) {
2854 struct md_rdev *rdev = conf->disks[i].rdev;
2855 if (rdev
2856 && !test_bit(Faulty, &rdev->flags)
2857 && !test_bit(In_sync, &rdev->flags)
2858 && !rdev_set_badblocks(rdev, sh->sector,
2859 STRIPE_SECTORS, 0))
2860 abort = 1;
2861 rdev = conf->disks[i].replacement;
2862 if (rdev
2863 && !test_bit(Faulty, &rdev->flags)
2864 && !test_bit(In_sync, &rdev->flags)
2865 && !rdev_set_badblocks(rdev, sh->sector,
2866 STRIPE_SECTORS, 0))
2867 abort = 1;
2868 }
2869 if (abort)
2870 conf->recovery_disabled =
2871 conf->mddev->recovery_disabled;
NeilBrown7f0da592011-07-28 11:39:22 +10002872 }
NeilBrown18b98372012-04-01 23:48:38 +10002873 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
NeilBrown7f0da592011-07-28 11:39:22 +10002874}
2875
NeilBrown9a3e1102011-12-23 10:17:53 +11002876static int want_replace(struct stripe_head *sh, int disk_idx)
2877{
2878 struct md_rdev *rdev;
2879 int rv = 0;
2880 /* Doing recovery so rcu locking not required */
2881 rdev = sh->raid_conf->disks[disk_idx].replacement;
2882 if (rdev
2883 && !test_bit(Faulty, &rdev->flags)
2884 && !test_bit(In_sync, &rdev->flags)
2885 && (rdev->recovery_offset <= sh->sector
2886 || rdev->mddev->recovery_cp <= sh->sector))
2887 rv = 1;
2888
2889 return rv;
2890}
2891
NeilBrown93b3dbc2011-07-27 11:00:36 +10002892/* fetch_block - checks the given member device to see if its data needs
Dan Williams1fe797e2008-06-28 09:16:30 +10002893 * to be read or computed to satisfy a request.
2894 *
2895 * Returns 1 when no more member devices need to be checked, otherwise returns
NeilBrown93b3dbc2011-07-27 11:00:36 +10002896 * 0 to tell the loop in handle_stripe_fill to continue
Dan Williamsf38e1212007-01-02 13:52:30 -07002897 */
NeilBrown2c58f062015-02-02 11:32:23 +11002898
2899static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
2900 int disk_idx, int disks)
Dan Williamsf38e1212007-01-02 13:52:30 -07002901{
2902 struct r5dev *dev = &sh->dev[disk_idx];
NeilBrown93b3dbc2011-07-27 11:00:36 +10002903 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2904 &sh->dev[s->failed_num[1]] };
NeilBrownea664c82015-02-02 14:03:28 +11002905 int i;
Dan Williamsf38e1212007-01-02 13:52:30 -07002906
NeilBrowna79cfe12015-02-02 11:37:59 +11002907
2908 if (test_bit(R5_LOCKED, &dev->flags) ||
2909 test_bit(R5_UPTODATE, &dev->flags))
2910 /* No point reading this as we already have it or have
2911 * decided to get it.
2912 */
2913 return 0;
2914
2915 if (dev->toread ||
2916 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
2917 /* We need this block to directly satisfy a request */
2918 return 1;
2919
2920 if (s->syncing || s->expanding ||
2921 (s->replacing && want_replace(sh, disk_idx)))
2922 /* When syncing, or expanding we read everything.
2923 * When replacing, we need the replaced block.
2924 */
2925 return 1;
2926
2927 if ((s->failed >= 1 && fdev[0]->toread) ||
2928 (s->failed >= 2 && fdev[1]->toread))
2929 /* If we want to read from a failed device, then
2930 * we need to actually read every other device.
2931 */
2932 return 1;
2933
NeilBrowna9d56952015-02-02 11:49:10 +11002934 /* Sometimes neither read-modify-write nor reconstruct-write
2935 * cycles can work. In those cases we read every block we
2936 * can. Then the parity-update is certain to have enough to
2937 * work with.
2938 * This can only be a problem when we need to write something,
2939 * and some device has failed. If either of those tests
2940 * fail we need look no further.
2941 */
2942 if (!s->failed || !s->to_write)
2943 return 0;
2944
2945 if (test_bit(R5_Insync, &dev->flags) &&
2946 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
2947 /* Pre-reads at not permitted until after short delay
2948 * to gather multiple requests. However if this
2949 * device is no Insync, the block could only be be computed
2950 * and there is no need to delay that.
2951 */
2952 return 0;
NeilBrownea664c82015-02-02 14:03:28 +11002953
2954 for (i = 0; i < s->failed; i++) {
2955 if (fdev[i]->towrite &&
2956 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
2957 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
2958 /* If we have a partial write to a failed
2959 * device, then we will need to reconstruct
2960 * the content of that device, so all other
2961 * devices must be read.
2962 */
2963 return 1;
2964 }
2965
2966 /* If we are forced to do a reconstruct-write, either because
2967 * the current RAID6 implementation only supports that, or
2968 * or because parity cannot be trusted and we are currently
2969 * recovering it, there is extra need to be careful.
2970 * If one of the devices that we would need to read, because
2971 * it is not being overwritten (and maybe not written at all)
2972 * is missing/faulty, then we need to read everything we can.
2973 */
2974 if (sh->raid_conf->level != 6 &&
2975 sh->sector < sh->raid_conf->mddev->recovery_cp)
2976 /* reconstruct-write isn't being forced */
2977 return 0;
2978 for (i = 0; i < s->failed; i++) {
2979 if (!test_bit(R5_UPTODATE, &fdev[i]->flags) &&
2980 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
2981 return 1;
2982 }
2983
NeilBrown2c58f062015-02-02 11:32:23 +11002984 return 0;
2985}
2986
2987static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2988 int disk_idx, int disks)
2989{
2990 struct r5dev *dev = &sh->dev[disk_idx];
2991
2992 /* is the data in this block needed, and can we get it? */
2993 if (need_this_block(sh, s, disk_idx, disks)) {
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07002994 /* we would like to get this block, possibly by computing it,
2995 * otherwise read it if the backing disk is insync
2996 */
2997 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2998 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2999 if ((s->uptodate == disks - 1) &&
NeilBrownf2b3b442011-07-26 11:35:19 +10003000 (s->failed && (disk_idx == s->failed_num[0] ||
3001 disk_idx == s->failed_num[1]))) {
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07003002 /* have disk failed, and we're requested to fetch it;
3003 * do compute it
3004 */
3005 pr_debug("Computing stripe %llu block %d\n",
3006 (unsigned long long)sh->sector, disk_idx);
3007 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3008 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3009 set_bit(R5_Wantcompute, &dev->flags);
3010 sh->ops.target = disk_idx;
3011 sh->ops.target2 = -1; /* no 2nd target */
3012 s->req_compute = 1;
NeilBrown93b3dbc2011-07-27 11:00:36 +10003013 /* Careful: from this point on 'uptodate' is in the eye
3014 * of raid_run_ops which services 'compute' operations
3015 * before writes. R5_Wantcompute flags a block that will
3016 * be R5_UPTODATE by the time it is needed for a
3017 * subsequent operation.
3018 */
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07003019 s->uptodate++;
3020 return 1;
3021 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3022 /* Computing 2-failure is *very* expensive; only
3023 * do it if failed >= 2
3024 */
3025 int other;
3026 for (other = disks; other--; ) {
3027 if (other == disk_idx)
3028 continue;
3029 if (!test_bit(R5_UPTODATE,
3030 &sh->dev[other].flags))
3031 break;
3032 }
3033 BUG_ON(other < 0);
3034 pr_debug("Computing stripe %llu blocks %d,%d\n",
3035 (unsigned long long)sh->sector,
3036 disk_idx, other);
3037 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3038 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3039 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3040 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3041 sh->ops.target = disk_idx;
3042 sh->ops.target2 = other;
3043 s->uptodate += 2;
3044 s->req_compute = 1;
3045 return 1;
3046 } else if (test_bit(R5_Insync, &dev->flags)) {
3047 set_bit(R5_LOCKED, &dev->flags);
3048 set_bit(R5_Wantread, &dev->flags);
3049 s->locked++;
3050 pr_debug("Reading block %d (sync=%d)\n",
3051 disk_idx, s->syncing);
3052 }
3053 }
3054
3055 return 0;
3056}
3057
3058/**
NeilBrown93b3dbc2011-07-27 11:00:36 +10003059 * handle_stripe_fill - read or compute data to satisfy pending requests.
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07003060 */
NeilBrown93b3dbc2011-07-27 11:00:36 +10003061static void handle_stripe_fill(struct stripe_head *sh,
3062 struct stripe_head_state *s,
3063 int disks)
Dan Williamsa4456852007-07-09 11:56:43 -07003064{
3065 int i;
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07003066
3067 /* look for blocks to read/compute, skip this if a compute
3068 * is already in flight, or if the stripe contents are in the
3069 * midst of changing due to a write
3070 */
3071 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3072 !sh->reconstruct_state)
3073 for (i = disks; i--; )
NeilBrown93b3dbc2011-07-27 11:00:36 +10003074 if (fetch_block(sh, s, i, disks))
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07003075 break;
Dan Williamsa4456852007-07-09 11:56:43 -07003076 set_bit(STRIPE_HANDLE, &sh->state);
3077}
3078
Dan Williams1fe797e2008-06-28 09:16:30 +10003079/* handle_stripe_clean_event
Dan Williamsa4456852007-07-09 11:56:43 -07003080 * any written block on an uptodate or failed drive can be returned.
3081 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3082 * never LOCKED, so we don't need to test 'failed' directly.
3083 */
NeilBrownd1688a62011-10-11 16:49:52 +11003084static void handle_stripe_clean_event(struct r5conf *conf,
Dan Williamsa4456852007-07-09 11:56:43 -07003085 struct stripe_head *sh, int disks, struct bio **return_bi)
3086{
3087 int i;
3088 struct r5dev *dev;
NeilBrownf8dfcff2013-03-12 12:18:06 +11003089 int discard_pending = 0;
Dan Williamsa4456852007-07-09 11:56:43 -07003090
3091 for (i = disks; i--; )
3092 if (sh->dev[i].written) {
3093 dev = &sh->dev[i];
3094 if (!test_bit(R5_LOCKED, &dev->flags) &&
Shaohua Li9e4447682012-10-11 13:49:49 +11003095 (test_bit(R5_UPTODATE, &dev->flags) ||
Shaohua Lid592a992014-05-21 17:57:44 +08003096 test_bit(R5_Discard, &dev->flags) ||
3097 test_bit(R5_SkipCopy, &dev->flags))) {
Dan Williamsa4456852007-07-09 11:56:43 -07003098 /* We can return any write requests */
3099 struct bio *wbi, *wbi2;
Dan Williams45b42332007-07-09 11:56:43 -07003100 pr_debug("Return write for disc %d\n", i);
NeilBrownca64cae2012-11-21 16:33:40 +11003101 if (test_and_clear_bit(R5_Discard, &dev->flags))
3102 clear_bit(R5_UPTODATE, &dev->flags);
Shaohua Lid592a992014-05-21 17:57:44 +08003103 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3104 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3105 dev->page = dev->orig_page;
3106 }
Dan Williamsa4456852007-07-09 11:56:43 -07003107 wbi = dev->written;
3108 dev->written = NULL;
Kent Overstreet4f024f32013-10-11 15:44:27 -07003109 while (wbi && wbi->bi_iter.bi_sector <
Dan Williamsa4456852007-07-09 11:56:43 -07003110 dev->sector + STRIPE_SECTORS) {
3111 wbi2 = r5_next_bio(wbi, dev->sector);
Shaohua Lie7836bd62012-07-19 16:01:31 +10003112 if (!raid5_dec_bi_active_stripes(wbi)) {
Dan Williamsa4456852007-07-09 11:56:43 -07003113 md_write_end(conf->mddev);
3114 wbi->bi_next = *return_bi;
3115 *return_bi = wbi;
3116 }
3117 wbi = wbi2;
3118 }
Shaohua Li7eaf7e82012-07-19 16:01:31 +10003119 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3120 STRIPE_SECTORS,
Dan Williamsa4456852007-07-09 11:56:43 -07003121 !test_bit(STRIPE_DEGRADED, &sh->state),
Shaohua Li7eaf7e82012-07-19 16:01:31 +10003122 0);
NeilBrownf8dfcff2013-03-12 12:18:06 +11003123 } else if (test_bit(R5_Discard, &dev->flags))
3124 discard_pending = 1;
Shaohua Lid592a992014-05-21 17:57:44 +08003125 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3126 WARN_ON(dev->page != dev->orig_page);
NeilBrownf8dfcff2013-03-12 12:18:06 +11003127 }
3128 if (!discard_pending &&
3129 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3130 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3131 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3132 if (sh->qd_idx >= 0) {
3133 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3134 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3135 }
3136 /* now that discard is done we can proceed with any sync */
3137 clear_bit(STRIPE_DISCARD, &sh->state);
Shaohua Lid47648f2013-10-19 14:51:42 +08003138 /*
3139 * SCSI discard will change some bio fields and the stripe has
3140 * no updated data, so remove it from hash list and the stripe
3141 * will be reinitialized
3142 */
3143 spin_lock_irq(&conf->device_lock);
3144 remove_hash(sh);
3145 spin_unlock_irq(&conf->device_lock);
NeilBrownf8dfcff2013-03-12 12:18:06 +11003146 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3147 set_bit(STRIPE_HANDLE, &sh->state);
3148
3149 }
Dan Williams8b3e6cd2008-04-28 02:15:53 -07003150
3151 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3152 if (atomic_dec_and_test(&conf->pending_full_writes))
3153 md_wakeup_thread(conf->mddev->thread);
Dan Williamsa4456852007-07-09 11:56:43 -07003154}
3155
NeilBrownd1688a62011-10-11 16:49:52 +11003156static void handle_stripe_dirtying(struct r5conf *conf,
NeilBrownc8ac1802011-07-27 11:00:36 +10003157 struct stripe_head *sh,
3158 struct stripe_head_state *s,
3159 int disks)
Dan Williamsa4456852007-07-09 11:56:43 -07003160{
3161 int rmw = 0, rcw = 0, i;
Alexander Lyakasa7854482012-10-11 13:50:12 +11003162 sector_t recovery_cp = conf->mddev->recovery_cp;
3163
3164 /* RAID6 requires 'rcw' in current implementation.
3165 * Otherwise, check whether resync is now happening or should start.
3166 * If yes, then the array is dirty (after unclean shutdown or
3167 * initial creation), so parity in some stripes might be inconsistent.
3168 * In this case, we need to always do reconstruct-write, to ensure
3169 * that in case of drive failure or read-error correction, we
3170 * generate correct data from the parity.
3171 */
3172 if (conf->max_degraded == 2 ||
NeilBrown26ac1072015-02-18 11:35:14 +11003173 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3174 s->failed == 0)) {
Alexander Lyakasa7854482012-10-11 13:50:12 +11003175 /* Calculate the real rcw later - for now make it
NeilBrownc8ac1802011-07-27 11:00:36 +10003176 * look like rcw is cheaper
3177 */
3178 rcw = 1; rmw = 2;
Alexander Lyakasa7854482012-10-11 13:50:12 +11003179 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
3180 conf->max_degraded, (unsigned long long)recovery_cp,
3181 (unsigned long long)sh->sector);
NeilBrownc8ac1802011-07-27 11:00:36 +10003182 } else for (i = disks; i--; ) {
Dan Williamsa4456852007-07-09 11:56:43 -07003183 /* would I have to read this buffer for read_modify_write */
3184 struct r5dev *dev = &sh->dev[i];
3185 if ((dev->towrite || i == sh->pd_idx) &&
3186 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07003187 !(test_bit(R5_UPTODATE, &dev->flags) ||
3188 test_bit(R5_Wantcompute, &dev->flags))) {
Dan Williamsa4456852007-07-09 11:56:43 -07003189 if (test_bit(R5_Insync, &dev->flags))
3190 rmw++;
3191 else
3192 rmw += 2*disks; /* cannot read it */
3193 }
3194 /* Would I have to read this buffer for reconstruct_write */
3195 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
3196 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07003197 !(test_bit(R5_UPTODATE, &dev->flags) ||
3198 test_bit(R5_Wantcompute, &dev->flags))) {
NeilBrown67f45542014-05-28 13:39:22 +10003199 if (test_bit(R5_Insync, &dev->flags))
3200 rcw++;
Dan Williamsa4456852007-07-09 11:56:43 -07003201 else
3202 rcw += 2*disks;
3203 }
3204 }
Dan Williams45b42332007-07-09 11:56:43 -07003205 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
Dan Williamsa4456852007-07-09 11:56:43 -07003206 (unsigned long long)sh->sector, rmw, rcw);
3207 set_bit(STRIPE_HANDLE, &sh->state);
NeilBrowna9add5d2012-10-31 11:59:09 +11003208 if (rmw < rcw && rmw > 0) {
Dan Williamsa4456852007-07-09 11:56:43 -07003209 /* prefer read-modify-write, but need to get some data */
Jonathan Brassowe3620a32013-03-07 16:22:01 -06003210 if (conf->mddev->queue)
3211 blk_add_trace_msg(conf->mddev->queue,
3212 "raid5 rmw %llu %d",
3213 (unsigned long long)sh->sector, rmw);
Dan Williamsa4456852007-07-09 11:56:43 -07003214 for (i = disks; i--; ) {
3215 struct r5dev *dev = &sh->dev[i];
3216 if ((dev->towrite || i == sh->pd_idx) &&
3217 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07003218 !(test_bit(R5_UPTODATE, &dev->flags) ||
3219 test_bit(R5_Wantcompute, &dev->flags)) &&
Dan Williamsa4456852007-07-09 11:56:43 -07003220 test_bit(R5_Insync, &dev->flags)) {
NeilBrown67f45542014-05-28 13:39:22 +10003221 if (test_bit(STRIPE_PREREAD_ACTIVE,
3222 &sh->state)) {
3223 pr_debug("Read_old block %d for r-m-w\n",
3224 i);
Dan Williamsa4456852007-07-09 11:56:43 -07003225 set_bit(R5_LOCKED, &dev->flags);
3226 set_bit(R5_Wantread, &dev->flags);
3227 s->locked++;
3228 } else {
3229 set_bit(STRIPE_DELAYED, &sh->state);
3230 set_bit(STRIPE_HANDLE, &sh->state);
3231 }
3232 }
3233 }
NeilBrowna9add5d2012-10-31 11:59:09 +11003234 }
NeilBrownc8ac1802011-07-27 11:00:36 +10003235 if (rcw <= rmw && rcw > 0) {
Dan Williamsa4456852007-07-09 11:56:43 -07003236 /* want reconstruct write, but need to get some data */
NeilBrowna9add5d2012-10-31 11:59:09 +11003237 int qread =0;
NeilBrownc8ac1802011-07-27 11:00:36 +10003238 rcw = 0;
Dan Williamsa4456852007-07-09 11:56:43 -07003239 for (i = disks; i--; ) {
3240 struct r5dev *dev = &sh->dev[i];
3241 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
NeilBrownc8ac1802011-07-27 11:00:36 +10003242 i != sh->pd_idx && i != sh->qd_idx &&
Dan Williamsa4456852007-07-09 11:56:43 -07003243 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07003244 !(test_bit(R5_UPTODATE, &dev->flags) ||
NeilBrownc8ac1802011-07-27 11:00:36 +10003245 test_bit(R5_Wantcompute, &dev->flags))) {
3246 rcw++;
NeilBrown67f45542014-05-28 13:39:22 +10003247 if (test_bit(R5_Insync, &dev->flags) &&
3248 test_bit(STRIPE_PREREAD_ACTIVE,
3249 &sh->state)) {
Dan Williams45b42332007-07-09 11:56:43 -07003250 pr_debug("Read_old block "
Dan Williamsa4456852007-07-09 11:56:43 -07003251 "%d for Reconstruct\n", i);
3252 set_bit(R5_LOCKED, &dev->flags);
3253 set_bit(R5_Wantread, &dev->flags);
3254 s->locked++;
NeilBrowna9add5d2012-10-31 11:59:09 +11003255 qread++;
Dan Williamsa4456852007-07-09 11:56:43 -07003256 } else {
3257 set_bit(STRIPE_DELAYED, &sh->state);
3258 set_bit(STRIPE_HANDLE, &sh->state);
3259 }
3260 }
3261 }
Jonathan Brassowe3620a32013-03-07 16:22:01 -06003262 if (rcw && conf->mddev->queue)
NeilBrowna9add5d2012-10-31 11:59:09 +11003263 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3264 (unsigned long long)sh->sector,
3265 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
NeilBrownc8ac1802011-07-27 11:00:36 +10003266 }
NeilBrownb1b02fe2015-02-02 10:44:29 +11003267
3268 if (rcw > disks && rmw > disks &&
3269 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3270 set_bit(STRIPE_DELAYED, &sh->state);
3271
Dan Williamsa4456852007-07-09 11:56:43 -07003272 /* now if nothing is locked, and if we have enough data,
3273 * we can start a write request
3274 */
Dan Williamsf38e1212007-01-02 13:52:30 -07003275 /* since handle_stripe can be called at any time we need to handle the
3276 * case where a compute block operation has been submitted and then a
Dan Williamsac6b53b2009-07-14 13:40:19 -07003277 * subsequent call wants to start a write request. raid_run_ops only
3278 * handles the case where compute block and reconstruct are requested
Dan Williamsf38e1212007-01-02 13:52:30 -07003279 * simultaneously. If this is not the case then new writes need to be
3280 * held off until the compute completes.
3281 */
Dan Williams976ea8d2008-06-28 08:32:03 +10003282 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3283 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3284 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07003285 schedule_reconstruction(sh, s, rcw == 0, 0);
Dan Williamsa4456852007-07-09 11:56:43 -07003286}
3287
NeilBrownd1688a62011-10-11 16:49:52 +11003288static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
Dan Williamsa4456852007-07-09 11:56:43 -07003289 struct stripe_head_state *s, int disks)
3290{
Dan Williamsecc65c92008-06-28 08:31:57 +10003291 struct r5dev *dev = NULL;
Dan Williamse89f8962007-01-02 13:52:31 -07003292
Dan Williamsbd2ab672008-04-10 21:29:27 -07003293 set_bit(STRIPE_HANDLE, &sh->state);
3294
Dan Williamsecc65c92008-06-28 08:31:57 +10003295 switch (sh->check_state) {
3296 case check_state_idle:
3297 /* start a new check operation if there are no failures */
Dan Williamsbd2ab672008-04-10 21:29:27 -07003298 if (s->failed == 0) {
Dan Williamsbd2ab672008-04-10 21:29:27 -07003299 BUG_ON(s->uptodate != disks);
Dan Williamsecc65c92008-06-28 08:31:57 +10003300 sh->check_state = check_state_run;
3301 set_bit(STRIPE_OP_CHECK, &s->ops_request);
Dan Williamsbd2ab672008-04-10 21:29:27 -07003302 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
Dan Williamsbd2ab672008-04-10 21:29:27 -07003303 s->uptodate--;
Dan Williamsecc65c92008-06-28 08:31:57 +10003304 break;
Dan Williamsbd2ab672008-04-10 21:29:27 -07003305 }
NeilBrownf2b3b442011-07-26 11:35:19 +10003306 dev = &sh->dev[s->failed_num[0]];
Dan Williamsecc65c92008-06-28 08:31:57 +10003307 /* fall through */
3308 case check_state_compute_result:
3309 sh->check_state = check_state_idle;
3310 if (!dev)
3311 dev = &sh->dev[sh->pd_idx];
3312
3313 /* check that a write has not made the stripe insync */
3314 if (test_bit(STRIPE_INSYNC, &sh->state))
3315 break;
3316
3317 /* either failed parity check, or recovery is happening */
Dan Williamsa4456852007-07-09 11:56:43 -07003318 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3319 BUG_ON(s->uptodate != disks);
3320
3321 set_bit(R5_LOCKED, &dev->flags);
Dan Williamsecc65c92008-06-28 08:31:57 +10003322 s->locked++;
Dan Williamsa4456852007-07-09 11:56:43 -07003323 set_bit(R5_Wantwrite, &dev->flags);
Dan Williams830ea012007-01-02 13:52:31 -07003324
Dan Williamsa4456852007-07-09 11:56:43 -07003325 clear_bit(STRIPE_DEGRADED, &sh->state);
Dan Williamsa4456852007-07-09 11:56:43 -07003326 set_bit(STRIPE_INSYNC, &sh->state);
Dan Williamsecc65c92008-06-28 08:31:57 +10003327 break;
3328 case check_state_run:
3329 break; /* we will be called again upon completion */
3330 case check_state_check_result:
3331 sh->check_state = check_state_idle;
3332
3333 /* if a failure occurred during the check operation, leave
3334 * STRIPE_INSYNC not set and let the stripe be handled again
3335 */
3336 if (s->failed)
3337 break;
3338
3339 /* handle a successful check operation, if parity is correct
3340 * we are done. Otherwise update the mismatch count and repair
3341 * parity if !MD_RECOVERY_CHECK
3342 */
Dan Williamsad283ea2009-08-29 19:09:26 -07003343 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
Dan Williamsecc65c92008-06-28 08:31:57 +10003344 /* parity is correct (on disc,
3345 * not in buffer any more)
3346 */
3347 set_bit(STRIPE_INSYNC, &sh->state);
3348 else {
Jianpeng Ma7f7583d2012-10-11 14:17:59 +11003349 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
Dan Williamsecc65c92008-06-28 08:31:57 +10003350 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3351 /* don't try to repair!! */
3352 set_bit(STRIPE_INSYNC, &sh->state);
3353 else {
3354 sh->check_state = check_state_compute_run;
Dan Williams976ea8d2008-06-28 08:32:03 +10003355 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
Dan Williamsecc65c92008-06-28 08:31:57 +10003356 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3357 set_bit(R5_Wantcompute,
3358 &sh->dev[sh->pd_idx].flags);
3359 sh->ops.target = sh->pd_idx;
Dan Williamsac6b53b2009-07-14 13:40:19 -07003360 sh->ops.target2 = -1;
Dan Williamsecc65c92008-06-28 08:31:57 +10003361 s->uptodate++;
3362 }
3363 }
3364 break;
3365 case check_state_compute_run:
3366 break;
3367 default:
3368 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3369 __func__, sh->check_state,
3370 (unsigned long long) sh->sector);
3371 BUG();
Dan Williamsa4456852007-07-09 11:56:43 -07003372 }
3373}
3374
NeilBrownd1688a62011-10-11 16:49:52 +11003375static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
Dan Williams36d1c642009-07-14 11:48:22 -07003376 struct stripe_head_state *s,
NeilBrownf2b3b442011-07-26 11:35:19 +10003377 int disks)
Dan Williamsa4456852007-07-09 11:56:43 -07003378{
Dan Williamsa4456852007-07-09 11:56:43 -07003379 int pd_idx = sh->pd_idx;
NeilBrown34e04e82009-03-31 15:10:16 +11003380 int qd_idx = sh->qd_idx;
Dan Williamsd82dfee2009-07-14 13:40:57 -07003381 struct r5dev *dev;
Dan Williamsa4456852007-07-09 11:56:43 -07003382
3383 set_bit(STRIPE_HANDLE, &sh->state);
3384
3385 BUG_ON(s->failed > 2);
Dan Williamsd82dfee2009-07-14 13:40:57 -07003386
Dan Williamsa4456852007-07-09 11:56:43 -07003387 /* Want to check and possibly repair P and Q.
3388 * However there could be one 'failed' device, in which
3389 * case we can only check one of them, possibly using the
3390 * other to generate missing data
3391 */
3392
Dan Williamsd82dfee2009-07-14 13:40:57 -07003393 switch (sh->check_state) {
3394 case check_state_idle:
3395 /* start a new check operation if there are < 2 failures */
NeilBrownf2b3b442011-07-26 11:35:19 +10003396 if (s->failed == s->q_failed) {
Dan Williamsd82dfee2009-07-14 13:40:57 -07003397 /* The only possible failed device holds Q, so it
Dan Williamsa4456852007-07-09 11:56:43 -07003398 * makes sense to check P (If anything else were failed,
3399 * we would have used P to recreate it).
3400 */
Dan Williamsd82dfee2009-07-14 13:40:57 -07003401 sh->check_state = check_state_run;
Dan Williamsa4456852007-07-09 11:56:43 -07003402 }
NeilBrownf2b3b442011-07-26 11:35:19 +10003403 if (!s->q_failed && s->failed < 2) {
Dan Williamsd82dfee2009-07-14 13:40:57 -07003404 /* Q is not failed, and we didn't use it to generate
Dan Williamsa4456852007-07-09 11:56:43 -07003405 * anything, so it makes sense to check it
3406 */
Dan Williamsd82dfee2009-07-14 13:40:57 -07003407 if (sh->check_state == check_state_run)
3408 sh->check_state = check_state_run_pq;
3409 else
3410 sh->check_state = check_state_run_q;
Dan Williamsa4456852007-07-09 11:56:43 -07003411 }
Dan Williams36d1c642009-07-14 11:48:22 -07003412
Dan Williamsd82dfee2009-07-14 13:40:57 -07003413 /* discard potentially stale zero_sum_result */
3414 sh->ops.zero_sum_result = 0;
Dan Williams36d1c642009-07-14 11:48:22 -07003415
Dan Williamsd82dfee2009-07-14 13:40:57 -07003416 if (sh->check_state == check_state_run) {
3417 /* async_xor_zero_sum destroys the contents of P */
3418 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3419 s->uptodate--;
Dan Williamsa4456852007-07-09 11:56:43 -07003420 }
Dan Williamsd82dfee2009-07-14 13:40:57 -07003421 if (sh->check_state >= check_state_run &&
3422 sh->check_state <= check_state_run_pq) {
3423 /* async_syndrome_zero_sum preserves P and Q, so
3424 * no need to mark them !uptodate here
3425 */
3426 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3427 break;
3428 }
Dan Williams36d1c642009-07-14 11:48:22 -07003429
Dan Williamsd82dfee2009-07-14 13:40:57 -07003430 /* we have 2-disk failure */
3431 BUG_ON(s->failed != 2);
3432 /* fall through */
3433 case check_state_compute_result:
3434 sh->check_state = check_state_idle;
Dan Williams36d1c642009-07-14 11:48:22 -07003435
Dan Williamsd82dfee2009-07-14 13:40:57 -07003436 /* check that a write has not made the stripe insync */
3437 if (test_bit(STRIPE_INSYNC, &sh->state))
3438 break;
Dan Williamsa4456852007-07-09 11:56:43 -07003439
3440 /* now write out any block on a failed drive,
Dan Williamsd82dfee2009-07-14 13:40:57 -07003441 * or P or Q if they were recomputed
Dan Williamsa4456852007-07-09 11:56:43 -07003442 */
Dan Williamsd82dfee2009-07-14 13:40:57 -07003443 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
Dan Williamsa4456852007-07-09 11:56:43 -07003444 if (s->failed == 2) {
NeilBrownf2b3b442011-07-26 11:35:19 +10003445 dev = &sh->dev[s->failed_num[1]];
Dan Williamsa4456852007-07-09 11:56:43 -07003446 s->locked++;
3447 set_bit(R5_LOCKED, &dev->flags);
3448 set_bit(R5_Wantwrite, &dev->flags);
3449 }
3450 if (s->failed >= 1) {
NeilBrownf2b3b442011-07-26 11:35:19 +10003451 dev = &sh->dev[s->failed_num[0]];
Dan Williamsa4456852007-07-09 11:56:43 -07003452 s->locked++;
3453 set_bit(R5_LOCKED, &dev->flags);
3454 set_bit(R5_Wantwrite, &dev->flags);
3455 }
Dan Williamsd82dfee2009-07-14 13:40:57 -07003456 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
Dan Williamsa4456852007-07-09 11:56:43 -07003457 dev = &sh->dev[pd_idx];
3458 s->locked++;
3459 set_bit(R5_LOCKED, &dev->flags);
3460 set_bit(R5_Wantwrite, &dev->flags);
3461 }
Dan Williamsd82dfee2009-07-14 13:40:57 -07003462 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
Dan Williamsa4456852007-07-09 11:56:43 -07003463 dev = &sh->dev[qd_idx];
3464 s->locked++;
3465 set_bit(R5_LOCKED, &dev->flags);
3466 set_bit(R5_Wantwrite, &dev->flags);
3467 }
3468 clear_bit(STRIPE_DEGRADED, &sh->state);
3469
3470 set_bit(STRIPE_INSYNC, &sh->state);
Dan Williamsd82dfee2009-07-14 13:40:57 -07003471 break;
3472 case check_state_run:
3473 case check_state_run_q:
3474 case check_state_run_pq:
3475 break; /* we will be called again upon completion */
3476 case check_state_check_result:
3477 sh->check_state = check_state_idle;
3478
3479 /* handle a successful check operation, if parity is correct
3480 * we are done. Otherwise update the mismatch count and repair
3481 * parity if !MD_RECOVERY_CHECK
3482 */
3483 if (sh->ops.zero_sum_result == 0) {
3484 /* both parities are correct */
3485 if (!s->failed)
3486 set_bit(STRIPE_INSYNC, &sh->state);
3487 else {
3488 /* in contrast to the raid5 case we can validate
3489 * parity, but still have a failure to write
3490 * back
3491 */
3492 sh->check_state = check_state_compute_result;
3493 /* Returning at this point means that we may go
3494 * off and bring p and/or q uptodate again so
3495 * we make sure to check zero_sum_result again
3496 * to verify if p or q need writeback
3497 */
3498 }
3499 } else {
Jianpeng Ma7f7583d2012-10-11 14:17:59 +11003500 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
Dan Williamsd82dfee2009-07-14 13:40:57 -07003501 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3502 /* don't try to repair!! */
3503 set_bit(STRIPE_INSYNC, &sh->state);
3504 else {
3505 int *target = &sh->ops.target;
3506
3507 sh->ops.target = -1;
3508 sh->ops.target2 = -1;
3509 sh->check_state = check_state_compute_run;
3510 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3511 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3512 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3513 set_bit(R5_Wantcompute,
3514 &sh->dev[pd_idx].flags);
3515 *target = pd_idx;
3516 target = &sh->ops.target2;
3517 s->uptodate++;
3518 }
3519 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3520 set_bit(R5_Wantcompute,
3521 &sh->dev[qd_idx].flags);
3522 *target = qd_idx;
3523 s->uptodate++;
3524 }
3525 }
3526 }
3527 break;
3528 case check_state_compute_run:
3529 break;
3530 default:
3531 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3532 __func__, sh->check_state,
3533 (unsigned long long) sh->sector);
3534 BUG();
Dan Williamsa4456852007-07-09 11:56:43 -07003535 }
3536}
3537
NeilBrownd1688a62011-10-11 16:49:52 +11003538static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
Dan Williamsa4456852007-07-09 11:56:43 -07003539{
3540 int i;
3541
3542 /* We have read all the blocks in this stripe and now we need to
3543 * copy some of them into a target stripe for expand.
3544 */
Dan Williamsf0a50d32007-01-02 13:52:31 -07003545 struct dma_async_tx_descriptor *tx = NULL;
Dan Williamsa4456852007-07-09 11:56:43 -07003546 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3547 for (i = 0; i < sh->disks; i++)
NeilBrown34e04e82009-03-31 15:10:16 +11003548 if (i != sh->pd_idx && i != sh->qd_idx) {
NeilBrown911d4ee2009-03-31 14:39:38 +11003549 int dd_idx, j;
Dan Williamsa4456852007-07-09 11:56:43 -07003550 struct stripe_head *sh2;
Dan Williamsa08abd82009-06-03 11:43:59 -07003551 struct async_submit_ctl submit;
Dan Williamsa4456852007-07-09 11:56:43 -07003552
NeilBrown784052e2009-03-31 15:19:07 +11003553 sector_t bn = compute_blocknr(sh, i, 1);
NeilBrown911d4ee2009-03-31 14:39:38 +11003554 sector_t s = raid5_compute_sector(conf, bn, 0,
3555 &dd_idx, NULL);
NeilBrowna8c906c2009-06-09 14:39:59 +10003556 sh2 = get_active_stripe(conf, s, 0, 1, 1);
Dan Williamsa4456852007-07-09 11:56:43 -07003557 if (sh2 == NULL)
3558 /* so far only the early blocks of this stripe
3559 * have been requested. When later blocks
3560 * get requested, we will try again
3561 */
3562 continue;
3563 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3564 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3565 /* must have already done this block */
3566 release_stripe(sh2);
3567 continue;
3568 }
Dan Williamsf0a50d32007-01-02 13:52:31 -07003569
3570 /* place all the copies on one channel */
Dan Williamsa08abd82009-06-03 11:43:59 -07003571 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
Dan Williamsf0a50d32007-01-02 13:52:31 -07003572 tx = async_memcpy(sh2->dev[dd_idx].page,
Dan Williams88ba2aa2009-04-09 16:16:18 -07003573 sh->dev[i].page, 0, 0, STRIPE_SIZE,
Dan Williamsa08abd82009-06-03 11:43:59 -07003574 &submit);
Dan Williamsf0a50d32007-01-02 13:52:31 -07003575
Dan Williamsa4456852007-07-09 11:56:43 -07003576 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3577 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3578 for (j = 0; j < conf->raid_disks; j++)
3579 if (j != sh2->pd_idx &&
NeilBrown86c374b2011-07-27 11:00:36 +10003580 j != sh2->qd_idx &&
Dan Williamsa4456852007-07-09 11:56:43 -07003581 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3582 break;
3583 if (j == conf->raid_disks) {
3584 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3585 set_bit(STRIPE_HANDLE, &sh2->state);
3586 }
3587 release_stripe(sh2);
Dan Williamsf0a50d32007-01-02 13:52:31 -07003588
Dan Williamsa4456852007-07-09 11:56:43 -07003589 }
NeilBrowna2e08552007-09-11 15:23:36 -07003590 /* done submitting copies, wait for them to complete */
NeilBrown749586b2012-11-20 14:11:15 +11003591 async_tx_quiesce(&tx);
Dan Williamsa4456852007-07-09 11:56:43 -07003592}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593
3594/*
3595 * handle_stripe - do things to a stripe.
3596 *
NeilBrown9a3e1102011-12-23 10:17:53 +11003597 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3598 * state of various bits to see what needs to be done.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599 * Possible results:
NeilBrown9a3e1102011-12-23 10:17:53 +11003600 * return some read requests which now have data
3601 * return some write requests which are safely on storage
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602 * schedule a read on some buffers
3603 * schedule a write of some buffers
3604 * return confirmation of parity correctness
3605 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606 */
Dan Williamsa4456852007-07-09 11:56:43 -07003607
NeilBrownacfe7262011-07-27 11:00:36 +10003608static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
NeilBrown16a53ec2006-06-26 00:27:38 -07003609{
NeilBrownd1688a62011-10-11 16:49:52 +11003610 struct r5conf *conf = sh->raid_conf;
NeilBrownf4168852007-02-28 20:11:53 -08003611 int disks = sh->disks;
NeilBrown474af965fe2011-07-27 11:00:36 +10003612 struct r5dev *dev;
3613 int i;
NeilBrown9a3e1102011-12-23 10:17:53 +11003614 int do_recovery = 0;
NeilBrown16a53ec2006-06-26 00:27:38 -07003615
NeilBrownacfe7262011-07-27 11:00:36 +10003616 memset(s, 0, sizeof(*s));
NeilBrown16a53ec2006-06-26 00:27:38 -07003617
NeilBrownacfe7262011-07-27 11:00:36 +10003618 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3619 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3620 s->failed_num[0] = -1;
3621 s->failed_num[1] = -1;
3622
3623 /* Now to look around and see what can be done */
NeilBrown16a53ec2006-06-26 00:27:38 -07003624 rcu_read_lock();
3625 for (i=disks; i--; ) {
NeilBrown3cb03002011-10-11 16:45:26 +11003626 struct md_rdev *rdev;
NeilBrown31c176e2011-07-28 11:39:22 +10003627 sector_t first_bad;
3628 int bad_sectors;
3629 int is_bad = 0;
NeilBrownacfe7262011-07-27 11:00:36 +10003630
NeilBrown16a53ec2006-06-26 00:27:38 -07003631 dev = &sh->dev[i];
NeilBrown16a53ec2006-06-26 00:27:38 -07003632
Dan Williams45b42332007-07-09 11:56:43 -07003633 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
NeilBrown9a3e1102011-12-23 10:17:53 +11003634 i, dev->flags,
3635 dev->toread, dev->towrite, dev->written);
Yuri Tikhonov6c0069c2009-08-29 19:13:13 -07003636 /* maybe we can reply to a read
3637 *
3638 * new wantfill requests are only permitted while
3639 * ops_complete_biofill is guaranteed to be inactive
3640 */
3641 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3642 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3643 set_bit(R5_Wantfill, &dev->flags);
NeilBrown16a53ec2006-06-26 00:27:38 -07003644
3645 /* now count some things */
NeilBrowncc940152011-07-26 11:35:35 +10003646 if (test_bit(R5_LOCKED, &dev->flags))
3647 s->locked++;
3648 if (test_bit(R5_UPTODATE, &dev->flags))
3649 s->uptodate++;
Dan Williams2d6e4ec2009-09-16 12:11:54 -07003650 if (test_bit(R5_Wantcompute, &dev->flags)) {
NeilBrowncc940152011-07-26 11:35:35 +10003651 s->compute++;
3652 BUG_ON(s->compute > 2);
Dan Williams2d6e4ec2009-09-16 12:11:54 -07003653 }
NeilBrown16a53ec2006-06-26 00:27:38 -07003654
NeilBrownacfe7262011-07-27 11:00:36 +10003655 if (test_bit(R5_Wantfill, &dev->flags))
NeilBrowncc940152011-07-26 11:35:35 +10003656 s->to_fill++;
NeilBrownacfe7262011-07-27 11:00:36 +10003657 else if (dev->toread)
NeilBrowncc940152011-07-26 11:35:35 +10003658 s->to_read++;
NeilBrown16a53ec2006-06-26 00:27:38 -07003659 if (dev->towrite) {
NeilBrowncc940152011-07-26 11:35:35 +10003660 s->to_write++;
NeilBrown16a53ec2006-06-26 00:27:38 -07003661 if (!test_bit(R5_OVERWRITE, &dev->flags))
NeilBrowncc940152011-07-26 11:35:35 +10003662 s->non_overwrite++;
NeilBrown16a53ec2006-06-26 00:27:38 -07003663 }
Dan Williamsa4456852007-07-09 11:56:43 -07003664 if (dev->written)
NeilBrowncc940152011-07-26 11:35:35 +10003665 s->written++;
NeilBrown14a75d32011-12-23 10:17:52 +11003666 /* Prefer to use the replacement for reads, but only
3667 * if it is recovered enough and has no bad blocks.
3668 */
3669 rdev = rcu_dereference(conf->disks[i].replacement);
3670 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3671 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3672 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3673 &first_bad, &bad_sectors))
3674 set_bit(R5_ReadRepl, &dev->flags);
3675 else {
NeilBrown9a3e1102011-12-23 10:17:53 +11003676 if (rdev)
3677 set_bit(R5_NeedReplace, &dev->flags);
NeilBrown14a75d32011-12-23 10:17:52 +11003678 rdev = rcu_dereference(conf->disks[i].rdev);
3679 clear_bit(R5_ReadRepl, &dev->flags);
3680 }
NeilBrown9283d8c2011-12-08 16:27:57 +11003681 if (rdev && test_bit(Faulty, &rdev->flags))
3682 rdev = NULL;
NeilBrown31c176e2011-07-28 11:39:22 +10003683 if (rdev) {
3684 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3685 &first_bad, &bad_sectors);
3686 if (s->blocked_rdev == NULL
3687 && (test_bit(Blocked, &rdev->flags)
3688 || is_bad < 0)) {
3689 if (is_bad < 0)
3690 set_bit(BlockedBadBlocks,
3691 &rdev->flags);
3692 s->blocked_rdev = rdev;
3693 atomic_inc(&rdev->nr_pending);
3694 }
Dan Williams6bfe0b42008-04-30 00:52:32 -07003695 }
NeilBrown415e72d2010-06-17 17:25:21 +10003696 clear_bit(R5_Insync, &dev->flags);
3697 if (!rdev)
3698 /* Not in-sync */;
NeilBrown31c176e2011-07-28 11:39:22 +10003699 else if (is_bad) {
3700 /* also not in-sync */
NeilBrown18b98372012-04-01 23:48:38 +10003701 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3702 test_bit(R5_UPTODATE, &dev->flags)) {
NeilBrown31c176e2011-07-28 11:39:22 +10003703 /* treat as in-sync, but with a read error
3704 * which we can now try to correct
3705 */
3706 set_bit(R5_Insync, &dev->flags);
3707 set_bit(R5_ReadError, &dev->flags);
3708 }
3709 } else if (test_bit(In_sync, &rdev->flags))
NeilBrown415e72d2010-06-17 17:25:21 +10003710 set_bit(R5_Insync, &dev->flags);
NeilBrown30d7a482011-12-23 09:57:00 +11003711 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
NeilBrown415e72d2010-06-17 17:25:21 +10003712 /* in sync if before recovery_offset */
NeilBrown30d7a482011-12-23 09:57:00 +11003713 set_bit(R5_Insync, &dev->flags);
3714 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3715 test_bit(R5_Expanded, &dev->flags))
3716 /* If we've reshaped into here, we assume it is Insync.
3717 * We will shortly update recovery_offset to make
3718 * it official.
3719 */
3720 set_bit(R5_Insync, &dev->flags);
3721
NeilBrown1cc03eb2014-01-06 13:19:42 +11003722 if (test_bit(R5_WriteError, &dev->flags)) {
NeilBrown14a75d32011-12-23 10:17:52 +11003723 /* This flag does not apply to '.replacement'
3724 * only to .rdev, so make sure to check that*/
3725 struct md_rdev *rdev2 = rcu_dereference(
3726 conf->disks[i].rdev);
3727 if (rdev2 == rdev)
3728 clear_bit(R5_Insync, &dev->flags);
3729 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
NeilBrownbc2607f2011-07-28 11:39:22 +10003730 s->handle_bad_blocks = 1;
NeilBrown14a75d32011-12-23 10:17:52 +11003731 atomic_inc(&rdev2->nr_pending);
NeilBrownbc2607f2011-07-28 11:39:22 +10003732 } else
3733 clear_bit(R5_WriteError, &dev->flags);
3734 }
NeilBrown1cc03eb2014-01-06 13:19:42 +11003735 if (test_bit(R5_MadeGood, &dev->flags)) {
NeilBrown14a75d32011-12-23 10:17:52 +11003736 /* This flag does not apply to '.replacement'
3737 * only to .rdev, so make sure to check that*/
3738 struct md_rdev *rdev2 = rcu_dereference(
3739 conf->disks[i].rdev);
3740 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
NeilBrownb84db562011-07-28 11:39:23 +10003741 s->handle_bad_blocks = 1;
NeilBrown14a75d32011-12-23 10:17:52 +11003742 atomic_inc(&rdev2->nr_pending);
NeilBrownb84db562011-07-28 11:39:23 +10003743 } else
3744 clear_bit(R5_MadeGood, &dev->flags);
3745 }
NeilBrown977df362011-12-23 10:17:53 +11003746 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3747 struct md_rdev *rdev2 = rcu_dereference(
3748 conf->disks[i].replacement);
3749 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3750 s->handle_bad_blocks = 1;
3751 atomic_inc(&rdev2->nr_pending);
3752 } else
3753 clear_bit(R5_MadeGoodRepl, &dev->flags);
3754 }
NeilBrown415e72d2010-06-17 17:25:21 +10003755 if (!test_bit(R5_Insync, &dev->flags)) {
NeilBrown16a53ec2006-06-26 00:27:38 -07003756 /* The ReadError flag will just be confusing now */
3757 clear_bit(R5_ReadError, &dev->flags);
3758 clear_bit(R5_ReWrite, &dev->flags);
3759 }
NeilBrown415e72d2010-06-17 17:25:21 +10003760 if (test_bit(R5_ReadError, &dev->flags))
3761 clear_bit(R5_Insync, &dev->flags);
3762 if (!test_bit(R5_Insync, &dev->flags)) {
NeilBrowncc940152011-07-26 11:35:35 +10003763 if (s->failed < 2)
3764 s->failed_num[s->failed] = i;
3765 s->failed++;
NeilBrown9a3e1102011-12-23 10:17:53 +11003766 if (rdev && !test_bit(Faulty, &rdev->flags))
3767 do_recovery = 1;
NeilBrown415e72d2010-06-17 17:25:21 +10003768 }
NeilBrown16a53ec2006-06-26 00:27:38 -07003769 }
NeilBrown9a3e1102011-12-23 10:17:53 +11003770 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3771 /* If there is a failed device being replaced,
3772 * we must be recovering.
3773 * else if we are after recovery_cp, we must be syncing
majianpengc6d2e082012-04-02 01:16:59 +10003774 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
NeilBrown9a3e1102011-12-23 10:17:53 +11003775 * else we can only be replacing
3776 * sync and recovery both need to read all devices, and so
3777 * use the same flag.
3778 */
3779 if (do_recovery ||
majianpengc6d2e082012-04-02 01:16:59 +10003780 sh->sector >= conf->mddev->recovery_cp ||
3781 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
NeilBrown9a3e1102011-12-23 10:17:53 +11003782 s->syncing = 1;
3783 else
3784 s->replacing = 1;
3785 }
NeilBrown16a53ec2006-06-26 00:27:38 -07003786 rcu_read_unlock();
NeilBrowncc940152011-07-26 11:35:35 +10003787}
NeilBrownf4168852007-02-28 20:11:53 -08003788
NeilBrowncc940152011-07-26 11:35:35 +10003789static void handle_stripe(struct stripe_head *sh)
3790{
3791 struct stripe_head_state s;
NeilBrownd1688a62011-10-11 16:49:52 +11003792 struct r5conf *conf = sh->raid_conf;
NeilBrown3687c062011-07-27 11:00:36 +10003793 int i;
NeilBrown84789552011-07-27 11:00:36 +10003794 int prexor;
3795 int disks = sh->disks;
NeilBrown474af965fe2011-07-27 11:00:36 +10003796 struct r5dev *pdev, *qdev;
NeilBrowncc940152011-07-26 11:35:35 +10003797
3798 clear_bit(STRIPE_HANDLE, &sh->state);
Dan Williams257a4b42011-11-08 16:22:06 +11003799 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
NeilBrowncc940152011-07-26 11:35:35 +10003800 /* already being handled, ensure it gets handled
3801 * again when current action finishes */
3802 set_bit(STRIPE_HANDLE, &sh->state);
3803 return;
3804 }
3805
NeilBrownf8dfcff2013-03-12 12:18:06 +11003806 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3807 spin_lock(&sh->stripe_lock);
3808 /* Cannot process 'sync' concurrently with 'discard' */
3809 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3810 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3811 set_bit(STRIPE_SYNCING, &sh->state);
3812 clear_bit(STRIPE_INSYNC, &sh->state);
NeilBrownf94c0b62013-07-22 12:57:21 +10003813 clear_bit(STRIPE_REPLACED, &sh->state);
NeilBrownf8dfcff2013-03-12 12:18:06 +11003814 }
3815 spin_unlock(&sh->stripe_lock);
NeilBrowncc940152011-07-26 11:35:35 +10003816 }
3817 clear_bit(STRIPE_DELAYED, &sh->state);
3818
3819 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3820 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3821 (unsigned long long)sh->sector, sh->state,
3822 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3823 sh->check_state, sh->reconstruct_state);
NeilBrowncc940152011-07-26 11:35:35 +10003824
NeilBrownacfe7262011-07-27 11:00:36 +10003825 analyse_stripe(sh, &s);
NeilBrownc5a31002011-07-27 11:00:36 +10003826
NeilBrownbc2607f2011-07-28 11:39:22 +10003827 if (s.handle_bad_blocks) {
3828 set_bit(STRIPE_HANDLE, &sh->state);
3829 goto finish;
3830 }
3831
NeilBrown474af965fe2011-07-27 11:00:36 +10003832 if (unlikely(s.blocked_rdev)) {
3833 if (s.syncing || s.expanding || s.expanded ||
NeilBrown9a3e1102011-12-23 10:17:53 +11003834 s.replacing || s.to_write || s.written) {
NeilBrown474af965fe2011-07-27 11:00:36 +10003835 set_bit(STRIPE_HANDLE, &sh->state);
3836 goto finish;
3837 }
3838 /* There is nothing for the blocked_rdev to block */
3839 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3840 s.blocked_rdev = NULL;
3841 }
3842
3843 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3844 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3845 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3846 }
3847
3848 pr_debug("locked=%d uptodate=%d to_read=%d"
3849 " to_write=%d failed=%d failed_num=%d,%d\n",
3850 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3851 s.failed_num[0], s.failed_num[1]);
3852 /* check if the array has lost more than max_degraded devices and,
3853 * if so, some requests might need to be failed.
3854 */
NeilBrown9a3f5302011-11-08 16:22:01 +11003855 if (s.failed > conf->max_degraded) {
3856 sh->check_state = 0;
3857 sh->reconstruct_state = 0;
3858 if (s.to_read+s.to_write+s.written)
3859 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
NeilBrown9a3e1102011-12-23 10:17:53 +11003860 if (s.syncing + s.replacing)
NeilBrown9a3f5302011-11-08 16:22:01 +11003861 handle_failed_sync(conf, sh, &s);
3862 }
NeilBrown474af965fe2011-07-27 11:00:36 +10003863
NeilBrown84789552011-07-27 11:00:36 +10003864 /* Now we check to see if any write operations have recently
3865 * completed
3866 */
3867 prexor = 0;
3868 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3869 prexor = 1;
3870 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3871 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3872 sh->reconstruct_state = reconstruct_state_idle;
3873
3874 /* All the 'written' buffers and the parity block are ready to
3875 * be written back to disk
3876 */
Shaohua Li9e4447682012-10-11 13:49:49 +11003877 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3878 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
NeilBrown84789552011-07-27 11:00:36 +10003879 BUG_ON(sh->qd_idx >= 0 &&
Shaohua Li9e4447682012-10-11 13:49:49 +11003880 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3881 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
NeilBrown84789552011-07-27 11:00:36 +10003882 for (i = disks; i--; ) {
3883 struct r5dev *dev = &sh->dev[i];
3884 if (test_bit(R5_LOCKED, &dev->flags) &&
3885 (i == sh->pd_idx || i == sh->qd_idx ||
3886 dev->written)) {
3887 pr_debug("Writing block %d\n", i);
3888 set_bit(R5_Wantwrite, &dev->flags);
3889 if (prexor)
3890 continue;
NeilBrown9c4bdf62014-08-13 09:57:07 +10003891 if (s.failed > 1)
3892 continue;
NeilBrown84789552011-07-27 11:00:36 +10003893 if (!test_bit(R5_Insync, &dev->flags) ||
3894 ((i == sh->pd_idx || i == sh->qd_idx) &&
3895 s.failed == 0))
3896 set_bit(STRIPE_INSYNC, &sh->state);
3897 }
3898 }
3899 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3900 s.dec_preread_active = 1;
3901 }
3902
NeilBrownef5b7c62012-11-22 09:13:36 +11003903 /*
3904 * might be able to return some write requests if the parity blocks
3905 * are safe, or on a failed drive
3906 */
3907 pdev = &sh->dev[sh->pd_idx];
3908 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3909 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3910 qdev = &sh->dev[sh->qd_idx];
3911 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3912 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3913 || conf->level < 6;
3914
3915 if (s.written &&
3916 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3917 && !test_bit(R5_LOCKED, &pdev->flags)
3918 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3919 test_bit(R5_Discard, &pdev->flags))))) &&
3920 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3921 && !test_bit(R5_LOCKED, &qdev->flags)
3922 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3923 test_bit(R5_Discard, &qdev->flags))))))
3924 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3925
3926 /* Now we might consider reading some blocks, either to check/generate
3927 * parity, or to satisfy requests
3928 * or to load a block that is being partially written.
3929 */
3930 if (s.to_read || s.non_overwrite
3931 || (conf->level == 6 && s.to_write && s.failed)
3932 || (s.syncing && (s.uptodate + s.compute < disks))
3933 || s.replacing
3934 || s.expanding)
3935 handle_stripe_fill(sh, &s, disks);
3936
NeilBrown84789552011-07-27 11:00:36 +10003937 /* Now to consider new write requests and what else, if anything
3938 * should be read. We do not handle new writes when:
3939 * 1/ A 'write' operation (copy+xor) is already in flight.
3940 * 2/ A 'check' operation is in flight, as it may clobber the parity
3941 * block.
3942 */
3943 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3944 handle_stripe_dirtying(conf, sh, &s, disks);
3945
3946 /* maybe we need to check and possibly fix the parity for this stripe
3947 * Any reads will already have been scheduled, so we just see if enough
3948 * data is available. The parity check is held off while parity
3949 * dependent operations are in flight.
3950 */
3951 if (sh->check_state ||
3952 (s.syncing && s.locked == 0 &&
3953 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3954 !test_bit(STRIPE_INSYNC, &sh->state))) {
3955 if (conf->level == 6)
3956 handle_parity_checks6(conf, sh, &s, disks);
3957 else
3958 handle_parity_checks5(conf, sh, &s, disks);
3959 }
NeilBrownc5a31002011-07-27 11:00:36 +10003960
NeilBrownf94c0b62013-07-22 12:57:21 +10003961 if ((s.replacing || s.syncing) && s.locked == 0
3962 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3963 && !test_bit(STRIPE_REPLACED, &sh->state)) {
NeilBrown9a3e1102011-12-23 10:17:53 +11003964 /* Write out to replacement devices where possible */
3965 for (i = 0; i < conf->raid_disks; i++)
NeilBrownf94c0b62013-07-22 12:57:21 +10003966 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3967 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
NeilBrown9a3e1102011-12-23 10:17:53 +11003968 set_bit(R5_WantReplace, &sh->dev[i].flags);
3969 set_bit(R5_LOCKED, &sh->dev[i].flags);
3970 s.locked++;
3971 }
NeilBrownf94c0b62013-07-22 12:57:21 +10003972 if (s.replacing)
3973 set_bit(STRIPE_INSYNC, &sh->state);
3974 set_bit(STRIPE_REPLACED, &sh->state);
NeilBrown9a3e1102011-12-23 10:17:53 +11003975 }
3976 if ((s.syncing || s.replacing) && s.locked == 0 &&
NeilBrownf94c0b62013-07-22 12:57:21 +10003977 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
NeilBrown9a3e1102011-12-23 10:17:53 +11003978 test_bit(STRIPE_INSYNC, &sh->state)) {
NeilBrownc5a31002011-07-27 11:00:36 +10003979 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3980 clear_bit(STRIPE_SYNCING, &sh->state);
NeilBrownf8dfcff2013-03-12 12:18:06 +11003981 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3982 wake_up(&conf->wait_for_overlap);
NeilBrownc5a31002011-07-27 11:00:36 +10003983 }
3984
3985 /* If the failed drives are just a ReadError, then we might need
3986 * to progress the repair/check process
3987 */
3988 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3989 for (i = 0; i < s.failed; i++) {
3990 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3991 if (test_bit(R5_ReadError, &dev->flags)
3992 && !test_bit(R5_LOCKED, &dev->flags)
3993 && test_bit(R5_UPTODATE, &dev->flags)
3994 ) {
3995 if (!test_bit(R5_ReWrite, &dev->flags)) {
3996 set_bit(R5_Wantwrite, &dev->flags);
3997 set_bit(R5_ReWrite, &dev->flags);
3998 set_bit(R5_LOCKED, &dev->flags);
3999 s.locked++;
4000 } else {
4001 /* let's read it back */
4002 set_bit(R5_Wantread, &dev->flags);
4003 set_bit(R5_LOCKED, &dev->flags);
4004 s.locked++;
4005 }
4006 }
4007 }
4008
NeilBrown3687c062011-07-27 11:00:36 +10004009 /* Finish reconstruct operations initiated by the expansion process */
4010 if (sh->reconstruct_state == reconstruct_state_result) {
4011 struct stripe_head *sh_src
4012 = get_active_stripe(conf, sh->sector, 1, 1, 1);
4013 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4014 /* sh cannot be written until sh_src has been read.
4015 * so arrange for sh to be delayed a little
4016 */
4017 set_bit(STRIPE_DELAYED, &sh->state);
4018 set_bit(STRIPE_HANDLE, &sh->state);
4019 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4020 &sh_src->state))
4021 atomic_inc(&conf->preread_active_stripes);
4022 release_stripe(sh_src);
4023 goto finish;
4024 }
4025 if (sh_src)
4026 release_stripe(sh_src);
NeilBrown16a53ec2006-06-26 00:27:38 -07004027
NeilBrown3687c062011-07-27 11:00:36 +10004028 sh->reconstruct_state = reconstruct_state_idle;
4029 clear_bit(STRIPE_EXPANDING, &sh->state);
4030 for (i = conf->raid_disks; i--; ) {
4031 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4032 set_bit(R5_LOCKED, &sh->dev[i].flags);
4033 s.locked++;
4034 }
4035 }
4036
4037 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4038 !sh->reconstruct_state) {
4039 /* Need to write out all blocks after computing parity */
4040 sh->disks = conf->raid_disks;
4041 stripe_set_idx(sh->sector, conf, 0, sh);
4042 schedule_reconstruction(sh, &s, 1, 1);
4043 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4044 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4045 atomic_dec(&conf->reshape_stripes);
4046 wake_up(&conf->wait_for_overlap);
4047 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4048 }
4049
4050 if (s.expanding && s.locked == 0 &&
4051 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4052 handle_stripe_expansion(conf, sh);
4053
4054finish:
Dan Williams6bfe0b42008-04-30 00:52:32 -07004055 /* wait for this device to become unblocked */
NeilBrown5f066c62012-07-03 12:13:29 +10004056 if (unlikely(s.blocked_rdev)) {
4057 if (conf->mddev->external)
4058 md_wait_for_blocked_rdev(s.blocked_rdev,
4059 conf->mddev);
4060 else
4061 /* Internal metadata will immediately
4062 * be written by raid5d, so we don't
4063 * need to wait here.
4064 */
4065 rdev_dec_pending(s.blocked_rdev,
4066 conf->mddev);
4067 }
Dan Williams6bfe0b42008-04-30 00:52:32 -07004068
NeilBrownbc2607f2011-07-28 11:39:22 +10004069 if (s.handle_bad_blocks)
4070 for (i = disks; i--; ) {
NeilBrown3cb03002011-10-11 16:45:26 +11004071 struct md_rdev *rdev;
NeilBrownbc2607f2011-07-28 11:39:22 +10004072 struct r5dev *dev = &sh->dev[i];
4073 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4074 /* We own a safe reference to the rdev */
4075 rdev = conf->disks[i].rdev;
4076 if (!rdev_set_badblocks(rdev, sh->sector,
4077 STRIPE_SECTORS, 0))
4078 md_error(conf->mddev, rdev);
4079 rdev_dec_pending(rdev, conf->mddev);
4080 }
NeilBrownb84db562011-07-28 11:39:23 +10004081 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4082 rdev = conf->disks[i].rdev;
4083 rdev_clear_badblocks(rdev, sh->sector,
NeilBrownc6563a82012-05-21 09:27:00 +10004084 STRIPE_SECTORS, 0);
NeilBrownb84db562011-07-28 11:39:23 +10004085 rdev_dec_pending(rdev, conf->mddev);
4086 }
NeilBrown977df362011-12-23 10:17:53 +11004087 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4088 rdev = conf->disks[i].replacement;
NeilBrowndd054fc2011-12-23 10:17:53 +11004089 if (!rdev)
4090 /* rdev have been moved down */
4091 rdev = conf->disks[i].rdev;
NeilBrown977df362011-12-23 10:17:53 +11004092 rdev_clear_badblocks(rdev, sh->sector,
NeilBrownc6563a82012-05-21 09:27:00 +10004093 STRIPE_SECTORS, 0);
NeilBrown977df362011-12-23 10:17:53 +11004094 rdev_dec_pending(rdev, conf->mddev);
4095 }
NeilBrownbc2607f2011-07-28 11:39:22 +10004096 }
4097
Yuri Tikhonov6c0069c2009-08-29 19:13:13 -07004098 if (s.ops_request)
4099 raid_run_ops(sh, s.ops_request);
4100
Dan Williamsf0e43bc2008-06-28 08:31:55 +10004101 ops_run_io(sh, &s);
4102
NeilBrownc5709ef2011-07-26 11:35:20 +10004103 if (s.dec_preread_active) {
NeilBrown729a1862009-12-14 12:49:50 +11004104 /* We delay this until after ops_run_io so that if make_request
Tejun Heoe9c74692010-09-03 11:56:18 +02004105 * is waiting on a flush, it won't continue until the writes
NeilBrown729a1862009-12-14 12:49:50 +11004106 * have actually been submitted.
4107 */
4108 atomic_dec(&conf->preread_active_stripes);
4109 if (atomic_read(&conf->preread_active_stripes) <
4110 IO_THRESHOLD)
4111 md_wakeup_thread(conf->mddev->thread);
4112 }
4113
NeilBrownc5709ef2011-07-26 11:35:20 +10004114 return_io(s.return_bi);
NeilBrown16a53ec2006-06-26 00:27:38 -07004115
Dan Williams257a4b42011-11-08 16:22:06 +11004116 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
NeilBrown16a53ec2006-06-26 00:27:38 -07004117}
4118
NeilBrownd1688a62011-10-11 16:49:52 +11004119static void raid5_activate_delayed(struct r5conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120{
4121 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4122 while (!list_empty(&conf->delayed_list)) {
4123 struct list_head *l = conf->delayed_list.next;
4124 struct stripe_head *sh;
4125 sh = list_entry(l, struct stripe_head, lru);
4126 list_del_init(l);
4127 clear_bit(STRIPE_DELAYED, &sh->state);
4128 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4129 atomic_inc(&conf->preread_active_stripes);
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004130 list_add_tail(&sh->lru, &conf->hold_list);
Shaohua Li851c30c2013-08-28 14:30:16 +08004131 raid5_wakeup_stripe_thread(sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004132 }
NeilBrown482c0832011-04-18 18:25:42 +10004133 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004134}
4135
Shaohua Li566c09c2013-11-14 15:16:17 +11004136static void activate_bit_delay(struct r5conf *conf,
4137 struct list_head *temp_inactive_list)
NeilBrown72626682005-09-09 16:23:54 -07004138{
4139 /* device_lock is held */
4140 struct list_head head;
4141 list_add(&head, &conf->bitmap_list);
4142 list_del_init(&conf->bitmap_list);
4143 while (!list_empty(&head)) {
4144 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
Shaohua Li566c09c2013-11-14 15:16:17 +11004145 int hash;
NeilBrown72626682005-09-09 16:23:54 -07004146 list_del_init(&sh->lru);
4147 atomic_inc(&sh->count);
Shaohua Li566c09c2013-11-14 15:16:17 +11004148 hash = sh->hash_lock_index;
4149 __release_stripe(conf, sh, &temp_inactive_list[hash]);
NeilBrown72626682005-09-09 16:23:54 -07004150 }
4151}
4152
NeilBrown5c675f82014-12-15 12:56:56 +11004153static int raid5_congested(struct mddev *mddev, int bits)
NeilBrownf022b2f2006-10-03 01:15:56 -07004154{
NeilBrownd1688a62011-10-11 16:49:52 +11004155 struct r5conf *conf = mddev->private;
NeilBrownf022b2f2006-10-03 01:15:56 -07004156
4157 /* No difference between reads and writes. Just check
4158 * how busy the stripe_cache is
4159 */
NeilBrown3fa841d2009-09-23 18:10:29 +10004160
NeilBrownf022b2f2006-10-03 01:15:56 -07004161 if (conf->inactive_blocked)
4162 return 1;
4163 if (conf->quiesce)
4164 return 1;
Shaohua Li4bda5562013-11-14 15:16:17 +11004165 if (atomic_read(&conf->empty_inactive_list_nr))
NeilBrownf022b2f2006-10-03 01:15:56 -07004166 return 1;
4167
4168 return 0;
4169}
4170
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08004171/* We want read requests to align with chunks where possible,
4172 * but write requests don't need to.
4173 */
NeilBrown64590f42014-12-15 12:56:57 +11004174static int raid5_mergeable_bvec(struct mddev *mddev,
Alasdair G Kergoncc371e62008-07-03 09:53:43 +02004175 struct bvec_merge_data *bvm,
4176 struct bio_vec *biovec)
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08004177{
Alasdair G Kergoncc371e62008-07-03 09:53:43 +02004178 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08004179 int max;
Andre Noll9d8f0362009-06-18 08:45:01 +10004180 unsigned int chunk_sectors = mddev->chunk_sectors;
Alasdair G Kergoncc371e62008-07-03 09:53:43 +02004181 unsigned int bio_sectors = bvm->bi_size >> 9;
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08004182
Alasdair G Kergoncc371e62008-07-03 09:53:43 +02004183 if ((bvm->bi_rw & 1) == WRITE)
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08004184 return biovec->bv_len; /* always allow writes to be mergeable */
4185
Andre Noll664e7c42009-06-18 08:45:27 +10004186 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4187 chunk_sectors = mddev->new_chunk_sectors;
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08004188 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4189 if (max < 0) max = 0;
4190 if (max <= biovec->bv_len && bio_sectors == 0)
4191 return biovec->bv_len;
4192 else
4193 return max;
4194}
4195
NeilBrownfd01b882011-10-11 16:47:53 +11004196static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004197{
Kent Overstreet4f024f32013-10-11 15:44:27 -07004198 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
Andre Noll9d8f0362009-06-18 08:45:01 +10004199 unsigned int chunk_sectors = mddev->chunk_sectors;
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08004200 unsigned int bio_sectors = bio_sectors(bio);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004201
Andre Noll664e7c42009-06-18 08:45:27 +10004202 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4203 chunk_sectors = mddev->new_chunk_sectors;
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004204 return chunk_sectors >=
4205 ((sector & (chunk_sectors - 1)) + bio_sectors);
4206}
4207
4208/*
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004209 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4210 * later sampled by raid5d.
4211 */
NeilBrownd1688a62011-10-11 16:49:52 +11004212static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004213{
4214 unsigned long flags;
4215
4216 spin_lock_irqsave(&conf->device_lock, flags);
4217
4218 bi->bi_next = conf->retry_read_aligned_list;
4219 conf->retry_read_aligned_list = bi;
4220
4221 spin_unlock_irqrestore(&conf->device_lock, flags);
4222 md_wakeup_thread(conf->mddev->thread);
4223}
4224
NeilBrownd1688a62011-10-11 16:49:52 +11004225static struct bio *remove_bio_from_retry(struct r5conf *conf)
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004226{
4227 struct bio *bi;
4228
4229 bi = conf->retry_read_aligned;
4230 if (bi) {
4231 conf->retry_read_aligned = NULL;
4232 return bi;
4233 }
4234 bi = conf->retry_read_aligned_list;
4235 if(bi) {
Neil Brown387bb172007-02-08 14:20:29 -08004236 conf->retry_read_aligned_list = bi->bi_next;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004237 bi->bi_next = NULL;
Jens Axboe960e7392008-08-15 10:41:18 +02004238 /*
4239 * this sets the active strip count to 1 and the processed
4240 * strip count to zero (upper 8 bits)
4241 */
Shaohua Lie7836bd62012-07-19 16:01:31 +10004242 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004243 }
4244
4245 return bi;
4246}
4247
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004248/*
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004249 * The "raid5_align_endio" should check if the read succeeded and if it
4250 * did, call bio_endio on the original bio (having bio_put the new bio
4251 * first).
4252 * If the read failed..
4253 */
NeilBrown6712ecf2007-09-27 12:47:43 +02004254static void raid5_align_endio(struct bio *bi, int error)
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004255{
4256 struct bio* raid_bi = bi->bi_private;
NeilBrownfd01b882011-10-11 16:47:53 +11004257 struct mddev *mddev;
NeilBrownd1688a62011-10-11 16:49:52 +11004258 struct r5conf *conf;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004259 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
NeilBrown3cb03002011-10-11 16:45:26 +11004260 struct md_rdev *rdev;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004261
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004262 bio_put(bi);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004263
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004264 rdev = (void*)raid_bi->bi_next;
4265 raid_bi->bi_next = NULL;
NeilBrown2b7f2222010-03-25 16:06:03 +11004266 mddev = rdev->mddev;
4267 conf = mddev->private;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004268
4269 rdev_dec_pending(rdev, conf->mddev);
4270
4271 if (!error && uptodate) {
Linus Torvalds0a82a8d2013-04-18 09:00:26 -07004272 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4273 raid_bi, 0);
NeilBrown6712ecf2007-09-27 12:47:43 +02004274 bio_endio(raid_bi, 0);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004275 if (atomic_dec_and_test(&conf->active_aligned_reads))
4276 wake_up(&conf->wait_for_stripe);
NeilBrown6712ecf2007-09-27 12:47:43 +02004277 return;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004278 }
4279
Dan Williams45b42332007-07-09 11:56:43 -07004280 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004281
4282 add_bio_to_retry(raid_bi, conf);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004283}
4284
Neil Brown387bb172007-02-08 14:20:29 -08004285static int bio_fits_rdev(struct bio *bi)
4286{
Jens Axboe165125e2007-07-24 09:28:11 +02004287 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
Neil Brown387bb172007-02-08 14:20:29 -08004288
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08004289 if (bio_sectors(bi) > queue_max_sectors(q))
Neil Brown387bb172007-02-08 14:20:29 -08004290 return 0;
4291 blk_recount_segments(q, bi);
Martin K. Petersen8a783622010-02-26 00:20:39 -05004292 if (bi->bi_phys_segments > queue_max_segments(q))
Neil Brown387bb172007-02-08 14:20:29 -08004293 return 0;
4294
4295 if (q->merge_bvec_fn)
4296 /* it's too hard to apply the merge_bvec_fn at this stage,
4297 * just just give up
4298 */
4299 return 0;
4300
4301 return 1;
4302}
4303
NeilBrownfd01b882011-10-11 16:47:53 +11004304static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004305{
NeilBrownd1688a62011-10-11 16:49:52 +11004306 struct r5conf *conf = mddev->private;
NeilBrown8553fe7ec2009-12-14 12:49:47 +11004307 int dd_idx;
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004308 struct bio* align_bi;
NeilBrown3cb03002011-10-11 16:45:26 +11004309 struct md_rdev *rdev;
NeilBrown671488c2011-12-23 10:17:52 +11004310 sector_t end_sector;
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004311
4312 if (!in_chunk_boundary(mddev, raid_bio)) {
Dan Williams45b42332007-07-09 11:56:43 -07004313 pr_debug("chunk_aligned_read : non aligned\n");
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004314 return 0;
4315 }
4316 /*
NeilBrowna167f662010-10-26 18:31:13 +11004317 * use bio_clone_mddev to make a copy of the bio
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004318 */
NeilBrowna167f662010-10-26 18:31:13 +11004319 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004320 if (!align_bi)
4321 return 0;
4322 /*
4323 * set bi_end_io to a new function, and set bi_private to the
4324 * original bio.
4325 */
4326 align_bi->bi_end_io = raid5_align_endio;
4327 align_bi->bi_private = raid_bio;
4328 /*
4329 * compute position
4330 */
Kent Overstreet4f024f32013-10-11 15:44:27 -07004331 align_bi->bi_iter.bi_sector =
4332 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4333 0, &dd_idx, NULL);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004334
Kent Overstreetf73a1c72012-09-25 15:05:12 -07004335 end_sector = bio_end_sector(align_bi);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004336 rcu_read_lock();
NeilBrown671488c2011-12-23 10:17:52 +11004337 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4338 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4339 rdev->recovery_offset < end_sector) {
4340 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4341 if (rdev &&
4342 (test_bit(Faulty, &rdev->flags) ||
4343 !(test_bit(In_sync, &rdev->flags) ||
4344 rdev->recovery_offset >= end_sector)))
4345 rdev = NULL;
4346 }
4347 if (rdev) {
NeilBrown31c176e2011-07-28 11:39:22 +10004348 sector_t first_bad;
4349 int bad_sectors;
4350
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004351 atomic_inc(&rdev->nr_pending);
4352 rcu_read_unlock();
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004353 raid_bio->bi_next = (void*)rdev;
4354 align_bi->bi_bdev = rdev->bdev;
NeilBrown3fd83712014-08-23 20:19:26 +10004355 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004356
NeilBrown31c176e2011-07-28 11:39:22 +10004357 if (!bio_fits_rdev(align_bi) ||
Kent Overstreet4f024f32013-10-11 15:44:27 -07004358 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4359 bio_sectors(align_bi),
NeilBrown31c176e2011-07-28 11:39:22 +10004360 &first_bad, &bad_sectors)) {
4361 /* too big in some way, or has a known bad block */
Neil Brown387bb172007-02-08 14:20:29 -08004362 bio_put(align_bi);
4363 rdev_dec_pending(rdev, mddev);
4364 return 0;
4365 }
4366
majianpeng6c0544e2012-06-12 08:31:10 +08004367 /* No reshape active, so we can trust rdev->data_offset */
Kent Overstreet4f024f32013-10-11 15:44:27 -07004368 align_bi->bi_iter.bi_sector += rdev->data_offset;
majianpeng6c0544e2012-06-12 08:31:10 +08004369
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004370 spin_lock_irq(&conf->device_lock);
4371 wait_event_lock_irq(conf->wait_for_stripe,
4372 conf->quiesce == 0,
Lukas Czernereed8c022012-11-30 11:42:40 +01004373 conf->device_lock);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004374 atomic_inc(&conf->active_aligned_reads);
4375 spin_unlock_irq(&conf->device_lock);
4376
Jonathan Brassowe3620a32013-03-07 16:22:01 -06004377 if (mddev->gendisk)
4378 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4379 align_bi, disk_devt(mddev->gendisk),
Kent Overstreet4f024f32013-10-11 15:44:27 -07004380 raid_bio->bi_iter.bi_sector);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004381 generic_make_request(align_bi);
4382 return 1;
4383 } else {
4384 rcu_read_unlock();
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004385 bio_put(align_bi);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004386 return 0;
4387 }
4388}
4389
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004390/* __get_priority_stripe - get the next stripe to process
4391 *
4392 * Full stripe writes are allowed to pass preread active stripes up until
4393 * the bypass_threshold is exceeded. In general the bypass_count
4394 * increments when the handle_list is handled before the hold_list; however, it
4395 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4396 * stripe with in flight i/o. The bypass_count will be reset when the
4397 * head of the hold_list has changed, i.e. the head was promoted to the
4398 * handle_list.
4399 */
Shaohua Li851c30c2013-08-28 14:30:16 +08004400static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004401{
Shaohua Li851c30c2013-08-28 14:30:16 +08004402 struct stripe_head *sh = NULL, *tmp;
4403 struct list_head *handle_list = NULL;
Shaohua Libfc90cb2013-08-29 15:40:32 +08004404 struct r5worker_group *wg = NULL;
Shaohua Li851c30c2013-08-28 14:30:16 +08004405
4406 if (conf->worker_cnt_per_group == 0) {
4407 handle_list = &conf->handle_list;
4408 } else if (group != ANY_GROUP) {
4409 handle_list = &conf->worker_groups[group].handle_list;
Shaohua Libfc90cb2013-08-29 15:40:32 +08004410 wg = &conf->worker_groups[group];
Shaohua Li851c30c2013-08-28 14:30:16 +08004411 } else {
4412 int i;
4413 for (i = 0; i < conf->group_cnt; i++) {
4414 handle_list = &conf->worker_groups[i].handle_list;
Shaohua Libfc90cb2013-08-29 15:40:32 +08004415 wg = &conf->worker_groups[i];
Shaohua Li851c30c2013-08-28 14:30:16 +08004416 if (!list_empty(handle_list))
4417 break;
4418 }
4419 }
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004420
4421 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4422 __func__,
Shaohua Li851c30c2013-08-28 14:30:16 +08004423 list_empty(handle_list) ? "empty" : "busy",
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004424 list_empty(&conf->hold_list) ? "empty" : "busy",
4425 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4426
Shaohua Li851c30c2013-08-28 14:30:16 +08004427 if (!list_empty(handle_list)) {
4428 sh = list_entry(handle_list->next, typeof(*sh), lru);
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004429
4430 if (list_empty(&conf->hold_list))
4431 conf->bypass_count = 0;
4432 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4433 if (conf->hold_list.next == conf->last_hold)
4434 conf->bypass_count++;
4435 else {
4436 conf->last_hold = conf->hold_list.next;
4437 conf->bypass_count -= conf->bypass_threshold;
4438 if (conf->bypass_count < 0)
4439 conf->bypass_count = 0;
4440 }
4441 }
4442 } else if (!list_empty(&conf->hold_list) &&
4443 ((conf->bypass_threshold &&
4444 conf->bypass_count > conf->bypass_threshold) ||
4445 atomic_read(&conf->pending_full_writes) == 0)) {
Shaohua Li851c30c2013-08-28 14:30:16 +08004446
4447 list_for_each_entry(tmp, &conf->hold_list, lru) {
4448 if (conf->worker_cnt_per_group == 0 ||
4449 group == ANY_GROUP ||
4450 !cpu_online(tmp->cpu) ||
4451 cpu_to_group(tmp->cpu) == group) {
4452 sh = tmp;
4453 break;
4454 }
4455 }
4456
4457 if (sh) {
4458 conf->bypass_count -= conf->bypass_threshold;
4459 if (conf->bypass_count < 0)
4460 conf->bypass_count = 0;
4461 }
Shaohua Libfc90cb2013-08-29 15:40:32 +08004462 wg = NULL;
Shaohua Li851c30c2013-08-28 14:30:16 +08004463 }
4464
4465 if (!sh)
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004466 return NULL;
4467
Shaohua Libfc90cb2013-08-29 15:40:32 +08004468 if (wg) {
4469 wg->stripes_cnt--;
4470 sh->group = NULL;
4471 }
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004472 list_del_init(&sh->lru);
Shaohua Lic7a6d352014-04-15 09:12:54 +08004473 BUG_ON(atomic_inc_return(&sh->count) != 1);
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004474 return sh;
4475}
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004476
Shaohua Li8811b592012-08-02 08:33:00 +10004477struct raid5_plug_cb {
4478 struct blk_plug_cb cb;
4479 struct list_head list;
Shaohua Li566c09c2013-11-14 15:16:17 +11004480 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
Shaohua Li8811b592012-08-02 08:33:00 +10004481};
4482
4483static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4484{
4485 struct raid5_plug_cb *cb = container_of(
4486 blk_cb, struct raid5_plug_cb, cb);
4487 struct stripe_head *sh;
4488 struct mddev *mddev = cb->cb.data;
4489 struct r5conf *conf = mddev->private;
NeilBrowna9add5d2012-10-31 11:59:09 +11004490 int cnt = 0;
Shaohua Li566c09c2013-11-14 15:16:17 +11004491 int hash;
Shaohua Li8811b592012-08-02 08:33:00 +10004492
4493 if (cb->list.next && !list_empty(&cb->list)) {
4494 spin_lock_irq(&conf->device_lock);
4495 while (!list_empty(&cb->list)) {
4496 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4497 list_del_init(&sh->lru);
4498 /*
4499 * avoid race release_stripe_plug() sees
4500 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4501 * is still in our list
4502 */
Peter Zijlstra4e857c52014-03-17 18:06:10 +01004503 smp_mb__before_atomic();
Shaohua Li8811b592012-08-02 08:33:00 +10004504 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
Shaohua Li773ca822013-08-27 17:50:39 +08004505 /*
4506 * STRIPE_ON_RELEASE_LIST could be set here. In that
4507 * case, the count is always > 1 here
4508 */
Shaohua Li566c09c2013-11-14 15:16:17 +11004509 hash = sh->hash_lock_index;
4510 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
NeilBrowna9add5d2012-10-31 11:59:09 +11004511 cnt++;
Shaohua Li8811b592012-08-02 08:33:00 +10004512 }
4513 spin_unlock_irq(&conf->device_lock);
4514 }
Shaohua Li566c09c2013-11-14 15:16:17 +11004515 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4516 NR_STRIPE_HASH_LOCKS);
Jonathan Brassowe3620a32013-03-07 16:22:01 -06004517 if (mddev->queue)
4518 trace_block_unplug(mddev->queue, cnt, !from_schedule);
Shaohua Li8811b592012-08-02 08:33:00 +10004519 kfree(cb);
4520}
4521
4522static void release_stripe_plug(struct mddev *mddev,
4523 struct stripe_head *sh)
4524{
4525 struct blk_plug_cb *blk_cb = blk_check_plugged(
4526 raid5_unplug, mddev,
4527 sizeof(struct raid5_plug_cb));
4528 struct raid5_plug_cb *cb;
4529
4530 if (!blk_cb) {
4531 release_stripe(sh);
4532 return;
4533 }
4534
4535 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4536
Shaohua Li566c09c2013-11-14 15:16:17 +11004537 if (cb->list.next == NULL) {
4538 int i;
Shaohua Li8811b592012-08-02 08:33:00 +10004539 INIT_LIST_HEAD(&cb->list);
Shaohua Li566c09c2013-11-14 15:16:17 +11004540 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
4541 INIT_LIST_HEAD(cb->temp_inactive_list + i);
4542 }
Shaohua Li8811b592012-08-02 08:33:00 +10004543
4544 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4545 list_add_tail(&sh->lru, &cb->list);
4546 else
4547 release_stripe(sh);
4548}
4549
Shaohua Li620125f2012-10-11 13:49:05 +11004550static void make_discard_request(struct mddev *mddev, struct bio *bi)
4551{
4552 struct r5conf *conf = mddev->private;
4553 sector_t logical_sector, last_sector;
4554 struct stripe_head *sh;
4555 int remaining;
4556 int stripe_sectors;
4557
4558 if (mddev->reshape_position != MaxSector)
4559 /* Skip discard while reshape is happening */
4560 return;
4561
Kent Overstreet4f024f32013-10-11 15:44:27 -07004562 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4563 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
Shaohua Li620125f2012-10-11 13:49:05 +11004564
4565 bi->bi_next = NULL;
4566 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4567
4568 stripe_sectors = conf->chunk_sectors *
4569 (conf->raid_disks - conf->max_degraded);
4570 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4571 stripe_sectors);
4572 sector_div(last_sector, stripe_sectors);
4573
4574 logical_sector *= conf->chunk_sectors;
4575 last_sector *= conf->chunk_sectors;
4576
4577 for (; logical_sector < last_sector;
4578 logical_sector += STRIPE_SECTORS) {
4579 DEFINE_WAIT(w);
4580 int d;
4581 again:
4582 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4583 prepare_to_wait(&conf->wait_for_overlap, &w,
4584 TASK_UNINTERRUPTIBLE);
NeilBrownf8dfcff2013-03-12 12:18:06 +11004585 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4586 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4587 release_stripe(sh);
4588 schedule();
4589 goto again;
4590 }
4591 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
Shaohua Li620125f2012-10-11 13:49:05 +11004592 spin_lock_irq(&sh->stripe_lock);
4593 for (d = 0; d < conf->raid_disks; d++) {
4594 if (d == sh->pd_idx || d == sh->qd_idx)
4595 continue;
4596 if (sh->dev[d].towrite || sh->dev[d].toread) {
4597 set_bit(R5_Overlap, &sh->dev[d].flags);
4598 spin_unlock_irq(&sh->stripe_lock);
4599 release_stripe(sh);
4600 schedule();
4601 goto again;
4602 }
4603 }
NeilBrownf8dfcff2013-03-12 12:18:06 +11004604 set_bit(STRIPE_DISCARD, &sh->state);
Shaohua Li620125f2012-10-11 13:49:05 +11004605 finish_wait(&conf->wait_for_overlap, &w);
4606 for (d = 0; d < conf->raid_disks; d++) {
4607 if (d == sh->pd_idx || d == sh->qd_idx)
4608 continue;
4609 sh->dev[d].towrite = bi;
4610 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4611 raid5_inc_bi_active_stripes(bi);
4612 }
4613 spin_unlock_irq(&sh->stripe_lock);
4614 if (conf->mddev->bitmap) {
4615 for (d = 0;
4616 d < conf->raid_disks - conf->max_degraded;
4617 d++)
4618 bitmap_startwrite(mddev->bitmap,
4619 sh->sector,
4620 STRIPE_SECTORS,
4621 0);
4622 sh->bm_seq = conf->seq_flush + 1;
4623 set_bit(STRIPE_BIT_DELAY, &sh->state);
4624 }
4625
4626 set_bit(STRIPE_HANDLE, &sh->state);
4627 clear_bit(STRIPE_DELAYED, &sh->state);
4628 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4629 atomic_inc(&conf->preread_active_stripes);
4630 release_stripe_plug(mddev, sh);
4631 }
4632
4633 remaining = raid5_dec_bi_active_stripes(bi);
4634 if (remaining == 0) {
4635 md_write_end(mddev);
4636 bio_endio(bi, 0);
4637 }
4638}
4639
Linus Torvaldsb4fdcb02011-11-04 17:06:58 -07004640static void make_request(struct mddev *mddev, struct bio * bi)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004641{
NeilBrownd1688a62011-10-11 16:49:52 +11004642 struct r5conf *conf = mddev->private;
NeilBrown911d4ee2009-03-31 14:39:38 +11004643 int dd_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004644 sector_t new_sector;
4645 sector_t logical_sector, last_sector;
4646 struct stripe_head *sh;
Jens Axboea3623572005-11-01 09:26:16 +01004647 const int rw = bio_data_dir(bi);
NeilBrown49077322010-03-25 16:20:56 +11004648 int remaining;
Shaohua Li27c0f682014-04-09 11:25:47 +08004649 DEFINE_WAIT(w);
4650 bool do_prepare;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004651
Tejun Heoe9c74692010-09-03 11:56:18 +02004652 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4653 md_flush_request(mddev, bi);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02004654 return;
NeilBrowne5dcdd82005-09-09 16:23:41 -07004655 }
4656
NeilBrown3d310eb2005-06-21 17:17:26 -07004657 md_write_start(mddev, bi);
NeilBrown06d91a52005-06-21 17:17:12 -07004658
NeilBrown802ba062006-12-13 00:34:13 -08004659 if (rw == READ &&
Raz Ben-Jehuda(caro)52488612006-12-10 02:20:48 -08004660 mddev->reshape_position == MaxSector &&
NeilBrown21a52c62010-04-01 15:02:13 +11004661 chunk_aligned_read(mddev,bi))
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02004662 return;
Raz Ben-Jehuda(caro)52488612006-12-10 02:20:48 -08004663
Shaohua Li620125f2012-10-11 13:49:05 +11004664 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4665 make_discard_request(mddev, bi);
4666 return;
4667 }
4668
Kent Overstreet4f024f32013-10-11 15:44:27 -07004669 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
Kent Overstreetf73a1c72012-09-25 15:05:12 -07004670 last_sector = bio_end_sector(bi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004671 bi->bi_next = NULL;
4672 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
NeilBrown06d91a52005-06-21 17:17:12 -07004673
Shaohua Li27c0f682014-04-09 11:25:47 +08004674 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004675 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
NeilBrownb5663ba2009-03-31 14:39:38 +11004676 int previous;
NeilBrownc46501b2013-08-27 15:52:13 +10004677 int seq;
NeilBrownb578d552006-03-27 01:18:12 -08004678
Shaohua Li27c0f682014-04-09 11:25:47 +08004679 do_prepare = false;
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004680 retry:
NeilBrownc46501b2013-08-27 15:52:13 +10004681 seq = read_seqcount_begin(&conf->gen_lock);
NeilBrownb5663ba2009-03-31 14:39:38 +11004682 previous = 0;
Shaohua Li27c0f682014-04-09 11:25:47 +08004683 if (do_prepare)
4684 prepare_to_wait(&conf->wait_for_overlap, &w,
4685 TASK_UNINTERRUPTIBLE);
NeilBrownb0f9ec02009-03-31 15:27:18 +11004686 if (unlikely(conf->reshape_progress != MaxSector)) {
NeilBrownfef9c612009-03-31 15:16:46 +11004687 /* spinlock is needed as reshape_progress may be
NeilBrowndf8e7f72006-03-27 01:18:15 -08004688 * 64bit on a 32bit platform, and so it might be
4689 * possible to see a half-updated value
Jesper Juhlaeb878b2011-04-10 18:06:17 +02004690 * Of course reshape_progress could change after
NeilBrowndf8e7f72006-03-27 01:18:15 -08004691 * the lock is dropped, so once we get a reference
4692 * to the stripe that we think it is, we will have
4693 * to check again.
4694 */
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004695 spin_lock_irq(&conf->device_lock);
NeilBrown2c810cd2012-05-21 09:27:00 +10004696 if (mddev->reshape_backwards
NeilBrownfef9c612009-03-31 15:16:46 +11004697 ? logical_sector < conf->reshape_progress
4698 : logical_sector >= conf->reshape_progress) {
NeilBrownb5663ba2009-03-31 14:39:38 +11004699 previous = 1;
4700 } else {
NeilBrown2c810cd2012-05-21 09:27:00 +10004701 if (mddev->reshape_backwards
NeilBrownfef9c612009-03-31 15:16:46 +11004702 ? logical_sector < conf->reshape_safe
4703 : logical_sector >= conf->reshape_safe) {
NeilBrownb578d552006-03-27 01:18:12 -08004704 spin_unlock_irq(&conf->device_lock);
4705 schedule();
Shaohua Li27c0f682014-04-09 11:25:47 +08004706 do_prepare = true;
NeilBrownb578d552006-03-27 01:18:12 -08004707 goto retry;
4708 }
4709 }
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004710 spin_unlock_irq(&conf->device_lock);
4711 }
NeilBrown16a53ec2006-06-26 00:27:38 -07004712
NeilBrown112bf892009-03-31 14:39:38 +11004713 new_sector = raid5_compute_sector(conf, logical_sector,
4714 previous,
NeilBrown911d4ee2009-03-31 14:39:38 +11004715 &dd_idx, NULL);
NeilBrown0c55e022010-05-03 14:09:02 +10004716 pr_debug("raid456: make_request, sector %llu logical %llu\n",
NeilBrownc46501b2013-08-27 15:52:13 +10004717 (unsigned long long)new_sector,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004718 (unsigned long long)logical_sector);
4719
NeilBrownb5663ba2009-03-31 14:39:38 +11004720 sh = get_active_stripe(conf, new_sector, previous,
NeilBrowna8c906c2009-06-09 14:39:59 +10004721 (bi->bi_rw&RWA_MASK), 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004722 if (sh) {
NeilBrownb0f9ec02009-03-31 15:27:18 +11004723 if (unlikely(previous)) {
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004724 /* expansion might have moved on while waiting for a
NeilBrowndf8e7f72006-03-27 01:18:15 -08004725 * stripe, so we must do the range check again.
4726 * Expansion could still move past after this
4727 * test, but as we are holding a reference to
4728 * 'sh', we know that if that happens,
4729 * STRIPE_EXPANDING will get set and the expansion
4730 * won't proceed until we finish with the stripe.
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004731 */
4732 int must_retry = 0;
4733 spin_lock_irq(&conf->device_lock);
NeilBrown2c810cd2012-05-21 09:27:00 +10004734 if (mddev->reshape_backwards
NeilBrownb0f9ec02009-03-31 15:27:18 +11004735 ? logical_sector >= conf->reshape_progress
4736 : logical_sector < conf->reshape_progress)
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004737 /* mismatch, need to try again */
4738 must_retry = 1;
4739 spin_unlock_irq(&conf->device_lock);
4740 if (must_retry) {
4741 release_stripe(sh);
Dan Williams7a3ab902009-06-16 16:00:33 -07004742 schedule();
Shaohua Li27c0f682014-04-09 11:25:47 +08004743 do_prepare = true;
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004744 goto retry;
4745 }
4746 }
NeilBrownc46501b2013-08-27 15:52:13 +10004747 if (read_seqcount_retry(&conf->gen_lock, seq)) {
4748 /* Might have got the wrong stripe_head
4749 * by accident
4750 */
4751 release_stripe(sh);
4752 goto retry;
4753 }
NeilBrowne62e58a2009-07-01 13:15:35 +10004754
Namhyung Kimffd96e32011-07-18 17:38:51 +10004755 if (rw == WRITE &&
NeilBrowna5c308d2009-07-01 13:15:35 +10004756 logical_sector >= mddev->suspend_lo &&
NeilBrowne464eaf2006-03-27 01:18:14 -08004757 logical_sector < mddev->suspend_hi) {
4758 release_stripe(sh);
NeilBrowne62e58a2009-07-01 13:15:35 +10004759 /* As the suspend_* range is controlled by
4760 * userspace, we want an interruptible
4761 * wait.
4762 */
4763 flush_signals(current);
4764 prepare_to_wait(&conf->wait_for_overlap,
4765 &w, TASK_INTERRUPTIBLE);
4766 if (logical_sector >= mddev->suspend_lo &&
Shaohua Li27c0f682014-04-09 11:25:47 +08004767 logical_sector < mddev->suspend_hi) {
NeilBrowne62e58a2009-07-01 13:15:35 +10004768 schedule();
Shaohua Li27c0f682014-04-09 11:25:47 +08004769 do_prepare = true;
4770 }
NeilBrowne464eaf2006-03-27 01:18:14 -08004771 goto retry;
4772 }
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004773
4774 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
Namhyung Kimffd96e32011-07-18 17:38:51 +10004775 !add_stripe_bio(sh, bi, dd_idx, rw)) {
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004776 /* Stripe is busy expanding or
4777 * add failed due to overlap. Flush everything
Linus Torvalds1da177e2005-04-16 15:20:36 -07004778 * and wait a while
4779 */
NeilBrown482c0832011-04-18 18:25:42 +10004780 md_wakeup_thread(mddev->thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004781 release_stripe(sh);
4782 schedule();
Shaohua Li27c0f682014-04-09 11:25:47 +08004783 do_prepare = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004784 goto retry;
4785 }
NeilBrown6ed30032008-02-06 01:40:00 -08004786 set_bit(STRIPE_HANDLE, &sh->state);
4787 clear_bit(STRIPE_DELAYED, &sh->state);
NeilBrowna852d7b2012-09-19 12:48:30 +10004788 if ((bi->bi_rw & REQ_SYNC) &&
NeilBrown729a1862009-12-14 12:49:50 +11004789 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4790 atomic_inc(&conf->preread_active_stripes);
Shaohua Li8811b592012-08-02 08:33:00 +10004791 release_stripe_plug(mddev, sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004792 } else {
4793 /* cannot get stripe for read-ahead, just give-up */
4794 clear_bit(BIO_UPTODATE, &bi->bi_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004795 break;
4796 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004797 }
Shaohua Li27c0f682014-04-09 11:25:47 +08004798 finish_wait(&conf->wait_for_overlap, &w);
NeilBrown7c13edc2011-04-18 18:25:43 +10004799
Shaohua Lie7836bd62012-07-19 16:01:31 +10004800 remaining = raid5_dec_bi_active_stripes(bi);
NeilBrownf6344752006-03-27 01:18:17 -08004801 if (remaining == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004802
NeilBrown16a53ec2006-06-26 00:27:38 -07004803 if ( rw == WRITE )
Linus Torvalds1da177e2005-04-16 15:20:36 -07004804 md_write_end(mddev);
NeilBrown6712ecf2007-09-27 12:47:43 +02004805
Linus Torvalds0a82a8d2013-04-18 09:00:26 -07004806 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4807 bi, 0);
Neil Brown0e13fe232008-06-28 08:31:20 +10004808 bio_endio(bi, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004809 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004810}
4811
NeilBrownfd01b882011-10-11 16:47:53 +11004812static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
Dan Williamsb522adc2009-03-31 15:00:31 +11004813
NeilBrownfd01b882011-10-11 16:47:53 +11004814static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004815{
NeilBrown52c03292006-06-26 00:27:43 -07004816 /* reshaping is quite different to recovery/resync so it is
4817 * handled quite separately ... here.
4818 *
4819 * On each call to sync_request, we gather one chunk worth of
4820 * destination stripes and flag them as expanding.
4821 * Then we find all the source stripes and request reads.
4822 * As the reads complete, handle_stripe will copy the data
4823 * into the destination stripe and release that stripe.
4824 */
NeilBrownd1688a62011-10-11 16:49:52 +11004825 struct r5conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004826 struct stripe_head *sh;
NeilBrownccfcc3c2006-03-27 01:18:09 -08004827 sector_t first_sector, last_sector;
NeilBrownf4168852007-02-28 20:11:53 -08004828 int raid_disks = conf->previous_raid_disks;
4829 int data_disks = raid_disks - conf->max_degraded;
4830 int new_data_disks = conf->raid_disks - conf->max_degraded;
NeilBrown52c03292006-06-26 00:27:43 -07004831 int i;
4832 int dd_idx;
NeilBrownc8f517c2009-03-31 15:28:40 +11004833 sector_t writepos, readpos, safepos;
NeilBrownec32a2b2009-03-31 15:17:38 +11004834 sector_t stripe_addr;
NeilBrown7a661382009-03-31 15:21:40 +11004835 int reshape_sectors;
NeilBrownab69ae12009-03-31 15:26:47 +11004836 struct list_head stripes;
NeilBrown52c03292006-06-26 00:27:43 -07004837
NeilBrownfef9c612009-03-31 15:16:46 +11004838 if (sector_nr == 0) {
4839 /* If restarting in the middle, skip the initial sectors */
NeilBrown2c810cd2012-05-21 09:27:00 +10004840 if (mddev->reshape_backwards &&
NeilBrownfef9c612009-03-31 15:16:46 +11004841 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4842 sector_nr = raid5_size(mddev, 0, 0)
4843 - conf->reshape_progress;
NeilBrown2c810cd2012-05-21 09:27:00 +10004844 } else if (!mddev->reshape_backwards &&
NeilBrownfef9c612009-03-31 15:16:46 +11004845 conf->reshape_progress > 0)
4846 sector_nr = conf->reshape_progress;
NeilBrownf4168852007-02-28 20:11:53 -08004847 sector_div(sector_nr, new_data_disks);
NeilBrownfef9c612009-03-31 15:16:46 +11004848 if (sector_nr) {
NeilBrown8dee7212009-11-06 14:59:29 +11004849 mddev->curr_resync_completed = sector_nr;
4850 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
NeilBrownfef9c612009-03-31 15:16:46 +11004851 *skipped = 1;
4852 return sector_nr;
4853 }
NeilBrown52c03292006-06-26 00:27:43 -07004854 }
4855
NeilBrown7a661382009-03-31 15:21:40 +11004856 /* We need to process a full chunk at a time.
4857 * If old and new chunk sizes differ, we need to process the
4858 * largest of these
4859 */
Andre Noll664e7c42009-06-18 08:45:27 +10004860 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4861 reshape_sectors = mddev->new_chunk_sectors;
NeilBrown7a661382009-03-31 15:21:40 +11004862 else
Andre Noll9d8f0362009-06-18 08:45:01 +10004863 reshape_sectors = mddev->chunk_sectors;
NeilBrown7a661382009-03-31 15:21:40 +11004864
NeilBrownb5254dd2012-05-21 09:27:01 +10004865 /* We update the metadata at least every 10 seconds, or when
4866 * the data about to be copied would over-write the source of
4867 * the data at the front of the range. i.e. one new_stripe
4868 * along from reshape_progress new_maps to after where
4869 * reshape_safe old_maps to
NeilBrown52c03292006-06-26 00:27:43 -07004870 */
NeilBrownfef9c612009-03-31 15:16:46 +11004871 writepos = conf->reshape_progress;
NeilBrownf4168852007-02-28 20:11:53 -08004872 sector_div(writepos, new_data_disks);
NeilBrownc8f517c2009-03-31 15:28:40 +11004873 readpos = conf->reshape_progress;
4874 sector_div(readpos, data_disks);
NeilBrownfef9c612009-03-31 15:16:46 +11004875 safepos = conf->reshape_safe;
NeilBrownf4168852007-02-28 20:11:53 -08004876 sector_div(safepos, data_disks);
NeilBrown2c810cd2012-05-21 09:27:00 +10004877 if (mddev->reshape_backwards) {
NeilBrowned37d832009-05-27 21:39:05 +10004878 writepos -= min_t(sector_t, reshape_sectors, writepos);
NeilBrownc8f517c2009-03-31 15:28:40 +11004879 readpos += reshape_sectors;
NeilBrown7a661382009-03-31 15:21:40 +11004880 safepos += reshape_sectors;
NeilBrownfef9c612009-03-31 15:16:46 +11004881 } else {
NeilBrown7a661382009-03-31 15:21:40 +11004882 writepos += reshape_sectors;
NeilBrowned37d832009-05-27 21:39:05 +10004883 readpos -= min_t(sector_t, reshape_sectors, readpos);
4884 safepos -= min_t(sector_t, reshape_sectors, safepos);
NeilBrownfef9c612009-03-31 15:16:46 +11004885 }
NeilBrown52c03292006-06-26 00:27:43 -07004886
NeilBrownb5254dd2012-05-21 09:27:01 +10004887 /* Having calculated the 'writepos' possibly use it
4888 * to set 'stripe_addr' which is where we will write to.
4889 */
4890 if (mddev->reshape_backwards) {
4891 BUG_ON(conf->reshape_progress == 0);
4892 stripe_addr = writepos;
4893 BUG_ON((mddev->dev_sectors &
4894 ~((sector_t)reshape_sectors - 1))
4895 - reshape_sectors - stripe_addr
4896 != sector_nr);
4897 } else {
4898 BUG_ON(writepos != sector_nr + reshape_sectors);
4899 stripe_addr = sector_nr;
4900 }
4901
NeilBrownc8f517c2009-03-31 15:28:40 +11004902 /* 'writepos' is the most advanced device address we might write.
4903 * 'readpos' is the least advanced device address we might read.
4904 * 'safepos' is the least address recorded in the metadata as having
4905 * been reshaped.
NeilBrownb5254dd2012-05-21 09:27:01 +10004906 * If there is a min_offset_diff, these are adjusted either by
4907 * increasing the safepos/readpos if diff is negative, or
4908 * increasing writepos if diff is positive.
4909 * If 'readpos' is then behind 'writepos', there is no way that we can
NeilBrownc8f517c2009-03-31 15:28:40 +11004910 * ensure safety in the face of a crash - that must be done by userspace
4911 * making a backup of the data. So in that case there is no particular
4912 * rush to update metadata.
4913 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4914 * update the metadata to advance 'safepos' to match 'readpos' so that
4915 * we can be safe in the event of a crash.
4916 * So we insist on updating metadata if safepos is behind writepos and
4917 * readpos is beyond writepos.
4918 * In any case, update the metadata every 10 seconds.
4919 * Maybe that number should be configurable, but I'm not sure it is
4920 * worth it.... maybe it could be a multiple of safemode_delay???
4921 */
NeilBrownb5254dd2012-05-21 09:27:01 +10004922 if (conf->min_offset_diff < 0) {
4923 safepos += -conf->min_offset_diff;
4924 readpos += -conf->min_offset_diff;
4925 } else
4926 writepos += conf->min_offset_diff;
4927
NeilBrown2c810cd2012-05-21 09:27:00 +10004928 if ((mddev->reshape_backwards
NeilBrownc8f517c2009-03-31 15:28:40 +11004929 ? (safepos > writepos && readpos < writepos)
4930 : (safepos < writepos && readpos > writepos)) ||
4931 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
NeilBrown52c03292006-06-26 00:27:43 -07004932 /* Cannot proceed until we've updated the superblock... */
4933 wait_event(conf->wait_for_overlap,
NeilBrownc91abf52013-11-19 12:02:01 +11004934 atomic_read(&conf->reshape_stripes)==0
4935 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4936 if (atomic_read(&conf->reshape_stripes) != 0)
4937 return 0;
NeilBrownfef9c612009-03-31 15:16:46 +11004938 mddev->reshape_position = conf->reshape_progress;
NeilBrown75d3da42011-01-14 09:14:34 +11004939 mddev->curr_resync_completed = sector_nr;
NeilBrownc8f517c2009-03-31 15:28:40 +11004940 conf->reshape_checkpoint = jiffies;
NeilBrown850b2b42006-10-03 01:15:46 -07004941 set_bit(MD_CHANGE_DEVS, &mddev->flags);
NeilBrown52c03292006-06-26 00:27:43 -07004942 md_wakeup_thread(mddev->thread);
NeilBrown850b2b42006-10-03 01:15:46 -07004943 wait_event(mddev->sb_wait, mddev->flags == 0 ||
NeilBrownc91abf52013-11-19 12:02:01 +11004944 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4945 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4946 return 0;
NeilBrown52c03292006-06-26 00:27:43 -07004947 spin_lock_irq(&conf->device_lock);
NeilBrownfef9c612009-03-31 15:16:46 +11004948 conf->reshape_safe = mddev->reshape_position;
NeilBrown52c03292006-06-26 00:27:43 -07004949 spin_unlock_irq(&conf->device_lock);
4950 wake_up(&conf->wait_for_overlap);
NeilBrownacb180b2009-04-14 16:28:34 +10004951 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
NeilBrown52c03292006-06-26 00:27:43 -07004952 }
4953
NeilBrownab69ae12009-03-31 15:26:47 +11004954 INIT_LIST_HEAD(&stripes);
NeilBrown7a661382009-03-31 15:21:40 +11004955 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
NeilBrown52c03292006-06-26 00:27:43 -07004956 int j;
NeilBrowna9f326e2009-09-23 18:06:41 +10004957 int skipped_disk = 0;
NeilBrowna8c906c2009-06-09 14:39:59 +10004958 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
NeilBrown52c03292006-06-26 00:27:43 -07004959 set_bit(STRIPE_EXPANDING, &sh->state);
4960 atomic_inc(&conf->reshape_stripes);
4961 /* If any of this stripe is beyond the end of the old
4962 * array, then we need to zero those blocks
4963 */
4964 for (j=sh->disks; j--;) {
4965 sector_t s;
4966 if (j == sh->pd_idx)
4967 continue;
NeilBrownf4168852007-02-28 20:11:53 -08004968 if (conf->level == 6 &&
NeilBrownd0dabf72009-03-31 14:39:38 +11004969 j == sh->qd_idx)
NeilBrownf4168852007-02-28 20:11:53 -08004970 continue;
NeilBrown784052e2009-03-31 15:19:07 +11004971 s = compute_blocknr(sh, j, 0);
Dan Williamsb522adc2009-03-31 15:00:31 +11004972 if (s < raid5_size(mddev, 0, 0)) {
NeilBrowna9f326e2009-09-23 18:06:41 +10004973 skipped_disk = 1;
NeilBrown52c03292006-06-26 00:27:43 -07004974 continue;
4975 }
4976 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4977 set_bit(R5_Expanded, &sh->dev[j].flags);
4978 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4979 }
NeilBrowna9f326e2009-09-23 18:06:41 +10004980 if (!skipped_disk) {
NeilBrown52c03292006-06-26 00:27:43 -07004981 set_bit(STRIPE_EXPAND_READY, &sh->state);
4982 set_bit(STRIPE_HANDLE, &sh->state);
4983 }
NeilBrownab69ae12009-03-31 15:26:47 +11004984 list_add(&sh->lru, &stripes);
NeilBrown52c03292006-06-26 00:27:43 -07004985 }
4986 spin_lock_irq(&conf->device_lock);
NeilBrown2c810cd2012-05-21 09:27:00 +10004987 if (mddev->reshape_backwards)
NeilBrown7a661382009-03-31 15:21:40 +11004988 conf->reshape_progress -= reshape_sectors * new_data_disks;
NeilBrownfef9c612009-03-31 15:16:46 +11004989 else
NeilBrown7a661382009-03-31 15:21:40 +11004990 conf->reshape_progress += reshape_sectors * new_data_disks;
NeilBrown52c03292006-06-26 00:27:43 -07004991 spin_unlock_irq(&conf->device_lock);
4992 /* Ok, those stripe are ready. We can start scheduling
4993 * reads on the source stripes.
4994 * The source stripes are determined by mapping the first and last
4995 * block on the destination stripes.
4996 */
NeilBrown52c03292006-06-26 00:27:43 -07004997 first_sector =
NeilBrownec32a2b2009-03-31 15:17:38 +11004998 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
NeilBrown911d4ee2009-03-31 14:39:38 +11004999 1, &dd_idx, NULL);
NeilBrown52c03292006-06-26 00:27:43 -07005000 last_sector =
NeilBrown0e6e0272009-06-09 16:32:22 +10005001 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
Andre Noll09c9e5f2009-06-18 08:45:55 +10005002 * new_data_disks - 1),
NeilBrown911d4ee2009-03-31 14:39:38 +11005003 1, &dd_idx, NULL);
Andre Noll58c0fed2009-03-31 14:33:13 +11005004 if (last_sector >= mddev->dev_sectors)
5005 last_sector = mddev->dev_sectors - 1;
NeilBrown52c03292006-06-26 00:27:43 -07005006 while (first_sector <= last_sector) {
NeilBrowna8c906c2009-06-09 14:39:59 +10005007 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
NeilBrown52c03292006-06-26 00:27:43 -07005008 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5009 set_bit(STRIPE_HANDLE, &sh->state);
5010 release_stripe(sh);
5011 first_sector += STRIPE_SECTORS;
5012 }
NeilBrownab69ae12009-03-31 15:26:47 +11005013 /* Now that the sources are clearly marked, we can release
5014 * the destination stripes
5015 */
5016 while (!list_empty(&stripes)) {
5017 sh = list_entry(stripes.next, struct stripe_head, lru);
5018 list_del_init(&sh->lru);
5019 release_stripe(sh);
5020 }
NeilBrownc6207272008-02-06 01:39:52 -08005021 /* If this takes us to the resync_max point where we have to pause,
5022 * then we need to write out the superblock.
5023 */
NeilBrown7a661382009-03-31 15:21:40 +11005024 sector_nr += reshape_sectors;
NeilBrownc03f6a12009-04-17 11:06:30 +10005025 if ((sector_nr - mddev->curr_resync_completed) * 2
5026 >= mddev->resync_max - mddev->curr_resync_completed) {
NeilBrownc6207272008-02-06 01:39:52 -08005027 /* Cannot proceed until we've updated the superblock... */
5028 wait_event(conf->wait_for_overlap,
NeilBrownc91abf52013-11-19 12:02:01 +11005029 atomic_read(&conf->reshape_stripes) == 0
5030 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5031 if (atomic_read(&conf->reshape_stripes) != 0)
5032 goto ret;
NeilBrownfef9c612009-03-31 15:16:46 +11005033 mddev->reshape_position = conf->reshape_progress;
NeilBrown75d3da42011-01-14 09:14:34 +11005034 mddev->curr_resync_completed = sector_nr;
NeilBrownc8f517c2009-03-31 15:28:40 +11005035 conf->reshape_checkpoint = jiffies;
NeilBrownc6207272008-02-06 01:39:52 -08005036 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5037 md_wakeup_thread(mddev->thread);
5038 wait_event(mddev->sb_wait,
5039 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
NeilBrownc91abf52013-11-19 12:02:01 +11005040 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5041 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5042 goto ret;
NeilBrownc6207272008-02-06 01:39:52 -08005043 spin_lock_irq(&conf->device_lock);
NeilBrownfef9c612009-03-31 15:16:46 +11005044 conf->reshape_safe = mddev->reshape_position;
NeilBrownc6207272008-02-06 01:39:52 -08005045 spin_unlock_irq(&conf->device_lock);
5046 wake_up(&conf->wait_for_overlap);
NeilBrownacb180b2009-04-14 16:28:34 +10005047 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
NeilBrownc6207272008-02-06 01:39:52 -08005048 }
NeilBrownc91abf52013-11-19 12:02:01 +11005049ret:
NeilBrown7a661382009-03-31 15:21:40 +11005050 return reshape_sectors;
NeilBrown52c03292006-06-26 00:27:43 -07005051}
5052
5053/* FIXME go_faster isn't used */
NeilBrownfd01b882011-10-11 16:47:53 +11005054static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
NeilBrown52c03292006-06-26 00:27:43 -07005055{
NeilBrownd1688a62011-10-11 16:49:52 +11005056 struct r5conf *conf = mddev->private;
NeilBrown52c03292006-06-26 00:27:43 -07005057 struct stripe_head *sh;
Andre Noll58c0fed2009-03-31 14:33:13 +11005058 sector_t max_sector = mddev->dev_sectors;
NeilBrown57dab0b2010-10-19 10:03:39 +11005059 sector_t sync_blocks;
NeilBrown16a53ec2006-06-26 00:27:38 -07005060 int still_degraded = 0;
5061 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005062
NeilBrown72626682005-09-09 16:23:54 -07005063 if (sector_nr >= max_sector) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005064 /* just being told to finish up .. nothing much to do */
NeilBrowncea9c222009-03-31 15:15:05 +11005065
NeilBrown29269552006-03-27 01:18:10 -08005066 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5067 end_reshape(conf);
5068 return 0;
5069 }
NeilBrown72626682005-09-09 16:23:54 -07005070
5071 if (mddev->curr_resync < max_sector) /* aborted */
5072 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5073 &sync_blocks, 1);
NeilBrown16a53ec2006-06-26 00:27:38 -07005074 else /* completed sync */
NeilBrown72626682005-09-09 16:23:54 -07005075 conf->fullsync = 0;
5076 bitmap_close_sync(mddev->bitmap);
5077
Linus Torvalds1da177e2005-04-16 15:20:36 -07005078 return 0;
5079 }
NeilBrownccfcc3c2006-03-27 01:18:09 -08005080
NeilBrown64bd6602009-08-03 10:59:58 +10005081 /* Allow raid5_quiesce to complete */
5082 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5083
NeilBrown52c03292006-06-26 00:27:43 -07005084 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5085 return reshape_request(mddev, sector_nr, skipped);
NeilBrownf6705572006-03-27 01:18:11 -08005086
NeilBrownc6207272008-02-06 01:39:52 -08005087 /* No need to check resync_max as we never do more than one
5088 * stripe, and as resync_max will always be on a chunk boundary,
5089 * if the check in md_do_sync didn't fire, there is no chance
5090 * of overstepping resync_max here
5091 */
5092
NeilBrown16a53ec2006-06-26 00:27:38 -07005093 /* if there is too many failed drives and we are trying
Linus Torvalds1da177e2005-04-16 15:20:36 -07005094 * to resync, then assert that we are finished, because there is
5095 * nothing we can do.
5096 */
NeilBrown3285edf2006-06-26 00:27:55 -07005097 if (mddev->degraded >= conf->max_degraded &&
NeilBrown16a53ec2006-06-26 00:27:38 -07005098 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
Andre Noll58c0fed2009-03-31 14:33:13 +11005099 sector_t rv = mddev->dev_sectors - sector_nr;
NeilBrown57afd892005-06-21 17:17:13 -07005100 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005101 return rv;
5102 }
majianpeng6f608042013-04-24 11:42:41 +10005103 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5104 !conf->fullsync &&
5105 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5106 sync_blocks >= STRIPE_SECTORS) {
NeilBrown72626682005-09-09 16:23:54 -07005107 /* we can skip this block, and probably more */
5108 sync_blocks /= STRIPE_SECTORS;
5109 *skipped = 1;
5110 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5111 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005112
NeilBrownb47490c2008-02-06 01:39:50 -08005113 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5114
NeilBrowna8c906c2009-06-09 14:39:59 +10005115 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005116 if (sh == NULL) {
NeilBrowna8c906c2009-06-09 14:39:59 +10005117 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005118 /* make sure we don't swamp the stripe cache if someone else
NeilBrown16a53ec2006-06-26 00:27:38 -07005119 * is trying to get access
Linus Torvalds1da177e2005-04-16 15:20:36 -07005120 */
Nishanth Aravamudan66c006a2005-11-07 01:01:17 -08005121 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005122 }
NeilBrown16a53ec2006-06-26 00:27:38 -07005123 /* Need to check if array will still be degraded after recovery/resync
5124 * We don't need to check the 'failed' flag as when that gets set,
5125 * recovery aborts.
5126 */
NeilBrownf001a702009-06-09 14:30:31 +10005127 for (i = 0; i < conf->raid_disks; i++)
NeilBrown16a53ec2006-06-26 00:27:38 -07005128 if (conf->disks[i].rdev == NULL)
5129 still_degraded = 1;
5130
5131 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5132
NeilBrown83206d62011-07-26 11:19:49 +10005133 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
Eivind Sarto053f5b62014-06-09 17:06:19 -07005134 set_bit(STRIPE_HANDLE, &sh->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005135
Linus Torvalds1da177e2005-04-16 15:20:36 -07005136 release_stripe(sh);
5137
5138 return STRIPE_SECTORS;
5139}
5140
NeilBrownd1688a62011-10-11 16:49:52 +11005141static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005142{
5143 /* We may not be able to submit a whole bio at once as there
5144 * may not be enough stripe_heads available.
5145 * We cannot pre-allocate enough stripe_heads as we may need
5146 * more than exist in the cache (if we allow ever large chunks).
5147 * So we do one stripe head at a time and record in
5148 * ->bi_hw_segments how many have been done.
5149 *
5150 * We *know* that this entire raid_bio is in one chunk, so
5151 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5152 */
5153 struct stripe_head *sh;
NeilBrown911d4ee2009-03-31 14:39:38 +11005154 int dd_idx;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005155 sector_t sector, logical_sector, last_sector;
5156 int scnt = 0;
5157 int remaining;
5158 int handled = 0;
5159
Kent Overstreet4f024f32013-10-11 15:44:27 -07005160 logical_sector = raid_bio->bi_iter.bi_sector &
5161 ~((sector_t)STRIPE_SECTORS-1);
NeilBrown112bf892009-03-31 14:39:38 +11005162 sector = raid5_compute_sector(conf, logical_sector,
NeilBrown911d4ee2009-03-31 14:39:38 +11005163 0, &dd_idx, NULL);
Kent Overstreetf73a1c72012-09-25 15:05:12 -07005164 last_sector = bio_end_sector(raid_bio);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005165
5166 for (; logical_sector < last_sector;
Neil Brown387bb172007-02-08 14:20:29 -08005167 logical_sector += STRIPE_SECTORS,
5168 sector += STRIPE_SECTORS,
5169 scnt++) {
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005170
Shaohua Lie7836bd62012-07-19 16:01:31 +10005171 if (scnt < raid5_bi_processed_stripes(raid_bio))
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005172 /* already done this stripe */
5173 continue;
5174
hui jiao2844dc32014-06-05 11:34:24 +08005175 sh = get_active_stripe(conf, sector, 0, 1, 1);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005176
5177 if (!sh) {
5178 /* failed to get a stripe - must wait */
Shaohua Lie7836bd62012-07-19 16:01:31 +10005179 raid5_set_bi_processed_stripes(raid_bio, scnt);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005180 conf->retry_read_aligned = raid_bio;
5181 return handled;
5182 }
5183
Neil Brown387bb172007-02-08 14:20:29 -08005184 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
5185 release_stripe(sh);
Shaohua Lie7836bd62012-07-19 16:01:31 +10005186 raid5_set_bi_processed_stripes(raid_bio, scnt);
Neil Brown387bb172007-02-08 14:20:29 -08005187 conf->retry_read_aligned = raid_bio;
5188 return handled;
5189 }
5190
majianpeng3f9e7c12012-07-31 10:04:21 +10005191 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
Dan Williams36d1c642009-07-14 11:48:22 -07005192 handle_stripe(sh);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005193 release_stripe(sh);
5194 handled++;
5195 }
Shaohua Lie7836bd62012-07-19 16:01:31 +10005196 remaining = raid5_dec_bi_active_stripes(raid_bio);
Linus Torvalds0a82a8d2013-04-18 09:00:26 -07005197 if (remaining == 0) {
5198 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5199 raid_bio, 0);
Neil Brown0e13fe232008-06-28 08:31:20 +10005200 bio_endio(raid_bio, 0);
Linus Torvalds0a82a8d2013-04-18 09:00:26 -07005201 }
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005202 if (atomic_dec_and_test(&conf->active_aligned_reads))
5203 wake_up(&conf->wait_for_stripe);
5204 return handled;
5205}
5206
Shaohua Libfc90cb2013-08-29 15:40:32 +08005207static int handle_active_stripes(struct r5conf *conf, int group,
Shaohua Li566c09c2013-11-14 15:16:17 +11005208 struct r5worker *worker,
5209 struct list_head *temp_inactive_list)
Shaohua Li46a06402012-08-02 08:33:15 +10005210{
5211 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
Shaohua Li566c09c2013-11-14 15:16:17 +11005212 int i, batch_size = 0, hash;
5213 bool release_inactive = false;
Shaohua Li46a06402012-08-02 08:33:15 +10005214
5215 while (batch_size < MAX_STRIPE_BATCH &&
Shaohua Li851c30c2013-08-28 14:30:16 +08005216 (sh = __get_priority_stripe(conf, group)) != NULL)
Shaohua Li46a06402012-08-02 08:33:15 +10005217 batch[batch_size++] = sh;
5218
Shaohua Li566c09c2013-11-14 15:16:17 +11005219 if (batch_size == 0) {
5220 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5221 if (!list_empty(temp_inactive_list + i))
5222 break;
5223 if (i == NR_STRIPE_HASH_LOCKS)
5224 return batch_size;
5225 release_inactive = true;
5226 }
Shaohua Li46a06402012-08-02 08:33:15 +10005227 spin_unlock_irq(&conf->device_lock);
5228
Shaohua Li566c09c2013-11-14 15:16:17 +11005229 release_inactive_stripe_list(conf, temp_inactive_list,
5230 NR_STRIPE_HASH_LOCKS);
5231
5232 if (release_inactive) {
5233 spin_lock_irq(&conf->device_lock);
5234 return 0;
5235 }
5236
Shaohua Li46a06402012-08-02 08:33:15 +10005237 for (i = 0; i < batch_size; i++)
5238 handle_stripe(batch[i]);
5239
5240 cond_resched();
5241
5242 spin_lock_irq(&conf->device_lock);
Shaohua Li566c09c2013-11-14 15:16:17 +11005243 for (i = 0; i < batch_size; i++) {
5244 hash = batch[i]->hash_lock_index;
5245 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5246 }
Shaohua Li46a06402012-08-02 08:33:15 +10005247 return batch_size;
5248}
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005249
Shaohua Li851c30c2013-08-28 14:30:16 +08005250static void raid5_do_work(struct work_struct *work)
5251{
5252 struct r5worker *worker = container_of(work, struct r5worker, work);
5253 struct r5worker_group *group = worker->group;
5254 struct r5conf *conf = group->conf;
5255 int group_id = group - conf->worker_groups;
5256 int handled;
5257 struct blk_plug plug;
5258
5259 pr_debug("+++ raid5worker active\n");
5260
5261 blk_start_plug(&plug);
5262 handled = 0;
5263 spin_lock_irq(&conf->device_lock);
5264 while (1) {
5265 int batch_size, released;
5266
Shaohua Li566c09c2013-11-14 15:16:17 +11005267 released = release_stripe_list(conf, worker->temp_inactive_list);
Shaohua Li851c30c2013-08-28 14:30:16 +08005268
Shaohua Li566c09c2013-11-14 15:16:17 +11005269 batch_size = handle_active_stripes(conf, group_id, worker,
5270 worker->temp_inactive_list);
Shaohua Libfc90cb2013-08-29 15:40:32 +08005271 worker->working = false;
Shaohua Li851c30c2013-08-28 14:30:16 +08005272 if (!batch_size && !released)
5273 break;
5274 handled += batch_size;
5275 }
5276 pr_debug("%d stripes handled\n", handled);
5277
5278 spin_unlock_irq(&conf->device_lock);
5279 blk_finish_plug(&plug);
5280
5281 pr_debug("--- raid5worker inactive\n");
5282}
5283
Linus Torvalds1da177e2005-04-16 15:20:36 -07005284/*
5285 * This is our raid5 kernel thread.
5286 *
5287 * We scan the hash table for stripes which can be handled now.
5288 * During the scan, completed stripes are saved for us by the interrupt
5289 * handler, so that they will not have to wait for our next wakeup.
5290 */
Shaohua Li4ed87312012-10-11 13:34:00 +11005291static void raid5d(struct md_thread *thread)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005292{
Shaohua Li4ed87312012-10-11 13:34:00 +11005293 struct mddev *mddev = thread->mddev;
NeilBrownd1688a62011-10-11 16:49:52 +11005294 struct r5conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005295 int handled;
NeilBrowne1dfa0a2011-04-18 18:25:41 +10005296 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005297
Dan Williams45b42332007-07-09 11:56:43 -07005298 pr_debug("+++ raid5d active\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005299
5300 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005301
NeilBrowne1dfa0a2011-04-18 18:25:41 +10005302 blk_start_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005303 handled = 0;
5304 spin_lock_irq(&conf->device_lock);
5305 while (1) {
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005306 struct bio *bio;
Shaohua Li773ca822013-08-27 17:50:39 +08005307 int batch_size, released;
5308
Shaohua Li566c09c2013-11-14 15:16:17 +11005309 released = release_stripe_list(conf, conf->temp_inactive_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005310
NeilBrown0021b7b2012-07-31 09:08:14 +02005311 if (
NeilBrown7c13edc2011-04-18 18:25:43 +10005312 !list_empty(&conf->bitmap_list)) {
5313 /* Now is a good time to flush some bitmap updates */
5314 conf->seq_flush++;
NeilBrown700e4322005-11-28 13:44:10 -08005315 spin_unlock_irq(&conf->device_lock);
NeilBrown72626682005-09-09 16:23:54 -07005316 bitmap_unplug(mddev->bitmap);
NeilBrown700e4322005-11-28 13:44:10 -08005317 spin_lock_irq(&conf->device_lock);
NeilBrown7c13edc2011-04-18 18:25:43 +10005318 conf->seq_write = conf->seq_flush;
Shaohua Li566c09c2013-11-14 15:16:17 +11005319 activate_bit_delay(conf, conf->temp_inactive_list);
NeilBrown72626682005-09-09 16:23:54 -07005320 }
NeilBrown0021b7b2012-07-31 09:08:14 +02005321 raid5_activate_delayed(conf);
NeilBrown72626682005-09-09 16:23:54 -07005322
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08005323 while ((bio = remove_bio_from_retry(conf))) {
5324 int ok;
5325 spin_unlock_irq(&conf->device_lock);
5326 ok = retry_aligned_read(conf, bio);
5327 spin_lock_irq(&conf->device_lock);
5328 if (!ok)
5329 break;
5330 handled++;
5331 }
5332
Shaohua Li566c09c2013-11-14 15:16:17 +11005333 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5334 conf->temp_inactive_list);
Shaohua Li773ca822013-08-27 17:50:39 +08005335 if (!batch_size && !released)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005336 break;
Shaohua Li46a06402012-08-02 08:33:15 +10005337 handled += batch_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005338
Shaohua Li46a06402012-08-02 08:33:15 +10005339 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5340 spin_unlock_irq(&conf->device_lock);
NeilBrownde393cd2011-07-28 11:31:48 +10005341 md_check_recovery(mddev);
Shaohua Li46a06402012-08-02 08:33:15 +10005342 spin_lock_irq(&conf->device_lock);
5343 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005344 }
Dan Williams45b42332007-07-09 11:56:43 -07005345 pr_debug("%d stripes handled\n", handled);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005346
5347 spin_unlock_irq(&conf->device_lock);
5348
Dan Williamsc9f21aa2008-07-23 12:05:51 -07005349 async_tx_issue_pending_all();
NeilBrowne1dfa0a2011-04-18 18:25:41 +10005350 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005351
Dan Williams45b42332007-07-09 11:56:43 -07005352 pr_debug("--- raid5d inactive\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005353}
5354
NeilBrown3f294f42005-11-08 21:39:25 -08005355static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11005356raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
NeilBrown3f294f42005-11-08 21:39:25 -08005357{
NeilBrown7b1485b2014-12-15 12:56:59 +11005358 struct r5conf *conf;
5359 int ret = 0;
5360 spin_lock(&mddev->lock);
5361 conf = mddev->private;
NeilBrown96de1e62005-11-08 21:39:39 -08005362 if (conf)
NeilBrown7b1485b2014-12-15 12:56:59 +11005363 ret = sprintf(page, "%d\n", conf->max_nr_stripes);
5364 spin_unlock(&mddev->lock);
5365 return ret;
NeilBrown3f294f42005-11-08 21:39:25 -08005366}
5367
NeilBrownc41d4ac2010-06-01 19:37:24 +10005368int
NeilBrownfd01b882011-10-11 16:47:53 +11005369raid5_set_cache_size(struct mddev *mddev, int size)
NeilBrownc41d4ac2010-06-01 19:37:24 +10005370{
NeilBrownd1688a62011-10-11 16:49:52 +11005371 struct r5conf *conf = mddev->private;
NeilBrownc41d4ac2010-06-01 19:37:24 +10005372 int err;
Shaohua Li566c09c2013-11-14 15:16:17 +11005373 int hash;
NeilBrownc41d4ac2010-06-01 19:37:24 +10005374
5375 if (size <= 16 || size > 32768)
5376 return -EINVAL;
Shaohua Li566c09c2013-11-14 15:16:17 +11005377 hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
NeilBrownc41d4ac2010-06-01 19:37:24 +10005378 while (size < conf->max_nr_stripes) {
Shaohua Li566c09c2013-11-14 15:16:17 +11005379 if (drop_one_stripe(conf, hash))
NeilBrownc41d4ac2010-06-01 19:37:24 +10005380 conf->max_nr_stripes--;
5381 else
5382 break;
Shaohua Li566c09c2013-11-14 15:16:17 +11005383 hash--;
5384 if (hash < 0)
5385 hash = NR_STRIPE_HASH_LOCKS - 1;
NeilBrownc41d4ac2010-06-01 19:37:24 +10005386 }
5387 err = md_allow_write(mddev);
5388 if (err)
5389 return err;
Shaohua Li566c09c2013-11-14 15:16:17 +11005390 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
NeilBrownc41d4ac2010-06-01 19:37:24 +10005391 while (size > conf->max_nr_stripes) {
Shaohua Li566c09c2013-11-14 15:16:17 +11005392 if (grow_one_stripe(conf, hash))
NeilBrownc41d4ac2010-06-01 19:37:24 +10005393 conf->max_nr_stripes++;
5394 else break;
Shaohua Li566c09c2013-11-14 15:16:17 +11005395 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
NeilBrownc41d4ac2010-06-01 19:37:24 +10005396 }
5397 return 0;
5398}
5399EXPORT_SYMBOL(raid5_set_cache_size);
5400
NeilBrown3f294f42005-11-08 21:39:25 -08005401static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11005402raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
NeilBrown3f294f42005-11-08 21:39:25 -08005403{
NeilBrown67918752014-12-15 12:57:01 +11005404 struct r5conf *conf;
Dan Williams4ef197d82008-04-28 02:15:54 -07005405 unsigned long new;
Dan Williamsb5470dc2008-06-27 21:44:04 -07005406 int err;
5407
NeilBrown3f294f42005-11-08 21:39:25 -08005408 if (len >= PAGE_SIZE)
5409 return -EINVAL;
Jingoo Hanb29bebd2013-06-01 16:15:16 +09005410 if (kstrtoul(page, 10, &new))
NeilBrown3f294f42005-11-08 21:39:25 -08005411 return -EINVAL;
NeilBrown67918752014-12-15 12:57:01 +11005412 err = mddev_lock(mddev);
Dan Williamsb5470dc2008-06-27 21:44:04 -07005413 if (err)
5414 return err;
NeilBrown67918752014-12-15 12:57:01 +11005415 conf = mddev->private;
5416 if (!conf)
5417 err = -ENODEV;
5418 else
5419 err = raid5_set_cache_size(mddev, new);
5420 mddev_unlock(mddev);
5421
5422 return err ?: len;
NeilBrown3f294f42005-11-08 21:39:25 -08005423}
NeilBrown007583c2005-11-08 21:39:30 -08005424
NeilBrown96de1e62005-11-08 21:39:39 -08005425static struct md_sysfs_entry
5426raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5427 raid5_show_stripe_cache_size,
5428 raid5_store_stripe_cache_size);
NeilBrown3f294f42005-11-08 21:39:25 -08005429
5430static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11005431raid5_show_preread_threshold(struct mddev *mddev, char *page)
Dan Williams8b3e6cd2008-04-28 02:15:53 -07005432{
NeilBrown7b1485b2014-12-15 12:56:59 +11005433 struct r5conf *conf;
5434 int ret = 0;
5435 spin_lock(&mddev->lock);
5436 conf = mddev->private;
Dan Williams8b3e6cd2008-04-28 02:15:53 -07005437 if (conf)
NeilBrown7b1485b2014-12-15 12:56:59 +11005438 ret = sprintf(page, "%d\n", conf->bypass_threshold);
5439 spin_unlock(&mddev->lock);
5440 return ret;
Dan Williams8b3e6cd2008-04-28 02:15:53 -07005441}
5442
5443static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11005444raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
Dan Williams8b3e6cd2008-04-28 02:15:53 -07005445{
NeilBrown67918752014-12-15 12:57:01 +11005446 struct r5conf *conf;
Dan Williams4ef197d82008-04-28 02:15:54 -07005447 unsigned long new;
NeilBrown67918752014-12-15 12:57:01 +11005448 int err;
5449
Dan Williams8b3e6cd2008-04-28 02:15:53 -07005450 if (len >= PAGE_SIZE)
5451 return -EINVAL;
Jingoo Hanb29bebd2013-06-01 16:15:16 +09005452 if (kstrtoul(page, 10, &new))
Dan Williams8b3e6cd2008-04-28 02:15:53 -07005453 return -EINVAL;
NeilBrown67918752014-12-15 12:57:01 +11005454
5455 err = mddev_lock(mddev);
5456 if (err)
5457 return err;
5458 conf = mddev->private;
5459 if (!conf)
5460 err = -ENODEV;
5461 else if (new > conf->max_nr_stripes)
5462 err = -EINVAL;
5463 else
5464 conf->bypass_threshold = new;
5465 mddev_unlock(mddev);
5466 return err ?: len;
Dan Williams8b3e6cd2008-04-28 02:15:53 -07005467}
5468
5469static struct md_sysfs_entry
5470raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5471 S_IRUGO | S_IWUSR,
5472 raid5_show_preread_threshold,
5473 raid5_store_preread_threshold);
5474
5475static ssize_t
Shaohua Lid592a992014-05-21 17:57:44 +08005476raid5_show_skip_copy(struct mddev *mddev, char *page)
5477{
NeilBrown7b1485b2014-12-15 12:56:59 +11005478 struct r5conf *conf;
5479 int ret = 0;
5480 spin_lock(&mddev->lock);
5481 conf = mddev->private;
Shaohua Lid592a992014-05-21 17:57:44 +08005482 if (conf)
NeilBrown7b1485b2014-12-15 12:56:59 +11005483 ret = sprintf(page, "%d\n", conf->skip_copy);
5484 spin_unlock(&mddev->lock);
5485 return ret;
Shaohua Lid592a992014-05-21 17:57:44 +08005486}
5487
5488static ssize_t
5489raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
5490{
NeilBrown67918752014-12-15 12:57:01 +11005491 struct r5conf *conf;
Shaohua Lid592a992014-05-21 17:57:44 +08005492 unsigned long new;
NeilBrown67918752014-12-15 12:57:01 +11005493 int err;
5494
Shaohua Lid592a992014-05-21 17:57:44 +08005495 if (len >= PAGE_SIZE)
5496 return -EINVAL;
Shaohua Lid592a992014-05-21 17:57:44 +08005497 if (kstrtoul(page, 10, &new))
5498 return -EINVAL;
5499 new = !!new;
Shaohua Lid592a992014-05-21 17:57:44 +08005500
NeilBrown67918752014-12-15 12:57:01 +11005501 err = mddev_lock(mddev);
5502 if (err)
5503 return err;
5504 conf = mddev->private;
5505 if (!conf)
5506 err = -ENODEV;
5507 else if (new != conf->skip_copy) {
5508 mddev_suspend(mddev);
5509 conf->skip_copy = new;
5510 if (new)
5511 mddev->queue->backing_dev_info.capabilities |=
5512 BDI_CAP_STABLE_WRITES;
5513 else
5514 mddev->queue->backing_dev_info.capabilities &=
5515 ~BDI_CAP_STABLE_WRITES;
5516 mddev_resume(mddev);
5517 }
5518 mddev_unlock(mddev);
5519 return err ?: len;
Shaohua Lid592a992014-05-21 17:57:44 +08005520}
5521
5522static struct md_sysfs_entry
5523raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
5524 raid5_show_skip_copy,
5525 raid5_store_skip_copy);
5526
Shaohua Lid592a992014-05-21 17:57:44 +08005527static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11005528stripe_cache_active_show(struct mddev *mddev, char *page)
NeilBrown3f294f42005-11-08 21:39:25 -08005529{
NeilBrownd1688a62011-10-11 16:49:52 +11005530 struct r5conf *conf = mddev->private;
NeilBrown96de1e62005-11-08 21:39:39 -08005531 if (conf)
5532 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5533 else
5534 return 0;
NeilBrown3f294f42005-11-08 21:39:25 -08005535}
5536
NeilBrown96de1e62005-11-08 21:39:39 -08005537static struct md_sysfs_entry
5538raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
NeilBrown3f294f42005-11-08 21:39:25 -08005539
Shaohua Lib7214202013-08-27 17:50:42 +08005540static ssize_t
5541raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
5542{
NeilBrown7b1485b2014-12-15 12:56:59 +11005543 struct r5conf *conf;
5544 int ret = 0;
5545 spin_lock(&mddev->lock);
5546 conf = mddev->private;
Shaohua Lib7214202013-08-27 17:50:42 +08005547 if (conf)
NeilBrown7b1485b2014-12-15 12:56:59 +11005548 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
5549 spin_unlock(&mddev->lock);
5550 return ret;
Shaohua Lib7214202013-08-27 17:50:42 +08005551}
5552
majianpeng60aaf932013-11-14 15:16:20 +11005553static int alloc_thread_groups(struct r5conf *conf, int cnt,
5554 int *group_cnt,
5555 int *worker_cnt_per_group,
5556 struct r5worker_group **worker_groups);
Shaohua Lib7214202013-08-27 17:50:42 +08005557static ssize_t
5558raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
5559{
NeilBrown67918752014-12-15 12:57:01 +11005560 struct r5conf *conf;
Shaohua Lib7214202013-08-27 17:50:42 +08005561 unsigned long new;
5562 int err;
majianpeng60aaf932013-11-14 15:16:20 +11005563 struct r5worker_group *new_groups, *old_groups;
5564 int group_cnt, worker_cnt_per_group;
Shaohua Lib7214202013-08-27 17:50:42 +08005565
5566 if (len >= PAGE_SIZE)
5567 return -EINVAL;
Shaohua Lib7214202013-08-27 17:50:42 +08005568 if (kstrtoul(page, 10, &new))
5569 return -EINVAL;
5570
NeilBrown67918752014-12-15 12:57:01 +11005571 err = mddev_lock(mddev);
Shaohua Lib7214202013-08-27 17:50:42 +08005572 if (err)
5573 return err;
NeilBrown67918752014-12-15 12:57:01 +11005574 conf = mddev->private;
5575 if (!conf)
5576 err = -ENODEV;
5577 else if (new != conf->worker_cnt_per_group) {
5578 mddev_suspend(mddev);
5579
5580 old_groups = conf->worker_groups;
5581 if (old_groups)
5582 flush_workqueue(raid5_wq);
5583
5584 err = alloc_thread_groups(conf, new,
5585 &group_cnt, &worker_cnt_per_group,
5586 &new_groups);
5587 if (!err) {
5588 spin_lock_irq(&conf->device_lock);
5589 conf->group_cnt = group_cnt;
5590 conf->worker_cnt_per_group = worker_cnt_per_group;
5591 conf->worker_groups = new_groups;
5592 spin_unlock_irq(&conf->device_lock);
5593
5594 if (old_groups)
5595 kfree(old_groups[0].workers);
5596 kfree(old_groups);
5597 }
5598 mddev_resume(mddev);
5599 }
5600 mddev_unlock(mddev);
5601
5602 return err ?: len;
Shaohua Lib7214202013-08-27 17:50:42 +08005603}
5604
5605static struct md_sysfs_entry
5606raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
5607 raid5_show_group_thread_cnt,
5608 raid5_store_group_thread_cnt);
5609
NeilBrown007583c2005-11-08 21:39:30 -08005610static struct attribute *raid5_attrs[] = {
NeilBrown3f294f42005-11-08 21:39:25 -08005611 &raid5_stripecache_size.attr,
5612 &raid5_stripecache_active.attr,
Dan Williams8b3e6cd2008-04-28 02:15:53 -07005613 &raid5_preread_bypass_threshold.attr,
Shaohua Lib7214202013-08-27 17:50:42 +08005614 &raid5_group_thread_cnt.attr,
Shaohua Lid592a992014-05-21 17:57:44 +08005615 &raid5_skip_copy.attr,
NeilBrown3f294f42005-11-08 21:39:25 -08005616 NULL,
5617};
NeilBrown007583c2005-11-08 21:39:30 -08005618static struct attribute_group raid5_attrs_group = {
5619 .name = NULL,
5620 .attrs = raid5_attrs,
NeilBrown3f294f42005-11-08 21:39:25 -08005621};
5622
majianpeng60aaf932013-11-14 15:16:20 +11005623static int alloc_thread_groups(struct r5conf *conf, int cnt,
5624 int *group_cnt,
5625 int *worker_cnt_per_group,
5626 struct r5worker_group **worker_groups)
Shaohua Li851c30c2013-08-28 14:30:16 +08005627{
Shaohua Li566c09c2013-11-14 15:16:17 +11005628 int i, j, k;
Shaohua Li851c30c2013-08-28 14:30:16 +08005629 ssize_t size;
5630 struct r5worker *workers;
5631
majianpeng60aaf932013-11-14 15:16:20 +11005632 *worker_cnt_per_group = cnt;
Shaohua Li851c30c2013-08-28 14:30:16 +08005633 if (cnt == 0) {
majianpeng60aaf932013-11-14 15:16:20 +11005634 *group_cnt = 0;
5635 *worker_groups = NULL;
Shaohua Li851c30c2013-08-28 14:30:16 +08005636 return 0;
5637 }
majianpeng60aaf932013-11-14 15:16:20 +11005638 *group_cnt = num_possible_nodes();
Shaohua Li851c30c2013-08-28 14:30:16 +08005639 size = sizeof(struct r5worker) * cnt;
majianpeng60aaf932013-11-14 15:16:20 +11005640 workers = kzalloc(size * *group_cnt, GFP_NOIO);
5641 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
5642 *group_cnt, GFP_NOIO);
5643 if (!*worker_groups || !workers) {
Shaohua Li851c30c2013-08-28 14:30:16 +08005644 kfree(workers);
majianpeng60aaf932013-11-14 15:16:20 +11005645 kfree(*worker_groups);
Shaohua Li851c30c2013-08-28 14:30:16 +08005646 return -ENOMEM;
5647 }
5648
majianpeng60aaf932013-11-14 15:16:20 +11005649 for (i = 0; i < *group_cnt; i++) {
Shaohua Li851c30c2013-08-28 14:30:16 +08005650 struct r5worker_group *group;
5651
NeilBrown0c775d52013-11-25 11:12:43 +11005652 group = &(*worker_groups)[i];
Shaohua Li851c30c2013-08-28 14:30:16 +08005653 INIT_LIST_HEAD(&group->handle_list);
5654 group->conf = conf;
5655 group->workers = workers + i * cnt;
5656
5657 for (j = 0; j < cnt; j++) {
Shaohua Li566c09c2013-11-14 15:16:17 +11005658 struct r5worker *worker = group->workers + j;
5659 worker->group = group;
5660 INIT_WORK(&worker->work, raid5_do_work);
5661
5662 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
5663 INIT_LIST_HEAD(worker->temp_inactive_list + k);
Shaohua Li851c30c2013-08-28 14:30:16 +08005664 }
5665 }
5666
5667 return 0;
5668}
5669
5670static void free_thread_groups(struct r5conf *conf)
5671{
5672 if (conf->worker_groups)
5673 kfree(conf->worker_groups[0].workers);
5674 kfree(conf->worker_groups);
5675 conf->worker_groups = NULL;
5676}
5677
Dan Williams80c3a6c2009-03-17 18:10:40 -07005678static sector_t
NeilBrownfd01b882011-10-11 16:47:53 +11005679raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
Dan Williams80c3a6c2009-03-17 18:10:40 -07005680{
NeilBrownd1688a62011-10-11 16:49:52 +11005681 struct r5conf *conf = mddev->private;
Dan Williams80c3a6c2009-03-17 18:10:40 -07005682
5683 if (!sectors)
5684 sectors = mddev->dev_sectors;
NeilBrown5e5e3e72009-10-16 16:35:30 +11005685 if (!raid_disks)
NeilBrown7ec05472009-03-31 15:10:36 +11005686 /* size is defined by the smallest of previous and new size */
NeilBrown5e5e3e72009-10-16 16:35:30 +11005687 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
Dan Williams80c3a6c2009-03-17 18:10:40 -07005688
Andre Noll9d8f0362009-06-18 08:45:01 +10005689 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
Andre Noll664e7c42009-06-18 08:45:27 +10005690 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
Dan Williams80c3a6c2009-03-17 18:10:40 -07005691 return sectors * (raid_disks - conf->max_degraded);
5692}
5693
Oleg Nesterov789b5e02014-02-06 03:42:45 +05305694static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5695{
5696 safe_put_page(percpu->spare_page);
5697 kfree(percpu->scribble);
5698 percpu->spare_page = NULL;
5699 percpu->scribble = NULL;
5700}
5701
5702static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5703{
5704 if (conf->level == 6 && !percpu->spare_page)
5705 percpu->spare_page = alloc_page(GFP_KERNEL);
5706 if (!percpu->scribble)
5707 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5708
5709 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
5710 free_scratch_buffer(conf, percpu);
5711 return -ENOMEM;
5712 }
5713
5714 return 0;
5715}
5716
NeilBrownd1688a62011-10-11 16:49:52 +11005717static void raid5_free_percpu(struct r5conf *conf)
Dan Williams36d1c642009-07-14 11:48:22 -07005718{
Dan Williams36d1c642009-07-14 11:48:22 -07005719 unsigned long cpu;
5720
5721 if (!conf->percpu)
5722 return;
5723
Dan Williams36d1c642009-07-14 11:48:22 -07005724#ifdef CONFIG_HOTPLUG_CPU
5725 unregister_cpu_notifier(&conf->cpu_notify);
5726#endif
Oleg Nesterov789b5e02014-02-06 03:42:45 +05305727
5728 get_online_cpus();
5729 for_each_possible_cpu(cpu)
5730 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
Dan Williams36d1c642009-07-14 11:48:22 -07005731 put_online_cpus();
5732
5733 free_percpu(conf->percpu);
5734}
5735
NeilBrownd1688a62011-10-11 16:49:52 +11005736static void free_conf(struct r5conf *conf)
Dan Williams95fc17a2009-07-31 12:39:15 +10005737{
Shaohua Li851c30c2013-08-28 14:30:16 +08005738 free_thread_groups(conf);
Dan Williams95fc17a2009-07-31 12:39:15 +10005739 shrink_stripes(conf);
Dan Williams36d1c642009-07-14 11:48:22 -07005740 raid5_free_percpu(conf);
Dan Williams95fc17a2009-07-31 12:39:15 +10005741 kfree(conf->disks);
5742 kfree(conf->stripe_hashtbl);
5743 kfree(conf);
5744}
5745
Dan Williams36d1c642009-07-14 11:48:22 -07005746#ifdef CONFIG_HOTPLUG_CPU
5747static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5748 void *hcpu)
5749{
NeilBrownd1688a62011-10-11 16:49:52 +11005750 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
Dan Williams36d1c642009-07-14 11:48:22 -07005751 long cpu = (long)hcpu;
5752 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5753
5754 switch (action) {
5755 case CPU_UP_PREPARE:
5756 case CPU_UP_PREPARE_FROZEN:
Oleg Nesterov789b5e02014-02-06 03:42:45 +05305757 if (alloc_scratch_buffer(conf, percpu)) {
Dan Williams36d1c642009-07-14 11:48:22 -07005758 pr_err("%s: failed memory allocation for cpu%ld\n",
5759 __func__, cpu);
Akinobu Mita55af6bb2010-05-26 14:43:35 -07005760 return notifier_from_errno(-ENOMEM);
Dan Williams36d1c642009-07-14 11:48:22 -07005761 }
5762 break;
5763 case CPU_DEAD:
5764 case CPU_DEAD_FROZEN:
Oleg Nesterov789b5e02014-02-06 03:42:45 +05305765 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
Dan Williams36d1c642009-07-14 11:48:22 -07005766 break;
5767 default:
5768 break;
5769 }
5770 return NOTIFY_OK;
5771}
5772#endif
5773
NeilBrownd1688a62011-10-11 16:49:52 +11005774static int raid5_alloc_percpu(struct r5conf *conf)
Dan Williams36d1c642009-07-14 11:48:22 -07005775{
5776 unsigned long cpu;
Oleg Nesterov789b5e02014-02-06 03:42:45 +05305777 int err = 0;
Dan Williams36d1c642009-07-14 11:48:22 -07005778
Oleg Nesterov789b5e02014-02-06 03:42:45 +05305779 conf->percpu = alloc_percpu(struct raid5_percpu);
5780 if (!conf->percpu)
Dan Williams36d1c642009-07-14 11:48:22 -07005781 return -ENOMEM;
Dan Williams36d1c642009-07-14 11:48:22 -07005782
Dan Williams36d1c642009-07-14 11:48:22 -07005783#ifdef CONFIG_HOTPLUG_CPU
5784 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5785 conf->cpu_notify.priority = 0;
Oleg Nesterov789b5e02014-02-06 03:42:45 +05305786 err = register_cpu_notifier(&conf->cpu_notify);
5787 if (err)
5788 return err;
Dan Williams36d1c642009-07-14 11:48:22 -07005789#endif
Oleg Nesterov789b5e02014-02-06 03:42:45 +05305790
5791 get_online_cpus();
5792 for_each_present_cpu(cpu) {
5793 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5794 if (err) {
5795 pr_err("%s: failed memory allocation for cpu%ld\n",
5796 __func__, cpu);
5797 break;
5798 }
5799 }
Dan Williams36d1c642009-07-14 11:48:22 -07005800 put_online_cpus();
5801
5802 return err;
5803}
5804
NeilBrownd1688a62011-10-11 16:49:52 +11005805static struct r5conf *setup_conf(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005806{
NeilBrownd1688a62011-10-11 16:49:52 +11005807 struct r5conf *conf;
NeilBrown5e5e3e72009-10-16 16:35:30 +11005808 int raid_disk, memory, max_disks;
NeilBrown3cb03002011-10-11 16:45:26 +11005809 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005810 struct disk_info *disk;
NeilBrown02326052012-07-03 15:56:52 +10005811 char pers_name[6];
Shaohua Li566c09c2013-11-14 15:16:17 +11005812 int i;
majianpeng60aaf932013-11-14 15:16:20 +11005813 int group_cnt, worker_cnt_per_group;
5814 struct r5worker_group *new_group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005815
NeilBrown91adb562009-03-31 14:39:39 +11005816 if (mddev->new_level != 5
5817 && mddev->new_level != 4
5818 && mddev->new_level != 6) {
NeilBrown0c55e022010-05-03 14:09:02 +10005819 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
NeilBrown91adb562009-03-31 14:39:39 +11005820 mdname(mddev), mddev->new_level);
5821 return ERR_PTR(-EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005822 }
NeilBrown91adb562009-03-31 14:39:39 +11005823 if ((mddev->new_level == 5
5824 && !algorithm_valid_raid5(mddev->new_layout)) ||
5825 (mddev->new_level == 6
5826 && !algorithm_valid_raid6(mddev->new_layout))) {
NeilBrown0c55e022010-05-03 14:09:02 +10005827 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
NeilBrown91adb562009-03-31 14:39:39 +11005828 mdname(mddev), mddev->new_layout);
5829 return ERR_PTR(-EIO);
5830 }
5831 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
NeilBrown0c55e022010-05-03 14:09:02 +10005832 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
NeilBrown91adb562009-03-31 14:39:39 +11005833 mdname(mddev), mddev->raid_disks);
5834 return ERR_PTR(-EINVAL);
NeilBrown99c0fb52009-03-31 14:39:38 +11005835 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005836
Andre Noll664e7c42009-06-18 08:45:27 +10005837 if (!mddev->new_chunk_sectors ||
5838 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5839 !is_power_of_2(mddev->new_chunk_sectors)) {
NeilBrown0c55e022010-05-03 14:09:02 +10005840 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5841 mdname(mddev), mddev->new_chunk_sectors << 9);
NeilBrown91adb562009-03-31 14:39:39 +11005842 return ERR_PTR(-EINVAL);
NeilBrown4bbf3772008-10-13 11:55:12 +11005843 }
5844
NeilBrownd1688a62011-10-11 16:49:52 +11005845 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
NeilBrown91adb562009-03-31 14:39:39 +11005846 if (conf == NULL)
5847 goto abort;
Shaohua Li851c30c2013-08-28 14:30:16 +08005848 /* Don't enable multi-threading by default*/
majianpeng60aaf932013-11-14 15:16:20 +11005849 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
5850 &new_group)) {
5851 conf->group_cnt = group_cnt;
5852 conf->worker_cnt_per_group = worker_cnt_per_group;
5853 conf->worker_groups = new_group;
5854 } else
Shaohua Li851c30c2013-08-28 14:30:16 +08005855 goto abort;
Dan Williamsf5efd452009-10-16 15:55:38 +11005856 spin_lock_init(&conf->device_lock);
NeilBrownc46501b2013-08-27 15:52:13 +10005857 seqcount_init(&conf->gen_lock);
Dan Williamsf5efd452009-10-16 15:55:38 +11005858 init_waitqueue_head(&conf->wait_for_stripe);
5859 init_waitqueue_head(&conf->wait_for_overlap);
5860 INIT_LIST_HEAD(&conf->handle_list);
5861 INIT_LIST_HEAD(&conf->hold_list);
5862 INIT_LIST_HEAD(&conf->delayed_list);
5863 INIT_LIST_HEAD(&conf->bitmap_list);
Shaohua Li773ca822013-08-27 17:50:39 +08005864 init_llist_head(&conf->released_stripes);
Dan Williamsf5efd452009-10-16 15:55:38 +11005865 atomic_set(&conf->active_stripes, 0);
5866 atomic_set(&conf->preread_active_stripes, 0);
5867 atomic_set(&conf->active_aligned_reads, 0);
5868 conf->bypass_threshold = BYPASS_THRESHOLD;
NeilBrownd890fa22011-10-26 11:54:39 +11005869 conf->recovery_disabled = mddev->recovery_disabled - 1;
NeilBrown91adb562009-03-31 14:39:39 +11005870
5871 conf->raid_disks = mddev->raid_disks;
5872 if (mddev->reshape_position == MaxSector)
5873 conf->previous_raid_disks = mddev->raid_disks;
5874 else
5875 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
NeilBrown5e5e3e72009-10-16 16:35:30 +11005876 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5877 conf->scribble_len = scribble_len(max_disks);
NeilBrown91adb562009-03-31 14:39:39 +11005878
NeilBrown5e5e3e72009-10-16 16:35:30 +11005879 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
NeilBrown91adb562009-03-31 14:39:39 +11005880 GFP_KERNEL);
5881 if (!conf->disks)
5882 goto abort;
5883
5884 conf->mddev = mddev;
5885
5886 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5887 goto abort;
5888
Shaohua Li566c09c2013-11-14 15:16:17 +11005889 /* We init hash_locks[0] separately to that it can be used
5890 * as the reference lock in the spin_lock_nest_lock() call
5891 * in lock_all_device_hash_locks_irq in order to convince
5892 * lockdep that we know what we are doing.
5893 */
5894 spin_lock_init(conf->hash_locks);
5895 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
5896 spin_lock_init(conf->hash_locks + i);
5897
5898 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5899 INIT_LIST_HEAD(conf->inactive_list + i);
5900
5901 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5902 INIT_LIST_HEAD(conf->temp_inactive_list + i);
5903
Dan Williams36d1c642009-07-14 11:48:22 -07005904 conf->level = mddev->new_level;
5905 if (raid5_alloc_percpu(conf) != 0)
5906 goto abort;
5907
NeilBrown0c55e022010-05-03 14:09:02 +10005908 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
NeilBrown91adb562009-03-31 14:39:39 +11005909
NeilBrowndafb20f2012-03-19 12:46:39 +11005910 rdev_for_each(rdev, mddev) {
NeilBrown91adb562009-03-31 14:39:39 +11005911 raid_disk = rdev->raid_disk;
NeilBrown5e5e3e72009-10-16 16:35:30 +11005912 if (raid_disk >= max_disks
NeilBrown91adb562009-03-31 14:39:39 +11005913 || raid_disk < 0)
5914 continue;
5915 disk = conf->disks + raid_disk;
5916
NeilBrown17045f52011-12-23 10:17:53 +11005917 if (test_bit(Replacement, &rdev->flags)) {
5918 if (disk->replacement)
5919 goto abort;
5920 disk->replacement = rdev;
5921 } else {
5922 if (disk->rdev)
5923 goto abort;
5924 disk->rdev = rdev;
5925 }
NeilBrown91adb562009-03-31 14:39:39 +11005926
5927 if (test_bit(In_sync, &rdev->flags)) {
5928 char b[BDEVNAME_SIZE];
NeilBrown0c55e022010-05-03 14:09:02 +10005929 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5930 " disk %d\n",
5931 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
Jonathan Brassowd6b212f2011-06-08 18:00:28 -05005932 } else if (rdev->saved_raid_disk != raid_disk)
NeilBrown91adb562009-03-31 14:39:39 +11005933 /* Cannot rely on bitmap to complete recovery */
5934 conf->fullsync = 1;
5935 }
5936
Andre Noll09c9e5f2009-06-18 08:45:55 +10005937 conf->chunk_sectors = mddev->new_chunk_sectors;
NeilBrown91adb562009-03-31 14:39:39 +11005938 conf->level = mddev->new_level;
5939 if (conf->level == 6)
5940 conf->max_degraded = 2;
5941 else
5942 conf->max_degraded = 1;
5943 conf->algorithm = mddev->new_layout;
NeilBrownfef9c612009-03-31 15:16:46 +11005944 conf->reshape_progress = mddev->reshape_position;
NeilBrowne183eae2009-03-31 15:20:22 +11005945 if (conf->reshape_progress != MaxSector) {
Andre Noll09c9e5f2009-06-18 08:45:55 +10005946 conf->prev_chunk_sectors = mddev->chunk_sectors;
NeilBrowne183eae2009-03-31 15:20:22 +11005947 conf->prev_algo = mddev->layout;
5948 }
NeilBrown91adb562009-03-31 14:39:39 +11005949
5950 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
NeilBrown5e5e3e72009-10-16 16:35:30 +11005951 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
Shaohua Li4bda5562013-11-14 15:16:17 +11005952 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
Shaohua Li566c09c2013-11-14 15:16:17 +11005953 if (grow_stripes(conf, NR_STRIPES)) {
NeilBrown91adb562009-03-31 14:39:39 +11005954 printk(KERN_ERR
NeilBrown0c55e022010-05-03 14:09:02 +10005955 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5956 mdname(mddev), memory);
NeilBrown91adb562009-03-31 14:39:39 +11005957 goto abort;
5958 } else
NeilBrown0c55e022010-05-03 14:09:02 +10005959 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5960 mdname(mddev), memory);
NeilBrown91adb562009-03-31 14:39:39 +11005961
NeilBrown02326052012-07-03 15:56:52 +10005962 sprintf(pers_name, "raid%d", mddev->new_level);
5963 conf->thread = md_register_thread(raid5d, mddev, pers_name);
NeilBrown91adb562009-03-31 14:39:39 +11005964 if (!conf->thread) {
5965 printk(KERN_ERR
NeilBrown0c55e022010-05-03 14:09:02 +10005966 "md/raid:%s: couldn't allocate thread.\n",
NeilBrown91adb562009-03-31 14:39:39 +11005967 mdname(mddev));
5968 goto abort;
5969 }
5970
5971 return conf;
5972
5973 abort:
5974 if (conf) {
Dan Williams95fc17a2009-07-31 12:39:15 +10005975 free_conf(conf);
NeilBrown91adb562009-03-31 14:39:39 +11005976 return ERR_PTR(-EIO);
5977 } else
5978 return ERR_PTR(-ENOMEM);
5979}
5980
NeilBrownc148ffd2009-11-13 17:47:00 +11005981static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5982{
5983 switch (algo) {
5984 case ALGORITHM_PARITY_0:
5985 if (raid_disk < max_degraded)
5986 return 1;
5987 break;
5988 case ALGORITHM_PARITY_N:
5989 if (raid_disk >= raid_disks - max_degraded)
5990 return 1;
5991 break;
5992 case ALGORITHM_PARITY_0_6:
NeilBrownf72ffdd2014-09-30 14:23:59 +10005993 if (raid_disk == 0 ||
NeilBrownc148ffd2009-11-13 17:47:00 +11005994 raid_disk == raid_disks - 1)
5995 return 1;
5996 break;
5997 case ALGORITHM_LEFT_ASYMMETRIC_6:
5998 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5999 case ALGORITHM_LEFT_SYMMETRIC_6:
6000 case ALGORITHM_RIGHT_SYMMETRIC_6:
6001 if (raid_disk == raid_disks - 1)
6002 return 1;
6003 }
6004 return 0;
6005}
6006
NeilBrownfd01b882011-10-11 16:47:53 +11006007static int run(struct mddev *mddev)
NeilBrown91adb562009-03-31 14:39:39 +11006008{
NeilBrownd1688a62011-10-11 16:49:52 +11006009 struct r5conf *conf;
NeilBrown9f7c2222010-07-26 12:04:13 +10006010 int working_disks = 0;
NeilBrownc148ffd2009-11-13 17:47:00 +11006011 int dirty_parity_disks = 0;
NeilBrown3cb03002011-10-11 16:45:26 +11006012 struct md_rdev *rdev;
NeilBrownc148ffd2009-11-13 17:47:00 +11006013 sector_t reshape_offset = 0;
NeilBrown17045f52011-12-23 10:17:53 +11006014 int i;
NeilBrownb5254dd2012-05-21 09:27:01 +10006015 long long min_offset_diff = 0;
6016 int first = 1;
NeilBrown91adb562009-03-31 14:39:39 +11006017
Andre Noll8c6ac862009-06-18 08:48:06 +10006018 if (mddev->recovery_cp != MaxSector)
NeilBrown0c55e022010-05-03 14:09:02 +10006019 printk(KERN_NOTICE "md/raid:%s: not clean"
Andre Noll8c6ac862009-06-18 08:48:06 +10006020 " -- starting background reconstruction\n",
6021 mdname(mddev));
NeilBrownb5254dd2012-05-21 09:27:01 +10006022
6023 rdev_for_each(rdev, mddev) {
6024 long long diff;
6025 if (rdev->raid_disk < 0)
6026 continue;
6027 diff = (rdev->new_data_offset - rdev->data_offset);
6028 if (first) {
6029 min_offset_diff = diff;
6030 first = 0;
6031 } else if (mddev->reshape_backwards &&
6032 diff < min_offset_diff)
6033 min_offset_diff = diff;
6034 else if (!mddev->reshape_backwards &&
6035 diff > min_offset_diff)
6036 min_offset_diff = diff;
6037 }
6038
NeilBrownf6705572006-03-27 01:18:11 -08006039 if (mddev->reshape_position != MaxSector) {
6040 /* Check that we can continue the reshape.
NeilBrownb5254dd2012-05-21 09:27:01 +10006041 * Difficulties arise if the stripe we would write to
6042 * next is at or after the stripe we would read from next.
6043 * For a reshape that changes the number of devices, this
6044 * is only possible for a very short time, and mdadm makes
6045 * sure that time appears to have past before assembling
6046 * the array. So we fail if that time hasn't passed.
6047 * For a reshape that keeps the number of devices the same
6048 * mdadm must be monitoring the reshape can keeping the
6049 * critical areas read-only and backed up. It will start
6050 * the array in read-only mode, so we check for that.
NeilBrownf6705572006-03-27 01:18:11 -08006051 */
6052 sector_t here_new, here_old;
6053 int old_disks;
Andre Noll18b00332009-03-31 15:00:56 +11006054 int max_degraded = (mddev->level == 6 ? 2 : 1);
NeilBrownf6705572006-03-27 01:18:11 -08006055
NeilBrown88ce4932009-03-31 15:24:23 +11006056 if (mddev->new_level != mddev->level) {
NeilBrown0c55e022010-05-03 14:09:02 +10006057 printk(KERN_ERR "md/raid:%s: unsupported reshape "
NeilBrownf4168852007-02-28 20:11:53 -08006058 "required - aborting.\n",
NeilBrownf6705572006-03-27 01:18:11 -08006059 mdname(mddev));
6060 return -EINVAL;
6061 }
NeilBrownf6705572006-03-27 01:18:11 -08006062 old_disks = mddev->raid_disks - mddev->delta_disks;
6063 /* reshape_position must be on a new-stripe boundary, and one
NeilBrownf4168852007-02-28 20:11:53 -08006064 * further up in new geometry must map after here in old
6065 * geometry.
NeilBrownf6705572006-03-27 01:18:11 -08006066 */
6067 here_new = mddev->reshape_position;
Andre Noll664e7c42009-06-18 08:45:27 +10006068 if (sector_div(here_new, mddev->new_chunk_sectors *
NeilBrownf4168852007-02-28 20:11:53 -08006069 (mddev->raid_disks - max_degraded))) {
NeilBrown0c55e022010-05-03 14:09:02 +10006070 printk(KERN_ERR "md/raid:%s: reshape_position not "
6071 "on a stripe boundary\n", mdname(mddev));
NeilBrownf6705572006-03-27 01:18:11 -08006072 return -EINVAL;
6073 }
NeilBrownc148ffd2009-11-13 17:47:00 +11006074 reshape_offset = here_new * mddev->new_chunk_sectors;
NeilBrownf6705572006-03-27 01:18:11 -08006075 /* here_new is the stripe we will write to */
6076 here_old = mddev->reshape_position;
Andre Noll9d8f0362009-06-18 08:45:01 +10006077 sector_div(here_old, mddev->chunk_sectors *
NeilBrownf4168852007-02-28 20:11:53 -08006078 (old_disks-max_degraded));
6079 /* here_old is the first stripe that we might need to read
6080 * from */
NeilBrown67ac6012009-08-13 10:06:24 +10006081 if (mddev->delta_disks == 0) {
NeilBrownb5254dd2012-05-21 09:27:01 +10006082 if ((here_new * mddev->new_chunk_sectors !=
6083 here_old * mddev->chunk_sectors)) {
6084 printk(KERN_ERR "md/raid:%s: reshape position is"
6085 " confused - aborting\n", mdname(mddev));
6086 return -EINVAL;
6087 }
NeilBrown67ac6012009-08-13 10:06:24 +10006088 /* We cannot be sure it is safe to start an in-place
NeilBrownb5254dd2012-05-21 09:27:01 +10006089 * reshape. It is only safe if user-space is monitoring
NeilBrown67ac6012009-08-13 10:06:24 +10006090 * and taking constant backups.
6091 * mdadm always starts a situation like this in
6092 * readonly mode so it can take control before
6093 * allowing any writes. So just check for that.
6094 */
NeilBrownb5254dd2012-05-21 09:27:01 +10006095 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6096 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6097 /* not really in-place - so OK */;
6098 else if (mddev->ro == 0) {
6099 printk(KERN_ERR "md/raid:%s: in-place reshape "
6100 "must be started in read-only mode "
6101 "- aborting\n",
NeilBrown0c55e022010-05-03 14:09:02 +10006102 mdname(mddev));
NeilBrown67ac6012009-08-13 10:06:24 +10006103 return -EINVAL;
6104 }
NeilBrown2c810cd2012-05-21 09:27:00 +10006105 } else if (mddev->reshape_backwards
NeilBrownb5254dd2012-05-21 09:27:01 +10006106 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
NeilBrown67ac6012009-08-13 10:06:24 +10006107 here_old * mddev->chunk_sectors)
6108 : (here_new * mddev->new_chunk_sectors >=
NeilBrownb5254dd2012-05-21 09:27:01 +10006109 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
NeilBrownf6705572006-03-27 01:18:11 -08006110 /* Reading from the same stripe as writing to - bad */
NeilBrown0c55e022010-05-03 14:09:02 +10006111 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6112 "auto-recovery - aborting.\n",
6113 mdname(mddev));
NeilBrownf6705572006-03-27 01:18:11 -08006114 return -EINVAL;
6115 }
NeilBrown0c55e022010-05-03 14:09:02 +10006116 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6117 mdname(mddev));
NeilBrownf6705572006-03-27 01:18:11 -08006118 /* OK, we should be able to continue; */
NeilBrownf6705572006-03-27 01:18:11 -08006119 } else {
NeilBrown91adb562009-03-31 14:39:39 +11006120 BUG_ON(mddev->level != mddev->new_level);
6121 BUG_ON(mddev->layout != mddev->new_layout);
Andre Noll664e7c42009-06-18 08:45:27 +10006122 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
NeilBrown91adb562009-03-31 14:39:39 +11006123 BUG_ON(mddev->delta_disks != 0);
NeilBrownf6705572006-03-27 01:18:11 -08006124 }
6125
NeilBrown245f46c2009-03-31 14:39:39 +11006126 if (mddev->private == NULL)
6127 conf = setup_conf(mddev);
6128 else
6129 conf = mddev->private;
6130
NeilBrown91adb562009-03-31 14:39:39 +11006131 if (IS_ERR(conf))
6132 return PTR_ERR(conf);
NeilBrown9ffae0c2006-01-06 00:20:32 -08006133
NeilBrownb5254dd2012-05-21 09:27:01 +10006134 conf->min_offset_diff = min_offset_diff;
NeilBrown91adb562009-03-31 14:39:39 +11006135 mddev->thread = conf->thread;
6136 conf->thread = NULL;
6137 mddev->private = conf;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006138
NeilBrown17045f52011-12-23 10:17:53 +11006139 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6140 i++) {
6141 rdev = conf->disks[i].rdev;
6142 if (!rdev && conf->disks[i].replacement) {
6143 /* The replacement is all we have yet */
6144 rdev = conf->disks[i].replacement;
6145 conf->disks[i].replacement = NULL;
6146 clear_bit(Replacement, &rdev->flags);
6147 conf->disks[i].rdev = rdev;
6148 }
6149 if (!rdev)
NeilBrownc148ffd2009-11-13 17:47:00 +11006150 continue;
NeilBrown17045f52011-12-23 10:17:53 +11006151 if (conf->disks[i].replacement &&
6152 conf->reshape_progress != MaxSector) {
6153 /* replacements and reshape simply do not mix. */
6154 printk(KERN_ERR "md: cannot handle concurrent "
6155 "replacement and reshape.\n");
6156 goto abort;
6157 }
NeilBrown2f115882010-06-17 17:41:03 +10006158 if (test_bit(In_sync, &rdev->flags)) {
NeilBrown91adb562009-03-31 14:39:39 +11006159 working_disks++;
NeilBrown2f115882010-06-17 17:41:03 +10006160 continue;
6161 }
NeilBrownc148ffd2009-11-13 17:47:00 +11006162 /* This disc is not fully in-sync. However if it
6163 * just stored parity (beyond the recovery_offset),
6164 * when we don't need to be concerned about the
6165 * array being dirty.
6166 * When reshape goes 'backwards', we never have
6167 * partially completed devices, so we only need
6168 * to worry about reshape going forwards.
6169 */
6170 /* Hack because v0.91 doesn't store recovery_offset properly. */
6171 if (mddev->major_version == 0 &&
6172 mddev->minor_version > 90)
6173 rdev->recovery_offset = reshape_offset;
H. Peter Anvin5026d7a2013-06-12 07:37:43 -07006174
NeilBrownc148ffd2009-11-13 17:47:00 +11006175 if (rdev->recovery_offset < reshape_offset) {
6176 /* We need to check old and new layout */
6177 if (!only_parity(rdev->raid_disk,
6178 conf->algorithm,
6179 conf->raid_disks,
6180 conf->max_degraded))
6181 continue;
6182 }
6183 if (!only_parity(rdev->raid_disk,
6184 conf->prev_algo,
6185 conf->previous_raid_disks,
6186 conf->max_degraded))
6187 continue;
6188 dirty_parity_disks++;
6189 }
NeilBrown91adb562009-03-31 14:39:39 +11006190
NeilBrown17045f52011-12-23 10:17:53 +11006191 /*
6192 * 0 for a fully functional array, 1 or 2 for a degraded array.
6193 */
NeilBrown908f4fb2011-12-23 10:17:50 +11006194 mddev->degraded = calc_degraded(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006195
NeilBrown674806d2010-06-16 17:17:53 +10006196 if (has_failed(conf)) {
NeilBrown0c55e022010-05-03 14:09:02 +10006197 printk(KERN_ERR "md/raid:%s: not enough operational devices"
Linus Torvalds1da177e2005-04-16 15:20:36 -07006198 " (%d/%d failed)\n",
NeilBrown02c2de82006-10-03 01:15:47 -07006199 mdname(mddev), mddev->degraded, conf->raid_disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006200 goto abort;
6201 }
6202
NeilBrown91adb562009-03-31 14:39:39 +11006203 /* device size must be a multiple of chunk size */
Andre Noll9d8f0362009-06-18 08:45:01 +10006204 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
NeilBrown91adb562009-03-31 14:39:39 +11006205 mddev->resync_max_sectors = mddev->dev_sectors;
6206
NeilBrownc148ffd2009-11-13 17:47:00 +11006207 if (mddev->degraded > dirty_parity_disks &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07006208 mddev->recovery_cp != MaxSector) {
NeilBrown6ff8d8ec2006-01-06 00:20:15 -08006209 if (mddev->ok_start_degraded)
6210 printk(KERN_WARNING
NeilBrown0c55e022010-05-03 14:09:02 +10006211 "md/raid:%s: starting dirty degraded array"
6212 " - data corruption possible.\n",
NeilBrown6ff8d8ec2006-01-06 00:20:15 -08006213 mdname(mddev));
6214 else {
6215 printk(KERN_ERR
NeilBrown0c55e022010-05-03 14:09:02 +10006216 "md/raid:%s: cannot start dirty degraded array.\n",
NeilBrown6ff8d8ec2006-01-06 00:20:15 -08006217 mdname(mddev));
6218 goto abort;
6219 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006220 }
6221
Linus Torvalds1da177e2005-04-16 15:20:36 -07006222 if (mddev->degraded == 0)
NeilBrown0c55e022010-05-03 14:09:02 +10006223 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6224 " devices, algorithm %d\n", mdname(mddev), conf->level,
NeilBrowne183eae2009-03-31 15:20:22 +11006225 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6226 mddev->new_layout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006227 else
NeilBrown0c55e022010-05-03 14:09:02 +10006228 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6229 " out of %d devices, algorithm %d\n",
6230 mdname(mddev), conf->level,
6231 mddev->raid_disks - mddev->degraded,
6232 mddev->raid_disks, mddev->new_layout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006233
6234 print_raid5_conf(conf);
6235
NeilBrownfef9c612009-03-31 15:16:46 +11006236 if (conf->reshape_progress != MaxSector) {
NeilBrownfef9c612009-03-31 15:16:46 +11006237 conf->reshape_safe = conf->reshape_progress;
NeilBrownf6705572006-03-27 01:18:11 -08006238 atomic_set(&conf->reshape_stripes, 0);
6239 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6240 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6241 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6242 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6243 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
NeilBrown0da3c612009-09-23 18:09:45 +10006244 "reshape");
NeilBrownf6705572006-03-27 01:18:11 -08006245 }
6246
Linus Torvalds1da177e2005-04-16 15:20:36 -07006247 /* Ok, everything is just fine now */
NeilBrowna64c8762010-04-14 17:15:37 +10006248 if (mddev->to_remove == &raid5_attrs_group)
6249 mddev->to_remove = NULL;
NeilBrown00bcb4a2010-06-01 19:37:23 +10006250 else if (mddev->kobj.sd &&
6251 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
NeilBrown5e55e2f2007-03-26 21:32:14 -08006252 printk(KERN_WARNING
NeilBrown4a5add42010-06-01 19:37:28 +10006253 "raid5: failed to create sysfs attributes for %s\n",
NeilBrown5e55e2f2007-03-26 21:32:14 -08006254 mdname(mddev));
NeilBrown4a5add42010-06-01 19:37:28 +10006255 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6256
6257 if (mddev->queue) {
NeilBrown9f7c2222010-07-26 12:04:13 +10006258 int chunk_size;
Shaohua Li620125f2012-10-11 13:49:05 +11006259 bool discard_supported = true;
NeilBrown4a5add42010-06-01 19:37:28 +10006260 /* read-ahead size must cover two whole stripes, which
6261 * is 2 * (datadisks) * chunksize where 'n' is the
6262 * number of raid devices
6263 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006264 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6265 int stripe = data_disks *
6266 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6267 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6268 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
NeilBrown4a5add42010-06-01 19:37:28 +10006269
NeilBrown9f7c2222010-07-26 12:04:13 +10006270 chunk_size = mddev->chunk_sectors << 9;
6271 blk_queue_io_min(mddev->queue, chunk_size);
6272 blk_queue_io_opt(mddev->queue, chunk_size *
6273 (conf->raid_disks - conf->max_degraded));
Kent Overstreetc78afc62013-07-11 22:39:53 -07006274 mddev->queue->limits.raid_partial_stripes_expensive = 1;
Shaohua Li620125f2012-10-11 13:49:05 +11006275 /*
6276 * We can only discard a whole stripe. It doesn't make sense to
6277 * discard data disk but write parity disk
6278 */
6279 stripe = stripe * PAGE_SIZE;
NeilBrown4ac68752012-11-19 13:11:26 +11006280 /* Round up to power of 2, as discard handling
6281 * currently assumes that */
6282 while ((stripe-1) & stripe)
6283 stripe = (stripe | (stripe-1)) + 1;
Shaohua Li620125f2012-10-11 13:49:05 +11006284 mddev->queue->limits.discard_alignment = stripe;
6285 mddev->queue->limits.discard_granularity = stripe;
6286 /*
6287 * unaligned part of discard request will be ignored, so can't
NeilBrown8e0e99b2014-10-02 13:45:00 +10006288 * guarantee discard_zeroes_data
Shaohua Li620125f2012-10-11 13:49:05 +11006289 */
6290 mddev->queue->limits.discard_zeroes_data = 0;
NeilBrown9f7c2222010-07-26 12:04:13 +10006291
H. Peter Anvin5026d7a2013-06-12 07:37:43 -07006292 blk_queue_max_write_same_sectors(mddev->queue, 0);
6293
NeilBrown05616be2012-05-21 09:27:00 +10006294 rdev_for_each(rdev, mddev) {
NeilBrown9f7c2222010-07-26 12:04:13 +10006295 disk_stack_limits(mddev->gendisk, rdev->bdev,
6296 rdev->data_offset << 9);
NeilBrown05616be2012-05-21 09:27:00 +10006297 disk_stack_limits(mddev->gendisk, rdev->bdev,
6298 rdev->new_data_offset << 9);
Shaohua Li620125f2012-10-11 13:49:05 +11006299 /*
6300 * discard_zeroes_data is required, otherwise data
6301 * could be lost. Consider a scenario: discard a stripe
6302 * (the stripe could be inconsistent if
6303 * discard_zeroes_data is 0); write one disk of the
6304 * stripe (the stripe could be inconsistent again
6305 * depending on which disks are used to calculate
6306 * parity); the disk is broken; The stripe data of this
6307 * disk is lost.
6308 */
6309 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6310 !bdev_get_queue(rdev->bdev)->
6311 limits.discard_zeroes_data)
6312 discard_supported = false;
NeilBrown8e0e99b2014-10-02 13:45:00 +10006313 /* Unfortunately, discard_zeroes_data is not currently
6314 * a guarantee - just a hint. So we only allow DISCARD
6315 * if the sysadmin has confirmed that only safe devices
6316 * are in use by setting a module parameter.
6317 */
6318 if (!devices_handle_discard_safely) {
6319 if (discard_supported) {
6320 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6321 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6322 }
6323 discard_supported = false;
6324 }
NeilBrown05616be2012-05-21 09:27:00 +10006325 }
Shaohua Li620125f2012-10-11 13:49:05 +11006326
6327 if (discard_supported &&
6328 mddev->queue->limits.max_discard_sectors >= stripe &&
6329 mddev->queue->limits.discard_granularity >= stripe)
6330 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6331 mddev->queue);
6332 else
6333 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6334 mddev->queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006335 }
6336
Linus Torvalds1da177e2005-04-16 15:20:36 -07006337 return 0;
6338abort:
NeilBrown01f96c02011-09-21 15:30:20 +10006339 md_unregister_thread(&mddev->thread);
NeilBrowne4f869d2011-10-07 14:22:49 +11006340 print_raid5_conf(conf);
6341 free_conf(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006342 mddev->private = NULL;
NeilBrown0c55e022010-05-03 14:09:02 +10006343 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006344 return -EIO;
6345}
6346
NeilBrownafa0f552014-12-15 12:56:58 +11006347static void raid5_free(struct mddev *mddev, void *priv)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006348{
NeilBrownafa0f552014-12-15 12:56:58 +11006349 struct r5conf *conf = priv;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006350
Dan Williams95fc17a2009-07-31 12:39:15 +10006351 free_conf(conf);
NeilBrowna64c8762010-04-14 17:15:37 +10006352 mddev->to_remove = &raid5_attrs_group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006353}
6354
NeilBrownfd01b882011-10-11 16:47:53 +11006355static void status(struct seq_file *seq, struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006356{
NeilBrownd1688a62011-10-11 16:49:52 +11006357 struct r5conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006358 int i;
6359
Andre Noll9d8f0362009-06-18 08:45:01 +10006360 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6361 mddev->chunk_sectors / 2, mddev->layout);
NeilBrown02c2de82006-10-03 01:15:47 -07006362 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006363 for (i = 0; i < conf->raid_disks; i++)
6364 seq_printf (seq, "%s",
6365 conf->disks[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08006366 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006367 seq_printf (seq, "]");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006368}
6369
NeilBrownd1688a62011-10-11 16:49:52 +11006370static void print_raid5_conf (struct r5conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006371{
6372 int i;
6373 struct disk_info *tmp;
6374
NeilBrown0c55e022010-05-03 14:09:02 +10006375 printk(KERN_DEBUG "RAID conf printout:\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006376 if (!conf) {
6377 printk("(conf==NULL)\n");
6378 return;
6379 }
NeilBrown0c55e022010-05-03 14:09:02 +10006380 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6381 conf->raid_disks,
6382 conf->raid_disks - conf->mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006383
6384 for (i = 0; i < conf->raid_disks; i++) {
6385 char b[BDEVNAME_SIZE];
6386 tmp = conf->disks + i;
6387 if (tmp->rdev)
NeilBrown0c55e022010-05-03 14:09:02 +10006388 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6389 i, !test_bit(Faulty, &tmp->rdev->flags),
6390 bdevname(tmp->rdev->bdev, b));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006391 }
6392}
6393
NeilBrownfd01b882011-10-11 16:47:53 +11006394static int raid5_spare_active(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006395{
6396 int i;
NeilBrownd1688a62011-10-11 16:49:52 +11006397 struct r5conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006398 struct disk_info *tmp;
NeilBrown6b965622010-08-18 11:56:59 +10006399 int count = 0;
6400 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006401
6402 for (i = 0; i < conf->raid_disks; i++) {
6403 tmp = conf->disks + i;
NeilBrowndd054fc2011-12-23 10:17:53 +11006404 if (tmp->replacement
6405 && tmp->replacement->recovery_offset == MaxSector
6406 && !test_bit(Faulty, &tmp->replacement->flags)
6407 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
6408 /* Replacement has just become active. */
6409 if (!tmp->rdev
6410 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
6411 count++;
6412 if (tmp->rdev) {
6413 /* Replaced device not technically faulty,
6414 * but we need to be sure it gets removed
6415 * and never re-added.
6416 */
6417 set_bit(Faulty, &tmp->rdev->flags);
6418 sysfs_notify_dirent_safe(
6419 tmp->rdev->sysfs_state);
6420 }
6421 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
6422 } else if (tmp->rdev
NeilBrown70fffd02010-06-16 17:01:25 +10006423 && tmp->rdev->recovery_offset == MaxSector
NeilBrownb2d444d2005-11-08 21:39:31 -08006424 && !test_bit(Faulty, &tmp->rdev->flags)
NeilBrownc04be0a2006-10-03 01:15:53 -07006425 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
NeilBrown6b965622010-08-18 11:56:59 +10006426 count++;
Jonathan Brassow43c73ca2011-01-14 09:14:33 +11006427 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006428 }
6429 }
NeilBrown6b965622010-08-18 11:56:59 +10006430 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrown908f4fb2011-12-23 10:17:50 +11006431 mddev->degraded = calc_degraded(conf);
NeilBrown6b965622010-08-18 11:56:59 +10006432 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006433 print_raid5_conf(conf);
NeilBrown6b965622010-08-18 11:56:59 +10006434 return count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006435}
6436
NeilBrownb8321b62011-12-23 10:17:51 +11006437static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006438{
NeilBrownd1688a62011-10-11 16:49:52 +11006439 struct r5conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006440 int err = 0;
NeilBrownb8321b62011-12-23 10:17:51 +11006441 int number = rdev->raid_disk;
NeilBrown657e3e42011-12-23 10:17:52 +11006442 struct md_rdev **rdevp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006443 struct disk_info *p = conf->disks + number;
6444
6445 print_raid5_conf(conf);
NeilBrown657e3e42011-12-23 10:17:52 +11006446 if (rdev == p->rdev)
6447 rdevp = &p->rdev;
6448 else if (rdev == p->replacement)
6449 rdevp = &p->replacement;
6450 else
6451 return 0;
NeilBrownec32a2b2009-03-31 15:17:38 +11006452
NeilBrown657e3e42011-12-23 10:17:52 +11006453 if (number >= conf->raid_disks &&
6454 conf->reshape_progress == MaxSector)
6455 clear_bit(In_sync, &rdev->flags);
6456
6457 if (test_bit(In_sync, &rdev->flags) ||
6458 atomic_read(&rdev->nr_pending)) {
6459 err = -EBUSY;
6460 goto abort;
6461 }
6462 /* Only remove non-faulty devices if recovery
6463 * isn't possible.
6464 */
6465 if (!test_bit(Faulty, &rdev->flags) &&
6466 mddev->recovery_disabled != conf->recovery_disabled &&
6467 !has_failed(conf) &&
NeilBrowndd054fc2011-12-23 10:17:53 +11006468 (!p->replacement || p->replacement == rdev) &&
NeilBrown657e3e42011-12-23 10:17:52 +11006469 number < conf->raid_disks) {
6470 err = -EBUSY;
6471 goto abort;
6472 }
6473 *rdevp = NULL;
6474 synchronize_rcu();
6475 if (atomic_read(&rdev->nr_pending)) {
6476 /* lost the race, try later */
6477 err = -EBUSY;
6478 *rdevp = rdev;
NeilBrowndd054fc2011-12-23 10:17:53 +11006479 } else if (p->replacement) {
6480 /* We must have just cleared 'rdev' */
6481 p->rdev = p->replacement;
6482 clear_bit(Replacement, &p->replacement->flags);
6483 smp_mb(); /* Make sure other CPUs may see both as identical
6484 * but will never see neither - if they are careful
6485 */
6486 p->replacement = NULL;
6487 clear_bit(WantReplacement, &rdev->flags);
6488 } else
6489 /* We might have just removed the Replacement as faulty-
6490 * clear the bit just in case
6491 */
6492 clear_bit(WantReplacement, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006493abort:
6494
6495 print_raid5_conf(conf);
6496 return err;
6497}
6498
NeilBrownfd01b882011-10-11 16:47:53 +11006499static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006500{
NeilBrownd1688a62011-10-11 16:49:52 +11006501 struct r5conf *conf = mddev->private;
Neil Brown199050e2008-06-28 08:31:33 +10006502 int err = -EEXIST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006503 int disk;
6504 struct disk_info *p;
Neil Brown6c2fce22008-06-28 08:31:31 +10006505 int first = 0;
6506 int last = conf->raid_disks - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006507
NeilBrown7f0da592011-07-28 11:39:22 +10006508 if (mddev->recovery_disabled == conf->recovery_disabled)
6509 return -EBUSY;
6510
NeilBrowndc10c642012-03-19 12:46:37 +11006511 if (rdev->saved_raid_disk < 0 && has_failed(conf))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006512 /* no point adding a device */
Neil Brown199050e2008-06-28 08:31:33 +10006513 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006514
Neil Brown6c2fce22008-06-28 08:31:31 +10006515 if (rdev->raid_disk >= 0)
6516 first = last = rdev->raid_disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006517
6518 /*
NeilBrown16a53ec2006-06-26 00:27:38 -07006519 * find the disk ... but prefer rdev->saved_raid_disk
6520 * if possible.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006521 */
NeilBrown16a53ec2006-06-26 00:27:38 -07006522 if (rdev->saved_raid_disk >= 0 &&
Neil Brown6c2fce22008-06-28 08:31:31 +10006523 rdev->saved_raid_disk >= first &&
NeilBrown16a53ec2006-06-26 00:27:38 -07006524 conf->disks[rdev->saved_raid_disk].rdev == NULL)
NeilBrown5cfb22a2012-07-03 11:46:53 +10006525 first = rdev->saved_raid_disk;
6526
6527 for (disk = first; disk <= last; disk++) {
NeilBrown7bfec5f2011-12-23 10:17:53 +11006528 p = conf->disks + disk;
6529 if (p->rdev == NULL) {
NeilBrownb2d444d2005-11-08 21:39:31 -08006530 clear_bit(In_sync, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006531 rdev->raid_disk = disk;
Neil Brown199050e2008-06-28 08:31:33 +10006532 err = 0;
NeilBrown72626682005-09-09 16:23:54 -07006533 if (rdev->saved_raid_disk != disk)
6534 conf->fullsync = 1;
Suzanne Woodd6065f72005-11-08 21:39:27 -08006535 rcu_assign_pointer(p->rdev, rdev);
NeilBrown5cfb22a2012-07-03 11:46:53 +10006536 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006537 }
NeilBrown5cfb22a2012-07-03 11:46:53 +10006538 }
6539 for (disk = first; disk <= last; disk++) {
6540 p = conf->disks + disk;
NeilBrown7bfec5f2011-12-23 10:17:53 +11006541 if (test_bit(WantReplacement, &p->rdev->flags) &&
6542 p->replacement == NULL) {
6543 clear_bit(In_sync, &rdev->flags);
6544 set_bit(Replacement, &rdev->flags);
6545 rdev->raid_disk = disk;
6546 err = 0;
6547 conf->fullsync = 1;
6548 rcu_assign_pointer(p->replacement, rdev);
6549 break;
6550 }
6551 }
NeilBrown5cfb22a2012-07-03 11:46:53 +10006552out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006553 print_raid5_conf(conf);
Neil Brown199050e2008-06-28 08:31:33 +10006554 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006555}
6556
NeilBrownfd01b882011-10-11 16:47:53 +11006557static int raid5_resize(struct mddev *mddev, sector_t sectors)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006558{
6559 /* no resync is happening, and there is enough space
6560 * on all devices, so we can resize.
6561 * We need to make sure resync covers any new space.
6562 * If the array is shrinking we should possibly wait until
6563 * any io in the removed space completes, but it hardly seems
6564 * worth it.
6565 */
NeilBrowna4a61252012-05-22 13:55:27 +10006566 sector_t newsize;
Andre Noll9d8f0362009-06-18 08:45:01 +10006567 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
NeilBrowna4a61252012-05-22 13:55:27 +10006568 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
6569 if (mddev->external_size &&
6570 mddev->array_sectors > newsize)
Dan Williamsb522adc2009-03-31 15:00:31 +11006571 return -EINVAL;
NeilBrowna4a61252012-05-22 13:55:27 +10006572 if (mddev->bitmap) {
6573 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
6574 if (ret)
6575 return ret;
6576 }
6577 md_set_array_sectors(mddev, newsize);
Andre Nollf233ea52008-07-21 17:05:22 +10006578 set_capacity(mddev->gendisk, mddev->array_sectors);
NeilBrown449aad32009-08-03 10:59:58 +10006579 revalidate_disk(mddev->gendisk);
NeilBrownb0986362011-05-11 15:52:21 +10006580 if (sectors > mddev->dev_sectors &&
6581 mddev->recovery_cp > mddev->dev_sectors) {
Andre Noll58c0fed2009-03-31 14:33:13 +11006582 mddev->recovery_cp = mddev->dev_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006583 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
6584 }
Andre Noll58c0fed2009-03-31 14:33:13 +11006585 mddev->dev_sectors = sectors;
NeilBrown4b5c7ae2005-07-27 11:43:28 -07006586 mddev->resync_max_sectors = sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006587 return 0;
6588}
6589
NeilBrownfd01b882011-10-11 16:47:53 +11006590static int check_stripe_cache(struct mddev *mddev)
NeilBrown01ee22b2009-06-18 08:47:20 +10006591{
6592 /* Can only proceed if there are plenty of stripe_heads.
6593 * We need a minimum of one full stripe,, and for sensible progress
6594 * it is best to have about 4 times that.
6595 * If we require 4 times, then the default 256 4K stripe_heads will
6596 * allow for chunk sizes up to 256K, which is probably OK.
6597 * If the chunk size is greater, user-space should request more
6598 * stripe_heads first.
6599 */
NeilBrownd1688a62011-10-11 16:49:52 +11006600 struct r5conf *conf = mddev->private;
NeilBrown01ee22b2009-06-18 08:47:20 +10006601 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
6602 > conf->max_nr_stripes ||
6603 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
6604 > conf->max_nr_stripes) {
NeilBrown0c55e022010-05-03 14:09:02 +10006605 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
6606 mdname(mddev),
NeilBrown01ee22b2009-06-18 08:47:20 +10006607 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
6608 / STRIPE_SIZE)*4);
6609 return 0;
6610 }
6611 return 1;
6612}
6613
NeilBrownfd01b882011-10-11 16:47:53 +11006614static int check_reshape(struct mddev *mddev)
NeilBrown29269552006-03-27 01:18:10 -08006615{
NeilBrownd1688a62011-10-11 16:49:52 +11006616 struct r5conf *conf = mddev->private;
NeilBrown29269552006-03-27 01:18:10 -08006617
NeilBrown88ce4932009-03-31 15:24:23 +11006618 if (mddev->delta_disks == 0 &&
6619 mddev->new_layout == mddev->layout &&
Andre Noll664e7c42009-06-18 08:45:27 +10006620 mddev->new_chunk_sectors == mddev->chunk_sectors)
NeilBrown50ac1682009-06-18 08:47:55 +10006621 return 0; /* nothing to do */
NeilBrown674806d2010-06-16 17:17:53 +10006622 if (has_failed(conf))
NeilBrownec32a2b2009-03-31 15:17:38 +11006623 return -EINVAL;
NeilBrownfdcfbbb2013-07-04 16:38:16 +10006624 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
NeilBrownec32a2b2009-03-31 15:17:38 +11006625 /* We might be able to shrink, but the devices must
6626 * be made bigger first.
6627 * For raid6, 4 is the minimum size.
6628 * Otherwise 2 is the minimum
6629 */
6630 int min = 2;
6631 if (mddev->level == 6)
6632 min = 4;
6633 if (mddev->raid_disks + mddev->delta_disks < min)
6634 return -EINVAL;
6635 }
NeilBrown29269552006-03-27 01:18:10 -08006636
NeilBrown01ee22b2009-06-18 08:47:20 +10006637 if (!check_stripe_cache(mddev))
NeilBrown29269552006-03-27 01:18:10 -08006638 return -ENOSPC;
NeilBrown29269552006-03-27 01:18:10 -08006639
NeilBrowne56108d62012-10-11 14:24:13 +11006640 return resize_stripes(conf, (conf->previous_raid_disks
6641 + mddev->delta_disks));
NeilBrown63c70c42006-03-27 01:18:13 -08006642}
6643
NeilBrownfd01b882011-10-11 16:47:53 +11006644static int raid5_start_reshape(struct mddev *mddev)
NeilBrown63c70c42006-03-27 01:18:13 -08006645{
NeilBrownd1688a62011-10-11 16:49:52 +11006646 struct r5conf *conf = mddev->private;
NeilBrown3cb03002011-10-11 16:45:26 +11006647 struct md_rdev *rdev;
NeilBrown63c70c42006-03-27 01:18:13 -08006648 int spares = 0;
NeilBrownc04be0a2006-10-03 01:15:53 -07006649 unsigned long flags;
NeilBrown63c70c42006-03-27 01:18:13 -08006650
NeilBrownf4168852007-02-28 20:11:53 -08006651 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
NeilBrown63c70c42006-03-27 01:18:13 -08006652 return -EBUSY;
6653
NeilBrown01ee22b2009-06-18 08:47:20 +10006654 if (!check_stripe_cache(mddev))
6655 return -ENOSPC;
6656
NeilBrown30b67642012-05-22 13:55:28 +10006657 if (has_failed(conf))
6658 return -EINVAL;
6659
NeilBrownc6563a82012-05-21 09:27:00 +10006660 rdev_for_each(rdev, mddev) {
NeilBrown469518a2011-01-31 11:57:43 +11006661 if (!test_bit(In_sync, &rdev->flags)
6662 && !test_bit(Faulty, &rdev->flags))
NeilBrown29269552006-03-27 01:18:10 -08006663 spares++;
NeilBrownc6563a82012-05-21 09:27:00 +10006664 }
NeilBrown63c70c42006-03-27 01:18:13 -08006665
NeilBrownf4168852007-02-28 20:11:53 -08006666 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
NeilBrown29269552006-03-27 01:18:10 -08006667 /* Not enough devices even to make a degraded array
6668 * of that size
6669 */
6670 return -EINVAL;
6671
NeilBrownec32a2b2009-03-31 15:17:38 +11006672 /* Refuse to reduce size of the array. Any reductions in
6673 * array size must be through explicit setting of array_size
6674 * attribute.
6675 */
6676 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
6677 < mddev->array_sectors) {
NeilBrown0c55e022010-05-03 14:09:02 +10006678 printk(KERN_ERR "md/raid:%s: array size must be reduced "
NeilBrownec32a2b2009-03-31 15:17:38 +11006679 "before number of disks\n", mdname(mddev));
6680 return -EINVAL;
6681 }
6682
NeilBrownf6705572006-03-27 01:18:11 -08006683 atomic_set(&conf->reshape_stripes, 0);
NeilBrown29269552006-03-27 01:18:10 -08006684 spin_lock_irq(&conf->device_lock);
NeilBrownc46501b2013-08-27 15:52:13 +10006685 write_seqcount_begin(&conf->gen_lock);
NeilBrown29269552006-03-27 01:18:10 -08006686 conf->previous_raid_disks = conf->raid_disks;
NeilBrown63c70c42006-03-27 01:18:13 -08006687 conf->raid_disks += mddev->delta_disks;
Andre Noll09c9e5f2009-06-18 08:45:55 +10006688 conf->prev_chunk_sectors = conf->chunk_sectors;
6689 conf->chunk_sectors = mddev->new_chunk_sectors;
NeilBrown88ce4932009-03-31 15:24:23 +11006690 conf->prev_algo = conf->algorithm;
6691 conf->algorithm = mddev->new_layout;
NeilBrown05616be2012-05-21 09:27:00 +10006692 conf->generation++;
6693 /* Code that selects data_offset needs to see the generation update
6694 * if reshape_progress has been set - so a memory barrier needed.
6695 */
6696 smp_mb();
NeilBrown2c810cd2012-05-21 09:27:00 +10006697 if (mddev->reshape_backwards)
NeilBrownfef9c612009-03-31 15:16:46 +11006698 conf->reshape_progress = raid5_size(mddev, 0, 0);
6699 else
6700 conf->reshape_progress = 0;
6701 conf->reshape_safe = conf->reshape_progress;
NeilBrownc46501b2013-08-27 15:52:13 +10006702 write_seqcount_end(&conf->gen_lock);
NeilBrown29269552006-03-27 01:18:10 -08006703 spin_unlock_irq(&conf->device_lock);
6704
NeilBrown4d77e3b2013-08-27 15:57:47 +10006705 /* Now make sure any requests that proceeded on the assumption
6706 * the reshape wasn't running - like Discard or Read - have
6707 * completed.
6708 */
6709 mddev_suspend(mddev);
6710 mddev_resume(mddev);
6711
NeilBrown29269552006-03-27 01:18:10 -08006712 /* Add some new drives, as many as will fit.
6713 * We know there are enough to make the newly sized array work.
NeilBrown3424bf62010-06-17 17:48:26 +10006714 * Don't add devices if we are reducing the number of
6715 * devices in the array. This is because it is not possible
6716 * to correctly record the "partially reconstructed" state of
6717 * such devices during the reshape and confusion could result.
NeilBrown29269552006-03-27 01:18:10 -08006718 */
NeilBrown87a8dec2011-01-31 11:57:43 +11006719 if (mddev->delta_disks >= 0) {
NeilBrowndafb20f2012-03-19 12:46:39 +11006720 rdev_for_each(rdev, mddev)
NeilBrown87a8dec2011-01-31 11:57:43 +11006721 if (rdev->raid_disk < 0 &&
6722 !test_bit(Faulty, &rdev->flags)) {
6723 if (raid5_add_disk(mddev, rdev) == 0) {
NeilBrown87a8dec2011-01-31 11:57:43 +11006724 if (rdev->raid_disk
NeilBrown9d4c7d82012-03-13 11:21:21 +11006725 >= conf->previous_raid_disks)
NeilBrown87a8dec2011-01-31 11:57:43 +11006726 set_bit(In_sync, &rdev->flags);
NeilBrown9d4c7d82012-03-13 11:21:21 +11006727 else
NeilBrown87a8dec2011-01-31 11:57:43 +11006728 rdev->recovery_offset = 0;
Namhyung Kim36fad852011-07-27 11:00:36 +10006729
6730 if (sysfs_link_rdev(mddev, rdev))
NeilBrown87a8dec2011-01-31 11:57:43 +11006731 /* Failure here is OK */;
NeilBrown50da0842011-01-31 11:57:43 +11006732 }
NeilBrown87a8dec2011-01-31 11:57:43 +11006733 } else if (rdev->raid_disk >= conf->previous_raid_disks
6734 && !test_bit(Faulty, &rdev->flags)) {
6735 /* This is a spare that was manually added */
6736 set_bit(In_sync, &rdev->flags);
NeilBrown87a8dec2011-01-31 11:57:43 +11006737 }
NeilBrown29269552006-03-27 01:18:10 -08006738
NeilBrown87a8dec2011-01-31 11:57:43 +11006739 /* When a reshape changes the number of devices,
6740 * ->degraded is measured against the larger of the
6741 * pre and post number of devices.
6742 */
NeilBrownec32a2b2009-03-31 15:17:38 +11006743 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrown908f4fb2011-12-23 10:17:50 +11006744 mddev->degraded = calc_degraded(conf);
NeilBrownec32a2b2009-03-31 15:17:38 +11006745 spin_unlock_irqrestore(&conf->device_lock, flags);
6746 }
NeilBrown63c70c42006-03-27 01:18:13 -08006747 mddev->raid_disks = conf->raid_disks;
NeilBrowne5164022009-08-03 10:59:57 +10006748 mddev->reshape_position = conf->reshape_progress;
NeilBrown850b2b42006-10-03 01:15:46 -07006749 set_bit(MD_CHANGE_DEVS, &mddev->flags);
NeilBrownf6705572006-03-27 01:18:11 -08006750
NeilBrown29269552006-03-27 01:18:10 -08006751 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6752 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6753 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6754 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6755 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
NeilBrown0da3c612009-09-23 18:09:45 +10006756 "reshape");
NeilBrown29269552006-03-27 01:18:10 -08006757 if (!mddev->sync_thread) {
6758 mddev->recovery = 0;
6759 spin_lock_irq(&conf->device_lock);
NeilBrownba8805b2013-11-14 15:16:15 +11006760 write_seqcount_begin(&conf->gen_lock);
NeilBrown29269552006-03-27 01:18:10 -08006761 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
NeilBrownba8805b2013-11-14 15:16:15 +11006762 mddev->new_chunk_sectors =
6763 conf->chunk_sectors = conf->prev_chunk_sectors;
6764 mddev->new_layout = conf->algorithm = conf->prev_algo;
NeilBrown05616be2012-05-21 09:27:00 +10006765 rdev_for_each(rdev, mddev)
6766 rdev->new_data_offset = rdev->data_offset;
6767 smp_wmb();
NeilBrownba8805b2013-11-14 15:16:15 +11006768 conf->generation --;
NeilBrownfef9c612009-03-31 15:16:46 +11006769 conf->reshape_progress = MaxSector;
NeilBrown1e3fa9b2012-03-13 11:21:18 +11006770 mddev->reshape_position = MaxSector;
NeilBrownba8805b2013-11-14 15:16:15 +11006771 write_seqcount_end(&conf->gen_lock);
NeilBrown29269552006-03-27 01:18:10 -08006772 spin_unlock_irq(&conf->device_lock);
6773 return -EAGAIN;
6774 }
NeilBrownc8f517c2009-03-31 15:28:40 +11006775 conf->reshape_checkpoint = jiffies;
NeilBrown29269552006-03-27 01:18:10 -08006776 md_wakeup_thread(mddev->sync_thread);
6777 md_new_event(mddev);
6778 return 0;
6779}
NeilBrown29269552006-03-27 01:18:10 -08006780
NeilBrownec32a2b2009-03-31 15:17:38 +11006781/* This is called from the reshape thread and should make any
6782 * changes needed in 'conf'
6783 */
NeilBrownd1688a62011-10-11 16:49:52 +11006784static void end_reshape(struct r5conf *conf)
NeilBrown29269552006-03-27 01:18:10 -08006785{
NeilBrown29269552006-03-27 01:18:10 -08006786
NeilBrownf6705572006-03-27 01:18:11 -08006787 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
NeilBrown05616be2012-05-21 09:27:00 +10006788 struct md_rdev *rdev;
Dan Williams80c3a6c2009-03-17 18:10:40 -07006789
NeilBrownf6705572006-03-27 01:18:11 -08006790 spin_lock_irq(&conf->device_lock);
NeilBrowncea9c222009-03-31 15:15:05 +11006791 conf->previous_raid_disks = conf->raid_disks;
NeilBrown05616be2012-05-21 09:27:00 +10006792 rdev_for_each(rdev, conf->mddev)
6793 rdev->data_offset = rdev->new_data_offset;
6794 smp_wmb();
NeilBrownfef9c612009-03-31 15:16:46 +11006795 conf->reshape_progress = MaxSector;
NeilBrownf6705572006-03-27 01:18:11 -08006796 spin_unlock_irq(&conf->device_lock);
NeilBrownb0f9ec02009-03-31 15:27:18 +11006797 wake_up(&conf->wait_for_overlap);
NeilBrown16a53ec2006-06-26 00:27:38 -07006798
6799 /* read-ahead size must cover two whole stripes, which is
6800 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6801 */
NeilBrown4a5add42010-06-01 19:37:28 +10006802 if (conf->mddev->queue) {
NeilBrowncea9c222009-03-31 15:15:05 +11006803 int data_disks = conf->raid_disks - conf->max_degraded;
Andre Noll09c9e5f2009-06-18 08:45:55 +10006804 int stripe = data_disks * ((conf->chunk_sectors << 9)
NeilBrowncea9c222009-03-31 15:15:05 +11006805 / PAGE_SIZE);
NeilBrown16a53ec2006-06-26 00:27:38 -07006806 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6807 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6808 }
NeilBrown29269552006-03-27 01:18:10 -08006809 }
NeilBrown29269552006-03-27 01:18:10 -08006810}
6811
NeilBrownec32a2b2009-03-31 15:17:38 +11006812/* This is called from the raid5d thread with mddev_lock held.
6813 * It makes config changes to the device.
6814 */
NeilBrownfd01b882011-10-11 16:47:53 +11006815static void raid5_finish_reshape(struct mddev *mddev)
NeilBrowncea9c222009-03-31 15:15:05 +11006816{
NeilBrownd1688a62011-10-11 16:49:52 +11006817 struct r5conf *conf = mddev->private;
NeilBrowncea9c222009-03-31 15:15:05 +11006818
6819 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6820
NeilBrownec32a2b2009-03-31 15:17:38 +11006821 if (mddev->delta_disks > 0) {
6822 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6823 set_capacity(mddev->gendisk, mddev->array_sectors);
NeilBrown449aad32009-08-03 10:59:58 +10006824 revalidate_disk(mddev->gendisk);
NeilBrownec32a2b2009-03-31 15:17:38 +11006825 } else {
6826 int d;
NeilBrown908f4fb2011-12-23 10:17:50 +11006827 spin_lock_irq(&conf->device_lock);
6828 mddev->degraded = calc_degraded(conf);
6829 spin_unlock_irq(&conf->device_lock);
NeilBrownec32a2b2009-03-31 15:17:38 +11006830 for (d = conf->raid_disks ;
6831 d < conf->raid_disks - mddev->delta_disks;
NeilBrown1a67dde2009-08-13 10:41:49 +10006832 d++) {
NeilBrown3cb03002011-10-11 16:45:26 +11006833 struct md_rdev *rdev = conf->disks[d].rdev;
NeilBrownda7613b2012-05-22 13:55:33 +10006834 if (rdev)
6835 clear_bit(In_sync, &rdev->flags);
6836 rdev = conf->disks[d].replacement;
6837 if (rdev)
6838 clear_bit(In_sync, &rdev->flags);
NeilBrown1a67dde2009-08-13 10:41:49 +10006839 }
NeilBrowncea9c222009-03-31 15:15:05 +11006840 }
NeilBrown88ce4932009-03-31 15:24:23 +11006841 mddev->layout = conf->algorithm;
Andre Noll09c9e5f2009-06-18 08:45:55 +10006842 mddev->chunk_sectors = conf->chunk_sectors;
NeilBrownec32a2b2009-03-31 15:17:38 +11006843 mddev->reshape_position = MaxSector;
6844 mddev->delta_disks = 0;
NeilBrown2c810cd2012-05-21 09:27:00 +10006845 mddev->reshape_backwards = 0;
NeilBrowncea9c222009-03-31 15:15:05 +11006846 }
6847}
6848
NeilBrownfd01b882011-10-11 16:47:53 +11006849static void raid5_quiesce(struct mddev *mddev, int state)
NeilBrown72626682005-09-09 16:23:54 -07006850{
NeilBrownd1688a62011-10-11 16:49:52 +11006851 struct r5conf *conf = mddev->private;
NeilBrown72626682005-09-09 16:23:54 -07006852
6853 switch(state) {
NeilBrowne464eaf2006-03-27 01:18:14 -08006854 case 2: /* resume for a suspend */
6855 wake_up(&conf->wait_for_overlap);
6856 break;
6857
NeilBrown72626682005-09-09 16:23:54 -07006858 case 1: /* stop all writes */
Shaohua Li566c09c2013-11-14 15:16:17 +11006859 lock_all_device_hash_locks_irq(conf);
NeilBrown64bd6602009-08-03 10:59:58 +10006860 /* '2' tells resync/reshape to pause so that all
6861 * active stripes can drain
6862 */
6863 conf->quiesce = 2;
Shaohua Li566c09c2013-11-14 15:16:17 +11006864 wait_event_cmd(conf->wait_for_stripe,
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08006865 atomic_read(&conf->active_stripes) == 0 &&
6866 atomic_read(&conf->active_aligned_reads) == 0,
Shaohua Li566c09c2013-11-14 15:16:17 +11006867 unlock_all_device_hash_locks_irq(conf),
6868 lock_all_device_hash_locks_irq(conf));
NeilBrown64bd6602009-08-03 10:59:58 +10006869 conf->quiesce = 1;
Shaohua Li566c09c2013-11-14 15:16:17 +11006870 unlock_all_device_hash_locks_irq(conf);
NeilBrown64bd6602009-08-03 10:59:58 +10006871 /* allow reshape to continue */
6872 wake_up(&conf->wait_for_overlap);
NeilBrown72626682005-09-09 16:23:54 -07006873 break;
6874
6875 case 0: /* re-enable writes */
Shaohua Li566c09c2013-11-14 15:16:17 +11006876 lock_all_device_hash_locks_irq(conf);
NeilBrown72626682005-09-09 16:23:54 -07006877 conf->quiesce = 0;
6878 wake_up(&conf->wait_for_stripe);
NeilBrowne464eaf2006-03-27 01:18:14 -08006879 wake_up(&conf->wait_for_overlap);
Shaohua Li566c09c2013-11-14 15:16:17 +11006880 unlock_all_device_hash_locks_irq(conf);
NeilBrown72626682005-09-09 16:23:54 -07006881 break;
6882 }
NeilBrown72626682005-09-09 16:23:54 -07006883}
NeilBrownb15c2e52006-01-06 00:20:16 -08006884
NeilBrownfd01b882011-10-11 16:47:53 +11006885static void *raid45_takeover_raid0(struct mddev *mddev, int level)
Trela Maciej54071b32010-03-08 16:02:42 +11006886{
NeilBrowne373ab12011-10-11 16:48:59 +11006887 struct r0conf *raid0_conf = mddev->private;
Randy Dunlapd76c8422011-04-21 09:07:26 -07006888 sector_t sectors;
Trela Maciej54071b32010-03-08 16:02:42 +11006889
Dan Williamsf1b29bc2010-05-01 18:09:05 -07006890 /* for raid0 takeover only one zone is supported */
NeilBrowne373ab12011-10-11 16:48:59 +11006891 if (raid0_conf->nr_strip_zones > 1) {
NeilBrown0c55e022010-05-03 14:09:02 +10006892 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6893 mdname(mddev));
Dan Williamsf1b29bc2010-05-01 18:09:05 -07006894 return ERR_PTR(-EINVAL);
6895 }
6896
NeilBrowne373ab12011-10-11 16:48:59 +11006897 sectors = raid0_conf->strip_zone[0].zone_end;
6898 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
NeilBrown3b71bd92011-04-20 15:38:18 +10006899 mddev->dev_sectors = sectors;
Dan Williamsf1b29bc2010-05-01 18:09:05 -07006900 mddev->new_level = level;
Trela Maciej54071b32010-03-08 16:02:42 +11006901 mddev->new_layout = ALGORITHM_PARITY_N;
6902 mddev->new_chunk_sectors = mddev->chunk_sectors;
6903 mddev->raid_disks += 1;
6904 mddev->delta_disks = 1;
6905 /* make sure it will be not marked as dirty */
6906 mddev->recovery_cp = MaxSector;
6907
6908 return setup_conf(mddev);
6909}
6910
NeilBrownfd01b882011-10-11 16:47:53 +11006911static void *raid5_takeover_raid1(struct mddev *mddev)
NeilBrownd562b0c2009-03-31 14:39:39 +11006912{
6913 int chunksect;
6914
6915 if (mddev->raid_disks != 2 ||
6916 mddev->degraded > 1)
6917 return ERR_PTR(-EINVAL);
6918
6919 /* Should check if there are write-behind devices? */
6920
6921 chunksect = 64*2; /* 64K by default */
6922
6923 /* The array must be an exact multiple of chunksize */
6924 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6925 chunksect >>= 1;
6926
6927 if ((chunksect<<9) < STRIPE_SIZE)
6928 /* array size does not allow a suitable chunk size */
6929 return ERR_PTR(-EINVAL);
6930
6931 mddev->new_level = 5;
6932 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
Andre Noll664e7c42009-06-18 08:45:27 +10006933 mddev->new_chunk_sectors = chunksect;
NeilBrownd562b0c2009-03-31 14:39:39 +11006934
6935 return setup_conf(mddev);
6936}
6937
NeilBrownfd01b882011-10-11 16:47:53 +11006938static void *raid5_takeover_raid6(struct mddev *mddev)
NeilBrownfc9739c2009-03-31 14:57:20 +11006939{
6940 int new_layout;
6941
6942 switch (mddev->layout) {
6943 case ALGORITHM_LEFT_ASYMMETRIC_6:
6944 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6945 break;
6946 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6947 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6948 break;
6949 case ALGORITHM_LEFT_SYMMETRIC_6:
6950 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6951 break;
6952 case ALGORITHM_RIGHT_SYMMETRIC_6:
6953 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6954 break;
6955 case ALGORITHM_PARITY_0_6:
6956 new_layout = ALGORITHM_PARITY_0;
6957 break;
6958 case ALGORITHM_PARITY_N:
6959 new_layout = ALGORITHM_PARITY_N;
6960 break;
6961 default:
6962 return ERR_PTR(-EINVAL);
6963 }
6964 mddev->new_level = 5;
6965 mddev->new_layout = new_layout;
6966 mddev->delta_disks = -1;
6967 mddev->raid_disks -= 1;
6968 return setup_conf(mddev);
6969}
6970
NeilBrownfd01b882011-10-11 16:47:53 +11006971static int raid5_check_reshape(struct mddev *mddev)
NeilBrownb3546032009-03-31 14:56:41 +11006972{
NeilBrown88ce4932009-03-31 15:24:23 +11006973 /* For a 2-drive array, the layout and chunk size can be changed
6974 * immediately as not restriping is needed.
6975 * For larger arrays we record the new value - after validation
6976 * to be used by a reshape pass.
NeilBrownb3546032009-03-31 14:56:41 +11006977 */
NeilBrownd1688a62011-10-11 16:49:52 +11006978 struct r5conf *conf = mddev->private;
NeilBrown597a7112009-06-18 08:47:42 +10006979 int new_chunk = mddev->new_chunk_sectors;
NeilBrownb3546032009-03-31 14:56:41 +11006980
NeilBrown597a7112009-06-18 08:47:42 +10006981 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
NeilBrownb3546032009-03-31 14:56:41 +11006982 return -EINVAL;
6983 if (new_chunk > 0) {
Andre Noll0ba459d2009-06-18 08:46:10 +10006984 if (!is_power_of_2(new_chunk))
NeilBrownb3546032009-03-31 14:56:41 +11006985 return -EINVAL;
NeilBrown597a7112009-06-18 08:47:42 +10006986 if (new_chunk < (PAGE_SIZE>>9))
NeilBrownb3546032009-03-31 14:56:41 +11006987 return -EINVAL;
NeilBrown597a7112009-06-18 08:47:42 +10006988 if (mddev->array_sectors & (new_chunk-1))
NeilBrownb3546032009-03-31 14:56:41 +11006989 /* not factor of array size */
6990 return -EINVAL;
6991 }
6992
6993 /* They look valid */
6994
NeilBrown88ce4932009-03-31 15:24:23 +11006995 if (mddev->raid_disks == 2) {
NeilBrown597a7112009-06-18 08:47:42 +10006996 /* can make the change immediately */
6997 if (mddev->new_layout >= 0) {
6998 conf->algorithm = mddev->new_layout;
6999 mddev->layout = mddev->new_layout;
NeilBrown88ce4932009-03-31 15:24:23 +11007000 }
7001 if (new_chunk > 0) {
NeilBrown597a7112009-06-18 08:47:42 +10007002 conf->chunk_sectors = new_chunk ;
7003 mddev->chunk_sectors = new_chunk;
NeilBrown88ce4932009-03-31 15:24:23 +11007004 }
7005 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7006 md_wakeup_thread(mddev->thread);
NeilBrownb3546032009-03-31 14:56:41 +11007007 }
NeilBrown50ac1682009-06-18 08:47:55 +10007008 return check_reshape(mddev);
NeilBrown88ce4932009-03-31 15:24:23 +11007009}
7010
NeilBrownfd01b882011-10-11 16:47:53 +11007011static int raid6_check_reshape(struct mddev *mddev)
NeilBrown88ce4932009-03-31 15:24:23 +11007012{
NeilBrown597a7112009-06-18 08:47:42 +10007013 int new_chunk = mddev->new_chunk_sectors;
NeilBrown50ac1682009-06-18 08:47:55 +10007014
NeilBrown597a7112009-06-18 08:47:42 +10007015 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
NeilBrown88ce4932009-03-31 15:24:23 +11007016 return -EINVAL;
NeilBrownb3546032009-03-31 14:56:41 +11007017 if (new_chunk > 0) {
Andre Noll0ba459d2009-06-18 08:46:10 +10007018 if (!is_power_of_2(new_chunk))
NeilBrown88ce4932009-03-31 15:24:23 +11007019 return -EINVAL;
NeilBrown597a7112009-06-18 08:47:42 +10007020 if (new_chunk < (PAGE_SIZE >> 9))
NeilBrown88ce4932009-03-31 15:24:23 +11007021 return -EINVAL;
NeilBrown597a7112009-06-18 08:47:42 +10007022 if (mddev->array_sectors & (new_chunk-1))
NeilBrown88ce4932009-03-31 15:24:23 +11007023 /* not factor of array size */
7024 return -EINVAL;
NeilBrownb3546032009-03-31 14:56:41 +11007025 }
NeilBrown88ce4932009-03-31 15:24:23 +11007026
7027 /* They look valid */
NeilBrown50ac1682009-06-18 08:47:55 +10007028 return check_reshape(mddev);
NeilBrownb3546032009-03-31 14:56:41 +11007029}
7030
NeilBrownfd01b882011-10-11 16:47:53 +11007031static void *raid5_takeover(struct mddev *mddev)
NeilBrownd562b0c2009-03-31 14:39:39 +11007032{
7033 /* raid5 can take over:
Dan Williamsf1b29bc2010-05-01 18:09:05 -07007034 * raid0 - if there is only one strip zone - make it a raid4 layout
NeilBrownd562b0c2009-03-31 14:39:39 +11007035 * raid1 - if there are two drives. We need to know the chunk size
7036 * raid4 - trivial - just use a raid4 layout.
7037 * raid6 - Providing it is a *_6 layout
NeilBrownd562b0c2009-03-31 14:39:39 +11007038 */
Dan Williamsf1b29bc2010-05-01 18:09:05 -07007039 if (mddev->level == 0)
7040 return raid45_takeover_raid0(mddev, 5);
NeilBrownd562b0c2009-03-31 14:39:39 +11007041 if (mddev->level == 1)
7042 return raid5_takeover_raid1(mddev);
NeilBrowne9d47582009-03-31 14:57:09 +11007043 if (mddev->level == 4) {
7044 mddev->new_layout = ALGORITHM_PARITY_N;
7045 mddev->new_level = 5;
7046 return setup_conf(mddev);
7047 }
NeilBrownfc9739c2009-03-31 14:57:20 +11007048 if (mddev->level == 6)
7049 return raid5_takeover_raid6(mddev);
NeilBrownd562b0c2009-03-31 14:39:39 +11007050
7051 return ERR_PTR(-EINVAL);
7052}
7053
NeilBrownfd01b882011-10-11 16:47:53 +11007054static void *raid4_takeover(struct mddev *mddev)
NeilBrowna78d38a2010-03-22 16:53:49 +11007055{
Dan Williamsf1b29bc2010-05-01 18:09:05 -07007056 /* raid4 can take over:
7057 * raid0 - if there is only one strip zone
7058 * raid5 - if layout is right
NeilBrowna78d38a2010-03-22 16:53:49 +11007059 */
Dan Williamsf1b29bc2010-05-01 18:09:05 -07007060 if (mddev->level == 0)
7061 return raid45_takeover_raid0(mddev, 4);
NeilBrowna78d38a2010-03-22 16:53:49 +11007062 if (mddev->level == 5 &&
7063 mddev->layout == ALGORITHM_PARITY_N) {
7064 mddev->new_layout = 0;
7065 mddev->new_level = 4;
7066 return setup_conf(mddev);
7067 }
7068 return ERR_PTR(-EINVAL);
7069}
NeilBrownd562b0c2009-03-31 14:39:39 +11007070
NeilBrown84fc4b52011-10-11 16:49:58 +11007071static struct md_personality raid5_personality;
NeilBrown245f46c2009-03-31 14:39:39 +11007072
NeilBrownfd01b882011-10-11 16:47:53 +11007073static void *raid6_takeover(struct mddev *mddev)
NeilBrown245f46c2009-03-31 14:39:39 +11007074{
7075 /* Currently can only take over a raid5. We map the
7076 * personality to an equivalent raid6 personality
7077 * with the Q block at the end.
7078 */
7079 int new_layout;
7080
7081 if (mddev->pers != &raid5_personality)
7082 return ERR_PTR(-EINVAL);
7083 if (mddev->degraded > 1)
7084 return ERR_PTR(-EINVAL);
7085 if (mddev->raid_disks > 253)
7086 return ERR_PTR(-EINVAL);
7087 if (mddev->raid_disks < 3)
7088 return ERR_PTR(-EINVAL);
7089
7090 switch (mddev->layout) {
7091 case ALGORITHM_LEFT_ASYMMETRIC:
7092 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7093 break;
7094 case ALGORITHM_RIGHT_ASYMMETRIC:
7095 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7096 break;
7097 case ALGORITHM_LEFT_SYMMETRIC:
7098 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7099 break;
7100 case ALGORITHM_RIGHT_SYMMETRIC:
7101 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7102 break;
7103 case ALGORITHM_PARITY_0:
7104 new_layout = ALGORITHM_PARITY_0_6;
7105 break;
7106 case ALGORITHM_PARITY_N:
7107 new_layout = ALGORITHM_PARITY_N;
7108 break;
7109 default:
7110 return ERR_PTR(-EINVAL);
7111 }
7112 mddev->new_level = 6;
7113 mddev->new_layout = new_layout;
7114 mddev->delta_disks = 1;
7115 mddev->raid_disks += 1;
7116 return setup_conf(mddev);
7117}
7118
NeilBrown84fc4b52011-10-11 16:49:58 +11007119static struct md_personality raid6_personality =
NeilBrown16a53ec2006-06-26 00:27:38 -07007120{
7121 .name = "raid6",
7122 .level = 6,
7123 .owner = THIS_MODULE,
7124 .make_request = make_request,
7125 .run = run,
NeilBrownafa0f552014-12-15 12:56:58 +11007126 .free = raid5_free,
NeilBrown16a53ec2006-06-26 00:27:38 -07007127 .status = status,
7128 .error_handler = error,
7129 .hot_add_disk = raid5_add_disk,
7130 .hot_remove_disk= raid5_remove_disk,
7131 .spare_active = raid5_spare_active,
7132 .sync_request = sync_request,
7133 .resize = raid5_resize,
Dan Williams80c3a6c2009-03-17 18:10:40 -07007134 .size = raid5_size,
NeilBrown50ac1682009-06-18 08:47:55 +10007135 .check_reshape = raid6_check_reshape,
NeilBrownf4168852007-02-28 20:11:53 -08007136 .start_reshape = raid5_start_reshape,
NeilBrowncea9c222009-03-31 15:15:05 +11007137 .finish_reshape = raid5_finish_reshape,
NeilBrown16a53ec2006-06-26 00:27:38 -07007138 .quiesce = raid5_quiesce,
NeilBrown245f46c2009-03-31 14:39:39 +11007139 .takeover = raid6_takeover,
NeilBrown5c675f82014-12-15 12:56:56 +11007140 .congested = raid5_congested,
NeilBrown64590f42014-12-15 12:56:57 +11007141 .mergeable_bvec = raid5_mergeable_bvec,
NeilBrown16a53ec2006-06-26 00:27:38 -07007142};
NeilBrown84fc4b52011-10-11 16:49:58 +11007143static struct md_personality raid5_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07007144{
7145 .name = "raid5",
NeilBrown2604b702006-01-06 00:20:36 -08007146 .level = 5,
Linus Torvalds1da177e2005-04-16 15:20:36 -07007147 .owner = THIS_MODULE,
7148 .make_request = make_request,
7149 .run = run,
NeilBrownafa0f552014-12-15 12:56:58 +11007150 .free = raid5_free,
Linus Torvalds1da177e2005-04-16 15:20:36 -07007151 .status = status,
7152 .error_handler = error,
7153 .hot_add_disk = raid5_add_disk,
7154 .hot_remove_disk= raid5_remove_disk,
7155 .spare_active = raid5_spare_active,
7156 .sync_request = sync_request,
7157 .resize = raid5_resize,
Dan Williams80c3a6c2009-03-17 18:10:40 -07007158 .size = raid5_size,
NeilBrown63c70c42006-03-27 01:18:13 -08007159 .check_reshape = raid5_check_reshape,
7160 .start_reshape = raid5_start_reshape,
NeilBrowncea9c222009-03-31 15:15:05 +11007161 .finish_reshape = raid5_finish_reshape,
NeilBrown72626682005-09-09 16:23:54 -07007162 .quiesce = raid5_quiesce,
NeilBrownd562b0c2009-03-31 14:39:39 +11007163 .takeover = raid5_takeover,
NeilBrown5c675f82014-12-15 12:56:56 +11007164 .congested = raid5_congested,
NeilBrown64590f42014-12-15 12:56:57 +11007165 .mergeable_bvec = raid5_mergeable_bvec,
Linus Torvalds1da177e2005-04-16 15:20:36 -07007166};
7167
NeilBrown84fc4b52011-10-11 16:49:58 +11007168static struct md_personality raid4_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07007169{
NeilBrown2604b702006-01-06 00:20:36 -08007170 .name = "raid4",
7171 .level = 4,
7172 .owner = THIS_MODULE,
7173 .make_request = make_request,
7174 .run = run,
NeilBrownafa0f552014-12-15 12:56:58 +11007175 .free = raid5_free,
NeilBrown2604b702006-01-06 00:20:36 -08007176 .status = status,
7177 .error_handler = error,
7178 .hot_add_disk = raid5_add_disk,
7179 .hot_remove_disk= raid5_remove_disk,
7180 .spare_active = raid5_spare_active,
7181 .sync_request = sync_request,
7182 .resize = raid5_resize,
Dan Williams80c3a6c2009-03-17 18:10:40 -07007183 .size = raid5_size,
NeilBrown3d378902007-03-26 21:32:13 -08007184 .check_reshape = raid5_check_reshape,
7185 .start_reshape = raid5_start_reshape,
NeilBrowncea9c222009-03-31 15:15:05 +11007186 .finish_reshape = raid5_finish_reshape,
NeilBrown2604b702006-01-06 00:20:36 -08007187 .quiesce = raid5_quiesce,
NeilBrowna78d38a2010-03-22 16:53:49 +11007188 .takeover = raid4_takeover,
NeilBrown5c675f82014-12-15 12:56:56 +11007189 .congested = raid5_congested,
NeilBrown64590f42014-12-15 12:56:57 +11007190 .mergeable_bvec = raid5_mergeable_bvec,
NeilBrown2604b702006-01-06 00:20:36 -08007191};
7192
7193static int __init raid5_init(void)
7194{
Shaohua Li851c30c2013-08-28 14:30:16 +08007195 raid5_wq = alloc_workqueue("raid5wq",
7196 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7197 if (!raid5_wq)
7198 return -ENOMEM;
NeilBrown16a53ec2006-06-26 00:27:38 -07007199 register_md_personality(&raid6_personality);
NeilBrown2604b702006-01-06 00:20:36 -08007200 register_md_personality(&raid5_personality);
7201 register_md_personality(&raid4_personality);
7202 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007203}
7204
NeilBrown2604b702006-01-06 00:20:36 -08007205static void raid5_exit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007206{
NeilBrown16a53ec2006-06-26 00:27:38 -07007207 unregister_md_personality(&raid6_personality);
NeilBrown2604b702006-01-06 00:20:36 -08007208 unregister_md_personality(&raid5_personality);
7209 unregister_md_personality(&raid4_personality);
Shaohua Li851c30c2013-08-28 14:30:16 +08007210 destroy_workqueue(raid5_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007211}
7212
7213module_init(raid5_init);
7214module_exit(raid5_exit);
7215MODULE_LICENSE("GPL");
NeilBrown0efb9e62009-12-14 12:49:58 +11007216MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
Linus Torvalds1da177e2005-04-16 15:20:36 -07007217MODULE_ALIAS("md-personality-4"); /* RAID5 */
NeilBrownd9d166c2006-01-06 00:20:51 -08007218MODULE_ALIAS("md-raid5");
7219MODULE_ALIAS("md-raid4");
NeilBrown2604b702006-01-06 00:20:36 -08007220MODULE_ALIAS("md-level-5");
7221MODULE_ALIAS("md-level-4");
NeilBrown16a53ec2006-06-26 00:27:38 -07007222MODULE_ALIAS("md-personality-8"); /* RAID6 */
7223MODULE_ALIAS("md-raid6");
7224MODULE_ALIAS("md-level-6");
7225
7226/* This used to be two separate modules, they were: */
7227MODULE_ALIAS("raid5");
7228MODULE_ALIAS("raid6");