1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * raid5.c : Multiple Devices driver for Linux
4 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
5 * Copyright (C) 1999, 2000 Ingo Molnar
6 * Copyright (C) 2002, 2003 H. Peter Anvin
8 * RAID-4/5/6 management functions.
9 * Thanks to Penguin Computing for making the RAID-6 development possible
10 * by donating a test server!
16 * The sequencing for updating the bitmap reliably is a little
17 * subtle (and I got it wrong the first time) so it deserves some
20 * We group bitmap updates into batches. Each batch has a number.
21 * We may write out several batches at once, but that isn't very important.
22 * conf->seq_write is the number of the last batch successfully written.
23 * conf->seq_flush is the number of the last batch that was closed to
25 * When we discover that we will need to write to any block in a stripe
26 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
27 * the number of the batch it will be in. This is seq_flush+1.
28 * When we are ready to do a write, if that batch hasn't been written yet,
29 * we plug the array and queue the stripe for later.
30 * When an unplug happens, we increment bm_flush, thus closing the current
32 * When we notice that bm_flush > bm_write, we write out all pending updates
33 * to the bitmap, and advance bm_write to where bm_flush was.
34 * This may occasionally write a bit out twice, but is sure never to
38 #include <linux/blkdev.h>
39 #include <linux/kthread.h>
40 #include <linux/raid/pq.h>
41 #include <linux/async_tx.h>
42 #include <linux/module.h>
43 #include <linux/async.h>
44 #include <linux/seq_file.h>
45 #include <linux/cpu.h>
46 #include <linux/slab.h>
47 #include <linux/ratelimit.h>
48 #include <linux/nodemask.h>
50 #include <trace/events/block.h>
51 #include <linux/list_sort.h>
56 #include "md-bitmap.h"
57 #include "raid5-log.h"
59 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
61 #define cpu_to_group(cpu) cpu_to_node(cpu)
62 #define ANY_GROUP NUMA_NO_NODE
64 static bool devices_handle_discard_safely = false;
65 module_param(devices_handle_discard_safely, bool, 0644);
66 MODULE_PARM_DESC(devices_handle_discard_safely,
67 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
68 static struct workqueue_struct *raid5_wq;
70 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
72 int hash = (sect >> RAID5_STRIPE_SHIFT(conf)) & HASH_MASK;
73 return &conf->stripe_hashtbl[hash];
76 static inline int stripe_hash_locks_hash(struct r5conf *conf, sector_t sect)
78 return (sect >> RAID5_STRIPE_SHIFT(conf)) & STRIPE_HASH_LOCKS_MASK;
81 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
83 spin_lock_irq(conf->hash_locks + hash);
84 spin_lock(&conf->device_lock);
87 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
89 spin_unlock(&conf->device_lock);
90 spin_unlock_irq(conf->hash_locks + hash);
93 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
96 spin_lock_irq(conf->hash_locks);
97 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
98 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
99 spin_lock(&conf->device_lock);
102 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
105 spin_unlock(&conf->device_lock);
106 for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
107 spin_unlock(conf->hash_locks + i);
108 spin_unlock_irq(conf->hash_locks);
111 /* Find first data disk in a raid6 stripe */
112 static inline int raid6_d0(struct stripe_head *sh)
115 /* ddf always start from first device */
117 /* md starts just after Q block */
118 if (sh->qd_idx == sh->disks - 1)
121 return sh->qd_idx + 1;
123 static inline int raid6_next_disk(int disk, int raid_disks)
126 return (disk < raid_disks) ? disk : 0;
129 /* When walking through the disks in a raid5, starting at raid6_d0,
130 * We need to map each disk to a 'slot', where the data disks are slot
131 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
132 * is raid_disks-1. This help does that mapping.
134 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
135 int *count, int syndrome_disks)
141 if (idx == sh->pd_idx)
142 return syndrome_disks;
143 if (idx == sh->qd_idx)
144 return syndrome_disks + 1;
150 static void print_raid5_conf (struct r5conf *conf);
152 static int stripe_operations_active(struct stripe_head *sh)
154 return sh->check_state || sh->reconstruct_state ||
155 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
156 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
159 static bool stripe_is_lowprio(struct stripe_head *sh)
161 return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
162 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
163 !test_bit(STRIPE_R5C_CACHING, &sh->state);
166 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
168 struct r5conf *conf = sh->raid_conf;
169 struct r5worker_group *group;
171 int i, cpu = sh->cpu;
173 if (!cpu_online(cpu)) {
174 cpu = cpumask_any(cpu_online_mask);
178 if (list_empty(&sh->lru)) {
179 struct r5worker_group *group;
180 group = conf->worker_groups + cpu_to_group(cpu);
181 if (stripe_is_lowprio(sh))
182 list_add_tail(&sh->lru, &group->loprio_list);
184 list_add_tail(&sh->lru, &group->handle_list);
185 group->stripes_cnt++;
189 if (conf->worker_cnt_per_group == 0) {
190 md_wakeup_thread(conf->mddev->thread);
194 group = conf->worker_groups + cpu_to_group(sh->cpu);
196 group->workers[0].working = true;
197 /* at least one worker should run to avoid race */
198 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
200 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
201 /* wakeup more workers */
202 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
203 if (group->workers[i].working == false) {
204 group->workers[i].working = true;
205 queue_work_on(sh->cpu, raid5_wq,
206 &group->workers[i].work);
212 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
213 struct list_head *temp_inactive_list)
216 int injournal = 0; /* number of date pages with R5_InJournal */
218 BUG_ON(!list_empty(&sh->lru));
219 BUG_ON(atomic_read(&conf->active_stripes)==0);
221 if (r5c_is_writeback(conf->log))
222 for (i = sh->disks; i--; )
223 if (test_bit(R5_InJournal, &sh->dev[i].flags))
226 * In the following cases, the stripe cannot be released to cached
227 * lists. Therefore, we make the stripe write out and set
229 * 1. when quiesce in r5c write back;
230 * 2. when resync is requested fot the stripe.
232 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
233 (conf->quiesce && r5c_is_writeback(conf->log) &&
234 !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
235 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
236 r5c_make_stripe_write_out(sh);
237 set_bit(STRIPE_HANDLE, &sh->state);
240 if (test_bit(STRIPE_HANDLE, &sh->state)) {
241 if (test_bit(STRIPE_DELAYED, &sh->state) &&
242 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
243 list_add_tail(&sh->lru, &conf->delayed_list);
244 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
245 sh->bm_seq - conf->seq_write > 0)
246 list_add_tail(&sh->lru, &conf->bitmap_list);
248 clear_bit(STRIPE_DELAYED, &sh->state);
249 clear_bit(STRIPE_BIT_DELAY, &sh->state);
250 if (conf->worker_cnt_per_group == 0) {
251 if (stripe_is_lowprio(sh))
252 list_add_tail(&sh->lru,
255 list_add_tail(&sh->lru,
258 raid5_wakeup_stripe_thread(sh);
262 md_wakeup_thread(conf->mddev->thread);
264 BUG_ON(stripe_operations_active(sh));
265 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
266 if (atomic_dec_return(&conf->preread_active_stripes)
268 md_wakeup_thread(conf->mddev->thread);
269 atomic_dec(&conf->active_stripes);
270 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
271 if (!r5c_is_writeback(conf->log))
272 list_add_tail(&sh->lru, temp_inactive_list);
274 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
276 list_add_tail(&sh->lru, temp_inactive_list);
277 else if (injournal == conf->raid_disks - conf->max_degraded) {
279 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
280 atomic_inc(&conf->r5c_cached_full_stripes);
281 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
282 atomic_dec(&conf->r5c_cached_partial_stripes);
283 list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
284 r5c_check_cached_full_stripe(conf);
287 * STRIPE_R5C_PARTIAL_STRIPE is set in
288 * r5c_try_caching_write(). No need to
291 list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
297 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
298 struct list_head *temp_inactive_list)
300 if (atomic_dec_and_test(&sh->count))
301 do_release_stripe(conf, sh, temp_inactive_list);
305 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
307 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
308 * given time. Adding stripes only takes device lock, while deleting stripes
309 * only takes hash lock.
311 static void release_inactive_stripe_list(struct r5conf *conf,
312 struct list_head *temp_inactive_list,
316 bool do_wakeup = false;
319 if (hash == NR_STRIPE_HASH_LOCKS) {
320 size = NR_STRIPE_HASH_LOCKS;
321 hash = NR_STRIPE_HASH_LOCKS - 1;
325 struct list_head *list = &temp_inactive_list[size - 1];
328 * We don't hold any lock here yet, raid5_get_active_stripe() might
329 * remove stripes from the list
331 if (!list_empty_careful(list)) {
332 spin_lock_irqsave(conf->hash_locks + hash, flags);
333 if (list_empty(conf->inactive_list + hash) &&
335 atomic_dec(&conf->empty_inactive_list_nr);
336 list_splice_tail_init(list, conf->inactive_list + hash);
338 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
345 wake_up(&conf->wait_for_stripe);
346 if (atomic_read(&conf->active_stripes) == 0)
347 wake_up(&conf->wait_for_quiescent);
348 if (conf->retry_read_aligned)
349 md_wakeup_thread(conf->mddev->thread);
353 /* should hold conf->device_lock already */
354 static int release_stripe_list(struct r5conf *conf,
355 struct list_head *temp_inactive_list)
357 struct stripe_head *sh, *t;
359 struct llist_node *head;
361 head = llist_del_all(&conf->released_stripes);
362 head = llist_reverse_order(head);
363 llist_for_each_entry_safe(sh, t, head, release_list) {
366 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
368 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
370 * Don't worry the bit is set here, because if the bit is set
371 * again, the count is always > 1. This is true for
372 * STRIPE_ON_UNPLUG_LIST bit too.
374 hash = sh->hash_lock_index;
375 __release_stripe(conf, sh, &temp_inactive_list[hash]);
382 void raid5_release_stripe(struct stripe_head *sh)
384 struct r5conf *conf = sh->raid_conf;
386 struct list_head list;
390 /* Avoid release_list until the last reference.
392 if (atomic_add_unless(&sh->count, -1, 1))
395 if (unlikely(!conf->mddev->thread) ||
396 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
398 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
400 md_wakeup_thread(conf->mddev->thread);
403 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
404 if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
405 INIT_LIST_HEAD(&list);
406 hash = sh->hash_lock_index;
407 do_release_stripe(conf, sh, &list);
408 spin_unlock_irqrestore(&conf->device_lock, flags);
409 release_inactive_stripe_list(conf, &list, hash);
413 static inline void remove_hash(struct stripe_head *sh)
415 pr_debug("remove_hash(), stripe %llu\n",
416 (unsigned long long)sh->sector);
418 hlist_del_init(&sh->hash);
421 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
423 struct hlist_head *hp = stripe_hash(conf, sh->sector);
425 pr_debug("insert_hash(), stripe %llu\n",
426 (unsigned long long)sh->sector);
428 hlist_add_head(&sh->hash, hp);
431 /* find an idle stripe, make sure it is unhashed, and return it. */
432 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
434 struct stripe_head *sh = NULL;
435 struct list_head *first;
437 if (list_empty(conf->inactive_list + hash))
439 first = (conf->inactive_list + hash)->next;
440 sh = list_entry(first, struct stripe_head, lru);
441 list_del_init(first);
443 atomic_inc(&conf->active_stripes);
444 BUG_ON(hash != sh->hash_lock_index);
445 if (list_empty(conf->inactive_list + hash))
446 atomic_inc(&conf->empty_inactive_list_nr);
451 static void shrink_buffers(struct stripe_head *sh)
455 int num = sh->raid_conf->pool_size;
457 for (i = 0; i < num ; i++) {
458 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
462 sh->dev[i].page = NULL;
467 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
470 int num = sh->raid_conf->pool_size;
472 for (i = 0; i < num; i++) {
475 if (!(page = alloc_page(gfp))) {
478 sh->dev[i].page = page;
479 sh->dev[i].orig_page = page;
485 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
486 struct stripe_head *sh);
488 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
490 struct r5conf *conf = sh->raid_conf;
493 BUG_ON(atomic_read(&sh->count) != 0);
494 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
495 BUG_ON(stripe_operations_active(sh));
496 BUG_ON(sh->batch_head);
498 pr_debug("init_stripe called, stripe %llu\n",
499 (unsigned long long)sector);
501 seq = read_seqcount_begin(&conf->gen_lock);
502 sh->generation = conf->generation - previous;
503 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
505 stripe_set_idx(sector, conf, previous, sh);
508 for (i = sh->disks; i--; ) {
509 struct r5dev *dev = &sh->dev[i];
511 if (dev->toread || dev->read || dev->towrite || dev->written ||
512 test_bit(R5_LOCKED, &dev->flags)) {
513 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
514 (unsigned long long)sh->sector, i, dev->toread,
515 dev->read, dev->towrite, dev->written,
516 test_bit(R5_LOCKED, &dev->flags));
520 dev->sector = raid5_compute_blocknr(sh, i, previous);
522 if (read_seqcount_retry(&conf->gen_lock, seq))
524 sh->overwrite_disks = 0;
525 insert_hash(conf, sh);
526 sh->cpu = smp_processor_id();
527 set_bit(STRIPE_BATCH_READY, &sh->state);
530 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
533 struct stripe_head *sh;
535 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
536 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
537 if (sh->sector == sector && sh->generation == generation)
539 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
544 * Need to check if array has failed when deciding whether to:
546 * - remove non-faulty devices
549 * This determination is simple when no reshape is happening.
550 * However if there is a reshape, we need to carefully check
551 * both the before and after sections.
552 * This is because some failed devices may only affect one
553 * of the two sections, and some non-in_sync devices may
554 * be insync in the section most affected by failed devices.
556 int raid5_calc_degraded(struct r5conf *conf)
558 int degraded, degraded2;
563 for (i = 0; i < conf->previous_raid_disks; i++) {
564 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
565 if (rdev && test_bit(Faulty, &rdev->flags))
566 rdev = rcu_dereference(conf->disks[i].replacement);
567 if (!rdev || test_bit(Faulty, &rdev->flags))
569 else if (test_bit(In_sync, &rdev->flags))
572 /* not in-sync or faulty.
573 * If the reshape increases the number of devices,
574 * this is being recovered by the reshape, so
575 * this 'previous' section is not in_sync.
576 * If the number of devices is being reduced however,
577 * the device can only be part of the array if
578 * we are reverting a reshape, so this section will
581 if (conf->raid_disks >= conf->previous_raid_disks)
585 if (conf->raid_disks == conf->previous_raid_disks)
589 for (i = 0; i < conf->raid_disks; i++) {
590 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
591 if (rdev && test_bit(Faulty, &rdev->flags))
592 rdev = rcu_dereference(conf->disks[i].replacement);
593 if (!rdev || test_bit(Faulty, &rdev->flags))
595 else if (test_bit(In_sync, &rdev->flags))
598 /* not in-sync or faulty.
599 * If reshape increases the number of devices, this
600 * section has already been recovered, else it
601 * almost certainly hasn't.
603 if (conf->raid_disks <= conf->previous_raid_disks)
607 if (degraded2 > degraded)
612 static int has_failed(struct r5conf *conf)
616 if (conf->mddev->reshape_position == MaxSector)
617 return conf->mddev->degraded > conf->max_degraded;
619 degraded = raid5_calc_degraded(conf);
620 if (degraded > conf->max_degraded)
626 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
627 int previous, int noblock, int noquiesce)
629 struct stripe_head *sh;
630 int hash = stripe_hash_locks_hash(conf, sector);
631 int inc_empty_inactive_list_flag;
633 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
635 spin_lock_irq(conf->hash_locks + hash);
638 wait_event_lock_irq(conf->wait_for_quiescent,
639 conf->quiesce == 0 || noquiesce,
640 *(conf->hash_locks + hash));
641 sh = __find_stripe(conf, sector, conf->generation - previous);
643 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
644 sh = get_free_stripe(conf, hash);
645 if (!sh && !test_bit(R5_DID_ALLOC,
647 set_bit(R5_ALLOC_MORE,
650 if (noblock && sh == NULL)
653 r5c_check_stripe_cache_usage(conf);
655 set_bit(R5_INACTIVE_BLOCKED,
657 r5l_wake_reclaim(conf->log, 0);
659 conf->wait_for_stripe,
660 !list_empty(conf->inactive_list + hash) &&
661 (atomic_read(&conf->active_stripes)
662 < (conf->max_nr_stripes * 3 / 4)
663 || !test_bit(R5_INACTIVE_BLOCKED,
664 &conf->cache_state)),
665 *(conf->hash_locks + hash));
666 clear_bit(R5_INACTIVE_BLOCKED,
669 init_stripe(sh, sector, previous);
670 atomic_inc(&sh->count);
672 } else if (!atomic_inc_not_zero(&sh->count)) {
673 spin_lock(&conf->device_lock);
674 if (!atomic_read(&sh->count)) {
675 if (!test_bit(STRIPE_HANDLE, &sh->state))
676 atomic_inc(&conf->active_stripes);
677 BUG_ON(list_empty(&sh->lru) &&
678 !test_bit(STRIPE_EXPANDING, &sh->state));
679 inc_empty_inactive_list_flag = 0;
680 if (!list_empty(conf->inactive_list + hash))
681 inc_empty_inactive_list_flag = 1;
682 list_del_init(&sh->lru);
683 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
684 atomic_inc(&conf->empty_inactive_list_nr);
686 sh->group->stripes_cnt--;
690 atomic_inc(&sh->count);
691 spin_unlock(&conf->device_lock);
693 } while (sh == NULL);
695 spin_unlock_irq(conf->hash_locks + hash);
699 static bool is_full_stripe_write(struct stripe_head *sh)
701 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
702 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
705 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
706 __acquires(&sh1->stripe_lock)
707 __acquires(&sh2->stripe_lock)
710 spin_lock_irq(&sh2->stripe_lock);
711 spin_lock_nested(&sh1->stripe_lock, 1);
713 spin_lock_irq(&sh1->stripe_lock);
714 spin_lock_nested(&sh2->stripe_lock, 1);
718 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
719 __releases(&sh1->stripe_lock)
720 __releases(&sh2->stripe_lock)
722 spin_unlock(&sh1->stripe_lock);
723 spin_unlock_irq(&sh2->stripe_lock);
726 /* Only freshly new full stripe normal write stripe can be added to a batch list */
727 static bool stripe_can_batch(struct stripe_head *sh)
729 struct r5conf *conf = sh->raid_conf;
731 if (raid5_has_log(conf) || raid5_has_ppl(conf))
733 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
734 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
735 is_full_stripe_write(sh);
738 /* we only do back search */
739 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
741 struct stripe_head *head;
742 sector_t head_sector, tmp_sec;
745 int inc_empty_inactive_list_flag;
747 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
748 tmp_sec = sh->sector;
749 if (!sector_div(tmp_sec, conf->chunk_sectors))
751 head_sector = sh->sector - RAID5_STRIPE_SECTORS(conf);
753 hash = stripe_hash_locks_hash(conf, head_sector);
754 spin_lock_irq(conf->hash_locks + hash);
755 head = __find_stripe(conf, head_sector, conf->generation);
756 if (head && !atomic_inc_not_zero(&head->count)) {
757 spin_lock(&conf->device_lock);
758 if (!atomic_read(&head->count)) {
759 if (!test_bit(STRIPE_HANDLE, &head->state))
760 atomic_inc(&conf->active_stripes);
761 BUG_ON(list_empty(&head->lru) &&
762 !test_bit(STRIPE_EXPANDING, &head->state));
763 inc_empty_inactive_list_flag = 0;
764 if (!list_empty(conf->inactive_list + hash))
765 inc_empty_inactive_list_flag = 1;
766 list_del_init(&head->lru);
767 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
768 atomic_inc(&conf->empty_inactive_list_nr);
770 head->group->stripes_cnt--;
774 atomic_inc(&head->count);
775 spin_unlock(&conf->device_lock);
777 spin_unlock_irq(conf->hash_locks + hash);
781 if (!stripe_can_batch(head))
784 lock_two_stripes(head, sh);
785 /* clear_batch_ready clear the flag */
786 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
793 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
795 if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
796 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
799 if (head->batch_head) {
800 spin_lock(&head->batch_head->batch_lock);
801 /* This batch list is already running */
802 if (!stripe_can_batch(head)) {
803 spin_unlock(&head->batch_head->batch_lock);
807 * We must assign batch_head of this stripe within the
808 * batch_lock, otherwise clear_batch_ready of batch head
809 * stripe could clear BATCH_READY bit of this stripe and
810 * this stripe->batch_head doesn't get assigned, which
811 * could confuse clear_batch_ready for this stripe
813 sh->batch_head = head->batch_head;
816 * at this point, head's BATCH_READY could be cleared, but we
817 * can still add the stripe to batch list
819 list_add(&sh->batch_list, &head->batch_list);
820 spin_unlock(&head->batch_head->batch_lock);
822 head->batch_head = head;
823 sh->batch_head = head->batch_head;
824 spin_lock(&head->batch_lock);
825 list_add_tail(&sh->batch_list, &head->batch_list);
826 spin_unlock(&head->batch_lock);
829 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
830 if (atomic_dec_return(&conf->preread_active_stripes)
832 md_wakeup_thread(conf->mddev->thread);
834 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
835 int seq = sh->bm_seq;
836 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
837 sh->batch_head->bm_seq > seq)
838 seq = sh->batch_head->bm_seq;
839 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
840 sh->batch_head->bm_seq = seq;
843 atomic_inc(&sh->count);
845 unlock_two_stripes(head, sh);
847 raid5_release_stripe(head);
850 /* Determine if 'data_offset' or 'new_data_offset' should be used
851 * in this stripe_head.
853 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
855 sector_t progress = conf->reshape_progress;
856 /* Need a memory barrier to make sure we see the value
857 * of conf->generation, or ->data_offset that was set before
858 * reshape_progress was updated.
861 if (progress == MaxSector)
863 if (sh->generation == conf->generation - 1)
865 /* We are in a reshape, and this is a new-generation stripe,
866 * so use new_data_offset.
871 static void dispatch_bio_list(struct bio_list *tmp)
875 while ((bio = bio_list_pop(tmp)))
876 submit_bio_noacct(bio);
879 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
881 const struct r5pending_data *da = list_entry(a,
882 struct r5pending_data, sibling);
883 const struct r5pending_data *db = list_entry(b,
884 struct r5pending_data, sibling);
885 if (da->sector > db->sector)
887 if (da->sector < db->sector)
892 static void dispatch_defer_bios(struct r5conf *conf, int target,
893 struct bio_list *list)
895 struct r5pending_data *data;
896 struct list_head *first, *next = NULL;
899 if (conf->pending_data_cnt == 0)
902 list_sort(NULL, &conf->pending_list, cmp_stripe);
904 first = conf->pending_list.next;
906 /* temporarily move the head */
907 if (conf->next_pending_data)
908 list_move_tail(&conf->pending_list,
909 &conf->next_pending_data->sibling);
911 while (!list_empty(&conf->pending_list)) {
912 data = list_first_entry(&conf->pending_list,
913 struct r5pending_data, sibling);
914 if (&data->sibling == first)
915 first = data->sibling.next;
916 next = data->sibling.next;
918 bio_list_merge(list, &data->bios);
919 list_move(&data->sibling, &conf->free_list);
924 conf->pending_data_cnt -= cnt;
925 BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
927 if (next != &conf->pending_list)
928 conf->next_pending_data = list_entry(next,
929 struct r5pending_data, sibling);
931 conf->next_pending_data = NULL;
932 /* list isn't empty */
933 if (first != &conf->pending_list)
934 list_move_tail(&conf->pending_list, first);
937 static void flush_deferred_bios(struct r5conf *conf)
939 struct bio_list tmp = BIO_EMPTY_LIST;
941 if (conf->pending_data_cnt == 0)
944 spin_lock(&conf->pending_bios_lock);
945 dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
946 BUG_ON(conf->pending_data_cnt != 0);
947 spin_unlock(&conf->pending_bios_lock);
949 dispatch_bio_list(&tmp);
952 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
953 struct bio_list *bios)
955 struct bio_list tmp = BIO_EMPTY_LIST;
956 struct r5pending_data *ent;
958 spin_lock(&conf->pending_bios_lock);
959 ent = list_first_entry(&conf->free_list, struct r5pending_data,
961 list_move_tail(&ent->sibling, &conf->pending_list);
962 ent->sector = sector;
963 bio_list_init(&ent->bios);
964 bio_list_merge(&ent->bios, bios);
965 conf->pending_data_cnt++;
966 if (conf->pending_data_cnt >= PENDING_IO_MAX)
967 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
969 spin_unlock(&conf->pending_bios_lock);
971 dispatch_bio_list(&tmp);
975 raid5_end_read_request(struct bio *bi);
977 raid5_end_write_request(struct bio *bi);
979 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
981 struct r5conf *conf = sh->raid_conf;
982 int i, disks = sh->disks;
983 struct stripe_head *head_sh = sh;
984 struct bio_list pending_bios = BIO_EMPTY_LIST;
989 if (log_stripe(sh, s) == 0)
992 should_defer = conf->batch_bio_dispatch && conf->group_cnt;
994 for (i = disks; i--; ) {
995 int op, op_flags = 0;
996 int replace_only = 0;
997 struct bio *bi, *rbi;
998 struct md_rdev *rdev, *rrdev = NULL;
1001 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1003 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1005 if (test_bit(R5_Discard, &sh->dev[i].flags))
1006 op = REQ_OP_DISCARD;
1007 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1009 else if (test_and_clear_bit(R5_WantReplace,
1010 &sh->dev[i].flags)) {
1015 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1016 op_flags |= REQ_SYNC;
1019 bi = &sh->dev[i].req;
1020 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1023 rrdev = rcu_dereference(conf->disks[i].replacement);
1024 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1025 rdev = rcu_dereference(conf->disks[i].rdev);
1030 if (op_is_write(op)) {
1034 /* We raced and saw duplicates */
1037 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1042 if (rdev && test_bit(Faulty, &rdev->flags))
1045 atomic_inc(&rdev->nr_pending);
1046 if (rrdev && test_bit(Faulty, &rrdev->flags))
1049 atomic_inc(&rrdev->nr_pending);
1052 /* We have already checked bad blocks for reads. Now
1053 * need to check for writes. We never accept write errors
1054 * on the replacement, so we don't to check rrdev.
1056 while (op_is_write(op) && rdev &&
1057 test_bit(WriteErrorSeen, &rdev->flags)) {
1060 int bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
1061 &first_bad, &bad_sectors);
1066 set_bit(BlockedBadBlocks, &rdev->flags);
1067 if (!conf->mddev->external &&
1068 conf->mddev->sb_flags) {
1069 /* It is very unlikely, but we might
1070 * still need to write out the
1071 * bad block log - better give it
1073 md_check_recovery(conf->mddev);
1076 * Because md_wait_for_blocked_rdev
1077 * will dec nr_pending, we must
1078 * increment it first.
1080 atomic_inc(&rdev->nr_pending);
1081 md_wait_for_blocked_rdev(rdev, conf->mddev);
1083 /* Acknowledged bad block - skip the write */
1084 rdev_dec_pending(rdev, conf->mddev);
1090 if (s->syncing || s->expanding || s->expanded
1092 md_sync_acct(rdev->bdev, RAID5_STRIPE_SECTORS(conf));
1094 set_bit(STRIPE_IO_STARTED, &sh->state);
1096 bio_set_dev(bi, rdev->bdev);
1097 bio_set_op_attrs(bi, op, op_flags);
1098 bi->bi_end_io = op_is_write(op)
1099 ? raid5_end_write_request
1100 : raid5_end_read_request;
1101 bi->bi_private = sh;
1103 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1104 __func__, (unsigned long long)sh->sector,
1106 atomic_inc(&sh->count);
1108 atomic_inc(&head_sh->count);
1109 if (use_new_offset(conf, sh))
1110 bi->bi_iter.bi_sector = (sh->sector
1111 + rdev->new_data_offset);
1113 bi->bi_iter.bi_sector = (sh->sector
1114 + rdev->data_offset);
1115 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1116 bi->bi_opf |= REQ_NOMERGE;
1118 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1119 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1121 if (!op_is_write(op) &&
1122 test_bit(R5_InJournal, &sh->dev[i].flags))
1124 * issuing read for a page in journal, this
1125 * must be preparing for prexor in rmw; read
1126 * the data into orig_page
1128 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1130 sh->dev[i].vec.bv_page = sh->dev[i].page;
1132 bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1133 bi->bi_io_vec[0].bv_offset = 0;
1134 bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1135 bi->bi_write_hint = sh->dev[i].write_hint;
1137 sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1139 * If this is discard request, set bi_vcnt 0. We don't
1140 * want to confuse SCSI because SCSI will replace payload
1142 if (op == REQ_OP_DISCARD)
1145 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1147 if (conf->mddev->gendisk)
1148 trace_block_bio_remap(bi->bi_disk->queue,
1149 bi, disk_devt(conf->mddev->gendisk),
1151 if (should_defer && op_is_write(op))
1152 bio_list_add(&pending_bios, bi);
1154 submit_bio_noacct(bi);
1157 if (s->syncing || s->expanding || s->expanded
1159 md_sync_acct(rrdev->bdev, RAID5_STRIPE_SECTORS(conf));
1161 set_bit(STRIPE_IO_STARTED, &sh->state);
1163 bio_set_dev(rbi, rrdev->bdev);
1164 bio_set_op_attrs(rbi, op, op_flags);
1165 BUG_ON(!op_is_write(op));
1166 rbi->bi_end_io = raid5_end_write_request;
1167 rbi->bi_private = sh;
1169 pr_debug("%s: for %llu schedule op %d on "
1170 "replacement disc %d\n",
1171 __func__, (unsigned long long)sh->sector,
1173 atomic_inc(&sh->count);
1175 atomic_inc(&head_sh->count);
1176 if (use_new_offset(conf, sh))
1177 rbi->bi_iter.bi_sector = (sh->sector
1178 + rrdev->new_data_offset);
1180 rbi->bi_iter.bi_sector = (sh->sector
1181 + rrdev->data_offset);
1182 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1183 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1184 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1186 rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1187 rbi->bi_io_vec[0].bv_offset = 0;
1188 rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1189 rbi->bi_write_hint = sh->dev[i].write_hint;
1190 sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1192 * If this is discard request, set bi_vcnt 0. We don't
1193 * want to confuse SCSI because SCSI will replace payload
1195 if (op == REQ_OP_DISCARD)
1197 if (conf->mddev->gendisk)
1198 trace_block_bio_remap(rbi->bi_disk->queue,
1199 rbi, disk_devt(conf->mddev->gendisk),
1201 if (should_defer && op_is_write(op))
1202 bio_list_add(&pending_bios, rbi);
1204 submit_bio_noacct(rbi);
1206 if (!rdev && !rrdev) {
1207 if (op_is_write(op))
1208 set_bit(STRIPE_DEGRADED, &sh->state);
1209 pr_debug("skip op %d on disc %d for sector %llu\n",
1210 bi->bi_opf, i, (unsigned long long)sh->sector);
1211 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1212 set_bit(STRIPE_HANDLE, &sh->state);
1215 if (!head_sh->batch_head)
1217 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1223 if (should_defer && !bio_list_empty(&pending_bios))
1224 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1227 static struct dma_async_tx_descriptor *
1228 async_copy_data(int frombio, struct bio *bio, struct page **page,
1229 sector_t sector, struct dma_async_tx_descriptor *tx,
1230 struct stripe_head *sh, int no_skipcopy)
1233 struct bvec_iter iter;
1234 struct page *bio_page;
1236 struct async_submit_ctl submit;
1237 enum async_tx_flags flags = 0;
1238 struct r5conf *conf = sh->raid_conf;
1240 if (bio->bi_iter.bi_sector >= sector)
1241 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1243 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1246 flags |= ASYNC_TX_FENCE;
1247 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1249 bio_for_each_segment(bvl, bio, iter) {
1250 int len = bvl.bv_len;
1254 if (page_offset < 0) {
1255 b_offset = -page_offset;
1256 page_offset += b_offset;
1260 if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
1261 clen = RAID5_STRIPE_SIZE(conf) - page_offset;
1266 b_offset += bvl.bv_offset;
1267 bio_page = bvl.bv_page;
1269 if (conf->skip_copy &&
1270 b_offset == 0 && page_offset == 0 &&
1271 clen == RAID5_STRIPE_SIZE(conf) &&
1275 tx = async_memcpy(*page, bio_page, page_offset,
1276 b_offset, clen, &submit);
1278 tx = async_memcpy(bio_page, *page, b_offset,
1279 page_offset, clen, &submit);
1281 /* chain the operations */
1282 submit.depend_tx = tx;
1284 if (clen < len) /* hit end of page */
1292 static void ops_complete_biofill(void *stripe_head_ref)
1294 struct stripe_head *sh = stripe_head_ref;
1296 struct r5conf *conf = sh->raid_conf;
1298 pr_debug("%s: stripe %llu\n", __func__,
1299 (unsigned long long)sh->sector);
1301 /* clear completed biofills */
1302 for (i = sh->disks; i--; ) {
1303 struct r5dev *dev = &sh->dev[i];
1305 /* acknowledge completion of a biofill operation */
1306 /* and check if we need to reply to a read request,
1307 * new R5_Wantfill requests are held off until
1308 * !STRIPE_BIOFILL_RUN
1310 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1311 struct bio *rbi, *rbi2;
1316 while (rbi && rbi->bi_iter.bi_sector <
1317 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1318 rbi2 = r5_next_bio(conf, rbi, dev->sector);
1324 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1326 set_bit(STRIPE_HANDLE, &sh->state);
1327 raid5_release_stripe(sh);
1330 static void ops_run_biofill(struct stripe_head *sh)
1332 struct dma_async_tx_descriptor *tx = NULL;
1333 struct async_submit_ctl submit;
1335 struct r5conf *conf = sh->raid_conf;
1337 BUG_ON(sh->batch_head);
1338 pr_debug("%s: stripe %llu\n", __func__,
1339 (unsigned long long)sh->sector);
1341 for (i = sh->disks; i--; ) {
1342 struct r5dev *dev = &sh->dev[i];
1343 if (test_bit(R5_Wantfill, &dev->flags)) {
1345 spin_lock_irq(&sh->stripe_lock);
1346 dev->read = rbi = dev->toread;
1348 spin_unlock_irq(&sh->stripe_lock);
1349 while (rbi && rbi->bi_iter.bi_sector <
1350 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1351 tx = async_copy_data(0, rbi, &dev->page,
1352 dev->sector, tx, sh, 0);
1353 rbi = r5_next_bio(conf, rbi, dev->sector);
1358 atomic_inc(&sh->count);
1359 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1360 async_trigger_callback(&submit);
1363 static void mark_target_uptodate(struct stripe_head *sh, int target)
1370 tgt = &sh->dev[target];
1371 set_bit(R5_UPTODATE, &tgt->flags);
1372 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1373 clear_bit(R5_Wantcompute, &tgt->flags);
1376 static void ops_complete_compute(void *stripe_head_ref)
1378 struct stripe_head *sh = stripe_head_ref;
1380 pr_debug("%s: stripe %llu\n", __func__,
1381 (unsigned long long)sh->sector);
1383 /* mark the computed target(s) as uptodate */
1384 mark_target_uptodate(sh, sh->ops.target);
1385 mark_target_uptodate(sh, sh->ops.target2);
1387 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1388 if (sh->check_state == check_state_compute_run)
1389 sh->check_state = check_state_compute_result;
1390 set_bit(STRIPE_HANDLE, &sh->state);
1391 raid5_release_stripe(sh);
1394 /* return a pointer to the address conversion region of the scribble buffer */
1395 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1397 return percpu->scribble + i * percpu->scribble_obj_size;
1400 /* return a pointer to the address conversion region of the scribble buffer */
1401 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1402 struct raid5_percpu *percpu, int i)
1404 return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1407 static struct dma_async_tx_descriptor *
1408 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1410 int disks = sh->disks;
1411 struct page **xor_srcs = to_addr_page(percpu, 0);
1412 int target = sh->ops.target;
1413 struct r5dev *tgt = &sh->dev[target];
1414 struct page *xor_dest = tgt->page;
1416 struct dma_async_tx_descriptor *tx;
1417 struct async_submit_ctl submit;
1420 BUG_ON(sh->batch_head);
1422 pr_debug("%s: stripe %llu block: %d\n",
1423 __func__, (unsigned long long)sh->sector, target);
1424 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1426 for (i = disks; i--; )
1428 xor_srcs[count++] = sh->dev[i].page;
1430 atomic_inc(&sh->count);
1432 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1433 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1434 if (unlikely(count == 1))
1435 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0,
1436 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1438 tx = async_xor(xor_dest, xor_srcs, 0, count,
1439 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1444 /* set_syndrome_sources - populate source buffers for gen_syndrome
1445 * @srcs - (struct page *) array of size sh->disks
1446 * @sh - stripe_head to parse
1448 * Populates srcs in proper layout order for the stripe and returns the
1449 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1450 * destination buffer is recorded in srcs[count] and the Q destination
1451 * is recorded in srcs[count+1]].
1453 static int set_syndrome_sources(struct page **srcs,
1454 struct stripe_head *sh,
1457 int disks = sh->disks;
1458 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1459 int d0_idx = raid6_d0(sh);
1463 for (i = 0; i < disks; i++)
1469 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1470 struct r5dev *dev = &sh->dev[i];
1472 if (i == sh->qd_idx || i == sh->pd_idx ||
1473 (srctype == SYNDROME_SRC_ALL) ||
1474 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1475 (test_bit(R5_Wantdrain, &dev->flags) ||
1476 test_bit(R5_InJournal, &dev->flags))) ||
1477 (srctype == SYNDROME_SRC_WRITTEN &&
1479 test_bit(R5_InJournal, &dev->flags)))) {
1480 if (test_bit(R5_InJournal, &dev->flags))
1481 srcs[slot] = sh->dev[i].orig_page;
1483 srcs[slot] = sh->dev[i].page;
1485 i = raid6_next_disk(i, disks);
1486 } while (i != d0_idx);
1488 return syndrome_disks;
1491 static struct dma_async_tx_descriptor *
1492 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1494 int disks = sh->disks;
1495 struct page **blocks = to_addr_page(percpu, 0);
1497 int qd_idx = sh->qd_idx;
1498 struct dma_async_tx_descriptor *tx;
1499 struct async_submit_ctl submit;
1505 BUG_ON(sh->batch_head);
1506 if (sh->ops.target < 0)
1507 target = sh->ops.target2;
1508 else if (sh->ops.target2 < 0)
1509 target = sh->ops.target;
1511 /* we should only have one valid target */
1514 pr_debug("%s: stripe %llu block: %d\n",
1515 __func__, (unsigned long long)sh->sector, target);
1517 tgt = &sh->dev[target];
1518 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1521 atomic_inc(&sh->count);
1523 if (target == qd_idx) {
1524 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1525 blocks[count] = NULL; /* regenerating p is not necessary */
1526 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1527 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1528 ops_complete_compute, sh,
1529 to_addr_conv(sh, percpu, 0));
1530 tx = async_gen_syndrome(blocks, 0, count+2,
1531 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1533 /* Compute any data- or p-drive using XOR */
1535 for (i = disks; i-- ; ) {
1536 if (i == target || i == qd_idx)
1538 blocks[count++] = sh->dev[i].page;
1541 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1542 NULL, ops_complete_compute, sh,
1543 to_addr_conv(sh, percpu, 0));
1544 tx = async_xor(dest, blocks, 0, count,
1545 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1551 static struct dma_async_tx_descriptor *
1552 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1554 int i, count, disks = sh->disks;
1555 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1556 int d0_idx = raid6_d0(sh);
1557 int faila = -1, failb = -1;
1558 int target = sh->ops.target;
1559 int target2 = sh->ops.target2;
1560 struct r5dev *tgt = &sh->dev[target];
1561 struct r5dev *tgt2 = &sh->dev[target2];
1562 struct dma_async_tx_descriptor *tx;
1563 struct page **blocks = to_addr_page(percpu, 0);
1564 struct async_submit_ctl submit;
1566 BUG_ON(sh->batch_head);
1567 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1568 __func__, (unsigned long long)sh->sector, target, target2);
1569 BUG_ON(target < 0 || target2 < 0);
1570 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1571 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1573 /* we need to open-code set_syndrome_sources to handle the
1574 * slot number conversion for 'faila' and 'failb'
1576 for (i = 0; i < disks ; i++)
1581 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1583 blocks[slot] = sh->dev[i].page;
1589 i = raid6_next_disk(i, disks);
1590 } while (i != d0_idx);
1592 BUG_ON(faila == failb);
1595 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1596 __func__, (unsigned long long)sh->sector, faila, failb);
1598 atomic_inc(&sh->count);
1600 if (failb == syndrome_disks+1) {
1601 /* Q disk is one of the missing disks */
1602 if (faila == syndrome_disks) {
1603 /* Missing P+Q, just recompute */
1604 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1605 ops_complete_compute, sh,
1606 to_addr_conv(sh, percpu, 0));
1607 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1608 RAID5_STRIPE_SIZE(sh->raid_conf),
1613 int qd_idx = sh->qd_idx;
1615 /* Missing D+Q: recompute D from P, then recompute Q */
1616 if (target == qd_idx)
1617 data_target = target2;
1619 data_target = target;
1622 for (i = disks; i-- ; ) {
1623 if (i == data_target || i == qd_idx)
1625 blocks[count++] = sh->dev[i].page;
1627 dest = sh->dev[data_target].page;
1628 init_async_submit(&submit,
1629 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1631 to_addr_conv(sh, percpu, 0));
1632 tx = async_xor(dest, blocks, 0, count,
1633 RAID5_STRIPE_SIZE(sh->raid_conf),
1636 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1637 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1638 ops_complete_compute, sh,
1639 to_addr_conv(sh, percpu, 0));
1640 return async_gen_syndrome(blocks, 0, count+2,
1641 RAID5_STRIPE_SIZE(sh->raid_conf),
1645 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1646 ops_complete_compute, sh,
1647 to_addr_conv(sh, percpu, 0));
1648 if (failb == syndrome_disks) {
1649 /* We're missing D+P. */
1650 return async_raid6_datap_recov(syndrome_disks+2,
1651 RAID5_STRIPE_SIZE(sh->raid_conf),
1655 /* We're missing D+D. */
1656 return async_raid6_2data_recov(syndrome_disks+2,
1657 RAID5_STRIPE_SIZE(sh->raid_conf),
1664 static void ops_complete_prexor(void *stripe_head_ref)
1666 struct stripe_head *sh = stripe_head_ref;
1668 pr_debug("%s: stripe %llu\n", __func__,
1669 (unsigned long long)sh->sector);
1671 if (r5c_is_writeback(sh->raid_conf->log))
1673 * raid5-cache write back uses orig_page during prexor.
1674 * After prexor, it is time to free orig_page
1676 r5c_release_extra_page(sh);
1679 static struct dma_async_tx_descriptor *
1680 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1681 struct dma_async_tx_descriptor *tx)
1683 int disks = sh->disks;
1684 struct page **xor_srcs = to_addr_page(percpu, 0);
1685 int count = 0, pd_idx = sh->pd_idx, i;
1686 struct async_submit_ctl submit;
1688 /* existing parity data subtracted */
1689 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1691 BUG_ON(sh->batch_head);
1692 pr_debug("%s: stripe %llu\n", __func__,
1693 (unsigned long long)sh->sector);
1695 for (i = disks; i--; ) {
1696 struct r5dev *dev = &sh->dev[i];
1697 /* Only process blocks that are known to be uptodate */
1698 if (test_bit(R5_InJournal, &dev->flags))
1699 xor_srcs[count++] = dev->orig_page;
1700 else if (test_bit(R5_Wantdrain, &dev->flags))
1701 xor_srcs[count++] = dev->page;
1704 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1705 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1706 tx = async_xor(xor_dest, xor_srcs, 0, count,
1707 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1712 static struct dma_async_tx_descriptor *
1713 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1714 struct dma_async_tx_descriptor *tx)
1716 struct page **blocks = to_addr_page(percpu, 0);
1718 struct async_submit_ctl submit;
1720 pr_debug("%s: stripe %llu\n", __func__,
1721 (unsigned long long)sh->sector);
1723 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1725 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1726 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1727 tx = async_gen_syndrome(blocks, 0, count+2,
1728 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1733 static struct dma_async_tx_descriptor *
1734 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1736 struct r5conf *conf = sh->raid_conf;
1737 int disks = sh->disks;
1739 struct stripe_head *head_sh = sh;
1741 pr_debug("%s: stripe %llu\n", __func__,
1742 (unsigned long long)sh->sector);
1744 for (i = disks; i--; ) {
1749 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1755 * clear R5_InJournal, so when rewriting a page in
1756 * journal, it is not skipped by r5l_log_stripe()
1758 clear_bit(R5_InJournal, &dev->flags);
1759 spin_lock_irq(&sh->stripe_lock);
1760 chosen = dev->towrite;
1761 dev->towrite = NULL;
1762 sh->overwrite_disks = 0;
1763 BUG_ON(dev->written);
1764 wbi = dev->written = chosen;
1765 spin_unlock_irq(&sh->stripe_lock);
1766 WARN_ON(dev->page != dev->orig_page);
1768 while (wbi && wbi->bi_iter.bi_sector <
1769 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1770 if (wbi->bi_opf & REQ_FUA)
1771 set_bit(R5_WantFUA, &dev->flags);
1772 if (wbi->bi_opf & REQ_SYNC)
1773 set_bit(R5_SyncIO, &dev->flags);
1774 if (bio_op(wbi) == REQ_OP_DISCARD)
1775 set_bit(R5_Discard, &dev->flags);
1777 tx = async_copy_data(1, wbi, &dev->page,
1778 dev->sector, tx, sh,
1779 r5c_is_writeback(conf->log));
1780 if (dev->page != dev->orig_page &&
1781 !r5c_is_writeback(conf->log)) {
1782 set_bit(R5_SkipCopy, &dev->flags);
1783 clear_bit(R5_UPTODATE, &dev->flags);
1784 clear_bit(R5_OVERWRITE, &dev->flags);
1787 wbi = r5_next_bio(conf, wbi, dev->sector);
1790 if (head_sh->batch_head) {
1791 sh = list_first_entry(&sh->batch_list,
1804 static void ops_complete_reconstruct(void *stripe_head_ref)
1806 struct stripe_head *sh = stripe_head_ref;
1807 int disks = sh->disks;
1808 int pd_idx = sh->pd_idx;
1809 int qd_idx = sh->qd_idx;
1811 bool fua = false, sync = false, discard = false;
1813 pr_debug("%s: stripe %llu\n", __func__,
1814 (unsigned long long)sh->sector);
1816 for (i = disks; i--; ) {
1817 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1818 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1819 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1822 for (i = disks; i--; ) {
1823 struct r5dev *dev = &sh->dev[i];
1825 if (dev->written || i == pd_idx || i == qd_idx) {
1826 if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1827 set_bit(R5_UPTODATE, &dev->flags);
1828 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1829 set_bit(R5_Expanded, &dev->flags);
1832 set_bit(R5_WantFUA, &dev->flags);
1834 set_bit(R5_SyncIO, &dev->flags);
1838 if (sh->reconstruct_state == reconstruct_state_drain_run)
1839 sh->reconstruct_state = reconstruct_state_drain_result;
1840 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1841 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1843 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1844 sh->reconstruct_state = reconstruct_state_result;
1847 set_bit(STRIPE_HANDLE, &sh->state);
1848 raid5_release_stripe(sh);
1852 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1853 struct dma_async_tx_descriptor *tx)
1855 int disks = sh->disks;
1856 struct page **xor_srcs;
1857 struct async_submit_ctl submit;
1858 int count, pd_idx = sh->pd_idx, i;
1859 struct page *xor_dest;
1861 unsigned long flags;
1863 struct stripe_head *head_sh = sh;
1866 pr_debug("%s: stripe %llu\n", __func__,
1867 (unsigned long long)sh->sector);
1869 for (i = 0; i < sh->disks; i++) {
1872 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1875 if (i >= sh->disks) {
1876 atomic_inc(&sh->count);
1877 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1878 ops_complete_reconstruct(sh);
1883 xor_srcs = to_addr_page(percpu, j);
1884 /* check if prexor is active which means only process blocks
1885 * that are part of a read-modify-write (written)
1887 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1889 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1890 for (i = disks; i--; ) {
1891 struct r5dev *dev = &sh->dev[i];
1892 if (head_sh->dev[i].written ||
1893 test_bit(R5_InJournal, &head_sh->dev[i].flags))
1894 xor_srcs[count++] = dev->page;
1897 xor_dest = sh->dev[pd_idx].page;
1898 for (i = disks; i--; ) {
1899 struct r5dev *dev = &sh->dev[i];
1901 xor_srcs[count++] = dev->page;
1905 /* 1/ if we prexor'd then the dest is reused as a source
1906 * 2/ if we did not prexor then we are redoing the parity
1907 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1908 * for the synchronous xor case
1910 last_stripe = !head_sh->batch_head ||
1911 list_first_entry(&sh->batch_list,
1912 struct stripe_head, batch_list) == head_sh;
1914 flags = ASYNC_TX_ACK |
1915 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1917 atomic_inc(&head_sh->count);
1918 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1919 to_addr_conv(sh, percpu, j));
1921 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1922 init_async_submit(&submit, flags, tx, NULL, NULL,
1923 to_addr_conv(sh, percpu, j));
1926 if (unlikely(count == 1))
1927 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0,
1928 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1930 tx = async_xor(xor_dest, xor_srcs, 0, count,
1931 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1934 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1941 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1942 struct dma_async_tx_descriptor *tx)
1944 struct async_submit_ctl submit;
1945 struct page **blocks;
1946 int count, i, j = 0;
1947 struct stripe_head *head_sh = sh;
1950 unsigned long txflags;
1952 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1954 for (i = 0; i < sh->disks; i++) {
1955 if (sh->pd_idx == i || sh->qd_idx == i)
1957 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1960 if (i >= sh->disks) {
1961 atomic_inc(&sh->count);
1962 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1963 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1964 ops_complete_reconstruct(sh);
1969 blocks = to_addr_page(percpu, j);
1971 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1972 synflags = SYNDROME_SRC_WRITTEN;
1973 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1975 synflags = SYNDROME_SRC_ALL;
1976 txflags = ASYNC_TX_ACK;
1979 count = set_syndrome_sources(blocks, sh, synflags);
1980 last_stripe = !head_sh->batch_head ||
1981 list_first_entry(&sh->batch_list,
1982 struct stripe_head, batch_list) == head_sh;
1985 atomic_inc(&head_sh->count);
1986 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1987 head_sh, to_addr_conv(sh, percpu, j));
1989 init_async_submit(&submit, 0, tx, NULL, NULL,
1990 to_addr_conv(sh, percpu, j));
1991 tx = async_gen_syndrome(blocks, 0, count+2,
1992 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1995 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2001 static void ops_complete_check(void *stripe_head_ref)
2003 struct stripe_head *sh = stripe_head_ref;
2005 pr_debug("%s: stripe %llu\n", __func__,
2006 (unsigned long long)sh->sector);
2008 sh->check_state = check_state_check_result;
2009 set_bit(STRIPE_HANDLE, &sh->state);
2010 raid5_release_stripe(sh);
2013 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2015 int disks = sh->disks;
2016 int pd_idx = sh->pd_idx;
2017 int qd_idx = sh->qd_idx;
2018 struct page *xor_dest;
2019 struct page **xor_srcs = to_addr_page(percpu, 0);
2020 struct dma_async_tx_descriptor *tx;
2021 struct async_submit_ctl submit;
2025 pr_debug("%s: stripe %llu\n", __func__,
2026 (unsigned long long)sh->sector);
2028 BUG_ON(sh->batch_head);
2030 xor_dest = sh->dev[pd_idx].page;
2031 xor_srcs[count++] = xor_dest;
2032 for (i = disks; i--; ) {
2033 if (i == pd_idx || i == qd_idx)
2035 xor_srcs[count++] = sh->dev[i].page;
2038 init_async_submit(&submit, 0, NULL, NULL, NULL,
2039 to_addr_conv(sh, percpu, 0));
2040 tx = async_xor_val(xor_dest, xor_srcs, 0, count,
2041 RAID5_STRIPE_SIZE(sh->raid_conf),
2042 &sh->ops.zero_sum_result, &submit);
2044 atomic_inc(&sh->count);
2045 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2046 tx = async_trigger_callback(&submit);
2049 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2051 struct page **srcs = to_addr_page(percpu, 0);
2052 struct async_submit_ctl submit;
2055 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2056 (unsigned long long)sh->sector, checkp);
2058 BUG_ON(sh->batch_head);
2059 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2063 atomic_inc(&sh->count);
2064 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2065 sh, to_addr_conv(sh, percpu, 0));
2066 async_syndrome_val(srcs, 0, count+2,
2067 RAID5_STRIPE_SIZE(sh->raid_conf),
2068 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2071 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2073 int overlap_clear = 0, i, disks = sh->disks;
2074 struct dma_async_tx_descriptor *tx = NULL;
2075 struct r5conf *conf = sh->raid_conf;
2076 int level = conf->level;
2077 struct raid5_percpu *percpu;
2081 percpu = per_cpu_ptr(conf->percpu, cpu);
2082 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2083 ops_run_biofill(sh);
2087 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2089 tx = ops_run_compute5(sh, percpu);
2091 if (sh->ops.target2 < 0 || sh->ops.target < 0)
2092 tx = ops_run_compute6_1(sh, percpu);
2094 tx = ops_run_compute6_2(sh, percpu);
2096 /* terminate the chain if reconstruct is not set to be run */
2097 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2101 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2103 tx = ops_run_prexor5(sh, percpu, tx);
2105 tx = ops_run_prexor6(sh, percpu, tx);
2108 if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2109 tx = ops_run_partial_parity(sh, percpu, tx);
2111 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2112 tx = ops_run_biodrain(sh, tx);
2116 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2118 ops_run_reconstruct5(sh, percpu, tx);
2120 ops_run_reconstruct6(sh, percpu, tx);
2123 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2124 if (sh->check_state == check_state_run)
2125 ops_run_check_p(sh, percpu);
2126 else if (sh->check_state == check_state_run_q)
2127 ops_run_check_pq(sh, percpu, 0);
2128 else if (sh->check_state == check_state_run_pq)
2129 ops_run_check_pq(sh, percpu, 1);
2134 if (overlap_clear && !sh->batch_head)
2135 for (i = disks; i--; ) {
2136 struct r5dev *dev = &sh->dev[i];
2137 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2138 wake_up(&sh->raid_conf->wait_for_overlap);
2143 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2146 __free_page(sh->ppl_page);
2147 kmem_cache_free(sc, sh);
2150 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2151 int disks, struct r5conf *conf)
2153 struct stripe_head *sh;
2156 sh = kmem_cache_zalloc(sc, gfp);
2158 spin_lock_init(&sh->stripe_lock);
2159 spin_lock_init(&sh->batch_lock);
2160 INIT_LIST_HEAD(&sh->batch_list);
2161 INIT_LIST_HEAD(&sh->lru);
2162 INIT_LIST_HEAD(&sh->r5c);
2163 INIT_LIST_HEAD(&sh->log_list);
2164 atomic_set(&sh->count, 1);
2165 sh->raid_conf = conf;
2166 sh->log_start = MaxSector;
2167 for (i = 0; i < disks; i++) {
2168 struct r5dev *dev = &sh->dev[i];
2170 bio_init(&dev->req, &dev->vec, 1);
2171 bio_init(&dev->rreq, &dev->rvec, 1);
2174 if (raid5_has_ppl(conf)) {
2175 sh->ppl_page = alloc_page(gfp);
2176 if (!sh->ppl_page) {
2177 free_stripe(sc, sh);
2184 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2186 struct stripe_head *sh;
2188 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2192 if (grow_buffers(sh, gfp)) {
2194 free_stripe(conf->slab_cache, sh);
2197 sh->hash_lock_index =
2198 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2199 /* we just created an active stripe so... */
2200 atomic_inc(&conf->active_stripes);
2202 raid5_release_stripe(sh);
2203 conf->max_nr_stripes++;
2207 static int grow_stripes(struct r5conf *conf, int num)
2209 struct kmem_cache *sc;
2210 size_t namelen = sizeof(conf->cache_name[0]);
2211 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2213 if (conf->mddev->gendisk)
2214 snprintf(conf->cache_name[0], namelen,
2215 "raid%d-%s", conf->level, mdname(conf->mddev));
2217 snprintf(conf->cache_name[0], namelen,
2218 "raid%d-%p", conf->level, conf->mddev);
2219 snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2221 conf->active_name = 0;
2222 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2223 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2227 conf->slab_cache = sc;
2228 conf->pool_size = devs;
2230 if (!grow_one_stripe(conf, GFP_KERNEL))
2237 * scribble_alloc - allocate percpu scribble buffer for required size
2238 * of the scribble region
2239 * @percpu: from for_each_present_cpu() of the caller
2240 * @num: total number of disks in the array
2241 * @cnt: scribble objs count for required size of the scribble region
2243 * The scribble buffer size must be enough to contain:
2244 * 1/ a struct page pointer for each device in the array +2
2245 * 2/ room to convert each entry in (1) to its corresponding dma
2246 * (dma_map_page()) or page (page_address()) address.
2248 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2249 * calculate over all devices (not just the data blocks), using zeros in place
2250 * of the P and Q blocks.
2252 static int scribble_alloc(struct raid5_percpu *percpu,
2256 sizeof(struct page *) * (num+2) +
2257 sizeof(addr_conv_t) * (num+2);
2261 * If here is in raid array suspend context, it is in memalloc noio
2262 * context as well, there is no potential recursive memory reclaim
2263 * I/Os with the GFP_KERNEL flag.
2265 scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2269 kvfree(percpu->scribble);
2271 percpu->scribble = scribble;
2272 percpu->scribble_obj_size = obj_size;
2276 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2282 * Never shrink. And mddev_suspend() could deadlock if this is called
2283 * from raid5d. In that case, scribble_disks and scribble_sectors
2284 * should equal to new_disks and new_sectors
2286 if (conf->scribble_disks >= new_disks &&
2287 conf->scribble_sectors >= new_sectors)
2289 mddev_suspend(conf->mddev);
2292 for_each_present_cpu(cpu) {
2293 struct raid5_percpu *percpu;
2295 percpu = per_cpu_ptr(conf->percpu, cpu);
2296 err = scribble_alloc(percpu, new_disks,
2297 new_sectors / RAID5_STRIPE_SECTORS(conf));
2303 mddev_resume(conf->mddev);
2305 conf->scribble_disks = new_disks;
2306 conf->scribble_sectors = new_sectors;
2311 static int resize_stripes(struct r5conf *conf, int newsize)
2313 /* Make all the stripes able to hold 'newsize' devices.
2314 * New slots in each stripe get 'page' set to a new page.
2316 * This happens in stages:
2317 * 1/ create a new kmem_cache and allocate the required number of
2319 * 2/ gather all the old stripe_heads and transfer the pages across
2320 * to the new stripe_heads. This will have the side effect of
2321 * freezing the array as once all stripe_heads have been collected,
2322 * no IO will be possible. Old stripe heads are freed once their
2323 * pages have been transferred over, and the old kmem_cache is
2324 * freed when all stripes are done.
2325 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2326 * we simple return a failure status - no need to clean anything up.
2327 * 4/ allocate new pages for the new slots in the new stripe_heads.
2328 * If this fails, we don't bother trying the shrink the
2329 * stripe_heads down again, we just leave them as they are.
2330 * As each stripe_head is processed the new one is released into
2333 * Once step2 is started, we cannot afford to wait for a write,
2334 * so we use GFP_NOIO allocations.
2336 struct stripe_head *osh, *nsh;
2337 LIST_HEAD(newstripes);
2338 struct disk_info *ndisks;
2340 struct kmem_cache *sc;
2344 md_allow_write(conf->mddev);
2347 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2348 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2353 /* Need to ensure auto-resizing doesn't interfere */
2354 mutex_lock(&conf->cache_size_mutex);
2356 for (i = conf->max_nr_stripes; i; i--) {
2357 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2361 list_add(&nsh->lru, &newstripes);
2364 /* didn't get enough, give up */
2365 while (!list_empty(&newstripes)) {
2366 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2367 list_del(&nsh->lru);
2368 free_stripe(sc, nsh);
2370 kmem_cache_destroy(sc);
2371 mutex_unlock(&conf->cache_size_mutex);
2374 /* Step 2 - Must use GFP_NOIO now.
2375 * OK, we have enough stripes, start collecting inactive
2376 * stripes and copying them over
2380 list_for_each_entry(nsh, &newstripes, lru) {
2381 lock_device_hash_lock(conf, hash);
2382 wait_event_cmd(conf->wait_for_stripe,
2383 !list_empty(conf->inactive_list + hash),
2384 unlock_device_hash_lock(conf, hash),
2385 lock_device_hash_lock(conf, hash));
2386 osh = get_free_stripe(conf, hash);
2387 unlock_device_hash_lock(conf, hash);
2389 for(i=0; i<conf->pool_size; i++) {
2390 nsh->dev[i].page = osh->dev[i].page;
2391 nsh->dev[i].orig_page = osh->dev[i].page;
2393 nsh->hash_lock_index = hash;
2394 free_stripe(conf->slab_cache, osh);
2396 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2397 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2402 kmem_cache_destroy(conf->slab_cache);
2405 * At this point, we are holding all the stripes so the array
2406 * is completely stalled, so now is a good time to resize
2407 * conf->disks and the scribble region
2409 ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2411 for (i = 0; i < conf->pool_size; i++)
2412 ndisks[i] = conf->disks[i];
2414 for (i = conf->pool_size; i < newsize; i++) {
2415 ndisks[i].extra_page = alloc_page(GFP_NOIO);
2416 if (!ndisks[i].extra_page)
2421 for (i = conf->pool_size; i < newsize; i++)
2422 if (ndisks[i].extra_page)
2423 put_page(ndisks[i].extra_page);
2427 conf->disks = ndisks;
2432 mutex_unlock(&conf->cache_size_mutex);
2434 conf->slab_cache = sc;
2435 conf->active_name = 1-conf->active_name;
2437 /* Step 4, return new stripes to service */
2438 while(!list_empty(&newstripes)) {
2439 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2440 list_del_init(&nsh->lru);
2442 for (i=conf->raid_disks; i < newsize; i++)
2443 if (nsh->dev[i].page == NULL) {
2444 struct page *p = alloc_page(GFP_NOIO);
2445 nsh->dev[i].page = p;
2446 nsh->dev[i].orig_page = p;
2450 raid5_release_stripe(nsh);
2452 /* critical section pass, GFP_NOIO no longer needed */
2455 conf->pool_size = newsize;
2459 static int drop_one_stripe(struct r5conf *conf)
2461 struct stripe_head *sh;
2462 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2464 spin_lock_irq(conf->hash_locks + hash);
2465 sh = get_free_stripe(conf, hash);
2466 spin_unlock_irq(conf->hash_locks + hash);
2469 BUG_ON(atomic_read(&sh->count));
2471 free_stripe(conf->slab_cache, sh);
2472 atomic_dec(&conf->active_stripes);
2473 conf->max_nr_stripes--;
2477 static void shrink_stripes(struct r5conf *conf)
2479 while (conf->max_nr_stripes &&
2480 drop_one_stripe(conf))
2483 kmem_cache_destroy(conf->slab_cache);
2484 conf->slab_cache = NULL;
2487 static void raid5_end_read_request(struct bio * bi)
2489 struct stripe_head *sh = bi->bi_private;
2490 struct r5conf *conf = sh->raid_conf;
2491 int disks = sh->disks, i;
2492 char b[BDEVNAME_SIZE];
2493 struct md_rdev *rdev = NULL;
2496 for (i=0 ; i<disks; i++)
2497 if (bi == &sh->dev[i].req)
2500 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2501 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2508 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2509 /* If replacement finished while this request was outstanding,
2510 * 'replacement' might be NULL already.
2511 * In that case it moved down to 'rdev'.
2512 * rdev is not removed until all requests are finished.
2514 rdev = conf->disks[i].replacement;
2516 rdev = conf->disks[i].rdev;
2518 if (use_new_offset(conf, sh))
2519 s = sh->sector + rdev->new_data_offset;
2521 s = sh->sector + rdev->data_offset;
2522 if (!bi->bi_status) {
2523 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2524 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2525 /* Note that this cannot happen on a
2526 * replacement device. We just fail those on
2529 pr_info_ratelimited(
2530 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2531 mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
2532 (unsigned long long)s,
2533 bdevname(rdev->bdev, b));
2534 atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
2535 clear_bit(R5_ReadError, &sh->dev[i].flags);
2536 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2537 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2538 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2540 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2542 * end read for a page in journal, this
2543 * must be preparing for prexor in rmw
2545 set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2547 if (atomic_read(&rdev->read_errors))
2548 atomic_set(&rdev->read_errors, 0);
2550 const char *bdn = bdevname(rdev->bdev, b);
2554 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2555 if (!(bi->bi_status == BLK_STS_PROTECTION))
2556 atomic_inc(&rdev->read_errors);
2557 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2558 pr_warn_ratelimited(
2559 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2560 mdname(conf->mddev),
2561 (unsigned long long)s,
2563 else if (conf->mddev->degraded >= conf->max_degraded) {
2565 pr_warn_ratelimited(
2566 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2567 mdname(conf->mddev),
2568 (unsigned long long)s,
2570 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2573 pr_warn_ratelimited(
2574 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2575 mdname(conf->mddev),
2576 (unsigned long long)s,
2578 } else if (atomic_read(&rdev->read_errors)
2579 > conf->max_nr_stripes) {
2580 if (!test_bit(Faulty, &rdev->flags)) {
2581 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2582 mdname(conf->mddev),
2583 atomic_read(&rdev->read_errors),
2584 conf->max_nr_stripes);
2585 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2586 mdname(conf->mddev), bdn);
2590 if (set_bad && test_bit(In_sync, &rdev->flags)
2591 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2594 if (sh->qd_idx >= 0 && sh->pd_idx == i)
2595 set_bit(R5_ReadError, &sh->dev[i].flags);
2596 else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2597 set_bit(R5_ReadError, &sh->dev[i].flags);
2598 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2600 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2602 clear_bit(R5_ReadError, &sh->dev[i].flags);
2603 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2605 && test_bit(In_sync, &rdev->flags)
2606 && rdev_set_badblocks(
2607 rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
2608 md_error(conf->mddev, rdev);
2611 rdev_dec_pending(rdev, conf->mddev);
2613 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2614 set_bit(STRIPE_HANDLE, &sh->state);
2615 raid5_release_stripe(sh);
2618 static void raid5_end_write_request(struct bio *bi)
2620 struct stripe_head *sh = bi->bi_private;
2621 struct r5conf *conf = sh->raid_conf;
2622 int disks = sh->disks, i;
2623 struct md_rdev *rdev;
2626 int replacement = 0;
2628 for (i = 0 ; i < disks; i++) {
2629 if (bi == &sh->dev[i].req) {
2630 rdev = conf->disks[i].rdev;
2633 if (bi == &sh->dev[i].rreq) {
2634 rdev = conf->disks[i].replacement;
2638 /* rdev was removed and 'replacement'
2639 * replaced it. rdev is not removed
2640 * until all requests are finished.
2642 rdev = conf->disks[i].rdev;
2646 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2647 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2657 md_error(conf->mddev, rdev);
2658 else if (is_badblock(rdev, sh->sector,
2659 RAID5_STRIPE_SECTORS(conf),
2660 &first_bad, &bad_sectors))
2661 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2663 if (bi->bi_status) {
2664 set_bit(STRIPE_DEGRADED, &sh->state);
2665 set_bit(WriteErrorSeen, &rdev->flags);
2666 set_bit(R5_WriteError, &sh->dev[i].flags);
2667 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2668 set_bit(MD_RECOVERY_NEEDED,
2669 &rdev->mddev->recovery);
2670 } else if (is_badblock(rdev, sh->sector,
2671 RAID5_STRIPE_SECTORS(conf),
2672 &first_bad, &bad_sectors)) {
2673 set_bit(R5_MadeGood, &sh->dev[i].flags);
2674 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2675 /* That was a successful write so make
2676 * sure it looks like we already did
2679 set_bit(R5_ReWrite, &sh->dev[i].flags);
2682 rdev_dec_pending(rdev, conf->mddev);
2684 if (sh->batch_head && bi->bi_status && !replacement)
2685 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2688 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2689 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2690 set_bit(STRIPE_HANDLE, &sh->state);
2691 raid5_release_stripe(sh);
2693 if (sh->batch_head && sh != sh->batch_head)
2694 raid5_release_stripe(sh->batch_head);
2697 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2699 char b[BDEVNAME_SIZE];
2700 struct r5conf *conf = mddev->private;
2701 unsigned long flags;
2702 pr_debug("raid456: error called\n");
2704 spin_lock_irqsave(&conf->device_lock, flags);
2706 if (test_bit(In_sync, &rdev->flags) &&
2707 mddev->degraded == conf->max_degraded) {
2709 * Don't allow to achieve failed state
2710 * Don't try to recover this device
2712 conf->recovery_disabled = mddev->recovery_disabled;
2713 spin_unlock_irqrestore(&conf->device_lock, flags);
2717 set_bit(Faulty, &rdev->flags);
2718 clear_bit(In_sync, &rdev->flags);
2719 mddev->degraded = raid5_calc_degraded(conf);
2720 spin_unlock_irqrestore(&conf->device_lock, flags);
2721 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2723 set_bit(Blocked, &rdev->flags);
2724 set_mask_bits(&mddev->sb_flags, 0,
2725 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2726 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2727 "md/raid:%s: Operation continuing on %d devices.\n",
2729 bdevname(rdev->bdev, b),
2731 conf->raid_disks - mddev->degraded);
2732 r5c_update_on_rdev_error(mddev, rdev);
2736 * Input: a 'big' sector number,
2737 * Output: index of the data and parity disk, and the sector # in them.
2739 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2740 int previous, int *dd_idx,
2741 struct stripe_head *sh)
2743 sector_t stripe, stripe2;
2744 sector_t chunk_number;
2745 unsigned int chunk_offset;
2748 sector_t new_sector;
2749 int algorithm = previous ? conf->prev_algo
2751 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2752 : conf->chunk_sectors;
2753 int raid_disks = previous ? conf->previous_raid_disks
2755 int data_disks = raid_disks - conf->max_degraded;
2757 /* First compute the information on this sector */
2760 * Compute the chunk number and the sector offset inside the chunk
2762 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2763 chunk_number = r_sector;
2766 * Compute the stripe number
2768 stripe = chunk_number;
2769 *dd_idx = sector_div(stripe, data_disks);
2772 * Select the parity disk based on the user selected algorithm.
2774 pd_idx = qd_idx = -1;
2775 switch(conf->level) {
2777 pd_idx = data_disks;
2780 switch (algorithm) {
2781 case ALGORITHM_LEFT_ASYMMETRIC:
2782 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2783 if (*dd_idx >= pd_idx)
2786 case ALGORITHM_RIGHT_ASYMMETRIC:
2787 pd_idx = sector_div(stripe2, raid_disks);
2788 if (*dd_idx >= pd_idx)
2791 case ALGORITHM_LEFT_SYMMETRIC:
2792 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2793 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2795 case ALGORITHM_RIGHT_SYMMETRIC:
2796 pd_idx = sector_div(stripe2, raid_disks);
2797 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2799 case ALGORITHM_PARITY_0:
2803 case ALGORITHM_PARITY_N:
2804 pd_idx = data_disks;
2812 switch (algorithm) {
2813 case ALGORITHM_LEFT_ASYMMETRIC:
2814 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2815 qd_idx = pd_idx + 1;
2816 if (pd_idx == raid_disks-1) {
2817 (*dd_idx)++; /* Q D D D P */
2819 } else if (*dd_idx >= pd_idx)
2820 (*dd_idx) += 2; /* D D P Q D */
2822 case ALGORITHM_RIGHT_ASYMMETRIC:
2823 pd_idx = sector_div(stripe2, raid_disks);
2824 qd_idx = pd_idx + 1;
2825 if (pd_idx == raid_disks-1) {
2826 (*dd_idx)++; /* Q D D D P */
2828 } else if (*dd_idx >= pd_idx)
2829 (*dd_idx) += 2; /* D D P Q D */
2831 case ALGORITHM_LEFT_SYMMETRIC:
2832 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2833 qd_idx = (pd_idx + 1) % raid_disks;
2834 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2836 case ALGORITHM_RIGHT_SYMMETRIC:
2837 pd_idx = sector_div(stripe2, raid_disks);
2838 qd_idx = (pd_idx + 1) % raid_disks;
2839 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2842 case ALGORITHM_PARITY_0:
2847 case ALGORITHM_PARITY_N:
2848 pd_idx = data_disks;
2849 qd_idx = data_disks + 1;
2852 case ALGORITHM_ROTATING_ZERO_RESTART:
2853 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2854 * of blocks for computing Q is different.
2856 pd_idx = sector_div(stripe2, raid_disks);
2857 qd_idx = pd_idx + 1;
2858 if (pd_idx == raid_disks-1) {
2859 (*dd_idx)++; /* Q D D D P */
2861 } else if (*dd_idx >= pd_idx)
2862 (*dd_idx) += 2; /* D D P Q D */
2866 case ALGORITHM_ROTATING_N_RESTART:
2867 /* Same a left_asymmetric, by first stripe is
2868 * D D D P Q rather than
2872 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2873 qd_idx = pd_idx + 1;
2874 if (pd_idx == raid_disks-1) {
2875 (*dd_idx)++; /* Q D D D P */
2877 } else if (*dd_idx >= pd_idx)
2878 (*dd_idx) += 2; /* D D P Q D */
2882 case ALGORITHM_ROTATING_N_CONTINUE:
2883 /* Same as left_symmetric but Q is before P */
2884 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2885 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2886 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2890 case ALGORITHM_LEFT_ASYMMETRIC_6:
2891 /* RAID5 left_asymmetric, with Q on last device */
2892 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2893 if (*dd_idx >= pd_idx)
2895 qd_idx = raid_disks - 1;
2898 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2899 pd_idx = sector_div(stripe2, raid_disks-1);
2900 if (*dd_idx >= pd_idx)
2902 qd_idx = raid_disks - 1;
2905 case ALGORITHM_LEFT_SYMMETRIC_6:
2906 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2907 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2908 qd_idx = raid_disks - 1;
2911 case ALGORITHM_RIGHT_SYMMETRIC_6:
2912 pd_idx = sector_div(stripe2, raid_disks-1);
2913 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2914 qd_idx = raid_disks - 1;
2917 case ALGORITHM_PARITY_0_6:
2920 qd_idx = raid_disks - 1;
2930 sh->pd_idx = pd_idx;
2931 sh->qd_idx = qd_idx;
2932 sh->ddf_layout = ddf_layout;
2935 * Finally, compute the new sector number
2937 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2941 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2943 struct r5conf *conf = sh->raid_conf;
2944 int raid_disks = sh->disks;
2945 int data_disks = raid_disks - conf->max_degraded;
2946 sector_t new_sector = sh->sector, check;
2947 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2948 : conf->chunk_sectors;
2949 int algorithm = previous ? conf->prev_algo
2953 sector_t chunk_number;
2954 int dummy1, dd_idx = i;
2956 struct stripe_head sh2;
2958 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2959 stripe = new_sector;
2961 if (i == sh->pd_idx)
2963 switch(conf->level) {
2966 switch (algorithm) {
2967 case ALGORITHM_LEFT_ASYMMETRIC:
2968 case ALGORITHM_RIGHT_ASYMMETRIC:
2972 case ALGORITHM_LEFT_SYMMETRIC:
2973 case ALGORITHM_RIGHT_SYMMETRIC:
2976 i -= (sh->pd_idx + 1);
2978 case ALGORITHM_PARITY_0:
2981 case ALGORITHM_PARITY_N:
2988 if (i == sh->qd_idx)
2989 return 0; /* It is the Q disk */
2990 switch (algorithm) {
2991 case ALGORITHM_LEFT_ASYMMETRIC:
2992 case ALGORITHM_RIGHT_ASYMMETRIC:
2993 case ALGORITHM_ROTATING_ZERO_RESTART:
2994 case ALGORITHM_ROTATING_N_RESTART:
2995 if (sh->pd_idx == raid_disks-1)
2996 i--; /* Q D D D P */
2997 else if (i > sh->pd_idx)
2998 i -= 2; /* D D P Q D */
3000 case ALGORITHM_LEFT_SYMMETRIC:
3001 case ALGORITHM_RIGHT_SYMMETRIC:
3002 if (sh->pd_idx == raid_disks-1)
3003 i--; /* Q D D D P */
3008 i -= (sh->pd_idx + 2);
3011 case ALGORITHM_PARITY_0:
3014 case ALGORITHM_PARITY_N:
3016 case ALGORITHM_ROTATING_N_CONTINUE:
3017 /* Like left_symmetric, but P is before Q */
3018 if (sh->pd_idx == 0)
3019 i--; /* P D D D Q */
3024 i -= (sh->pd_idx + 1);
3027 case ALGORITHM_LEFT_ASYMMETRIC_6:
3028 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3032 case ALGORITHM_LEFT_SYMMETRIC_6:
3033 case ALGORITHM_RIGHT_SYMMETRIC_6:
3035 i += data_disks + 1;
3036 i -= (sh->pd_idx + 1);
3038 case ALGORITHM_PARITY_0_6:
3047 chunk_number = stripe * data_disks + i;
3048 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3050 check = raid5_compute_sector(conf, r_sector,
3051 previous, &dummy1, &sh2);
3052 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3053 || sh2.qd_idx != sh->qd_idx) {
3054 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3055 mdname(conf->mddev));
3062 * There are cases where we want handle_stripe_dirtying() and
3063 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3065 * This function checks whether we want to delay the towrite. Specifically,
3066 * we delay the towrite when:
3068 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3069 * stripe has data in journal (for other devices).
3071 * In this case, when reading data for the non-overwrite dev, it is
3072 * necessary to handle complex rmw of write back cache (prexor with
3073 * orig_page, and xor with page). To keep read path simple, we would
3074 * like to flush data in journal to RAID disks first, so complex rmw
3075 * is handled in the write patch (handle_stripe_dirtying).
3077 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3079 * It is important to be able to flush all stripes in raid5-cache.
3080 * Therefore, we need reserve some space on the journal device for
3081 * these flushes. If flush operation includes pending writes to the
3082 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3083 * for the flush out. If we exclude these pending writes from flush
3084 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3085 * Therefore, excluding pending writes in these cases enables more
3086 * efficient use of the journal device.
3088 * Note: To make sure the stripe makes progress, we only delay
3089 * towrite for stripes with data already in journal (injournal > 0).
3090 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3091 * no_space_stripes list.
3093 * 3. during journal failure
3094 * In journal failure, we try to flush all cached data to raid disks
3095 * based on data in stripe cache. The array is read-only to upper
3096 * layers, so we would skip all pending writes.
3099 static inline bool delay_towrite(struct r5conf *conf,
3101 struct stripe_head_state *s)
3104 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3105 !test_bit(R5_Insync, &dev->flags) && s->injournal)
3108 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3112 if (s->log_failed && s->injournal)
3118 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3119 int rcw, int expand)
3121 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3122 struct r5conf *conf = sh->raid_conf;
3123 int level = conf->level;
3127 * In some cases, handle_stripe_dirtying initially decided to
3128 * run rmw and allocates extra page for prexor. However, rcw is
3129 * cheaper later on. We need to free the extra page now,
3130 * because we won't be able to do that in ops_complete_prexor().
3132 r5c_release_extra_page(sh);
3134 for (i = disks; i--; ) {
3135 struct r5dev *dev = &sh->dev[i];
3137 if (dev->towrite && !delay_towrite(conf, dev, s)) {
3138 set_bit(R5_LOCKED, &dev->flags);
3139 set_bit(R5_Wantdrain, &dev->flags);
3141 clear_bit(R5_UPTODATE, &dev->flags);
3143 } else if (test_bit(R5_InJournal, &dev->flags)) {
3144 set_bit(R5_LOCKED, &dev->flags);
3148 /* if we are not expanding this is a proper write request, and
3149 * there will be bios with new data to be drained into the
3154 /* False alarm, nothing to do */
3156 sh->reconstruct_state = reconstruct_state_drain_run;
3157 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3159 sh->reconstruct_state = reconstruct_state_run;
3161 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3163 if (s->locked + conf->max_degraded == disks)
3164 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3165 atomic_inc(&conf->pending_full_writes);
3167 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3168 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3169 BUG_ON(level == 6 &&
3170 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3171 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3173 for (i = disks; i--; ) {
3174 struct r5dev *dev = &sh->dev[i];
3175 if (i == pd_idx || i == qd_idx)
3179 (test_bit(R5_UPTODATE, &dev->flags) ||
3180 test_bit(R5_Wantcompute, &dev->flags))) {
3181 set_bit(R5_Wantdrain, &dev->flags);
3182 set_bit(R5_LOCKED, &dev->flags);
3183 clear_bit(R5_UPTODATE, &dev->flags);
3185 } else if (test_bit(R5_InJournal, &dev->flags)) {
3186 set_bit(R5_LOCKED, &dev->flags);
3191 /* False alarm - nothing to do */
3193 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3194 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3195 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3196 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3199 /* keep the parity disk(s) locked while asynchronous operations
3202 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3203 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3207 int qd_idx = sh->qd_idx;
3208 struct r5dev *dev = &sh->dev[qd_idx];
3210 set_bit(R5_LOCKED, &dev->flags);
3211 clear_bit(R5_UPTODATE, &dev->flags);
3215 if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3216 test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3217 !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3218 test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3219 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3221 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3222 __func__, (unsigned long long)sh->sector,
3223 s->locked, s->ops_request);
3227 * Each stripe/dev can have one or more bion attached.
3228 * toread/towrite point to the first in a chain.
3229 * The bi_next chain must be in order.
3231 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3232 int forwrite, int previous)
3235 struct r5conf *conf = sh->raid_conf;
3238 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3239 (unsigned long long)bi->bi_iter.bi_sector,
3240 (unsigned long long)sh->sector);
3242 spin_lock_irq(&sh->stripe_lock);
3243 sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3244 /* Don't allow new IO added to stripes in batch list */
3248 bip = &sh->dev[dd_idx].towrite;
3252 bip = &sh->dev[dd_idx].toread;
3253 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3254 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3256 bip = & (*bip)->bi_next;
3258 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3261 if (forwrite && raid5_has_ppl(conf)) {
3263 * With PPL only writes to consecutive data chunks within a
3264 * stripe are allowed because for a single stripe_head we can
3265 * only have one PPL entry at a time, which describes one data
3266 * range. Not really an overlap, but wait_for_overlap can be
3267 * used to handle this.
3275 for (i = 0; i < sh->disks; i++) {
3276 if (i != sh->pd_idx &&
3277 (i == dd_idx || sh->dev[i].towrite)) {
3278 sector = sh->dev[i].sector;
3279 if (count == 0 || sector < first)
3287 if (first + conf->chunk_sectors * (count - 1) != last)
3291 if (!forwrite || previous)
3292 clear_bit(STRIPE_BATCH_READY, &sh->state);
3294 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3298 bio_inc_remaining(bi);
3299 md_write_inc(conf->mddev, bi);
3302 /* check if page is covered */
3303 sector_t sector = sh->dev[dd_idx].sector;
3304 for (bi=sh->dev[dd_idx].towrite;
3305 sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
3306 bi && bi->bi_iter.bi_sector <= sector;
3307 bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
3308 if (bio_end_sector(bi) >= sector)
3309 sector = bio_end_sector(bi);
3311 if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
3312 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3313 sh->overwrite_disks++;
3316 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3317 (unsigned long long)(*bip)->bi_iter.bi_sector,
3318 (unsigned long long)sh->sector, dd_idx);
3320 if (conf->mddev->bitmap && firstwrite) {
3321 /* Cannot hold spinlock over bitmap_startwrite,
3322 * but must ensure this isn't added to a batch until
3323 * we have added to the bitmap and set bm_seq.
3324 * So set STRIPE_BITMAP_PENDING to prevent
3326 * If multiple add_stripe_bio() calls race here they
3327 * much all set STRIPE_BITMAP_PENDING. So only the first one
3328 * to complete "bitmap_startwrite" gets to set
3329 * STRIPE_BIT_DELAY. This is important as once a stripe
3330 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3333 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3334 spin_unlock_irq(&sh->stripe_lock);
3335 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3336 RAID5_STRIPE_SECTORS(conf), 0);
3337 spin_lock_irq(&sh->stripe_lock);
3338 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3339 if (!sh->batch_head) {
3340 sh->bm_seq = conf->seq_flush+1;
3341 set_bit(STRIPE_BIT_DELAY, &sh->state);
3344 spin_unlock_irq(&sh->stripe_lock);
3346 if (stripe_can_batch(sh))
3347 stripe_add_to_batch_list(conf, sh);
3351 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3352 spin_unlock_irq(&sh->stripe_lock);
3356 static void end_reshape(struct r5conf *conf);
3358 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3359 struct stripe_head *sh)
3361 int sectors_per_chunk =
3362 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3364 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3365 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3367 raid5_compute_sector(conf,
3368 stripe * (disks - conf->max_degraded)
3369 *sectors_per_chunk + chunk_offset,
3375 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3376 struct stripe_head_state *s, int disks)
3379 BUG_ON(sh->batch_head);
3380 for (i = disks; i--; ) {
3384 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3385 struct md_rdev *rdev;
3387 rdev = rcu_dereference(conf->disks[i].rdev);
3388 if (rdev && test_bit(In_sync, &rdev->flags) &&
3389 !test_bit(Faulty, &rdev->flags))
3390 atomic_inc(&rdev->nr_pending);
3395 if (!rdev_set_badblocks(
3398 RAID5_STRIPE_SECTORS(conf), 0))
3399 md_error(conf->mddev, rdev);
3400 rdev_dec_pending(rdev, conf->mddev);
3403 spin_lock_irq(&sh->stripe_lock);
3404 /* fail all writes first */
3405 bi = sh->dev[i].towrite;
3406 sh->dev[i].towrite = NULL;
3407 sh->overwrite_disks = 0;
3408 spin_unlock_irq(&sh->stripe_lock);
3412 log_stripe_write_finished(sh);
3414 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3415 wake_up(&conf->wait_for_overlap);
3417 while (bi && bi->bi_iter.bi_sector <
3418 sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3419 struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
3421 md_write_end(conf->mddev);
3426 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3427 RAID5_STRIPE_SECTORS(conf), 0, 0);
3429 /* and fail all 'written' */
3430 bi = sh->dev[i].written;
3431 sh->dev[i].written = NULL;
3432 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3433 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3434 sh->dev[i].page = sh->dev[i].orig_page;
3437 if (bi) bitmap_end = 1;
3438 while (bi && bi->bi_iter.bi_sector <
3439 sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3440 struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
3442 md_write_end(conf->mddev);
3447 /* fail any reads if this device is non-operational and
3448 * the data has not reached the cache yet.
3450 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3451 s->failed > conf->max_degraded &&
3452 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3453 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3454 spin_lock_irq(&sh->stripe_lock);
3455 bi = sh->dev[i].toread;
3456 sh->dev[i].toread = NULL;
3457 spin_unlock_irq(&sh->stripe_lock);
3458 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3459 wake_up(&conf->wait_for_overlap);
3462 while (bi && bi->bi_iter.bi_sector <
3463 sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3464 struct bio *nextbi =
3465 r5_next_bio(conf, bi, sh->dev[i].sector);
3472 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3473 RAID5_STRIPE_SECTORS(conf), 0, 0);
3474 /* If we were in the middle of a write the parity block might
3475 * still be locked - so just clear all R5_LOCKED flags
3477 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3482 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3483 if (atomic_dec_and_test(&conf->pending_full_writes))
3484 md_wakeup_thread(conf->mddev->thread);
3488 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3489 struct stripe_head_state *s)
3494 BUG_ON(sh->batch_head);
3495 clear_bit(STRIPE_SYNCING, &sh->state);
3496 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3497 wake_up(&conf->wait_for_overlap);
3500 /* There is nothing more to do for sync/check/repair.
3501 * Don't even need to abort as that is handled elsewhere
3502 * if needed, and not always wanted e.g. if there is a known
3504 * For recover/replace we need to record a bad block on all
3505 * non-sync devices, or abort the recovery
3507 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3508 /* During recovery devices cannot be removed, so
3509 * locking and refcounting of rdevs is not needed
3512 for (i = 0; i < conf->raid_disks; i++) {
3513 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3515 && !test_bit(Faulty, &rdev->flags)
3516 && !test_bit(In_sync, &rdev->flags)
3517 && !rdev_set_badblocks(rdev, sh->sector,
3518 RAID5_STRIPE_SECTORS(conf), 0))
3520 rdev = rcu_dereference(conf->disks[i].replacement);
3522 && !test_bit(Faulty, &rdev->flags)
3523 && !test_bit(In_sync, &rdev->flags)
3524 && !rdev_set_badblocks(rdev, sh->sector,
3525 RAID5_STRIPE_SECTORS(conf), 0))
3530 conf->recovery_disabled =
3531 conf->mddev->recovery_disabled;
3533 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
3536 static int want_replace(struct stripe_head *sh, int disk_idx)
3538 struct md_rdev *rdev;
3542 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3544 && !test_bit(Faulty, &rdev->flags)
3545 && !test_bit(In_sync, &rdev->flags)
3546 && (rdev->recovery_offset <= sh->sector
3547 || rdev->mddev->recovery_cp <= sh->sector))
3553 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3554 int disk_idx, int disks)
3556 struct r5dev *dev = &sh->dev[disk_idx];
3557 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3558 &sh->dev[s->failed_num[1]] };
3560 bool force_rcw = (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW);
3563 if (test_bit(R5_LOCKED, &dev->flags) ||
3564 test_bit(R5_UPTODATE, &dev->flags))
3565 /* No point reading this as we already have it or have
3566 * decided to get it.
3571 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3572 /* We need this block to directly satisfy a request */
3575 if (s->syncing || s->expanding ||
3576 (s->replacing && want_replace(sh, disk_idx)))
3577 /* When syncing, or expanding we read everything.
3578 * When replacing, we need the replaced block.
3582 if ((s->failed >= 1 && fdev[0]->toread) ||
3583 (s->failed >= 2 && fdev[1]->toread))
3584 /* If we want to read from a failed device, then
3585 * we need to actually read every other device.
3589 /* Sometimes neither read-modify-write nor reconstruct-write
3590 * cycles can work. In those cases we read every block we
3591 * can. Then the parity-update is certain to have enough to
3593 * This can only be a problem when we need to write something,
3594 * and some device has failed. If either of those tests
3595 * fail we need look no further.
3597 if (!s->failed || !s->to_write)
3600 if (test_bit(R5_Insync, &dev->flags) &&
3601 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3602 /* Pre-reads at not permitted until after short delay
3603 * to gather multiple requests. However if this
3604 * device is no Insync, the block could only be computed
3605 * and there is no need to delay that.
3609 for (i = 0; i < s->failed && i < 2; i++) {
3610 if (fdev[i]->towrite &&
3611 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3612 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3613 /* If we have a partial write to a failed
3614 * device, then we will need to reconstruct
3615 * the content of that device, so all other
3616 * devices must be read.
3620 if (s->failed >= 2 &&
3621 (fdev[i]->towrite ||
3622 s->failed_num[i] == sh->pd_idx ||
3623 s->failed_num[i] == sh->qd_idx) &&
3624 !test_bit(R5_UPTODATE, &fdev[i]->flags))
3625 /* In max degraded raid6, If the failed disk is P, Q,
3626 * or we want to read the failed disk, we need to do
3627 * reconstruct-write.
3632 /* If we are forced to do a reconstruct-write, because parity
3633 * cannot be trusted and we are currently recovering it, there
3634 * is extra need to be careful.
3635 * If one of the devices that we would need to read, because
3636 * it is not being overwritten (and maybe not written at all)
3637 * is missing/faulty, then we need to read everything we can.
3640 sh->sector < sh->raid_conf->mddev->recovery_cp)
3641 /* reconstruct-write isn't being forced */
3643 for (i = 0; i < s->failed && i < 2; i++) {
3644 if (s->failed_num[i] != sh->pd_idx &&
3645 s->failed_num[i] != sh->qd_idx &&
3646 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3647 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3654 /* fetch_block - checks the given member device to see if its data needs
3655 * to be read or computed to satisfy a request.
3657 * Returns 1 when no more member devices need to be checked, otherwise returns
3658 * 0 to tell the loop in handle_stripe_fill to continue
3660 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3661 int disk_idx, int disks)
3663 struct r5dev *dev = &sh->dev[disk_idx];
3665 /* is the data in this block needed, and can we get it? */
3666 if (need_this_block(sh, s, disk_idx, disks)) {
3667 /* we would like to get this block, possibly by computing it,
3668 * otherwise read it if the backing disk is insync
3670 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3671 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3672 BUG_ON(sh->batch_head);
3675 * In the raid6 case if the only non-uptodate disk is P
3676 * then we already trusted P to compute the other failed
3677 * drives. It is safe to compute rather than re-read P.
3678 * In other cases we only compute blocks from failed
3679 * devices, otherwise check/repair might fail to detect
3680 * a real inconsistency.
3683 if ((s->uptodate == disks - 1) &&
3684 ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3685 (s->failed && (disk_idx == s->failed_num[0] ||
3686 disk_idx == s->failed_num[1])))) {
3687 /* have disk failed, and we're requested to fetch it;
3690 pr_debug("Computing stripe %llu block %d\n",
3691 (unsigned long long)sh->sector, disk_idx);
3692 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3693 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3694 set_bit(R5_Wantcompute, &dev->flags);
3695 sh->ops.target = disk_idx;
3696 sh->ops.target2 = -1; /* no 2nd target */
3698 /* Careful: from this point on 'uptodate' is in the eye
3699 * of raid_run_ops which services 'compute' operations
3700 * before writes. R5_Wantcompute flags a block that will
3701 * be R5_UPTODATE by the time it is needed for a
3702 * subsequent operation.
3706 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3707 /* Computing 2-failure is *very* expensive; only
3708 * do it if failed >= 2
3711 for (other = disks; other--; ) {
3712 if (other == disk_idx)
3714 if (!test_bit(R5_UPTODATE,
3715 &sh->dev[other].flags))
3719 pr_debug("Computing stripe %llu blocks %d,%d\n",
3720 (unsigned long long)sh->sector,
3722 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3723 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3724 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3725 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3726 sh->ops.target = disk_idx;
3727 sh->ops.target2 = other;
3731 } else if (test_bit(R5_Insync, &dev->flags)) {
3732 set_bit(R5_LOCKED, &dev->flags);
3733 set_bit(R5_Wantread, &dev->flags);
3735 pr_debug("Reading block %d (sync=%d)\n",
3736 disk_idx, s->syncing);
3744 * handle_stripe_fill - read or compute data to satisfy pending requests.
3746 static void handle_stripe_fill(struct stripe_head *sh,
3747 struct stripe_head_state *s,
3752 /* look for blocks to read/compute, skip this if a compute
3753 * is already in flight, or if the stripe contents are in the
3754 * midst of changing due to a write
3756 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3757 !sh->reconstruct_state) {
3760 * For degraded stripe with data in journal, do not handle
3761 * read requests yet, instead, flush the stripe to raid
3762 * disks first, this avoids handling complex rmw of write
3763 * back cache (prexor with orig_page, and then xor with
3764 * page) in the read path
3766 if (s->injournal && s->failed) {
3767 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3768 r5c_make_stripe_write_out(sh);
3772 for (i = disks; i--; )
3773 if (fetch_block(sh, s, i, disks))
3777 set_bit(STRIPE_HANDLE, &sh->state);
3780 static void break_stripe_batch_list(struct stripe_head *head_sh,
3781 unsigned long handle_flags);
3782 /* handle_stripe_clean_event
3783 * any written block on an uptodate or failed drive can be returned.
3784 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3785 * never LOCKED, so we don't need to test 'failed' directly.
3787 static void handle_stripe_clean_event(struct r5conf *conf,
3788 struct stripe_head *sh, int disks)
3792 int discard_pending = 0;
3793 struct stripe_head *head_sh = sh;
3794 bool do_endio = false;
3796 for (i = disks; i--; )
3797 if (sh->dev[i].written) {
3799 if (!test_bit(R5_LOCKED, &dev->flags) &&
3800 (test_bit(R5_UPTODATE, &dev->flags) ||
3801 test_bit(R5_Discard, &dev->flags) ||
3802 test_bit(R5_SkipCopy, &dev->flags))) {
3803 /* We can return any write requests */
3804 struct bio *wbi, *wbi2;
3805 pr_debug("Return write for disc %d\n", i);
3806 if (test_and_clear_bit(R5_Discard, &dev->flags))
3807 clear_bit(R5_UPTODATE, &dev->flags);
3808 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3809 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3814 dev->page = dev->orig_page;
3816 dev->written = NULL;
3817 while (wbi && wbi->bi_iter.bi_sector <
3818 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
3819 wbi2 = r5_next_bio(conf, wbi, dev->sector);
3820 md_write_end(conf->mddev);
3824 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3825 RAID5_STRIPE_SECTORS(conf),
3826 !test_bit(STRIPE_DEGRADED, &sh->state),
3828 if (head_sh->batch_head) {
3829 sh = list_first_entry(&sh->batch_list,
3832 if (sh != head_sh) {
3839 } else if (test_bit(R5_Discard, &dev->flags))
3840 discard_pending = 1;
3843 log_stripe_write_finished(sh);
3845 if (!discard_pending &&
3846 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3848 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3849 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3850 if (sh->qd_idx >= 0) {
3851 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3852 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3854 /* now that discard is done we can proceed with any sync */
3855 clear_bit(STRIPE_DISCARD, &sh->state);
3857 * SCSI discard will change some bio fields and the stripe has
3858 * no updated data, so remove it from hash list and the stripe
3859 * will be reinitialized
3862 hash = sh->hash_lock_index;
3863 spin_lock_irq(conf->hash_locks + hash);
3865 spin_unlock_irq(conf->hash_locks + hash);
3866 if (head_sh->batch_head) {
3867 sh = list_first_entry(&sh->batch_list,
3868 struct stripe_head, batch_list);
3874 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3875 set_bit(STRIPE_HANDLE, &sh->state);
3879 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3880 if (atomic_dec_and_test(&conf->pending_full_writes))
3881 md_wakeup_thread(conf->mddev->thread);
3883 if (head_sh->batch_head && do_endio)
3884 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3888 * For RMW in write back cache, we need extra page in prexor to store the
3889 * old data. This page is stored in dev->orig_page.
3891 * This function checks whether we have data for prexor. The exact logic
3893 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3895 static inline bool uptodate_for_rmw(struct r5dev *dev)
3897 return (test_bit(R5_UPTODATE, &dev->flags)) &&
3898 (!test_bit(R5_InJournal, &dev->flags) ||
3899 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3902 static int handle_stripe_dirtying(struct r5conf *conf,
3903 struct stripe_head *sh,
3904 struct stripe_head_state *s,
3907 int rmw = 0, rcw = 0, i;
3908 sector_t recovery_cp = conf->mddev->recovery_cp;
3910 /* Check whether resync is now happening or should start.
3911 * If yes, then the array is dirty (after unclean shutdown or
3912 * initial creation), so parity in some stripes might be inconsistent.
3913 * In this case, we need to always do reconstruct-write, to ensure
3914 * that in case of drive failure or read-error correction, we
3915 * generate correct data from the parity.
3917 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3918 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3920 /* Calculate the real rcw later - for now make it
3921 * look like rcw is cheaper
3924 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3925 conf->rmw_level, (unsigned long long)recovery_cp,
3926 (unsigned long long)sh->sector);
3927 } else for (i = disks; i--; ) {
3928 /* would I have to read this buffer for read_modify_write */
3929 struct r5dev *dev = &sh->dev[i];
3930 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3931 i == sh->pd_idx || i == sh->qd_idx ||
3932 test_bit(R5_InJournal, &dev->flags)) &&
3933 !test_bit(R5_LOCKED, &dev->flags) &&
3934 !(uptodate_for_rmw(dev) ||
3935 test_bit(R5_Wantcompute, &dev->flags))) {
3936 if (test_bit(R5_Insync, &dev->flags))
3939 rmw += 2*disks; /* cannot read it */
3941 /* Would I have to read this buffer for reconstruct_write */
3942 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3943 i != sh->pd_idx && i != sh->qd_idx &&
3944 !test_bit(R5_LOCKED, &dev->flags) &&
3945 !(test_bit(R5_UPTODATE, &dev->flags) ||
3946 test_bit(R5_Wantcompute, &dev->flags))) {
3947 if (test_bit(R5_Insync, &dev->flags))
3954 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3955 (unsigned long long)sh->sector, sh->state, rmw, rcw);
3956 set_bit(STRIPE_HANDLE, &sh->state);
3957 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3958 /* prefer read-modify-write, but need to get some data */
3959 if (conf->mddev->queue)
3960 blk_add_trace_msg(conf->mddev->queue,
3961 "raid5 rmw %llu %d",
3962 (unsigned long long)sh->sector, rmw);
3963 for (i = disks; i--; ) {
3964 struct r5dev *dev = &sh->dev[i];
3965 if (test_bit(R5_InJournal, &dev->flags) &&
3966 dev->page == dev->orig_page &&
3967 !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3968 /* alloc page for prexor */
3969 struct page *p = alloc_page(GFP_NOIO);
3977 * alloc_page() failed, try use
3978 * disk_info->extra_page
3980 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3981 &conf->cache_state)) {
3982 r5c_use_extra_page(sh);
3986 /* extra_page in use, add to delayed_list */
3987 set_bit(STRIPE_DELAYED, &sh->state);
3988 s->waiting_extra_page = 1;
3993 for (i = disks; i--; ) {
3994 struct r5dev *dev = &sh->dev[i];
3995 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3996 i == sh->pd_idx || i == sh->qd_idx ||
3997 test_bit(R5_InJournal, &dev->flags)) &&
3998 !test_bit(R5_LOCKED, &dev->flags) &&
3999 !(uptodate_for_rmw(dev) ||
4000 test_bit(R5_Wantcompute, &dev->flags)) &&
4001 test_bit(R5_Insync, &dev->flags)) {
4002 if (test_bit(STRIPE_PREREAD_ACTIVE,
4004 pr_debug("Read_old block %d for r-m-w\n",
4006 set_bit(R5_LOCKED, &dev->flags);
4007 set_bit(R5_Wantread, &dev->flags);
4010 set_bit(STRIPE_DELAYED, &sh->state);
4014 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
4015 /* want reconstruct write, but need to get some data */
4018 for (i = disks; i--; ) {
4019 struct r5dev *dev = &sh->dev[i];
4020 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4021 i != sh->pd_idx && i != sh->qd_idx &&
4022 !test_bit(R5_LOCKED, &dev->flags) &&
4023 !(test_bit(R5_UPTODATE, &dev->flags) ||
4024 test_bit(R5_Wantcompute, &dev->flags))) {
4026 if (test_bit(R5_Insync, &dev->flags) &&
4027 test_bit(STRIPE_PREREAD_ACTIVE,
4029 pr_debug("Read_old block "
4030 "%d for Reconstruct\n", i);
4031 set_bit(R5_LOCKED, &dev->flags);
4032 set_bit(R5_Wantread, &dev->flags);
4036 set_bit(STRIPE_DELAYED, &sh->state);
4039 if (rcw && conf->mddev->queue)
4040 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4041 (unsigned long long)sh->sector,
4042 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4045 if (rcw > disks && rmw > disks &&
4046 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4047 set_bit(STRIPE_DELAYED, &sh->state);
4049 /* now if nothing is locked, and if we have enough data,
4050 * we can start a write request
4052 /* since handle_stripe can be called at any time we need to handle the
4053 * case where a compute block operation has been submitted and then a
4054 * subsequent call wants to start a write request. raid_run_ops only
4055 * handles the case where compute block and reconstruct are requested
4056 * simultaneously. If this is not the case then new writes need to be
4057 * held off until the compute completes.
4059 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4060 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4061 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4062 schedule_reconstruction(sh, s, rcw == 0, 0);
4066 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4067 struct stripe_head_state *s, int disks)
4069 struct r5dev *dev = NULL;
4071 BUG_ON(sh->batch_head);
4072 set_bit(STRIPE_HANDLE, &sh->state);
4074 switch (sh->check_state) {
4075 case check_state_idle:
4076 /* start a new check operation if there are no failures */
4077 if (s->failed == 0) {
4078 BUG_ON(s->uptodate != disks);
4079 sh->check_state = check_state_run;
4080 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4081 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4085 dev = &sh->dev[s->failed_num[0]];
4087 case check_state_compute_result:
4088 sh->check_state = check_state_idle;
4090 dev = &sh->dev[sh->pd_idx];
4092 /* check that a write has not made the stripe insync */
4093 if (test_bit(STRIPE_INSYNC, &sh->state))
4096 /* either failed parity check, or recovery is happening */
4097 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4098 BUG_ON(s->uptodate != disks);
4100 set_bit(R5_LOCKED, &dev->flags);
4102 set_bit(R5_Wantwrite, &dev->flags);
4104 clear_bit(STRIPE_DEGRADED, &sh->state);
4105 set_bit(STRIPE_INSYNC, &sh->state);
4107 case check_state_run:
4108 break; /* we will be called again upon completion */
4109 case check_state_check_result:
4110 sh->check_state = check_state_idle;
4112 /* if a failure occurred during the check operation, leave
4113 * STRIPE_INSYNC not set and let the stripe be handled again
4118 /* handle a successful check operation, if parity is correct
4119 * we are done. Otherwise update the mismatch count and repair
4120 * parity if !MD_RECOVERY_CHECK
4122 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4123 /* parity is correct (on disc,
4124 * not in buffer any more)
4126 set_bit(STRIPE_INSYNC, &sh->state);
4128 atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4129 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4130 /* don't try to repair!! */
4131 set_bit(STRIPE_INSYNC, &sh->state);
4132 pr_warn_ratelimited("%s: mismatch sector in range "
4133 "%llu-%llu\n", mdname(conf->mddev),
4134 (unsigned long long) sh->sector,
4135 (unsigned long long) sh->sector +
4136 RAID5_STRIPE_SECTORS(conf));
4138 sh->check_state = check_state_compute_run;
4139 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4140 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4141 set_bit(R5_Wantcompute,
4142 &sh->dev[sh->pd_idx].flags);
4143 sh->ops.target = sh->pd_idx;
4144 sh->ops.target2 = -1;
4149 case check_state_compute_run:
4152 pr_err("%s: unknown check_state: %d sector: %llu\n",
4153 __func__, sh->check_state,
4154 (unsigned long long) sh->sector);
4159 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4160 struct stripe_head_state *s,
4163 int pd_idx = sh->pd_idx;
4164 int qd_idx = sh->qd_idx;
4167 BUG_ON(sh->batch_head);
4168 set_bit(STRIPE_HANDLE, &sh->state);
4170 BUG_ON(s->failed > 2);
4172 /* Want to check and possibly repair P and Q.
4173 * However there could be one 'failed' device, in which
4174 * case we can only check one of them, possibly using the
4175 * other to generate missing data
4178 switch (sh->check_state) {
4179 case check_state_idle:
4180 /* start a new check operation if there are < 2 failures */
4181 if (s->failed == s->q_failed) {
4182 /* The only possible failed device holds Q, so it
4183 * makes sense to check P (If anything else were failed,
4184 * we would have used P to recreate it).
4186 sh->check_state = check_state_run;
4188 if (!s->q_failed && s->failed < 2) {
4189 /* Q is not failed, and we didn't use it to generate
4190 * anything, so it makes sense to check it
4192 if (sh->check_state == check_state_run)
4193 sh->check_state = check_state_run_pq;
4195 sh->check_state = check_state_run_q;
4198 /* discard potentially stale zero_sum_result */
4199 sh->ops.zero_sum_result = 0;
4201 if (sh->check_state == check_state_run) {
4202 /* async_xor_zero_sum destroys the contents of P */
4203 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4206 if (sh->check_state >= check_state_run &&
4207 sh->check_state <= check_state_run_pq) {
4208 /* async_syndrome_zero_sum preserves P and Q, so
4209 * no need to mark them !uptodate here
4211 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4215 /* we have 2-disk failure */
4216 BUG_ON(s->failed != 2);
4218 case check_state_compute_result:
4219 sh->check_state = check_state_idle;
4221 /* check that a write has not made the stripe insync */
4222 if (test_bit(STRIPE_INSYNC, &sh->state))
4225 /* now write out any block on a failed drive,
4226 * or P or Q if they were recomputed
4229 if (s->failed == 2) {
4230 dev = &sh->dev[s->failed_num[1]];
4232 set_bit(R5_LOCKED, &dev->flags);
4233 set_bit(R5_Wantwrite, &dev->flags);
4235 if (s->failed >= 1) {
4236 dev = &sh->dev[s->failed_num[0]];
4238 set_bit(R5_LOCKED, &dev->flags);
4239 set_bit(R5_Wantwrite, &dev->flags);
4241 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4242 dev = &sh->dev[pd_idx];
4244 set_bit(R5_LOCKED, &dev->flags);
4245 set_bit(R5_Wantwrite, &dev->flags);
4247 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4248 dev = &sh->dev[qd_idx];
4250 set_bit(R5_LOCKED, &dev->flags);
4251 set_bit(R5_Wantwrite, &dev->flags);
4253 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4254 "%s: disk%td not up to date\n",
4255 mdname(conf->mddev),
4256 dev - (struct r5dev *) &sh->dev)) {
4257 clear_bit(R5_LOCKED, &dev->flags);
4258 clear_bit(R5_Wantwrite, &dev->flags);
4261 clear_bit(STRIPE_DEGRADED, &sh->state);
4263 set_bit(STRIPE_INSYNC, &sh->state);
4265 case check_state_run:
4266 case check_state_run_q:
4267 case check_state_run_pq:
4268 break; /* we will be called again upon completion */
4269 case check_state_check_result:
4270 sh->check_state = check_state_idle;
4272 /* handle a successful check operation, if parity is correct
4273 * we are done. Otherwise update the mismatch count and repair
4274 * parity if !MD_RECOVERY_CHECK
4276 if (sh->ops.zero_sum_result == 0) {
4277 /* both parities are correct */
4279 set_bit(STRIPE_INSYNC, &sh->state);
4281 /* in contrast to the raid5 case we can validate
4282 * parity, but still have a failure to write
4285 sh->check_state = check_state_compute_result;
4286 /* Returning at this point means that we may go
4287 * off and bring p and/or q uptodate again so
4288 * we make sure to check zero_sum_result again
4289 * to verify if p or q need writeback
4293 atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4294 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4295 /* don't try to repair!! */
4296 set_bit(STRIPE_INSYNC, &sh->state);
4297 pr_warn_ratelimited("%s: mismatch sector in range "
4298 "%llu-%llu\n", mdname(conf->mddev),
4299 (unsigned long long) sh->sector,
4300 (unsigned long long) sh->sector +
4301 RAID5_STRIPE_SECTORS(conf));
4303 int *target = &sh->ops.target;
4305 sh->ops.target = -1;
4306 sh->ops.target2 = -1;
4307 sh->check_state = check_state_compute_run;
4308 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4309 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4310 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4311 set_bit(R5_Wantcompute,
4312 &sh->dev[pd_idx].flags);
4314 target = &sh->ops.target2;
4317 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4318 set_bit(R5_Wantcompute,
4319 &sh->dev[qd_idx].flags);
4326 case check_state_compute_run:
4329 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4330 __func__, sh->check_state,
4331 (unsigned long long) sh->sector);
4336 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4340 /* We have read all the blocks in this stripe and now we need to
4341 * copy some of them into a target stripe for expand.
4343 struct dma_async_tx_descriptor *tx = NULL;
4344 BUG_ON(sh->batch_head);
4345 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4346 for (i = 0; i < sh->disks; i++)
4347 if (i != sh->pd_idx && i != sh->qd_idx) {
4349 struct stripe_head *sh2;
4350 struct async_submit_ctl submit;
4352 sector_t bn = raid5_compute_blocknr(sh, i, 1);
4353 sector_t s = raid5_compute_sector(conf, bn, 0,
4355 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4357 /* so far only the early blocks of this stripe
4358 * have been requested. When later blocks
4359 * get requested, we will try again
4362 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4363 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4364 /* must have already done this block */
4365 raid5_release_stripe(sh2);
4369 /* place all the copies on one channel */
4370 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4371 tx = async_memcpy(sh2->dev[dd_idx].page,
4372 sh->dev[i].page, 0, 0, RAID5_STRIPE_SIZE(conf),
4375 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4376 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4377 for (j = 0; j < conf->raid_disks; j++)
4378 if (j != sh2->pd_idx &&
4380 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4382 if (j == conf->raid_disks) {
4383 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4384 set_bit(STRIPE_HANDLE, &sh2->state);
4386 raid5_release_stripe(sh2);
4389 /* done submitting copies, wait for them to complete */
4390 async_tx_quiesce(&tx);
4394 * handle_stripe - do things to a stripe.
4396 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4397 * state of various bits to see what needs to be done.
4399 * return some read requests which now have data
4400 * return some write requests which are safely on storage
4401 * schedule a read on some buffers
4402 * schedule a write of some buffers
4403 * return confirmation of parity correctness
4407 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4409 struct r5conf *conf = sh->raid_conf;
4410 int disks = sh->disks;
4413 int do_recovery = 0;
4415 memset(s, 0, sizeof(*s));
4417 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4418 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4419 s->failed_num[0] = -1;
4420 s->failed_num[1] = -1;
4421 s->log_failed = r5l_log_disk_error(conf);
4423 /* Now to look around and see what can be done */
4425 for (i=disks; i--; ) {
4426 struct md_rdev *rdev;
4433 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4435 dev->toread, dev->towrite, dev->written);
4436 /* maybe we can reply to a read
4438 * new wantfill requests are only permitted while
4439 * ops_complete_biofill is guaranteed to be inactive
4441 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4442 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4443 set_bit(R5_Wantfill, &dev->flags);
4445 /* now count some things */
4446 if (test_bit(R5_LOCKED, &dev->flags))
4448 if (test_bit(R5_UPTODATE, &dev->flags))
4450 if (test_bit(R5_Wantcompute, &dev->flags)) {
4452 BUG_ON(s->compute > 2);
4455 if (test_bit(R5_Wantfill, &dev->flags))
4457 else if (dev->toread)
4461 if (!test_bit(R5_OVERWRITE, &dev->flags))
4466 /* Prefer to use the replacement for reads, but only
4467 * if it is recovered enough and has no bad blocks.
4469 rdev = rcu_dereference(conf->disks[i].replacement);
4470 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4471 rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
4472 !is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4473 &first_bad, &bad_sectors))
4474 set_bit(R5_ReadRepl, &dev->flags);
4476 if (rdev && !test_bit(Faulty, &rdev->flags))
4477 set_bit(R5_NeedReplace, &dev->flags);
4479 clear_bit(R5_NeedReplace, &dev->flags);
4480 rdev = rcu_dereference(conf->disks[i].rdev);
4481 clear_bit(R5_ReadRepl, &dev->flags);
4483 if (rdev && test_bit(Faulty, &rdev->flags))
4486 is_bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4487 &first_bad, &bad_sectors);
4488 if (s->blocked_rdev == NULL
4489 && (test_bit(Blocked, &rdev->flags)
4492 set_bit(BlockedBadBlocks,
4494 s->blocked_rdev = rdev;
4495 atomic_inc(&rdev->nr_pending);
4498 clear_bit(R5_Insync, &dev->flags);
4502 /* also not in-sync */
4503 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4504 test_bit(R5_UPTODATE, &dev->flags)) {
4505 /* treat as in-sync, but with a read error
4506 * which we can now try to correct
4508 set_bit(R5_Insync, &dev->flags);
4509 set_bit(R5_ReadError, &dev->flags);
4511 } else if (test_bit(In_sync, &rdev->flags))
4512 set_bit(R5_Insync, &dev->flags);
4513 else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <= rdev->recovery_offset)
4514 /* in sync if before recovery_offset */
4515 set_bit(R5_Insync, &dev->flags);
4516 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4517 test_bit(R5_Expanded, &dev->flags))
4518 /* If we've reshaped into here, we assume it is Insync.
4519 * We will shortly update recovery_offset to make
4522 set_bit(R5_Insync, &dev->flags);
4524 if (test_bit(R5_WriteError, &dev->flags)) {
4525 /* This flag does not apply to '.replacement'
4526 * only to .rdev, so make sure to check that*/
4527 struct md_rdev *rdev2 = rcu_dereference(
4528 conf->disks[i].rdev);
4530 clear_bit(R5_Insync, &dev->flags);
4531 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4532 s->handle_bad_blocks = 1;
4533 atomic_inc(&rdev2->nr_pending);
4535 clear_bit(R5_WriteError, &dev->flags);
4537 if (test_bit(R5_MadeGood, &dev->flags)) {
4538 /* This flag does not apply to '.replacement'
4539 * only to .rdev, so make sure to check that*/
4540 struct md_rdev *rdev2 = rcu_dereference(
4541 conf->disks[i].rdev);
4542 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4543 s->handle_bad_blocks = 1;
4544 atomic_inc(&rdev2->nr_pending);
4546 clear_bit(R5_MadeGood, &dev->flags);
4548 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4549 struct md_rdev *rdev2 = rcu_dereference(
4550 conf->disks[i].replacement);
4551 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4552 s->handle_bad_blocks = 1;
4553 atomic_inc(&rdev2->nr_pending);
4555 clear_bit(R5_MadeGoodRepl, &dev->flags);
4557 if (!test_bit(R5_Insync, &dev->flags)) {
4558 /* The ReadError flag will just be confusing now */
4559 clear_bit(R5_ReadError, &dev->flags);
4560 clear_bit(R5_ReWrite, &dev->flags);
4562 if (test_bit(R5_ReadError, &dev->flags))
4563 clear_bit(R5_Insync, &dev->flags);
4564 if (!test_bit(R5_Insync, &dev->flags)) {
4566 s->failed_num[s->failed] = i;
4568 if (rdev && !test_bit(Faulty, &rdev->flags))
4571 rdev = rcu_dereference(
4572 conf->disks[i].replacement);
4573 if (rdev && !test_bit(Faulty, &rdev->flags))
4578 if (test_bit(R5_InJournal, &dev->flags))
4580 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4583 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4584 /* If there is a failed device being replaced,
4585 * we must be recovering.
4586 * else if we are after recovery_cp, we must be syncing
4587 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4588 * else we can only be replacing
4589 * sync and recovery both need to read all devices, and so
4590 * use the same flag.
4593 sh->sector >= conf->mddev->recovery_cp ||
4594 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4603 * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4604 * a head which can now be handled.
4606 static int clear_batch_ready(struct stripe_head *sh)
4608 struct stripe_head *tmp;
4609 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4610 return (sh->batch_head && sh->batch_head != sh);
4611 spin_lock(&sh->stripe_lock);
4612 if (!sh->batch_head) {
4613 spin_unlock(&sh->stripe_lock);
4618 * this stripe could be added to a batch list before we check
4619 * BATCH_READY, skips it
4621 if (sh->batch_head != sh) {
4622 spin_unlock(&sh->stripe_lock);
4625 spin_lock(&sh->batch_lock);
4626 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4627 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4628 spin_unlock(&sh->batch_lock);
4629 spin_unlock(&sh->stripe_lock);
4632 * BATCH_READY is cleared, no new stripes can be added.
4633 * batch_list can be accessed without lock
4638 static void break_stripe_batch_list(struct stripe_head *head_sh,
4639 unsigned long handle_flags)
4641 struct stripe_head *sh, *next;
4645 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4647 list_del_init(&sh->batch_list);
4649 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4650 (1 << STRIPE_SYNCING) |
4651 (1 << STRIPE_REPLACED) |
4652 (1 << STRIPE_DELAYED) |
4653 (1 << STRIPE_BIT_DELAY) |
4654 (1 << STRIPE_FULL_WRITE) |
4655 (1 << STRIPE_BIOFILL_RUN) |
4656 (1 << STRIPE_COMPUTE_RUN) |
4657 (1 << STRIPE_DISCARD) |
4658 (1 << STRIPE_BATCH_READY) |
4659 (1 << STRIPE_BATCH_ERR) |
4660 (1 << STRIPE_BITMAP_PENDING)),
4661 "stripe state: %lx\n", sh->state);
4662 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4663 (1 << STRIPE_REPLACED)),
4664 "head stripe state: %lx\n", head_sh->state);
4666 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4667 (1 << STRIPE_PREREAD_ACTIVE) |
4668 (1 << STRIPE_DEGRADED) |
4669 (1 << STRIPE_ON_UNPLUG_LIST)),
4670 head_sh->state & (1 << STRIPE_INSYNC));
4672 sh->check_state = head_sh->check_state;
4673 sh->reconstruct_state = head_sh->reconstruct_state;
4674 spin_lock_irq(&sh->stripe_lock);
4675 sh->batch_head = NULL;
4676 spin_unlock_irq(&sh->stripe_lock);
4677 for (i = 0; i < sh->disks; i++) {
4678 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4680 sh->dev[i].flags = head_sh->dev[i].flags &
4681 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4683 if (handle_flags == 0 ||
4684 sh->state & handle_flags)
4685 set_bit(STRIPE_HANDLE, &sh->state);
4686 raid5_release_stripe(sh);
4688 spin_lock_irq(&head_sh->stripe_lock);
4689 head_sh->batch_head = NULL;
4690 spin_unlock_irq(&head_sh->stripe_lock);
4691 for (i = 0; i < head_sh->disks; i++)
4692 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4694 if (head_sh->state & handle_flags)
4695 set_bit(STRIPE_HANDLE, &head_sh->state);
4698 wake_up(&head_sh->raid_conf->wait_for_overlap);
4701 static void handle_stripe(struct stripe_head *sh)
4703 struct stripe_head_state s;
4704 struct r5conf *conf = sh->raid_conf;
4707 int disks = sh->disks;
4708 struct r5dev *pdev, *qdev;
4710 clear_bit(STRIPE_HANDLE, &sh->state);
4713 * handle_stripe should not continue handle the batched stripe, only
4714 * the head of batch list or lone stripe can continue. Otherwise we
4715 * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
4716 * is set for the batched stripe.
4718 if (clear_batch_ready(sh))
4721 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4722 /* already being handled, ensure it gets handled
4723 * again when current action finishes */
4724 set_bit(STRIPE_HANDLE, &sh->state);
4728 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4729 break_stripe_batch_list(sh, 0);
4731 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4732 spin_lock(&sh->stripe_lock);
4734 * Cannot process 'sync' concurrently with 'discard'.
4735 * Flush data in r5cache before 'sync'.
4737 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4738 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4739 !test_bit(STRIPE_DISCARD, &sh->state) &&
4740 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4741 set_bit(STRIPE_SYNCING, &sh->state);
4742 clear_bit(STRIPE_INSYNC, &sh->state);
4743 clear_bit(STRIPE_REPLACED, &sh->state);
4745 spin_unlock(&sh->stripe_lock);
4747 clear_bit(STRIPE_DELAYED, &sh->state);
4749 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4750 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4751 (unsigned long long)sh->sector, sh->state,
4752 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4753 sh->check_state, sh->reconstruct_state);
4755 analyse_stripe(sh, &s);
4757 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4760 if (s.handle_bad_blocks ||
4761 test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4762 set_bit(STRIPE_HANDLE, &sh->state);
4766 if (unlikely(s.blocked_rdev)) {
4767 if (s.syncing || s.expanding || s.expanded ||
4768 s.replacing || s.to_write || s.written) {
4769 set_bit(STRIPE_HANDLE, &sh->state);
4772 /* There is nothing for the blocked_rdev to block */
4773 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4774 s.blocked_rdev = NULL;
4777 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4778 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4779 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4782 pr_debug("locked=%d uptodate=%d to_read=%d"
4783 " to_write=%d failed=%d failed_num=%d,%d\n",
4784 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4785 s.failed_num[0], s.failed_num[1]);
4787 * check if the array has lost more than max_degraded devices and,
4788 * if so, some requests might need to be failed.
4790 * When journal device failed (log_failed), we will only process
4791 * the stripe if there is data need write to raid disks
4793 if (s.failed > conf->max_degraded ||
4794 (s.log_failed && s.injournal == 0)) {
4795 sh->check_state = 0;
4796 sh->reconstruct_state = 0;
4797 break_stripe_batch_list(sh, 0);
4798 if (s.to_read+s.to_write+s.written)
4799 handle_failed_stripe(conf, sh, &s, disks);
4800 if (s.syncing + s.replacing)
4801 handle_failed_sync(conf, sh, &s);
4804 /* Now we check to see if any write operations have recently
4808 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4810 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4811 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4812 sh->reconstruct_state = reconstruct_state_idle;
4814 /* All the 'written' buffers and the parity block are ready to
4815 * be written back to disk
4817 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4818 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4819 BUG_ON(sh->qd_idx >= 0 &&
4820 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4821 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4822 for (i = disks; i--; ) {
4823 struct r5dev *dev = &sh->dev[i];
4824 if (test_bit(R5_LOCKED, &dev->flags) &&
4825 (i == sh->pd_idx || i == sh->qd_idx ||
4826 dev->written || test_bit(R5_InJournal,
4828 pr_debug("Writing block %d\n", i);
4829 set_bit(R5_Wantwrite, &dev->flags);
4834 if (!test_bit(R5_Insync, &dev->flags) ||
4835 ((i == sh->pd_idx || i == sh->qd_idx) &&
4837 set_bit(STRIPE_INSYNC, &sh->state);
4840 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4841 s.dec_preread_active = 1;
4845 * might be able to return some write requests if the parity blocks
4846 * are safe, or on a failed drive
4848 pdev = &sh->dev[sh->pd_idx];
4849 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4850 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4851 qdev = &sh->dev[sh->qd_idx];
4852 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4853 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4857 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4858 && !test_bit(R5_LOCKED, &pdev->flags)
4859 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4860 test_bit(R5_Discard, &pdev->flags))))) &&
4861 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4862 && !test_bit(R5_LOCKED, &qdev->flags)
4863 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4864 test_bit(R5_Discard, &qdev->flags))))))
4865 handle_stripe_clean_event(conf, sh, disks);
4868 r5c_handle_cached_data_endio(conf, sh, disks);
4869 log_stripe_write_finished(sh);
4871 /* Now we might consider reading some blocks, either to check/generate
4872 * parity, or to satisfy requests
4873 * or to load a block that is being partially written.
4875 if (s.to_read || s.non_overwrite
4876 || (s.to_write && s.failed)
4877 || (s.syncing && (s.uptodate + s.compute < disks))
4880 handle_stripe_fill(sh, &s, disks);
4883 * When the stripe finishes full journal write cycle (write to journal
4884 * and raid disk), this is the clean up procedure so it is ready for
4887 r5c_finish_stripe_write_out(conf, sh, &s);
4890 * Now to consider new write requests, cache write back and what else,
4891 * if anything should be read. We do not handle new writes when:
4892 * 1/ A 'write' operation (copy+xor) is already in flight.
4893 * 2/ A 'check' operation is in flight, as it may clobber the parity
4895 * 3/ A r5c cache log write is in flight.
4898 if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4899 if (!r5c_is_writeback(conf->log)) {
4901 handle_stripe_dirtying(conf, sh, &s, disks);
4902 } else { /* write back cache */
4905 /* First, try handle writes in caching phase */
4907 ret = r5c_try_caching_write(conf, sh, &s,
4910 * If caching phase failed: ret == -EAGAIN
4912 * stripe under reclaim: !caching && injournal
4914 * fall back to handle_stripe_dirtying()
4916 if (ret == -EAGAIN ||
4917 /* stripe under reclaim: !caching && injournal */
4918 (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4920 ret = handle_stripe_dirtying(conf, sh, &s,
4928 /* maybe we need to check and possibly fix the parity for this stripe
4929 * Any reads will already have been scheduled, so we just see if enough
4930 * data is available. The parity check is held off while parity
4931 * dependent operations are in flight.
4933 if (sh->check_state ||
4934 (s.syncing && s.locked == 0 &&
4935 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4936 !test_bit(STRIPE_INSYNC, &sh->state))) {
4937 if (conf->level == 6)
4938 handle_parity_checks6(conf, sh, &s, disks);
4940 handle_parity_checks5(conf, sh, &s, disks);
4943 if ((s.replacing || s.syncing) && s.locked == 0
4944 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4945 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4946 /* Write out to replacement devices where possible */
4947 for (i = 0; i < conf->raid_disks; i++)
4948 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4949 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4950 set_bit(R5_WantReplace, &sh->dev[i].flags);
4951 set_bit(R5_LOCKED, &sh->dev[i].flags);
4955 set_bit(STRIPE_INSYNC, &sh->state);
4956 set_bit(STRIPE_REPLACED, &sh->state);
4958 if ((s.syncing || s.replacing) && s.locked == 0 &&
4959 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4960 test_bit(STRIPE_INSYNC, &sh->state)) {
4961 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
4962 clear_bit(STRIPE_SYNCING, &sh->state);
4963 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4964 wake_up(&conf->wait_for_overlap);
4967 /* If the failed drives are just a ReadError, then we might need
4968 * to progress the repair/check process
4970 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4971 for (i = 0; i < s.failed; i++) {
4972 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4973 if (test_bit(R5_ReadError, &dev->flags)
4974 && !test_bit(R5_LOCKED, &dev->flags)
4975 && test_bit(R5_UPTODATE, &dev->flags)
4977 if (!test_bit(R5_ReWrite, &dev->flags)) {
4978 set_bit(R5_Wantwrite, &dev->flags);
4979 set_bit(R5_ReWrite, &dev->flags);
4981 /* let's read it back */
4982 set_bit(R5_Wantread, &dev->flags);
4983 set_bit(R5_LOCKED, &dev->flags);
4988 /* Finish reconstruct operations initiated by the expansion process */
4989 if (sh->reconstruct_state == reconstruct_state_result) {
4990 struct stripe_head *sh_src
4991 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4992 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4993 /* sh cannot be written until sh_src has been read.
4994 * so arrange for sh to be delayed a little
4996 set_bit(STRIPE_DELAYED, &sh->state);
4997 set_bit(STRIPE_HANDLE, &sh->state);
4998 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
5000 atomic_inc(&conf->preread_active_stripes);
5001 raid5_release_stripe(sh_src);
5005 raid5_release_stripe(sh_src);
5007 sh->reconstruct_state = reconstruct_state_idle;
5008 clear_bit(STRIPE_EXPANDING, &sh->state);
5009 for (i = conf->raid_disks; i--; ) {
5010 set_bit(R5_Wantwrite, &sh->dev[i].flags);
5011 set_bit(R5_LOCKED, &sh->dev[i].flags);
5016 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
5017 !sh->reconstruct_state) {
5018 /* Need to write out all blocks after computing parity */
5019 sh->disks = conf->raid_disks;
5020 stripe_set_idx(sh->sector, conf, 0, sh);
5021 schedule_reconstruction(sh, &s, 1, 1);
5022 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
5023 clear_bit(STRIPE_EXPAND_READY, &sh->state);
5024 atomic_dec(&conf->reshape_stripes);
5025 wake_up(&conf->wait_for_overlap);
5026 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5029 if (s.expanding && s.locked == 0 &&
5030 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
5031 handle_stripe_expansion(conf, sh);
5034 /* wait for this device to become unblocked */
5035 if (unlikely(s.blocked_rdev)) {
5036 if (conf->mddev->external)
5037 md_wait_for_blocked_rdev(s.blocked_rdev,
5040 /* Internal metadata will immediately
5041 * be written by raid5d, so we don't
5042 * need to wait here.
5044 rdev_dec_pending(s.blocked_rdev,
5048 if (s.handle_bad_blocks)
5049 for (i = disks; i--; ) {
5050 struct md_rdev *rdev;
5051 struct r5dev *dev = &sh->dev[i];
5052 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5053 /* We own a safe reference to the rdev */
5054 rdev = conf->disks[i].rdev;
5055 if (!rdev_set_badblocks(rdev, sh->sector,
5056 RAID5_STRIPE_SECTORS(conf), 0))
5057 md_error(conf->mddev, rdev);
5058 rdev_dec_pending(rdev, conf->mddev);
5060 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5061 rdev = conf->disks[i].rdev;
5062 rdev_clear_badblocks(rdev, sh->sector,
5063 RAID5_STRIPE_SECTORS(conf), 0);
5064 rdev_dec_pending(rdev, conf->mddev);
5066 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5067 rdev = conf->disks[i].replacement;
5069 /* rdev have been moved down */
5070 rdev = conf->disks[i].rdev;
5071 rdev_clear_badblocks(rdev, sh->sector,
5072 RAID5_STRIPE_SECTORS(conf), 0);
5073 rdev_dec_pending(rdev, conf->mddev);
5078 raid_run_ops(sh, s.ops_request);
5082 if (s.dec_preread_active) {
5083 /* We delay this until after ops_run_io so that if make_request
5084 * is waiting on a flush, it won't continue until the writes
5085 * have actually been submitted.
5087 atomic_dec(&conf->preread_active_stripes);
5088 if (atomic_read(&conf->preread_active_stripes) <
5090 md_wakeup_thread(conf->mddev->thread);
5093 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5096 static void raid5_activate_delayed(struct r5conf *conf)
5098 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5099 while (!list_empty(&conf->delayed_list)) {
5100 struct list_head *l = conf->delayed_list.next;
5101 struct stripe_head *sh;
5102 sh = list_entry(l, struct stripe_head, lru);
5104 clear_bit(STRIPE_DELAYED, &sh->state);
5105 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5106 atomic_inc(&conf->preread_active_stripes);
5107 list_add_tail(&sh->lru, &conf->hold_list);
5108 raid5_wakeup_stripe_thread(sh);
5113 static void activate_bit_delay(struct r5conf *conf,
5114 struct list_head *temp_inactive_list)
5116 /* device_lock is held */
5117 struct list_head head;
5118 list_add(&head, &conf->bitmap_list);
5119 list_del_init(&conf->bitmap_list);
5120 while (!list_empty(&head)) {
5121 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5123 list_del_init(&sh->lru);
5124 atomic_inc(&sh->count);
5125 hash = sh->hash_lock_index;
5126 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5130 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5132 struct r5conf *conf = mddev->private;
5133 sector_t sector = bio->bi_iter.bi_sector;
5134 unsigned int chunk_sectors;
5135 unsigned int bio_sectors = bio_sectors(bio);
5137 WARN_ON_ONCE(bio->bi_partno);
5139 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5140 return chunk_sectors >=
5141 ((sector & (chunk_sectors - 1)) + bio_sectors);
5145 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5146 * later sampled by raid5d.
5148 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5150 unsigned long flags;
5152 spin_lock_irqsave(&conf->device_lock, flags);
5154 bi->bi_next = conf->retry_read_aligned_list;
5155 conf->retry_read_aligned_list = bi;
5157 spin_unlock_irqrestore(&conf->device_lock, flags);
5158 md_wakeup_thread(conf->mddev->thread);
5161 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5162 unsigned int *offset)
5166 bi = conf->retry_read_aligned;
5168 *offset = conf->retry_read_offset;
5169 conf->retry_read_aligned = NULL;
5172 bi = conf->retry_read_aligned_list;
5174 conf->retry_read_aligned_list = bi->bi_next;
5183 * The "raid5_align_endio" should check if the read succeeded and if it
5184 * did, call bio_endio on the original bio (having bio_put the new bio
5186 * If the read failed..
5188 static void raid5_align_endio(struct bio *bi)
5190 struct bio* raid_bi = bi->bi_private;
5191 struct mddev *mddev;
5192 struct r5conf *conf;
5193 struct md_rdev *rdev;
5194 blk_status_t error = bi->bi_status;
5198 rdev = (void*)raid_bi->bi_next;
5199 raid_bi->bi_next = NULL;
5200 mddev = rdev->mddev;
5201 conf = mddev->private;
5203 rdev_dec_pending(rdev, conf->mddev);
5207 if (atomic_dec_and_test(&conf->active_aligned_reads))
5208 wake_up(&conf->wait_for_quiescent);
5212 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5214 add_bio_to_retry(raid_bi, conf);
5217 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5219 struct r5conf *conf = mddev->private;
5221 struct bio* align_bi;
5222 struct md_rdev *rdev;
5223 sector_t end_sector;
5225 if (!in_chunk_boundary(mddev, raid_bio)) {
5226 pr_debug("%s: non aligned\n", __func__);
5230 * use bio_clone_fast to make a copy of the bio
5232 align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5236 * set bi_end_io to a new function, and set bi_private to the
5239 align_bi->bi_end_io = raid5_align_endio;
5240 align_bi->bi_private = raid_bio;
5244 align_bi->bi_iter.bi_sector =
5245 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5248 end_sector = bio_end_sector(align_bi);
5250 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5251 if (!rdev || test_bit(Faulty, &rdev->flags) ||
5252 rdev->recovery_offset < end_sector) {
5253 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5255 (test_bit(Faulty, &rdev->flags) ||
5256 !(test_bit(In_sync, &rdev->flags) ||
5257 rdev->recovery_offset >= end_sector)))
5261 if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5271 atomic_inc(&rdev->nr_pending);
5273 raid_bio->bi_next = (void*)rdev;
5274 bio_set_dev(align_bi, rdev->bdev);
5276 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5277 bio_sectors(align_bi),
5278 &first_bad, &bad_sectors)) {
5280 rdev_dec_pending(rdev, mddev);
5284 /* No reshape active, so we can trust rdev->data_offset */
5285 align_bi->bi_iter.bi_sector += rdev->data_offset;
5287 spin_lock_irq(&conf->device_lock);
5288 wait_event_lock_irq(conf->wait_for_quiescent,
5291 atomic_inc(&conf->active_aligned_reads);
5292 spin_unlock_irq(&conf->device_lock);
5295 trace_block_bio_remap(align_bi->bi_disk->queue,
5296 align_bi, disk_devt(mddev->gendisk),
5297 raid_bio->bi_iter.bi_sector);
5298 submit_bio_noacct(align_bi);
5307 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5310 sector_t sector = raid_bio->bi_iter.bi_sector;
5311 unsigned chunk_sects = mddev->chunk_sectors;
5312 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5314 if (sectors < bio_sectors(raid_bio)) {
5315 struct r5conf *conf = mddev->private;
5316 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5317 bio_chain(split, raid_bio);
5318 submit_bio_noacct(raid_bio);
5322 if (!raid5_read_one_chunk(mddev, raid_bio))
5328 /* __get_priority_stripe - get the next stripe to process
5330 * Full stripe writes are allowed to pass preread active stripes up until
5331 * the bypass_threshold is exceeded. In general the bypass_count
5332 * increments when the handle_list is handled before the hold_list; however, it
5333 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5334 * stripe with in flight i/o. The bypass_count will be reset when the
5335 * head of the hold_list has changed, i.e. the head was promoted to the
5338 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5340 struct stripe_head *sh, *tmp;
5341 struct list_head *handle_list = NULL;
5342 struct r5worker_group *wg;
5343 bool second_try = !r5c_is_writeback(conf->log) &&
5344 !r5l_log_disk_error(conf);
5345 bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5346 r5l_log_disk_error(conf);
5351 if (conf->worker_cnt_per_group == 0) {
5352 handle_list = try_loprio ? &conf->loprio_list :
5354 } else if (group != ANY_GROUP) {
5355 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5356 &conf->worker_groups[group].handle_list;
5357 wg = &conf->worker_groups[group];
5360 for (i = 0; i < conf->group_cnt; i++) {
5361 handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5362 &conf->worker_groups[i].handle_list;
5363 wg = &conf->worker_groups[i];
5364 if (!list_empty(handle_list))
5369 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5371 list_empty(handle_list) ? "empty" : "busy",
5372 list_empty(&conf->hold_list) ? "empty" : "busy",
5373 atomic_read(&conf->pending_full_writes), conf->bypass_count);
5375 if (!list_empty(handle_list)) {
5376 sh = list_entry(handle_list->next, typeof(*sh), lru);
5378 if (list_empty(&conf->hold_list))
5379 conf->bypass_count = 0;
5380 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5381 if (conf->hold_list.next == conf->last_hold)
5382 conf->bypass_count++;
5384 conf->last_hold = conf->hold_list.next;
5385 conf->bypass_count -= conf->bypass_threshold;
5386 if (conf->bypass_count < 0)
5387 conf->bypass_count = 0;
5390 } else if (!list_empty(&conf->hold_list) &&
5391 ((conf->bypass_threshold &&
5392 conf->bypass_count > conf->bypass_threshold) ||
5393 atomic_read(&conf->pending_full_writes) == 0)) {
5395 list_for_each_entry(tmp, &conf->hold_list, lru) {
5396 if (conf->worker_cnt_per_group == 0 ||
5397 group == ANY_GROUP ||
5398 !cpu_online(tmp->cpu) ||
5399 cpu_to_group(tmp->cpu) == group) {
5406 conf->bypass_count -= conf->bypass_threshold;
5407 if (conf->bypass_count < 0)
5408 conf->bypass_count = 0;
5417 try_loprio = !try_loprio;
5425 list_del_init(&sh->lru);
5426 BUG_ON(atomic_inc_return(&sh->count) != 1);
5430 struct raid5_plug_cb {
5431 struct blk_plug_cb cb;
5432 struct list_head list;
5433 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5436 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5438 struct raid5_plug_cb *cb = container_of(
5439 blk_cb, struct raid5_plug_cb, cb);
5440 struct stripe_head *sh;
5441 struct mddev *mddev = cb->cb.data;
5442 struct r5conf *conf = mddev->private;
5446 if (cb->list.next && !list_empty(&cb->list)) {
5447 spin_lock_irq(&conf->device_lock);
5448 while (!list_empty(&cb->list)) {
5449 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5450 list_del_init(&sh->lru);
5452 * avoid race release_stripe_plug() sees
5453 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5454 * is still in our list
5456 smp_mb__before_atomic();
5457 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5459 * STRIPE_ON_RELEASE_LIST could be set here. In that
5460 * case, the count is always > 1 here
5462 hash = sh->hash_lock_index;
5463 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5466 spin_unlock_irq(&conf->device_lock);
5468 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5469 NR_STRIPE_HASH_LOCKS);
5471 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5475 static void release_stripe_plug(struct mddev *mddev,
5476 struct stripe_head *sh)
5478 struct blk_plug_cb *blk_cb = blk_check_plugged(
5479 raid5_unplug, mddev,
5480 sizeof(struct raid5_plug_cb));
5481 struct raid5_plug_cb *cb;
5484 raid5_release_stripe(sh);
5488 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5490 if (cb->list.next == NULL) {
5492 INIT_LIST_HEAD(&cb->list);
5493 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5494 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5497 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5498 list_add_tail(&sh->lru, &cb->list);
5500 raid5_release_stripe(sh);
5503 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5505 struct r5conf *conf = mddev->private;
5506 sector_t logical_sector, last_sector;
5507 struct stripe_head *sh;
5510 if (mddev->reshape_position != MaxSector)
5511 /* Skip discard while reshape is happening */
5514 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5515 last_sector = bio_end_sector(bi);
5519 stripe_sectors = conf->chunk_sectors *
5520 (conf->raid_disks - conf->max_degraded);
5521 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5523 sector_div(last_sector, stripe_sectors);
5525 logical_sector *= conf->chunk_sectors;
5526 last_sector *= conf->chunk_sectors;
5528 for (; logical_sector < last_sector;
5529 logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5533 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5534 prepare_to_wait(&conf->wait_for_overlap, &w,
5535 TASK_UNINTERRUPTIBLE);
5536 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5537 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5538 raid5_release_stripe(sh);
5542 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5543 spin_lock_irq(&sh->stripe_lock);
5544 for (d = 0; d < conf->raid_disks; d++) {
5545 if (d == sh->pd_idx || d == sh->qd_idx)
5547 if (sh->dev[d].towrite || sh->dev[d].toread) {
5548 set_bit(R5_Overlap, &sh->dev[d].flags);
5549 spin_unlock_irq(&sh->stripe_lock);
5550 raid5_release_stripe(sh);
5555 set_bit(STRIPE_DISCARD, &sh->state);
5556 finish_wait(&conf->wait_for_overlap, &w);
5557 sh->overwrite_disks = 0;
5558 for (d = 0; d < conf->raid_disks; d++) {
5559 if (d == sh->pd_idx || d == sh->qd_idx)
5561 sh->dev[d].towrite = bi;
5562 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5563 bio_inc_remaining(bi);
5564 md_write_inc(mddev, bi);
5565 sh->overwrite_disks++;
5567 spin_unlock_irq(&sh->stripe_lock);
5568 if (conf->mddev->bitmap) {
5570 d < conf->raid_disks - conf->max_degraded;
5572 md_bitmap_startwrite(mddev->bitmap,
5574 RAID5_STRIPE_SECTORS(conf),
5576 sh->bm_seq = conf->seq_flush + 1;
5577 set_bit(STRIPE_BIT_DELAY, &sh->state);
5580 set_bit(STRIPE_HANDLE, &sh->state);
5581 clear_bit(STRIPE_DELAYED, &sh->state);
5582 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5583 atomic_inc(&conf->preread_active_stripes);
5584 release_stripe_plug(mddev, sh);
5590 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5592 struct r5conf *conf = mddev->private;
5594 sector_t new_sector;
5595 sector_t logical_sector, last_sector;
5596 struct stripe_head *sh;
5597 const int rw = bio_data_dir(bi);
5600 bool do_flush = false;
5602 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5603 int ret = log_handle_flush_request(conf, bi);
5607 if (ret == -ENODEV) {
5608 if (md_flush_request(mddev, bi))
5611 /* ret == -EAGAIN, fallback */
5613 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5614 * we need to flush journal device
5616 do_flush = bi->bi_opf & REQ_PREFLUSH;
5619 if (!md_write_start(mddev, bi))
5622 * If array is degraded, better not do chunk aligned read because
5623 * later we might have to read it again in order to reconstruct
5624 * data on failed drives.
5626 if (rw == READ && mddev->degraded == 0 &&
5627 mddev->reshape_position == MaxSector) {
5628 bi = chunk_aligned_read(mddev, bi);
5633 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5634 make_discard_request(mddev, bi);
5635 md_write_end(mddev);
5639 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5640 last_sector = bio_end_sector(bi);
5643 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5644 for (; logical_sector < last_sector; logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5650 seq = read_seqcount_begin(&conf->gen_lock);
5653 prepare_to_wait(&conf->wait_for_overlap, &w,
5654 TASK_UNINTERRUPTIBLE);
5655 if (unlikely(conf->reshape_progress != MaxSector)) {
5656 /* spinlock is needed as reshape_progress may be
5657 * 64bit on a 32bit platform, and so it might be
5658 * possible to see a half-updated value
5659 * Of course reshape_progress could change after
5660 * the lock is dropped, so once we get a reference
5661 * to the stripe that we think it is, we will have
5664 spin_lock_irq(&conf->device_lock);
5665 if (mddev->reshape_backwards
5666 ? logical_sector < conf->reshape_progress
5667 : logical_sector >= conf->reshape_progress) {
5670 if (mddev->reshape_backwards
5671 ? logical_sector < conf->reshape_safe
5672 : logical_sector >= conf->reshape_safe) {
5673 spin_unlock_irq(&conf->device_lock);
5679 spin_unlock_irq(&conf->device_lock);
5682 new_sector = raid5_compute_sector(conf, logical_sector,
5685 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5686 (unsigned long long)new_sector,
5687 (unsigned long long)logical_sector);
5689 sh = raid5_get_active_stripe(conf, new_sector, previous,
5690 (bi->bi_opf & REQ_RAHEAD), 0);
5692 if (unlikely(previous)) {
5693 /* expansion might have moved on while waiting for a
5694 * stripe, so we must do the range check again.
5695 * Expansion could still move past after this
5696 * test, but as we are holding a reference to
5697 * 'sh', we know that if that happens,
5698 * STRIPE_EXPANDING will get set and the expansion
5699 * won't proceed until we finish with the stripe.
5702 spin_lock_irq(&conf->device_lock);
5703 if (mddev->reshape_backwards
5704 ? logical_sector >= conf->reshape_progress
5705 : logical_sector < conf->reshape_progress)
5706 /* mismatch, need to try again */
5708 spin_unlock_irq(&conf->device_lock);
5710 raid5_release_stripe(sh);
5716 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5717 /* Might have got the wrong stripe_head
5720 raid5_release_stripe(sh);
5724 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5725 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5726 /* Stripe is busy expanding or
5727 * add failed due to overlap. Flush everything
5730 md_wakeup_thread(mddev->thread);
5731 raid5_release_stripe(sh);
5737 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5738 /* we only need flush for one stripe */
5742 set_bit(STRIPE_HANDLE, &sh->state);
5743 clear_bit(STRIPE_DELAYED, &sh->state);
5744 if ((!sh->batch_head || sh == sh->batch_head) &&
5745 (bi->bi_opf & REQ_SYNC) &&
5746 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5747 atomic_inc(&conf->preread_active_stripes);
5748 release_stripe_plug(mddev, sh);
5750 /* cannot get stripe for read-ahead, just give-up */
5751 bi->bi_status = BLK_STS_IOERR;
5755 finish_wait(&conf->wait_for_overlap, &w);
5758 md_write_end(mddev);
5763 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5765 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5767 /* reshaping is quite different to recovery/resync so it is
5768 * handled quite separately ... here.
5770 * On each call to sync_request, we gather one chunk worth of
5771 * destination stripes and flag them as expanding.
5772 * Then we find all the source stripes and request reads.
5773 * As the reads complete, handle_stripe will copy the data
5774 * into the destination stripe and release that stripe.
5776 struct r5conf *conf = mddev->private;
5777 struct stripe_head *sh;
5778 struct md_rdev *rdev;
5779 sector_t first_sector, last_sector;
5780 int raid_disks = conf->previous_raid_disks;
5781 int data_disks = raid_disks - conf->max_degraded;
5782 int new_data_disks = conf->raid_disks - conf->max_degraded;
5785 sector_t writepos, readpos, safepos;
5786 sector_t stripe_addr;
5787 int reshape_sectors;
5788 struct list_head stripes;
5791 if (sector_nr == 0) {
5792 /* If restarting in the middle, skip the initial sectors */
5793 if (mddev->reshape_backwards &&
5794 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5795 sector_nr = raid5_size(mddev, 0, 0)
5796 - conf->reshape_progress;
5797 } else if (mddev->reshape_backwards &&
5798 conf->reshape_progress == MaxSector) {
5799 /* shouldn't happen, but just in case, finish up.*/
5800 sector_nr = MaxSector;
5801 } else if (!mddev->reshape_backwards &&
5802 conf->reshape_progress > 0)
5803 sector_nr = conf->reshape_progress;
5804 sector_div(sector_nr, new_data_disks);
5806 mddev->curr_resync_completed = sector_nr;
5807 sysfs_notify_dirent_safe(mddev->sysfs_completed);
5814 /* We need to process a full chunk at a time.
5815 * If old and new chunk sizes differ, we need to process the
5819 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5821 /* We update the metadata at least every 10 seconds, or when
5822 * the data about to be copied would over-write the source of
5823 * the data at the front of the range. i.e. one new_stripe
5824 * along from reshape_progress new_maps to after where
5825 * reshape_safe old_maps to
5827 writepos = conf->reshape_progress;
5828 sector_div(writepos, new_data_disks);
5829 readpos = conf->reshape_progress;
5830 sector_div(readpos, data_disks);
5831 safepos = conf->reshape_safe;
5832 sector_div(safepos, data_disks);
5833 if (mddev->reshape_backwards) {
5834 BUG_ON(writepos < reshape_sectors);
5835 writepos -= reshape_sectors;
5836 readpos += reshape_sectors;
5837 safepos += reshape_sectors;
5839 writepos += reshape_sectors;
5840 /* readpos and safepos are worst-case calculations.
5841 * A negative number is overly pessimistic, and causes
5842 * obvious problems for unsigned storage. So clip to 0.
5844 readpos -= min_t(sector_t, reshape_sectors, readpos);
5845 safepos -= min_t(sector_t, reshape_sectors, safepos);
5848 /* Having calculated the 'writepos' possibly use it
5849 * to set 'stripe_addr' which is where we will write to.
5851 if (mddev->reshape_backwards) {
5852 BUG_ON(conf->reshape_progress == 0);
5853 stripe_addr = writepos;
5854 BUG_ON((mddev->dev_sectors &
5855 ~((sector_t)reshape_sectors - 1))
5856 - reshape_sectors - stripe_addr
5859 BUG_ON(writepos != sector_nr + reshape_sectors);
5860 stripe_addr = sector_nr;
5863 /* 'writepos' is the most advanced device address we might write.
5864 * 'readpos' is the least advanced device address we might read.
5865 * 'safepos' is the least address recorded in the metadata as having
5867 * If there is a min_offset_diff, these are adjusted either by
5868 * increasing the safepos/readpos if diff is negative, or
5869 * increasing writepos if diff is positive.
5870 * If 'readpos' is then behind 'writepos', there is no way that we can
5871 * ensure safety in the face of a crash - that must be done by userspace
5872 * making a backup of the data. So in that case there is no particular
5873 * rush to update metadata.
5874 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5875 * update the metadata to advance 'safepos' to match 'readpos' so that
5876 * we can be safe in the event of a crash.
5877 * So we insist on updating metadata if safepos is behind writepos and
5878 * readpos is beyond writepos.
5879 * In any case, update the metadata every 10 seconds.
5880 * Maybe that number should be configurable, but I'm not sure it is
5881 * worth it.... maybe it could be a multiple of safemode_delay???
5883 if (conf->min_offset_diff < 0) {
5884 safepos += -conf->min_offset_diff;
5885 readpos += -conf->min_offset_diff;
5887 writepos += conf->min_offset_diff;
5889 if ((mddev->reshape_backwards
5890 ? (safepos > writepos && readpos < writepos)
5891 : (safepos < writepos && readpos > writepos)) ||
5892 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5893 /* Cannot proceed until we've updated the superblock... */
5894 wait_event(conf->wait_for_overlap,
5895 atomic_read(&conf->reshape_stripes)==0
5896 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5897 if (atomic_read(&conf->reshape_stripes) != 0)
5899 mddev->reshape_position = conf->reshape_progress;
5900 mddev->curr_resync_completed = sector_nr;
5901 if (!mddev->reshape_backwards)
5902 /* Can update recovery_offset */
5903 rdev_for_each(rdev, mddev)
5904 if (rdev->raid_disk >= 0 &&
5905 !test_bit(Journal, &rdev->flags) &&
5906 !test_bit(In_sync, &rdev->flags) &&
5907 rdev->recovery_offset < sector_nr)
5908 rdev->recovery_offset = sector_nr;
5910 conf->reshape_checkpoint = jiffies;
5911 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5912 md_wakeup_thread(mddev->thread);
5913 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5914 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5915 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5917 spin_lock_irq(&conf->device_lock);
5918 conf->reshape_safe = mddev->reshape_position;
5919 spin_unlock_irq(&conf->device_lock);
5920 wake_up(&conf->wait_for_overlap);
5921 sysfs_notify_dirent_safe(mddev->sysfs_completed);
5924 INIT_LIST_HEAD(&stripes);
5925 for (i = 0; i < reshape_sectors; i += RAID5_STRIPE_SECTORS(conf)) {
5927 int skipped_disk = 0;
5928 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5929 set_bit(STRIPE_EXPANDING, &sh->state);
5930 atomic_inc(&conf->reshape_stripes);
5931 /* If any of this stripe is beyond the end of the old
5932 * array, then we need to zero those blocks
5934 for (j=sh->disks; j--;) {
5936 if (j == sh->pd_idx)
5938 if (conf->level == 6 &&
5941 s = raid5_compute_blocknr(sh, j, 0);
5942 if (s < raid5_size(mddev, 0, 0)) {
5946 memset(page_address(sh->dev[j].page), 0, RAID5_STRIPE_SIZE(conf));
5947 set_bit(R5_Expanded, &sh->dev[j].flags);
5948 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5950 if (!skipped_disk) {
5951 set_bit(STRIPE_EXPAND_READY, &sh->state);
5952 set_bit(STRIPE_HANDLE, &sh->state);
5954 list_add(&sh->lru, &stripes);
5956 spin_lock_irq(&conf->device_lock);
5957 if (mddev->reshape_backwards)
5958 conf->reshape_progress -= reshape_sectors * new_data_disks;
5960 conf->reshape_progress += reshape_sectors * new_data_disks;
5961 spin_unlock_irq(&conf->device_lock);
5962 /* Ok, those stripe are ready. We can start scheduling
5963 * reads on the source stripes.
5964 * The source stripes are determined by mapping the first and last
5965 * block on the destination stripes.
5968 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5971 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5972 * new_data_disks - 1),
5974 if (last_sector >= mddev->dev_sectors)
5975 last_sector = mddev->dev_sectors - 1;
5976 while (first_sector <= last_sector) {
5977 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5978 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5979 set_bit(STRIPE_HANDLE, &sh->state);
5980 raid5_release_stripe(sh);
5981 first_sector += RAID5_STRIPE_SECTORS(conf);
5983 /* Now that the sources are clearly marked, we can release
5984 * the destination stripes
5986 while (!list_empty(&stripes)) {
5987 sh = list_entry(stripes.next, struct stripe_head, lru);
5988 list_del_init(&sh->lru);
5989 raid5_release_stripe(sh);
5991 /* If this takes us to the resync_max point where we have to pause,
5992 * then we need to write out the superblock.
5994 sector_nr += reshape_sectors;
5995 retn = reshape_sectors;
5997 if (mddev->curr_resync_completed > mddev->resync_max ||
5998 (sector_nr - mddev->curr_resync_completed) * 2
5999 >= mddev->resync_max - mddev->curr_resync_completed) {
6000 /* Cannot proceed until we've updated the superblock... */
6001 wait_event(conf->wait_for_overlap,
6002 atomic_read(&conf->reshape_stripes) == 0
6003 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6004 if (atomic_read(&conf->reshape_stripes) != 0)
6006 mddev->reshape_position = conf->reshape_progress;
6007 mddev->curr_resync_completed = sector_nr;
6008 if (!mddev->reshape_backwards)
6009 /* Can update recovery_offset */
6010 rdev_for_each(rdev, mddev)
6011 if (rdev->raid_disk >= 0 &&
6012 !test_bit(Journal, &rdev->flags) &&
6013 !test_bit(In_sync, &rdev->flags) &&
6014 rdev->recovery_offset < sector_nr)
6015 rdev->recovery_offset = sector_nr;
6016 conf->reshape_checkpoint = jiffies;
6017 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6018 md_wakeup_thread(mddev->thread);
6019 wait_event(mddev->sb_wait,
6020 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6021 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6022 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6024 spin_lock_irq(&conf->device_lock);
6025 conf->reshape_safe = mddev->reshape_position;
6026 spin_unlock_irq(&conf->device_lock);
6027 wake_up(&conf->wait_for_overlap);
6028 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6034 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6037 struct r5conf *conf = mddev->private;
6038 struct stripe_head *sh;
6039 sector_t max_sector = mddev->dev_sectors;
6040 sector_t sync_blocks;
6041 int still_degraded = 0;
6044 if (sector_nr >= max_sector) {
6045 /* just being told to finish up .. nothing much to do */
6047 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6052 if (mddev->curr_resync < max_sector) /* aborted */
6053 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6055 else /* completed sync */
6057 md_bitmap_close_sync(mddev->bitmap);
6062 /* Allow raid5_quiesce to complete */
6063 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6065 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6066 return reshape_request(mddev, sector_nr, skipped);
6068 /* No need to check resync_max as we never do more than one
6069 * stripe, and as resync_max will always be on a chunk boundary,
6070 * if the check in md_do_sync didn't fire, there is no chance
6071 * of overstepping resync_max here
6074 /* if there is too many failed drives and we are trying
6075 * to resync, then assert that we are finished, because there is
6076 * nothing we can do.
6078 if (mddev->degraded >= conf->max_degraded &&
6079 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6080 sector_t rv = mddev->dev_sectors - sector_nr;
6084 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6086 !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6087 sync_blocks >= RAID5_STRIPE_SECTORS(conf)) {
6088 /* we can skip this block, and probably more */
6089 do_div(sync_blocks, RAID5_STRIPE_SECTORS(conf));
6091 /* keep things rounded to whole stripes */
6092 return sync_blocks * RAID5_STRIPE_SECTORS(conf);
6095 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6097 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6099 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6100 /* make sure we don't swamp the stripe cache if someone else
6101 * is trying to get access
6103 schedule_timeout_uninterruptible(1);
6105 /* Need to check if array will still be degraded after recovery/resync
6106 * Note in case of > 1 drive failures it's possible we're rebuilding
6107 * one drive while leaving another faulty drive in array.
6110 for (i = 0; i < conf->raid_disks; i++) {
6111 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6113 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6118 md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6120 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6121 set_bit(STRIPE_HANDLE, &sh->state);
6123 raid5_release_stripe(sh);
6125 return RAID5_STRIPE_SECTORS(conf);
6128 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6129 unsigned int offset)
6131 /* We may not be able to submit a whole bio at once as there
6132 * may not be enough stripe_heads available.
6133 * We cannot pre-allocate enough stripe_heads as we may need
6134 * more than exist in the cache (if we allow ever large chunks).
6135 * So we do one stripe head at a time and record in
6136 * ->bi_hw_segments how many have been done.
6138 * We *know* that this entire raid_bio is in one chunk, so
6139 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6141 struct stripe_head *sh;
6143 sector_t sector, logical_sector, last_sector;
6147 logical_sector = raid_bio->bi_iter.bi_sector &
6148 ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6149 sector = raid5_compute_sector(conf, logical_sector,
6151 last_sector = bio_end_sector(raid_bio);
6153 for (; logical_sector < last_sector;
6154 logical_sector += RAID5_STRIPE_SECTORS(conf),
6155 sector += RAID5_STRIPE_SECTORS(conf),
6159 /* already done this stripe */
6162 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6165 /* failed to get a stripe - must wait */
6166 conf->retry_read_aligned = raid_bio;
6167 conf->retry_read_offset = scnt;
6171 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6172 raid5_release_stripe(sh);
6173 conf->retry_read_aligned = raid_bio;
6174 conf->retry_read_offset = scnt;
6178 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6180 raid5_release_stripe(sh);
6184 bio_endio(raid_bio);
6186 if (atomic_dec_and_test(&conf->active_aligned_reads))
6187 wake_up(&conf->wait_for_quiescent);
6191 static int handle_active_stripes(struct r5conf *conf, int group,
6192 struct r5worker *worker,
6193 struct list_head *temp_inactive_list)
6194 __releases(&conf->device_lock)
6195 __acquires(&conf->device_lock)
6197 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6198 int i, batch_size = 0, hash;
6199 bool release_inactive = false;
6201 while (batch_size < MAX_STRIPE_BATCH &&
6202 (sh = __get_priority_stripe(conf, group)) != NULL)
6203 batch[batch_size++] = sh;
6205 if (batch_size == 0) {
6206 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6207 if (!list_empty(temp_inactive_list + i))
6209 if (i == NR_STRIPE_HASH_LOCKS) {
6210 spin_unlock_irq(&conf->device_lock);
6211 log_flush_stripe_to_raid(conf);
6212 spin_lock_irq(&conf->device_lock);
6215 release_inactive = true;
6217 spin_unlock_irq(&conf->device_lock);
6219 release_inactive_stripe_list(conf, temp_inactive_list,
6220 NR_STRIPE_HASH_LOCKS);
6222 r5l_flush_stripe_to_raid(conf->log);
6223 if (release_inactive) {
6224 spin_lock_irq(&conf->device_lock);
6228 for (i = 0; i < batch_size; i++)
6229 handle_stripe(batch[i]);
6230 log_write_stripe_run(conf);
6234 spin_lock_irq(&conf->device_lock);
6235 for (i = 0; i < batch_size; i++) {
6236 hash = batch[i]->hash_lock_index;
6237 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6242 static void raid5_do_work(struct work_struct *work)
6244 struct r5worker *worker = container_of(work, struct r5worker, work);
6245 struct r5worker_group *group = worker->group;
6246 struct r5conf *conf = group->conf;
6247 struct mddev *mddev = conf->mddev;
6248 int group_id = group - conf->worker_groups;
6250 struct blk_plug plug;
6252 pr_debug("+++ raid5worker active\n");
6254 blk_start_plug(&plug);
6256 spin_lock_irq(&conf->device_lock);
6258 int batch_size, released;
6260 released = release_stripe_list(conf, worker->temp_inactive_list);
6262 batch_size = handle_active_stripes(conf, group_id, worker,
6263 worker->temp_inactive_list);
6264 worker->working = false;
6265 if (!batch_size && !released)
6267 handled += batch_size;
6268 wait_event_lock_irq(mddev->sb_wait,
6269 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6272 pr_debug("%d stripes handled\n", handled);
6274 spin_unlock_irq(&conf->device_lock);
6276 flush_deferred_bios(conf);
6278 r5l_flush_stripe_to_raid(conf->log);
6280 async_tx_issue_pending_all();
6281 blk_finish_plug(&plug);
6283 pr_debug("--- raid5worker inactive\n");
6287 * This is our raid5 kernel thread.
6289 * We scan the hash table for stripes which can be handled now.
6290 * During the scan, completed stripes are saved for us by the interrupt
6291 * handler, so that they will not have to wait for our next wakeup.
6293 static void raid5d(struct md_thread *thread)
6295 struct mddev *mddev = thread->mddev;
6296 struct r5conf *conf = mddev->private;
6298 struct blk_plug plug;
6300 pr_debug("+++ raid5d active\n");
6302 md_check_recovery(mddev);
6304 blk_start_plug(&plug);
6306 spin_lock_irq(&conf->device_lock);
6309 int batch_size, released;
6310 unsigned int offset;
6312 released = release_stripe_list(conf, conf->temp_inactive_list);
6314 clear_bit(R5_DID_ALLOC, &conf->cache_state);
6317 !list_empty(&conf->bitmap_list)) {
6318 /* Now is a good time to flush some bitmap updates */
6320 spin_unlock_irq(&conf->device_lock);
6321 md_bitmap_unplug(mddev->bitmap);
6322 spin_lock_irq(&conf->device_lock);
6323 conf->seq_write = conf->seq_flush;
6324 activate_bit_delay(conf, conf->temp_inactive_list);
6326 raid5_activate_delayed(conf);
6328 while ((bio = remove_bio_from_retry(conf, &offset))) {
6330 spin_unlock_irq(&conf->device_lock);
6331 ok = retry_aligned_read(conf, bio, offset);
6332 spin_lock_irq(&conf->device_lock);
6338 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6339 conf->temp_inactive_list);
6340 if (!batch_size && !released)
6342 handled += batch_size;
6344 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6345 spin_unlock_irq(&conf->device_lock);
6346 md_check_recovery(mddev);
6347 spin_lock_irq(&conf->device_lock);
6350 pr_debug("%d stripes handled\n", handled);
6352 spin_unlock_irq(&conf->device_lock);
6353 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6354 mutex_trylock(&conf->cache_size_mutex)) {
6355 grow_one_stripe(conf, __GFP_NOWARN);
6356 /* Set flag even if allocation failed. This helps
6357 * slow down allocation requests when mem is short
6359 set_bit(R5_DID_ALLOC, &conf->cache_state);
6360 mutex_unlock(&conf->cache_size_mutex);
6363 flush_deferred_bios(conf);
6365 r5l_flush_stripe_to_raid(conf->log);
6367 async_tx_issue_pending_all();
6368 blk_finish_plug(&plug);
6370 pr_debug("--- raid5d inactive\n");
6374 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6376 struct r5conf *conf;
6378 spin_lock(&mddev->lock);
6379 conf = mddev->private;
6381 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6382 spin_unlock(&mddev->lock);
6387 raid5_set_cache_size(struct mddev *mddev, int size)
6390 struct r5conf *conf = mddev->private;
6392 if (size <= 16 || size > 32768)
6395 conf->min_nr_stripes = size;
6396 mutex_lock(&conf->cache_size_mutex);
6397 while (size < conf->max_nr_stripes &&
6398 drop_one_stripe(conf))
6400 mutex_unlock(&conf->cache_size_mutex);
6402 md_allow_write(mddev);
6404 mutex_lock(&conf->cache_size_mutex);
6405 while (size > conf->max_nr_stripes)
6406 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6407 conf->min_nr_stripes = conf->max_nr_stripes;
6411 mutex_unlock(&conf->cache_size_mutex);
6415 EXPORT_SYMBOL(raid5_set_cache_size);
6418 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6420 struct r5conf *conf;
6424 if (len >= PAGE_SIZE)
6426 if (kstrtoul(page, 10, &new))
6428 err = mddev_lock(mddev);
6431 conf = mddev->private;
6435 err = raid5_set_cache_size(mddev, new);
6436 mddev_unlock(mddev);
6441 static struct md_sysfs_entry
6442 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6443 raid5_show_stripe_cache_size,
6444 raid5_store_stripe_cache_size);
6447 raid5_show_rmw_level(struct mddev *mddev, char *page)
6449 struct r5conf *conf = mddev->private;
6451 return sprintf(page, "%d\n", conf->rmw_level);
6457 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6459 struct r5conf *conf = mddev->private;
6465 if (len >= PAGE_SIZE)
6468 if (kstrtoul(page, 10, &new))
6471 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6474 if (new != PARITY_DISABLE_RMW &&
6475 new != PARITY_ENABLE_RMW &&
6476 new != PARITY_PREFER_RMW)
6479 conf->rmw_level = new;
6483 static struct md_sysfs_entry
6484 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6485 raid5_show_rmw_level,
6486 raid5_store_rmw_level);
6489 raid5_show_stripe_size(struct mddev *mddev, char *page)
6491 struct r5conf *conf;
6494 spin_lock(&mddev->lock);
6495 conf = mddev->private;
6497 ret = sprintf(page, "%lu\n", RAID5_STRIPE_SIZE(conf));
6498 spin_unlock(&mddev->lock);
6502 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
6504 raid5_store_stripe_size(struct mddev *mddev, const char *page, size_t len)
6506 struct r5conf *conf;
6510 if (len >= PAGE_SIZE)
6512 if (kstrtoul(page, 10, &new))
6516 * The value should not be bigger than PAGE_SIZE. It requires to
6517 * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
6520 if (new % DEFAULT_STRIPE_SIZE != 0 ||
6521 new > PAGE_SIZE || new == 0 ||
6522 new != roundup_pow_of_two(new))
6525 err = mddev_lock(mddev);
6529 conf = mddev->private;
6535 if (new == conf->stripe_size)
6538 pr_debug("md/raid: change stripe_size from %lu to %lu\n",
6539 conf->stripe_size, new);
6541 mddev_suspend(mddev);
6542 conf->stripe_size = new;
6543 conf->stripe_shift = ilog2(new) - 9;
6544 conf->stripe_sectors = new >> 9;
6545 mddev_resume(mddev);
6548 mddev_unlock(mddev);
6552 static struct md_sysfs_entry
6553 raid5_stripe_size = __ATTR(stripe_size, 0644,
6554 raid5_show_stripe_size,
6555 raid5_store_stripe_size);
6557 static struct md_sysfs_entry
6558 raid5_stripe_size = __ATTR(stripe_size, 0444,
6559 raid5_show_stripe_size,
6564 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6566 struct r5conf *conf;
6568 spin_lock(&mddev->lock);
6569 conf = mddev->private;
6571 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6572 spin_unlock(&mddev->lock);
6577 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6579 struct r5conf *conf;
6583 if (len >= PAGE_SIZE)
6585 if (kstrtoul(page, 10, &new))
6588 err = mddev_lock(mddev);
6591 conf = mddev->private;
6594 else if (new > conf->min_nr_stripes)
6597 conf->bypass_threshold = new;
6598 mddev_unlock(mddev);
6602 static struct md_sysfs_entry
6603 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6605 raid5_show_preread_threshold,
6606 raid5_store_preread_threshold);
6609 raid5_show_skip_copy(struct mddev *mddev, char *page)
6611 struct r5conf *conf;
6613 spin_lock(&mddev->lock);
6614 conf = mddev->private;
6616 ret = sprintf(page, "%d\n", conf->skip_copy);
6617 spin_unlock(&mddev->lock);
6622 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6624 struct r5conf *conf;
6628 if (len >= PAGE_SIZE)
6630 if (kstrtoul(page, 10, &new))
6634 err = mddev_lock(mddev);
6637 conf = mddev->private;
6640 else if (new != conf->skip_copy) {
6641 mddev_suspend(mddev);
6642 conf->skip_copy = new;
6644 mddev->queue->backing_dev_info->capabilities |=
6645 BDI_CAP_STABLE_WRITES;
6647 mddev->queue->backing_dev_info->capabilities &=
6648 ~BDI_CAP_STABLE_WRITES;
6649 mddev_resume(mddev);
6651 mddev_unlock(mddev);
6655 static struct md_sysfs_entry
6656 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6657 raid5_show_skip_copy,
6658 raid5_store_skip_copy);
6661 stripe_cache_active_show(struct mddev *mddev, char *page)
6663 struct r5conf *conf = mddev->private;
6665 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6670 static struct md_sysfs_entry
6671 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6674 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6676 struct r5conf *conf;
6678 spin_lock(&mddev->lock);
6679 conf = mddev->private;
6681 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6682 spin_unlock(&mddev->lock);
6686 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6688 struct r5worker_group **worker_groups);
6690 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6692 struct r5conf *conf;
6695 struct r5worker_group *new_groups, *old_groups;
6698 if (len >= PAGE_SIZE)
6700 if (kstrtouint(page, 10, &new))
6702 /* 8192 should be big enough */
6706 err = mddev_lock(mddev);
6709 conf = mddev->private;
6712 else if (new != conf->worker_cnt_per_group) {
6713 mddev_suspend(mddev);
6715 old_groups = conf->worker_groups;
6717 flush_workqueue(raid5_wq);
6719 err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
6721 spin_lock_irq(&conf->device_lock);
6722 conf->group_cnt = group_cnt;
6723 conf->worker_cnt_per_group = new;
6724 conf->worker_groups = new_groups;
6725 spin_unlock_irq(&conf->device_lock);
6728 kfree(old_groups[0].workers);
6731 mddev_resume(mddev);
6733 mddev_unlock(mddev);
6738 static struct md_sysfs_entry
6739 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6740 raid5_show_group_thread_cnt,
6741 raid5_store_group_thread_cnt);
6743 static struct attribute *raid5_attrs[] = {
6744 &raid5_stripecache_size.attr,
6745 &raid5_stripecache_active.attr,
6746 &raid5_preread_bypass_threshold.attr,
6747 &raid5_group_thread_cnt.attr,
6748 &raid5_skip_copy.attr,
6749 &raid5_rmw_level.attr,
6750 &raid5_stripe_size.attr,
6751 &r5c_journal_mode.attr,
6752 &ppl_write_hint.attr,
6755 static struct attribute_group raid5_attrs_group = {
6757 .attrs = raid5_attrs,
6760 static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
6761 struct r5worker_group **worker_groups)
6765 struct r5worker *workers;
6769 *worker_groups = NULL;
6772 *group_cnt = num_possible_nodes();
6773 size = sizeof(struct r5worker) * cnt;
6774 workers = kcalloc(size, *group_cnt, GFP_NOIO);
6775 *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6777 if (!*worker_groups || !workers) {
6779 kfree(*worker_groups);
6783 for (i = 0; i < *group_cnt; i++) {
6784 struct r5worker_group *group;
6786 group = &(*worker_groups)[i];
6787 INIT_LIST_HEAD(&group->handle_list);
6788 INIT_LIST_HEAD(&group->loprio_list);
6790 group->workers = workers + i * cnt;
6792 for (j = 0; j < cnt; j++) {
6793 struct r5worker *worker = group->workers + j;
6794 worker->group = group;
6795 INIT_WORK(&worker->work, raid5_do_work);
6797 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6798 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6805 static void free_thread_groups(struct r5conf *conf)
6807 if (conf->worker_groups)
6808 kfree(conf->worker_groups[0].workers);
6809 kfree(conf->worker_groups);
6810 conf->worker_groups = NULL;
6814 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6816 struct r5conf *conf = mddev->private;
6819 sectors = mddev->dev_sectors;
6821 /* size is defined by the smallest of previous and new size */
6822 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6824 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6825 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6826 return sectors * (raid_disks - conf->max_degraded);
6829 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6831 safe_put_page(percpu->spare_page);
6832 percpu->spare_page = NULL;
6833 kvfree(percpu->scribble);
6834 percpu->scribble = NULL;
6837 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6839 if (conf->level == 6 && !percpu->spare_page) {
6840 percpu->spare_page = alloc_page(GFP_KERNEL);
6841 if (!percpu->spare_page)
6845 if (scribble_alloc(percpu,
6846 max(conf->raid_disks,
6847 conf->previous_raid_disks),
6848 max(conf->chunk_sectors,
6849 conf->prev_chunk_sectors)
6850 / RAID5_STRIPE_SECTORS(conf))) {
6851 free_scratch_buffer(conf, percpu);
6858 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6860 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6862 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6866 static void raid5_free_percpu(struct r5conf *conf)
6871 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6872 free_percpu(conf->percpu);
6875 static void free_conf(struct r5conf *conf)
6881 unregister_shrinker(&conf->shrinker);
6882 free_thread_groups(conf);
6883 shrink_stripes(conf);
6884 raid5_free_percpu(conf);
6885 for (i = 0; i < conf->pool_size; i++)
6886 if (conf->disks[i].extra_page)
6887 put_page(conf->disks[i].extra_page);
6889 bioset_exit(&conf->bio_split);
6890 kfree(conf->stripe_hashtbl);
6891 kfree(conf->pending_data);
6895 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6897 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6898 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6900 if (alloc_scratch_buffer(conf, percpu)) {
6901 pr_warn("%s: failed memory allocation for cpu%u\n",
6908 static int raid5_alloc_percpu(struct r5conf *conf)
6912 conf->percpu = alloc_percpu(struct raid5_percpu);
6916 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6918 conf->scribble_disks = max(conf->raid_disks,
6919 conf->previous_raid_disks);
6920 conf->scribble_sectors = max(conf->chunk_sectors,
6921 conf->prev_chunk_sectors);
6926 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6927 struct shrink_control *sc)
6929 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6930 unsigned long ret = SHRINK_STOP;
6932 if (mutex_trylock(&conf->cache_size_mutex)) {
6934 while (ret < sc->nr_to_scan &&
6935 conf->max_nr_stripes > conf->min_nr_stripes) {
6936 if (drop_one_stripe(conf) == 0) {
6942 mutex_unlock(&conf->cache_size_mutex);
6947 static unsigned long raid5_cache_count(struct shrinker *shrink,
6948 struct shrink_control *sc)
6950 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6952 if (conf->max_nr_stripes < conf->min_nr_stripes)
6953 /* unlikely, but not impossible */
6955 return conf->max_nr_stripes - conf->min_nr_stripes;
6958 static struct r5conf *setup_conf(struct mddev *mddev)
6960 struct r5conf *conf;
6961 int raid_disk, memory, max_disks;
6962 struct md_rdev *rdev;
6963 struct disk_info *disk;
6967 struct r5worker_group *new_group;
6970 if (mddev->new_level != 5
6971 && mddev->new_level != 4
6972 && mddev->new_level != 6) {
6973 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6974 mdname(mddev), mddev->new_level);
6975 return ERR_PTR(-EIO);
6977 if ((mddev->new_level == 5
6978 && !algorithm_valid_raid5(mddev->new_layout)) ||
6979 (mddev->new_level == 6
6980 && !algorithm_valid_raid6(mddev->new_layout))) {
6981 pr_warn("md/raid:%s: layout %d not supported\n",
6982 mdname(mddev), mddev->new_layout);
6983 return ERR_PTR(-EIO);
6985 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6986 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6987 mdname(mddev), mddev->raid_disks);
6988 return ERR_PTR(-EINVAL);
6991 if (!mddev->new_chunk_sectors ||
6992 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6993 !is_power_of_2(mddev->new_chunk_sectors)) {
6994 pr_warn("md/raid:%s: invalid chunk size %d\n",
6995 mdname(mddev), mddev->new_chunk_sectors << 9);
6996 return ERR_PTR(-EINVAL);
6999 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
7003 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7004 conf->stripe_size = DEFAULT_STRIPE_SIZE;
7005 conf->stripe_shift = ilog2(DEFAULT_STRIPE_SIZE) - 9;
7006 conf->stripe_sectors = DEFAULT_STRIPE_SIZE >> 9;
7008 INIT_LIST_HEAD(&conf->free_list);
7009 INIT_LIST_HEAD(&conf->pending_list);
7010 conf->pending_data = kcalloc(PENDING_IO_MAX,
7011 sizeof(struct r5pending_data),
7013 if (!conf->pending_data)
7015 for (i = 0; i < PENDING_IO_MAX; i++)
7016 list_add(&conf->pending_data[i].sibling, &conf->free_list);
7017 /* Don't enable multi-threading by default*/
7018 if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
7019 conf->group_cnt = group_cnt;
7020 conf->worker_cnt_per_group = 0;
7021 conf->worker_groups = new_group;
7024 spin_lock_init(&conf->device_lock);
7025 seqcount_spinlock_init(&conf->gen_lock, &conf->device_lock);
7026 mutex_init(&conf->cache_size_mutex);
7027 init_waitqueue_head(&conf->wait_for_quiescent);
7028 init_waitqueue_head(&conf->wait_for_stripe);
7029 init_waitqueue_head(&conf->wait_for_overlap);
7030 INIT_LIST_HEAD(&conf->handle_list);
7031 INIT_LIST_HEAD(&conf->loprio_list);
7032 INIT_LIST_HEAD(&conf->hold_list);
7033 INIT_LIST_HEAD(&conf->delayed_list);
7034 INIT_LIST_HEAD(&conf->bitmap_list);
7035 init_llist_head(&conf->released_stripes);
7036 atomic_set(&conf->active_stripes, 0);
7037 atomic_set(&conf->preread_active_stripes, 0);
7038 atomic_set(&conf->active_aligned_reads, 0);
7039 spin_lock_init(&conf->pending_bios_lock);
7040 conf->batch_bio_dispatch = true;
7041 rdev_for_each(rdev, mddev) {
7042 if (test_bit(Journal, &rdev->flags))
7044 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
7045 conf->batch_bio_dispatch = false;
7050 conf->bypass_threshold = BYPASS_THRESHOLD;
7051 conf->recovery_disabled = mddev->recovery_disabled - 1;
7053 conf->raid_disks = mddev->raid_disks;
7054 if (mddev->reshape_position == MaxSector)
7055 conf->previous_raid_disks = mddev->raid_disks;
7057 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
7058 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
7060 conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
7066 for (i = 0; i < max_disks; i++) {
7067 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
7068 if (!conf->disks[i].extra_page)
7072 ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
7075 conf->mddev = mddev;
7077 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
7080 /* We init hash_locks[0] separately to that it can be used
7081 * as the reference lock in the spin_lock_nest_lock() call
7082 * in lock_all_device_hash_locks_irq in order to convince
7083 * lockdep that we know what we are doing.
7085 spin_lock_init(conf->hash_locks);
7086 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7087 spin_lock_init(conf->hash_locks + i);
7089 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7090 INIT_LIST_HEAD(conf->inactive_list + i);
7092 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7093 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7095 atomic_set(&conf->r5c_cached_full_stripes, 0);
7096 INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7097 atomic_set(&conf->r5c_cached_partial_stripes, 0);
7098 INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7099 atomic_set(&conf->r5c_flushing_full_stripes, 0);
7100 atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7102 conf->level = mddev->new_level;
7103 conf->chunk_sectors = mddev->new_chunk_sectors;
7104 if (raid5_alloc_percpu(conf) != 0)
7107 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7109 rdev_for_each(rdev, mddev) {
7110 raid_disk = rdev->raid_disk;
7111 if (raid_disk >= max_disks
7112 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7114 disk = conf->disks + raid_disk;
7116 if (test_bit(Replacement, &rdev->flags)) {
7117 if (disk->replacement)
7119 disk->replacement = rdev;
7126 if (test_bit(In_sync, &rdev->flags)) {
7127 char b[BDEVNAME_SIZE];
7128 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7129 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7130 } else if (rdev->saved_raid_disk != raid_disk)
7131 /* Cannot rely on bitmap to complete recovery */
7135 conf->level = mddev->new_level;
7136 if (conf->level == 6) {
7137 conf->max_degraded = 2;
7138 if (raid6_call.xor_syndrome)
7139 conf->rmw_level = PARITY_ENABLE_RMW;
7141 conf->rmw_level = PARITY_DISABLE_RMW;
7143 conf->max_degraded = 1;
7144 conf->rmw_level = PARITY_ENABLE_RMW;
7146 conf->algorithm = mddev->new_layout;
7147 conf->reshape_progress = mddev->reshape_position;
7148 if (conf->reshape_progress != MaxSector) {
7149 conf->prev_chunk_sectors = mddev->chunk_sectors;
7150 conf->prev_algo = mddev->layout;
7152 conf->prev_chunk_sectors = conf->chunk_sectors;
7153 conf->prev_algo = conf->algorithm;
7156 conf->min_nr_stripes = NR_STRIPES;
7157 if (mddev->reshape_position != MaxSector) {
7158 int stripes = max_t(int,
7159 ((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4,
7160 ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4);
7161 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7162 if (conf->min_nr_stripes != NR_STRIPES)
7163 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7164 mdname(mddev), conf->min_nr_stripes);
7166 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7167 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7168 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7169 if (grow_stripes(conf, conf->min_nr_stripes)) {
7170 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7171 mdname(mddev), memory);
7174 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7176 * Losing a stripe head costs more than the time to refill it,
7177 * it reduces the queue depth and so can hurt throughput.
7178 * So set it rather large, scaled by number of devices.
7180 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7181 conf->shrinker.scan_objects = raid5_cache_scan;
7182 conf->shrinker.count_objects = raid5_cache_count;
7183 conf->shrinker.batch = 128;
7184 conf->shrinker.flags = 0;
7185 if (register_shrinker(&conf->shrinker)) {
7186 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7191 sprintf(pers_name, "raid%d", mddev->new_level);
7192 conf->thread = md_register_thread(raid5d, mddev, pers_name);
7193 if (!conf->thread) {
7194 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7204 return ERR_PTR(-EIO);
7206 return ERR_PTR(-ENOMEM);
7209 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7212 case ALGORITHM_PARITY_0:
7213 if (raid_disk < max_degraded)
7216 case ALGORITHM_PARITY_N:
7217 if (raid_disk >= raid_disks - max_degraded)
7220 case ALGORITHM_PARITY_0_6:
7221 if (raid_disk == 0 ||
7222 raid_disk == raid_disks - 1)
7225 case ALGORITHM_LEFT_ASYMMETRIC_6:
7226 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7227 case ALGORITHM_LEFT_SYMMETRIC_6:
7228 case ALGORITHM_RIGHT_SYMMETRIC_6:
7229 if (raid_disk == raid_disks - 1)
7235 static int raid5_run(struct mddev *mddev)
7237 struct r5conf *conf;
7238 int working_disks = 0;
7239 int dirty_parity_disks = 0;
7240 struct md_rdev *rdev;
7241 struct md_rdev *journal_dev = NULL;
7242 sector_t reshape_offset = 0;
7244 long long min_offset_diff = 0;
7247 if (mddev_init_writes_pending(mddev) < 0)
7250 if (mddev->recovery_cp != MaxSector)
7251 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7254 rdev_for_each(rdev, mddev) {
7257 if (test_bit(Journal, &rdev->flags)) {
7261 if (rdev->raid_disk < 0)
7263 diff = (rdev->new_data_offset - rdev->data_offset);
7265 min_offset_diff = diff;
7267 } else if (mddev->reshape_backwards &&
7268 diff < min_offset_diff)
7269 min_offset_diff = diff;
7270 else if (!mddev->reshape_backwards &&
7271 diff > min_offset_diff)
7272 min_offset_diff = diff;
7275 if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7276 (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7277 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7282 if (mddev->reshape_position != MaxSector) {
7283 /* Check that we can continue the reshape.
7284 * Difficulties arise if the stripe we would write to
7285 * next is at or after the stripe we would read from next.
7286 * For a reshape that changes the number of devices, this
7287 * is only possible for a very short time, and mdadm makes
7288 * sure that time appears to have past before assembling
7289 * the array. So we fail if that time hasn't passed.
7290 * For a reshape that keeps the number of devices the same
7291 * mdadm must be monitoring the reshape can keeping the
7292 * critical areas read-only and backed up. It will start
7293 * the array in read-only mode, so we check for that.
7295 sector_t here_new, here_old;
7297 int max_degraded = (mddev->level == 6 ? 2 : 1);
7302 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7307 if (mddev->new_level != mddev->level) {
7308 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7312 old_disks = mddev->raid_disks - mddev->delta_disks;
7313 /* reshape_position must be on a new-stripe boundary, and one
7314 * further up in new geometry must map after here in old
7316 * If the chunk sizes are different, then as we perform reshape
7317 * in units of the largest of the two, reshape_position needs
7318 * be a multiple of the largest chunk size times new data disks.
7320 here_new = mddev->reshape_position;
7321 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7322 new_data_disks = mddev->raid_disks - max_degraded;
7323 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7324 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7328 reshape_offset = here_new * chunk_sectors;
7329 /* here_new is the stripe we will write to */
7330 here_old = mddev->reshape_position;
7331 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7332 /* here_old is the first stripe that we might need to read
7334 if (mddev->delta_disks == 0) {
7335 /* We cannot be sure it is safe to start an in-place
7336 * reshape. It is only safe if user-space is monitoring
7337 * and taking constant backups.
7338 * mdadm always starts a situation like this in
7339 * readonly mode so it can take control before
7340 * allowing any writes. So just check for that.
7342 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7343 abs(min_offset_diff) >= mddev->new_chunk_sectors)
7344 /* not really in-place - so OK */;
7345 else if (mddev->ro == 0) {
7346 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7350 } else if (mddev->reshape_backwards
7351 ? (here_new * chunk_sectors + min_offset_diff <=
7352 here_old * chunk_sectors)
7353 : (here_new * chunk_sectors >=
7354 here_old * chunk_sectors + (-min_offset_diff))) {
7355 /* Reading from the same stripe as writing to - bad */
7356 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7360 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7361 /* OK, we should be able to continue; */
7363 BUG_ON(mddev->level != mddev->new_level);
7364 BUG_ON(mddev->layout != mddev->new_layout);
7365 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7366 BUG_ON(mddev->delta_disks != 0);
7369 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7370 test_bit(MD_HAS_PPL, &mddev->flags)) {
7371 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7373 clear_bit(MD_HAS_PPL, &mddev->flags);
7374 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7377 if (mddev->private == NULL)
7378 conf = setup_conf(mddev);
7380 conf = mddev->private;
7383 return PTR_ERR(conf);
7385 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7387 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7390 set_disk_ro(mddev->gendisk, 1);
7391 } else if (mddev->recovery_cp == MaxSector)
7392 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7395 conf->min_offset_diff = min_offset_diff;
7396 mddev->thread = conf->thread;
7397 conf->thread = NULL;
7398 mddev->private = conf;
7400 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7402 rdev = conf->disks[i].rdev;
7403 if (!rdev && conf->disks[i].replacement) {
7404 /* The replacement is all we have yet */
7405 rdev = conf->disks[i].replacement;
7406 conf->disks[i].replacement = NULL;
7407 clear_bit(Replacement, &rdev->flags);
7408 conf->disks[i].rdev = rdev;
7412 if (conf->disks[i].replacement &&
7413 conf->reshape_progress != MaxSector) {
7414 /* replacements and reshape simply do not mix. */
7415 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7418 if (test_bit(In_sync, &rdev->flags)) {
7422 /* This disc is not fully in-sync. However if it
7423 * just stored parity (beyond the recovery_offset),
7424 * when we don't need to be concerned about the
7425 * array being dirty.
7426 * When reshape goes 'backwards', we never have
7427 * partially completed devices, so we only need
7428 * to worry about reshape going forwards.
7430 /* Hack because v0.91 doesn't store recovery_offset properly. */
7431 if (mddev->major_version == 0 &&
7432 mddev->minor_version > 90)
7433 rdev->recovery_offset = reshape_offset;
7435 if (rdev->recovery_offset < reshape_offset) {
7436 /* We need to check old and new layout */
7437 if (!only_parity(rdev->raid_disk,
7440 conf->max_degraded))
7443 if (!only_parity(rdev->raid_disk,
7445 conf->previous_raid_disks,
7446 conf->max_degraded))
7448 dirty_parity_disks++;
7452 * 0 for a fully functional array, 1 or 2 for a degraded array.
7454 mddev->degraded = raid5_calc_degraded(conf);
7456 if (has_failed(conf)) {
7457 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7458 mdname(mddev), mddev->degraded, conf->raid_disks);
7462 /* device size must be a multiple of chunk size */
7463 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7464 mddev->resync_max_sectors = mddev->dev_sectors;
7466 if (mddev->degraded > dirty_parity_disks &&
7467 mddev->recovery_cp != MaxSector) {
7468 if (test_bit(MD_HAS_PPL, &mddev->flags))
7469 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7471 else if (mddev->ok_start_degraded)
7472 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7475 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7481 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7482 mdname(mddev), conf->level,
7483 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7486 print_raid5_conf(conf);
7488 if (conf->reshape_progress != MaxSector) {
7489 conf->reshape_safe = conf->reshape_progress;
7490 atomic_set(&conf->reshape_stripes, 0);
7491 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7492 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7493 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7494 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7495 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7497 if (!mddev->sync_thread)
7501 /* Ok, everything is just fine now */
7502 if (mddev->to_remove == &raid5_attrs_group)
7503 mddev->to_remove = NULL;
7504 else if (mddev->kobj.sd &&
7505 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7506 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7508 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7512 /* read-ahead size must cover two whole stripes, which
7513 * is 2 * (datadisks) * chunksize where 'n' is the
7514 * number of raid devices
7516 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7517 int stripe = data_disks *
7518 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7519 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7520 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7522 chunk_size = mddev->chunk_sectors << 9;
7523 blk_queue_io_min(mddev->queue, chunk_size);
7524 blk_queue_io_opt(mddev->queue, chunk_size *
7525 (conf->raid_disks - conf->max_degraded));
7526 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7528 * We can only discard a whole stripe. It doesn't make sense to
7529 * discard data disk but write parity disk
7531 stripe = stripe * PAGE_SIZE;
7532 /* Round up to power of 2, as discard handling
7533 * currently assumes that */
7534 while ((stripe-1) & stripe)
7535 stripe = (stripe | (stripe-1)) + 1;
7536 mddev->queue->limits.discard_alignment = stripe;
7537 mddev->queue->limits.discard_granularity = stripe;
7539 blk_queue_max_write_same_sectors(mddev->queue, 0);
7540 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7542 rdev_for_each(rdev, mddev) {
7543 disk_stack_limits(mddev->gendisk, rdev->bdev,
7544 rdev->data_offset << 9);
7545 disk_stack_limits(mddev->gendisk, rdev->bdev,
7546 rdev->new_data_offset << 9);
7550 * zeroing is required, otherwise data
7551 * could be lost. Consider a scenario: discard a stripe
7552 * (the stripe could be inconsistent if
7553 * discard_zeroes_data is 0); write one disk of the
7554 * stripe (the stripe could be inconsistent again
7555 * depending on which disks are used to calculate
7556 * parity); the disk is broken; The stripe data of this
7559 * We only allow DISCARD if the sysadmin has confirmed that
7560 * only safe devices are in use by setting a module parameter.
7561 * A better idea might be to turn DISCARD into WRITE_ZEROES
7562 * requests, as that is required to be safe.
7564 if (devices_handle_discard_safely &&
7565 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7566 mddev->queue->limits.discard_granularity >= stripe)
7567 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7570 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7573 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7576 if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7581 md_unregister_thread(&mddev->thread);
7582 print_raid5_conf(conf);
7584 mddev->private = NULL;
7585 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7589 static void raid5_free(struct mddev *mddev, void *priv)
7591 struct r5conf *conf = priv;
7594 mddev->to_remove = &raid5_attrs_group;
7597 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7599 struct r5conf *conf = mddev->private;
7602 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7603 conf->chunk_sectors / 2, mddev->layout);
7604 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7606 for (i = 0; i < conf->raid_disks; i++) {
7607 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7608 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7611 seq_printf (seq, "]");
7614 static void print_raid5_conf (struct r5conf *conf)
7617 struct disk_info *tmp;
7619 pr_debug("RAID conf printout:\n");
7621 pr_debug("(conf==NULL)\n");
7624 pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7626 conf->raid_disks - conf->mddev->degraded);
7628 for (i = 0; i < conf->raid_disks; i++) {
7629 char b[BDEVNAME_SIZE];
7630 tmp = conf->disks + i;
7632 pr_debug(" disk %d, o:%d, dev:%s\n",
7633 i, !test_bit(Faulty, &tmp->rdev->flags),
7634 bdevname(tmp->rdev->bdev, b));
7638 static int raid5_spare_active(struct mddev *mddev)
7641 struct r5conf *conf = mddev->private;
7642 struct disk_info *tmp;
7644 unsigned long flags;
7646 for (i = 0; i < conf->raid_disks; i++) {
7647 tmp = conf->disks + i;
7648 if (tmp->replacement
7649 && tmp->replacement->recovery_offset == MaxSector
7650 && !test_bit(Faulty, &tmp->replacement->flags)
7651 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7652 /* Replacement has just become active. */
7654 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7657 /* Replaced device not technically faulty,
7658 * but we need to be sure it gets removed
7659 * and never re-added.
7661 set_bit(Faulty, &tmp->rdev->flags);
7662 sysfs_notify_dirent_safe(
7663 tmp->rdev->sysfs_state);
7665 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7666 } else if (tmp->rdev
7667 && tmp->rdev->recovery_offset == MaxSector
7668 && !test_bit(Faulty, &tmp->rdev->flags)
7669 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7671 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7674 spin_lock_irqsave(&conf->device_lock, flags);
7675 mddev->degraded = raid5_calc_degraded(conf);
7676 spin_unlock_irqrestore(&conf->device_lock, flags);
7677 print_raid5_conf(conf);
7681 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7683 struct r5conf *conf = mddev->private;
7685 int number = rdev->raid_disk;
7686 struct md_rdev **rdevp;
7687 struct disk_info *p = conf->disks + number;
7689 print_raid5_conf(conf);
7690 if (test_bit(Journal, &rdev->flags) && conf->log) {
7692 * we can't wait pending write here, as this is called in
7693 * raid5d, wait will deadlock.
7694 * neilb: there is no locking about new writes here,
7695 * so this cannot be safe.
7697 if (atomic_read(&conf->active_stripes) ||
7698 atomic_read(&conf->r5c_cached_full_stripes) ||
7699 atomic_read(&conf->r5c_cached_partial_stripes)) {
7705 if (rdev == p->rdev)
7707 else if (rdev == p->replacement)
7708 rdevp = &p->replacement;
7712 if (number >= conf->raid_disks &&
7713 conf->reshape_progress == MaxSector)
7714 clear_bit(In_sync, &rdev->flags);
7716 if (test_bit(In_sync, &rdev->flags) ||
7717 atomic_read(&rdev->nr_pending)) {
7721 /* Only remove non-faulty devices if recovery
7724 if (!test_bit(Faulty, &rdev->flags) &&
7725 mddev->recovery_disabled != conf->recovery_disabled &&
7726 !has_failed(conf) &&
7727 (!p->replacement || p->replacement == rdev) &&
7728 number < conf->raid_disks) {
7733 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7735 if (atomic_read(&rdev->nr_pending)) {
7736 /* lost the race, try later */
7742 err = log_modify(conf, rdev, false);
7746 if (p->replacement) {
7747 /* We must have just cleared 'rdev' */
7748 p->rdev = p->replacement;
7749 clear_bit(Replacement, &p->replacement->flags);
7750 smp_mb(); /* Make sure other CPUs may see both as identical
7751 * but will never see neither - if they are careful
7753 p->replacement = NULL;
7756 err = log_modify(conf, p->rdev, true);
7759 clear_bit(WantReplacement, &rdev->flags);
7762 print_raid5_conf(conf);
7766 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7768 struct r5conf *conf = mddev->private;
7769 int ret, err = -EEXIST;
7771 struct disk_info *p;
7773 int last = conf->raid_disks - 1;
7775 if (test_bit(Journal, &rdev->flags)) {
7779 rdev->raid_disk = 0;
7781 * The array is in readonly mode if journal is missing, so no
7782 * write requests running. We should be safe
7784 ret = log_init(conf, rdev, false);
7788 ret = r5l_start(conf->log);
7794 if (mddev->recovery_disabled == conf->recovery_disabled)
7797 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7798 /* no point adding a device */
7801 if (rdev->raid_disk >= 0)
7802 first = last = rdev->raid_disk;
7805 * find the disk ... but prefer rdev->saved_raid_disk
7808 if (rdev->saved_raid_disk >= 0 &&
7809 rdev->saved_raid_disk >= first &&
7810 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7811 first = rdev->saved_raid_disk;
7813 for (disk = first; disk <= last; disk++) {
7814 p = conf->disks + disk;
7815 if (p->rdev == NULL) {
7816 clear_bit(In_sync, &rdev->flags);
7817 rdev->raid_disk = disk;
7818 if (rdev->saved_raid_disk != disk)
7820 rcu_assign_pointer(p->rdev, rdev);
7822 err = log_modify(conf, rdev, true);
7827 for (disk = first; disk <= last; disk++) {
7828 p = conf->disks + disk;
7829 if (test_bit(WantReplacement, &p->rdev->flags) &&
7830 p->replacement == NULL) {
7831 clear_bit(In_sync, &rdev->flags);
7832 set_bit(Replacement, &rdev->flags);
7833 rdev->raid_disk = disk;
7836 rcu_assign_pointer(p->replacement, rdev);
7841 print_raid5_conf(conf);
7845 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7847 /* no resync is happening, and there is enough space
7848 * on all devices, so we can resize.
7849 * We need to make sure resync covers any new space.
7850 * If the array is shrinking we should possibly wait until
7851 * any io in the removed space completes, but it hardly seems
7855 struct r5conf *conf = mddev->private;
7857 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7859 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7860 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7861 if (mddev->external_size &&
7862 mddev->array_sectors > newsize)
7864 if (mddev->bitmap) {
7865 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
7869 md_set_array_sectors(mddev, newsize);
7870 if (sectors > mddev->dev_sectors &&
7871 mddev->recovery_cp > mddev->dev_sectors) {
7872 mddev->recovery_cp = mddev->dev_sectors;
7873 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7875 mddev->dev_sectors = sectors;
7876 mddev->resync_max_sectors = sectors;
7880 static int check_stripe_cache(struct mddev *mddev)
7882 /* Can only proceed if there are plenty of stripe_heads.
7883 * We need a minimum of one full stripe,, and for sensible progress
7884 * it is best to have about 4 times that.
7885 * If we require 4 times, then the default 256 4K stripe_heads will
7886 * allow for chunk sizes up to 256K, which is probably OK.
7887 * If the chunk size is greater, user-space should request more
7888 * stripe_heads first.
7890 struct r5conf *conf = mddev->private;
7891 if (((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
7892 > conf->min_nr_stripes ||
7893 ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
7894 > conf->min_nr_stripes) {
7895 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7897 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7898 / RAID5_STRIPE_SIZE(conf))*4);
7904 static int check_reshape(struct mddev *mddev)
7906 struct r5conf *conf = mddev->private;
7908 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7910 if (mddev->delta_disks == 0 &&
7911 mddev->new_layout == mddev->layout &&
7912 mddev->new_chunk_sectors == mddev->chunk_sectors)
7913 return 0; /* nothing to do */
7914 if (has_failed(conf))
7916 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7917 /* We might be able to shrink, but the devices must
7918 * be made bigger first.
7919 * For raid6, 4 is the minimum size.
7920 * Otherwise 2 is the minimum
7923 if (mddev->level == 6)
7925 if (mddev->raid_disks + mddev->delta_disks < min)
7929 if (!check_stripe_cache(mddev))
7932 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7933 mddev->delta_disks > 0)
7934 if (resize_chunks(conf,
7935 conf->previous_raid_disks
7936 + max(0, mddev->delta_disks),
7937 max(mddev->new_chunk_sectors,
7938 mddev->chunk_sectors)
7942 if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7943 return 0; /* never bother to shrink */
7944 return resize_stripes(conf, (conf->previous_raid_disks
7945 + mddev->delta_disks));
7948 static int raid5_start_reshape(struct mddev *mddev)
7950 struct r5conf *conf = mddev->private;
7951 struct md_rdev *rdev;
7953 unsigned long flags;
7955 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7958 if (!check_stripe_cache(mddev))
7961 if (has_failed(conf))
7964 rdev_for_each(rdev, mddev) {
7965 if (!test_bit(In_sync, &rdev->flags)
7966 && !test_bit(Faulty, &rdev->flags))
7970 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7971 /* Not enough devices even to make a degraded array
7976 /* Refuse to reduce size of the array. Any reductions in
7977 * array size must be through explicit setting of array_size
7980 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7981 < mddev->array_sectors) {
7982 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7987 atomic_set(&conf->reshape_stripes, 0);
7988 spin_lock_irq(&conf->device_lock);
7989 write_seqcount_begin(&conf->gen_lock);
7990 conf->previous_raid_disks = conf->raid_disks;
7991 conf->raid_disks += mddev->delta_disks;
7992 conf->prev_chunk_sectors = conf->chunk_sectors;
7993 conf->chunk_sectors = mddev->new_chunk_sectors;
7994 conf->prev_algo = conf->algorithm;
7995 conf->algorithm = mddev->new_layout;
7997 /* Code that selects data_offset needs to see the generation update
7998 * if reshape_progress has been set - so a memory barrier needed.
8001 if (mddev->reshape_backwards)
8002 conf->reshape_progress = raid5_size(mddev, 0, 0);
8004 conf->reshape_progress = 0;
8005 conf->reshape_safe = conf->reshape_progress;
8006 write_seqcount_end(&conf->gen_lock);
8007 spin_unlock_irq(&conf->device_lock);
8009 /* Now make sure any requests that proceeded on the assumption
8010 * the reshape wasn't running - like Discard or Read - have
8013 mddev_suspend(mddev);
8014 mddev_resume(mddev);
8016 /* Add some new drives, as many as will fit.
8017 * We know there are enough to make the newly sized array work.
8018 * Don't add devices if we are reducing the number of
8019 * devices in the array. This is because it is not possible
8020 * to correctly record the "partially reconstructed" state of
8021 * such devices during the reshape and confusion could result.
8023 if (mddev->delta_disks >= 0) {
8024 rdev_for_each(rdev, mddev)
8025 if (rdev->raid_disk < 0 &&
8026 !test_bit(Faulty, &rdev->flags)) {
8027 if (raid5_add_disk(mddev, rdev) == 0) {
8029 >= conf->previous_raid_disks)
8030 set_bit(In_sync, &rdev->flags);
8032 rdev->recovery_offset = 0;
8034 /* Failure here is OK */
8035 sysfs_link_rdev(mddev, rdev);
8037 } else if (rdev->raid_disk >= conf->previous_raid_disks
8038 && !test_bit(Faulty, &rdev->flags)) {
8039 /* This is a spare that was manually added */
8040 set_bit(In_sync, &rdev->flags);
8043 /* When a reshape changes the number of devices,
8044 * ->degraded is measured against the larger of the
8045 * pre and post number of devices.
8047 spin_lock_irqsave(&conf->device_lock, flags);
8048 mddev->degraded = raid5_calc_degraded(conf);
8049 spin_unlock_irqrestore(&conf->device_lock, flags);
8051 mddev->raid_disks = conf->raid_disks;
8052 mddev->reshape_position = conf->reshape_progress;
8053 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8055 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8056 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8057 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
8058 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8059 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
8060 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
8062 if (!mddev->sync_thread) {
8063 mddev->recovery = 0;
8064 spin_lock_irq(&conf->device_lock);
8065 write_seqcount_begin(&conf->gen_lock);
8066 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
8067 mddev->new_chunk_sectors =
8068 conf->chunk_sectors = conf->prev_chunk_sectors;
8069 mddev->new_layout = conf->algorithm = conf->prev_algo;
8070 rdev_for_each(rdev, mddev)
8071 rdev->new_data_offset = rdev->data_offset;
8073 conf->generation --;
8074 conf->reshape_progress = MaxSector;
8075 mddev->reshape_position = MaxSector;
8076 write_seqcount_end(&conf->gen_lock);
8077 spin_unlock_irq(&conf->device_lock);
8080 conf->reshape_checkpoint = jiffies;
8081 md_wakeup_thread(mddev->sync_thread);
8082 md_new_event(mddev);
8086 /* This is called from the reshape thread and should make any
8087 * changes needed in 'conf'
8089 static void end_reshape(struct r5conf *conf)
8092 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8093 struct md_rdev *rdev;
8095 spin_lock_irq(&conf->device_lock);
8096 conf->previous_raid_disks = conf->raid_disks;
8097 md_finish_reshape(conf->mddev);
8099 conf->reshape_progress = MaxSector;
8100 conf->mddev->reshape_position = MaxSector;
8101 rdev_for_each(rdev, conf->mddev)
8102 if (rdev->raid_disk >= 0 &&
8103 !test_bit(Journal, &rdev->flags) &&
8104 !test_bit(In_sync, &rdev->flags))
8105 rdev->recovery_offset = MaxSector;
8106 spin_unlock_irq(&conf->device_lock);
8107 wake_up(&conf->wait_for_overlap);
8109 /* read-ahead size must cover two whole stripes, which is
8110 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
8112 if (conf->mddev->queue) {
8113 int data_disks = conf->raid_disks - conf->max_degraded;
8114 int stripe = data_disks * ((conf->chunk_sectors << 9)
8116 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
8117 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
8122 /* This is called from the raid5d thread with mddev_lock held.
8123 * It makes config changes to the device.
8125 static void raid5_finish_reshape(struct mddev *mddev)
8127 struct r5conf *conf = mddev->private;
8129 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8131 if (mddev->delta_disks <= 0) {
8133 spin_lock_irq(&conf->device_lock);
8134 mddev->degraded = raid5_calc_degraded(conf);
8135 spin_unlock_irq(&conf->device_lock);
8136 for (d = conf->raid_disks ;
8137 d < conf->raid_disks - mddev->delta_disks;
8139 struct md_rdev *rdev = conf->disks[d].rdev;
8141 clear_bit(In_sync, &rdev->flags);
8142 rdev = conf->disks[d].replacement;
8144 clear_bit(In_sync, &rdev->flags);
8147 mddev->layout = conf->algorithm;
8148 mddev->chunk_sectors = conf->chunk_sectors;
8149 mddev->reshape_position = MaxSector;
8150 mddev->delta_disks = 0;
8151 mddev->reshape_backwards = 0;
8155 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8157 struct r5conf *conf = mddev->private;
8160 /* stop all writes */
8161 lock_all_device_hash_locks_irq(conf);
8162 /* '2' tells resync/reshape to pause so that all
8163 * active stripes can drain
8165 r5c_flush_cache(conf, INT_MAX);
8167 wait_event_cmd(conf->wait_for_quiescent,
8168 atomic_read(&conf->active_stripes) == 0 &&
8169 atomic_read(&conf->active_aligned_reads) == 0,
8170 unlock_all_device_hash_locks_irq(conf),
8171 lock_all_device_hash_locks_irq(conf));
8173 unlock_all_device_hash_locks_irq(conf);
8174 /* allow reshape to continue */
8175 wake_up(&conf->wait_for_overlap);
8177 /* re-enable writes */
8178 lock_all_device_hash_locks_irq(conf);
8180 wake_up(&conf->wait_for_quiescent);
8181 wake_up(&conf->wait_for_overlap);
8182 unlock_all_device_hash_locks_irq(conf);
8184 log_quiesce(conf, quiesce);
8187 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8189 struct r0conf *raid0_conf = mddev->private;
8192 /* for raid0 takeover only one zone is supported */
8193 if (raid0_conf->nr_strip_zones > 1) {
8194 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8196 return ERR_PTR(-EINVAL);
8199 sectors = raid0_conf->strip_zone[0].zone_end;
8200 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8201 mddev->dev_sectors = sectors;
8202 mddev->new_level = level;
8203 mddev->new_layout = ALGORITHM_PARITY_N;
8204 mddev->new_chunk_sectors = mddev->chunk_sectors;
8205 mddev->raid_disks += 1;
8206 mddev->delta_disks = 1;
8207 /* make sure it will be not marked as dirty */
8208 mddev->recovery_cp = MaxSector;
8210 return setup_conf(mddev);
8213 static void *raid5_takeover_raid1(struct mddev *mddev)
8218 if (mddev->raid_disks != 2 ||
8219 mddev->degraded > 1)
8220 return ERR_PTR(-EINVAL);
8222 /* Should check if there are write-behind devices? */
8224 chunksect = 64*2; /* 64K by default */
8226 /* The array must be an exact multiple of chunksize */
8227 while (chunksect && (mddev->array_sectors & (chunksect-1)))
8230 if ((chunksect<<9) < RAID5_STRIPE_SIZE((struct r5conf *)mddev->private))
8231 /* array size does not allow a suitable chunk size */
8232 return ERR_PTR(-EINVAL);
8234 mddev->new_level = 5;
8235 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8236 mddev->new_chunk_sectors = chunksect;
8238 ret = setup_conf(mddev);
8240 mddev_clear_unsupported_flags(mddev,
8241 UNSUPPORTED_MDDEV_FLAGS);
8245 static void *raid5_takeover_raid6(struct mddev *mddev)
8249 switch (mddev->layout) {
8250 case ALGORITHM_LEFT_ASYMMETRIC_6:
8251 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8253 case ALGORITHM_RIGHT_ASYMMETRIC_6:
8254 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8256 case ALGORITHM_LEFT_SYMMETRIC_6:
8257 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8259 case ALGORITHM_RIGHT_SYMMETRIC_6:
8260 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8262 case ALGORITHM_PARITY_0_6:
8263 new_layout = ALGORITHM_PARITY_0;
8265 case ALGORITHM_PARITY_N:
8266 new_layout = ALGORITHM_PARITY_N;
8269 return ERR_PTR(-EINVAL);
8271 mddev->new_level = 5;
8272 mddev->new_layout = new_layout;
8273 mddev->delta_disks = -1;
8274 mddev->raid_disks -= 1;
8275 return setup_conf(mddev);
8278 static int raid5_check_reshape(struct mddev *mddev)
8280 /* For a 2-drive array, the layout and chunk size can be changed
8281 * immediately as not restriping is needed.
8282 * For larger arrays we record the new value - after validation
8283 * to be used by a reshape pass.
8285 struct r5conf *conf = mddev->private;
8286 int new_chunk = mddev->new_chunk_sectors;
8288 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8290 if (new_chunk > 0) {
8291 if (!is_power_of_2(new_chunk))
8293 if (new_chunk < (PAGE_SIZE>>9))
8295 if (mddev->array_sectors & (new_chunk-1))
8296 /* not factor of array size */
8300 /* They look valid */
8302 if (mddev->raid_disks == 2) {
8303 /* can make the change immediately */
8304 if (mddev->new_layout >= 0) {
8305 conf->algorithm = mddev->new_layout;
8306 mddev->layout = mddev->new_layout;
8308 if (new_chunk > 0) {
8309 conf->chunk_sectors = new_chunk ;
8310 mddev->chunk_sectors = new_chunk;
8312 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8313 md_wakeup_thread(mddev->thread);
8315 return check_reshape(mddev);
8318 static int raid6_check_reshape(struct mddev *mddev)
8320 int new_chunk = mddev->new_chunk_sectors;
8322 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8324 if (new_chunk > 0) {
8325 if (!is_power_of_2(new_chunk))
8327 if (new_chunk < (PAGE_SIZE >> 9))
8329 if (mddev->array_sectors & (new_chunk-1))
8330 /* not factor of array size */
8334 /* They look valid */
8335 return check_reshape(mddev);
8338 static void *raid5_takeover(struct mddev *mddev)
8340 /* raid5 can take over:
8341 * raid0 - if there is only one strip zone - make it a raid4 layout
8342 * raid1 - if there are two drives. We need to know the chunk size
8343 * raid4 - trivial - just use a raid4 layout.
8344 * raid6 - Providing it is a *_6 layout
8346 if (mddev->level == 0)
8347 return raid45_takeover_raid0(mddev, 5);
8348 if (mddev->level == 1)
8349 return raid5_takeover_raid1(mddev);
8350 if (mddev->level == 4) {
8351 mddev->new_layout = ALGORITHM_PARITY_N;
8352 mddev->new_level = 5;
8353 return setup_conf(mddev);
8355 if (mddev->level == 6)
8356 return raid5_takeover_raid6(mddev);
8358 return ERR_PTR(-EINVAL);
8361 static void *raid4_takeover(struct mddev *mddev)
8363 /* raid4 can take over:
8364 * raid0 - if there is only one strip zone
8365 * raid5 - if layout is right
8367 if (mddev->level == 0)
8368 return raid45_takeover_raid0(mddev, 4);
8369 if (mddev->level == 5 &&
8370 mddev->layout == ALGORITHM_PARITY_N) {
8371 mddev->new_layout = 0;
8372 mddev->new_level = 4;
8373 return setup_conf(mddev);
8375 return ERR_PTR(-EINVAL);
8378 static struct md_personality raid5_personality;
8380 static void *raid6_takeover(struct mddev *mddev)
8382 /* Currently can only take over a raid5. We map the
8383 * personality to an equivalent raid6 personality
8384 * with the Q block at the end.
8388 if (mddev->pers != &raid5_personality)
8389 return ERR_PTR(-EINVAL);
8390 if (mddev->degraded > 1)
8391 return ERR_PTR(-EINVAL);
8392 if (mddev->raid_disks > 253)
8393 return ERR_PTR(-EINVAL);
8394 if (mddev->raid_disks < 3)
8395 return ERR_PTR(-EINVAL);
8397 switch (mddev->layout) {
8398 case ALGORITHM_LEFT_ASYMMETRIC:
8399 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8401 case ALGORITHM_RIGHT_ASYMMETRIC:
8402 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8404 case ALGORITHM_LEFT_SYMMETRIC:
8405 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8407 case ALGORITHM_RIGHT_SYMMETRIC:
8408 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8410 case ALGORITHM_PARITY_0:
8411 new_layout = ALGORITHM_PARITY_0_6;
8413 case ALGORITHM_PARITY_N:
8414 new_layout = ALGORITHM_PARITY_N;
8417 return ERR_PTR(-EINVAL);
8419 mddev->new_level = 6;
8420 mddev->new_layout = new_layout;
8421 mddev->delta_disks = 1;
8422 mddev->raid_disks += 1;
8423 return setup_conf(mddev);
8426 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8428 struct r5conf *conf;
8431 err = mddev_lock(mddev);
8434 conf = mddev->private;
8436 mddev_unlock(mddev);
8440 if (strncmp(buf, "ppl", 3) == 0) {
8441 /* ppl only works with RAID 5 */
8442 if (!raid5_has_ppl(conf) && conf->level == 5) {
8443 err = log_init(conf, NULL, true);
8445 err = resize_stripes(conf, conf->pool_size);
8451 } else if (strncmp(buf, "resync", 6) == 0) {
8452 if (raid5_has_ppl(conf)) {
8453 mddev_suspend(mddev);
8455 mddev_resume(mddev);
8456 err = resize_stripes(conf, conf->pool_size);
8457 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8458 r5l_log_disk_error(conf)) {
8459 bool journal_dev_exists = false;
8460 struct md_rdev *rdev;
8462 rdev_for_each(rdev, mddev)
8463 if (test_bit(Journal, &rdev->flags)) {
8464 journal_dev_exists = true;
8468 if (!journal_dev_exists) {
8469 mddev_suspend(mddev);
8470 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8471 mddev_resume(mddev);
8472 } else /* need remove journal device first */
8481 md_update_sb(mddev, 1);
8483 mddev_unlock(mddev);
8488 static int raid5_start(struct mddev *mddev)
8490 struct r5conf *conf = mddev->private;
8492 return r5l_start(conf->log);
8495 static struct md_personality raid6_personality =
8499 .owner = THIS_MODULE,
8500 .make_request = raid5_make_request,
8502 .start = raid5_start,
8504 .status = raid5_status,
8505 .error_handler = raid5_error,
8506 .hot_add_disk = raid5_add_disk,
8507 .hot_remove_disk= raid5_remove_disk,
8508 .spare_active = raid5_spare_active,
8509 .sync_request = raid5_sync_request,
8510 .resize = raid5_resize,
8512 .check_reshape = raid6_check_reshape,
8513 .start_reshape = raid5_start_reshape,
8514 .finish_reshape = raid5_finish_reshape,
8515 .quiesce = raid5_quiesce,
8516 .takeover = raid6_takeover,
8517 .change_consistency_policy = raid5_change_consistency_policy,
8519 static struct md_personality raid5_personality =
8523 .owner = THIS_MODULE,
8524 .make_request = raid5_make_request,
8526 .start = raid5_start,
8528 .status = raid5_status,
8529 .error_handler = raid5_error,
8530 .hot_add_disk = raid5_add_disk,
8531 .hot_remove_disk= raid5_remove_disk,
8532 .spare_active = raid5_spare_active,
8533 .sync_request = raid5_sync_request,
8534 .resize = raid5_resize,
8536 .check_reshape = raid5_check_reshape,
8537 .start_reshape = raid5_start_reshape,
8538 .finish_reshape = raid5_finish_reshape,
8539 .quiesce = raid5_quiesce,
8540 .takeover = raid5_takeover,
8541 .change_consistency_policy = raid5_change_consistency_policy,
8544 static struct md_personality raid4_personality =
8548 .owner = THIS_MODULE,
8549 .make_request = raid5_make_request,
8551 .start = raid5_start,
8553 .status = raid5_status,
8554 .error_handler = raid5_error,
8555 .hot_add_disk = raid5_add_disk,
8556 .hot_remove_disk= raid5_remove_disk,
8557 .spare_active = raid5_spare_active,
8558 .sync_request = raid5_sync_request,
8559 .resize = raid5_resize,
8561 .check_reshape = raid5_check_reshape,
8562 .start_reshape = raid5_start_reshape,
8563 .finish_reshape = raid5_finish_reshape,
8564 .quiesce = raid5_quiesce,
8565 .takeover = raid4_takeover,
8566 .change_consistency_policy = raid5_change_consistency_policy,
8569 static int __init raid5_init(void)
8573 raid5_wq = alloc_workqueue("raid5wq",
8574 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8578 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8580 raid456_cpu_up_prepare,
8583 destroy_workqueue(raid5_wq);
8586 register_md_personality(&raid6_personality);
8587 register_md_personality(&raid5_personality);
8588 register_md_personality(&raid4_personality);
8592 static void raid5_exit(void)
8594 unregister_md_personality(&raid6_personality);
8595 unregister_md_personality(&raid5_personality);
8596 unregister_md_personality(&raid4_personality);
8597 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8598 destroy_workqueue(raid5_wq);
8601 module_init(raid5_init);
8602 module_exit(raid5_exit);
8603 MODULE_LICENSE("GPL");
8604 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8605 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8606 MODULE_ALIAS("md-raid5");
8607 MODULE_ALIAS("md-raid4");
8608 MODULE_ALIAS("md-level-5");
8609 MODULE_ALIAS("md-level-4");
8610 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8611 MODULE_ALIAS("md-raid6");
8612 MODULE_ALIAS("md-level-6");
8614 /* This used to be two separate modules, they were: */
8615 MODULE_ALIAS("raid5");
8616 MODULE_ALIAS("raid6");