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
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
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)
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.
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
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.
30 * 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
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
35 * the number of the batch it will be in. This is seq_flush+1.
36 * 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
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
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
58 #include <trace/events/block.h>
59 #include <linux/list_sort.h>
64 #include "md-bitmap.h"
65 #include "raid5-log.h"
67 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
69 #define cpu_to_group(cpu) cpu_to_node(cpu)
70 #define ANY_GROUP NUMA_NO_NODE
72 static bool devices_handle_discard_safely = false;
73 module_param(devices_handle_discard_safely, bool, 0644);
74 MODULE_PARM_DESC(devices_handle_discard_safely,
75 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
76 static struct workqueue_struct *raid5_wq;
78 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
80 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
81 return &conf->stripe_hashtbl[hash];
84 static inline int stripe_hash_locks_hash(sector_t sect)
86 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
89 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
91 spin_lock_irq(conf->hash_locks + hash);
92 spin_lock(&conf->device_lock);
95 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
97 spin_unlock(&conf->device_lock);
98 spin_unlock_irq(conf->hash_locks + hash);
101 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
104 spin_lock_irq(conf->hash_locks);
105 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
106 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
107 spin_lock(&conf->device_lock);
110 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
113 spin_unlock(&conf->device_lock);
114 for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
115 spin_unlock(conf->hash_locks + i);
116 spin_unlock_irq(conf->hash_locks);
119 /* Find first data disk in a raid6 stripe */
120 static inline int raid6_d0(struct stripe_head *sh)
123 /* ddf always start from first device */
125 /* md starts just after Q block */
126 if (sh->qd_idx == sh->disks - 1)
129 return sh->qd_idx + 1;
131 static inline int raid6_next_disk(int disk, int raid_disks)
134 return (disk < raid_disks) ? disk : 0;
137 /* When walking through the disks in a raid5, starting at raid6_d0,
138 * We need to map each disk to a 'slot', where the data disks are slot
139 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
140 * is raid_disks-1. This help does that mapping.
142 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
143 int *count, int syndrome_disks)
149 if (idx == sh->pd_idx)
150 return syndrome_disks;
151 if (idx == sh->qd_idx)
152 return syndrome_disks + 1;
158 static void print_raid5_conf (struct r5conf *conf);
160 static int stripe_operations_active(struct stripe_head *sh)
162 return sh->check_state || sh->reconstruct_state ||
163 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
164 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
167 static bool stripe_is_lowprio(struct stripe_head *sh)
169 return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
170 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
171 !test_bit(STRIPE_R5C_CACHING, &sh->state);
174 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
176 struct r5conf *conf = sh->raid_conf;
177 struct r5worker_group *group;
179 int i, cpu = sh->cpu;
181 if (!cpu_online(cpu)) {
182 cpu = cpumask_any(cpu_online_mask);
186 if (list_empty(&sh->lru)) {
187 struct r5worker_group *group;
188 group = conf->worker_groups + cpu_to_group(cpu);
189 if (stripe_is_lowprio(sh))
190 list_add_tail(&sh->lru, &group->loprio_list);
192 list_add_tail(&sh->lru, &group->handle_list);
193 group->stripes_cnt++;
197 if (conf->worker_cnt_per_group == 0) {
198 md_wakeup_thread(conf->mddev->thread);
202 group = conf->worker_groups + cpu_to_group(sh->cpu);
204 group->workers[0].working = true;
205 /* at least one worker should run to avoid race */
206 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
208 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
209 /* wakeup more workers */
210 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
211 if (group->workers[i].working == false) {
212 group->workers[i].working = true;
213 queue_work_on(sh->cpu, raid5_wq,
214 &group->workers[i].work);
220 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
221 struct list_head *temp_inactive_list)
224 int injournal = 0; /* number of date pages with R5_InJournal */
226 BUG_ON(!list_empty(&sh->lru));
227 BUG_ON(atomic_read(&conf->active_stripes)==0);
229 if (r5c_is_writeback(conf->log))
230 for (i = sh->disks; i--; )
231 if (test_bit(R5_InJournal, &sh->dev[i].flags))
234 * In the following cases, the stripe cannot be released to cached
235 * lists. Therefore, we make the stripe write out and set
237 * 1. when quiesce in r5c write back;
238 * 2. when resync is requested fot the stripe.
240 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
241 (conf->quiesce && r5c_is_writeback(conf->log) &&
242 !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
243 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
244 r5c_make_stripe_write_out(sh);
245 set_bit(STRIPE_HANDLE, &sh->state);
248 if (test_bit(STRIPE_HANDLE, &sh->state)) {
249 if (test_bit(STRIPE_DELAYED, &sh->state) &&
250 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
251 list_add_tail(&sh->lru, &conf->delayed_list);
252 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
253 sh->bm_seq - conf->seq_write > 0)
254 list_add_tail(&sh->lru, &conf->bitmap_list);
256 clear_bit(STRIPE_DELAYED, &sh->state);
257 clear_bit(STRIPE_BIT_DELAY, &sh->state);
258 if (conf->worker_cnt_per_group == 0) {
259 if (stripe_is_lowprio(sh))
260 list_add_tail(&sh->lru,
263 list_add_tail(&sh->lru,
266 raid5_wakeup_stripe_thread(sh);
270 md_wakeup_thread(conf->mddev->thread);
272 BUG_ON(stripe_operations_active(sh));
273 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
274 if (atomic_dec_return(&conf->preread_active_stripes)
276 md_wakeup_thread(conf->mddev->thread);
277 atomic_dec(&conf->active_stripes);
278 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
279 if (!r5c_is_writeback(conf->log))
280 list_add_tail(&sh->lru, temp_inactive_list);
282 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
284 list_add_tail(&sh->lru, temp_inactive_list);
285 else if (injournal == conf->raid_disks - conf->max_degraded) {
287 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
288 atomic_inc(&conf->r5c_cached_full_stripes);
289 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
290 atomic_dec(&conf->r5c_cached_partial_stripes);
291 list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
292 r5c_check_cached_full_stripe(conf);
295 * STRIPE_R5C_PARTIAL_STRIPE is set in
296 * r5c_try_caching_write(). No need to
299 list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
305 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
306 struct list_head *temp_inactive_list)
308 if (atomic_dec_and_test(&sh->count))
309 do_release_stripe(conf, sh, temp_inactive_list);
313 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
315 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
316 * given time. Adding stripes only takes device lock, while deleting stripes
317 * only takes hash lock.
319 static void release_inactive_stripe_list(struct r5conf *conf,
320 struct list_head *temp_inactive_list,
324 bool do_wakeup = false;
327 if (hash == NR_STRIPE_HASH_LOCKS) {
328 size = NR_STRIPE_HASH_LOCKS;
329 hash = NR_STRIPE_HASH_LOCKS - 1;
333 struct list_head *list = &temp_inactive_list[size - 1];
336 * We don't hold any lock here yet, raid5_get_active_stripe() might
337 * remove stripes from the list
339 if (!list_empty_careful(list)) {
340 spin_lock_irqsave(conf->hash_locks + hash, flags);
341 if (list_empty(conf->inactive_list + hash) &&
343 atomic_dec(&conf->empty_inactive_list_nr);
344 list_splice_tail_init(list, conf->inactive_list + hash);
346 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
353 wake_up(&conf->wait_for_stripe);
354 if (atomic_read(&conf->active_stripes) == 0)
355 wake_up(&conf->wait_for_quiescent);
356 if (conf->retry_read_aligned)
357 md_wakeup_thread(conf->mddev->thread);
361 /* should hold conf->device_lock already */
362 static int release_stripe_list(struct r5conf *conf,
363 struct list_head *temp_inactive_list)
365 struct stripe_head *sh, *t;
367 struct llist_node *head;
369 head = llist_del_all(&conf->released_stripes);
370 head = llist_reverse_order(head);
371 llist_for_each_entry_safe(sh, t, head, release_list) {
374 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
376 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
378 * Don't worry the bit is set here, because if the bit is set
379 * again, the count is always > 1. This is true for
380 * STRIPE_ON_UNPLUG_LIST bit too.
382 hash = sh->hash_lock_index;
383 __release_stripe(conf, sh, &temp_inactive_list[hash]);
390 void raid5_release_stripe(struct stripe_head *sh)
392 struct r5conf *conf = sh->raid_conf;
394 struct list_head list;
398 /* Avoid release_list until the last reference.
400 if (atomic_add_unless(&sh->count, -1, 1))
403 if (unlikely(!conf->mddev->thread) ||
404 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
406 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
408 md_wakeup_thread(conf->mddev->thread);
411 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
412 if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
413 INIT_LIST_HEAD(&list);
414 hash = sh->hash_lock_index;
415 do_release_stripe(conf, sh, &list);
416 spin_unlock_irqrestore(&conf->device_lock, flags);
417 release_inactive_stripe_list(conf, &list, hash);
421 static inline void remove_hash(struct stripe_head *sh)
423 pr_debug("remove_hash(), stripe %llu\n",
424 (unsigned long long)sh->sector);
426 hlist_del_init(&sh->hash);
429 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
431 struct hlist_head *hp = stripe_hash(conf, sh->sector);
433 pr_debug("insert_hash(), stripe %llu\n",
434 (unsigned long long)sh->sector);
436 hlist_add_head(&sh->hash, hp);
439 /* find an idle stripe, make sure it is unhashed, and return it. */
440 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
442 struct stripe_head *sh = NULL;
443 struct list_head *first;
445 if (list_empty(conf->inactive_list + hash))
447 first = (conf->inactive_list + hash)->next;
448 sh = list_entry(first, struct stripe_head, lru);
449 list_del_init(first);
451 atomic_inc(&conf->active_stripes);
452 BUG_ON(hash != sh->hash_lock_index);
453 if (list_empty(conf->inactive_list + hash))
454 atomic_inc(&conf->empty_inactive_list_nr);
459 static void shrink_buffers(struct stripe_head *sh)
463 int num = sh->raid_conf->pool_size;
465 for (i = 0; i < num ; i++) {
466 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
470 sh->dev[i].page = NULL;
475 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
478 int num = sh->raid_conf->pool_size;
480 for (i = 0; i < num; i++) {
483 if (!(page = alloc_page(gfp))) {
486 sh->dev[i].page = page;
487 sh->dev[i].orig_page = page;
493 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
494 struct stripe_head *sh);
496 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
498 struct r5conf *conf = sh->raid_conf;
501 BUG_ON(atomic_read(&sh->count) != 0);
502 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
503 BUG_ON(stripe_operations_active(sh));
504 BUG_ON(sh->batch_head);
506 pr_debug("init_stripe called, stripe %llu\n",
507 (unsigned long long)sector);
509 seq = read_seqcount_begin(&conf->gen_lock);
510 sh->generation = conf->generation - previous;
511 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
513 stripe_set_idx(sector, conf, previous, sh);
516 for (i = sh->disks; i--; ) {
517 struct r5dev *dev = &sh->dev[i];
519 if (dev->toread || dev->read || dev->towrite || dev->written ||
520 test_bit(R5_LOCKED, &dev->flags)) {
521 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
522 (unsigned long long)sh->sector, i, dev->toread,
523 dev->read, dev->towrite, dev->written,
524 test_bit(R5_LOCKED, &dev->flags));
528 dev->sector = raid5_compute_blocknr(sh, i, previous);
530 if (read_seqcount_retry(&conf->gen_lock, seq))
532 sh->overwrite_disks = 0;
533 insert_hash(conf, sh);
534 sh->cpu = smp_processor_id();
535 set_bit(STRIPE_BATCH_READY, &sh->state);
538 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
541 struct stripe_head *sh;
543 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
544 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
545 if (sh->sector == sector && sh->generation == generation)
547 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
552 * Need to check if array has failed when deciding whether to:
554 * - remove non-faulty devices
557 * This determination is simple when no reshape is happening.
558 * However if there is a reshape, we need to carefully check
559 * both the before and after sections.
560 * This is because some failed devices may only affect one
561 * of the two sections, and some non-in_sync devices may
562 * be insync in the section most affected by failed devices.
564 int raid5_calc_degraded(struct r5conf *conf)
566 int degraded, degraded2;
571 for (i = 0; i < conf->previous_raid_disks; i++) {
572 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
573 if (rdev && test_bit(Faulty, &rdev->flags))
574 rdev = rcu_dereference(conf->disks[i].replacement);
575 if (!rdev || test_bit(Faulty, &rdev->flags))
577 else if (test_bit(In_sync, &rdev->flags))
580 /* not in-sync or faulty.
581 * If the reshape increases the number of devices,
582 * this is being recovered by the reshape, so
583 * this 'previous' section is not in_sync.
584 * If the number of devices is being reduced however,
585 * the device can only be part of the array if
586 * we are reverting a reshape, so this section will
589 if (conf->raid_disks >= conf->previous_raid_disks)
593 if (conf->raid_disks == conf->previous_raid_disks)
597 for (i = 0; i < conf->raid_disks; i++) {
598 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
599 if (rdev && test_bit(Faulty, &rdev->flags))
600 rdev = rcu_dereference(conf->disks[i].replacement);
601 if (!rdev || test_bit(Faulty, &rdev->flags))
603 else if (test_bit(In_sync, &rdev->flags))
606 /* not in-sync or faulty.
607 * If reshape increases the number of devices, this
608 * section has already been recovered, else it
609 * almost certainly hasn't.
611 if (conf->raid_disks <= conf->previous_raid_disks)
615 if (degraded2 > degraded)
620 static int has_failed(struct r5conf *conf)
624 if (conf->mddev->reshape_position == MaxSector)
625 return conf->mddev->degraded > conf->max_degraded;
627 degraded = raid5_calc_degraded(conf);
628 if (degraded > conf->max_degraded)
634 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
635 int previous, int noblock, int noquiesce)
637 struct stripe_head *sh;
638 int hash = stripe_hash_locks_hash(sector);
639 int inc_empty_inactive_list_flag;
641 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
643 spin_lock_irq(conf->hash_locks + hash);
646 wait_event_lock_irq(conf->wait_for_quiescent,
647 conf->quiesce == 0 || noquiesce,
648 *(conf->hash_locks + hash));
649 sh = __find_stripe(conf, sector, conf->generation - previous);
651 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
652 sh = get_free_stripe(conf, hash);
653 if (!sh && !test_bit(R5_DID_ALLOC,
655 set_bit(R5_ALLOC_MORE,
658 if (noblock && sh == NULL)
661 r5c_check_stripe_cache_usage(conf);
663 set_bit(R5_INACTIVE_BLOCKED,
665 r5l_wake_reclaim(conf->log, 0);
667 conf->wait_for_stripe,
668 !list_empty(conf->inactive_list + hash) &&
669 (atomic_read(&conf->active_stripes)
670 < (conf->max_nr_stripes * 3 / 4)
671 || !test_bit(R5_INACTIVE_BLOCKED,
672 &conf->cache_state)),
673 *(conf->hash_locks + hash));
674 clear_bit(R5_INACTIVE_BLOCKED,
677 init_stripe(sh, sector, previous);
678 atomic_inc(&sh->count);
680 } else if (!atomic_inc_not_zero(&sh->count)) {
681 spin_lock(&conf->device_lock);
682 if (!atomic_read(&sh->count)) {
683 if (!test_bit(STRIPE_HANDLE, &sh->state))
684 atomic_inc(&conf->active_stripes);
685 BUG_ON(list_empty(&sh->lru) &&
686 !test_bit(STRIPE_EXPANDING, &sh->state));
687 inc_empty_inactive_list_flag = 0;
688 if (!list_empty(conf->inactive_list + hash))
689 inc_empty_inactive_list_flag = 1;
690 list_del_init(&sh->lru);
691 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
692 atomic_inc(&conf->empty_inactive_list_nr);
694 sh->group->stripes_cnt--;
698 atomic_inc(&sh->count);
699 spin_unlock(&conf->device_lock);
701 } while (sh == NULL);
703 spin_unlock_irq(conf->hash_locks + hash);
707 static bool is_full_stripe_write(struct stripe_head *sh)
709 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
710 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
713 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
716 spin_lock_irq(&sh2->stripe_lock);
717 spin_lock_nested(&sh1->stripe_lock, 1);
719 spin_lock_irq(&sh1->stripe_lock);
720 spin_lock_nested(&sh2->stripe_lock, 1);
724 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
726 spin_unlock(&sh1->stripe_lock);
727 spin_unlock_irq(&sh2->stripe_lock);
730 /* Only freshly new full stripe normal write stripe can be added to a batch list */
731 static bool stripe_can_batch(struct stripe_head *sh)
733 struct r5conf *conf = sh->raid_conf;
735 if (raid5_has_log(conf) || raid5_has_ppl(conf))
737 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
738 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
739 is_full_stripe_write(sh);
742 /* we only do back search */
743 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
745 struct stripe_head *head;
746 sector_t head_sector, tmp_sec;
749 int inc_empty_inactive_list_flag;
751 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
752 tmp_sec = sh->sector;
753 if (!sector_div(tmp_sec, conf->chunk_sectors))
755 head_sector = sh->sector - STRIPE_SECTORS;
757 hash = stripe_hash_locks_hash(head_sector);
758 spin_lock_irq(conf->hash_locks + hash);
759 head = __find_stripe(conf, head_sector, conf->generation);
760 if (head && !atomic_inc_not_zero(&head->count)) {
761 spin_lock(&conf->device_lock);
762 if (!atomic_read(&head->count)) {
763 if (!test_bit(STRIPE_HANDLE, &head->state))
764 atomic_inc(&conf->active_stripes);
765 BUG_ON(list_empty(&head->lru) &&
766 !test_bit(STRIPE_EXPANDING, &head->state));
767 inc_empty_inactive_list_flag = 0;
768 if (!list_empty(conf->inactive_list + hash))
769 inc_empty_inactive_list_flag = 1;
770 list_del_init(&head->lru);
771 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
772 atomic_inc(&conf->empty_inactive_list_nr);
774 head->group->stripes_cnt--;
778 atomic_inc(&head->count);
779 spin_unlock(&conf->device_lock);
781 spin_unlock_irq(conf->hash_locks + hash);
785 if (!stripe_can_batch(head))
788 lock_two_stripes(head, sh);
789 /* clear_batch_ready clear the flag */
790 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
797 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
799 if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
800 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
803 if (head->batch_head) {
804 spin_lock(&head->batch_head->batch_lock);
805 /* This batch list is already running */
806 if (!stripe_can_batch(head)) {
807 spin_unlock(&head->batch_head->batch_lock);
811 * We must assign batch_head of this stripe within the
812 * batch_lock, otherwise clear_batch_ready of batch head
813 * stripe could clear BATCH_READY bit of this stripe and
814 * this stripe->batch_head doesn't get assigned, which
815 * could confuse clear_batch_ready for this stripe
817 sh->batch_head = head->batch_head;
820 * at this point, head's BATCH_READY could be cleared, but we
821 * can still add the stripe to batch list
823 list_add(&sh->batch_list, &head->batch_list);
824 spin_unlock(&head->batch_head->batch_lock);
826 head->batch_head = head;
827 sh->batch_head = head->batch_head;
828 spin_lock(&head->batch_lock);
829 list_add_tail(&sh->batch_list, &head->batch_list);
830 spin_unlock(&head->batch_lock);
833 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
834 if (atomic_dec_return(&conf->preread_active_stripes)
836 md_wakeup_thread(conf->mddev->thread);
838 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
839 int seq = sh->bm_seq;
840 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
841 sh->batch_head->bm_seq > seq)
842 seq = sh->batch_head->bm_seq;
843 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
844 sh->batch_head->bm_seq = seq;
847 atomic_inc(&sh->count);
849 unlock_two_stripes(head, sh);
851 raid5_release_stripe(head);
854 /* Determine if 'data_offset' or 'new_data_offset' should be used
855 * in this stripe_head.
857 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
859 sector_t progress = conf->reshape_progress;
860 /* Need a memory barrier to make sure we see the value
861 * of conf->generation, or ->data_offset that was set before
862 * reshape_progress was updated.
865 if (progress == MaxSector)
867 if (sh->generation == conf->generation - 1)
869 /* We are in a reshape, and this is a new-generation stripe,
870 * so use new_data_offset.
875 static void dispatch_bio_list(struct bio_list *tmp)
879 while ((bio = bio_list_pop(tmp)))
880 generic_make_request(bio);
883 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
885 const struct r5pending_data *da = list_entry(a,
886 struct r5pending_data, sibling);
887 const struct r5pending_data *db = list_entry(b,
888 struct r5pending_data, sibling);
889 if (da->sector > db->sector)
891 if (da->sector < db->sector)
896 static void dispatch_defer_bios(struct r5conf *conf, int target,
897 struct bio_list *list)
899 struct r5pending_data *data;
900 struct list_head *first, *next = NULL;
903 if (conf->pending_data_cnt == 0)
906 list_sort(NULL, &conf->pending_list, cmp_stripe);
908 first = conf->pending_list.next;
910 /* temporarily move the head */
911 if (conf->next_pending_data)
912 list_move_tail(&conf->pending_list,
913 &conf->next_pending_data->sibling);
915 while (!list_empty(&conf->pending_list)) {
916 data = list_first_entry(&conf->pending_list,
917 struct r5pending_data, sibling);
918 if (&data->sibling == first)
919 first = data->sibling.next;
920 next = data->sibling.next;
922 bio_list_merge(list, &data->bios);
923 list_move(&data->sibling, &conf->free_list);
928 conf->pending_data_cnt -= cnt;
929 BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
931 if (next != &conf->pending_list)
932 conf->next_pending_data = list_entry(next,
933 struct r5pending_data, sibling);
935 conf->next_pending_data = NULL;
936 /* list isn't empty */
937 if (first != &conf->pending_list)
938 list_move_tail(&conf->pending_list, first);
941 static void flush_deferred_bios(struct r5conf *conf)
943 struct bio_list tmp = BIO_EMPTY_LIST;
945 if (conf->pending_data_cnt == 0)
948 spin_lock(&conf->pending_bios_lock);
949 dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
950 BUG_ON(conf->pending_data_cnt != 0);
951 spin_unlock(&conf->pending_bios_lock);
953 dispatch_bio_list(&tmp);
956 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
957 struct bio_list *bios)
959 struct bio_list tmp = BIO_EMPTY_LIST;
960 struct r5pending_data *ent;
962 spin_lock(&conf->pending_bios_lock);
963 ent = list_first_entry(&conf->free_list, struct r5pending_data,
965 list_move_tail(&ent->sibling, &conf->pending_list);
966 ent->sector = sector;
967 bio_list_init(&ent->bios);
968 bio_list_merge(&ent->bios, bios);
969 conf->pending_data_cnt++;
970 if (conf->pending_data_cnt >= PENDING_IO_MAX)
971 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
973 spin_unlock(&conf->pending_bios_lock);
975 dispatch_bio_list(&tmp);
979 raid5_end_read_request(struct bio *bi);
981 raid5_end_write_request(struct bio *bi);
983 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
985 struct r5conf *conf = sh->raid_conf;
986 int i, disks = sh->disks;
987 struct stripe_head *head_sh = sh;
988 struct bio_list pending_bios = BIO_EMPTY_LIST;
993 if (log_stripe(sh, s) == 0)
996 should_defer = conf->batch_bio_dispatch && conf->group_cnt;
998 for (i = disks; i--; ) {
999 int op, op_flags = 0;
1000 int replace_only = 0;
1001 struct bio *bi, *rbi;
1002 struct md_rdev *rdev, *rrdev = NULL;
1005 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1007 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1009 if (test_bit(R5_Discard, &sh->dev[i].flags))
1010 op = REQ_OP_DISCARD;
1011 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1013 else if (test_and_clear_bit(R5_WantReplace,
1014 &sh->dev[i].flags)) {
1019 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1020 op_flags |= REQ_SYNC;
1023 bi = &sh->dev[i].req;
1024 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1027 rrdev = rcu_dereference(conf->disks[i].replacement);
1028 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1029 rdev = rcu_dereference(conf->disks[i].rdev);
1034 if (op_is_write(op)) {
1038 /* We raced and saw duplicates */
1041 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1046 if (rdev && test_bit(Faulty, &rdev->flags))
1049 atomic_inc(&rdev->nr_pending);
1050 if (rrdev && test_bit(Faulty, &rrdev->flags))
1053 atomic_inc(&rrdev->nr_pending);
1056 /* We have already checked bad blocks for reads. Now
1057 * need to check for writes. We never accept write errors
1058 * on the replacement, so we don't to check rrdev.
1060 while (op_is_write(op) && rdev &&
1061 test_bit(WriteErrorSeen, &rdev->flags)) {
1064 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1065 &first_bad, &bad_sectors);
1070 set_bit(BlockedBadBlocks, &rdev->flags);
1071 if (!conf->mddev->external &&
1072 conf->mddev->sb_flags) {
1073 /* It is very unlikely, but we might
1074 * still need to write out the
1075 * bad block log - better give it
1077 md_check_recovery(conf->mddev);
1080 * Because md_wait_for_blocked_rdev
1081 * will dec nr_pending, we must
1082 * increment it first.
1084 atomic_inc(&rdev->nr_pending);
1085 md_wait_for_blocked_rdev(rdev, conf->mddev);
1087 /* Acknowledged bad block - skip the write */
1088 rdev_dec_pending(rdev, conf->mddev);
1094 if (s->syncing || s->expanding || s->expanded
1096 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1098 set_bit(STRIPE_IO_STARTED, &sh->state);
1100 bio_set_dev(bi, rdev->bdev);
1101 bio_set_op_attrs(bi, op, op_flags);
1102 bi->bi_end_io = op_is_write(op)
1103 ? raid5_end_write_request
1104 : raid5_end_read_request;
1105 bi->bi_private = sh;
1107 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1108 __func__, (unsigned long long)sh->sector,
1110 atomic_inc(&sh->count);
1112 atomic_inc(&head_sh->count);
1113 if (use_new_offset(conf, sh))
1114 bi->bi_iter.bi_sector = (sh->sector
1115 + rdev->new_data_offset);
1117 bi->bi_iter.bi_sector = (sh->sector
1118 + rdev->data_offset);
1119 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1120 bi->bi_opf |= REQ_NOMERGE;
1122 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1123 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1125 if (!op_is_write(op) &&
1126 test_bit(R5_InJournal, &sh->dev[i].flags))
1128 * issuing read for a page in journal, this
1129 * must be preparing for prexor in rmw; read
1130 * the data into orig_page
1132 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1134 sh->dev[i].vec.bv_page = sh->dev[i].page;
1136 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1137 bi->bi_io_vec[0].bv_offset = 0;
1138 bi->bi_iter.bi_size = STRIPE_SIZE;
1139 bi->bi_write_hint = sh->dev[i].write_hint;
1141 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1143 * If this is discard request, set bi_vcnt 0. We don't
1144 * want to confuse SCSI because SCSI will replace payload
1146 if (op == REQ_OP_DISCARD)
1149 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1151 if (conf->mddev->gendisk)
1152 trace_block_bio_remap(bi->bi_disk->queue,
1153 bi, disk_devt(conf->mddev->gendisk),
1155 if (should_defer && op_is_write(op))
1156 bio_list_add(&pending_bios, bi);
1158 generic_make_request(bi);
1161 if (s->syncing || s->expanding || s->expanded
1163 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1165 set_bit(STRIPE_IO_STARTED, &sh->state);
1167 bio_set_dev(rbi, rrdev->bdev);
1168 bio_set_op_attrs(rbi, op, op_flags);
1169 BUG_ON(!op_is_write(op));
1170 rbi->bi_end_io = raid5_end_write_request;
1171 rbi->bi_private = sh;
1173 pr_debug("%s: for %llu schedule op %d on "
1174 "replacement disc %d\n",
1175 __func__, (unsigned long long)sh->sector,
1177 atomic_inc(&sh->count);
1179 atomic_inc(&head_sh->count);
1180 if (use_new_offset(conf, sh))
1181 rbi->bi_iter.bi_sector = (sh->sector
1182 + rrdev->new_data_offset);
1184 rbi->bi_iter.bi_sector = (sh->sector
1185 + rrdev->data_offset);
1186 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1187 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1188 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1190 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1191 rbi->bi_io_vec[0].bv_offset = 0;
1192 rbi->bi_iter.bi_size = STRIPE_SIZE;
1193 rbi->bi_write_hint = sh->dev[i].write_hint;
1194 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1196 * If this is discard request, set bi_vcnt 0. We don't
1197 * want to confuse SCSI because SCSI will replace payload
1199 if (op == REQ_OP_DISCARD)
1201 if (conf->mddev->gendisk)
1202 trace_block_bio_remap(rbi->bi_disk->queue,
1203 rbi, disk_devt(conf->mddev->gendisk),
1205 if (should_defer && op_is_write(op))
1206 bio_list_add(&pending_bios, rbi);
1208 generic_make_request(rbi);
1210 if (!rdev && !rrdev) {
1211 if (op_is_write(op))
1212 set_bit(STRIPE_DEGRADED, &sh->state);
1213 pr_debug("skip op %d on disc %d for sector %llu\n",
1214 bi->bi_opf, i, (unsigned long long)sh->sector);
1215 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1216 set_bit(STRIPE_HANDLE, &sh->state);
1219 if (!head_sh->batch_head)
1221 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1227 if (should_defer && !bio_list_empty(&pending_bios))
1228 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1231 static struct dma_async_tx_descriptor *
1232 async_copy_data(int frombio, struct bio *bio, struct page **page,
1233 sector_t sector, struct dma_async_tx_descriptor *tx,
1234 struct stripe_head *sh, int no_skipcopy)
1237 struct bvec_iter iter;
1238 struct page *bio_page;
1240 struct async_submit_ctl submit;
1241 enum async_tx_flags flags = 0;
1243 if (bio->bi_iter.bi_sector >= sector)
1244 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1246 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1249 flags |= ASYNC_TX_FENCE;
1250 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1252 bio_for_each_segment(bvl, bio, iter) {
1253 int len = bvl.bv_len;
1257 if (page_offset < 0) {
1258 b_offset = -page_offset;
1259 page_offset += b_offset;
1263 if (len > 0 && page_offset + len > STRIPE_SIZE)
1264 clen = STRIPE_SIZE - page_offset;
1269 b_offset += bvl.bv_offset;
1270 bio_page = bvl.bv_page;
1272 if (sh->raid_conf->skip_copy &&
1273 b_offset == 0 && page_offset == 0 &&
1274 clen == STRIPE_SIZE &&
1278 tx = async_memcpy(*page, bio_page, page_offset,
1279 b_offset, clen, &submit);
1281 tx = async_memcpy(bio_page, *page, b_offset,
1282 page_offset, clen, &submit);
1284 /* chain the operations */
1285 submit.depend_tx = tx;
1287 if (clen < len) /* hit end of page */
1295 static void ops_complete_biofill(void *stripe_head_ref)
1297 struct stripe_head *sh = stripe_head_ref;
1300 pr_debug("%s: stripe %llu\n", __func__,
1301 (unsigned long long)sh->sector);
1303 /* clear completed biofills */
1304 for (i = sh->disks; i--; ) {
1305 struct r5dev *dev = &sh->dev[i];
1307 /* acknowledge completion of a biofill operation */
1308 /* and check if we need to reply to a read request,
1309 * new R5_Wantfill requests are held off until
1310 * !STRIPE_BIOFILL_RUN
1312 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1313 struct bio *rbi, *rbi2;
1318 while (rbi && rbi->bi_iter.bi_sector <
1319 dev->sector + STRIPE_SECTORS) {
1320 rbi2 = r5_next_bio(rbi, dev->sector);
1326 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1328 set_bit(STRIPE_HANDLE, &sh->state);
1329 raid5_release_stripe(sh);
1332 static void ops_run_biofill(struct stripe_head *sh)
1334 struct dma_async_tx_descriptor *tx = NULL;
1335 struct async_submit_ctl submit;
1338 BUG_ON(sh->batch_head);
1339 pr_debug("%s: stripe %llu\n", __func__,
1340 (unsigned long long)sh->sector);
1342 for (i = sh->disks; i--; ) {
1343 struct r5dev *dev = &sh->dev[i];
1344 if (test_bit(R5_Wantfill, &dev->flags)) {
1346 spin_lock_irq(&sh->stripe_lock);
1347 dev->read = rbi = dev->toread;
1349 spin_unlock_irq(&sh->stripe_lock);
1350 while (rbi && rbi->bi_iter.bi_sector <
1351 dev->sector + STRIPE_SECTORS) {
1352 tx = async_copy_data(0, rbi, &dev->page,
1353 dev->sector, tx, sh, 0);
1354 rbi = r5_next_bio(rbi, dev->sector);
1359 atomic_inc(&sh->count);
1360 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1361 async_trigger_callback(&submit);
1364 static void mark_target_uptodate(struct stripe_head *sh, int target)
1371 tgt = &sh->dev[target];
1372 set_bit(R5_UPTODATE, &tgt->flags);
1373 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1374 clear_bit(R5_Wantcompute, &tgt->flags);
1377 static void ops_complete_compute(void *stripe_head_ref)
1379 struct stripe_head *sh = stripe_head_ref;
1381 pr_debug("%s: stripe %llu\n", __func__,
1382 (unsigned long long)sh->sector);
1384 /* mark the computed target(s) as uptodate */
1385 mark_target_uptodate(sh, sh->ops.target);
1386 mark_target_uptodate(sh, sh->ops.target2);
1388 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1389 if (sh->check_state == check_state_compute_run)
1390 sh->check_state = check_state_compute_result;
1391 set_bit(STRIPE_HANDLE, &sh->state);
1392 raid5_release_stripe(sh);
1395 /* return a pointer to the address conversion region of the scribble buffer */
1396 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1398 return percpu->scribble + i * percpu->scribble_obj_size;
1401 /* return a pointer to the address conversion region of the scribble buffer */
1402 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1403 struct raid5_percpu *percpu, int i)
1405 return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1408 static struct dma_async_tx_descriptor *
1409 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1411 int disks = sh->disks;
1412 struct page **xor_srcs = to_addr_page(percpu, 0);
1413 int target = sh->ops.target;
1414 struct r5dev *tgt = &sh->dev[target];
1415 struct page *xor_dest = tgt->page;
1417 struct dma_async_tx_descriptor *tx;
1418 struct async_submit_ctl submit;
1421 BUG_ON(sh->batch_head);
1423 pr_debug("%s: stripe %llu block: %d\n",
1424 __func__, (unsigned long long)sh->sector, target);
1425 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1427 for (i = disks; i--; )
1429 xor_srcs[count++] = sh->dev[i].page;
1431 atomic_inc(&sh->count);
1433 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1434 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1435 if (unlikely(count == 1))
1436 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1438 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1443 /* set_syndrome_sources - populate source buffers for gen_syndrome
1444 * @srcs - (struct page *) array of size sh->disks
1445 * @sh - stripe_head to parse
1447 * Populates srcs in proper layout order for the stripe and returns the
1448 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1449 * destination buffer is recorded in srcs[count] and the Q destination
1450 * is recorded in srcs[count+1]].
1452 static int set_syndrome_sources(struct page **srcs,
1453 struct stripe_head *sh,
1456 int disks = sh->disks;
1457 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1458 int d0_idx = raid6_d0(sh);
1462 for (i = 0; i < disks; i++)
1468 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1469 struct r5dev *dev = &sh->dev[i];
1471 if (i == sh->qd_idx || i == sh->pd_idx ||
1472 (srctype == SYNDROME_SRC_ALL) ||
1473 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1474 (test_bit(R5_Wantdrain, &dev->flags) ||
1475 test_bit(R5_InJournal, &dev->flags))) ||
1476 (srctype == SYNDROME_SRC_WRITTEN &&
1478 test_bit(R5_InJournal, &dev->flags)))) {
1479 if (test_bit(R5_InJournal, &dev->flags))
1480 srcs[slot] = sh->dev[i].orig_page;
1482 srcs[slot] = sh->dev[i].page;
1484 i = raid6_next_disk(i, disks);
1485 } while (i != d0_idx);
1487 return syndrome_disks;
1490 static struct dma_async_tx_descriptor *
1491 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1493 int disks = sh->disks;
1494 struct page **blocks = to_addr_page(percpu, 0);
1496 int qd_idx = sh->qd_idx;
1497 struct dma_async_tx_descriptor *tx;
1498 struct async_submit_ctl submit;
1504 BUG_ON(sh->batch_head);
1505 if (sh->ops.target < 0)
1506 target = sh->ops.target2;
1507 else if (sh->ops.target2 < 0)
1508 target = sh->ops.target;
1510 /* we should only have one valid target */
1513 pr_debug("%s: stripe %llu block: %d\n",
1514 __func__, (unsigned long long)sh->sector, target);
1516 tgt = &sh->dev[target];
1517 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1520 atomic_inc(&sh->count);
1522 if (target == qd_idx) {
1523 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1524 blocks[count] = NULL; /* regenerating p is not necessary */
1525 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1526 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1527 ops_complete_compute, sh,
1528 to_addr_conv(sh, percpu, 0));
1529 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1531 /* Compute any data- or p-drive using XOR */
1533 for (i = disks; i-- ; ) {
1534 if (i == target || i == qd_idx)
1536 blocks[count++] = sh->dev[i].page;
1539 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1540 NULL, ops_complete_compute, sh,
1541 to_addr_conv(sh, percpu, 0));
1542 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1548 static struct dma_async_tx_descriptor *
1549 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1551 int i, count, disks = sh->disks;
1552 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1553 int d0_idx = raid6_d0(sh);
1554 int faila = -1, failb = -1;
1555 int target = sh->ops.target;
1556 int target2 = sh->ops.target2;
1557 struct r5dev *tgt = &sh->dev[target];
1558 struct r5dev *tgt2 = &sh->dev[target2];
1559 struct dma_async_tx_descriptor *tx;
1560 struct page **blocks = to_addr_page(percpu, 0);
1561 struct async_submit_ctl submit;
1563 BUG_ON(sh->batch_head);
1564 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1565 __func__, (unsigned long long)sh->sector, target, target2);
1566 BUG_ON(target < 0 || target2 < 0);
1567 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1568 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1570 /* we need to open-code set_syndrome_sources to handle the
1571 * slot number conversion for 'faila' and 'failb'
1573 for (i = 0; i < disks ; i++)
1578 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1580 blocks[slot] = sh->dev[i].page;
1586 i = raid6_next_disk(i, disks);
1587 } while (i != d0_idx);
1589 BUG_ON(faila == failb);
1592 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1593 __func__, (unsigned long long)sh->sector, faila, failb);
1595 atomic_inc(&sh->count);
1597 if (failb == syndrome_disks+1) {
1598 /* Q disk is one of the missing disks */
1599 if (faila == syndrome_disks) {
1600 /* Missing P+Q, just recompute */
1601 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1602 ops_complete_compute, sh,
1603 to_addr_conv(sh, percpu, 0));
1604 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1605 STRIPE_SIZE, &submit);
1609 int qd_idx = sh->qd_idx;
1611 /* Missing D+Q: recompute D from P, then recompute Q */
1612 if (target == qd_idx)
1613 data_target = target2;
1615 data_target = target;
1618 for (i = disks; i-- ; ) {
1619 if (i == data_target || i == qd_idx)
1621 blocks[count++] = sh->dev[i].page;
1623 dest = sh->dev[data_target].page;
1624 init_async_submit(&submit,
1625 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1627 to_addr_conv(sh, percpu, 0));
1628 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1631 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1632 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1633 ops_complete_compute, sh,
1634 to_addr_conv(sh, percpu, 0));
1635 return async_gen_syndrome(blocks, 0, count+2,
1636 STRIPE_SIZE, &submit);
1639 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1640 ops_complete_compute, sh,
1641 to_addr_conv(sh, percpu, 0));
1642 if (failb == syndrome_disks) {
1643 /* We're missing D+P. */
1644 return async_raid6_datap_recov(syndrome_disks+2,
1648 /* We're missing D+D. */
1649 return async_raid6_2data_recov(syndrome_disks+2,
1650 STRIPE_SIZE, faila, failb,
1656 static void ops_complete_prexor(void *stripe_head_ref)
1658 struct stripe_head *sh = stripe_head_ref;
1660 pr_debug("%s: stripe %llu\n", __func__,
1661 (unsigned long long)sh->sector);
1663 if (r5c_is_writeback(sh->raid_conf->log))
1665 * raid5-cache write back uses orig_page during prexor.
1666 * After prexor, it is time to free orig_page
1668 r5c_release_extra_page(sh);
1671 static struct dma_async_tx_descriptor *
1672 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1673 struct dma_async_tx_descriptor *tx)
1675 int disks = sh->disks;
1676 struct page **xor_srcs = to_addr_page(percpu, 0);
1677 int count = 0, pd_idx = sh->pd_idx, i;
1678 struct async_submit_ctl submit;
1680 /* existing parity data subtracted */
1681 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1683 BUG_ON(sh->batch_head);
1684 pr_debug("%s: stripe %llu\n", __func__,
1685 (unsigned long long)sh->sector);
1687 for (i = disks; i--; ) {
1688 struct r5dev *dev = &sh->dev[i];
1689 /* Only process blocks that are known to be uptodate */
1690 if (test_bit(R5_InJournal, &dev->flags))
1691 xor_srcs[count++] = dev->orig_page;
1692 else if (test_bit(R5_Wantdrain, &dev->flags))
1693 xor_srcs[count++] = dev->page;
1696 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1697 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1698 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1703 static struct dma_async_tx_descriptor *
1704 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1705 struct dma_async_tx_descriptor *tx)
1707 struct page **blocks = to_addr_page(percpu, 0);
1709 struct async_submit_ctl submit;
1711 pr_debug("%s: stripe %llu\n", __func__,
1712 (unsigned long long)sh->sector);
1714 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1716 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1717 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1718 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1723 static struct dma_async_tx_descriptor *
1724 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1726 struct r5conf *conf = sh->raid_conf;
1727 int disks = sh->disks;
1729 struct stripe_head *head_sh = sh;
1731 pr_debug("%s: stripe %llu\n", __func__,
1732 (unsigned long long)sh->sector);
1734 for (i = disks; i--; ) {
1739 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1745 * clear R5_InJournal, so when rewriting a page in
1746 * journal, it is not skipped by r5l_log_stripe()
1748 clear_bit(R5_InJournal, &dev->flags);
1749 spin_lock_irq(&sh->stripe_lock);
1750 chosen = dev->towrite;
1751 dev->towrite = NULL;
1752 sh->overwrite_disks = 0;
1753 BUG_ON(dev->written);
1754 wbi = dev->written = chosen;
1755 spin_unlock_irq(&sh->stripe_lock);
1756 WARN_ON(dev->page != dev->orig_page);
1758 while (wbi && wbi->bi_iter.bi_sector <
1759 dev->sector + STRIPE_SECTORS) {
1760 if (wbi->bi_opf & REQ_FUA)
1761 set_bit(R5_WantFUA, &dev->flags);
1762 if (wbi->bi_opf & REQ_SYNC)
1763 set_bit(R5_SyncIO, &dev->flags);
1764 if (bio_op(wbi) == REQ_OP_DISCARD)
1765 set_bit(R5_Discard, &dev->flags);
1767 tx = async_copy_data(1, wbi, &dev->page,
1768 dev->sector, tx, sh,
1769 r5c_is_writeback(conf->log));
1770 if (dev->page != dev->orig_page &&
1771 !r5c_is_writeback(conf->log)) {
1772 set_bit(R5_SkipCopy, &dev->flags);
1773 clear_bit(R5_UPTODATE, &dev->flags);
1774 clear_bit(R5_OVERWRITE, &dev->flags);
1777 wbi = r5_next_bio(wbi, dev->sector);
1780 if (head_sh->batch_head) {
1781 sh = list_first_entry(&sh->batch_list,
1794 static void ops_complete_reconstruct(void *stripe_head_ref)
1796 struct stripe_head *sh = stripe_head_ref;
1797 int disks = sh->disks;
1798 int pd_idx = sh->pd_idx;
1799 int qd_idx = sh->qd_idx;
1801 bool fua = false, sync = false, discard = false;
1803 pr_debug("%s: stripe %llu\n", __func__,
1804 (unsigned long long)sh->sector);
1806 for (i = disks; i--; ) {
1807 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1808 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1809 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1812 for (i = disks; i--; ) {
1813 struct r5dev *dev = &sh->dev[i];
1815 if (dev->written || i == pd_idx || i == qd_idx) {
1816 if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1817 set_bit(R5_UPTODATE, &dev->flags);
1818 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1819 set_bit(R5_Expanded, &dev->flags);
1822 set_bit(R5_WantFUA, &dev->flags);
1824 set_bit(R5_SyncIO, &dev->flags);
1828 if (sh->reconstruct_state == reconstruct_state_drain_run)
1829 sh->reconstruct_state = reconstruct_state_drain_result;
1830 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1831 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1833 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1834 sh->reconstruct_state = reconstruct_state_result;
1837 set_bit(STRIPE_HANDLE, &sh->state);
1838 raid5_release_stripe(sh);
1842 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1843 struct dma_async_tx_descriptor *tx)
1845 int disks = sh->disks;
1846 struct page **xor_srcs;
1847 struct async_submit_ctl submit;
1848 int count, pd_idx = sh->pd_idx, i;
1849 struct page *xor_dest;
1851 unsigned long flags;
1853 struct stripe_head *head_sh = sh;
1856 pr_debug("%s: stripe %llu\n", __func__,
1857 (unsigned long long)sh->sector);
1859 for (i = 0; i < sh->disks; i++) {
1862 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1865 if (i >= sh->disks) {
1866 atomic_inc(&sh->count);
1867 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1868 ops_complete_reconstruct(sh);
1873 xor_srcs = to_addr_page(percpu, j);
1874 /* check if prexor is active which means only process blocks
1875 * that are part of a read-modify-write (written)
1877 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1879 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1880 for (i = disks; i--; ) {
1881 struct r5dev *dev = &sh->dev[i];
1882 if (head_sh->dev[i].written ||
1883 test_bit(R5_InJournal, &head_sh->dev[i].flags))
1884 xor_srcs[count++] = dev->page;
1887 xor_dest = sh->dev[pd_idx].page;
1888 for (i = disks; i--; ) {
1889 struct r5dev *dev = &sh->dev[i];
1891 xor_srcs[count++] = dev->page;
1895 /* 1/ if we prexor'd then the dest is reused as a source
1896 * 2/ if we did not prexor then we are redoing the parity
1897 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1898 * for the synchronous xor case
1900 last_stripe = !head_sh->batch_head ||
1901 list_first_entry(&sh->batch_list,
1902 struct stripe_head, batch_list) == head_sh;
1904 flags = ASYNC_TX_ACK |
1905 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1907 atomic_inc(&head_sh->count);
1908 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1909 to_addr_conv(sh, percpu, j));
1911 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1912 init_async_submit(&submit, flags, tx, NULL, NULL,
1913 to_addr_conv(sh, percpu, j));
1916 if (unlikely(count == 1))
1917 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1919 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1922 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1929 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1930 struct dma_async_tx_descriptor *tx)
1932 struct async_submit_ctl submit;
1933 struct page **blocks;
1934 int count, i, j = 0;
1935 struct stripe_head *head_sh = sh;
1938 unsigned long txflags;
1940 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1942 for (i = 0; i < sh->disks; i++) {
1943 if (sh->pd_idx == i || sh->qd_idx == i)
1945 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1948 if (i >= sh->disks) {
1949 atomic_inc(&sh->count);
1950 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1951 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1952 ops_complete_reconstruct(sh);
1957 blocks = to_addr_page(percpu, j);
1959 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1960 synflags = SYNDROME_SRC_WRITTEN;
1961 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1963 synflags = SYNDROME_SRC_ALL;
1964 txflags = ASYNC_TX_ACK;
1967 count = set_syndrome_sources(blocks, sh, synflags);
1968 last_stripe = !head_sh->batch_head ||
1969 list_first_entry(&sh->batch_list,
1970 struct stripe_head, batch_list) == head_sh;
1973 atomic_inc(&head_sh->count);
1974 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1975 head_sh, to_addr_conv(sh, percpu, j));
1977 init_async_submit(&submit, 0, tx, NULL, NULL,
1978 to_addr_conv(sh, percpu, j));
1979 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1982 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1988 static void ops_complete_check(void *stripe_head_ref)
1990 struct stripe_head *sh = stripe_head_ref;
1992 pr_debug("%s: stripe %llu\n", __func__,
1993 (unsigned long long)sh->sector);
1995 sh->check_state = check_state_check_result;
1996 set_bit(STRIPE_HANDLE, &sh->state);
1997 raid5_release_stripe(sh);
2000 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2002 int disks = sh->disks;
2003 int pd_idx = sh->pd_idx;
2004 int qd_idx = sh->qd_idx;
2005 struct page *xor_dest;
2006 struct page **xor_srcs = to_addr_page(percpu, 0);
2007 struct dma_async_tx_descriptor *tx;
2008 struct async_submit_ctl submit;
2012 pr_debug("%s: stripe %llu\n", __func__,
2013 (unsigned long long)sh->sector);
2015 BUG_ON(sh->batch_head);
2017 xor_dest = sh->dev[pd_idx].page;
2018 xor_srcs[count++] = xor_dest;
2019 for (i = disks; i--; ) {
2020 if (i == pd_idx || i == qd_idx)
2022 xor_srcs[count++] = sh->dev[i].page;
2025 init_async_submit(&submit, 0, NULL, NULL, NULL,
2026 to_addr_conv(sh, percpu, 0));
2027 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2028 &sh->ops.zero_sum_result, &submit);
2030 atomic_inc(&sh->count);
2031 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2032 tx = async_trigger_callback(&submit);
2035 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2037 struct page **srcs = to_addr_page(percpu, 0);
2038 struct async_submit_ctl submit;
2041 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2042 (unsigned long long)sh->sector, checkp);
2044 BUG_ON(sh->batch_head);
2045 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2049 atomic_inc(&sh->count);
2050 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2051 sh, to_addr_conv(sh, percpu, 0));
2052 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2053 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2056 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2058 int overlap_clear = 0, i, disks = sh->disks;
2059 struct dma_async_tx_descriptor *tx = NULL;
2060 struct r5conf *conf = sh->raid_conf;
2061 int level = conf->level;
2062 struct raid5_percpu *percpu;
2066 percpu = per_cpu_ptr(conf->percpu, cpu);
2067 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2068 ops_run_biofill(sh);
2072 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2074 tx = ops_run_compute5(sh, percpu);
2076 if (sh->ops.target2 < 0 || sh->ops.target < 0)
2077 tx = ops_run_compute6_1(sh, percpu);
2079 tx = ops_run_compute6_2(sh, percpu);
2081 /* terminate the chain if reconstruct is not set to be run */
2082 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2086 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2088 tx = ops_run_prexor5(sh, percpu, tx);
2090 tx = ops_run_prexor6(sh, percpu, tx);
2093 if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2094 tx = ops_run_partial_parity(sh, percpu, tx);
2096 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2097 tx = ops_run_biodrain(sh, tx);
2101 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2103 ops_run_reconstruct5(sh, percpu, tx);
2105 ops_run_reconstruct6(sh, percpu, tx);
2108 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2109 if (sh->check_state == check_state_run)
2110 ops_run_check_p(sh, percpu);
2111 else if (sh->check_state == check_state_run_q)
2112 ops_run_check_pq(sh, percpu, 0);
2113 else if (sh->check_state == check_state_run_pq)
2114 ops_run_check_pq(sh, percpu, 1);
2119 if (overlap_clear && !sh->batch_head)
2120 for (i = disks; i--; ) {
2121 struct r5dev *dev = &sh->dev[i];
2122 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2123 wake_up(&sh->raid_conf->wait_for_overlap);
2128 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2131 __free_page(sh->ppl_page);
2132 kmem_cache_free(sc, sh);
2135 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2136 int disks, struct r5conf *conf)
2138 struct stripe_head *sh;
2141 sh = kmem_cache_zalloc(sc, gfp);
2143 spin_lock_init(&sh->stripe_lock);
2144 spin_lock_init(&sh->batch_lock);
2145 INIT_LIST_HEAD(&sh->batch_list);
2146 INIT_LIST_HEAD(&sh->lru);
2147 INIT_LIST_HEAD(&sh->r5c);
2148 INIT_LIST_HEAD(&sh->log_list);
2149 atomic_set(&sh->count, 1);
2150 sh->raid_conf = conf;
2151 sh->log_start = MaxSector;
2152 for (i = 0; i < disks; i++) {
2153 struct r5dev *dev = &sh->dev[i];
2155 bio_init(&dev->req, &dev->vec, 1);
2156 bio_init(&dev->rreq, &dev->rvec, 1);
2159 if (raid5_has_ppl(conf)) {
2160 sh->ppl_page = alloc_page(gfp);
2161 if (!sh->ppl_page) {
2162 free_stripe(sc, sh);
2169 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2171 struct stripe_head *sh;
2173 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2177 if (grow_buffers(sh, gfp)) {
2179 free_stripe(conf->slab_cache, sh);
2182 sh->hash_lock_index =
2183 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2184 /* we just created an active stripe so... */
2185 atomic_inc(&conf->active_stripes);
2187 raid5_release_stripe(sh);
2188 conf->max_nr_stripes++;
2192 static int grow_stripes(struct r5conf *conf, int num)
2194 struct kmem_cache *sc;
2195 size_t namelen = sizeof(conf->cache_name[0]);
2196 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2198 if (conf->mddev->gendisk)
2199 snprintf(conf->cache_name[0], namelen,
2200 "raid%d-%s", conf->level, mdname(conf->mddev));
2202 snprintf(conf->cache_name[0], namelen,
2203 "raid%d-%p", conf->level, conf->mddev);
2204 snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2206 conf->active_name = 0;
2207 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2208 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2212 conf->slab_cache = sc;
2213 conf->pool_size = devs;
2215 if (!grow_one_stripe(conf, GFP_KERNEL))
2222 * scribble_len - return the required size of the scribble region
2223 * @num - total number of disks in the array
2225 * The size must be enough to contain:
2226 * 1/ a struct page pointer for each device in the array +2
2227 * 2/ room to convert each entry in (1) to its corresponding dma
2228 * (dma_map_page()) or page (page_address()) address.
2230 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2231 * calculate over all devices (not just the data blocks), using zeros in place
2232 * of the P and Q blocks.
2234 static int scribble_alloc(struct raid5_percpu *percpu,
2235 int num, int cnt, gfp_t flags)
2238 sizeof(struct page *) * (num+2) +
2239 sizeof(addr_conv_t) * (num+2);
2242 scribble = kvmalloc_array(cnt, obj_size, flags);
2246 kvfree(percpu->scribble);
2248 percpu->scribble = scribble;
2249 percpu->scribble_obj_size = obj_size;
2253 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2259 * Never shrink. And mddev_suspend() could deadlock if this is called
2260 * from raid5d. In that case, scribble_disks and scribble_sectors
2261 * should equal to new_disks and new_sectors
2263 if (conf->scribble_disks >= new_disks &&
2264 conf->scribble_sectors >= new_sectors)
2266 mddev_suspend(conf->mddev);
2269 for_each_present_cpu(cpu) {
2270 struct raid5_percpu *percpu;
2272 percpu = per_cpu_ptr(conf->percpu, cpu);
2273 err = scribble_alloc(percpu, new_disks,
2274 new_sectors / STRIPE_SECTORS,
2281 mddev_resume(conf->mddev);
2283 conf->scribble_disks = new_disks;
2284 conf->scribble_sectors = new_sectors;
2289 static int resize_stripes(struct r5conf *conf, int newsize)
2291 /* Make all the stripes able to hold 'newsize' devices.
2292 * New slots in each stripe get 'page' set to a new page.
2294 * This happens in stages:
2295 * 1/ create a new kmem_cache and allocate the required number of
2297 * 2/ gather all the old stripe_heads and transfer the pages across
2298 * to the new stripe_heads. This will have the side effect of
2299 * freezing the array as once all stripe_heads have been collected,
2300 * no IO will be possible. Old stripe heads are freed once their
2301 * pages have been transferred over, and the old kmem_cache is
2302 * freed when all stripes are done.
2303 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2304 * we simple return a failure status - no need to clean anything up.
2305 * 4/ allocate new pages for the new slots in the new stripe_heads.
2306 * If this fails, we don't bother trying the shrink the
2307 * stripe_heads down again, we just leave them as they are.
2308 * As each stripe_head is processed the new one is released into
2311 * Once step2 is started, we cannot afford to wait for a write,
2312 * so we use GFP_NOIO allocations.
2314 struct stripe_head *osh, *nsh;
2315 LIST_HEAD(newstripes);
2316 struct disk_info *ndisks;
2318 struct kmem_cache *sc;
2322 md_allow_write(conf->mddev);
2325 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2326 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2331 /* Need to ensure auto-resizing doesn't interfere */
2332 mutex_lock(&conf->cache_size_mutex);
2334 for (i = conf->max_nr_stripes; i; i--) {
2335 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2339 list_add(&nsh->lru, &newstripes);
2342 /* didn't get enough, give up */
2343 while (!list_empty(&newstripes)) {
2344 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2345 list_del(&nsh->lru);
2346 free_stripe(sc, nsh);
2348 kmem_cache_destroy(sc);
2349 mutex_unlock(&conf->cache_size_mutex);
2352 /* Step 2 - Must use GFP_NOIO now.
2353 * OK, we have enough stripes, start collecting inactive
2354 * stripes and copying them over
2358 list_for_each_entry(nsh, &newstripes, lru) {
2359 lock_device_hash_lock(conf, hash);
2360 wait_event_cmd(conf->wait_for_stripe,
2361 !list_empty(conf->inactive_list + hash),
2362 unlock_device_hash_lock(conf, hash),
2363 lock_device_hash_lock(conf, hash));
2364 osh = get_free_stripe(conf, hash);
2365 unlock_device_hash_lock(conf, hash);
2367 for(i=0; i<conf->pool_size; i++) {
2368 nsh->dev[i].page = osh->dev[i].page;
2369 nsh->dev[i].orig_page = osh->dev[i].page;
2371 nsh->hash_lock_index = hash;
2372 free_stripe(conf->slab_cache, osh);
2374 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2375 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2380 kmem_cache_destroy(conf->slab_cache);
2383 * At this point, we are holding all the stripes so the array
2384 * is completely stalled, so now is a good time to resize
2385 * conf->disks and the scribble region
2387 ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2389 for (i = 0; i < conf->pool_size; i++)
2390 ndisks[i] = conf->disks[i];
2392 for (i = conf->pool_size; i < newsize; i++) {
2393 ndisks[i].extra_page = alloc_page(GFP_NOIO);
2394 if (!ndisks[i].extra_page)
2399 for (i = conf->pool_size; i < newsize; i++)
2400 if (ndisks[i].extra_page)
2401 put_page(ndisks[i].extra_page);
2405 conf->disks = ndisks;
2410 mutex_unlock(&conf->cache_size_mutex);
2412 conf->slab_cache = sc;
2413 conf->active_name = 1-conf->active_name;
2415 /* Step 4, return new stripes to service */
2416 while(!list_empty(&newstripes)) {
2417 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2418 list_del_init(&nsh->lru);
2420 for (i=conf->raid_disks; i < newsize; i++)
2421 if (nsh->dev[i].page == NULL) {
2422 struct page *p = alloc_page(GFP_NOIO);
2423 nsh->dev[i].page = p;
2424 nsh->dev[i].orig_page = p;
2428 raid5_release_stripe(nsh);
2430 /* critical section pass, GFP_NOIO no longer needed */
2433 conf->pool_size = newsize;
2437 static int drop_one_stripe(struct r5conf *conf)
2439 struct stripe_head *sh;
2440 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2442 spin_lock_irq(conf->hash_locks + hash);
2443 sh = get_free_stripe(conf, hash);
2444 spin_unlock_irq(conf->hash_locks + hash);
2447 BUG_ON(atomic_read(&sh->count));
2449 free_stripe(conf->slab_cache, sh);
2450 atomic_dec(&conf->active_stripes);
2451 conf->max_nr_stripes--;
2455 static void shrink_stripes(struct r5conf *conf)
2457 while (conf->max_nr_stripes &&
2458 drop_one_stripe(conf))
2461 kmem_cache_destroy(conf->slab_cache);
2462 conf->slab_cache = NULL;
2465 static void raid5_end_read_request(struct bio * bi)
2467 struct stripe_head *sh = bi->bi_private;
2468 struct r5conf *conf = sh->raid_conf;
2469 int disks = sh->disks, i;
2470 char b[BDEVNAME_SIZE];
2471 struct md_rdev *rdev = NULL;
2474 for (i=0 ; i<disks; i++)
2475 if (bi == &sh->dev[i].req)
2478 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2479 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2486 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2487 /* If replacement finished while this request was outstanding,
2488 * 'replacement' might be NULL already.
2489 * In that case it moved down to 'rdev'.
2490 * rdev is not removed until all requests are finished.
2492 rdev = conf->disks[i].replacement;
2494 rdev = conf->disks[i].rdev;
2496 if (use_new_offset(conf, sh))
2497 s = sh->sector + rdev->new_data_offset;
2499 s = sh->sector + rdev->data_offset;
2500 if (!bi->bi_status) {
2501 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2502 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2503 /* Note that this cannot happen on a
2504 * replacement device. We just fail those on
2507 pr_info_ratelimited(
2508 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2509 mdname(conf->mddev), STRIPE_SECTORS,
2510 (unsigned long long)s,
2511 bdevname(rdev->bdev, b));
2512 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2513 clear_bit(R5_ReadError, &sh->dev[i].flags);
2514 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2515 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2516 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2518 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2520 * end read for a page in journal, this
2521 * must be preparing for prexor in rmw
2523 set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2525 if (atomic_read(&rdev->read_errors))
2526 atomic_set(&rdev->read_errors, 0);
2528 const char *bdn = bdevname(rdev->bdev, b);
2532 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2533 atomic_inc(&rdev->read_errors);
2534 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2535 pr_warn_ratelimited(
2536 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2537 mdname(conf->mddev),
2538 (unsigned long long)s,
2540 else if (conf->mddev->degraded >= conf->max_degraded) {
2542 pr_warn_ratelimited(
2543 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2544 mdname(conf->mddev),
2545 (unsigned long long)s,
2547 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2550 pr_warn_ratelimited(
2551 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2552 mdname(conf->mddev),
2553 (unsigned long long)s,
2555 } else if (atomic_read(&rdev->read_errors)
2556 > conf->max_nr_stripes)
2557 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2558 mdname(conf->mddev), bdn);
2561 if (set_bad && test_bit(In_sync, &rdev->flags)
2562 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2565 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2566 set_bit(R5_ReadError, &sh->dev[i].flags);
2567 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2569 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2571 clear_bit(R5_ReadError, &sh->dev[i].flags);
2572 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2574 && test_bit(In_sync, &rdev->flags)
2575 && rdev_set_badblocks(
2576 rdev, sh->sector, STRIPE_SECTORS, 0)))
2577 md_error(conf->mddev, rdev);
2580 rdev_dec_pending(rdev, conf->mddev);
2582 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2583 set_bit(STRIPE_HANDLE, &sh->state);
2584 raid5_release_stripe(sh);
2587 static void raid5_end_write_request(struct bio *bi)
2589 struct stripe_head *sh = bi->bi_private;
2590 struct r5conf *conf = sh->raid_conf;
2591 int disks = sh->disks, i;
2592 struct md_rdev *uninitialized_var(rdev);
2595 int replacement = 0;
2597 for (i = 0 ; i < disks; i++) {
2598 if (bi == &sh->dev[i].req) {
2599 rdev = conf->disks[i].rdev;
2602 if (bi == &sh->dev[i].rreq) {
2603 rdev = conf->disks[i].replacement;
2607 /* rdev was removed and 'replacement'
2608 * replaced it. rdev is not removed
2609 * until all requests are finished.
2611 rdev = conf->disks[i].rdev;
2615 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2616 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2626 md_error(conf->mddev, rdev);
2627 else if (is_badblock(rdev, sh->sector,
2629 &first_bad, &bad_sectors))
2630 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2632 if (bi->bi_status) {
2633 set_bit(STRIPE_DEGRADED, &sh->state);
2634 set_bit(WriteErrorSeen, &rdev->flags);
2635 set_bit(R5_WriteError, &sh->dev[i].flags);
2636 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2637 set_bit(MD_RECOVERY_NEEDED,
2638 &rdev->mddev->recovery);
2639 } else if (is_badblock(rdev, sh->sector,
2641 &first_bad, &bad_sectors)) {
2642 set_bit(R5_MadeGood, &sh->dev[i].flags);
2643 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2644 /* That was a successful write so make
2645 * sure it looks like we already did
2648 set_bit(R5_ReWrite, &sh->dev[i].flags);
2651 rdev_dec_pending(rdev, conf->mddev);
2653 if (sh->batch_head && bi->bi_status && !replacement)
2654 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2657 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2658 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2659 set_bit(STRIPE_HANDLE, &sh->state);
2660 raid5_release_stripe(sh);
2662 if (sh->batch_head && sh != sh->batch_head)
2663 raid5_release_stripe(sh->batch_head);
2666 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2668 char b[BDEVNAME_SIZE];
2669 struct r5conf *conf = mddev->private;
2670 unsigned long flags;
2671 pr_debug("raid456: error called\n");
2673 spin_lock_irqsave(&conf->device_lock, flags);
2675 if (test_bit(In_sync, &rdev->flags) &&
2676 mddev->degraded == conf->max_degraded) {
2678 * Don't allow to achieve failed state
2679 * Don't try to recover this device
2681 conf->recovery_disabled = mddev->recovery_disabled;
2682 spin_unlock_irqrestore(&conf->device_lock, flags);
2686 set_bit(Faulty, &rdev->flags);
2687 clear_bit(In_sync, &rdev->flags);
2688 mddev->degraded = raid5_calc_degraded(conf);
2689 spin_unlock_irqrestore(&conf->device_lock, flags);
2690 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2692 set_bit(Blocked, &rdev->flags);
2693 set_mask_bits(&mddev->sb_flags, 0,
2694 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2695 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2696 "md/raid:%s: Operation continuing on %d devices.\n",
2698 bdevname(rdev->bdev, b),
2700 conf->raid_disks - mddev->degraded);
2701 r5c_update_on_rdev_error(mddev, rdev);
2705 * Input: a 'big' sector number,
2706 * Output: index of the data and parity disk, and the sector # in them.
2708 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2709 int previous, int *dd_idx,
2710 struct stripe_head *sh)
2712 sector_t stripe, stripe2;
2713 sector_t chunk_number;
2714 unsigned int chunk_offset;
2717 sector_t new_sector;
2718 int algorithm = previous ? conf->prev_algo
2720 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2721 : conf->chunk_sectors;
2722 int raid_disks = previous ? conf->previous_raid_disks
2724 int data_disks = raid_disks - conf->max_degraded;
2726 /* First compute the information on this sector */
2729 * Compute the chunk number and the sector offset inside the chunk
2731 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2732 chunk_number = r_sector;
2735 * Compute the stripe number
2737 stripe = chunk_number;
2738 *dd_idx = sector_div(stripe, data_disks);
2741 * Select the parity disk based on the user selected algorithm.
2743 pd_idx = qd_idx = -1;
2744 switch(conf->level) {
2746 pd_idx = data_disks;
2749 switch (algorithm) {
2750 case ALGORITHM_LEFT_ASYMMETRIC:
2751 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2752 if (*dd_idx >= pd_idx)
2755 case ALGORITHM_RIGHT_ASYMMETRIC:
2756 pd_idx = sector_div(stripe2, raid_disks);
2757 if (*dd_idx >= pd_idx)
2760 case ALGORITHM_LEFT_SYMMETRIC:
2761 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2762 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2764 case ALGORITHM_RIGHT_SYMMETRIC:
2765 pd_idx = sector_div(stripe2, raid_disks);
2766 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2768 case ALGORITHM_PARITY_0:
2772 case ALGORITHM_PARITY_N:
2773 pd_idx = data_disks;
2781 switch (algorithm) {
2782 case ALGORITHM_LEFT_ASYMMETRIC:
2783 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2784 qd_idx = pd_idx + 1;
2785 if (pd_idx == raid_disks-1) {
2786 (*dd_idx)++; /* Q D D D P */
2788 } else if (*dd_idx >= pd_idx)
2789 (*dd_idx) += 2; /* D D P Q D */
2791 case ALGORITHM_RIGHT_ASYMMETRIC:
2792 pd_idx = sector_div(stripe2, raid_disks);
2793 qd_idx = pd_idx + 1;
2794 if (pd_idx == raid_disks-1) {
2795 (*dd_idx)++; /* Q D D D P */
2797 } else if (*dd_idx >= pd_idx)
2798 (*dd_idx) += 2; /* D D P Q D */
2800 case ALGORITHM_LEFT_SYMMETRIC:
2801 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2802 qd_idx = (pd_idx + 1) % raid_disks;
2803 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2805 case ALGORITHM_RIGHT_SYMMETRIC:
2806 pd_idx = sector_div(stripe2, raid_disks);
2807 qd_idx = (pd_idx + 1) % raid_disks;
2808 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2811 case ALGORITHM_PARITY_0:
2816 case ALGORITHM_PARITY_N:
2817 pd_idx = data_disks;
2818 qd_idx = data_disks + 1;
2821 case ALGORITHM_ROTATING_ZERO_RESTART:
2822 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2823 * of blocks for computing Q is different.
2825 pd_idx = sector_div(stripe2, raid_disks);
2826 qd_idx = pd_idx + 1;
2827 if (pd_idx == raid_disks-1) {
2828 (*dd_idx)++; /* Q D D D P */
2830 } else if (*dd_idx >= pd_idx)
2831 (*dd_idx) += 2; /* D D P Q D */
2835 case ALGORITHM_ROTATING_N_RESTART:
2836 /* Same a left_asymmetric, by first stripe is
2837 * D D D P Q rather than
2841 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2842 qd_idx = pd_idx + 1;
2843 if (pd_idx == raid_disks-1) {
2844 (*dd_idx)++; /* Q D D D P */
2846 } else if (*dd_idx >= pd_idx)
2847 (*dd_idx) += 2; /* D D P Q D */
2851 case ALGORITHM_ROTATING_N_CONTINUE:
2852 /* Same as left_symmetric but Q is before P */
2853 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2854 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2855 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2859 case ALGORITHM_LEFT_ASYMMETRIC_6:
2860 /* RAID5 left_asymmetric, with Q on last device */
2861 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2862 if (*dd_idx >= pd_idx)
2864 qd_idx = raid_disks - 1;
2867 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2868 pd_idx = sector_div(stripe2, raid_disks-1);
2869 if (*dd_idx >= pd_idx)
2871 qd_idx = raid_disks - 1;
2874 case ALGORITHM_LEFT_SYMMETRIC_6:
2875 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2876 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2877 qd_idx = raid_disks - 1;
2880 case ALGORITHM_RIGHT_SYMMETRIC_6:
2881 pd_idx = sector_div(stripe2, raid_disks-1);
2882 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2883 qd_idx = raid_disks - 1;
2886 case ALGORITHM_PARITY_0_6:
2889 qd_idx = raid_disks - 1;
2899 sh->pd_idx = pd_idx;
2900 sh->qd_idx = qd_idx;
2901 sh->ddf_layout = ddf_layout;
2904 * Finally, compute the new sector number
2906 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2910 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2912 struct r5conf *conf = sh->raid_conf;
2913 int raid_disks = sh->disks;
2914 int data_disks = raid_disks - conf->max_degraded;
2915 sector_t new_sector = sh->sector, check;
2916 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2917 : conf->chunk_sectors;
2918 int algorithm = previous ? conf->prev_algo
2922 sector_t chunk_number;
2923 int dummy1, dd_idx = i;
2925 struct stripe_head sh2;
2927 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2928 stripe = new_sector;
2930 if (i == sh->pd_idx)
2932 switch(conf->level) {
2935 switch (algorithm) {
2936 case ALGORITHM_LEFT_ASYMMETRIC:
2937 case ALGORITHM_RIGHT_ASYMMETRIC:
2941 case ALGORITHM_LEFT_SYMMETRIC:
2942 case ALGORITHM_RIGHT_SYMMETRIC:
2945 i -= (sh->pd_idx + 1);
2947 case ALGORITHM_PARITY_0:
2950 case ALGORITHM_PARITY_N:
2957 if (i == sh->qd_idx)
2958 return 0; /* It is the Q disk */
2959 switch (algorithm) {
2960 case ALGORITHM_LEFT_ASYMMETRIC:
2961 case ALGORITHM_RIGHT_ASYMMETRIC:
2962 case ALGORITHM_ROTATING_ZERO_RESTART:
2963 case ALGORITHM_ROTATING_N_RESTART:
2964 if (sh->pd_idx == raid_disks-1)
2965 i--; /* Q D D D P */
2966 else if (i > sh->pd_idx)
2967 i -= 2; /* D D P Q D */
2969 case ALGORITHM_LEFT_SYMMETRIC:
2970 case ALGORITHM_RIGHT_SYMMETRIC:
2971 if (sh->pd_idx == raid_disks-1)
2972 i--; /* Q D D D P */
2977 i -= (sh->pd_idx + 2);
2980 case ALGORITHM_PARITY_0:
2983 case ALGORITHM_PARITY_N:
2985 case ALGORITHM_ROTATING_N_CONTINUE:
2986 /* Like left_symmetric, but P is before Q */
2987 if (sh->pd_idx == 0)
2988 i--; /* P D D D Q */
2993 i -= (sh->pd_idx + 1);
2996 case ALGORITHM_LEFT_ASYMMETRIC_6:
2997 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3001 case ALGORITHM_LEFT_SYMMETRIC_6:
3002 case ALGORITHM_RIGHT_SYMMETRIC_6:
3004 i += data_disks + 1;
3005 i -= (sh->pd_idx + 1);
3007 case ALGORITHM_PARITY_0_6:
3016 chunk_number = stripe * data_disks + i;
3017 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3019 check = raid5_compute_sector(conf, r_sector,
3020 previous, &dummy1, &sh2);
3021 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3022 || sh2.qd_idx != sh->qd_idx) {
3023 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3024 mdname(conf->mddev));
3031 * There are cases where we want handle_stripe_dirtying() and
3032 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3034 * This function checks whether we want to delay the towrite. Specifically,
3035 * we delay the towrite when:
3037 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3038 * stripe has data in journal (for other devices).
3040 * In this case, when reading data for the non-overwrite dev, it is
3041 * necessary to handle complex rmw of write back cache (prexor with
3042 * orig_page, and xor with page). To keep read path simple, we would
3043 * like to flush data in journal to RAID disks first, so complex rmw
3044 * is handled in the write patch (handle_stripe_dirtying).
3046 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3048 * It is important to be able to flush all stripes in raid5-cache.
3049 * Therefore, we need reserve some space on the journal device for
3050 * these flushes. If flush operation includes pending writes to the
3051 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3052 * for the flush out. If we exclude these pending writes from flush
3053 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3054 * Therefore, excluding pending writes in these cases enables more
3055 * efficient use of the journal device.
3057 * Note: To make sure the stripe makes progress, we only delay
3058 * towrite for stripes with data already in journal (injournal > 0).
3059 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3060 * no_space_stripes list.
3062 * 3. during journal failure
3063 * In journal failure, we try to flush all cached data to raid disks
3064 * based on data in stripe cache. The array is read-only to upper
3065 * layers, so we would skip all pending writes.
3068 static inline bool delay_towrite(struct r5conf *conf,
3070 struct stripe_head_state *s)
3073 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3074 !test_bit(R5_Insync, &dev->flags) && s->injournal)
3077 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3081 if (s->log_failed && s->injournal)
3087 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3088 int rcw, int expand)
3090 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3091 struct r5conf *conf = sh->raid_conf;
3092 int level = conf->level;
3096 * In some cases, handle_stripe_dirtying initially decided to
3097 * run rmw and allocates extra page for prexor. However, rcw is
3098 * cheaper later on. We need to free the extra page now,
3099 * because we won't be able to do that in ops_complete_prexor().
3101 r5c_release_extra_page(sh);
3103 for (i = disks; i--; ) {
3104 struct r5dev *dev = &sh->dev[i];
3106 if (dev->towrite && !delay_towrite(conf, dev, s)) {
3107 set_bit(R5_LOCKED, &dev->flags);
3108 set_bit(R5_Wantdrain, &dev->flags);
3110 clear_bit(R5_UPTODATE, &dev->flags);
3112 } else if (test_bit(R5_InJournal, &dev->flags)) {
3113 set_bit(R5_LOCKED, &dev->flags);
3117 /* if we are not expanding this is a proper write request, and
3118 * there will be bios with new data to be drained into the
3123 /* False alarm, nothing to do */
3125 sh->reconstruct_state = reconstruct_state_drain_run;
3126 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3128 sh->reconstruct_state = reconstruct_state_run;
3130 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3132 if (s->locked + conf->max_degraded == disks)
3133 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3134 atomic_inc(&conf->pending_full_writes);
3136 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3137 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3138 BUG_ON(level == 6 &&
3139 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3140 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3142 for (i = disks; i--; ) {
3143 struct r5dev *dev = &sh->dev[i];
3144 if (i == pd_idx || i == qd_idx)
3148 (test_bit(R5_UPTODATE, &dev->flags) ||
3149 test_bit(R5_Wantcompute, &dev->flags))) {
3150 set_bit(R5_Wantdrain, &dev->flags);
3151 set_bit(R5_LOCKED, &dev->flags);
3152 clear_bit(R5_UPTODATE, &dev->flags);
3154 } else if (test_bit(R5_InJournal, &dev->flags)) {
3155 set_bit(R5_LOCKED, &dev->flags);
3160 /* False alarm - nothing to do */
3162 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3163 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3164 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3165 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3168 /* keep the parity disk(s) locked while asynchronous operations
3171 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3172 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3176 int qd_idx = sh->qd_idx;
3177 struct r5dev *dev = &sh->dev[qd_idx];
3179 set_bit(R5_LOCKED, &dev->flags);
3180 clear_bit(R5_UPTODATE, &dev->flags);
3184 if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3185 test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3186 !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3187 test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3188 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3190 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3191 __func__, (unsigned long long)sh->sector,
3192 s->locked, s->ops_request);
3196 * Each stripe/dev can have one or more bion attached.
3197 * toread/towrite point to the first in a chain.
3198 * The bi_next chain must be in order.
3200 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3201 int forwrite, int previous)
3204 struct r5conf *conf = sh->raid_conf;
3207 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3208 (unsigned long long)bi->bi_iter.bi_sector,
3209 (unsigned long long)sh->sector);
3211 spin_lock_irq(&sh->stripe_lock);
3212 sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3213 /* Don't allow new IO added to stripes in batch list */
3217 bip = &sh->dev[dd_idx].towrite;
3221 bip = &sh->dev[dd_idx].toread;
3222 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3223 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3225 bip = & (*bip)->bi_next;
3227 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3230 if (forwrite && raid5_has_ppl(conf)) {
3232 * With PPL only writes to consecutive data chunks within a
3233 * stripe are allowed because for a single stripe_head we can
3234 * only have one PPL entry at a time, which describes one data
3235 * range. Not really an overlap, but wait_for_overlap can be
3236 * used to handle this.
3244 for (i = 0; i < sh->disks; i++) {
3245 if (i != sh->pd_idx &&
3246 (i == dd_idx || sh->dev[i].towrite)) {
3247 sector = sh->dev[i].sector;
3248 if (count == 0 || sector < first)
3256 if (first + conf->chunk_sectors * (count - 1) != last)
3260 if (!forwrite || previous)
3261 clear_bit(STRIPE_BATCH_READY, &sh->state);
3263 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3267 bio_inc_remaining(bi);
3268 md_write_inc(conf->mddev, bi);
3271 /* check if page is covered */
3272 sector_t sector = sh->dev[dd_idx].sector;
3273 for (bi=sh->dev[dd_idx].towrite;
3274 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3275 bi && bi->bi_iter.bi_sector <= sector;
3276 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3277 if (bio_end_sector(bi) >= sector)
3278 sector = bio_end_sector(bi);
3280 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3281 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3282 sh->overwrite_disks++;
3285 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3286 (unsigned long long)(*bip)->bi_iter.bi_sector,
3287 (unsigned long long)sh->sector, dd_idx);
3289 if (conf->mddev->bitmap && firstwrite) {
3290 /* Cannot hold spinlock over bitmap_startwrite,
3291 * but must ensure this isn't added to a batch until
3292 * we have added to the bitmap and set bm_seq.
3293 * So set STRIPE_BITMAP_PENDING to prevent
3295 * If multiple add_stripe_bio() calls race here they
3296 * much all set STRIPE_BITMAP_PENDING. So only the first one
3297 * to complete "bitmap_startwrite" gets to set
3298 * STRIPE_BIT_DELAY. This is important as once a stripe
3299 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3302 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3303 spin_unlock_irq(&sh->stripe_lock);
3304 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3306 spin_lock_irq(&sh->stripe_lock);
3307 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3308 if (!sh->batch_head) {
3309 sh->bm_seq = conf->seq_flush+1;
3310 set_bit(STRIPE_BIT_DELAY, &sh->state);
3313 spin_unlock_irq(&sh->stripe_lock);
3315 if (stripe_can_batch(sh))
3316 stripe_add_to_batch_list(conf, sh);
3320 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3321 spin_unlock_irq(&sh->stripe_lock);
3325 static void end_reshape(struct r5conf *conf);
3327 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3328 struct stripe_head *sh)
3330 int sectors_per_chunk =
3331 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3333 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3334 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3336 raid5_compute_sector(conf,
3337 stripe * (disks - conf->max_degraded)
3338 *sectors_per_chunk + chunk_offset,
3344 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3345 struct stripe_head_state *s, int disks)
3348 BUG_ON(sh->batch_head);
3349 for (i = disks; i--; ) {
3353 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3354 struct md_rdev *rdev;
3356 rdev = rcu_dereference(conf->disks[i].rdev);
3357 if (rdev && test_bit(In_sync, &rdev->flags) &&
3358 !test_bit(Faulty, &rdev->flags))
3359 atomic_inc(&rdev->nr_pending);
3364 if (!rdev_set_badblocks(
3368 md_error(conf->mddev, rdev);
3369 rdev_dec_pending(rdev, conf->mddev);
3372 spin_lock_irq(&sh->stripe_lock);
3373 /* fail all writes first */
3374 bi = sh->dev[i].towrite;
3375 sh->dev[i].towrite = NULL;
3376 sh->overwrite_disks = 0;
3377 spin_unlock_irq(&sh->stripe_lock);
3381 log_stripe_write_finished(sh);
3383 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3384 wake_up(&conf->wait_for_overlap);
3386 while (bi && bi->bi_iter.bi_sector <
3387 sh->dev[i].sector + STRIPE_SECTORS) {
3388 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3390 md_write_end(conf->mddev);
3395 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3396 STRIPE_SECTORS, 0, 0);
3398 /* and fail all 'written' */
3399 bi = sh->dev[i].written;
3400 sh->dev[i].written = NULL;
3401 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3402 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3403 sh->dev[i].page = sh->dev[i].orig_page;
3406 if (bi) bitmap_end = 1;
3407 while (bi && bi->bi_iter.bi_sector <
3408 sh->dev[i].sector + STRIPE_SECTORS) {
3409 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3411 md_write_end(conf->mddev);
3416 /* fail any reads if this device is non-operational and
3417 * the data has not reached the cache yet.
3419 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3420 s->failed > conf->max_degraded &&
3421 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3422 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3423 spin_lock_irq(&sh->stripe_lock);
3424 bi = sh->dev[i].toread;
3425 sh->dev[i].toread = NULL;
3426 spin_unlock_irq(&sh->stripe_lock);
3427 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3428 wake_up(&conf->wait_for_overlap);
3431 while (bi && bi->bi_iter.bi_sector <
3432 sh->dev[i].sector + STRIPE_SECTORS) {
3433 struct bio *nextbi =
3434 r5_next_bio(bi, sh->dev[i].sector);
3441 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3442 STRIPE_SECTORS, 0, 0);
3443 /* If we were in the middle of a write the parity block might
3444 * still be locked - so just clear all R5_LOCKED flags
3446 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3451 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3452 if (atomic_dec_and_test(&conf->pending_full_writes))
3453 md_wakeup_thread(conf->mddev->thread);
3457 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3458 struct stripe_head_state *s)
3463 BUG_ON(sh->batch_head);
3464 clear_bit(STRIPE_SYNCING, &sh->state);
3465 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3466 wake_up(&conf->wait_for_overlap);
3469 /* There is nothing more to do for sync/check/repair.
3470 * Don't even need to abort as that is handled elsewhere
3471 * if needed, and not always wanted e.g. if there is a known
3473 * For recover/replace we need to record a bad block on all
3474 * non-sync devices, or abort the recovery
3476 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3477 /* During recovery devices cannot be removed, so
3478 * locking and refcounting of rdevs is not needed
3481 for (i = 0; i < conf->raid_disks; i++) {
3482 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3484 && !test_bit(Faulty, &rdev->flags)
3485 && !test_bit(In_sync, &rdev->flags)
3486 && !rdev_set_badblocks(rdev, sh->sector,
3489 rdev = rcu_dereference(conf->disks[i].replacement);
3491 && !test_bit(Faulty, &rdev->flags)
3492 && !test_bit(In_sync, &rdev->flags)
3493 && !rdev_set_badblocks(rdev, sh->sector,
3499 conf->recovery_disabled =
3500 conf->mddev->recovery_disabled;
3502 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3505 static int want_replace(struct stripe_head *sh, int disk_idx)
3507 struct md_rdev *rdev;
3511 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3513 && !test_bit(Faulty, &rdev->flags)
3514 && !test_bit(In_sync, &rdev->flags)
3515 && (rdev->recovery_offset <= sh->sector
3516 || rdev->mddev->recovery_cp <= sh->sector))
3522 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3523 int disk_idx, int disks)
3525 struct r5dev *dev = &sh->dev[disk_idx];
3526 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3527 &sh->dev[s->failed_num[1]] };
3531 if (test_bit(R5_LOCKED, &dev->flags) ||
3532 test_bit(R5_UPTODATE, &dev->flags))
3533 /* No point reading this as we already have it or have
3534 * decided to get it.
3539 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3540 /* We need this block to directly satisfy a request */
3543 if (s->syncing || s->expanding ||
3544 (s->replacing && want_replace(sh, disk_idx)))
3545 /* When syncing, or expanding we read everything.
3546 * When replacing, we need the replaced block.
3550 if ((s->failed >= 1 && fdev[0]->toread) ||
3551 (s->failed >= 2 && fdev[1]->toread))
3552 /* If we want to read from a failed device, then
3553 * we need to actually read every other device.
3557 /* Sometimes neither read-modify-write nor reconstruct-write
3558 * cycles can work. In those cases we read every block we
3559 * can. Then the parity-update is certain to have enough to
3561 * This can only be a problem when we need to write something,
3562 * and some device has failed. If either of those tests
3563 * fail we need look no further.
3565 if (!s->failed || !s->to_write)
3568 if (test_bit(R5_Insync, &dev->flags) &&
3569 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3570 /* Pre-reads at not permitted until after short delay
3571 * to gather multiple requests. However if this
3572 * device is no Insync, the block could only be computed
3573 * and there is no need to delay that.
3577 for (i = 0; i < s->failed && i < 2; i++) {
3578 if (fdev[i]->towrite &&
3579 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3580 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3581 /* If we have a partial write to a failed
3582 * device, then we will need to reconstruct
3583 * the content of that device, so all other
3584 * devices must be read.
3589 /* If we are forced to do a reconstruct-write, either because
3590 * the current RAID6 implementation only supports that, or
3591 * because parity cannot be trusted and we are currently
3592 * recovering it, there is extra need to be careful.
3593 * If one of the devices that we would need to read, because
3594 * it is not being overwritten (and maybe not written at all)
3595 * is missing/faulty, then we need to read everything we can.
3597 if (sh->raid_conf->level != 6 &&
3598 sh->sector < sh->raid_conf->mddev->recovery_cp)
3599 /* reconstruct-write isn't being forced */
3601 for (i = 0; i < s->failed && i < 2; i++) {
3602 if (s->failed_num[i] != sh->pd_idx &&
3603 s->failed_num[i] != sh->qd_idx &&
3604 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3605 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3612 /* fetch_block - checks the given member device to see if its data needs
3613 * to be read or computed to satisfy a request.
3615 * Returns 1 when no more member devices need to be checked, otherwise returns
3616 * 0 to tell the loop in handle_stripe_fill to continue
3618 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3619 int disk_idx, int disks)
3621 struct r5dev *dev = &sh->dev[disk_idx];
3623 /* is the data in this block needed, and can we get it? */
3624 if (need_this_block(sh, s, disk_idx, disks)) {
3625 /* we would like to get this block, possibly by computing it,
3626 * otherwise read it if the backing disk is insync
3628 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3629 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3630 BUG_ON(sh->batch_head);
3633 * In the raid6 case if the only non-uptodate disk is P
3634 * then we already trusted P to compute the other failed
3635 * drives. It is safe to compute rather than re-read P.
3636 * In other cases we only compute blocks from failed
3637 * devices, otherwise check/repair might fail to detect
3638 * a real inconsistency.
3641 if ((s->uptodate == disks - 1) &&
3642 ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3643 (s->failed && (disk_idx == s->failed_num[0] ||
3644 disk_idx == s->failed_num[1])))) {
3645 /* have disk failed, and we're requested to fetch it;
3648 pr_debug("Computing stripe %llu block %d\n",
3649 (unsigned long long)sh->sector, disk_idx);
3650 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3651 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3652 set_bit(R5_Wantcompute, &dev->flags);
3653 sh->ops.target = disk_idx;
3654 sh->ops.target2 = -1; /* no 2nd target */
3656 /* Careful: from this point on 'uptodate' is in the eye
3657 * of raid_run_ops which services 'compute' operations
3658 * before writes. R5_Wantcompute flags a block that will
3659 * be R5_UPTODATE by the time it is needed for a
3660 * subsequent operation.
3664 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3665 /* Computing 2-failure is *very* expensive; only
3666 * do it if failed >= 2
3669 for (other = disks; other--; ) {
3670 if (other == disk_idx)
3672 if (!test_bit(R5_UPTODATE,
3673 &sh->dev[other].flags))
3677 pr_debug("Computing stripe %llu blocks %d,%d\n",
3678 (unsigned long long)sh->sector,
3680 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3681 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3682 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3683 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3684 sh->ops.target = disk_idx;
3685 sh->ops.target2 = other;
3689 } else if (test_bit(R5_Insync, &dev->flags)) {
3690 set_bit(R5_LOCKED, &dev->flags);
3691 set_bit(R5_Wantread, &dev->flags);
3693 pr_debug("Reading block %d (sync=%d)\n",
3694 disk_idx, s->syncing);
3702 * handle_stripe_fill - read or compute data to satisfy pending requests.
3704 static void handle_stripe_fill(struct stripe_head *sh,
3705 struct stripe_head_state *s,
3710 /* look for blocks to read/compute, skip this if a compute
3711 * is already in flight, or if the stripe contents are in the
3712 * midst of changing due to a write
3714 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3715 !sh->reconstruct_state) {
3718 * For degraded stripe with data in journal, do not handle
3719 * read requests yet, instead, flush the stripe to raid
3720 * disks first, this avoids handling complex rmw of write
3721 * back cache (prexor with orig_page, and then xor with
3722 * page) in the read path
3724 if (s->injournal && s->failed) {
3725 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3726 r5c_make_stripe_write_out(sh);
3730 for (i = disks; i--; )
3731 if (fetch_block(sh, s, i, disks))
3735 set_bit(STRIPE_HANDLE, &sh->state);
3738 static void break_stripe_batch_list(struct stripe_head *head_sh,
3739 unsigned long handle_flags);
3740 /* handle_stripe_clean_event
3741 * any written block on an uptodate or failed drive can be returned.
3742 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3743 * never LOCKED, so we don't need to test 'failed' directly.
3745 static void handle_stripe_clean_event(struct r5conf *conf,
3746 struct stripe_head *sh, int disks)
3750 int discard_pending = 0;
3751 struct stripe_head *head_sh = sh;
3752 bool do_endio = false;
3754 for (i = disks; i--; )
3755 if (sh->dev[i].written) {
3757 if (!test_bit(R5_LOCKED, &dev->flags) &&
3758 (test_bit(R5_UPTODATE, &dev->flags) ||
3759 test_bit(R5_Discard, &dev->flags) ||
3760 test_bit(R5_SkipCopy, &dev->flags))) {
3761 /* We can return any write requests */
3762 struct bio *wbi, *wbi2;
3763 pr_debug("Return write for disc %d\n", i);
3764 if (test_and_clear_bit(R5_Discard, &dev->flags))
3765 clear_bit(R5_UPTODATE, &dev->flags);
3766 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3767 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3772 dev->page = dev->orig_page;
3774 dev->written = NULL;
3775 while (wbi && wbi->bi_iter.bi_sector <
3776 dev->sector + STRIPE_SECTORS) {
3777 wbi2 = r5_next_bio(wbi, dev->sector);
3778 md_write_end(conf->mddev);
3782 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3784 !test_bit(STRIPE_DEGRADED, &sh->state),
3786 if (head_sh->batch_head) {
3787 sh = list_first_entry(&sh->batch_list,
3790 if (sh != head_sh) {
3797 } else if (test_bit(R5_Discard, &dev->flags))
3798 discard_pending = 1;
3801 log_stripe_write_finished(sh);
3803 if (!discard_pending &&
3804 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3806 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3807 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3808 if (sh->qd_idx >= 0) {
3809 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3810 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3812 /* now that discard is done we can proceed with any sync */
3813 clear_bit(STRIPE_DISCARD, &sh->state);
3815 * SCSI discard will change some bio fields and the stripe has
3816 * no updated data, so remove it from hash list and the stripe
3817 * will be reinitialized
3820 hash = sh->hash_lock_index;
3821 spin_lock_irq(conf->hash_locks + hash);
3823 spin_unlock_irq(conf->hash_locks + hash);
3824 if (head_sh->batch_head) {
3825 sh = list_first_entry(&sh->batch_list,
3826 struct stripe_head, batch_list);
3832 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3833 set_bit(STRIPE_HANDLE, &sh->state);
3837 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3838 if (atomic_dec_and_test(&conf->pending_full_writes))
3839 md_wakeup_thread(conf->mddev->thread);
3841 if (head_sh->batch_head && do_endio)
3842 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3846 * For RMW in write back cache, we need extra page in prexor to store the
3847 * old data. This page is stored in dev->orig_page.
3849 * This function checks whether we have data for prexor. The exact logic
3851 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3853 static inline bool uptodate_for_rmw(struct r5dev *dev)
3855 return (test_bit(R5_UPTODATE, &dev->flags)) &&
3856 (!test_bit(R5_InJournal, &dev->flags) ||
3857 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3860 static int handle_stripe_dirtying(struct r5conf *conf,
3861 struct stripe_head *sh,
3862 struct stripe_head_state *s,
3865 int rmw = 0, rcw = 0, i;
3866 sector_t recovery_cp = conf->mddev->recovery_cp;
3868 /* Check whether resync is now happening or should start.
3869 * If yes, then the array is dirty (after unclean shutdown or
3870 * initial creation), so parity in some stripes might be inconsistent.
3871 * In this case, we need to always do reconstruct-write, to ensure
3872 * that in case of drive failure or read-error correction, we
3873 * generate correct data from the parity.
3875 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3876 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3878 /* Calculate the real rcw later - for now make it
3879 * look like rcw is cheaper
3882 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3883 conf->rmw_level, (unsigned long long)recovery_cp,
3884 (unsigned long long)sh->sector);
3885 } else for (i = disks; i--; ) {
3886 /* would I have to read this buffer for read_modify_write */
3887 struct r5dev *dev = &sh->dev[i];
3888 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3889 i == sh->pd_idx || i == sh->qd_idx ||
3890 test_bit(R5_InJournal, &dev->flags)) &&
3891 !test_bit(R5_LOCKED, &dev->flags) &&
3892 !(uptodate_for_rmw(dev) ||
3893 test_bit(R5_Wantcompute, &dev->flags))) {
3894 if (test_bit(R5_Insync, &dev->flags))
3897 rmw += 2*disks; /* cannot read it */
3899 /* Would I have to read this buffer for reconstruct_write */
3900 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3901 i != sh->pd_idx && i != sh->qd_idx &&
3902 !test_bit(R5_LOCKED, &dev->flags) &&
3903 !(test_bit(R5_UPTODATE, &dev->flags) ||
3904 test_bit(R5_Wantcompute, &dev->flags))) {
3905 if (test_bit(R5_Insync, &dev->flags))
3912 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3913 (unsigned long long)sh->sector, sh->state, rmw, rcw);
3914 set_bit(STRIPE_HANDLE, &sh->state);
3915 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3916 /* prefer read-modify-write, but need to get some data */
3917 if (conf->mddev->queue)
3918 blk_add_trace_msg(conf->mddev->queue,
3919 "raid5 rmw %llu %d",
3920 (unsigned long long)sh->sector, rmw);
3921 for (i = disks; i--; ) {
3922 struct r5dev *dev = &sh->dev[i];
3923 if (test_bit(R5_InJournal, &dev->flags) &&
3924 dev->page == dev->orig_page &&
3925 !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3926 /* alloc page for prexor */
3927 struct page *p = alloc_page(GFP_NOIO);
3935 * alloc_page() failed, try use
3936 * disk_info->extra_page
3938 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3939 &conf->cache_state)) {
3940 r5c_use_extra_page(sh);
3944 /* extra_page in use, add to delayed_list */
3945 set_bit(STRIPE_DELAYED, &sh->state);
3946 s->waiting_extra_page = 1;
3951 for (i = disks; i--; ) {
3952 struct r5dev *dev = &sh->dev[i];
3953 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3954 i == sh->pd_idx || i == sh->qd_idx ||
3955 test_bit(R5_InJournal, &dev->flags)) &&
3956 !test_bit(R5_LOCKED, &dev->flags) &&
3957 !(uptodate_for_rmw(dev) ||
3958 test_bit(R5_Wantcompute, &dev->flags)) &&
3959 test_bit(R5_Insync, &dev->flags)) {
3960 if (test_bit(STRIPE_PREREAD_ACTIVE,
3962 pr_debug("Read_old block %d for r-m-w\n",
3964 set_bit(R5_LOCKED, &dev->flags);
3965 set_bit(R5_Wantread, &dev->flags);
3968 set_bit(STRIPE_DELAYED, &sh->state);
3969 set_bit(STRIPE_HANDLE, &sh->state);
3974 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3975 /* want reconstruct write, but need to get some data */
3978 for (i = disks; i--; ) {
3979 struct r5dev *dev = &sh->dev[i];
3980 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3981 i != sh->pd_idx && i != sh->qd_idx &&
3982 !test_bit(R5_LOCKED, &dev->flags) &&
3983 !(test_bit(R5_UPTODATE, &dev->flags) ||
3984 test_bit(R5_Wantcompute, &dev->flags))) {
3986 if (test_bit(R5_Insync, &dev->flags) &&
3987 test_bit(STRIPE_PREREAD_ACTIVE,
3989 pr_debug("Read_old block "
3990 "%d for Reconstruct\n", i);
3991 set_bit(R5_LOCKED, &dev->flags);
3992 set_bit(R5_Wantread, &dev->flags);
3996 set_bit(STRIPE_DELAYED, &sh->state);
3997 set_bit(STRIPE_HANDLE, &sh->state);
4001 if (rcw && conf->mddev->queue)
4002 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4003 (unsigned long long)sh->sector,
4004 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4007 if (rcw > disks && rmw > disks &&
4008 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4009 set_bit(STRIPE_DELAYED, &sh->state);
4011 /* now if nothing is locked, and if we have enough data,
4012 * we can start a write request
4014 /* since handle_stripe can be called at any time we need to handle the
4015 * case where a compute block operation has been submitted and then a
4016 * subsequent call wants to start a write request. raid_run_ops only
4017 * handles the case where compute block and reconstruct are requested
4018 * simultaneously. If this is not the case then new writes need to be
4019 * held off until the compute completes.
4021 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4022 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4023 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4024 schedule_reconstruction(sh, s, rcw == 0, 0);
4028 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4029 struct stripe_head_state *s, int disks)
4031 struct r5dev *dev = NULL;
4033 BUG_ON(sh->batch_head);
4034 set_bit(STRIPE_HANDLE, &sh->state);
4036 switch (sh->check_state) {
4037 case check_state_idle:
4038 /* start a new check operation if there are no failures */
4039 if (s->failed == 0) {
4040 BUG_ON(s->uptodate != disks);
4041 sh->check_state = check_state_run;
4042 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4043 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4047 dev = &sh->dev[s->failed_num[0]];
4049 case check_state_compute_result:
4050 sh->check_state = check_state_idle;
4052 dev = &sh->dev[sh->pd_idx];
4054 /* check that a write has not made the stripe insync */
4055 if (test_bit(STRIPE_INSYNC, &sh->state))
4058 /* either failed parity check, or recovery is happening */
4059 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4060 BUG_ON(s->uptodate != disks);
4062 set_bit(R5_LOCKED, &dev->flags);
4064 set_bit(R5_Wantwrite, &dev->flags);
4066 clear_bit(STRIPE_DEGRADED, &sh->state);
4067 set_bit(STRIPE_INSYNC, &sh->state);
4069 case check_state_run:
4070 break; /* we will be called again upon completion */
4071 case check_state_check_result:
4072 sh->check_state = check_state_idle;
4074 /* if a failure occurred during the check operation, leave
4075 * STRIPE_INSYNC not set and let the stripe be handled again
4080 /* handle a successful check operation, if parity is correct
4081 * we are done. Otherwise update the mismatch count and repair
4082 * parity if !MD_RECOVERY_CHECK
4084 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4085 /* parity is correct (on disc,
4086 * not in buffer any more)
4088 set_bit(STRIPE_INSYNC, &sh->state);
4090 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4091 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4092 /* don't try to repair!! */
4093 set_bit(STRIPE_INSYNC, &sh->state);
4094 pr_warn_ratelimited("%s: mismatch sector in range "
4095 "%llu-%llu\n", mdname(conf->mddev),
4096 (unsigned long long) sh->sector,
4097 (unsigned long long) sh->sector +
4100 sh->check_state = check_state_compute_run;
4101 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4102 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4103 set_bit(R5_Wantcompute,
4104 &sh->dev[sh->pd_idx].flags);
4105 sh->ops.target = sh->pd_idx;
4106 sh->ops.target2 = -1;
4111 case check_state_compute_run:
4114 pr_err("%s: unknown check_state: %d sector: %llu\n",
4115 __func__, sh->check_state,
4116 (unsigned long long) sh->sector);
4121 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4122 struct stripe_head_state *s,
4125 int pd_idx = sh->pd_idx;
4126 int qd_idx = sh->qd_idx;
4129 BUG_ON(sh->batch_head);
4130 set_bit(STRIPE_HANDLE, &sh->state);
4132 BUG_ON(s->failed > 2);
4134 /* Want to check and possibly repair P and Q.
4135 * However there could be one 'failed' device, in which
4136 * case we can only check one of them, possibly using the
4137 * other to generate missing data
4140 switch (sh->check_state) {
4141 case check_state_idle:
4142 /* start a new check operation if there are < 2 failures */
4143 if (s->failed == s->q_failed) {
4144 /* The only possible failed device holds Q, so it
4145 * makes sense to check P (If anything else were failed,
4146 * we would have used P to recreate it).
4148 sh->check_state = check_state_run;
4150 if (!s->q_failed && s->failed < 2) {
4151 /* Q is not failed, and we didn't use it to generate
4152 * anything, so it makes sense to check it
4154 if (sh->check_state == check_state_run)
4155 sh->check_state = check_state_run_pq;
4157 sh->check_state = check_state_run_q;
4160 /* discard potentially stale zero_sum_result */
4161 sh->ops.zero_sum_result = 0;
4163 if (sh->check_state == check_state_run) {
4164 /* async_xor_zero_sum destroys the contents of P */
4165 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4168 if (sh->check_state >= check_state_run &&
4169 sh->check_state <= check_state_run_pq) {
4170 /* async_syndrome_zero_sum preserves P and Q, so
4171 * no need to mark them !uptodate here
4173 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4177 /* we have 2-disk failure */
4178 BUG_ON(s->failed != 2);
4180 case check_state_compute_result:
4181 sh->check_state = check_state_idle;
4183 /* check that a write has not made the stripe insync */
4184 if (test_bit(STRIPE_INSYNC, &sh->state))
4187 /* now write out any block on a failed drive,
4188 * or P or Q if they were recomputed
4190 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
4191 if (s->failed == 2) {
4192 dev = &sh->dev[s->failed_num[1]];
4194 set_bit(R5_LOCKED, &dev->flags);
4195 set_bit(R5_Wantwrite, &dev->flags);
4197 if (s->failed >= 1) {
4198 dev = &sh->dev[s->failed_num[0]];
4200 set_bit(R5_LOCKED, &dev->flags);
4201 set_bit(R5_Wantwrite, &dev->flags);
4203 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4204 dev = &sh->dev[pd_idx];
4206 set_bit(R5_LOCKED, &dev->flags);
4207 set_bit(R5_Wantwrite, &dev->flags);
4209 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4210 dev = &sh->dev[qd_idx];
4212 set_bit(R5_LOCKED, &dev->flags);
4213 set_bit(R5_Wantwrite, &dev->flags);
4215 clear_bit(STRIPE_DEGRADED, &sh->state);
4217 set_bit(STRIPE_INSYNC, &sh->state);
4219 case check_state_run:
4220 case check_state_run_q:
4221 case check_state_run_pq:
4222 break; /* we will be called again upon completion */
4223 case check_state_check_result:
4224 sh->check_state = check_state_idle;
4226 /* handle a successful check operation, if parity is correct
4227 * we are done. Otherwise update the mismatch count and repair
4228 * parity if !MD_RECOVERY_CHECK
4230 if (sh->ops.zero_sum_result == 0) {
4231 /* both parities are correct */
4233 set_bit(STRIPE_INSYNC, &sh->state);
4235 /* in contrast to the raid5 case we can validate
4236 * parity, but still have a failure to write
4239 sh->check_state = check_state_compute_result;
4240 /* Returning at this point means that we may go
4241 * off and bring p and/or q uptodate again so
4242 * we make sure to check zero_sum_result again
4243 * to verify if p or q need writeback
4247 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4248 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4249 /* don't try to repair!! */
4250 set_bit(STRIPE_INSYNC, &sh->state);
4251 pr_warn_ratelimited("%s: mismatch sector in range "
4252 "%llu-%llu\n", mdname(conf->mddev),
4253 (unsigned long long) sh->sector,
4254 (unsigned long long) sh->sector +
4257 int *target = &sh->ops.target;
4259 sh->ops.target = -1;
4260 sh->ops.target2 = -1;
4261 sh->check_state = check_state_compute_run;
4262 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4263 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4264 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4265 set_bit(R5_Wantcompute,
4266 &sh->dev[pd_idx].flags);
4268 target = &sh->ops.target2;
4271 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4272 set_bit(R5_Wantcompute,
4273 &sh->dev[qd_idx].flags);
4280 case check_state_compute_run:
4283 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4284 __func__, sh->check_state,
4285 (unsigned long long) sh->sector);
4290 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4294 /* We have read all the blocks in this stripe and now we need to
4295 * copy some of them into a target stripe for expand.
4297 struct dma_async_tx_descriptor *tx = NULL;
4298 BUG_ON(sh->batch_head);
4299 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4300 for (i = 0; i < sh->disks; i++)
4301 if (i != sh->pd_idx && i != sh->qd_idx) {
4303 struct stripe_head *sh2;
4304 struct async_submit_ctl submit;
4306 sector_t bn = raid5_compute_blocknr(sh, i, 1);
4307 sector_t s = raid5_compute_sector(conf, bn, 0,
4309 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4311 /* so far only the early blocks of this stripe
4312 * have been requested. When later blocks
4313 * get requested, we will try again
4316 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4317 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4318 /* must have already done this block */
4319 raid5_release_stripe(sh2);
4323 /* place all the copies on one channel */
4324 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4325 tx = async_memcpy(sh2->dev[dd_idx].page,
4326 sh->dev[i].page, 0, 0, STRIPE_SIZE,
4329 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4330 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4331 for (j = 0; j < conf->raid_disks; j++)
4332 if (j != sh2->pd_idx &&
4334 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4336 if (j == conf->raid_disks) {
4337 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4338 set_bit(STRIPE_HANDLE, &sh2->state);
4340 raid5_release_stripe(sh2);
4343 /* done submitting copies, wait for them to complete */
4344 async_tx_quiesce(&tx);
4348 * handle_stripe - do things to a stripe.
4350 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4351 * state of various bits to see what needs to be done.
4353 * return some read requests which now have data
4354 * return some write requests which are safely on storage
4355 * schedule a read on some buffers
4356 * schedule a write of some buffers
4357 * return confirmation of parity correctness
4361 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4363 struct r5conf *conf = sh->raid_conf;
4364 int disks = sh->disks;
4367 int do_recovery = 0;
4369 memset(s, 0, sizeof(*s));
4371 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4372 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4373 s->failed_num[0] = -1;
4374 s->failed_num[1] = -1;
4375 s->log_failed = r5l_log_disk_error(conf);
4377 /* Now to look around and see what can be done */
4379 for (i=disks; i--; ) {
4380 struct md_rdev *rdev;
4387 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4389 dev->toread, dev->towrite, dev->written);
4390 /* maybe we can reply to a read
4392 * new wantfill requests are only permitted while
4393 * ops_complete_biofill is guaranteed to be inactive
4395 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4396 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4397 set_bit(R5_Wantfill, &dev->flags);
4399 /* now count some things */
4400 if (test_bit(R5_LOCKED, &dev->flags))
4402 if (test_bit(R5_UPTODATE, &dev->flags))
4404 if (test_bit(R5_Wantcompute, &dev->flags)) {
4406 BUG_ON(s->compute > 2);
4409 if (test_bit(R5_Wantfill, &dev->flags))
4411 else if (dev->toread)
4415 if (!test_bit(R5_OVERWRITE, &dev->flags))
4420 /* Prefer to use the replacement for reads, but only
4421 * if it is recovered enough and has no bad blocks.
4423 rdev = rcu_dereference(conf->disks[i].replacement);
4424 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4425 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4426 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4427 &first_bad, &bad_sectors))
4428 set_bit(R5_ReadRepl, &dev->flags);
4430 if (rdev && !test_bit(Faulty, &rdev->flags))
4431 set_bit(R5_NeedReplace, &dev->flags);
4433 clear_bit(R5_NeedReplace, &dev->flags);
4434 rdev = rcu_dereference(conf->disks[i].rdev);
4435 clear_bit(R5_ReadRepl, &dev->flags);
4437 if (rdev && test_bit(Faulty, &rdev->flags))
4440 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4441 &first_bad, &bad_sectors);
4442 if (s->blocked_rdev == NULL
4443 && (test_bit(Blocked, &rdev->flags)
4446 set_bit(BlockedBadBlocks,
4448 s->blocked_rdev = rdev;
4449 atomic_inc(&rdev->nr_pending);
4452 clear_bit(R5_Insync, &dev->flags);
4456 /* also not in-sync */
4457 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4458 test_bit(R5_UPTODATE, &dev->flags)) {
4459 /* treat as in-sync, but with a read error
4460 * which we can now try to correct
4462 set_bit(R5_Insync, &dev->flags);
4463 set_bit(R5_ReadError, &dev->flags);
4465 } else if (test_bit(In_sync, &rdev->flags))
4466 set_bit(R5_Insync, &dev->flags);
4467 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4468 /* in sync if before recovery_offset */
4469 set_bit(R5_Insync, &dev->flags);
4470 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4471 test_bit(R5_Expanded, &dev->flags))
4472 /* If we've reshaped into here, we assume it is Insync.
4473 * We will shortly update recovery_offset to make
4476 set_bit(R5_Insync, &dev->flags);
4478 if (test_bit(R5_WriteError, &dev->flags)) {
4479 /* This flag does not apply to '.replacement'
4480 * only to .rdev, so make sure to check that*/
4481 struct md_rdev *rdev2 = rcu_dereference(
4482 conf->disks[i].rdev);
4484 clear_bit(R5_Insync, &dev->flags);
4485 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4486 s->handle_bad_blocks = 1;
4487 atomic_inc(&rdev2->nr_pending);
4489 clear_bit(R5_WriteError, &dev->flags);
4491 if (test_bit(R5_MadeGood, &dev->flags)) {
4492 /* This flag does not apply to '.replacement'
4493 * only to .rdev, so make sure to check that*/
4494 struct md_rdev *rdev2 = rcu_dereference(
4495 conf->disks[i].rdev);
4496 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4497 s->handle_bad_blocks = 1;
4498 atomic_inc(&rdev2->nr_pending);
4500 clear_bit(R5_MadeGood, &dev->flags);
4502 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4503 struct md_rdev *rdev2 = rcu_dereference(
4504 conf->disks[i].replacement);
4505 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4506 s->handle_bad_blocks = 1;
4507 atomic_inc(&rdev2->nr_pending);
4509 clear_bit(R5_MadeGoodRepl, &dev->flags);
4511 if (!test_bit(R5_Insync, &dev->flags)) {
4512 /* The ReadError flag will just be confusing now */
4513 clear_bit(R5_ReadError, &dev->flags);
4514 clear_bit(R5_ReWrite, &dev->flags);
4516 if (test_bit(R5_ReadError, &dev->flags))
4517 clear_bit(R5_Insync, &dev->flags);
4518 if (!test_bit(R5_Insync, &dev->flags)) {
4520 s->failed_num[s->failed] = i;
4522 if (rdev && !test_bit(Faulty, &rdev->flags))
4525 rdev = rcu_dereference(
4526 conf->disks[i].replacement);
4527 if (rdev && !test_bit(Faulty, &rdev->flags))
4532 if (test_bit(R5_InJournal, &dev->flags))
4534 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4537 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4538 /* If there is a failed device being replaced,
4539 * we must be recovering.
4540 * else if we are after recovery_cp, we must be syncing
4541 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4542 * else we can only be replacing
4543 * sync and recovery both need to read all devices, and so
4544 * use the same flag.
4547 sh->sector >= conf->mddev->recovery_cp ||
4548 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4556 static int clear_batch_ready(struct stripe_head *sh)
4558 /* Return '1' if this is a member of batch, or
4559 * '0' if it is a lone stripe or a head which can now be
4562 struct stripe_head *tmp;
4563 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4564 return (sh->batch_head && sh->batch_head != sh);
4565 spin_lock(&sh->stripe_lock);
4566 if (!sh->batch_head) {
4567 spin_unlock(&sh->stripe_lock);
4572 * this stripe could be added to a batch list before we check
4573 * BATCH_READY, skips it
4575 if (sh->batch_head != sh) {
4576 spin_unlock(&sh->stripe_lock);
4579 spin_lock(&sh->batch_lock);
4580 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4581 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4582 spin_unlock(&sh->batch_lock);
4583 spin_unlock(&sh->stripe_lock);
4586 * BATCH_READY is cleared, no new stripes can be added.
4587 * batch_list can be accessed without lock
4592 static void break_stripe_batch_list(struct stripe_head *head_sh,
4593 unsigned long handle_flags)
4595 struct stripe_head *sh, *next;
4599 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4601 list_del_init(&sh->batch_list);
4603 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4604 (1 << STRIPE_SYNCING) |
4605 (1 << STRIPE_REPLACED) |
4606 (1 << STRIPE_DELAYED) |
4607 (1 << STRIPE_BIT_DELAY) |
4608 (1 << STRIPE_FULL_WRITE) |
4609 (1 << STRIPE_BIOFILL_RUN) |
4610 (1 << STRIPE_COMPUTE_RUN) |
4611 (1 << STRIPE_OPS_REQ_PENDING) |
4612 (1 << STRIPE_DISCARD) |
4613 (1 << STRIPE_BATCH_READY) |
4614 (1 << STRIPE_BATCH_ERR) |
4615 (1 << STRIPE_BITMAP_PENDING)),
4616 "stripe state: %lx\n", sh->state);
4617 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4618 (1 << STRIPE_REPLACED)),
4619 "head stripe state: %lx\n", head_sh->state);
4621 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4622 (1 << STRIPE_PREREAD_ACTIVE) |
4623 (1 << STRIPE_DEGRADED) |
4624 (1 << STRIPE_ON_UNPLUG_LIST)),
4625 head_sh->state & (1 << STRIPE_INSYNC));
4627 sh->check_state = head_sh->check_state;
4628 sh->reconstruct_state = head_sh->reconstruct_state;
4629 spin_lock_irq(&sh->stripe_lock);
4630 sh->batch_head = NULL;
4631 spin_unlock_irq(&sh->stripe_lock);
4632 for (i = 0; i < sh->disks; i++) {
4633 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4635 sh->dev[i].flags = head_sh->dev[i].flags &
4636 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4638 if (handle_flags == 0 ||
4639 sh->state & handle_flags)
4640 set_bit(STRIPE_HANDLE, &sh->state);
4641 raid5_release_stripe(sh);
4643 spin_lock_irq(&head_sh->stripe_lock);
4644 head_sh->batch_head = NULL;
4645 spin_unlock_irq(&head_sh->stripe_lock);
4646 for (i = 0; i < head_sh->disks; i++)
4647 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4649 if (head_sh->state & handle_flags)
4650 set_bit(STRIPE_HANDLE, &head_sh->state);
4653 wake_up(&head_sh->raid_conf->wait_for_overlap);
4656 static void handle_stripe(struct stripe_head *sh)
4658 struct stripe_head_state s;
4659 struct r5conf *conf = sh->raid_conf;
4662 int disks = sh->disks;
4663 struct r5dev *pdev, *qdev;
4665 clear_bit(STRIPE_HANDLE, &sh->state);
4666 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4667 /* already being handled, ensure it gets handled
4668 * again when current action finishes */
4669 set_bit(STRIPE_HANDLE, &sh->state);
4673 if (clear_batch_ready(sh) ) {
4674 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4678 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4679 break_stripe_batch_list(sh, 0);
4681 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4682 spin_lock(&sh->stripe_lock);
4684 * Cannot process 'sync' concurrently with 'discard'.
4685 * Flush data in r5cache before 'sync'.
4687 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4688 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4689 !test_bit(STRIPE_DISCARD, &sh->state) &&
4690 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4691 set_bit(STRIPE_SYNCING, &sh->state);
4692 clear_bit(STRIPE_INSYNC, &sh->state);
4693 clear_bit(STRIPE_REPLACED, &sh->state);
4695 spin_unlock(&sh->stripe_lock);
4697 clear_bit(STRIPE_DELAYED, &sh->state);
4699 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4700 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4701 (unsigned long long)sh->sector, sh->state,
4702 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4703 sh->check_state, sh->reconstruct_state);
4705 analyse_stripe(sh, &s);
4707 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4710 if (s.handle_bad_blocks ||
4711 test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4712 set_bit(STRIPE_HANDLE, &sh->state);
4716 if (unlikely(s.blocked_rdev)) {
4717 if (s.syncing || s.expanding || s.expanded ||
4718 s.replacing || s.to_write || s.written) {
4719 set_bit(STRIPE_HANDLE, &sh->state);
4722 /* There is nothing for the blocked_rdev to block */
4723 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4724 s.blocked_rdev = NULL;
4727 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4728 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4729 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4732 pr_debug("locked=%d uptodate=%d to_read=%d"
4733 " to_write=%d failed=%d failed_num=%d,%d\n",
4734 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4735 s.failed_num[0], s.failed_num[1]);
4737 * check if the array has lost more than max_degraded devices and,
4738 * if so, some requests might need to be failed.
4740 * When journal device failed (log_failed), we will only process
4741 * the stripe if there is data need write to raid disks
4743 if (s.failed > conf->max_degraded ||
4744 (s.log_failed && s.injournal == 0)) {
4745 sh->check_state = 0;
4746 sh->reconstruct_state = 0;
4747 break_stripe_batch_list(sh, 0);
4748 if (s.to_read+s.to_write+s.written)
4749 handle_failed_stripe(conf, sh, &s, disks);
4750 if (s.syncing + s.replacing)
4751 handle_failed_sync(conf, sh, &s);
4754 /* Now we check to see if any write operations have recently
4758 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4760 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4761 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4762 sh->reconstruct_state = reconstruct_state_idle;
4764 /* All the 'written' buffers and the parity block are ready to
4765 * be written back to disk
4767 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4768 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4769 BUG_ON(sh->qd_idx >= 0 &&
4770 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4771 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4772 for (i = disks; i--; ) {
4773 struct r5dev *dev = &sh->dev[i];
4774 if (test_bit(R5_LOCKED, &dev->flags) &&
4775 (i == sh->pd_idx || i == sh->qd_idx ||
4776 dev->written || test_bit(R5_InJournal,
4778 pr_debug("Writing block %d\n", i);
4779 set_bit(R5_Wantwrite, &dev->flags);
4784 if (!test_bit(R5_Insync, &dev->flags) ||
4785 ((i == sh->pd_idx || i == sh->qd_idx) &&
4787 set_bit(STRIPE_INSYNC, &sh->state);
4790 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4791 s.dec_preread_active = 1;
4795 * might be able to return some write requests if the parity blocks
4796 * are safe, or on a failed drive
4798 pdev = &sh->dev[sh->pd_idx];
4799 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4800 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4801 qdev = &sh->dev[sh->qd_idx];
4802 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4803 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4807 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4808 && !test_bit(R5_LOCKED, &pdev->flags)
4809 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4810 test_bit(R5_Discard, &pdev->flags))))) &&
4811 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4812 && !test_bit(R5_LOCKED, &qdev->flags)
4813 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4814 test_bit(R5_Discard, &qdev->flags))))))
4815 handle_stripe_clean_event(conf, sh, disks);
4818 r5c_handle_cached_data_endio(conf, sh, disks);
4819 log_stripe_write_finished(sh);
4821 /* Now we might consider reading some blocks, either to check/generate
4822 * parity, or to satisfy requests
4823 * or to load a block that is being partially written.
4825 if (s.to_read || s.non_overwrite
4826 || (conf->level == 6 && s.to_write && s.failed)
4827 || (s.syncing && (s.uptodate + s.compute < disks))
4830 handle_stripe_fill(sh, &s, disks);
4833 * When the stripe finishes full journal write cycle (write to journal
4834 * and raid disk), this is the clean up procedure so it is ready for
4837 r5c_finish_stripe_write_out(conf, sh, &s);
4840 * Now to consider new write requests, cache write back and what else,
4841 * if anything should be read. We do not handle new writes when:
4842 * 1/ A 'write' operation (copy+xor) is already in flight.
4843 * 2/ A 'check' operation is in flight, as it may clobber the parity
4845 * 3/ A r5c cache log write is in flight.
4848 if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4849 if (!r5c_is_writeback(conf->log)) {
4851 handle_stripe_dirtying(conf, sh, &s, disks);
4852 } else { /* write back cache */
4855 /* First, try handle writes in caching phase */
4857 ret = r5c_try_caching_write(conf, sh, &s,
4860 * If caching phase failed: ret == -EAGAIN
4862 * stripe under reclaim: !caching && injournal
4864 * fall back to handle_stripe_dirtying()
4866 if (ret == -EAGAIN ||
4867 /* stripe under reclaim: !caching && injournal */
4868 (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4870 ret = handle_stripe_dirtying(conf, sh, &s,
4878 /* maybe we need to check and possibly fix the parity for this stripe
4879 * Any reads will already have been scheduled, so we just see if enough
4880 * data is available. The parity check is held off while parity
4881 * dependent operations are in flight.
4883 if (sh->check_state ||
4884 (s.syncing && s.locked == 0 &&
4885 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4886 !test_bit(STRIPE_INSYNC, &sh->state))) {
4887 if (conf->level == 6)
4888 handle_parity_checks6(conf, sh, &s, disks);
4890 handle_parity_checks5(conf, sh, &s, disks);
4893 if ((s.replacing || s.syncing) && s.locked == 0
4894 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4895 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4896 /* Write out to replacement devices where possible */
4897 for (i = 0; i < conf->raid_disks; i++)
4898 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4899 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4900 set_bit(R5_WantReplace, &sh->dev[i].flags);
4901 set_bit(R5_LOCKED, &sh->dev[i].flags);
4905 set_bit(STRIPE_INSYNC, &sh->state);
4906 set_bit(STRIPE_REPLACED, &sh->state);
4908 if ((s.syncing || s.replacing) && s.locked == 0 &&
4909 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4910 test_bit(STRIPE_INSYNC, &sh->state)) {
4911 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4912 clear_bit(STRIPE_SYNCING, &sh->state);
4913 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4914 wake_up(&conf->wait_for_overlap);
4917 /* If the failed drives are just a ReadError, then we might need
4918 * to progress the repair/check process
4920 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4921 for (i = 0; i < s.failed; i++) {
4922 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4923 if (test_bit(R5_ReadError, &dev->flags)
4924 && !test_bit(R5_LOCKED, &dev->flags)
4925 && test_bit(R5_UPTODATE, &dev->flags)
4927 if (!test_bit(R5_ReWrite, &dev->flags)) {
4928 set_bit(R5_Wantwrite, &dev->flags);
4929 set_bit(R5_ReWrite, &dev->flags);
4930 set_bit(R5_LOCKED, &dev->flags);
4933 /* let's read it back */
4934 set_bit(R5_Wantread, &dev->flags);
4935 set_bit(R5_LOCKED, &dev->flags);
4941 /* Finish reconstruct operations initiated by the expansion process */
4942 if (sh->reconstruct_state == reconstruct_state_result) {
4943 struct stripe_head *sh_src
4944 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4945 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4946 /* sh cannot be written until sh_src has been read.
4947 * so arrange for sh to be delayed a little
4949 set_bit(STRIPE_DELAYED, &sh->state);
4950 set_bit(STRIPE_HANDLE, &sh->state);
4951 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4953 atomic_inc(&conf->preread_active_stripes);
4954 raid5_release_stripe(sh_src);
4958 raid5_release_stripe(sh_src);
4960 sh->reconstruct_state = reconstruct_state_idle;
4961 clear_bit(STRIPE_EXPANDING, &sh->state);
4962 for (i = conf->raid_disks; i--; ) {
4963 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4964 set_bit(R5_LOCKED, &sh->dev[i].flags);
4969 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4970 !sh->reconstruct_state) {
4971 /* Need to write out all blocks after computing parity */
4972 sh->disks = conf->raid_disks;
4973 stripe_set_idx(sh->sector, conf, 0, sh);
4974 schedule_reconstruction(sh, &s, 1, 1);
4975 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4976 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4977 atomic_dec(&conf->reshape_stripes);
4978 wake_up(&conf->wait_for_overlap);
4979 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4982 if (s.expanding && s.locked == 0 &&
4983 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4984 handle_stripe_expansion(conf, sh);
4987 /* wait for this device to become unblocked */
4988 if (unlikely(s.blocked_rdev)) {
4989 if (conf->mddev->external)
4990 md_wait_for_blocked_rdev(s.blocked_rdev,
4993 /* Internal metadata will immediately
4994 * be written by raid5d, so we don't
4995 * need to wait here.
4997 rdev_dec_pending(s.blocked_rdev,
5001 if (s.handle_bad_blocks)
5002 for (i = disks; i--; ) {
5003 struct md_rdev *rdev;
5004 struct r5dev *dev = &sh->dev[i];
5005 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5006 /* We own a safe reference to the rdev */
5007 rdev = conf->disks[i].rdev;
5008 if (!rdev_set_badblocks(rdev, sh->sector,
5010 md_error(conf->mddev, rdev);
5011 rdev_dec_pending(rdev, conf->mddev);
5013 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5014 rdev = conf->disks[i].rdev;
5015 rdev_clear_badblocks(rdev, sh->sector,
5017 rdev_dec_pending(rdev, conf->mddev);
5019 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5020 rdev = conf->disks[i].replacement;
5022 /* rdev have been moved down */
5023 rdev = conf->disks[i].rdev;
5024 rdev_clear_badblocks(rdev, sh->sector,
5026 rdev_dec_pending(rdev, conf->mddev);
5031 raid_run_ops(sh, s.ops_request);
5035 if (s.dec_preread_active) {
5036 /* We delay this until after ops_run_io so that if make_request
5037 * is waiting on a flush, it won't continue until the writes
5038 * have actually been submitted.
5040 atomic_dec(&conf->preread_active_stripes);
5041 if (atomic_read(&conf->preread_active_stripes) <
5043 md_wakeup_thread(conf->mddev->thread);
5046 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5049 static void raid5_activate_delayed(struct r5conf *conf)
5051 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5052 while (!list_empty(&conf->delayed_list)) {
5053 struct list_head *l = conf->delayed_list.next;
5054 struct stripe_head *sh;
5055 sh = list_entry(l, struct stripe_head, lru);
5057 clear_bit(STRIPE_DELAYED, &sh->state);
5058 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5059 atomic_inc(&conf->preread_active_stripes);
5060 list_add_tail(&sh->lru, &conf->hold_list);
5061 raid5_wakeup_stripe_thread(sh);
5066 static void activate_bit_delay(struct r5conf *conf,
5067 struct list_head *temp_inactive_list)
5069 /* device_lock is held */
5070 struct list_head head;
5071 list_add(&head, &conf->bitmap_list);
5072 list_del_init(&conf->bitmap_list);
5073 while (!list_empty(&head)) {
5074 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5076 list_del_init(&sh->lru);
5077 atomic_inc(&sh->count);
5078 hash = sh->hash_lock_index;
5079 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5083 static int raid5_congested(struct mddev *mddev, int bits)
5085 struct r5conf *conf = mddev->private;
5087 /* No difference between reads and writes. Just check
5088 * how busy the stripe_cache is
5091 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5094 /* Also checks whether there is pressure on r5cache log space */
5095 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5099 if (atomic_read(&conf->empty_inactive_list_nr))
5105 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5107 struct r5conf *conf = mddev->private;
5108 sector_t sector = bio->bi_iter.bi_sector;
5109 unsigned int chunk_sectors;
5110 unsigned int bio_sectors = bio_sectors(bio);
5112 WARN_ON_ONCE(bio->bi_partno);
5114 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5115 return chunk_sectors >=
5116 ((sector & (chunk_sectors - 1)) + bio_sectors);
5120 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5121 * later sampled by raid5d.
5123 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5125 unsigned long flags;
5127 spin_lock_irqsave(&conf->device_lock, flags);
5129 bi->bi_next = conf->retry_read_aligned_list;
5130 conf->retry_read_aligned_list = bi;
5132 spin_unlock_irqrestore(&conf->device_lock, flags);
5133 md_wakeup_thread(conf->mddev->thread);
5136 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5137 unsigned int *offset)
5141 bi = conf->retry_read_aligned;
5143 *offset = conf->retry_read_offset;
5144 conf->retry_read_aligned = NULL;
5147 bi = conf->retry_read_aligned_list;
5149 conf->retry_read_aligned_list = bi->bi_next;
5158 * The "raid5_align_endio" should check if the read succeeded and if it
5159 * did, call bio_endio on the original bio (having bio_put the new bio
5161 * If the read failed..
5163 static void raid5_align_endio(struct bio *bi)
5165 struct bio* raid_bi = bi->bi_private;
5166 struct mddev *mddev;
5167 struct r5conf *conf;
5168 struct md_rdev *rdev;
5169 blk_status_t error = bi->bi_status;
5173 rdev = (void*)raid_bi->bi_next;
5174 raid_bi->bi_next = NULL;
5175 mddev = rdev->mddev;
5176 conf = mddev->private;
5178 rdev_dec_pending(rdev, conf->mddev);
5182 if (atomic_dec_and_test(&conf->active_aligned_reads))
5183 wake_up(&conf->wait_for_quiescent);
5187 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5189 add_bio_to_retry(raid_bi, conf);
5192 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5194 struct r5conf *conf = mddev->private;
5196 struct bio* align_bi;
5197 struct md_rdev *rdev;
5198 sector_t end_sector;
5200 if (!in_chunk_boundary(mddev, raid_bio)) {
5201 pr_debug("%s: non aligned\n", __func__);
5205 * use bio_clone_fast to make a copy of the bio
5207 align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5211 * set bi_end_io to a new function, and set bi_private to the
5214 align_bi->bi_end_io = raid5_align_endio;
5215 align_bi->bi_private = raid_bio;
5219 align_bi->bi_iter.bi_sector =
5220 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5223 end_sector = bio_end_sector(align_bi);
5225 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5226 if (!rdev || test_bit(Faulty, &rdev->flags) ||
5227 rdev->recovery_offset < end_sector) {
5228 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5230 (test_bit(Faulty, &rdev->flags) ||
5231 !(test_bit(In_sync, &rdev->flags) ||
5232 rdev->recovery_offset >= end_sector)))
5236 if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5246 atomic_inc(&rdev->nr_pending);
5248 raid_bio->bi_next = (void*)rdev;
5249 bio_set_dev(align_bi, rdev->bdev);
5250 bio_clear_flag(align_bi, BIO_SEG_VALID);
5252 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5253 bio_sectors(align_bi),
5254 &first_bad, &bad_sectors)) {
5256 rdev_dec_pending(rdev, mddev);
5260 /* No reshape active, so we can trust rdev->data_offset */
5261 align_bi->bi_iter.bi_sector += rdev->data_offset;
5263 spin_lock_irq(&conf->device_lock);
5264 wait_event_lock_irq(conf->wait_for_quiescent,
5267 atomic_inc(&conf->active_aligned_reads);
5268 spin_unlock_irq(&conf->device_lock);
5271 trace_block_bio_remap(align_bi->bi_disk->queue,
5272 align_bi, disk_devt(mddev->gendisk),
5273 raid_bio->bi_iter.bi_sector);
5274 generic_make_request(align_bi);
5283 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5286 sector_t sector = raid_bio->bi_iter.bi_sector;
5287 unsigned chunk_sects = mddev->chunk_sectors;
5288 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5290 if (sectors < bio_sectors(raid_bio)) {
5291 struct r5conf *conf = mddev->private;
5292 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5293 bio_chain(split, raid_bio);
5294 generic_make_request(raid_bio);
5298 if (!raid5_read_one_chunk(mddev, raid_bio))
5304 /* __get_priority_stripe - get the next stripe to process
5306 * Full stripe writes are allowed to pass preread active stripes up until
5307 * the bypass_threshold is exceeded. In general the bypass_count
5308 * increments when the handle_list is handled before the hold_list; however, it
5309 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5310 * stripe with in flight i/o. The bypass_count will be reset when the
5311 * head of the hold_list has changed, i.e. the head was promoted to the
5314 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5316 struct stripe_head *sh, *tmp;
5317 struct list_head *handle_list = NULL;
5318 struct r5worker_group *wg;
5319 bool second_try = !r5c_is_writeback(conf->log) &&
5320 !r5l_log_disk_error(conf);
5321 bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5322 r5l_log_disk_error(conf);
5327 if (conf->worker_cnt_per_group == 0) {
5328 handle_list = try_loprio ? &conf->loprio_list :
5330 } else if (group != ANY_GROUP) {
5331 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5332 &conf->worker_groups[group].handle_list;
5333 wg = &conf->worker_groups[group];
5336 for (i = 0; i < conf->group_cnt; i++) {
5337 handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5338 &conf->worker_groups[i].handle_list;
5339 wg = &conf->worker_groups[i];
5340 if (!list_empty(handle_list))
5345 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5347 list_empty(handle_list) ? "empty" : "busy",
5348 list_empty(&conf->hold_list) ? "empty" : "busy",
5349 atomic_read(&conf->pending_full_writes), conf->bypass_count);
5351 if (!list_empty(handle_list)) {
5352 sh = list_entry(handle_list->next, typeof(*sh), lru);
5354 if (list_empty(&conf->hold_list))
5355 conf->bypass_count = 0;
5356 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5357 if (conf->hold_list.next == conf->last_hold)
5358 conf->bypass_count++;
5360 conf->last_hold = conf->hold_list.next;
5361 conf->bypass_count -= conf->bypass_threshold;
5362 if (conf->bypass_count < 0)
5363 conf->bypass_count = 0;
5366 } else if (!list_empty(&conf->hold_list) &&
5367 ((conf->bypass_threshold &&
5368 conf->bypass_count > conf->bypass_threshold) ||
5369 atomic_read(&conf->pending_full_writes) == 0)) {
5371 list_for_each_entry(tmp, &conf->hold_list, lru) {
5372 if (conf->worker_cnt_per_group == 0 ||
5373 group == ANY_GROUP ||
5374 !cpu_online(tmp->cpu) ||
5375 cpu_to_group(tmp->cpu) == group) {
5382 conf->bypass_count -= conf->bypass_threshold;
5383 if (conf->bypass_count < 0)
5384 conf->bypass_count = 0;
5393 try_loprio = !try_loprio;
5401 list_del_init(&sh->lru);
5402 BUG_ON(atomic_inc_return(&sh->count) != 1);
5406 struct raid5_plug_cb {
5407 struct blk_plug_cb cb;
5408 struct list_head list;
5409 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5412 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5414 struct raid5_plug_cb *cb = container_of(
5415 blk_cb, struct raid5_plug_cb, cb);
5416 struct stripe_head *sh;
5417 struct mddev *mddev = cb->cb.data;
5418 struct r5conf *conf = mddev->private;
5422 if (cb->list.next && !list_empty(&cb->list)) {
5423 spin_lock_irq(&conf->device_lock);
5424 while (!list_empty(&cb->list)) {
5425 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5426 list_del_init(&sh->lru);
5428 * avoid race release_stripe_plug() sees
5429 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5430 * is still in our list
5432 smp_mb__before_atomic();
5433 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5435 * STRIPE_ON_RELEASE_LIST could be set here. In that
5436 * case, the count is always > 1 here
5438 hash = sh->hash_lock_index;
5439 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5442 spin_unlock_irq(&conf->device_lock);
5444 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5445 NR_STRIPE_HASH_LOCKS);
5447 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5451 static void release_stripe_plug(struct mddev *mddev,
5452 struct stripe_head *sh)
5454 struct blk_plug_cb *blk_cb = blk_check_plugged(
5455 raid5_unplug, mddev,
5456 sizeof(struct raid5_plug_cb));
5457 struct raid5_plug_cb *cb;
5460 raid5_release_stripe(sh);
5464 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5466 if (cb->list.next == NULL) {
5468 INIT_LIST_HEAD(&cb->list);
5469 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5470 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5473 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5474 list_add_tail(&sh->lru, &cb->list);
5476 raid5_release_stripe(sh);
5479 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5481 struct r5conf *conf = mddev->private;
5482 sector_t logical_sector, last_sector;
5483 struct stripe_head *sh;
5486 if (mddev->reshape_position != MaxSector)
5487 /* Skip discard while reshape is happening */
5490 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5491 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5495 stripe_sectors = conf->chunk_sectors *
5496 (conf->raid_disks - conf->max_degraded);
5497 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5499 sector_div(last_sector, stripe_sectors);
5501 logical_sector *= conf->chunk_sectors;
5502 last_sector *= conf->chunk_sectors;
5504 for (; logical_sector < last_sector;
5505 logical_sector += STRIPE_SECTORS) {
5509 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5510 prepare_to_wait(&conf->wait_for_overlap, &w,
5511 TASK_UNINTERRUPTIBLE);
5512 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5513 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5514 raid5_release_stripe(sh);
5518 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5519 spin_lock_irq(&sh->stripe_lock);
5520 for (d = 0; d < conf->raid_disks; d++) {
5521 if (d == sh->pd_idx || d == sh->qd_idx)
5523 if (sh->dev[d].towrite || sh->dev[d].toread) {
5524 set_bit(R5_Overlap, &sh->dev[d].flags);
5525 spin_unlock_irq(&sh->stripe_lock);
5526 raid5_release_stripe(sh);
5531 set_bit(STRIPE_DISCARD, &sh->state);
5532 finish_wait(&conf->wait_for_overlap, &w);
5533 sh->overwrite_disks = 0;
5534 for (d = 0; d < conf->raid_disks; d++) {
5535 if (d == sh->pd_idx || d == sh->qd_idx)
5537 sh->dev[d].towrite = bi;
5538 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5539 bio_inc_remaining(bi);
5540 md_write_inc(mddev, bi);
5541 sh->overwrite_disks++;
5543 spin_unlock_irq(&sh->stripe_lock);
5544 if (conf->mddev->bitmap) {
5546 d < conf->raid_disks - conf->max_degraded;
5548 md_bitmap_startwrite(mddev->bitmap,
5552 sh->bm_seq = conf->seq_flush + 1;
5553 set_bit(STRIPE_BIT_DELAY, &sh->state);
5556 set_bit(STRIPE_HANDLE, &sh->state);
5557 clear_bit(STRIPE_DELAYED, &sh->state);
5558 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5559 atomic_inc(&conf->preread_active_stripes);
5560 release_stripe_plug(mddev, sh);
5566 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5568 struct r5conf *conf = mddev->private;
5570 sector_t new_sector;
5571 sector_t logical_sector, last_sector;
5572 struct stripe_head *sh;
5573 const int rw = bio_data_dir(bi);
5576 bool do_flush = false;
5578 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5579 int ret = log_handle_flush_request(conf, bi);
5583 if (ret == -ENODEV) {
5584 md_flush_request(mddev, bi);
5587 /* ret == -EAGAIN, fallback */
5589 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5590 * we need to flush journal device
5592 do_flush = bi->bi_opf & REQ_PREFLUSH;
5595 if (!md_write_start(mddev, bi))
5598 * If array is degraded, better not do chunk aligned read because
5599 * later we might have to read it again in order to reconstruct
5600 * data on failed drives.
5602 if (rw == READ && mddev->degraded == 0 &&
5603 mddev->reshape_position == MaxSector) {
5604 bi = chunk_aligned_read(mddev, bi);
5609 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5610 make_discard_request(mddev, bi);
5611 md_write_end(mddev);
5615 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5616 last_sector = bio_end_sector(bi);
5619 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5620 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5626 seq = read_seqcount_begin(&conf->gen_lock);
5629 prepare_to_wait(&conf->wait_for_overlap, &w,
5630 TASK_UNINTERRUPTIBLE);
5631 if (unlikely(conf->reshape_progress != MaxSector)) {
5632 /* spinlock is needed as reshape_progress may be
5633 * 64bit on a 32bit platform, and so it might be
5634 * possible to see a half-updated value
5635 * Of course reshape_progress could change after
5636 * the lock is dropped, so once we get a reference
5637 * to the stripe that we think it is, we will have
5640 spin_lock_irq(&conf->device_lock);
5641 if (mddev->reshape_backwards
5642 ? logical_sector < conf->reshape_progress
5643 : logical_sector >= conf->reshape_progress) {
5646 if (mddev->reshape_backwards
5647 ? logical_sector < conf->reshape_safe
5648 : logical_sector >= conf->reshape_safe) {
5649 spin_unlock_irq(&conf->device_lock);
5655 spin_unlock_irq(&conf->device_lock);
5658 new_sector = raid5_compute_sector(conf, logical_sector,
5661 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5662 (unsigned long long)new_sector,
5663 (unsigned long long)logical_sector);
5665 sh = raid5_get_active_stripe(conf, new_sector, previous,
5666 (bi->bi_opf & REQ_RAHEAD), 0);
5668 if (unlikely(previous)) {
5669 /* expansion might have moved on while waiting for a
5670 * stripe, so we must do the range check again.
5671 * Expansion could still move past after this
5672 * test, but as we are holding a reference to
5673 * 'sh', we know that if that happens,
5674 * STRIPE_EXPANDING will get set and the expansion
5675 * won't proceed until we finish with the stripe.
5678 spin_lock_irq(&conf->device_lock);
5679 if (mddev->reshape_backwards
5680 ? logical_sector >= conf->reshape_progress
5681 : logical_sector < conf->reshape_progress)
5682 /* mismatch, need to try again */
5684 spin_unlock_irq(&conf->device_lock);
5686 raid5_release_stripe(sh);
5692 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5693 /* Might have got the wrong stripe_head
5696 raid5_release_stripe(sh);
5700 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5701 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5702 /* Stripe is busy expanding or
5703 * add failed due to overlap. Flush everything
5706 md_wakeup_thread(mddev->thread);
5707 raid5_release_stripe(sh);
5713 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5714 /* we only need flush for one stripe */
5718 set_bit(STRIPE_HANDLE, &sh->state);
5719 clear_bit(STRIPE_DELAYED, &sh->state);
5720 if ((!sh->batch_head || sh == sh->batch_head) &&
5721 (bi->bi_opf & REQ_SYNC) &&
5722 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5723 atomic_inc(&conf->preread_active_stripes);
5724 release_stripe_plug(mddev, sh);
5726 /* cannot get stripe for read-ahead, just give-up */
5727 bi->bi_status = BLK_STS_IOERR;
5731 finish_wait(&conf->wait_for_overlap, &w);
5734 md_write_end(mddev);
5739 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5741 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5743 /* reshaping is quite different to recovery/resync so it is
5744 * handled quite separately ... here.
5746 * On each call to sync_request, we gather one chunk worth of
5747 * destination stripes and flag them as expanding.
5748 * Then we find all the source stripes and request reads.
5749 * As the reads complete, handle_stripe will copy the data
5750 * into the destination stripe and release that stripe.
5752 struct r5conf *conf = mddev->private;
5753 struct stripe_head *sh;
5754 struct md_rdev *rdev;
5755 sector_t first_sector, last_sector;
5756 int raid_disks = conf->previous_raid_disks;
5757 int data_disks = raid_disks - conf->max_degraded;
5758 int new_data_disks = conf->raid_disks - conf->max_degraded;
5761 sector_t writepos, readpos, safepos;
5762 sector_t stripe_addr;
5763 int reshape_sectors;
5764 struct list_head stripes;
5767 if (sector_nr == 0) {
5768 /* If restarting in the middle, skip the initial sectors */
5769 if (mddev->reshape_backwards &&
5770 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5771 sector_nr = raid5_size(mddev, 0, 0)
5772 - conf->reshape_progress;
5773 } else if (mddev->reshape_backwards &&
5774 conf->reshape_progress == MaxSector) {
5775 /* shouldn't happen, but just in case, finish up.*/
5776 sector_nr = MaxSector;
5777 } else if (!mddev->reshape_backwards &&
5778 conf->reshape_progress > 0)
5779 sector_nr = conf->reshape_progress;
5780 sector_div(sector_nr, new_data_disks);
5782 mddev->curr_resync_completed = sector_nr;
5783 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5790 /* We need to process a full chunk at a time.
5791 * If old and new chunk sizes differ, we need to process the
5795 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5797 /* We update the metadata at least every 10 seconds, or when
5798 * the data about to be copied would over-write the source of
5799 * the data at the front of the range. i.e. one new_stripe
5800 * along from reshape_progress new_maps to after where
5801 * reshape_safe old_maps to
5803 writepos = conf->reshape_progress;
5804 sector_div(writepos, new_data_disks);
5805 readpos = conf->reshape_progress;
5806 sector_div(readpos, data_disks);
5807 safepos = conf->reshape_safe;
5808 sector_div(safepos, data_disks);
5809 if (mddev->reshape_backwards) {
5810 BUG_ON(writepos < reshape_sectors);
5811 writepos -= reshape_sectors;
5812 readpos += reshape_sectors;
5813 safepos += reshape_sectors;
5815 writepos += reshape_sectors;
5816 /* readpos and safepos are worst-case calculations.
5817 * A negative number is overly pessimistic, and causes
5818 * obvious problems for unsigned storage. So clip to 0.
5820 readpos -= min_t(sector_t, reshape_sectors, readpos);
5821 safepos -= min_t(sector_t, reshape_sectors, safepos);
5824 /* Having calculated the 'writepos' possibly use it
5825 * to set 'stripe_addr' which is where we will write to.
5827 if (mddev->reshape_backwards) {
5828 BUG_ON(conf->reshape_progress == 0);
5829 stripe_addr = writepos;
5830 BUG_ON((mddev->dev_sectors &
5831 ~((sector_t)reshape_sectors - 1))
5832 - reshape_sectors - stripe_addr
5835 BUG_ON(writepos != sector_nr + reshape_sectors);
5836 stripe_addr = sector_nr;
5839 /* 'writepos' is the most advanced device address we might write.
5840 * 'readpos' is the least advanced device address we might read.
5841 * 'safepos' is the least address recorded in the metadata as having
5843 * If there is a min_offset_diff, these are adjusted either by
5844 * increasing the safepos/readpos if diff is negative, or
5845 * increasing writepos if diff is positive.
5846 * If 'readpos' is then behind 'writepos', there is no way that we can
5847 * ensure safety in the face of a crash - that must be done by userspace
5848 * making a backup of the data. So in that case there is no particular
5849 * rush to update metadata.
5850 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5851 * update the metadata to advance 'safepos' to match 'readpos' so that
5852 * we can be safe in the event of a crash.
5853 * So we insist on updating metadata if safepos is behind writepos and
5854 * readpos is beyond writepos.
5855 * In any case, update the metadata every 10 seconds.
5856 * Maybe that number should be configurable, but I'm not sure it is
5857 * worth it.... maybe it could be a multiple of safemode_delay???
5859 if (conf->min_offset_diff < 0) {
5860 safepos += -conf->min_offset_diff;
5861 readpos += -conf->min_offset_diff;
5863 writepos += conf->min_offset_diff;
5865 if ((mddev->reshape_backwards
5866 ? (safepos > writepos && readpos < writepos)
5867 : (safepos < writepos && readpos > writepos)) ||
5868 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5869 /* Cannot proceed until we've updated the superblock... */
5870 wait_event(conf->wait_for_overlap,
5871 atomic_read(&conf->reshape_stripes)==0
5872 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5873 if (atomic_read(&conf->reshape_stripes) != 0)
5875 mddev->reshape_position = conf->reshape_progress;
5876 mddev->curr_resync_completed = sector_nr;
5877 if (!mddev->reshape_backwards)
5878 /* Can update recovery_offset */
5879 rdev_for_each(rdev, mddev)
5880 if (rdev->raid_disk >= 0 &&
5881 !test_bit(Journal, &rdev->flags) &&
5882 !test_bit(In_sync, &rdev->flags) &&
5883 rdev->recovery_offset < sector_nr)
5884 rdev->recovery_offset = sector_nr;
5886 conf->reshape_checkpoint = jiffies;
5887 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5888 md_wakeup_thread(mddev->thread);
5889 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5890 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5891 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5893 spin_lock_irq(&conf->device_lock);
5894 conf->reshape_safe = mddev->reshape_position;
5895 spin_unlock_irq(&conf->device_lock);
5896 wake_up(&conf->wait_for_overlap);
5897 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5900 INIT_LIST_HEAD(&stripes);
5901 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5903 int skipped_disk = 0;
5904 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5905 set_bit(STRIPE_EXPANDING, &sh->state);
5906 atomic_inc(&conf->reshape_stripes);
5907 /* If any of this stripe is beyond the end of the old
5908 * array, then we need to zero those blocks
5910 for (j=sh->disks; j--;) {
5912 if (j == sh->pd_idx)
5914 if (conf->level == 6 &&
5917 s = raid5_compute_blocknr(sh, j, 0);
5918 if (s < raid5_size(mddev, 0, 0)) {
5922 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5923 set_bit(R5_Expanded, &sh->dev[j].flags);
5924 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5926 if (!skipped_disk) {
5927 set_bit(STRIPE_EXPAND_READY, &sh->state);
5928 set_bit(STRIPE_HANDLE, &sh->state);
5930 list_add(&sh->lru, &stripes);
5932 spin_lock_irq(&conf->device_lock);
5933 if (mddev->reshape_backwards)
5934 conf->reshape_progress -= reshape_sectors * new_data_disks;
5936 conf->reshape_progress += reshape_sectors * new_data_disks;
5937 spin_unlock_irq(&conf->device_lock);
5938 /* Ok, those stripe are ready. We can start scheduling
5939 * reads on the source stripes.
5940 * The source stripes are determined by mapping the first and last
5941 * block on the destination stripes.
5944 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5947 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5948 * new_data_disks - 1),
5950 if (last_sector >= mddev->dev_sectors)
5951 last_sector = mddev->dev_sectors - 1;
5952 while (first_sector <= last_sector) {
5953 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5954 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5955 set_bit(STRIPE_HANDLE, &sh->state);
5956 raid5_release_stripe(sh);
5957 first_sector += STRIPE_SECTORS;
5959 /* Now that the sources are clearly marked, we can release
5960 * the destination stripes
5962 while (!list_empty(&stripes)) {
5963 sh = list_entry(stripes.next, struct stripe_head, lru);
5964 list_del_init(&sh->lru);
5965 raid5_release_stripe(sh);
5967 /* If this takes us to the resync_max point where we have to pause,
5968 * then we need to write out the superblock.
5970 sector_nr += reshape_sectors;
5971 retn = reshape_sectors;
5973 if (mddev->curr_resync_completed > mddev->resync_max ||
5974 (sector_nr - mddev->curr_resync_completed) * 2
5975 >= mddev->resync_max - mddev->curr_resync_completed) {
5976 /* Cannot proceed until we've updated the superblock... */
5977 wait_event(conf->wait_for_overlap,
5978 atomic_read(&conf->reshape_stripes) == 0
5979 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5980 if (atomic_read(&conf->reshape_stripes) != 0)
5982 mddev->reshape_position = conf->reshape_progress;
5983 mddev->curr_resync_completed = sector_nr;
5984 if (!mddev->reshape_backwards)
5985 /* Can update recovery_offset */
5986 rdev_for_each(rdev, mddev)
5987 if (rdev->raid_disk >= 0 &&
5988 !test_bit(Journal, &rdev->flags) &&
5989 !test_bit(In_sync, &rdev->flags) &&
5990 rdev->recovery_offset < sector_nr)
5991 rdev->recovery_offset = sector_nr;
5992 conf->reshape_checkpoint = jiffies;
5993 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5994 md_wakeup_thread(mddev->thread);
5995 wait_event(mddev->sb_wait,
5996 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
5997 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5998 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6000 spin_lock_irq(&conf->device_lock);
6001 conf->reshape_safe = mddev->reshape_position;
6002 spin_unlock_irq(&conf->device_lock);
6003 wake_up(&conf->wait_for_overlap);
6004 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6010 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6013 struct r5conf *conf = mddev->private;
6014 struct stripe_head *sh;
6015 sector_t max_sector = mddev->dev_sectors;
6016 sector_t sync_blocks;
6017 int still_degraded = 0;
6020 if (sector_nr >= max_sector) {
6021 /* just being told to finish up .. nothing much to do */
6023 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6028 if (mddev->curr_resync < max_sector) /* aborted */
6029 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6031 else /* completed sync */
6033 md_bitmap_close_sync(mddev->bitmap);
6038 /* Allow raid5_quiesce to complete */
6039 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6041 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6042 return reshape_request(mddev, sector_nr, skipped);
6044 /* No need to check resync_max as we never do more than one
6045 * stripe, and as resync_max will always be on a chunk boundary,
6046 * if the check in md_do_sync didn't fire, there is no chance
6047 * of overstepping resync_max here
6050 /* if there is too many failed drives and we are trying
6051 * to resync, then assert that we are finished, because there is
6052 * nothing we can do.
6054 if (mddev->degraded >= conf->max_degraded &&
6055 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6056 sector_t rv = mddev->dev_sectors - sector_nr;
6060 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6062 !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6063 sync_blocks >= STRIPE_SECTORS) {
6064 /* we can skip this block, and probably more */
6065 sync_blocks /= STRIPE_SECTORS;
6067 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6070 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6072 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6074 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6075 /* make sure we don't swamp the stripe cache if someone else
6076 * is trying to get access
6078 schedule_timeout_uninterruptible(1);
6080 /* Need to check if array will still be degraded after recovery/resync
6081 * Note in case of > 1 drive failures it's possible we're rebuilding
6082 * one drive while leaving another faulty drive in array.
6085 for (i = 0; i < conf->raid_disks; i++) {
6086 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6088 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6093 md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6095 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6096 set_bit(STRIPE_HANDLE, &sh->state);
6098 raid5_release_stripe(sh);
6100 return STRIPE_SECTORS;
6103 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6104 unsigned int offset)
6106 /* We may not be able to submit a whole bio at once as there
6107 * may not be enough stripe_heads available.
6108 * We cannot pre-allocate enough stripe_heads as we may need
6109 * more than exist in the cache (if we allow ever large chunks).
6110 * So we do one stripe head at a time and record in
6111 * ->bi_hw_segments how many have been done.
6113 * We *know* that this entire raid_bio is in one chunk, so
6114 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6116 struct stripe_head *sh;
6118 sector_t sector, logical_sector, last_sector;
6122 logical_sector = raid_bio->bi_iter.bi_sector &
6123 ~((sector_t)STRIPE_SECTORS-1);
6124 sector = raid5_compute_sector(conf, logical_sector,
6126 last_sector = bio_end_sector(raid_bio);
6128 for (; logical_sector < last_sector;
6129 logical_sector += STRIPE_SECTORS,
6130 sector += STRIPE_SECTORS,
6134 /* already done this stripe */
6137 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6140 /* failed to get a stripe - must wait */
6141 conf->retry_read_aligned = raid_bio;
6142 conf->retry_read_offset = scnt;
6146 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6147 raid5_release_stripe(sh);
6148 conf->retry_read_aligned = raid_bio;
6149 conf->retry_read_offset = scnt;
6153 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6155 raid5_release_stripe(sh);
6159 bio_endio(raid_bio);
6161 if (atomic_dec_and_test(&conf->active_aligned_reads))
6162 wake_up(&conf->wait_for_quiescent);
6166 static int handle_active_stripes(struct r5conf *conf, int group,
6167 struct r5worker *worker,
6168 struct list_head *temp_inactive_list)
6170 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6171 int i, batch_size = 0, hash;
6172 bool release_inactive = false;
6174 while (batch_size < MAX_STRIPE_BATCH &&
6175 (sh = __get_priority_stripe(conf, group)) != NULL)
6176 batch[batch_size++] = sh;
6178 if (batch_size == 0) {
6179 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6180 if (!list_empty(temp_inactive_list + i))
6182 if (i == NR_STRIPE_HASH_LOCKS) {
6183 spin_unlock_irq(&conf->device_lock);
6184 log_flush_stripe_to_raid(conf);
6185 spin_lock_irq(&conf->device_lock);
6188 release_inactive = true;
6190 spin_unlock_irq(&conf->device_lock);
6192 release_inactive_stripe_list(conf, temp_inactive_list,
6193 NR_STRIPE_HASH_LOCKS);
6195 r5l_flush_stripe_to_raid(conf->log);
6196 if (release_inactive) {
6197 spin_lock_irq(&conf->device_lock);
6201 for (i = 0; i < batch_size; i++)
6202 handle_stripe(batch[i]);
6203 log_write_stripe_run(conf);
6207 spin_lock_irq(&conf->device_lock);
6208 for (i = 0; i < batch_size; i++) {
6209 hash = batch[i]->hash_lock_index;
6210 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6215 static void raid5_do_work(struct work_struct *work)
6217 struct r5worker *worker = container_of(work, struct r5worker, work);
6218 struct r5worker_group *group = worker->group;
6219 struct r5conf *conf = group->conf;
6220 struct mddev *mddev = conf->mddev;
6221 int group_id = group - conf->worker_groups;
6223 struct blk_plug plug;
6225 pr_debug("+++ raid5worker active\n");
6227 blk_start_plug(&plug);
6229 spin_lock_irq(&conf->device_lock);
6231 int batch_size, released;
6233 released = release_stripe_list(conf, worker->temp_inactive_list);
6235 batch_size = handle_active_stripes(conf, group_id, worker,
6236 worker->temp_inactive_list);
6237 worker->working = false;
6238 if (!batch_size && !released)
6240 handled += batch_size;
6241 wait_event_lock_irq(mddev->sb_wait,
6242 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6245 pr_debug("%d stripes handled\n", handled);
6247 spin_unlock_irq(&conf->device_lock);
6249 flush_deferred_bios(conf);
6251 r5l_flush_stripe_to_raid(conf->log);
6253 async_tx_issue_pending_all();
6254 blk_finish_plug(&plug);
6256 pr_debug("--- raid5worker inactive\n");
6260 * This is our raid5 kernel thread.
6262 * We scan the hash table for stripes which can be handled now.
6263 * During the scan, completed stripes are saved for us by the interrupt
6264 * handler, so that they will not have to wait for our next wakeup.
6266 static void raid5d(struct md_thread *thread)
6268 struct mddev *mddev = thread->mddev;
6269 struct r5conf *conf = mddev->private;
6271 struct blk_plug plug;
6273 pr_debug("+++ raid5d active\n");
6275 md_check_recovery(mddev);
6277 blk_start_plug(&plug);
6279 spin_lock_irq(&conf->device_lock);
6282 int batch_size, released;
6283 unsigned int offset;
6285 released = release_stripe_list(conf, conf->temp_inactive_list);
6287 clear_bit(R5_DID_ALLOC, &conf->cache_state);
6290 !list_empty(&conf->bitmap_list)) {
6291 /* Now is a good time to flush some bitmap updates */
6293 spin_unlock_irq(&conf->device_lock);
6294 md_bitmap_unplug(mddev->bitmap);
6295 spin_lock_irq(&conf->device_lock);
6296 conf->seq_write = conf->seq_flush;
6297 activate_bit_delay(conf, conf->temp_inactive_list);
6299 raid5_activate_delayed(conf);
6301 while ((bio = remove_bio_from_retry(conf, &offset))) {
6303 spin_unlock_irq(&conf->device_lock);
6304 ok = retry_aligned_read(conf, bio, offset);
6305 spin_lock_irq(&conf->device_lock);
6311 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6312 conf->temp_inactive_list);
6313 if (!batch_size && !released)
6315 handled += batch_size;
6317 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6318 spin_unlock_irq(&conf->device_lock);
6319 md_check_recovery(mddev);
6320 spin_lock_irq(&conf->device_lock);
6323 pr_debug("%d stripes handled\n", handled);
6325 spin_unlock_irq(&conf->device_lock);
6326 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6327 mutex_trylock(&conf->cache_size_mutex)) {
6328 grow_one_stripe(conf, __GFP_NOWARN);
6329 /* Set flag even if allocation failed. This helps
6330 * slow down allocation requests when mem is short
6332 set_bit(R5_DID_ALLOC, &conf->cache_state);
6333 mutex_unlock(&conf->cache_size_mutex);
6336 flush_deferred_bios(conf);
6338 r5l_flush_stripe_to_raid(conf->log);
6340 async_tx_issue_pending_all();
6341 blk_finish_plug(&plug);
6343 pr_debug("--- raid5d inactive\n");
6347 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6349 struct r5conf *conf;
6351 spin_lock(&mddev->lock);
6352 conf = mddev->private;
6354 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6355 spin_unlock(&mddev->lock);
6360 raid5_set_cache_size(struct mddev *mddev, int size)
6363 struct r5conf *conf = mddev->private;
6365 if (size <= 16 || size > 32768)
6368 conf->min_nr_stripes = size;
6369 mutex_lock(&conf->cache_size_mutex);
6370 while (size < conf->max_nr_stripes &&
6371 drop_one_stripe(conf))
6373 mutex_unlock(&conf->cache_size_mutex);
6375 md_allow_write(mddev);
6377 mutex_lock(&conf->cache_size_mutex);
6378 while (size > conf->max_nr_stripes)
6379 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6380 conf->min_nr_stripes = conf->max_nr_stripes;
6384 mutex_unlock(&conf->cache_size_mutex);
6388 EXPORT_SYMBOL(raid5_set_cache_size);
6391 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6393 struct r5conf *conf;
6397 if (len >= PAGE_SIZE)
6399 if (kstrtoul(page, 10, &new))
6401 err = mddev_lock(mddev);
6404 conf = mddev->private;
6408 err = raid5_set_cache_size(mddev, new);
6409 mddev_unlock(mddev);
6414 static struct md_sysfs_entry
6415 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6416 raid5_show_stripe_cache_size,
6417 raid5_store_stripe_cache_size);
6420 raid5_show_rmw_level(struct mddev *mddev, char *page)
6422 struct r5conf *conf = mddev->private;
6424 return sprintf(page, "%d\n", conf->rmw_level);
6430 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6432 struct r5conf *conf = mddev->private;
6438 if (len >= PAGE_SIZE)
6441 if (kstrtoul(page, 10, &new))
6444 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6447 if (new != PARITY_DISABLE_RMW &&
6448 new != PARITY_ENABLE_RMW &&
6449 new != PARITY_PREFER_RMW)
6452 conf->rmw_level = new;
6456 static struct md_sysfs_entry
6457 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6458 raid5_show_rmw_level,
6459 raid5_store_rmw_level);
6463 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6465 struct r5conf *conf;
6467 spin_lock(&mddev->lock);
6468 conf = mddev->private;
6470 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6471 spin_unlock(&mddev->lock);
6476 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6478 struct r5conf *conf;
6482 if (len >= PAGE_SIZE)
6484 if (kstrtoul(page, 10, &new))
6487 err = mddev_lock(mddev);
6490 conf = mddev->private;
6493 else if (new > conf->min_nr_stripes)
6496 conf->bypass_threshold = new;
6497 mddev_unlock(mddev);
6501 static struct md_sysfs_entry
6502 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6504 raid5_show_preread_threshold,
6505 raid5_store_preread_threshold);
6508 raid5_show_skip_copy(struct mddev *mddev, char *page)
6510 struct r5conf *conf;
6512 spin_lock(&mddev->lock);
6513 conf = mddev->private;
6515 ret = sprintf(page, "%d\n", conf->skip_copy);
6516 spin_unlock(&mddev->lock);
6521 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6523 struct r5conf *conf;
6527 if (len >= PAGE_SIZE)
6529 if (kstrtoul(page, 10, &new))
6533 err = mddev_lock(mddev);
6536 conf = mddev->private;
6539 else if (new != conf->skip_copy) {
6540 mddev_suspend(mddev);
6541 conf->skip_copy = new;
6543 mddev->queue->backing_dev_info->capabilities |=
6544 BDI_CAP_STABLE_WRITES;
6546 mddev->queue->backing_dev_info->capabilities &=
6547 ~BDI_CAP_STABLE_WRITES;
6548 mddev_resume(mddev);
6550 mddev_unlock(mddev);
6554 static struct md_sysfs_entry
6555 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6556 raid5_show_skip_copy,
6557 raid5_store_skip_copy);
6560 stripe_cache_active_show(struct mddev *mddev, char *page)
6562 struct r5conf *conf = mddev->private;
6564 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6569 static struct md_sysfs_entry
6570 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6573 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6575 struct r5conf *conf;
6577 spin_lock(&mddev->lock);
6578 conf = mddev->private;
6580 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6581 spin_unlock(&mddev->lock);
6585 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6587 int *worker_cnt_per_group,
6588 struct r5worker_group **worker_groups);
6590 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6592 struct r5conf *conf;
6595 struct r5worker_group *new_groups, *old_groups;
6596 int group_cnt, worker_cnt_per_group;
6598 if (len >= PAGE_SIZE)
6600 if (kstrtouint(page, 10, &new))
6602 /* 8192 should be big enough */
6606 err = mddev_lock(mddev);
6609 conf = mddev->private;
6612 else if (new != conf->worker_cnt_per_group) {
6613 mddev_suspend(mddev);
6615 old_groups = conf->worker_groups;
6617 flush_workqueue(raid5_wq);
6619 err = alloc_thread_groups(conf, new,
6620 &group_cnt, &worker_cnt_per_group,
6623 spin_lock_irq(&conf->device_lock);
6624 conf->group_cnt = group_cnt;
6625 conf->worker_cnt_per_group = worker_cnt_per_group;
6626 conf->worker_groups = new_groups;
6627 spin_unlock_irq(&conf->device_lock);
6630 kfree(old_groups[0].workers);
6633 mddev_resume(mddev);
6635 mddev_unlock(mddev);
6640 static struct md_sysfs_entry
6641 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6642 raid5_show_group_thread_cnt,
6643 raid5_store_group_thread_cnt);
6645 static struct attribute *raid5_attrs[] = {
6646 &raid5_stripecache_size.attr,
6647 &raid5_stripecache_active.attr,
6648 &raid5_preread_bypass_threshold.attr,
6649 &raid5_group_thread_cnt.attr,
6650 &raid5_skip_copy.attr,
6651 &raid5_rmw_level.attr,
6652 &r5c_journal_mode.attr,
6653 &ppl_write_hint.attr,
6656 static struct attribute_group raid5_attrs_group = {
6658 .attrs = raid5_attrs,
6661 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6663 int *worker_cnt_per_group,
6664 struct r5worker_group **worker_groups)
6668 struct r5worker *workers;
6670 *worker_cnt_per_group = cnt;
6673 *worker_groups = NULL;
6676 *group_cnt = num_possible_nodes();
6677 size = sizeof(struct r5worker) * cnt;
6678 workers = kcalloc(size, *group_cnt, GFP_NOIO);
6679 *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6681 if (!*worker_groups || !workers) {
6683 kfree(*worker_groups);
6687 for (i = 0; i < *group_cnt; i++) {
6688 struct r5worker_group *group;
6690 group = &(*worker_groups)[i];
6691 INIT_LIST_HEAD(&group->handle_list);
6692 INIT_LIST_HEAD(&group->loprio_list);
6694 group->workers = workers + i * cnt;
6696 for (j = 0; j < cnt; j++) {
6697 struct r5worker *worker = group->workers + j;
6698 worker->group = group;
6699 INIT_WORK(&worker->work, raid5_do_work);
6701 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6702 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6709 static void free_thread_groups(struct r5conf *conf)
6711 if (conf->worker_groups)
6712 kfree(conf->worker_groups[0].workers);
6713 kfree(conf->worker_groups);
6714 conf->worker_groups = NULL;
6718 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6720 struct r5conf *conf = mddev->private;
6723 sectors = mddev->dev_sectors;
6725 /* size is defined by the smallest of previous and new size */
6726 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6728 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6729 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6730 return sectors * (raid_disks - conf->max_degraded);
6733 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6735 safe_put_page(percpu->spare_page);
6736 percpu->spare_page = NULL;
6737 kvfree(percpu->scribble);
6738 percpu->scribble = NULL;
6741 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6743 if (conf->level == 6 && !percpu->spare_page) {
6744 percpu->spare_page = alloc_page(GFP_KERNEL);
6745 if (!percpu->spare_page)
6749 if (scribble_alloc(percpu,
6750 max(conf->raid_disks,
6751 conf->previous_raid_disks),
6752 max(conf->chunk_sectors,
6753 conf->prev_chunk_sectors)
6756 free_scratch_buffer(conf, percpu);
6763 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6765 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6767 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6771 static void raid5_free_percpu(struct r5conf *conf)
6776 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6777 free_percpu(conf->percpu);
6780 static void free_conf(struct r5conf *conf)
6786 unregister_shrinker(&conf->shrinker);
6787 free_thread_groups(conf);
6788 shrink_stripes(conf);
6789 raid5_free_percpu(conf);
6790 for (i = 0; i < conf->pool_size; i++)
6791 if (conf->disks[i].extra_page)
6792 put_page(conf->disks[i].extra_page);
6794 bioset_exit(&conf->bio_split);
6795 kfree(conf->stripe_hashtbl);
6796 kfree(conf->pending_data);
6800 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6802 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6803 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6805 if (alloc_scratch_buffer(conf, percpu)) {
6806 pr_warn("%s: failed memory allocation for cpu%u\n",
6813 static int raid5_alloc_percpu(struct r5conf *conf)
6817 conf->percpu = alloc_percpu(struct raid5_percpu);
6821 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6823 conf->scribble_disks = max(conf->raid_disks,
6824 conf->previous_raid_disks);
6825 conf->scribble_sectors = max(conf->chunk_sectors,
6826 conf->prev_chunk_sectors);
6831 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6832 struct shrink_control *sc)
6834 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6835 unsigned long ret = SHRINK_STOP;
6837 if (mutex_trylock(&conf->cache_size_mutex)) {
6839 while (ret < sc->nr_to_scan &&
6840 conf->max_nr_stripes > conf->min_nr_stripes) {
6841 if (drop_one_stripe(conf) == 0) {
6847 mutex_unlock(&conf->cache_size_mutex);
6852 static unsigned long raid5_cache_count(struct shrinker *shrink,
6853 struct shrink_control *sc)
6855 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6857 if (conf->max_nr_stripes < conf->min_nr_stripes)
6858 /* unlikely, but not impossible */
6860 return conf->max_nr_stripes - conf->min_nr_stripes;
6863 static struct r5conf *setup_conf(struct mddev *mddev)
6865 struct r5conf *conf;
6866 int raid_disk, memory, max_disks;
6867 struct md_rdev *rdev;
6868 struct disk_info *disk;
6871 int group_cnt, worker_cnt_per_group;
6872 struct r5worker_group *new_group;
6875 if (mddev->new_level != 5
6876 && mddev->new_level != 4
6877 && mddev->new_level != 6) {
6878 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6879 mdname(mddev), mddev->new_level);
6880 return ERR_PTR(-EIO);
6882 if ((mddev->new_level == 5
6883 && !algorithm_valid_raid5(mddev->new_layout)) ||
6884 (mddev->new_level == 6
6885 && !algorithm_valid_raid6(mddev->new_layout))) {
6886 pr_warn("md/raid:%s: layout %d not supported\n",
6887 mdname(mddev), mddev->new_layout);
6888 return ERR_PTR(-EIO);
6890 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6891 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6892 mdname(mddev), mddev->raid_disks);
6893 return ERR_PTR(-EINVAL);
6896 if (!mddev->new_chunk_sectors ||
6897 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6898 !is_power_of_2(mddev->new_chunk_sectors)) {
6899 pr_warn("md/raid:%s: invalid chunk size %d\n",
6900 mdname(mddev), mddev->new_chunk_sectors << 9);
6901 return ERR_PTR(-EINVAL);
6904 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6907 INIT_LIST_HEAD(&conf->free_list);
6908 INIT_LIST_HEAD(&conf->pending_list);
6909 conf->pending_data = kcalloc(PENDING_IO_MAX,
6910 sizeof(struct r5pending_data),
6912 if (!conf->pending_data)
6914 for (i = 0; i < PENDING_IO_MAX; i++)
6915 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6916 /* Don't enable multi-threading by default*/
6917 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6919 conf->group_cnt = group_cnt;
6920 conf->worker_cnt_per_group = worker_cnt_per_group;
6921 conf->worker_groups = new_group;
6924 spin_lock_init(&conf->device_lock);
6925 seqcount_init(&conf->gen_lock);
6926 mutex_init(&conf->cache_size_mutex);
6927 init_waitqueue_head(&conf->wait_for_quiescent);
6928 init_waitqueue_head(&conf->wait_for_stripe);
6929 init_waitqueue_head(&conf->wait_for_overlap);
6930 INIT_LIST_HEAD(&conf->handle_list);
6931 INIT_LIST_HEAD(&conf->loprio_list);
6932 INIT_LIST_HEAD(&conf->hold_list);
6933 INIT_LIST_HEAD(&conf->delayed_list);
6934 INIT_LIST_HEAD(&conf->bitmap_list);
6935 init_llist_head(&conf->released_stripes);
6936 atomic_set(&conf->active_stripes, 0);
6937 atomic_set(&conf->preread_active_stripes, 0);
6938 atomic_set(&conf->active_aligned_reads, 0);
6939 spin_lock_init(&conf->pending_bios_lock);
6940 conf->batch_bio_dispatch = true;
6941 rdev_for_each(rdev, mddev) {
6942 if (test_bit(Journal, &rdev->flags))
6944 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6945 conf->batch_bio_dispatch = false;
6950 conf->bypass_threshold = BYPASS_THRESHOLD;
6951 conf->recovery_disabled = mddev->recovery_disabled - 1;
6953 conf->raid_disks = mddev->raid_disks;
6954 if (mddev->reshape_position == MaxSector)
6955 conf->previous_raid_disks = mddev->raid_disks;
6957 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6958 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6960 conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
6966 for (i = 0; i < max_disks; i++) {
6967 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6968 if (!conf->disks[i].extra_page)
6972 ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
6975 conf->mddev = mddev;
6977 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6980 /* We init hash_locks[0] separately to that it can be used
6981 * as the reference lock in the spin_lock_nest_lock() call
6982 * in lock_all_device_hash_locks_irq in order to convince
6983 * lockdep that we know what we are doing.
6985 spin_lock_init(conf->hash_locks);
6986 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6987 spin_lock_init(conf->hash_locks + i);
6989 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6990 INIT_LIST_HEAD(conf->inactive_list + i);
6992 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6993 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6995 atomic_set(&conf->r5c_cached_full_stripes, 0);
6996 INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
6997 atomic_set(&conf->r5c_cached_partial_stripes, 0);
6998 INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
6999 atomic_set(&conf->r5c_flushing_full_stripes, 0);
7000 atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7002 conf->level = mddev->new_level;
7003 conf->chunk_sectors = mddev->new_chunk_sectors;
7004 if (raid5_alloc_percpu(conf) != 0)
7007 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7009 rdev_for_each(rdev, mddev) {
7010 raid_disk = rdev->raid_disk;
7011 if (raid_disk >= max_disks
7012 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7014 disk = conf->disks + raid_disk;
7016 if (test_bit(Replacement, &rdev->flags)) {
7017 if (disk->replacement)
7019 disk->replacement = rdev;
7026 if (test_bit(In_sync, &rdev->flags)) {
7027 char b[BDEVNAME_SIZE];
7028 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7029 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7030 } else if (rdev->saved_raid_disk != raid_disk)
7031 /* Cannot rely on bitmap to complete recovery */
7035 conf->level = mddev->new_level;
7036 if (conf->level == 6) {
7037 conf->max_degraded = 2;
7038 if (raid6_call.xor_syndrome)
7039 conf->rmw_level = PARITY_ENABLE_RMW;
7041 conf->rmw_level = PARITY_DISABLE_RMW;
7043 conf->max_degraded = 1;
7044 conf->rmw_level = PARITY_ENABLE_RMW;
7046 conf->algorithm = mddev->new_layout;
7047 conf->reshape_progress = mddev->reshape_position;
7048 if (conf->reshape_progress != MaxSector) {
7049 conf->prev_chunk_sectors = mddev->chunk_sectors;
7050 conf->prev_algo = mddev->layout;
7052 conf->prev_chunk_sectors = conf->chunk_sectors;
7053 conf->prev_algo = conf->algorithm;
7056 conf->min_nr_stripes = NR_STRIPES;
7057 if (mddev->reshape_position != MaxSector) {
7058 int stripes = max_t(int,
7059 ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7060 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7061 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7062 if (conf->min_nr_stripes != NR_STRIPES)
7063 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7064 mdname(mddev), conf->min_nr_stripes);
7066 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7067 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7068 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7069 if (grow_stripes(conf, conf->min_nr_stripes)) {
7070 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7071 mdname(mddev), memory);
7074 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7076 * Losing a stripe head costs more than the time to refill it,
7077 * it reduces the queue depth and so can hurt throughput.
7078 * So set it rather large, scaled by number of devices.
7080 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7081 conf->shrinker.scan_objects = raid5_cache_scan;
7082 conf->shrinker.count_objects = raid5_cache_count;
7083 conf->shrinker.batch = 128;
7084 conf->shrinker.flags = 0;
7085 if (register_shrinker(&conf->shrinker)) {
7086 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7091 sprintf(pers_name, "raid%d", mddev->new_level);
7092 conf->thread = md_register_thread(raid5d, mddev, pers_name);
7093 if (!conf->thread) {
7094 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7104 return ERR_PTR(-EIO);
7106 return ERR_PTR(-ENOMEM);
7109 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7112 case ALGORITHM_PARITY_0:
7113 if (raid_disk < max_degraded)
7116 case ALGORITHM_PARITY_N:
7117 if (raid_disk >= raid_disks - max_degraded)
7120 case ALGORITHM_PARITY_0_6:
7121 if (raid_disk == 0 ||
7122 raid_disk == raid_disks - 1)
7125 case ALGORITHM_LEFT_ASYMMETRIC_6:
7126 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7127 case ALGORITHM_LEFT_SYMMETRIC_6:
7128 case ALGORITHM_RIGHT_SYMMETRIC_6:
7129 if (raid_disk == raid_disks - 1)
7135 static int raid5_run(struct mddev *mddev)
7137 struct r5conf *conf;
7138 int working_disks = 0;
7139 int dirty_parity_disks = 0;
7140 struct md_rdev *rdev;
7141 struct md_rdev *journal_dev = NULL;
7142 sector_t reshape_offset = 0;
7144 long long min_offset_diff = 0;
7147 if (mddev_init_writes_pending(mddev) < 0)
7150 if (mddev->recovery_cp != MaxSector)
7151 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7154 rdev_for_each(rdev, mddev) {
7157 if (test_bit(Journal, &rdev->flags)) {
7161 if (rdev->raid_disk < 0)
7163 diff = (rdev->new_data_offset - rdev->data_offset);
7165 min_offset_diff = diff;
7167 } else if (mddev->reshape_backwards &&
7168 diff < min_offset_diff)
7169 min_offset_diff = diff;
7170 else if (!mddev->reshape_backwards &&
7171 diff > min_offset_diff)
7172 min_offset_diff = diff;
7175 if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7176 (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7177 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7182 if (mddev->reshape_position != MaxSector) {
7183 /* Check that we can continue the reshape.
7184 * Difficulties arise if the stripe we would write to
7185 * next is at or after the stripe we would read from next.
7186 * For a reshape that changes the number of devices, this
7187 * is only possible for a very short time, and mdadm makes
7188 * sure that time appears to have past before assembling
7189 * the array. So we fail if that time hasn't passed.
7190 * For a reshape that keeps the number of devices the same
7191 * mdadm must be monitoring the reshape can keeping the
7192 * critical areas read-only and backed up. It will start
7193 * the array in read-only mode, so we check for that.
7195 sector_t here_new, here_old;
7197 int max_degraded = (mddev->level == 6 ? 2 : 1);
7202 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7207 if (mddev->new_level != mddev->level) {
7208 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7212 old_disks = mddev->raid_disks - mddev->delta_disks;
7213 /* reshape_position must be on a new-stripe boundary, and one
7214 * further up in new geometry must map after here in old
7216 * If the chunk sizes are different, then as we perform reshape
7217 * in units of the largest of the two, reshape_position needs
7218 * be a multiple of the largest chunk size times new data disks.
7220 here_new = mddev->reshape_position;
7221 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7222 new_data_disks = mddev->raid_disks - max_degraded;
7223 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7224 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7228 reshape_offset = here_new * chunk_sectors;
7229 /* here_new is the stripe we will write to */
7230 here_old = mddev->reshape_position;
7231 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7232 /* here_old is the first stripe that we might need to read
7234 if (mddev->delta_disks == 0) {
7235 /* We cannot be sure it is safe to start an in-place
7236 * reshape. It is only safe if user-space is monitoring
7237 * and taking constant backups.
7238 * mdadm always starts a situation like this in
7239 * readonly mode so it can take control before
7240 * allowing any writes. So just check for that.
7242 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7243 abs(min_offset_diff) >= mddev->new_chunk_sectors)
7244 /* not really in-place - so OK */;
7245 else if (mddev->ro == 0) {
7246 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7250 } else if (mddev->reshape_backwards
7251 ? (here_new * chunk_sectors + min_offset_diff <=
7252 here_old * chunk_sectors)
7253 : (here_new * chunk_sectors >=
7254 here_old * chunk_sectors + (-min_offset_diff))) {
7255 /* Reading from the same stripe as writing to - bad */
7256 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7260 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7261 /* OK, we should be able to continue; */
7263 BUG_ON(mddev->level != mddev->new_level);
7264 BUG_ON(mddev->layout != mddev->new_layout);
7265 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7266 BUG_ON(mddev->delta_disks != 0);
7269 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7270 test_bit(MD_HAS_PPL, &mddev->flags)) {
7271 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7273 clear_bit(MD_HAS_PPL, &mddev->flags);
7274 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7277 if (mddev->private == NULL)
7278 conf = setup_conf(mddev);
7280 conf = mddev->private;
7283 return PTR_ERR(conf);
7285 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7287 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7290 set_disk_ro(mddev->gendisk, 1);
7291 } else if (mddev->recovery_cp == MaxSector)
7292 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7295 conf->min_offset_diff = min_offset_diff;
7296 mddev->thread = conf->thread;
7297 conf->thread = NULL;
7298 mddev->private = conf;
7300 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7302 rdev = conf->disks[i].rdev;
7303 if (!rdev && conf->disks[i].replacement) {
7304 /* The replacement is all we have yet */
7305 rdev = conf->disks[i].replacement;
7306 conf->disks[i].replacement = NULL;
7307 clear_bit(Replacement, &rdev->flags);
7308 conf->disks[i].rdev = rdev;
7312 if (conf->disks[i].replacement &&
7313 conf->reshape_progress != MaxSector) {
7314 /* replacements and reshape simply do not mix. */
7315 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7318 if (test_bit(In_sync, &rdev->flags)) {
7322 /* This disc is not fully in-sync. However if it
7323 * just stored parity (beyond the recovery_offset),
7324 * when we don't need to be concerned about the
7325 * array being dirty.
7326 * When reshape goes 'backwards', we never have
7327 * partially completed devices, so we only need
7328 * to worry about reshape going forwards.
7330 /* Hack because v0.91 doesn't store recovery_offset properly. */
7331 if (mddev->major_version == 0 &&
7332 mddev->minor_version > 90)
7333 rdev->recovery_offset = reshape_offset;
7335 if (rdev->recovery_offset < reshape_offset) {
7336 /* We need to check old and new layout */
7337 if (!only_parity(rdev->raid_disk,
7340 conf->max_degraded))
7343 if (!only_parity(rdev->raid_disk,
7345 conf->previous_raid_disks,
7346 conf->max_degraded))
7348 dirty_parity_disks++;
7352 * 0 for a fully functional array, 1 or 2 for a degraded array.
7354 mddev->degraded = raid5_calc_degraded(conf);
7356 if (has_failed(conf)) {
7357 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7358 mdname(mddev), mddev->degraded, conf->raid_disks);
7362 /* device size must be a multiple of chunk size */
7363 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7364 mddev->resync_max_sectors = mddev->dev_sectors;
7366 if (mddev->degraded > dirty_parity_disks &&
7367 mddev->recovery_cp != MaxSector) {
7368 if (test_bit(MD_HAS_PPL, &mddev->flags))
7369 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7371 else if (mddev->ok_start_degraded)
7372 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7375 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7381 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7382 mdname(mddev), conf->level,
7383 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7386 print_raid5_conf(conf);
7388 if (conf->reshape_progress != MaxSector) {
7389 conf->reshape_safe = conf->reshape_progress;
7390 atomic_set(&conf->reshape_stripes, 0);
7391 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7392 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7393 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7394 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7395 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7397 if (!mddev->sync_thread)
7401 /* Ok, everything is just fine now */
7402 if (mddev->to_remove == &raid5_attrs_group)
7403 mddev->to_remove = NULL;
7404 else if (mddev->kobj.sd &&
7405 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7406 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7408 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7412 /* read-ahead size must cover two whole stripes, which
7413 * is 2 * (datadisks) * chunksize where 'n' is the
7414 * number of raid devices
7416 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7417 int stripe = data_disks *
7418 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7419 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7420 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7422 chunk_size = mddev->chunk_sectors << 9;
7423 blk_queue_io_min(mddev->queue, chunk_size);
7424 blk_queue_io_opt(mddev->queue, chunk_size *
7425 (conf->raid_disks - conf->max_degraded));
7426 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7428 * We can only discard a whole stripe. It doesn't make sense to
7429 * discard data disk but write parity disk
7431 stripe = stripe * PAGE_SIZE;
7432 /* Round up to power of 2, as discard handling
7433 * currently assumes that */
7434 while ((stripe-1) & stripe)
7435 stripe = (stripe | (stripe-1)) + 1;
7436 mddev->queue->limits.discard_alignment = stripe;
7437 mddev->queue->limits.discard_granularity = stripe;
7439 blk_queue_max_write_same_sectors(mddev->queue, 0);
7440 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7442 rdev_for_each(rdev, mddev) {
7443 disk_stack_limits(mddev->gendisk, rdev->bdev,
7444 rdev->data_offset << 9);
7445 disk_stack_limits(mddev->gendisk, rdev->bdev,
7446 rdev->new_data_offset << 9);
7450 * zeroing is required, otherwise data
7451 * could be lost. Consider a scenario: discard a stripe
7452 * (the stripe could be inconsistent if
7453 * discard_zeroes_data is 0); write one disk of the
7454 * stripe (the stripe could be inconsistent again
7455 * depending on which disks are used to calculate
7456 * parity); the disk is broken; The stripe data of this
7459 * We only allow DISCARD if the sysadmin has confirmed that
7460 * only safe devices are in use by setting a module parameter.
7461 * A better idea might be to turn DISCARD into WRITE_ZEROES
7462 * requests, as that is required to be safe.
7464 if (devices_handle_discard_safely &&
7465 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7466 mddev->queue->limits.discard_granularity >= stripe)
7467 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7470 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7473 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7476 if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7481 md_unregister_thread(&mddev->thread);
7482 print_raid5_conf(conf);
7484 mddev->private = NULL;
7485 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7489 static void raid5_free(struct mddev *mddev, void *priv)
7491 struct r5conf *conf = priv;
7494 mddev->to_remove = &raid5_attrs_group;
7497 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7499 struct r5conf *conf = mddev->private;
7502 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7503 conf->chunk_sectors / 2, mddev->layout);
7504 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7506 for (i = 0; i < conf->raid_disks; i++) {
7507 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7508 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7511 seq_printf (seq, "]");
7514 static void print_raid5_conf (struct r5conf *conf)
7517 struct disk_info *tmp;
7519 pr_debug("RAID conf printout:\n");
7521 pr_debug("(conf==NULL)\n");
7524 pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7526 conf->raid_disks - conf->mddev->degraded);
7528 for (i = 0; i < conf->raid_disks; i++) {
7529 char b[BDEVNAME_SIZE];
7530 tmp = conf->disks + i;
7532 pr_debug(" disk %d, o:%d, dev:%s\n",
7533 i, !test_bit(Faulty, &tmp->rdev->flags),
7534 bdevname(tmp->rdev->bdev, b));
7538 static int raid5_spare_active(struct mddev *mddev)
7541 struct r5conf *conf = mddev->private;
7542 struct disk_info *tmp;
7544 unsigned long flags;
7546 for (i = 0; i < conf->raid_disks; i++) {
7547 tmp = conf->disks + i;
7548 if (tmp->replacement
7549 && tmp->replacement->recovery_offset == MaxSector
7550 && !test_bit(Faulty, &tmp->replacement->flags)
7551 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7552 /* Replacement has just become active. */
7554 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7557 /* Replaced device not technically faulty,
7558 * but we need to be sure it gets removed
7559 * and never re-added.
7561 set_bit(Faulty, &tmp->rdev->flags);
7562 sysfs_notify_dirent_safe(
7563 tmp->rdev->sysfs_state);
7565 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7566 } else if (tmp->rdev
7567 && tmp->rdev->recovery_offset == MaxSector
7568 && !test_bit(Faulty, &tmp->rdev->flags)
7569 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7571 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7574 spin_lock_irqsave(&conf->device_lock, flags);
7575 mddev->degraded = raid5_calc_degraded(conf);
7576 spin_unlock_irqrestore(&conf->device_lock, flags);
7577 print_raid5_conf(conf);
7581 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7583 struct r5conf *conf = mddev->private;
7585 int number = rdev->raid_disk;
7586 struct md_rdev **rdevp;
7587 struct disk_info *p = conf->disks + number;
7589 print_raid5_conf(conf);
7590 if (test_bit(Journal, &rdev->flags) && conf->log) {
7592 * we can't wait pending write here, as this is called in
7593 * raid5d, wait will deadlock.
7594 * neilb: there is no locking about new writes here,
7595 * so this cannot be safe.
7597 if (atomic_read(&conf->active_stripes) ||
7598 atomic_read(&conf->r5c_cached_full_stripes) ||
7599 atomic_read(&conf->r5c_cached_partial_stripes)) {
7605 if (rdev == p->rdev)
7607 else if (rdev == p->replacement)
7608 rdevp = &p->replacement;
7612 if (number >= conf->raid_disks &&
7613 conf->reshape_progress == MaxSector)
7614 clear_bit(In_sync, &rdev->flags);
7616 if (test_bit(In_sync, &rdev->flags) ||
7617 atomic_read(&rdev->nr_pending)) {
7621 /* Only remove non-faulty devices if recovery
7624 if (!test_bit(Faulty, &rdev->flags) &&
7625 mddev->recovery_disabled != conf->recovery_disabled &&
7626 !has_failed(conf) &&
7627 (!p->replacement || p->replacement == rdev) &&
7628 number < conf->raid_disks) {
7633 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7635 if (atomic_read(&rdev->nr_pending)) {
7636 /* lost the race, try later */
7642 err = log_modify(conf, rdev, false);
7646 if (p->replacement) {
7647 /* We must have just cleared 'rdev' */
7648 p->rdev = p->replacement;
7649 clear_bit(Replacement, &p->replacement->flags);
7650 smp_mb(); /* Make sure other CPUs may see both as identical
7651 * but will never see neither - if they are careful
7653 p->replacement = NULL;
7656 err = log_modify(conf, p->rdev, true);
7659 clear_bit(WantReplacement, &rdev->flags);
7662 print_raid5_conf(conf);
7666 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7668 struct r5conf *conf = mddev->private;
7671 struct disk_info *p;
7673 int last = conf->raid_disks - 1;
7675 if (test_bit(Journal, &rdev->flags)) {
7679 rdev->raid_disk = 0;
7681 * The array is in readonly mode if journal is missing, so no
7682 * write requests running. We should be safe
7684 log_init(conf, rdev, false);
7687 if (mddev->recovery_disabled == conf->recovery_disabled)
7690 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7691 /* no point adding a device */
7694 if (rdev->raid_disk >= 0)
7695 first = last = rdev->raid_disk;
7698 * find the disk ... but prefer rdev->saved_raid_disk
7701 if (rdev->saved_raid_disk >= 0 &&
7702 rdev->saved_raid_disk >= first &&
7703 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7704 first = rdev->saved_raid_disk;
7706 for (disk = first; disk <= last; disk++) {
7707 p = conf->disks + disk;
7708 if (p->rdev == NULL) {
7709 clear_bit(In_sync, &rdev->flags);
7710 rdev->raid_disk = disk;
7711 if (rdev->saved_raid_disk != disk)
7713 rcu_assign_pointer(p->rdev, rdev);
7715 err = log_modify(conf, rdev, true);
7720 for (disk = first; disk <= last; disk++) {
7721 p = conf->disks + disk;
7722 if (test_bit(WantReplacement, &p->rdev->flags) &&
7723 p->replacement == NULL) {
7724 clear_bit(In_sync, &rdev->flags);
7725 set_bit(Replacement, &rdev->flags);
7726 rdev->raid_disk = disk;
7729 rcu_assign_pointer(p->replacement, rdev);
7734 print_raid5_conf(conf);
7738 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7740 /* no resync is happening, and there is enough space
7741 * on all devices, so we can resize.
7742 * We need to make sure resync covers any new space.
7743 * If the array is shrinking we should possibly wait until
7744 * any io in the removed space completes, but it hardly seems
7748 struct r5conf *conf = mddev->private;
7750 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7752 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7753 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7754 if (mddev->external_size &&
7755 mddev->array_sectors > newsize)
7757 if (mddev->bitmap) {
7758 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
7762 md_set_array_sectors(mddev, newsize);
7763 if (sectors > mddev->dev_sectors &&
7764 mddev->recovery_cp > mddev->dev_sectors) {
7765 mddev->recovery_cp = mddev->dev_sectors;
7766 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7768 mddev->dev_sectors = sectors;
7769 mddev->resync_max_sectors = sectors;
7773 static int check_stripe_cache(struct mddev *mddev)
7775 /* Can only proceed if there are plenty of stripe_heads.
7776 * We need a minimum of one full stripe,, and for sensible progress
7777 * it is best to have about 4 times that.
7778 * If we require 4 times, then the default 256 4K stripe_heads will
7779 * allow for chunk sizes up to 256K, which is probably OK.
7780 * If the chunk size is greater, user-space should request more
7781 * stripe_heads first.
7783 struct r5conf *conf = mddev->private;
7784 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7785 > conf->min_nr_stripes ||
7786 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7787 > conf->min_nr_stripes) {
7788 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7790 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7797 static int check_reshape(struct mddev *mddev)
7799 struct r5conf *conf = mddev->private;
7801 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7803 if (mddev->delta_disks == 0 &&
7804 mddev->new_layout == mddev->layout &&
7805 mddev->new_chunk_sectors == mddev->chunk_sectors)
7806 return 0; /* nothing to do */
7807 if (has_failed(conf))
7809 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7810 /* We might be able to shrink, but the devices must
7811 * be made bigger first.
7812 * For raid6, 4 is the minimum size.
7813 * Otherwise 2 is the minimum
7816 if (mddev->level == 6)
7818 if (mddev->raid_disks + mddev->delta_disks < min)
7822 if (!check_stripe_cache(mddev))
7825 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7826 mddev->delta_disks > 0)
7827 if (resize_chunks(conf,
7828 conf->previous_raid_disks
7829 + max(0, mddev->delta_disks),
7830 max(mddev->new_chunk_sectors,
7831 mddev->chunk_sectors)
7835 if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7836 return 0; /* never bother to shrink */
7837 return resize_stripes(conf, (conf->previous_raid_disks
7838 + mddev->delta_disks));
7841 static int raid5_start_reshape(struct mddev *mddev)
7843 struct r5conf *conf = mddev->private;
7844 struct md_rdev *rdev;
7846 unsigned long flags;
7848 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7851 if (!check_stripe_cache(mddev))
7854 if (has_failed(conf))
7857 rdev_for_each(rdev, mddev) {
7858 if (!test_bit(In_sync, &rdev->flags)
7859 && !test_bit(Faulty, &rdev->flags))
7863 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7864 /* Not enough devices even to make a degraded array
7869 /* Refuse to reduce size of the array. Any reductions in
7870 * array size must be through explicit setting of array_size
7873 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7874 < mddev->array_sectors) {
7875 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7880 atomic_set(&conf->reshape_stripes, 0);
7881 spin_lock_irq(&conf->device_lock);
7882 write_seqcount_begin(&conf->gen_lock);
7883 conf->previous_raid_disks = conf->raid_disks;
7884 conf->raid_disks += mddev->delta_disks;
7885 conf->prev_chunk_sectors = conf->chunk_sectors;
7886 conf->chunk_sectors = mddev->new_chunk_sectors;
7887 conf->prev_algo = conf->algorithm;
7888 conf->algorithm = mddev->new_layout;
7890 /* Code that selects data_offset needs to see the generation update
7891 * if reshape_progress has been set - so a memory barrier needed.
7894 if (mddev->reshape_backwards)
7895 conf->reshape_progress = raid5_size(mddev, 0, 0);
7897 conf->reshape_progress = 0;
7898 conf->reshape_safe = conf->reshape_progress;
7899 write_seqcount_end(&conf->gen_lock);
7900 spin_unlock_irq(&conf->device_lock);
7902 /* Now make sure any requests that proceeded on the assumption
7903 * the reshape wasn't running - like Discard or Read - have
7906 mddev_suspend(mddev);
7907 mddev_resume(mddev);
7909 /* Add some new drives, as many as will fit.
7910 * We know there are enough to make the newly sized array work.
7911 * Don't add devices if we are reducing the number of
7912 * devices in the array. This is because it is not possible
7913 * to correctly record the "partially reconstructed" state of
7914 * such devices during the reshape and confusion could result.
7916 if (mddev->delta_disks >= 0) {
7917 rdev_for_each(rdev, mddev)
7918 if (rdev->raid_disk < 0 &&
7919 !test_bit(Faulty, &rdev->flags)) {
7920 if (raid5_add_disk(mddev, rdev) == 0) {
7922 >= conf->previous_raid_disks)
7923 set_bit(In_sync, &rdev->flags);
7925 rdev->recovery_offset = 0;
7927 if (sysfs_link_rdev(mddev, rdev))
7928 /* Failure here is OK */;
7930 } else if (rdev->raid_disk >= conf->previous_raid_disks
7931 && !test_bit(Faulty, &rdev->flags)) {
7932 /* This is a spare that was manually added */
7933 set_bit(In_sync, &rdev->flags);
7936 /* When a reshape changes the number of devices,
7937 * ->degraded is measured against the larger of the
7938 * pre and post number of devices.
7940 spin_lock_irqsave(&conf->device_lock, flags);
7941 mddev->degraded = raid5_calc_degraded(conf);
7942 spin_unlock_irqrestore(&conf->device_lock, flags);
7944 mddev->raid_disks = conf->raid_disks;
7945 mddev->reshape_position = conf->reshape_progress;
7946 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7948 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7949 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7950 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7951 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7952 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7953 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7955 if (!mddev->sync_thread) {
7956 mddev->recovery = 0;
7957 spin_lock_irq(&conf->device_lock);
7958 write_seqcount_begin(&conf->gen_lock);
7959 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7960 mddev->new_chunk_sectors =
7961 conf->chunk_sectors = conf->prev_chunk_sectors;
7962 mddev->new_layout = conf->algorithm = conf->prev_algo;
7963 rdev_for_each(rdev, mddev)
7964 rdev->new_data_offset = rdev->data_offset;
7966 conf->generation --;
7967 conf->reshape_progress = MaxSector;
7968 mddev->reshape_position = MaxSector;
7969 write_seqcount_end(&conf->gen_lock);
7970 spin_unlock_irq(&conf->device_lock);
7973 conf->reshape_checkpoint = jiffies;
7974 md_wakeup_thread(mddev->sync_thread);
7975 md_new_event(mddev);
7979 /* This is called from the reshape thread and should make any
7980 * changes needed in 'conf'
7982 static void end_reshape(struct r5conf *conf)
7985 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7986 struct md_rdev *rdev;
7988 spin_lock_irq(&conf->device_lock);
7989 conf->previous_raid_disks = conf->raid_disks;
7990 md_finish_reshape(conf->mddev);
7992 conf->reshape_progress = MaxSector;
7993 conf->mddev->reshape_position = MaxSector;
7994 rdev_for_each(rdev, conf->mddev)
7995 if (rdev->raid_disk >= 0 &&
7996 !test_bit(Journal, &rdev->flags) &&
7997 !test_bit(In_sync, &rdev->flags))
7998 rdev->recovery_offset = MaxSector;
7999 spin_unlock_irq(&conf->device_lock);
8000 wake_up(&conf->wait_for_overlap);
8002 /* read-ahead size must cover two whole stripes, which is
8003 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
8005 if (conf->mddev->queue) {
8006 int data_disks = conf->raid_disks - conf->max_degraded;
8007 int stripe = data_disks * ((conf->chunk_sectors << 9)
8009 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
8010 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
8015 /* This is called from the raid5d thread with mddev_lock held.
8016 * It makes config changes to the device.
8018 static void raid5_finish_reshape(struct mddev *mddev)
8020 struct r5conf *conf = mddev->private;
8022 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8024 if (mddev->delta_disks <= 0) {
8026 spin_lock_irq(&conf->device_lock);
8027 mddev->degraded = raid5_calc_degraded(conf);
8028 spin_unlock_irq(&conf->device_lock);
8029 for (d = conf->raid_disks ;
8030 d < conf->raid_disks - mddev->delta_disks;
8032 struct md_rdev *rdev = conf->disks[d].rdev;
8034 clear_bit(In_sync, &rdev->flags);
8035 rdev = conf->disks[d].replacement;
8037 clear_bit(In_sync, &rdev->flags);
8040 mddev->layout = conf->algorithm;
8041 mddev->chunk_sectors = conf->chunk_sectors;
8042 mddev->reshape_position = MaxSector;
8043 mddev->delta_disks = 0;
8044 mddev->reshape_backwards = 0;
8048 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8050 struct r5conf *conf = mddev->private;
8053 /* stop all writes */
8054 lock_all_device_hash_locks_irq(conf);
8055 /* '2' tells resync/reshape to pause so that all
8056 * active stripes can drain
8058 r5c_flush_cache(conf, INT_MAX);
8060 wait_event_cmd(conf->wait_for_quiescent,
8061 atomic_read(&conf->active_stripes) == 0 &&
8062 atomic_read(&conf->active_aligned_reads) == 0,
8063 unlock_all_device_hash_locks_irq(conf),
8064 lock_all_device_hash_locks_irq(conf));
8066 unlock_all_device_hash_locks_irq(conf);
8067 /* allow reshape to continue */
8068 wake_up(&conf->wait_for_overlap);
8070 /* re-enable writes */
8071 lock_all_device_hash_locks_irq(conf);
8073 wake_up(&conf->wait_for_quiescent);
8074 wake_up(&conf->wait_for_overlap);
8075 unlock_all_device_hash_locks_irq(conf);
8077 log_quiesce(conf, quiesce);
8080 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8082 struct r0conf *raid0_conf = mddev->private;
8085 /* for raid0 takeover only one zone is supported */
8086 if (raid0_conf->nr_strip_zones > 1) {
8087 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8089 return ERR_PTR(-EINVAL);
8092 sectors = raid0_conf->strip_zone[0].zone_end;
8093 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8094 mddev->dev_sectors = sectors;
8095 mddev->new_level = level;
8096 mddev->new_layout = ALGORITHM_PARITY_N;
8097 mddev->new_chunk_sectors = mddev->chunk_sectors;
8098 mddev->raid_disks += 1;
8099 mddev->delta_disks = 1;
8100 /* make sure it will be not marked as dirty */
8101 mddev->recovery_cp = MaxSector;
8103 return setup_conf(mddev);
8106 static void *raid5_takeover_raid1(struct mddev *mddev)
8111 if (mddev->raid_disks != 2 ||
8112 mddev->degraded > 1)
8113 return ERR_PTR(-EINVAL);
8115 /* Should check if there are write-behind devices? */
8117 chunksect = 64*2; /* 64K by default */
8119 /* The array must be an exact multiple of chunksize */
8120 while (chunksect && (mddev->array_sectors & (chunksect-1)))
8123 if ((chunksect<<9) < STRIPE_SIZE)
8124 /* array size does not allow a suitable chunk size */
8125 return ERR_PTR(-EINVAL);
8127 mddev->new_level = 5;
8128 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8129 mddev->new_chunk_sectors = chunksect;
8131 ret = setup_conf(mddev);
8133 mddev_clear_unsupported_flags(mddev,
8134 UNSUPPORTED_MDDEV_FLAGS);
8138 static void *raid5_takeover_raid6(struct mddev *mddev)
8142 switch (mddev->layout) {
8143 case ALGORITHM_LEFT_ASYMMETRIC_6:
8144 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8146 case ALGORITHM_RIGHT_ASYMMETRIC_6:
8147 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8149 case ALGORITHM_LEFT_SYMMETRIC_6:
8150 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8152 case ALGORITHM_RIGHT_SYMMETRIC_6:
8153 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8155 case ALGORITHM_PARITY_0_6:
8156 new_layout = ALGORITHM_PARITY_0;
8158 case ALGORITHM_PARITY_N:
8159 new_layout = ALGORITHM_PARITY_N;
8162 return ERR_PTR(-EINVAL);
8164 mddev->new_level = 5;
8165 mddev->new_layout = new_layout;
8166 mddev->delta_disks = -1;
8167 mddev->raid_disks -= 1;
8168 return setup_conf(mddev);
8171 static int raid5_check_reshape(struct mddev *mddev)
8173 /* For a 2-drive array, the layout and chunk size can be changed
8174 * immediately as not restriping is needed.
8175 * For larger arrays we record the new value - after validation
8176 * to be used by a reshape pass.
8178 struct r5conf *conf = mddev->private;
8179 int new_chunk = mddev->new_chunk_sectors;
8181 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8183 if (new_chunk > 0) {
8184 if (!is_power_of_2(new_chunk))
8186 if (new_chunk < (PAGE_SIZE>>9))
8188 if (mddev->array_sectors & (new_chunk-1))
8189 /* not factor of array size */
8193 /* They look valid */
8195 if (mddev->raid_disks == 2) {
8196 /* can make the change immediately */
8197 if (mddev->new_layout >= 0) {
8198 conf->algorithm = mddev->new_layout;
8199 mddev->layout = mddev->new_layout;
8201 if (new_chunk > 0) {
8202 conf->chunk_sectors = new_chunk ;
8203 mddev->chunk_sectors = new_chunk;
8205 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8206 md_wakeup_thread(mddev->thread);
8208 return check_reshape(mddev);
8211 static int raid6_check_reshape(struct mddev *mddev)
8213 int new_chunk = mddev->new_chunk_sectors;
8215 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8217 if (new_chunk > 0) {
8218 if (!is_power_of_2(new_chunk))
8220 if (new_chunk < (PAGE_SIZE >> 9))
8222 if (mddev->array_sectors & (new_chunk-1))
8223 /* not factor of array size */
8227 /* They look valid */
8228 return check_reshape(mddev);
8231 static void *raid5_takeover(struct mddev *mddev)
8233 /* raid5 can take over:
8234 * raid0 - if there is only one strip zone - make it a raid4 layout
8235 * raid1 - if there are two drives. We need to know the chunk size
8236 * raid4 - trivial - just use a raid4 layout.
8237 * raid6 - Providing it is a *_6 layout
8239 if (mddev->level == 0)
8240 return raid45_takeover_raid0(mddev, 5);
8241 if (mddev->level == 1)
8242 return raid5_takeover_raid1(mddev);
8243 if (mddev->level == 4) {
8244 mddev->new_layout = ALGORITHM_PARITY_N;
8245 mddev->new_level = 5;
8246 return setup_conf(mddev);
8248 if (mddev->level == 6)
8249 return raid5_takeover_raid6(mddev);
8251 return ERR_PTR(-EINVAL);
8254 static void *raid4_takeover(struct mddev *mddev)
8256 /* raid4 can take over:
8257 * raid0 - if there is only one strip zone
8258 * raid5 - if layout is right
8260 if (mddev->level == 0)
8261 return raid45_takeover_raid0(mddev, 4);
8262 if (mddev->level == 5 &&
8263 mddev->layout == ALGORITHM_PARITY_N) {
8264 mddev->new_layout = 0;
8265 mddev->new_level = 4;
8266 return setup_conf(mddev);
8268 return ERR_PTR(-EINVAL);
8271 static struct md_personality raid5_personality;
8273 static void *raid6_takeover(struct mddev *mddev)
8275 /* Currently can only take over a raid5. We map the
8276 * personality to an equivalent raid6 personality
8277 * with the Q block at the end.
8281 if (mddev->pers != &raid5_personality)
8282 return ERR_PTR(-EINVAL);
8283 if (mddev->degraded > 1)
8284 return ERR_PTR(-EINVAL);
8285 if (mddev->raid_disks > 253)
8286 return ERR_PTR(-EINVAL);
8287 if (mddev->raid_disks < 3)
8288 return ERR_PTR(-EINVAL);
8290 switch (mddev->layout) {
8291 case ALGORITHM_LEFT_ASYMMETRIC:
8292 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8294 case ALGORITHM_RIGHT_ASYMMETRIC:
8295 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8297 case ALGORITHM_LEFT_SYMMETRIC:
8298 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8300 case ALGORITHM_RIGHT_SYMMETRIC:
8301 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8303 case ALGORITHM_PARITY_0:
8304 new_layout = ALGORITHM_PARITY_0_6;
8306 case ALGORITHM_PARITY_N:
8307 new_layout = ALGORITHM_PARITY_N;
8310 return ERR_PTR(-EINVAL);
8312 mddev->new_level = 6;
8313 mddev->new_layout = new_layout;
8314 mddev->delta_disks = 1;
8315 mddev->raid_disks += 1;
8316 return setup_conf(mddev);
8319 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8321 struct r5conf *conf;
8324 err = mddev_lock(mddev);
8327 conf = mddev->private;
8329 mddev_unlock(mddev);
8333 if (strncmp(buf, "ppl", 3) == 0) {
8334 /* ppl only works with RAID 5 */
8335 if (!raid5_has_ppl(conf) && conf->level == 5) {
8336 err = log_init(conf, NULL, true);
8338 err = resize_stripes(conf, conf->pool_size);
8344 } else if (strncmp(buf, "resync", 6) == 0) {
8345 if (raid5_has_ppl(conf)) {
8346 mddev_suspend(mddev);
8348 mddev_resume(mddev);
8349 err = resize_stripes(conf, conf->pool_size);
8350 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8351 r5l_log_disk_error(conf)) {
8352 bool journal_dev_exists = false;
8353 struct md_rdev *rdev;
8355 rdev_for_each(rdev, mddev)
8356 if (test_bit(Journal, &rdev->flags)) {
8357 journal_dev_exists = true;
8361 if (!journal_dev_exists) {
8362 mddev_suspend(mddev);
8363 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8364 mddev_resume(mddev);
8365 } else /* need remove journal device first */
8374 md_update_sb(mddev, 1);
8376 mddev_unlock(mddev);
8381 static int raid5_start(struct mddev *mddev)
8383 struct r5conf *conf = mddev->private;
8385 return r5l_start(conf->log);
8388 static struct md_personality raid6_personality =
8392 .owner = THIS_MODULE,
8393 .make_request = raid5_make_request,
8395 .start = raid5_start,
8397 .status = raid5_status,
8398 .error_handler = raid5_error,
8399 .hot_add_disk = raid5_add_disk,
8400 .hot_remove_disk= raid5_remove_disk,
8401 .spare_active = raid5_spare_active,
8402 .sync_request = raid5_sync_request,
8403 .resize = raid5_resize,
8405 .check_reshape = raid6_check_reshape,
8406 .start_reshape = raid5_start_reshape,
8407 .finish_reshape = raid5_finish_reshape,
8408 .quiesce = raid5_quiesce,
8409 .takeover = raid6_takeover,
8410 .congested = raid5_congested,
8411 .change_consistency_policy = raid5_change_consistency_policy,
8413 static struct md_personality raid5_personality =
8417 .owner = THIS_MODULE,
8418 .make_request = raid5_make_request,
8420 .start = raid5_start,
8422 .status = raid5_status,
8423 .error_handler = raid5_error,
8424 .hot_add_disk = raid5_add_disk,
8425 .hot_remove_disk= raid5_remove_disk,
8426 .spare_active = raid5_spare_active,
8427 .sync_request = raid5_sync_request,
8428 .resize = raid5_resize,
8430 .check_reshape = raid5_check_reshape,
8431 .start_reshape = raid5_start_reshape,
8432 .finish_reshape = raid5_finish_reshape,
8433 .quiesce = raid5_quiesce,
8434 .takeover = raid5_takeover,
8435 .congested = raid5_congested,
8436 .change_consistency_policy = raid5_change_consistency_policy,
8439 static struct md_personality raid4_personality =
8443 .owner = THIS_MODULE,
8444 .make_request = raid5_make_request,
8446 .start = raid5_start,
8448 .status = raid5_status,
8449 .error_handler = raid5_error,
8450 .hot_add_disk = raid5_add_disk,
8451 .hot_remove_disk= raid5_remove_disk,
8452 .spare_active = raid5_spare_active,
8453 .sync_request = raid5_sync_request,
8454 .resize = raid5_resize,
8456 .check_reshape = raid5_check_reshape,
8457 .start_reshape = raid5_start_reshape,
8458 .finish_reshape = raid5_finish_reshape,
8459 .quiesce = raid5_quiesce,
8460 .takeover = raid4_takeover,
8461 .congested = raid5_congested,
8462 .change_consistency_policy = raid5_change_consistency_policy,
8465 static int __init raid5_init(void)
8469 raid5_wq = alloc_workqueue("raid5wq",
8470 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8474 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8476 raid456_cpu_up_prepare,
8479 destroy_workqueue(raid5_wq);
8482 register_md_personality(&raid6_personality);
8483 register_md_personality(&raid5_personality);
8484 register_md_personality(&raid4_personality);
8488 static void raid5_exit(void)
8490 unregister_md_personality(&raid6_personality);
8491 unregister_md_personality(&raid5_personality);
8492 unregister_md_personality(&raid4_personality);
8493 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8494 destroy_workqueue(raid5_wq);
8497 module_init(raid5_init);
8498 module_exit(raid5_exit);
8499 MODULE_LICENSE("GPL");
8500 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8501 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8502 MODULE_ALIAS("md-raid5");
8503 MODULE_ALIAS("md-raid4");
8504 MODULE_ALIAS("md-level-5");
8505 MODULE_ALIAS("md-level-4");
8506 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8507 MODULE_ALIAS("md-raid6");
8508 MODULE_ALIAS("md-level-6");
8510 /* This used to be two separate modules, they were: */
8511 MODULE_ALIAS("raid5");
8512 MODULE_ALIAS("raid6");