2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
59 #include <trace/events/block.h>
60 #include <linux/list_sort.h>
65 #include "md-bitmap.h"
66 #include "raid5-log.h"
68 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
70 #define cpu_to_group(cpu) cpu_to_node(cpu)
71 #define ANY_GROUP NUMA_NO_NODE
73 static bool devices_handle_discard_safely = false;
74 module_param(devices_handle_discard_safely, bool, 0644);
75 MODULE_PARM_DESC(devices_handle_discard_safely,
76 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
77 static struct workqueue_struct *raid5_wq;
79 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
81 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
82 return &conf->stripe_hashtbl[hash];
85 static inline int stripe_hash_locks_hash(sector_t sect)
87 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
90 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
92 spin_lock_irq(conf->hash_locks + hash);
93 spin_lock(&conf->device_lock);
96 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
98 spin_unlock(&conf->device_lock);
99 spin_unlock_irq(conf->hash_locks + hash);
102 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
105 spin_lock_irq(conf->hash_locks);
106 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
107 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
108 spin_lock(&conf->device_lock);
111 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
114 spin_unlock(&conf->device_lock);
115 for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
116 spin_unlock(conf->hash_locks + i);
117 spin_unlock_irq(conf->hash_locks);
120 /* Find first data disk in a raid6 stripe */
121 static inline int raid6_d0(struct stripe_head *sh)
124 /* ddf always start from first device */
126 /* md starts just after Q block */
127 if (sh->qd_idx == sh->disks - 1)
130 return sh->qd_idx + 1;
132 static inline int raid6_next_disk(int disk, int raid_disks)
135 return (disk < raid_disks) ? disk : 0;
138 /* When walking through the disks in a raid5, starting at raid6_d0,
139 * We need to map each disk to a 'slot', where the data disks are slot
140 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
141 * is raid_disks-1. This help does that mapping.
143 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
144 int *count, int syndrome_disks)
150 if (idx == sh->pd_idx)
151 return syndrome_disks;
152 if (idx == sh->qd_idx)
153 return syndrome_disks + 1;
159 static void print_raid5_conf (struct r5conf *conf);
161 static int stripe_operations_active(struct stripe_head *sh)
163 return sh->check_state || sh->reconstruct_state ||
164 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
165 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
168 static bool stripe_is_lowprio(struct stripe_head *sh)
170 return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
171 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
172 !test_bit(STRIPE_R5C_CACHING, &sh->state);
175 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
177 struct r5conf *conf = sh->raid_conf;
178 struct r5worker_group *group;
180 int i, cpu = sh->cpu;
182 if (!cpu_online(cpu)) {
183 cpu = cpumask_any(cpu_online_mask);
187 if (list_empty(&sh->lru)) {
188 struct r5worker_group *group;
189 group = conf->worker_groups + cpu_to_group(cpu);
190 if (stripe_is_lowprio(sh))
191 list_add_tail(&sh->lru, &group->loprio_list);
193 list_add_tail(&sh->lru, &group->handle_list);
194 group->stripes_cnt++;
198 if (conf->worker_cnt_per_group == 0) {
199 md_wakeup_thread(conf->mddev->thread);
203 group = conf->worker_groups + cpu_to_group(sh->cpu);
205 group->workers[0].working = true;
206 /* at least one worker should run to avoid race */
207 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
209 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
210 /* wakeup more workers */
211 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
212 if (group->workers[i].working == false) {
213 group->workers[i].working = true;
214 queue_work_on(sh->cpu, raid5_wq,
215 &group->workers[i].work);
221 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
222 struct list_head *temp_inactive_list)
225 int injournal = 0; /* number of date pages with R5_InJournal */
227 BUG_ON(!list_empty(&sh->lru));
228 BUG_ON(atomic_read(&conf->active_stripes)==0);
230 if (r5c_is_writeback(conf->log))
231 for (i = sh->disks; i--; )
232 if (test_bit(R5_InJournal, &sh->dev[i].flags))
235 * In the following cases, the stripe cannot be released to cached
236 * lists. Therefore, we make the stripe write out and set
238 * 1. when quiesce in r5c write back;
239 * 2. when resync is requested fot the stripe.
241 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
242 (conf->quiesce && r5c_is_writeback(conf->log) &&
243 !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
244 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
245 r5c_make_stripe_write_out(sh);
246 set_bit(STRIPE_HANDLE, &sh->state);
249 if (test_bit(STRIPE_HANDLE, &sh->state)) {
250 if (test_bit(STRIPE_DELAYED, &sh->state) &&
251 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
252 list_add_tail(&sh->lru, &conf->delayed_list);
253 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
254 sh->bm_seq - conf->seq_write > 0)
255 list_add_tail(&sh->lru, &conf->bitmap_list);
257 clear_bit(STRIPE_DELAYED, &sh->state);
258 clear_bit(STRIPE_BIT_DELAY, &sh->state);
259 if (conf->worker_cnt_per_group == 0) {
260 if (stripe_is_lowprio(sh))
261 list_add_tail(&sh->lru,
264 list_add_tail(&sh->lru,
267 raid5_wakeup_stripe_thread(sh);
271 md_wakeup_thread(conf->mddev->thread);
273 BUG_ON(stripe_operations_active(sh));
274 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
275 if (atomic_dec_return(&conf->preread_active_stripes)
277 md_wakeup_thread(conf->mddev->thread);
278 atomic_dec(&conf->active_stripes);
279 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
280 if (!r5c_is_writeback(conf->log))
281 list_add_tail(&sh->lru, temp_inactive_list);
283 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
285 list_add_tail(&sh->lru, temp_inactive_list);
286 else if (injournal == conf->raid_disks - conf->max_degraded) {
288 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
289 atomic_inc(&conf->r5c_cached_full_stripes);
290 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
291 atomic_dec(&conf->r5c_cached_partial_stripes);
292 list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
293 r5c_check_cached_full_stripe(conf);
296 * STRIPE_R5C_PARTIAL_STRIPE is set in
297 * r5c_try_caching_write(). No need to
300 list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
306 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
307 struct list_head *temp_inactive_list)
309 if (atomic_dec_and_test(&sh->count))
310 do_release_stripe(conf, sh, temp_inactive_list);
314 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
316 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
317 * given time. Adding stripes only takes device lock, while deleting stripes
318 * only takes hash lock.
320 static void release_inactive_stripe_list(struct r5conf *conf,
321 struct list_head *temp_inactive_list,
325 bool do_wakeup = false;
328 if (hash == NR_STRIPE_HASH_LOCKS) {
329 size = NR_STRIPE_HASH_LOCKS;
330 hash = NR_STRIPE_HASH_LOCKS - 1;
334 struct list_head *list = &temp_inactive_list[size - 1];
337 * We don't hold any lock here yet, raid5_get_active_stripe() might
338 * remove stripes from the list
340 if (!list_empty_careful(list)) {
341 spin_lock_irqsave(conf->hash_locks + hash, flags);
342 if (list_empty(conf->inactive_list + hash) &&
344 atomic_dec(&conf->empty_inactive_list_nr);
345 list_splice_tail_init(list, conf->inactive_list + hash);
347 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
354 wake_up(&conf->wait_for_stripe);
355 if (atomic_read(&conf->active_stripes) == 0)
356 wake_up(&conf->wait_for_quiescent);
357 if (conf->retry_read_aligned)
358 md_wakeup_thread(conf->mddev->thread);
362 /* should hold conf->device_lock already */
363 static int release_stripe_list(struct r5conf *conf,
364 struct list_head *temp_inactive_list)
366 struct stripe_head *sh, *t;
368 struct llist_node *head;
370 head = llist_del_all(&conf->released_stripes);
371 head = llist_reverse_order(head);
372 llist_for_each_entry_safe(sh, t, head, release_list) {
375 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
377 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
379 * Don't worry the bit is set here, because if the bit is set
380 * again, the count is always > 1. This is true for
381 * STRIPE_ON_UNPLUG_LIST bit too.
383 hash = sh->hash_lock_index;
384 __release_stripe(conf, sh, &temp_inactive_list[hash]);
391 void raid5_release_stripe(struct stripe_head *sh)
393 struct r5conf *conf = sh->raid_conf;
395 struct list_head list;
399 /* Avoid release_list until the last reference.
401 if (atomic_add_unless(&sh->count, -1, 1))
404 if (unlikely(!conf->mddev->thread) ||
405 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
407 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
409 md_wakeup_thread(conf->mddev->thread);
412 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
413 if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
414 INIT_LIST_HEAD(&list);
415 hash = sh->hash_lock_index;
416 do_release_stripe(conf, sh, &list);
417 spin_unlock_irqrestore(&conf->device_lock, flags);
418 release_inactive_stripe_list(conf, &list, hash);
422 static inline void remove_hash(struct stripe_head *sh)
424 pr_debug("remove_hash(), stripe %llu\n",
425 (unsigned long long)sh->sector);
427 hlist_del_init(&sh->hash);
430 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
432 struct hlist_head *hp = stripe_hash(conf, sh->sector);
434 pr_debug("insert_hash(), stripe %llu\n",
435 (unsigned long long)sh->sector);
437 hlist_add_head(&sh->hash, hp);
440 /* find an idle stripe, make sure it is unhashed, and return it. */
441 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
443 struct stripe_head *sh = NULL;
444 struct list_head *first;
446 if (list_empty(conf->inactive_list + hash))
448 first = (conf->inactive_list + hash)->next;
449 sh = list_entry(first, struct stripe_head, lru);
450 list_del_init(first);
452 atomic_inc(&conf->active_stripes);
453 BUG_ON(hash != sh->hash_lock_index);
454 if (list_empty(conf->inactive_list + hash))
455 atomic_inc(&conf->empty_inactive_list_nr);
460 static void shrink_buffers(struct stripe_head *sh)
464 int num = sh->raid_conf->pool_size;
466 for (i = 0; i < num ; i++) {
467 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
471 sh->dev[i].page = NULL;
476 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
479 int num = sh->raid_conf->pool_size;
481 for (i = 0; i < num; i++) {
484 if (!(page = alloc_page(gfp))) {
487 sh->dev[i].page = page;
488 sh->dev[i].orig_page = page;
494 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
495 struct stripe_head *sh);
497 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
499 struct r5conf *conf = sh->raid_conf;
502 BUG_ON(atomic_read(&sh->count) != 0);
503 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
504 BUG_ON(stripe_operations_active(sh));
505 BUG_ON(sh->batch_head);
507 pr_debug("init_stripe called, stripe %llu\n",
508 (unsigned long long)sector);
510 seq = read_seqcount_begin(&conf->gen_lock);
511 sh->generation = conf->generation - previous;
512 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
514 stripe_set_idx(sector, conf, previous, sh);
517 for (i = sh->disks; i--; ) {
518 struct r5dev *dev = &sh->dev[i];
520 if (dev->toread || dev->read || dev->towrite || dev->written ||
521 test_bit(R5_LOCKED, &dev->flags)) {
522 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
523 (unsigned long long)sh->sector, i, dev->toread,
524 dev->read, dev->towrite, dev->written,
525 test_bit(R5_LOCKED, &dev->flags));
529 dev->sector = raid5_compute_blocknr(sh, i, previous);
531 if (read_seqcount_retry(&conf->gen_lock, seq))
533 sh->overwrite_disks = 0;
534 insert_hash(conf, sh);
535 sh->cpu = smp_processor_id();
536 set_bit(STRIPE_BATCH_READY, &sh->state);
539 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
542 struct stripe_head *sh;
544 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
545 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
546 if (sh->sector == sector && sh->generation == generation)
548 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
553 * Need to check if array has failed when deciding whether to:
555 * - remove non-faulty devices
558 * This determination is simple when no reshape is happening.
559 * However if there is a reshape, we need to carefully check
560 * both the before and after sections.
561 * This is because some failed devices may only affect one
562 * of the two sections, and some non-in_sync devices may
563 * be insync in the section most affected by failed devices.
565 int raid5_calc_degraded(struct r5conf *conf)
567 int degraded, degraded2;
572 for (i = 0; i < conf->previous_raid_disks; i++) {
573 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
574 if (rdev && test_bit(Faulty, &rdev->flags))
575 rdev = rcu_dereference(conf->disks[i].replacement);
576 if (!rdev || test_bit(Faulty, &rdev->flags))
578 else if (test_bit(In_sync, &rdev->flags))
581 /* not in-sync or faulty.
582 * If the reshape increases the number of devices,
583 * this is being recovered by the reshape, so
584 * this 'previous' section is not in_sync.
585 * If the number of devices is being reduced however,
586 * the device can only be part of the array if
587 * we are reverting a reshape, so this section will
590 if (conf->raid_disks >= conf->previous_raid_disks)
594 if (conf->raid_disks == conf->previous_raid_disks)
598 for (i = 0; i < conf->raid_disks; i++) {
599 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
600 if (rdev && test_bit(Faulty, &rdev->flags))
601 rdev = rcu_dereference(conf->disks[i].replacement);
602 if (!rdev || test_bit(Faulty, &rdev->flags))
604 else if (test_bit(In_sync, &rdev->flags))
607 /* not in-sync or faulty.
608 * If reshape increases the number of devices, this
609 * section has already been recovered, else it
610 * almost certainly hasn't.
612 if (conf->raid_disks <= conf->previous_raid_disks)
616 if (degraded2 > degraded)
621 static int has_failed(struct r5conf *conf)
625 if (conf->mddev->reshape_position == MaxSector)
626 return conf->mddev->degraded > conf->max_degraded;
628 degraded = raid5_calc_degraded(conf);
629 if (degraded > conf->max_degraded)
635 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
636 int previous, int noblock, int noquiesce)
638 struct stripe_head *sh;
639 int hash = stripe_hash_locks_hash(sector);
640 int inc_empty_inactive_list_flag;
642 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
644 spin_lock_irq(conf->hash_locks + hash);
647 wait_event_lock_irq(conf->wait_for_quiescent,
648 conf->quiesce == 0 || noquiesce,
649 *(conf->hash_locks + hash));
650 sh = __find_stripe(conf, sector, conf->generation - previous);
652 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
653 sh = get_free_stripe(conf, hash);
654 if (!sh && !test_bit(R5_DID_ALLOC,
656 set_bit(R5_ALLOC_MORE,
659 if (noblock && sh == NULL)
662 r5c_check_stripe_cache_usage(conf);
664 set_bit(R5_INACTIVE_BLOCKED,
666 r5l_wake_reclaim(conf->log, 0);
668 conf->wait_for_stripe,
669 !list_empty(conf->inactive_list + hash) &&
670 (atomic_read(&conf->active_stripes)
671 < (conf->max_nr_stripes * 3 / 4)
672 || !test_bit(R5_INACTIVE_BLOCKED,
673 &conf->cache_state)),
674 *(conf->hash_locks + hash));
675 clear_bit(R5_INACTIVE_BLOCKED,
678 init_stripe(sh, sector, previous);
679 atomic_inc(&sh->count);
681 } else if (!atomic_inc_not_zero(&sh->count)) {
682 spin_lock(&conf->device_lock);
683 if (!atomic_read(&sh->count)) {
684 if (!test_bit(STRIPE_HANDLE, &sh->state))
685 atomic_inc(&conf->active_stripes);
686 BUG_ON(list_empty(&sh->lru) &&
687 !test_bit(STRIPE_EXPANDING, &sh->state));
688 inc_empty_inactive_list_flag = 0;
689 if (!list_empty(conf->inactive_list + hash))
690 inc_empty_inactive_list_flag = 1;
691 list_del_init(&sh->lru);
692 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
693 atomic_inc(&conf->empty_inactive_list_nr);
695 sh->group->stripes_cnt--;
699 atomic_inc(&sh->count);
700 spin_unlock(&conf->device_lock);
702 } while (sh == NULL);
704 spin_unlock_irq(conf->hash_locks + hash);
708 static bool is_full_stripe_write(struct stripe_head *sh)
710 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
711 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
714 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
717 spin_lock_irq(&sh2->stripe_lock);
718 spin_lock_nested(&sh1->stripe_lock, 1);
720 spin_lock_irq(&sh1->stripe_lock);
721 spin_lock_nested(&sh2->stripe_lock, 1);
725 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
727 spin_unlock(&sh1->stripe_lock);
728 spin_unlock_irq(&sh2->stripe_lock);
731 /* Only freshly new full stripe normal write stripe can be added to a batch list */
732 static bool stripe_can_batch(struct stripe_head *sh)
734 struct r5conf *conf = sh->raid_conf;
736 if (raid5_has_log(conf) || raid5_has_ppl(conf))
738 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
739 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
740 is_full_stripe_write(sh);
743 /* we only do back search */
744 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
746 struct stripe_head *head;
747 sector_t head_sector, tmp_sec;
750 int inc_empty_inactive_list_flag;
752 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
753 tmp_sec = sh->sector;
754 if (!sector_div(tmp_sec, conf->chunk_sectors))
756 head_sector = sh->sector - STRIPE_SECTORS;
758 hash = stripe_hash_locks_hash(head_sector);
759 spin_lock_irq(conf->hash_locks + hash);
760 head = __find_stripe(conf, head_sector, conf->generation);
761 if (head && !atomic_inc_not_zero(&head->count)) {
762 spin_lock(&conf->device_lock);
763 if (!atomic_read(&head->count)) {
764 if (!test_bit(STRIPE_HANDLE, &head->state))
765 atomic_inc(&conf->active_stripes);
766 BUG_ON(list_empty(&head->lru) &&
767 !test_bit(STRIPE_EXPANDING, &head->state));
768 inc_empty_inactive_list_flag = 0;
769 if (!list_empty(conf->inactive_list + hash))
770 inc_empty_inactive_list_flag = 1;
771 list_del_init(&head->lru);
772 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
773 atomic_inc(&conf->empty_inactive_list_nr);
775 head->group->stripes_cnt--;
779 atomic_inc(&head->count);
780 spin_unlock(&conf->device_lock);
782 spin_unlock_irq(conf->hash_locks + hash);
786 if (!stripe_can_batch(head))
789 lock_two_stripes(head, sh);
790 /* clear_batch_ready clear the flag */
791 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
798 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
800 if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
801 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
804 if (head->batch_head) {
805 spin_lock(&head->batch_head->batch_lock);
806 /* This batch list is already running */
807 if (!stripe_can_batch(head)) {
808 spin_unlock(&head->batch_head->batch_lock);
812 * We must assign batch_head of this stripe within the
813 * batch_lock, otherwise clear_batch_ready of batch head
814 * stripe could clear BATCH_READY bit of this stripe and
815 * this stripe->batch_head doesn't get assigned, which
816 * could confuse clear_batch_ready for this stripe
818 sh->batch_head = head->batch_head;
821 * at this point, head's BATCH_READY could be cleared, but we
822 * can still add the stripe to batch list
824 list_add(&sh->batch_list, &head->batch_list);
825 spin_unlock(&head->batch_head->batch_lock);
827 head->batch_head = head;
828 sh->batch_head = head->batch_head;
829 spin_lock(&head->batch_lock);
830 list_add_tail(&sh->batch_list, &head->batch_list);
831 spin_unlock(&head->batch_lock);
834 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
835 if (atomic_dec_return(&conf->preread_active_stripes)
837 md_wakeup_thread(conf->mddev->thread);
839 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
840 int seq = sh->bm_seq;
841 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
842 sh->batch_head->bm_seq > seq)
843 seq = sh->batch_head->bm_seq;
844 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
845 sh->batch_head->bm_seq = seq;
848 atomic_inc(&sh->count);
850 unlock_two_stripes(head, sh);
852 raid5_release_stripe(head);
855 /* Determine if 'data_offset' or 'new_data_offset' should be used
856 * in this stripe_head.
858 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
860 sector_t progress = conf->reshape_progress;
861 /* Need a memory barrier to make sure we see the value
862 * of conf->generation, or ->data_offset that was set before
863 * reshape_progress was updated.
866 if (progress == MaxSector)
868 if (sh->generation == conf->generation - 1)
870 /* We are in a reshape, and this is a new-generation stripe,
871 * so use new_data_offset.
876 static void dispatch_bio_list(struct bio_list *tmp)
880 while ((bio = bio_list_pop(tmp)))
881 generic_make_request(bio);
884 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
886 const struct r5pending_data *da = list_entry(a,
887 struct r5pending_data, sibling);
888 const struct r5pending_data *db = list_entry(b,
889 struct r5pending_data, sibling);
890 if (da->sector > db->sector)
892 if (da->sector < db->sector)
897 static void dispatch_defer_bios(struct r5conf *conf, int target,
898 struct bio_list *list)
900 struct r5pending_data *data;
901 struct list_head *first, *next = NULL;
904 if (conf->pending_data_cnt == 0)
907 list_sort(NULL, &conf->pending_list, cmp_stripe);
909 first = conf->pending_list.next;
911 /* temporarily move the head */
912 if (conf->next_pending_data)
913 list_move_tail(&conf->pending_list,
914 &conf->next_pending_data->sibling);
916 while (!list_empty(&conf->pending_list)) {
917 data = list_first_entry(&conf->pending_list,
918 struct r5pending_data, sibling);
919 if (&data->sibling == first)
920 first = data->sibling.next;
921 next = data->sibling.next;
923 bio_list_merge(list, &data->bios);
924 list_move(&data->sibling, &conf->free_list);
929 conf->pending_data_cnt -= cnt;
930 BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
932 if (next != &conf->pending_list)
933 conf->next_pending_data = list_entry(next,
934 struct r5pending_data, sibling);
936 conf->next_pending_data = NULL;
937 /* list isn't empty */
938 if (first != &conf->pending_list)
939 list_move_tail(&conf->pending_list, first);
942 static void flush_deferred_bios(struct r5conf *conf)
944 struct bio_list tmp = BIO_EMPTY_LIST;
946 if (conf->pending_data_cnt == 0)
949 spin_lock(&conf->pending_bios_lock);
950 dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
951 BUG_ON(conf->pending_data_cnt != 0);
952 spin_unlock(&conf->pending_bios_lock);
954 dispatch_bio_list(&tmp);
957 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
958 struct bio_list *bios)
960 struct bio_list tmp = BIO_EMPTY_LIST;
961 struct r5pending_data *ent;
963 spin_lock(&conf->pending_bios_lock);
964 ent = list_first_entry(&conf->free_list, struct r5pending_data,
966 list_move_tail(&ent->sibling, &conf->pending_list);
967 ent->sector = sector;
968 bio_list_init(&ent->bios);
969 bio_list_merge(&ent->bios, bios);
970 conf->pending_data_cnt++;
971 if (conf->pending_data_cnt >= PENDING_IO_MAX)
972 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
974 spin_unlock(&conf->pending_bios_lock);
976 dispatch_bio_list(&tmp);
980 raid5_end_read_request(struct bio *bi);
982 raid5_end_write_request(struct bio *bi);
984 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
986 struct r5conf *conf = sh->raid_conf;
987 int i, disks = sh->disks;
988 struct stripe_head *head_sh = sh;
989 struct bio_list pending_bios = BIO_EMPTY_LIST;
994 if (log_stripe(sh, s) == 0)
997 should_defer = conf->batch_bio_dispatch && conf->group_cnt;
999 for (i = disks; i--; ) {
1000 int op, op_flags = 0;
1001 int replace_only = 0;
1002 struct bio *bi, *rbi;
1003 struct md_rdev *rdev, *rrdev = NULL;
1006 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1008 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1010 if (test_bit(R5_Discard, &sh->dev[i].flags))
1011 op = REQ_OP_DISCARD;
1012 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1014 else if (test_and_clear_bit(R5_WantReplace,
1015 &sh->dev[i].flags)) {
1020 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1021 op_flags |= REQ_SYNC;
1024 bi = &sh->dev[i].req;
1025 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1028 rrdev = rcu_dereference(conf->disks[i].replacement);
1029 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1030 rdev = rcu_dereference(conf->disks[i].rdev);
1035 if (op_is_write(op)) {
1039 /* We raced and saw duplicates */
1042 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1047 if (rdev && test_bit(Faulty, &rdev->flags))
1050 atomic_inc(&rdev->nr_pending);
1051 if (rrdev && test_bit(Faulty, &rrdev->flags))
1054 atomic_inc(&rrdev->nr_pending);
1057 /* We have already checked bad blocks for reads. Now
1058 * need to check for writes. We never accept write errors
1059 * on the replacement, so we don't to check rrdev.
1061 while (op_is_write(op) && rdev &&
1062 test_bit(WriteErrorSeen, &rdev->flags)) {
1065 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1066 &first_bad, &bad_sectors);
1071 set_bit(BlockedBadBlocks, &rdev->flags);
1072 if (!conf->mddev->external &&
1073 conf->mddev->sb_flags) {
1074 /* It is very unlikely, but we might
1075 * still need to write out the
1076 * bad block log - better give it
1078 md_check_recovery(conf->mddev);
1081 * Because md_wait_for_blocked_rdev
1082 * will dec nr_pending, we must
1083 * increment it first.
1085 atomic_inc(&rdev->nr_pending);
1086 md_wait_for_blocked_rdev(rdev, conf->mddev);
1088 /* Acknowledged bad block - skip the write */
1089 rdev_dec_pending(rdev, conf->mddev);
1095 if (s->syncing || s->expanding || s->expanded
1097 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1099 set_bit(STRIPE_IO_STARTED, &sh->state);
1101 bio_set_dev(bi, rdev->bdev);
1102 bio_set_op_attrs(bi, op, op_flags);
1103 bi->bi_end_io = op_is_write(op)
1104 ? raid5_end_write_request
1105 : raid5_end_read_request;
1106 bi->bi_private = sh;
1108 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1109 __func__, (unsigned long long)sh->sector,
1111 atomic_inc(&sh->count);
1113 atomic_inc(&head_sh->count);
1114 if (use_new_offset(conf, sh))
1115 bi->bi_iter.bi_sector = (sh->sector
1116 + rdev->new_data_offset);
1118 bi->bi_iter.bi_sector = (sh->sector
1119 + rdev->data_offset);
1120 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1121 bi->bi_opf |= REQ_NOMERGE;
1123 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1124 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1126 if (!op_is_write(op) &&
1127 test_bit(R5_InJournal, &sh->dev[i].flags))
1129 * issuing read for a page in journal, this
1130 * must be preparing for prexor in rmw; read
1131 * the data into orig_page
1133 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1135 sh->dev[i].vec.bv_page = sh->dev[i].page;
1137 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1138 bi->bi_io_vec[0].bv_offset = 0;
1139 bi->bi_iter.bi_size = STRIPE_SIZE;
1140 bi->bi_write_hint = sh->dev[i].write_hint;
1142 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1144 * If this is discard request, set bi_vcnt 0. We don't
1145 * want to confuse SCSI because SCSI will replace payload
1147 if (op == REQ_OP_DISCARD)
1150 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1152 if (conf->mddev->gendisk)
1153 trace_block_bio_remap(bi->bi_disk->queue,
1154 bi, disk_devt(conf->mddev->gendisk),
1156 if (should_defer && op_is_write(op))
1157 bio_list_add(&pending_bios, bi);
1159 generic_make_request(bi);
1162 if (s->syncing || s->expanding || s->expanded
1164 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1166 set_bit(STRIPE_IO_STARTED, &sh->state);
1168 bio_set_dev(rbi, rrdev->bdev);
1169 bio_set_op_attrs(rbi, op, op_flags);
1170 BUG_ON(!op_is_write(op));
1171 rbi->bi_end_io = raid5_end_write_request;
1172 rbi->bi_private = sh;
1174 pr_debug("%s: for %llu schedule op %d on "
1175 "replacement disc %d\n",
1176 __func__, (unsigned long long)sh->sector,
1178 atomic_inc(&sh->count);
1180 atomic_inc(&head_sh->count);
1181 if (use_new_offset(conf, sh))
1182 rbi->bi_iter.bi_sector = (sh->sector
1183 + rrdev->new_data_offset);
1185 rbi->bi_iter.bi_sector = (sh->sector
1186 + rrdev->data_offset);
1187 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1188 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1189 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1191 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1192 rbi->bi_io_vec[0].bv_offset = 0;
1193 rbi->bi_iter.bi_size = STRIPE_SIZE;
1194 rbi->bi_write_hint = sh->dev[i].write_hint;
1195 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1197 * If this is discard request, set bi_vcnt 0. We don't
1198 * want to confuse SCSI because SCSI will replace payload
1200 if (op == REQ_OP_DISCARD)
1202 if (conf->mddev->gendisk)
1203 trace_block_bio_remap(rbi->bi_disk->queue,
1204 rbi, disk_devt(conf->mddev->gendisk),
1206 if (should_defer && op_is_write(op))
1207 bio_list_add(&pending_bios, rbi);
1209 generic_make_request(rbi);
1211 if (!rdev && !rrdev) {
1212 if (op_is_write(op))
1213 set_bit(STRIPE_DEGRADED, &sh->state);
1214 pr_debug("skip op %d on disc %d for sector %llu\n",
1215 bi->bi_opf, i, (unsigned long long)sh->sector);
1216 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1217 set_bit(STRIPE_HANDLE, &sh->state);
1220 if (!head_sh->batch_head)
1222 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1228 if (should_defer && !bio_list_empty(&pending_bios))
1229 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1232 static struct dma_async_tx_descriptor *
1233 async_copy_data(int frombio, struct bio *bio, struct page **page,
1234 sector_t sector, struct dma_async_tx_descriptor *tx,
1235 struct stripe_head *sh, int no_skipcopy)
1238 struct bvec_iter iter;
1239 struct page *bio_page;
1241 struct async_submit_ctl submit;
1242 enum async_tx_flags flags = 0;
1244 if (bio->bi_iter.bi_sector >= sector)
1245 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1247 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1250 flags |= ASYNC_TX_FENCE;
1251 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1253 bio_for_each_segment(bvl, bio, iter) {
1254 int len = bvl.bv_len;
1258 if (page_offset < 0) {
1259 b_offset = -page_offset;
1260 page_offset += b_offset;
1264 if (len > 0 && page_offset + len > STRIPE_SIZE)
1265 clen = STRIPE_SIZE - page_offset;
1270 b_offset += bvl.bv_offset;
1271 bio_page = bvl.bv_page;
1273 if (sh->raid_conf->skip_copy &&
1274 b_offset == 0 && page_offset == 0 &&
1275 clen == STRIPE_SIZE &&
1279 tx = async_memcpy(*page, bio_page, page_offset,
1280 b_offset, clen, &submit);
1282 tx = async_memcpy(bio_page, *page, b_offset,
1283 page_offset, clen, &submit);
1285 /* chain the operations */
1286 submit.depend_tx = tx;
1288 if (clen < len) /* hit end of page */
1296 static void ops_complete_biofill(void *stripe_head_ref)
1298 struct stripe_head *sh = stripe_head_ref;
1301 pr_debug("%s: stripe %llu\n", __func__,
1302 (unsigned long long)sh->sector);
1304 /* clear completed biofills */
1305 for (i = sh->disks; i--; ) {
1306 struct r5dev *dev = &sh->dev[i];
1308 /* acknowledge completion of a biofill operation */
1309 /* and check if we need to reply to a read request,
1310 * new R5_Wantfill requests are held off until
1311 * !STRIPE_BIOFILL_RUN
1313 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1314 struct bio *rbi, *rbi2;
1319 while (rbi && rbi->bi_iter.bi_sector <
1320 dev->sector + STRIPE_SECTORS) {
1321 rbi2 = r5_next_bio(rbi, dev->sector);
1327 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1329 set_bit(STRIPE_HANDLE, &sh->state);
1330 raid5_release_stripe(sh);
1333 static void ops_run_biofill(struct stripe_head *sh)
1335 struct dma_async_tx_descriptor *tx = NULL;
1336 struct async_submit_ctl submit;
1339 BUG_ON(sh->batch_head);
1340 pr_debug("%s: stripe %llu\n", __func__,
1341 (unsigned long long)sh->sector);
1343 for (i = sh->disks; i--; ) {
1344 struct r5dev *dev = &sh->dev[i];
1345 if (test_bit(R5_Wantfill, &dev->flags)) {
1347 spin_lock_irq(&sh->stripe_lock);
1348 dev->read = rbi = dev->toread;
1350 spin_unlock_irq(&sh->stripe_lock);
1351 while (rbi && rbi->bi_iter.bi_sector <
1352 dev->sector + STRIPE_SECTORS) {
1353 tx = async_copy_data(0, rbi, &dev->page,
1354 dev->sector, tx, sh, 0);
1355 rbi = r5_next_bio(rbi, dev->sector);
1360 atomic_inc(&sh->count);
1361 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1362 async_trigger_callback(&submit);
1365 static void mark_target_uptodate(struct stripe_head *sh, int target)
1372 tgt = &sh->dev[target];
1373 set_bit(R5_UPTODATE, &tgt->flags);
1374 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1375 clear_bit(R5_Wantcompute, &tgt->flags);
1378 static void ops_complete_compute(void *stripe_head_ref)
1380 struct stripe_head *sh = stripe_head_ref;
1382 pr_debug("%s: stripe %llu\n", __func__,
1383 (unsigned long long)sh->sector);
1385 /* mark the computed target(s) as uptodate */
1386 mark_target_uptodate(sh, sh->ops.target);
1387 mark_target_uptodate(sh, sh->ops.target2);
1389 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1390 if (sh->check_state == check_state_compute_run)
1391 sh->check_state = check_state_compute_result;
1392 set_bit(STRIPE_HANDLE, &sh->state);
1393 raid5_release_stripe(sh);
1396 /* return a pointer to the address conversion region of the scribble buffer */
1397 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1398 struct raid5_percpu *percpu, int i)
1402 addr = flex_array_get(percpu->scribble, i);
1403 return addr + sizeof(struct page *) * (sh->disks + 2);
1406 /* return a pointer to the address conversion region of the scribble buffer */
1407 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1411 addr = flex_array_get(percpu->scribble, i);
1415 static struct dma_async_tx_descriptor *
1416 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1418 int disks = sh->disks;
1419 struct page **xor_srcs = to_addr_page(percpu, 0);
1420 int target = sh->ops.target;
1421 struct r5dev *tgt = &sh->dev[target];
1422 struct page *xor_dest = tgt->page;
1424 struct dma_async_tx_descriptor *tx;
1425 struct async_submit_ctl submit;
1428 BUG_ON(sh->batch_head);
1430 pr_debug("%s: stripe %llu block: %d\n",
1431 __func__, (unsigned long long)sh->sector, target);
1432 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1434 for (i = disks; i--; )
1436 xor_srcs[count++] = sh->dev[i].page;
1438 atomic_inc(&sh->count);
1440 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1441 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1442 if (unlikely(count == 1))
1443 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1445 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1450 /* set_syndrome_sources - populate source buffers for gen_syndrome
1451 * @srcs - (struct page *) array of size sh->disks
1452 * @sh - stripe_head to parse
1454 * Populates srcs in proper layout order for the stripe and returns the
1455 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1456 * destination buffer is recorded in srcs[count] and the Q destination
1457 * is recorded in srcs[count+1]].
1459 static int set_syndrome_sources(struct page **srcs,
1460 struct stripe_head *sh,
1463 int disks = sh->disks;
1464 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1465 int d0_idx = raid6_d0(sh);
1469 for (i = 0; i < disks; i++)
1475 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1476 struct r5dev *dev = &sh->dev[i];
1478 if (i == sh->qd_idx || i == sh->pd_idx ||
1479 (srctype == SYNDROME_SRC_ALL) ||
1480 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1481 (test_bit(R5_Wantdrain, &dev->flags) ||
1482 test_bit(R5_InJournal, &dev->flags))) ||
1483 (srctype == SYNDROME_SRC_WRITTEN &&
1485 test_bit(R5_InJournal, &dev->flags)))) {
1486 if (test_bit(R5_InJournal, &dev->flags))
1487 srcs[slot] = sh->dev[i].orig_page;
1489 srcs[slot] = sh->dev[i].page;
1491 i = raid6_next_disk(i, disks);
1492 } while (i != d0_idx);
1494 return syndrome_disks;
1497 static struct dma_async_tx_descriptor *
1498 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1500 int disks = sh->disks;
1501 struct page **blocks = to_addr_page(percpu, 0);
1503 int qd_idx = sh->qd_idx;
1504 struct dma_async_tx_descriptor *tx;
1505 struct async_submit_ctl submit;
1511 BUG_ON(sh->batch_head);
1512 if (sh->ops.target < 0)
1513 target = sh->ops.target2;
1514 else if (sh->ops.target2 < 0)
1515 target = sh->ops.target;
1517 /* we should only have one valid target */
1520 pr_debug("%s: stripe %llu block: %d\n",
1521 __func__, (unsigned long long)sh->sector, target);
1523 tgt = &sh->dev[target];
1524 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1527 atomic_inc(&sh->count);
1529 if (target == qd_idx) {
1530 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1531 blocks[count] = NULL; /* regenerating p is not necessary */
1532 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1533 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1534 ops_complete_compute, sh,
1535 to_addr_conv(sh, percpu, 0));
1536 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1538 /* Compute any data- or p-drive using XOR */
1540 for (i = disks; i-- ; ) {
1541 if (i == target || i == qd_idx)
1543 blocks[count++] = sh->dev[i].page;
1546 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1547 NULL, ops_complete_compute, sh,
1548 to_addr_conv(sh, percpu, 0));
1549 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1555 static struct dma_async_tx_descriptor *
1556 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1558 int i, count, disks = sh->disks;
1559 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1560 int d0_idx = raid6_d0(sh);
1561 int faila = -1, failb = -1;
1562 int target = sh->ops.target;
1563 int target2 = sh->ops.target2;
1564 struct r5dev *tgt = &sh->dev[target];
1565 struct r5dev *tgt2 = &sh->dev[target2];
1566 struct dma_async_tx_descriptor *tx;
1567 struct page **blocks = to_addr_page(percpu, 0);
1568 struct async_submit_ctl submit;
1570 BUG_ON(sh->batch_head);
1571 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1572 __func__, (unsigned long long)sh->sector, target, target2);
1573 BUG_ON(target < 0 || target2 < 0);
1574 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1575 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1577 /* we need to open-code set_syndrome_sources to handle the
1578 * slot number conversion for 'faila' and 'failb'
1580 for (i = 0; i < disks ; i++)
1585 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1587 blocks[slot] = sh->dev[i].page;
1593 i = raid6_next_disk(i, disks);
1594 } while (i != d0_idx);
1596 BUG_ON(faila == failb);
1599 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1600 __func__, (unsigned long long)sh->sector, faila, failb);
1602 atomic_inc(&sh->count);
1604 if (failb == syndrome_disks+1) {
1605 /* Q disk is one of the missing disks */
1606 if (faila == syndrome_disks) {
1607 /* Missing P+Q, just recompute */
1608 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1609 ops_complete_compute, sh,
1610 to_addr_conv(sh, percpu, 0));
1611 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1612 STRIPE_SIZE, &submit);
1616 int qd_idx = sh->qd_idx;
1618 /* Missing D+Q: recompute D from P, then recompute Q */
1619 if (target == qd_idx)
1620 data_target = target2;
1622 data_target = target;
1625 for (i = disks; i-- ; ) {
1626 if (i == data_target || i == qd_idx)
1628 blocks[count++] = sh->dev[i].page;
1630 dest = sh->dev[data_target].page;
1631 init_async_submit(&submit,
1632 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1634 to_addr_conv(sh, percpu, 0));
1635 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1638 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1639 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1640 ops_complete_compute, sh,
1641 to_addr_conv(sh, percpu, 0));
1642 return async_gen_syndrome(blocks, 0, count+2,
1643 STRIPE_SIZE, &submit);
1646 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1647 ops_complete_compute, sh,
1648 to_addr_conv(sh, percpu, 0));
1649 if (failb == syndrome_disks) {
1650 /* We're missing D+P. */
1651 return async_raid6_datap_recov(syndrome_disks+2,
1655 /* We're missing D+D. */
1656 return async_raid6_2data_recov(syndrome_disks+2,
1657 STRIPE_SIZE, faila, failb,
1663 static void ops_complete_prexor(void *stripe_head_ref)
1665 struct stripe_head *sh = stripe_head_ref;
1667 pr_debug("%s: stripe %llu\n", __func__,
1668 (unsigned long long)sh->sector);
1670 if (r5c_is_writeback(sh->raid_conf->log))
1672 * raid5-cache write back uses orig_page during prexor.
1673 * After prexor, it is time to free orig_page
1675 r5c_release_extra_page(sh);
1678 static struct dma_async_tx_descriptor *
1679 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1680 struct dma_async_tx_descriptor *tx)
1682 int disks = sh->disks;
1683 struct page **xor_srcs = to_addr_page(percpu, 0);
1684 int count = 0, pd_idx = sh->pd_idx, i;
1685 struct async_submit_ctl submit;
1687 /* existing parity data subtracted */
1688 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1690 BUG_ON(sh->batch_head);
1691 pr_debug("%s: stripe %llu\n", __func__,
1692 (unsigned long long)sh->sector);
1694 for (i = disks; i--; ) {
1695 struct r5dev *dev = &sh->dev[i];
1696 /* Only process blocks that are known to be uptodate */
1697 if (test_bit(R5_InJournal, &dev->flags))
1698 xor_srcs[count++] = dev->orig_page;
1699 else if (test_bit(R5_Wantdrain, &dev->flags))
1700 xor_srcs[count++] = dev->page;
1703 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1704 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1705 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1710 static struct dma_async_tx_descriptor *
1711 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1712 struct dma_async_tx_descriptor *tx)
1714 struct page **blocks = to_addr_page(percpu, 0);
1716 struct async_submit_ctl submit;
1718 pr_debug("%s: stripe %llu\n", __func__,
1719 (unsigned long long)sh->sector);
1721 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1723 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1724 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1725 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1730 static struct dma_async_tx_descriptor *
1731 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1733 struct r5conf *conf = sh->raid_conf;
1734 int disks = sh->disks;
1736 struct stripe_head *head_sh = sh;
1738 pr_debug("%s: stripe %llu\n", __func__,
1739 (unsigned long long)sh->sector);
1741 for (i = disks; i--; ) {
1746 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1752 * clear R5_InJournal, so when rewriting a page in
1753 * journal, it is not skipped by r5l_log_stripe()
1755 clear_bit(R5_InJournal, &dev->flags);
1756 spin_lock_irq(&sh->stripe_lock);
1757 chosen = dev->towrite;
1758 dev->towrite = NULL;
1759 sh->overwrite_disks = 0;
1760 BUG_ON(dev->written);
1761 wbi = dev->written = chosen;
1762 spin_unlock_irq(&sh->stripe_lock);
1763 WARN_ON(dev->page != dev->orig_page);
1765 while (wbi && wbi->bi_iter.bi_sector <
1766 dev->sector + STRIPE_SECTORS) {
1767 if (wbi->bi_opf & REQ_FUA)
1768 set_bit(R5_WantFUA, &dev->flags);
1769 if (wbi->bi_opf & REQ_SYNC)
1770 set_bit(R5_SyncIO, &dev->flags);
1771 if (bio_op(wbi) == REQ_OP_DISCARD)
1772 set_bit(R5_Discard, &dev->flags);
1774 tx = async_copy_data(1, wbi, &dev->page,
1775 dev->sector, tx, sh,
1776 r5c_is_writeback(conf->log));
1777 if (dev->page != dev->orig_page &&
1778 !r5c_is_writeback(conf->log)) {
1779 set_bit(R5_SkipCopy, &dev->flags);
1780 clear_bit(R5_UPTODATE, &dev->flags);
1781 clear_bit(R5_OVERWRITE, &dev->flags);
1784 wbi = r5_next_bio(wbi, dev->sector);
1787 if (head_sh->batch_head) {
1788 sh = list_first_entry(&sh->batch_list,
1801 static void ops_complete_reconstruct(void *stripe_head_ref)
1803 struct stripe_head *sh = stripe_head_ref;
1804 int disks = sh->disks;
1805 int pd_idx = sh->pd_idx;
1806 int qd_idx = sh->qd_idx;
1808 bool fua = false, sync = false, discard = false;
1810 pr_debug("%s: stripe %llu\n", __func__,
1811 (unsigned long long)sh->sector);
1813 for (i = disks; i--; ) {
1814 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1815 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1816 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1819 for (i = disks; i--; ) {
1820 struct r5dev *dev = &sh->dev[i];
1822 if (dev->written || i == pd_idx || i == qd_idx) {
1823 if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1824 set_bit(R5_UPTODATE, &dev->flags);
1825 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1826 set_bit(R5_Expanded, &dev->flags);
1829 set_bit(R5_WantFUA, &dev->flags);
1831 set_bit(R5_SyncIO, &dev->flags);
1835 if (sh->reconstruct_state == reconstruct_state_drain_run)
1836 sh->reconstruct_state = reconstruct_state_drain_result;
1837 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1838 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1840 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1841 sh->reconstruct_state = reconstruct_state_result;
1844 set_bit(STRIPE_HANDLE, &sh->state);
1845 raid5_release_stripe(sh);
1849 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1850 struct dma_async_tx_descriptor *tx)
1852 int disks = sh->disks;
1853 struct page **xor_srcs;
1854 struct async_submit_ctl submit;
1855 int count, pd_idx = sh->pd_idx, i;
1856 struct page *xor_dest;
1858 unsigned long flags;
1860 struct stripe_head *head_sh = sh;
1863 pr_debug("%s: stripe %llu\n", __func__,
1864 (unsigned long long)sh->sector);
1866 for (i = 0; i < sh->disks; i++) {
1869 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1872 if (i >= sh->disks) {
1873 atomic_inc(&sh->count);
1874 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1875 ops_complete_reconstruct(sh);
1880 xor_srcs = to_addr_page(percpu, j);
1881 /* check if prexor is active which means only process blocks
1882 * that are part of a read-modify-write (written)
1884 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1886 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1887 for (i = disks; i--; ) {
1888 struct r5dev *dev = &sh->dev[i];
1889 if (head_sh->dev[i].written ||
1890 test_bit(R5_InJournal, &head_sh->dev[i].flags))
1891 xor_srcs[count++] = dev->page;
1894 xor_dest = sh->dev[pd_idx].page;
1895 for (i = disks; i--; ) {
1896 struct r5dev *dev = &sh->dev[i];
1898 xor_srcs[count++] = dev->page;
1902 /* 1/ if we prexor'd then the dest is reused as a source
1903 * 2/ if we did not prexor then we are redoing the parity
1904 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1905 * for the synchronous xor case
1907 last_stripe = !head_sh->batch_head ||
1908 list_first_entry(&sh->batch_list,
1909 struct stripe_head, batch_list) == head_sh;
1911 flags = ASYNC_TX_ACK |
1912 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1914 atomic_inc(&head_sh->count);
1915 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1916 to_addr_conv(sh, percpu, j));
1918 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1919 init_async_submit(&submit, flags, tx, NULL, NULL,
1920 to_addr_conv(sh, percpu, j));
1923 if (unlikely(count == 1))
1924 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1926 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1929 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1936 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1937 struct dma_async_tx_descriptor *tx)
1939 struct async_submit_ctl submit;
1940 struct page **blocks;
1941 int count, i, j = 0;
1942 struct stripe_head *head_sh = sh;
1945 unsigned long txflags;
1947 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1949 for (i = 0; i < sh->disks; i++) {
1950 if (sh->pd_idx == i || sh->qd_idx == i)
1952 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1955 if (i >= sh->disks) {
1956 atomic_inc(&sh->count);
1957 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1958 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1959 ops_complete_reconstruct(sh);
1964 blocks = to_addr_page(percpu, j);
1966 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1967 synflags = SYNDROME_SRC_WRITTEN;
1968 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1970 synflags = SYNDROME_SRC_ALL;
1971 txflags = ASYNC_TX_ACK;
1974 count = set_syndrome_sources(blocks, sh, synflags);
1975 last_stripe = !head_sh->batch_head ||
1976 list_first_entry(&sh->batch_list,
1977 struct stripe_head, batch_list) == head_sh;
1980 atomic_inc(&head_sh->count);
1981 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1982 head_sh, to_addr_conv(sh, percpu, j));
1984 init_async_submit(&submit, 0, tx, NULL, NULL,
1985 to_addr_conv(sh, percpu, j));
1986 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1989 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1995 static void ops_complete_check(void *stripe_head_ref)
1997 struct stripe_head *sh = stripe_head_ref;
1999 pr_debug("%s: stripe %llu\n", __func__,
2000 (unsigned long long)sh->sector);
2002 sh->check_state = check_state_check_result;
2003 set_bit(STRIPE_HANDLE, &sh->state);
2004 raid5_release_stripe(sh);
2007 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2009 int disks = sh->disks;
2010 int pd_idx = sh->pd_idx;
2011 int qd_idx = sh->qd_idx;
2012 struct page *xor_dest;
2013 struct page **xor_srcs = to_addr_page(percpu, 0);
2014 struct dma_async_tx_descriptor *tx;
2015 struct async_submit_ctl submit;
2019 pr_debug("%s: stripe %llu\n", __func__,
2020 (unsigned long long)sh->sector);
2022 BUG_ON(sh->batch_head);
2024 xor_dest = sh->dev[pd_idx].page;
2025 xor_srcs[count++] = xor_dest;
2026 for (i = disks; i--; ) {
2027 if (i == pd_idx || i == qd_idx)
2029 xor_srcs[count++] = sh->dev[i].page;
2032 init_async_submit(&submit, 0, NULL, NULL, NULL,
2033 to_addr_conv(sh, percpu, 0));
2034 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2035 &sh->ops.zero_sum_result, &submit);
2037 atomic_inc(&sh->count);
2038 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2039 tx = async_trigger_callback(&submit);
2042 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2044 struct page **srcs = to_addr_page(percpu, 0);
2045 struct async_submit_ctl submit;
2048 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2049 (unsigned long long)sh->sector, checkp);
2051 BUG_ON(sh->batch_head);
2052 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2056 atomic_inc(&sh->count);
2057 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2058 sh, to_addr_conv(sh, percpu, 0));
2059 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2060 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2063 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2065 int overlap_clear = 0, i, disks = sh->disks;
2066 struct dma_async_tx_descriptor *tx = NULL;
2067 struct r5conf *conf = sh->raid_conf;
2068 int level = conf->level;
2069 struct raid5_percpu *percpu;
2073 percpu = per_cpu_ptr(conf->percpu, cpu);
2074 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2075 ops_run_biofill(sh);
2079 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2081 tx = ops_run_compute5(sh, percpu);
2083 if (sh->ops.target2 < 0 || sh->ops.target < 0)
2084 tx = ops_run_compute6_1(sh, percpu);
2086 tx = ops_run_compute6_2(sh, percpu);
2088 /* terminate the chain if reconstruct is not set to be run */
2089 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2093 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2095 tx = ops_run_prexor5(sh, percpu, tx);
2097 tx = ops_run_prexor6(sh, percpu, tx);
2100 if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2101 tx = ops_run_partial_parity(sh, percpu, tx);
2103 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2104 tx = ops_run_biodrain(sh, tx);
2108 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2110 ops_run_reconstruct5(sh, percpu, tx);
2112 ops_run_reconstruct6(sh, percpu, tx);
2115 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2116 if (sh->check_state == check_state_run)
2117 ops_run_check_p(sh, percpu);
2118 else if (sh->check_state == check_state_run_q)
2119 ops_run_check_pq(sh, percpu, 0);
2120 else if (sh->check_state == check_state_run_pq)
2121 ops_run_check_pq(sh, percpu, 1);
2126 if (overlap_clear && !sh->batch_head)
2127 for (i = disks; i--; ) {
2128 struct r5dev *dev = &sh->dev[i];
2129 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2130 wake_up(&sh->raid_conf->wait_for_overlap);
2135 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2138 __free_page(sh->ppl_page);
2139 kmem_cache_free(sc, sh);
2142 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2143 int disks, struct r5conf *conf)
2145 struct stripe_head *sh;
2148 sh = kmem_cache_zalloc(sc, gfp);
2150 spin_lock_init(&sh->stripe_lock);
2151 spin_lock_init(&sh->batch_lock);
2152 INIT_LIST_HEAD(&sh->batch_list);
2153 INIT_LIST_HEAD(&sh->lru);
2154 INIT_LIST_HEAD(&sh->r5c);
2155 INIT_LIST_HEAD(&sh->log_list);
2156 atomic_set(&sh->count, 1);
2157 sh->raid_conf = conf;
2158 sh->log_start = MaxSector;
2159 for (i = 0; i < disks; i++) {
2160 struct r5dev *dev = &sh->dev[i];
2162 bio_init(&dev->req, &dev->vec, 1);
2163 bio_init(&dev->rreq, &dev->rvec, 1);
2166 if (raid5_has_ppl(conf)) {
2167 sh->ppl_page = alloc_page(gfp);
2168 if (!sh->ppl_page) {
2169 free_stripe(sc, sh);
2176 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2178 struct stripe_head *sh;
2180 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2184 if (grow_buffers(sh, gfp)) {
2186 free_stripe(conf->slab_cache, sh);
2189 sh->hash_lock_index =
2190 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2191 /* we just created an active stripe so... */
2192 atomic_inc(&conf->active_stripes);
2194 raid5_release_stripe(sh);
2195 conf->max_nr_stripes++;
2199 static int grow_stripes(struct r5conf *conf, int num)
2201 struct kmem_cache *sc;
2202 size_t namelen = sizeof(conf->cache_name[0]);
2203 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2205 if (conf->mddev->gendisk)
2206 snprintf(conf->cache_name[0], namelen,
2207 "raid%d-%s", conf->level, mdname(conf->mddev));
2209 snprintf(conf->cache_name[0], namelen,
2210 "raid%d-%p", conf->level, conf->mddev);
2211 snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2213 conf->active_name = 0;
2214 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2215 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2219 conf->slab_cache = sc;
2220 conf->pool_size = devs;
2222 if (!grow_one_stripe(conf, GFP_KERNEL))
2229 * scribble_len - return the required size of the scribble region
2230 * @num - total number of disks in the array
2232 * The size must be enough to contain:
2233 * 1/ a struct page pointer for each device in the array +2
2234 * 2/ room to convert each entry in (1) to its corresponding dma
2235 * (dma_map_page()) or page (page_address()) address.
2237 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2238 * calculate over all devices (not just the data blocks), using zeros in place
2239 * of the P and Q blocks.
2241 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2243 struct flex_array *ret;
2246 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2247 ret = flex_array_alloc(len, cnt, flags);
2250 /* always prealloc all elements, so no locking is required */
2251 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2252 flex_array_free(ret);
2258 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2264 * Never shrink. And mddev_suspend() could deadlock if this is called
2265 * from raid5d. In that case, scribble_disks and scribble_sectors
2266 * should equal to new_disks and new_sectors
2268 if (conf->scribble_disks >= new_disks &&
2269 conf->scribble_sectors >= new_sectors)
2271 mddev_suspend(conf->mddev);
2273 for_each_present_cpu(cpu) {
2274 struct raid5_percpu *percpu;
2275 struct flex_array *scribble;
2277 percpu = per_cpu_ptr(conf->percpu, cpu);
2278 scribble = scribble_alloc(new_disks,
2279 new_sectors / STRIPE_SECTORS,
2283 flex_array_free(percpu->scribble);
2284 percpu->scribble = scribble;
2291 mddev_resume(conf->mddev);
2293 conf->scribble_disks = new_disks;
2294 conf->scribble_sectors = new_sectors;
2299 static int resize_stripes(struct r5conf *conf, int newsize)
2301 /* Make all the stripes able to hold 'newsize' devices.
2302 * New slots in each stripe get 'page' set to a new page.
2304 * This happens in stages:
2305 * 1/ create a new kmem_cache and allocate the required number of
2307 * 2/ gather all the old stripe_heads and transfer the pages across
2308 * to the new stripe_heads. This will have the side effect of
2309 * freezing the array as once all stripe_heads have been collected,
2310 * no IO will be possible. Old stripe heads are freed once their
2311 * pages have been transferred over, and the old kmem_cache is
2312 * freed when all stripes are done.
2313 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2314 * we simple return a failure status - no need to clean anything up.
2315 * 4/ allocate new pages for the new slots in the new stripe_heads.
2316 * If this fails, we don't bother trying the shrink the
2317 * stripe_heads down again, we just leave them as they are.
2318 * As each stripe_head is processed the new one is released into
2321 * Once step2 is started, we cannot afford to wait for a write,
2322 * so we use GFP_NOIO allocations.
2324 struct stripe_head *osh, *nsh;
2325 LIST_HEAD(newstripes);
2326 struct disk_info *ndisks;
2328 struct kmem_cache *sc;
2332 md_allow_write(conf->mddev);
2335 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2336 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2341 /* Need to ensure auto-resizing doesn't interfere */
2342 mutex_lock(&conf->cache_size_mutex);
2344 for (i = conf->max_nr_stripes; i; i--) {
2345 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2349 list_add(&nsh->lru, &newstripes);
2352 /* didn't get enough, give up */
2353 while (!list_empty(&newstripes)) {
2354 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2355 list_del(&nsh->lru);
2356 free_stripe(sc, nsh);
2358 kmem_cache_destroy(sc);
2359 mutex_unlock(&conf->cache_size_mutex);
2362 /* Step 2 - Must use GFP_NOIO now.
2363 * OK, we have enough stripes, start collecting inactive
2364 * stripes and copying them over
2368 list_for_each_entry(nsh, &newstripes, lru) {
2369 lock_device_hash_lock(conf, hash);
2370 wait_event_cmd(conf->wait_for_stripe,
2371 !list_empty(conf->inactive_list + hash),
2372 unlock_device_hash_lock(conf, hash),
2373 lock_device_hash_lock(conf, hash));
2374 osh = get_free_stripe(conf, hash);
2375 unlock_device_hash_lock(conf, hash);
2377 for(i=0; i<conf->pool_size; i++) {
2378 nsh->dev[i].page = osh->dev[i].page;
2379 nsh->dev[i].orig_page = osh->dev[i].page;
2381 nsh->hash_lock_index = hash;
2382 free_stripe(conf->slab_cache, osh);
2384 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2385 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2390 kmem_cache_destroy(conf->slab_cache);
2393 * At this point, we are holding all the stripes so the array
2394 * is completely stalled, so now is a good time to resize
2395 * conf->disks and the scribble region
2397 ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2399 for (i = 0; i < conf->pool_size; i++)
2400 ndisks[i] = conf->disks[i];
2402 for (i = conf->pool_size; i < newsize; i++) {
2403 ndisks[i].extra_page = alloc_page(GFP_NOIO);
2404 if (!ndisks[i].extra_page)
2409 for (i = conf->pool_size; i < newsize; i++)
2410 if (ndisks[i].extra_page)
2411 put_page(ndisks[i].extra_page);
2415 conf->disks = ndisks;
2420 conf->slab_cache = sc;
2421 conf->active_name = 1-conf->active_name;
2423 /* Step 4, return new stripes to service */
2424 while(!list_empty(&newstripes)) {
2425 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2426 list_del_init(&nsh->lru);
2428 for (i=conf->raid_disks; i < newsize; i++)
2429 if (nsh->dev[i].page == NULL) {
2430 struct page *p = alloc_page(GFP_NOIO);
2431 nsh->dev[i].page = p;
2432 nsh->dev[i].orig_page = p;
2436 raid5_release_stripe(nsh);
2438 /* critical section pass, GFP_NOIO no longer needed */
2441 conf->pool_size = newsize;
2442 mutex_unlock(&conf->cache_size_mutex);
2447 static int drop_one_stripe(struct r5conf *conf)
2449 struct stripe_head *sh;
2450 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2452 spin_lock_irq(conf->hash_locks + hash);
2453 sh = get_free_stripe(conf, hash);
2454 spin_unlock_irq(conf->hash_locks + hash);
2457 BUG_ON(atomic_read(&sh->count));
2459 free_stripe(conf->slab_cache, sh);
2460 atomic_dec(&conf->active_stripes);
2461 conf->max_nr_stripes--;
2465 static void shrink_stripes(struct r5conf *conf)
2467 while (conf->max_nr_stripes &&
2468 drop_one_stripe(conf))
2471 kmem_cache_destroy(conf->slab_cache);
2472 conf->slab_cache = NULL;
2475 static void raid5_end_read_request(struct bio * bi)
2477 struct stripe_head *sh = bi->bi_private;
2478 struct r5conf *conf = sh->raid_conf;
2479 int disks = sh->disks, i;
2480 char b[BDEVNAME_SIZE];
2481 struct md_rdev *rdev = NULL;
2484 for (i=0 ; i<disks; i++)
2485 if (bi == &sh->dev[i].req)
2488 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2489 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2496 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2497 /* If replacement finished while this request was outstanding,
2498 * 'replacement' might be NULL already.
2499 * In that case it moved down to 'rdev'.
2500 * rdev is not removed until all requests are finished.
2502 rdev = conf->disks[i].replacement;
2504 rdev = conf->disks[i].rdev;
2506 if (use_new_offset(conf, sh))
2507 s = sh->sector + rdev->new_data_offset;
2509 s = sh->sector + rdev->data_offset;
2510 if (!bi->bi_status) {
2511 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2512 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2513 /* Note that this cannot happen on a
2514 * replacement device. We just fail those on
2517 pr_info_ratelimited(
2518 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2519 mdname(conf->mddev), STRIPE_SECTORS,
2520 (unsigned long long)s,
2521 bdevname(rdev->bdev, b));
2522 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2523 clear_bit(R5_ReadError, &sh->dev[i].flags);
2524 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2525 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2526 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2528 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2530 * end read for a page in journal, this
2531 * must be preparing for prexor in rmw
2533 set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2535 if (atomic_read(&rdev->read_errors))
2536 atomic_set(&rdev->read_errors, 0);
2538 const char *bdn = bdevname(rdev->bdev, b);
2542 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2543 if (!(bi->bi_status == BLK_STS_PROTECTION))
2544 atomic_inc(&rdev->read_errors);
2545 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2546 pr_warn_ratelimited(
2547 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2548 mdname(conf->mddev),
2549 (unsigned long long)s,
2551 else if (conf->mddev->degraded >= conf->max_degraded) {
2553 pr_warn_ratelimited(
2554 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2555 mdname(conf->mddev),
2556 (unsigned long long)s,
2558 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2561 pr_warn_ratelimited(
2562 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2563 mdname(conf->mddev),
2564 (unsigned long long)s,
2566 } else if (atomic_read(&rdev->read_errors)
2567 > conf->max_nr_stripes)
2568 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2569 mdname(conf->mddev), bdn);
2572 if (set_bad && test_bit(In_sync, &rdev->flags)
2573 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2576 if (sh->qd_idx >= 0 && sh->pd_idx == i)
2577 set_bit(R5_ReadError, &sh->dev[i].flags);
2578 else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2579 set_bit(R5_ReadError, &sh->dev[i].flags);
2580 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2582 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2584 clear_bit(R5_ReadError, &sh->dev[i].flags);
2585 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2587 && test_bit(In_sync, &rdev->flags)
2588 && rdev_set_badblocks(
2589 rdev, sh->sector, STRIPE_SECTORS, 0)))
2590 md_error(conf->mddev, rdev);
2593 rdev_dec_pending(rdev, conf->mddev);
2595 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2596 set_bit(STRIPE_HANDLE, &sh->state);
2597 raid5_release_stripe(sh);
2600 static void raid5_end_write_request(struct bio *bi)
2602 struct stripe_head *sh = bi->bi_private;
2603 struct r5conf *conf = sh->raid_conf;
2604 int disks = sh->disks, i;
2605 struct md_rdev *uninitialized_var(rdev);
2608 int replacement = 0;
2610 for (i = 0 ; i < disks; i++) {
2611 if (bi == &sh->dev[i].req) {
2612 rdev = conf->disks[i].rdev;
2615 if (bi == &sh->dev[i].rreq) {
2616 rdev = conf->disks[i].replacement;
2620 /* rdev was removed and 'replacement'
2621 * replaced it. rdev is not removed
2622 * until all requests are finished.
2624 rdev = conf->disks[i].rdev;
2628 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2629 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2639 md_error(conf->mddev, rdev);
2640 else if (is_badblock(rdev, sh->sector,
2642 &first_bad, &bad_sectors))
2643 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2645 if (bi->bi_status) {
2646 set_bit(STRIPE_DEGRADED, &sh->state);
2647 set_bit(WriteErrorSeen, &rdev->flags);
2648 set_bit(R5_WriteError, &sh->dev[i].flags);
2649 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2650 set_bit(MD_RECOVERY_NEEDED,
2651 &rdev->mddev->recovery);
2652 } else if (is_badblock(rdev, sh->sector,
2654 &first_bad, &bad_sectors)) {
2655 set_bit(R5_MadeGood, &sh->dev[i].flags);
2656 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2657 /* That was a successful write so make
2658 * sure it looks like we already did
2661 set_bit(R5_ReWrite, &sh->dev[i].flags);
2664 rdev_dec_pending(rdev, conf->mddev);
2666 if (sh->batch_head && bi->bi_status && !replacement)
2667 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2670 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2671 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2672 set_bit(STRIPE_HANDLE, &sh->state);
2673 raid5_release_stripe(sh);
2675 if (sh->batch_head && sh != sh->batch_head)
2676 raid5_release_stripe(sh->batch_head);
2679 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2681 char b[BDEVNAME_SIZE];
2682 struct r5conf *conf = mddev->private;
2683 unsigned long flags;
2684 pr_debug("raid456: error called\n");
2686 spin_lock_irqsave(&conf->device_lock, flags);
2687 set_bit(Faulty, &rdev->flags);
2688 clear_bit(In_sync, &rdev->flags);
2689 mddev->degraded = raid5_calc_degraded(conf);
2690 spin_unlock_irqrestore(&conf->device_lock, flags);
2691 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2693 set_bit(Blocked, &rdev->flags);
2694 set_mask_bits(&mddev->sb_flags, 0,
2695 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2696 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2697 "md/raid:%s: Operation continuing on %d devices.\n",
2699 bdevname(rdev->bdev, b),
2701 conf->raid_disks - mddev->degraded);
2702 r5c_update_on_rdev_error(mddev, rdev);
2706 * Input: a 'big' sector number,
2707 * Output: index of the data and parity disk, and the sector # in them.
2709 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2710 int previous, int *dd_idx,
2711 struct stripe_head *sh)
2713 sector_t stripe, stripe2;
2714 sector_t chunk_number;
2715 unsigned int chunk_offset;
2718 sector_t new_sector;
2719 int algorithm = previous ? conf->prev_algo
2721 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2722 : conf->chunk_sectors;
2723 int raid_disks = previous ? conf->previous_raid_disks
2725 int data_disks = raid_disks - conf->max_degraded;
2727 /* First compute the information on this sector */
2730 * Compute the chunk number and the sector offset inside the chunk
2732 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2733 chunk_number = r_sector;
2736 * Compute the stripe number
2738 stripe = chunk_number;
2739 *dd_idx = sector_div(stripe, data_disks);
2742 * Select the parity disk based on the user selected algorithm.
2744 pd_idx = qd_idx = -1;
2745 switch(conf->level) {
2747 pd_idx = data_disks;
2750 switch (algorithm) {
2751 case ALGORITHM_LEFT_ASYMMETRIC:
2752 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2753 if (*dd_idx >= pd_idx)
2756 case ALGORITHM_RIGHT_ASYMMETRIC:
2757 pd_idx = sector_div(stripe2, raid_disks);
2758 if (*dd_idx >= pd_idx)
2761 case ALGORITHM_LEFT_SYMMETRIC:
2762 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2763 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2765 case ALGORITHM_RIGHT_SYMMETRIC:
2766 pd_idx = sector_div(stripe2, raid_disks);
2767 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2769 case ALGORITHM_PARITY_0:
2773 case ALGORITHM_PARITY_N:
2774 pd_idx = data_disks;
2782 switch (algorithm) {
2783 case ALGORITHM_LEFT_ASYMMETRIC:
2784 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2785 qd_idx = pd_idx + 1;
2786 if (pd_idx == raid_disks-1) {
2787 (*dd_idx)++; /* Q D D D P */
2789 } else if (*dd_idx >= pd_idx)
2790 (*dd_idx) += 2; /* D D P Q D */
2792 case ALGORITHM_RIGHT_ASYMMETRIC:
2793 pd_idx = sector_div(stripe2, raid_disks);
2794 qd_idx = pd_idx + 1;
2795 if (pd_idx == raid_disks-1) {
2796 (*dd_idx)++; /* Q D D D P */
2798 } else if (*dd_idx >= pd_idx)
2799 (*dd_idx) += 2; /* D D P Q D */
2801 case ALGORITHM_LEFT_SYMMETRIC:
2802 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2803 qd_idx = (pd_idx + 1) % raid_disks;
2804 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2806 case ALGORITHM_RIGHT_SYMMETRIC:
2807 pd_idx = sector_div(stripe2, raid_disks);
2808 qd_idx = (pd_idx + 1) % raid_disks;
2809 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2812 case ALGORITHM_PARITY_0:
2817 case ALGORITHM_PARITY_N:
2818 pd_idx = data_disks;
2819 qd_idx = data_disks + 1;
2822 case ALGORITHM_ROTATING_ZERO_RESTART:
2823 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2824 * of blocks for computing Q is different.
2826 pd_idx = sector_div(stripe2, raid_disks);
2827 qd_idx = pd_idx + 1;
2828 if (pd_idx == raid_disks-1) {
2829 (*dd_idx)++; /* Q D D D P */
2831 } else if (*dd_idx >= pd_idx)
2832 (*dd_idx) += 2; /* D D P Q D */
2836 case ALGORITHM_ROTATING_N_RESTART:
2837 /* Same a left_asymmetric, by first stripe is
2838 * D D D P Q rather than
2842 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2843 qd_idx = pd_idx + 1;
2844 if (pd_idx == raid_disks-1) {
2845 (*dd_idx)++; /* Q D D D P */
2847 } else if (*dd_idx >= pd_idx)
2848 (*dd_idx) += 2; /* D D P Q D */
2852 case ALGORITHM_ROTATING_N_CONTINUE:
2853 /* Same as left_symmetric but Q is before P */
2854 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2855 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2856 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2860 case ALGORITHM_LEFT_ASYMMETRIC_6:
2861 /* RAID5 left_asymmetric, with Q on last device */
2862 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2863 if (*dd_idx >= pd_idx)
2865 qd_idx = raid_disks - 1;
2868 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2869 pd_idx = sector_div(stripe2, raid_disks-1);
2870 if (*dd_idx >= pd_idx)
2872 qd_idx = raid_disks - 1;
2875 case ALGORITHM_LEFT_SYMMETRIC_6:
2876 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2877 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2878 qd_idx = raid_disks - 1;
2881 case ALGORITHM_RIGHT_SYMMETRIC_6:
2882 pd_idx = sector_div(stripe2, raid_disks-1);
2883 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2884 qd_idx = raid_disks - 1;
2887 case ALGORITHM_PARITY_0_6:
2890 qd_idx = raid_disks - 1;
2900 sh->pd_idx = pd_idx;
2901 sh->qd_idx = qd_idx;
2902 sh->ddf_layout = ddf_layout;
2905 * Finally, compute the new sector number
2907 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2911 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2913 struct r5conf *conf = sh->raid_conf;
2914 int raid_disks = sh->disks;
2915 int data_disks = raid_disks - conf->max_degraded;
2916 sector_t new_sector = sh->sector, check;
2917 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2918 : conf->chunk_sectors;
2919 int algorithm = previous ? conf->prev_algo
2923 sector_t chunk_number;
2924 int dummy1, dd_idx = i;
2926 struct stripe_head sh2;
2928 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2929 stripe = new_sector;
2931 if (i == sh->pd_idx)
2933 switch(conf->level) {
2936 switch (algorithm) {
2937 case ALGORITHM_LEFT_ASYMMETRIC:
2938 case ALGORITHM_RIGHT_ASYMMETRIC:
2942 case ALGORITHM_LEFT_SYMMETRIC:
2943 case ALGORITHM_RIGHT_SYMMETRIC:
2946 i -= (sh->pd_idx + 1);
2948 case ALGORITHM_PARITY_0:
2951 case ALGORITHM_PARITY_N:
2958 if (i == sh->qd_idx)
2959 return 0; /* It is the Q disk */
2960 switch (algorithm) {
2961 case ALGORITHM_LEFT_ASYMMETRIC:
2962 case ALGORITHM_RIGHT_ASYMMETRIC:
2963 case ALGORITHM_ROTATING_ZERO_RESTART:
2964 case ALGORITHM_ROTATING_N_RESTART:
2965 if (sh->pd_idx == raid_disks-1)
2966 i--; /* Q D D D P */
2967 else if (i > sh->pd_idx)
2968 i -= 2; /* D D P Q D */
2970 case ALGORITHM_LEFT_SYMMETRIC:
2971 case ALGORITHM_RIGHT_SYMMETRIC:
2972 if (sh->pd_idx == raid_disks-1)
2973 i--; /* Q D D D P */
2978 i -= (sh->pd_idx + 2);
2981 case ALGORITHM_PARITY_0:
2984 case ALGORITHM_PARITY_N:
2986 case ALGORITHM_ROTATING_N_CONTINUE:
2987 /* Like left_symmetric, but P is before Q */
2988 if (sh->pd_idx == 0)
2989 i--; /* P D D D Q */
2994 i -= (sh->pd_idx + 1);
2997 case ALGORITHM_LEFT_ASYMMETRIC_6:
2998 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3002 case ALGORITHM_LEFT_SYMMETRIC_6:
3003 case ALGORITHM_RIGHT_SYMMETRIC_6:
3005 i += data_disks + 1;
3006 i -= (sh->pd_idx + 1);
3008 case ALGORITHM_PARITY_0_6:
3017 chunk_number = stripe * data_disks + i;
3018 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3020 check = raid5_compute_sector(conf, r_sector,
3021 previous, &dummy1, &sh2);
3022 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3023 || sh2.qd_idx != sh->qd_idx) {
3024 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3025 mdname(conf->mddev));
3032 * There are cases where we want handle_stripe_dirtying() and
3033 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3035 * This function checks whether we want to delay the towrite. Specifically,
3036 * we delay the towrite when:
3038 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3039 * stripe has data in journal (for other devices).
3041 * In this case, when reading data for the non-overwrite dev, it is
3042 * necessary to handle complex rmw of write back cache (prexor with
3043 * orig_page, and xor with page). To keep read path simple, we would
3044 * like to flush data in journal to RAID disks first, so complex rmw
3045 * is handled in the write patch (handle_stripe_dirtying).
3047 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3049 * It is important to be able to flush all stripes in raid5-cache.
3050 * Therefore, we need reserve some space on the journal device for
3051 * these flushes. If flush operation includes pending writes to the
3052 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3053 * for the flush out. If we exclude these pending writes from flush
3054 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3055 * Therefore, excluding pending writes in these cases enables more
3056 * efficient use of the journal device.
3058 * Note: To make sure the stripe makes progress, we only delay
3059 * towrite for stripes with data already in journal (injournal > 0).
3060 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3061 * no_space_stripes list.
3063 * 3. during journal failure
3064 * In journal failure, we try to flush all cached data to raid disks
3065 * based on data in stripe cache. The array is read-only to upper
3066 * layers, so we would skip all pending writes.
3069 static inline bool delay_towrite(struct r5conf *conf,
3071 struct stripe_head_state *s)
3074 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3075 !test_bit(R5_Insync, &dev->flags) && s->injournal)
3078 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3082 if (s->log_failed && s->injournal)
3088 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3089 int rcw, int expand)
3091 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3092 struct r5conf *conf = sh->raid_conf;
3093 int level = conf->level;
3097 * In some cases, handle_stripe_dirtying initially decided to
3098 * run rmw and allocates extra page for prexor. However, rcw is
3099 * cheaper later on. We need to free the extra page now,
3100 * because we won't be able to do that in ops_complete_prexor().
3102 r5c_release_extra_page(sh);
3104 for (i = disks; i--; ) {
3105 struct r5dev *dev = &sh->dev[i];
3107 if (dev->towrite && !delay_towrite(conf, dev, s)) {
3108 set_bit(R5_LOCKED, &dev->flags);
3109 set_bit(R5_Wantdrain, &dev->flags);
3111 clear_bit(R5_UPTODATE, &dev->flags);
3113 } else if (test_bit(R5_InJournal, &dev->flags)) {
3114 set_bit(R5_LOCKED, &dev->flags);
3118 /* if we are not expanding this is a proper write request, and
3119 * there will be bios with new data to be drained into the
3124 /* False alarm, nothing to do */
3126 sh->reconstruct_state = reconstruct_state_drain_run;
3127 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3129 sh->reconstruct_state = reconstruct_state_run;
3131 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3133 if (s->locked + conf->max_degraded == disks)
3134 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3135 atomic_inc(&conf->pending_full_writes);
3137 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3138 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3139 BUG_ON(level == 6 &&
3140 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3141 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3143 for (i = disks; i--; ) {
3144 struct r5dev *dev = &sh->dev[i];
3145 if (i == pd_idx || i == qd_idx)
3149 (test_bit(R5_UPTODATE, &dev->flags) ||
3150 test_bit(R5_Wantcompute, &dev->flags))) {
3151 set_bit(R5_Wantdrain, &dev->flags);
3152 set_bit(R5_LOCKED, &dev->flags);
3153 clear_bit(R5_UPTODATE, &dev->flags);
3155 } else if (test_bit(R5_InJournal, &dev->flags)) {
3156 set_bit(R5_LOCKED, &dev->flags);
3161 /* False alarm - nothing to do */
3163 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3164 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3165 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3166 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3169 /* keep the parity disk(s) locked while asynchronous operations
3172 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3173 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3177 int qd_idx = sh->qd_idx;
3178 struct r5dev *dev = &sh->dev[qd_idx];
3180 set_bit(R5_LOCKED, &dev->flags);
3181 clear_bit(R5_UPTODATE, &dev->flags);
3185 if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3186 test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3187 !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3188 test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3189 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3191 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3192 __func__, (unsigned long long)sh->sector,
3193 s->locked, s->ops_request);
3197 * Each stripe/dev can have one or more bion attached.
3198 * toread/towrite point to the first in a chain.
3199 * The bi_next chain must be in order.
3201 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3202 int forwrite, int previous)
3205 struct r5conf *conf = sh->raid_conf;
3208 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3209 (unsigned long long)bi->bi_iter.bi_sector,
3210 (unsigned long long)sh->sector);
3212 spin_lock_irq(&sh->stripe_lock);
3213 sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3214 /* Don't allow new IO added to stripes in batch list */
3218 bip = &sh->dev[dd_idx].towrite;
3222 bip = &sh->dev[dd_idx].toread;
3223 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3224 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3226 bip = & (*bip)->bi_next;
3228 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3231 if (forwrite && raid5_has_ppl(conf)) {
3233 * With PPL only writes to consecutive data chunks within a
3234 * stripe are allowed because for a single stripe_head we can
3235 * only have one PPL entry at a time, which describes one data
3236 * range. Not really an overlap, but wait_for_overlap can be
3237 * used to handle this.
3245 for (i = 0; i < sh->disks; i++) {
3246 if (i != sh->pd_idx &&
3247 (i == dd_idx || sh->dev[i].towrite)) {
3248 sector = sh->dev[i].sector;
3249 if (count == 0 || sector < first)
3257 if (first + conf->chunk_sectors * (count - 1) != last)
3261 if (!forwrite || previous)
3262 clear_bit(STRIPE_BATCH_READY, &sh->state);
3264 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3268 bio_inc_remaining(bi);
3269 md_write_inc(conf->mddev, bi);
3272 /* check if page is covered */
3273 sector_t sector = sh->dev[dd_idx].sector;
3274 for (bi=sh->dev[dd_idx].towrite;
3275 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3276 bi && bi->bi_iter.bi_sector <= sector;
3277 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3278 if (bio_end_sector(bi) >= sector)
3279 sector = bio_end_sector(bi);
3281 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3282 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3283 sh->overwrite_disks++;
3286 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3287 (unsigned long long)(*bip)->bi_iter.bi_sector,
3288 (unsigned long long)sh->sector, dd_idx);
3290 if (conf->mddev->bitmap && firstwrite) {
3291 /* Cannot hold spinlock over bitmap_startwrite,
3292 * but must ensure this isn't added to a batch until
3293 * we have added to the bitmap and set bm_seq.
3294 * So set STRIPE_BITMAP_PENDING to prevent
3296 * If multiple add_stripe_bio() calls race here they
3297 * much all set STRIPE_BITMAP_PENDING. So only the first one
3298 * to complete "bitmap_startwrite" gets to set
3299 * STRIPE_BIT_DELAY. This is important as once a stripe
3300 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3303 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3304 spin_unlock_irq(&sh->stripe_lock);
3305 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3307 spin_lock_irq(&sh->stripe_lock);
3308 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3309 if (!sh->batch_head) {
3310 sh->bm_seq = conf->seq_flush+1;
3311 set_bit(STRIPE_BIT_DELAY, &sh->state);
3314 spin_unlock_irq(&sh->stripe_lock);
3316 if (stripe_can_batch(sh))
3317 stripe_add_to_batch_list(conf, sh);
3321 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3322 spin_unlock_irq(&sh->stripe_lock);
3326 static void end_reshape(struct r5conf *conf);
3328 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3329 struct stripe_head *sh)
3331 int sectors_per_chunk =
3332 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3334 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3335 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3337 raid5_compute_sector(conf,
3338 stripe * (disks - conf->max_degraded)
3339 *sectors_per_chunk + chunk_offset,
3345 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3346 struct stripe_head_state *s, int disks)
3349 BUG_ON(sh->batch_head);
3350 for (i = disks; i--; ) {
3354 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3355 struct md_rdev *rdev;
3357 rdev = rcu_dereference(conf->disks[i].rdev);
3358 if (rdev && test_bit(In_sync, &rdev->flags) &&
3359 !test_bit(Faulty, &rdev->flags))
3360 atomic_inc(&rdev->nr_pending);
3365 if (!rdev_set_badblocks(
3369 md_error(conf->mddev, rdev);
3370 rdev_dec_pending(rdev, conf->mddev);
3373 spin_lock_irq(&sh->stripe_lock);
3374 /* fail all writes first */
3375 bi = sh->dev[i].towrite;
3376 sh->dev[i].towrite = NULL;
3377 sh->overwrite_disks = 0;
3378 spin_unlock_irq(&sh->stripe_lock);
3382 log_stripe_write_finished(sh);
3384 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3385 wake_up(&conf->wait_for_overlap);
3387 while (bi && bi->bi_iter.bi_sector <
3388 sh->dev[i].sector + STRIPE_SECTORS) {
3389 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3391 md_write_end(conf->mddev);
3396 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3397 STRIPE_SECTORS, 0, 0);
3399 /* and fail all 'written' */
3400 bi = sh->dev[i].written;
3401 sh->dev[i].written = NULL;
3402 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3403 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3404 sh->dev[i].page = sh->dev[i].orig_page;
3407 if (bi) bitmap_end = 1;
3408 while (bi && bi->bi_iter.bi_sector <
3409 sh->dev[i].sector + STRIPE_SECTORS) {
3410 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3412 md_write_end(conf->mddev);
3417 /* fail any reads if this device is non-operational and
3418 * the data has not reached the cache yet.
3420 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3421 s->failed > conf->max_degraded &&
3422 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3423 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3424 spin_lock_irq(&sh->stripe_lock);
3425 bi = sh->dev[i].toread;
3426 sh->dev[i].toread = NULL;
3427 spin_unlock_irq(&sh->stripe_lock);
3428 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3429 wake_up(&conf->wait_for_overlap);
3432 while (bi && bi->bi_iter.bi_sector <
3433 sh->dev[i].sector + STRIPE_SECTORS) {
3434 struct bio *nextbi =
3435 r5_next_bio(bi, sh->dev[i].sector);
3442 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3443 STRIPE_SECTORS, 0, 0);
3444 /* If we were in the middle of a write the parity block might
3445 * still be locked - so just clear all R5_LOCKED flags
3447 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3452 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3453 if (atomic_dec_and_test(&conf->pending_full_writes))
3454 md_wakeup_thread(conf->mddev->thread);
3458 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3459 struct stripe_head_state *s)
3464 BUG_ON(sh->batch_head);
3465 clear_bit(STRIPE_SYNCING, &sh->state);
3466 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3467 wake_up(&conf->wait_for_overlap);
3470 /* There is nothing more to do for sync/check/repair.
3471 * Don't even need to abort as that is handled elsewhere
3472 * if needed, and not always wanted e.g. if there is a known
3474 * For recover/replace we need to record a bad block on all
3475 * non-sync devices, or abort the recovery
3477 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3478 /* During recovery devices cannot be removed, so
3479 * locking and refcounting of rdevs is not needed
3482 for (i = 0; i < conf->raid_disks; i++) {
3483 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3485 && !test_bit(Faulty, &rdev->flags)
3486 && !test_bit(In_sync, &rdev->flags)
3487 && !rdev_set_badblocks(rdev, sh->sector,
3490 rdev = rcu_dereference(conf->disks[i].replacement);
3492 && !test_bit(Faulty, &rdev->flags)
3493 && !test_bit(In_sync, &rdev->flags)
3494 && !rdev_set_badblocks(rdev, sh->sector,
3500 conf->recovery_disabled =
3501 conf->mddev->recovery_disabled;
3503 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3506 static int want_replace(struct stripe_head *sh, int disk_idx)
3508 struct md_rdev *rdev;
3512 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3514 && !test_bit(Faulty, &rdev->flags)
3515 && !test_bit(In_sync, &rdev->flags)
3516 && (rdev->recovery_offset <= sh->sector
3517 || rdev->mddev->recovery_cp <= sh->sector))
3523 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3524 int disk_idx, int disks)
3526 struct r5dev *dev = &sh->dev[disk_idx];
3527 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3528 &sh->dev[s->failed_num[1]] };
3532 if (test_bit(R5_LOCKED, &dev->flags) ||
3533 test_bit(R5_UPTODATE, &dev->flags))
3534 /* No point reading this as we already have it or have
3535 * decided to get it.
3540 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3541 /* We need this block to directly satisfy a request */
3544 if (s->syncing || s->expanding ||
3545 (s->replacing && want_replace(sh, disk_idx)))
3546 /* When syncing, or expanding we read everything.
3547 * When replacing, we need the replaced block.
3551 if ((s->failed >= 1 && fdev[0]->toread) ||
3552 (s->failed >= 2 && fdev[1]->toread))
3553 /* If we want to read from a failed device, then
3554 * we need to actually read every other device.
3558 /* Sometimes neither read-modify-write nor reconstruct-write
3559 * cycles can work. In those cases we read every block we
3560 * can. Then the parity-update is certain to have enough to
3562 * This can only be a problem when we need to write something,
3563 * and some device has failed. If either of those tests
3564 * fail we need look no further.
3566 if (!s->failed || !s->to_write)
3569 if (test_bit(R5_Insync, &dev->flags) &&
3570 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3571 /* Pre-reads at not permitted until after short delay
3572 * to gather multiple requests. However if this
3573 * device is no Insync, the block could only be computed
3574 * and there is no need to delay that.
3578 for (i = 0; i < s->failed && i < 2; i++) {
3579 if (fdev[i]->towrite &&
3580 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3581 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3582 /* If we have a partial write to a failed
3583 * device, then we will need to reconstruct
3584 * the content of that device, so all other
3585 * devices must be read.
3590 /* If we are forced to do a reconstruct-write, either because
3591 * the current RAID6 implementation only supports that, or
3592 * because parity cannot be trusted and we are currently
3593 * recovering it, there is extra need to be careful.
3594 * If one of the devices that we would need to read, because
3595 * it is not being overwritten (and maybe not written at all)
3596 * is missing/faulty, then we need to read everything we can.
3598 if (sh->raid_conf->level != 6 &&
3599 sh->raid_conf->rmw_level != PARITY_DISABLE_RMW &&
3600 sh->sector < sh->raid_conf->mddev->recovery_cp)
3601 /* reconstruct-write isn't being forced */
3603 for (i = 0; i < s->failed && i < 2; i++) {
3604 if (s->failed_num[i] != sh->pd_idx &&
3605 s->failed_num[i] != sh->qd_idx &&
3606 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3607 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3614 /* fetch_block - checks the given member device to see if its data needs
3615 * to be read or computed to satisfy a request.
3617 * Returns 1 when no more member devices need to be checked, otherwise returns
3618 * 0 to tell the loop in handle_stripe_fill to continue
3620 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3621 int disk_idx, int disks)
3623 struct r5dev *dev = &sh->dev[disk_idx];
3625 /* is the data in this block needed, and can we get it? */
3626 if (need_this_block(sh, s, disk_idx, disks)) {
3627 /* we would like to get this block, possibly by computing it,
3628 * otherwise read it if the backing disk is insync
3630 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3631 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3632 BUG_ON(sh->batch_head);
3635 * In the raid6 case if the only non-uptodate disk is P
3636 * then we already trusted P to compute the other failed
3637 * drives. It is safe to compute rather than re-read P.
3638 * In other cases we only compute blocks from failed
3639 * devices, otherwise check/repair might fail to detect
3640 * a real inconsistency.
3643 if ((s->uptodate == disks - 1) &&
3644 ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3645 (s->failed && (disk_idx == s->failed_num[0] ||
3646 disk_idx == s->failed_num[1])))) {
3647 /* have disk failed, and we're requested to fetch it;
3650 pr_debug("Computing stripe %llu block %d\n",
3651 (unsigned long long)sh->sector, disk_idx);
3652 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3653 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3654 set_bit(R5_Wantcompute, &dev->flags);
3655 sh->ops.target = disk_idx;
3656 sh->ops.target2 = -1; /* no 2nd target */
3658 /* Careful: from this point on 'uptodate' is in the eye
3659 * of raid_run_ops which services 'compute' operations
3660 * before writes. R5_Wantcompute flags a block that will
3661 * be R5_UPTODATE by the time it is needed for a
3662 * subsequent operation.
3666 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3667 /* Computing 2-failure is *very* expensive; only
3668 * do it if failed >= 2
3671 for (other = disks; other--; ) {
3672 if (other == disk_idx)
3674 if (!test_bit(R5_UPTODATE,
3675 &sh->dev[other].flags))
3679 pr_debug("Computing stripe %llu blocks %d,%d\n",
3680 (unsigned long long)sh->sector,
3682 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3683 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3684 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3685 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3686 sh->ops.target = disk_idx;
3687 sh->ops.target2 = other;
3691 } else if (test_bit(R5_Insync, &dev->flags)) {
3692 set_bit(R5_LOCKED, &dev->flags);
3693 set_bit(R5_Wantread, &dev->flags);
3695 pr_debug("Reading block %d (sync=%d)\n",
3696 disk_idx, s->syncing);
3704 * handle_stripe_fill - read or compute data to satisfy pending requests.
3706 static void handle_stripe_fill(struct stripe_head *sh,
3707 struct stripe_head_state *s,
3712 /* look for blocks to read/compute, skip this if a compute
3713 * is already in flight, or if the stripe contents are in the
3714 * midst of changing due to a write
3716 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3717 !sh->reconstruct_state) {
3720 * For degraded stripe with data in journal, do not handle
3721 * read requests yet, instead, flush the stripe to raid
3722 * disks first, this avoids handling complex rmw of write
3723 * back cache (prexor with orig_page, and then xor with
3724 * page) in the read path
3726 if (s->injournal && s->failed) {
3727 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3728 r5c_make_stripe_write_out(sh);
3732 for (i = disks; i--; )
3733 if (fetch_block(sh, s, i, disks))
3737 set_bit(STRIPE_HANDLE, &sh->state);
3740 static void break_stripe_batch_list(struct stripe_head *head_sh,
3741 unsigned long handle_flags);
3742 /* handle_stripe_clean_event
3743 * any written block on an uptodate or failed drive can be returned.
3744 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3745 * never LOCKED, so we don't need to test 'failed' directly.
3747 static void handle_stripe_clean_event(struct r5conf *conf,
3748 struct stripe_head *sh, int disks)
3752 int discard_pending = 0;
3753 struct stripe_head *head_sh = sh;
3754 bool do_endio = false;
3756 for (i = disks; i--; )
3757 if (sh->dev[i].written) {
3759 if (!test_bit(R5_LOCKED, &dev->flags) &&
3760 (test_bit(R5_UPTODATE, &dev->flags) ||
3761 test_bit(R5_Discard, &dev->flags) ||
3762 test_bit(R5_SkipCopy, &dev->flags))) {
3763 /* We can return any write requests */
3764 struct bio *wbi, *wbi2;
3765 pr_debug("Return write for disc %d\n", i);
3766 if (test_and_clear_bit(R5_Discard, &dev->flags))
3767 clear_bit(R5_UPTODATE, &dev->flags);
3768 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3769 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3774 dev->page = dev->orig_page;
3776 dev->written = NULL;
3777 while (wbi && wbi->bi_iter.bi_sector <
3778 dev->sector + STRIPE_SECTORS) {
3779 wbi2 = r5_next_bio(wbi, dev->sector);
3780 md_write_end(conf->mddev);
3784 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3786 !test_bit(STRIPE_DEGRADED, &sh->state),
3788 if (head_sh->batch_head) {
3789 sh = list_first_entry(&sh->batch_list,
3792 if (sh != head_sh) {
3799 } else if (test_bit(R5_Discard, &dev->flags))
3800 discard_pending = 1;
3803 log_stripe_write_finished(sh);
3805 if (!discard_pending &&
3806 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3808 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3809 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3810 if (sh->qd_idx >= 0) {
3811 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3812 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3814 /* now that discard is done we can proceed with any sync */
3815 clear_bit(STRIPE_DISCARD, &sh->state);
3817 * SCSI discard will change some bio fields and the stripe has
3818 * no updated data, so remove it from hash list and the stripe
3819 * will be reinitialized
3822 hash = sh->hash_lock_index;
3823 spin_lock_irq(conf->hash_locks + hash);
3825 spin_unlock_irq(conf->hash_locks + hash);
3826 if (head_sh->batch_head) {
3827 sh = list_first_entry(&sh->batch_list,
3828 struct stripe_head, batch_list);
3834 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3835 set_bit(STRIPE_HANDLE, &sh->state);
3839 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3840 if (atomic_dec_and_test(&conf->pending_full_writes))
3841 md_wakeup_thread(conf->mddev->thread);
3843 if (head_sh->batch_head && do_endio)
3844 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3848 * For RMW in write back cache, we need extra page in prexor to store the
3849 * old data. This page is stored in dev->orig_page.
3851 * This function checks whether we have data for prexor. The exact logic
3853 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3855 static inline bool uptodate_for_rmw(struct r5dev *dev)
3857 return (test_bit(R5_UPTODATE, &dev->flags)) &&
3858 (!test_bit(R5_InJournal, &dev->flags) ||
3859 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3862 static int handle_stripe_dirtying(struct r5conf *conf,
3863 struct stripe_head *sh,
3864 struct stripe_head_state *s,
3867 int rmw = 0, rcw = 0, i;
3868 sector_t recovery_cp = conf->mddev->recovery_cp;
3870 /* Check whether resync is now happening or should start.
3871 * If yes, then the array is dirty (after unclean shutdown or
3872 * initial creation), so parity in some stripes might be inconsistent.
3873 * In this case, we need to always do reconstruct-write, to ensure
3874 * that in case of drive failure or read-error correction, we
3875 * generate correct data from the parity.
3877 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3878 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3880 /* Calculate the real rcw later - for now make it
3881 * look like rcw is cheaper
3884 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3885 conf->rmw_level, (unsigned long long)recovery_cp,
3886 (unsigned long long)sh->sector);
3887 } else for (i = disks; i--; ) {
3888 /* would I have to read this buffer for read_modify_write */
3889 struct r5dev *dev = &sh->dev[i];
3890 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3891 i == sh->pd_idx || i == sh->qd_idx ||
3892 test_bit(R5_InJournal, &dev->flags)) &&
3893 !test_bit(R5_LOCKED, &dev->flags) &&
3894 !(uptodate_for_rmw(dev) ||
3895 test_bit(R5_Wantcompute, &dev->flags))) {
3896 if (test_bit(R5_Insync, &dev->flags))
3899 rmw += 2*disks; /* cannot read it */
3901 /* Would I have to read this buffer for reconstruct_write */
3902 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3903 i != sh->pd_idx && i != sh->qd_idx &&
3904 !test_bit(R5_LOCKED, &dev->flags) &&
3905 !(test_bit(R5_UPTODATE, &dev->flags) ||
3906 test_bit(R5_Wantcompute, &dev->flags))) {
3907 if (test_bit(R5_Insync, &dev->flags))
3914 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3915 (unsigned long long)sh->sector, sh->state, rmw, rcw);
3916 set_bit(STRIPE_HANDLE, &sh->state);
3917 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3918 /* prefer read-modify-write, but need to get some data */
3919 if (conf->mddev->queue)
3920 blk_add_trace_msg(conf->mddev->queue,
3921 "raid5 rmw %llu %d",
3922 (unsigned long long)sh->sector, rmw);
3923 for (i = disks; i--; ) {
3924 struct r5dev *dev = &sh->dev[i];
3925 if (test_bit(R5_InJournal, &dev->flags) &&
3926 dev->page == dev->orig_page &&
3927 !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3928 /* alloc page for prexor */
3929 struct page *p = alloc_page(GFP_NOIO);
3937 * alloc_page() failed, try use
3938 * disk_info->extra_page
3940 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3941 &conf->cache_state)) {
3942 r5c_use_extra_page(sh);
3946 /* extra_page in use, add to delayed_list */
3947 set_bit(STRIPE_DELAYED, &sh->state);
3948 s->waiting_extra_page = 1;
3953 for (i = disks; i--; ) {
3954 struct r5dev *dev = &sh->dev[i];
3955 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3956 i == sh->pd_idx || i == sh->qd_idx ||
3957 test_bit(R5_InJournal, &dev->flags)) &&
3958 !test_bit(R5_LOCKED, &dev->flags) &&
3959 !(uptodate_for_rmw(dev) ||
3960 test_bit(R5_Wantcompute, &dev->flags)) &&
3961 test_bit(R5_Insync, &dev->flags)) {
3962 if (test_bit(STRIPE_PREREAD_ACTIVE,
3964 pr_debug("Read_old block %d for r-m-w\n",
3966 set_bit(R5_LOCKED, &dev->flags);
3967 set_bit(R5_Wantread, &dev->flags);
3970 set_bit(STRIPE_DELAYED, &sh->state);
3971 set_bit(STRIPE_HANDLE, &sh->state);
3976 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3977 /* want reconstruct write, but need to get some data */
3980 for (i = disks; i--; ) {
3981 struct r5dev *dev = &sh->dev[i];
3982 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3983 i != sh->pd_idx && i != sh->qd_idx &&
3984 !test_bit(R5_LOCKED, &dev->flags) &&
3985 !(test_bit(R5_UPTODATE, &dev->flags) ||
3986 test_bit(R5_Wantcompute, &dev->flags))) {
3988 if (test_bit(R5_Insync, &dev->flags) &&
3989 test_bit(STRIPE_PREREAD_ACTIVE,
3991 pr_debug("Read_old block "
3992 "%d for Reconstruct\n", i);
3993 set_bit(R5_LOCKED, &dev->flags);
3994 set_bit(R5_Wantread, &dev->flags);
3998 set_bit(STRIPE_DELAYED, &sh->state);
3999 set_bit(STRIPE_HANDLE, &sh->state);
4003 if (rcw && conf->mddev->queue)
4004 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4005 (unsigned long long)sh->sector,
4006 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4009 if (rcw > disks && rmw > disks &&
4010 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4011 set_bit(STRIPE_DELAYED, &sh->state);
4013 /* now if nothing is locked, and if we have enough data,
4014 * we can start a write request
4016 /* since handle_stripe can be called at any time we need to handle the
4017 * case where a compute block operation has been submitted and then a
4018 * subsequent call wants to start a write request. raid_run_ops only
4019 * handles the case where compute block and reconstruct are requested
4020 * simultaneously. If this is not the case then new writes need to be
4021 * held off until the compute completes.
4023 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4024 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4025 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4026 schedule_reconstruction(sh, s, rcw == 0, 0);
4030 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4031 struct stripe_head_state *s, int disks)
4033 struct r5dev *dev = NULL;
4035 BUG_ON(sh->batch_head);
4036 set_bit(STRIPE_HANDLE, &sh->state);
4038 switch (sh->check_state) {
4039 case check_state_idle:
4040 /* start a new check operation if there are no failures */
4041 if (s->failed == 0) {
4042 BUG_ON(s->uptodate != disks);
4043 sh->check_state = check_state_run;
4044 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4045 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4049 dev = &sh->dev[s->failed_num[0]];
4051 case check_state_compute_result:
4052 sh->check_state = check_state_idle;
4054 dev = &sh->dev[sh->pd_idx];
4056 /* check that a write has not made the stripe insync */
4057 if (test_bit(STRIPE_INSYNC, &sh->state))
4060 /* either failed parity check, or recovery is happening */
4061 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4062 BUG_ON(s->uptodate != disks);
4064 set_bit(R5_LOCKED, &dev->flags);
4066 set_bit(R5_Wantwrite, &dev->flags);
4068 clear_bit(STRIPE_DEGRADED, &sh->state);
4069 set_bit(STRIPE_INSYNC, &sh->state);
4071 case check_state_run:
4072 break; /* we will be called again upon completion */
4073 case check_state_check_result:
4074 sh->check_state = check_state_idle;
4076 /* if a failure occurred during the check operation, leave
4077 * STRIPE_INSYNC not set and let the stripe be handled again
4082 /* handle a successful check operation, if parity is correct
4083 * we are done. Otherwise update the mismatch count and repair
4084 * parity if !MD_RECOVERY_CHECK
4086 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4087 /* parity is correct (on disc,
4088 * not in buffer any more)
4090 set_bit(STRIPE_INSYNC, &sh->state);
4092 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4093 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4094 /* don't try to repair!! */
4095 set_bit(STRIPE_INSYNC, &sh->state);
4096 pr_warn_ratelimited("%s: mismatch sector in range "
4097 "%llu-%llu\n", mdname(conf->mddev),
4098 (unsigned long long) sh->sector,
4099 (unsigned long long) sh->sector +
4102 sh->check_state = check_state_compute_run;
4103 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4104 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4105 set_bit(R5_Wantcompute,
4106 &sh->dev[sh->pd_idx].flags);
4107 sh->ops.target = sh->pd_idx;
4108 sh->ops.target2 = -1;
4113 case check_state_compute_run:
4116 pr_err("%s: unknown check_state: %d sector: %llu\n",
4117 __func__, sh->check_state,
4118 (unsigned long long) sh->sector);
4123 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4124 struct stripe_head_state *s,
4127 int pd_idx = sh->pd_idx;
4128 int qd_idx = sh->qd_idx;
4131 BUG_ON(sh->batch_head);
4132 set_bit(STRIPE_HANDLE, &sh->state);
4134 BUG_ON(s->failed > 2);
4136 /* Want to check and possibly repair P and Q.
4137 * However there could be one 'failed' device, in which
4138 * case we can only check one of them, possibly using the
4139 * other to generate missing data
4142 switch (sh->check_state) {
4143 case check_state_idle:
4144 /* start a new check operation if there are < 2 failures */
4145 if (s->failed == s->q_failed) {
4146 /* The only possible failed device holds Q, so it
4147 * makes sense to check P (If anything else were failed,
4148 * we would have used P to recreate it).
4150 sh->check_state = check_state_run;
4152 if (!s->q_failed && s->failed < 2) {
4153 /* Q is not failed, and we didn't use it to generate
4154 * anything, so it makes sense to check it
4156 if (sh->check_state == check_state_run)
4157 sh->check_state = check_state_run_pq;
4159 sh->check_state = check_state_run_q;
4162 /* discard potentially stale zero_sum_result */
4163 sh->ops.zero_sum_result = 0;
4165 if (sh->check_state == check_state_run) {
4166 /* async_xor_zero_sum destroys the contents of P */
4167 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4170 if (sh->check_state >= check_state_run &&
4171 sh->check_state <= check_state_run_pq) {
4172 /* async_syndrome_zero_sum preserves P and Q, so
4173 * no need to mark them !uptodate here
4175 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4179 /* we have 2-disk failure */
4180 BUG_ON(s->failed != 2);
4182 case check_state_compute_result:
4183 sh->check_state = check_state_idle;
4185 /* check that a write has not made the stripe insync */
4186 if (test_bit(STRIPE_INSYNC, &sh->state))
4189 /* now write out any block on a failed drive,
4190 * or P or Q if they were recomputed
4193 if (s->failed == 2) {
4194 dev = &sh->dev[s->failed_num[1]];
4196 set_bit(R5_LOCKED, &dev->flags);
4197 set_bit(R5_Wantwrite, &dev->flags);
4199 if (s->failed >= 1) {
4200 dev = &sh->dev[s->failed_num[0]];
4202 set_bit(R5_LOCKED, &dev->flags);
4203 set_bit(R5_Wantwrite, &dev->flags);
4205 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4206 dev = &sh->dev[pd_idx];
4208 set_bit(R5_LOCKED, &dev->flags);
4209 set_bit(R5_Wantwrite, &dev->flags);
4211 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4212 dev = &sh->dev[qd_idx];
4214 set_bit(R5_LOCKED, &dev->flags);
4215 set_bit(R5_Wantwrite, &dev->flags);
4217 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4218 "%s: disk%td not up to date\n",
4219 mdname(conf->mddev),
4220 dev - (struct r5dev *) &sh->dev)) {
4221 clear_bit(R5_LOCKED, &dev->flags);
4222 clear_bit(R5_Wantwrite, &dev->flags);
4225 clear_bit(STRIPE_DEGRADED, &sh->state);
4227 set_bit(STRIPE_INSYNC, &sh->state);
4229 case check_state_run:
4230 case check_state_run_q:
4231 case check_state_run_pq:
4232 break; /* we will be called again upon completion */
4233 case check_state_check_result:
4234 sh->check_state = check_state_idle;
4236 /* handle a successful check operation, if parity is correct
4237 * we are done. Otherwise update the mismatch count and repair
4238 * parity if !MD_RECOVERY_CHECK
4240 if (sh->ops.zero_sum_result == 0) {
4241 /* both parities are correct */
4243 set_bit(STRIPE_INSYNC, &sh->state);
4245 /* in contrast to the raid5 case we can validate
4246 * parity, but still have a failure to write
4249 sh->check_state = check_state_compute_result;
4250 /* Returning at this point means that we may go
4251 * off and bring p and/or q uptodate again so
4252 * we make sure to check zero_sum_result again
4253 * to verify if p or q need writeback
4257 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4258 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4259 /* don't try to repair!! */
4260 set_bit(STRIPE_INSYNC, &sh->state);
4261 pr_warn_ratelimited("%s: mismatch sector in range "
4262 "%llu-%llu\n", mdname(conf->mddev),
4263 (unsigned long long) sh->sector,
4264 (unsigned long long) sh->sector +
4267 int *target = &sh->ops.target;
4269 sh->ops.target = -1;
4270 sh->ops.target2 = -1;
4271 sh->check_state = check_state_compute_run;
4272 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4273 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4274 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4275 set_bit(R5_Wantcompute,
4276 &sh->dev[pd_idx].flags);
4278 target = &sh->ops.target2;
4281 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4282 set_bit(R5_Wantcompute,
4283 &sh->dev[qd_idx].flags);
4290 case check_state_compute_run:
4293 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4294 __func__, sh->check_state,
4295 (unsigned long long) sh->sector);
4300 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4304 /* We have read all the blocks in this stripe and now we need to
4305 * copy some of them into a target stripe for expand.
4307 struct dma_async_tx_descriptor *tx = NULL;
4308 BUG_ON(sh->batch_head);
4309 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4310 for (i = 0; i < sh->disks; i++)
4311 if (i != sh->pd_idx && i != sh->qd_idx) {
4313 struct stripe_head *sh2;
4314 struct async_submit_ctl submit;
4316 sector_t bn = raid5_compute_blocknr(sh, i, 1);
4317 sector_t s = raid5_compute_sector(conf, bn, 0,
4319 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4321 /* so far only the early blocks of this stripe
4322 * have been requested. When later blocks
4323 * get requested, we will try again
4326 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4327 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4328 /* must have already done this block */
4329 raid5_release_stripe(sh2);
4333 /* place all the copies on one channel */
4334 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4335 tx = async_memcpy(sh2->dev[dd_idx].page,
4336 sh->dev[i].page, 0, 0, STRIPE_SIZE,
4339 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4340 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4341 for (j = 0; j < conf->raid_disks; j++)
4342 if (j != sh2->pd_idx &&
4344 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4346 if (j == conf->raid_disks) {
4347 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4348 set_bit(STRIPE_HANDLE, &sh2->state);
4350 raid5_release_stripe(sh2);
4353 /* done submitting copies, wait for them to complete */
4354 async_tx_quiesce(&tx);
4358 * handle_stripe - do things to a stripe.
4360 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4361 * state of various bits to see what needs to be done.
4363 * return some read requests which now have data
4364 * return some write requests which are safely on storage
4365 * schedule a read on some buffers
4366 * schedule a write of some buffers
4367 * return confirmation of parity correctness
4371 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4373 struct r5conf *conf = sh->raid_conf;
4374 int disks = sh->disks;
4377 int do_recovery = 0;
4379 memset(s, 0, sizeof(*s));
4381 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4382 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4383 s->failed_num[0] = -1;
4384 s->failed_num[1] = -1;
4385 s->log_failed = r5l_log_disk_error(conf);
4387 /* Now to look around and see what can be done */
4389 for (i=disks; i--; ) {
4390 struct md_rdev *rdev;
4397 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4399 dev->toread, dev->towrite, dev->written);
4400 /* maybe we can reply to a read
4402 * new wantfill requests are only permitted while
4403 * ops_complete_biofill is guaranteed to be inactive
4405 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4406 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4407 set_bit(R5_Wantfill, &dev->flags);
4409 /* now count some things */
4410 if (test_bit(R5_LOCKED, &dev->flags))
4412 if (test_bit(R5_UPTODATE, &dev->flags))
4414 if (test_bit(R5_Wantcompute, &dev->flags)) {
4416 BUG_ON(s->compute > 2);
4419 if (test_bit(R5_Wantfill, &dev->flags))
4421 else if (dev->toread)
4425 if (!test_bit(R5_OVERWRITE, &dev->flags))
4430 /* Prefer to use the replacement for reads, but only
4431 * if it is recovered enough and has no bad blocks.
4433 rdev = rcu_dereference(conf->disks[i].replacement);
4434 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4435 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4436 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4437 &first_bad, &bad_sectors))
4438 set_bit(R5_ReadRepl, &dev->flags);
4440 if (rdev && !test_bit(Faulty, &rdev->flags))
4441 set_bit(R5_NeedReplace, &dev->flags);
4443 clear_bit(R5_NeedReplace, &dev->flags);
4444 rdev = rcu_dereference(conf->disks[i].rdev);
4445 clear_bit(R5_ReadRepl, &dev->flags);
4447 if (rdev && test_bit(Faulty, &rdev->flags))
4450 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4451 &first_bad, &bad_sectors);
4452 if (s->blocked_rdev == NULL
4453 && (test_bit(Blocked, &rdev->flags)
4456 set_bit(BlockedBadBlocks,
4458 s->blocked_rdev = rdev;
4459 atomic_inc(&rdev->nr_pending);
4462 clear_bit(R5_Insync, &dev->flags);
4466 /* also not in-sync */
4467 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4468 test_bit(R5_UPTODATE, &dev->flags)) {
4469 /* treat as in-sync, but with a read error
4470 * which we can now try to correct
4472 set_bit(R5_Insync, &dev->flags);
4473 set_bit(R5_ReadError, &dev->flags);
4475 } else if (test_bit(In_sync, &rdev->flags))
4476 set_bit(R5_Insync, &dev->flags);
4477 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4478 /* in sync if before recovery_offset */
4479 set_bit(R5_Insync, &dev->flags);
4480 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4481 test_bit(R5_Expanded, &dev->flags))
4482 /* If we've reshaped into here, we assume it is Insync.
4483 * We will shortly update recovery_offset to make
4486 set_bit(R5_Insync, &dev->flags);
4488 if (test_bit(R5_WriteError, &dev->flags)) {
4489 /* This flag does not apply to '.replacement'
4490 * only to .rdev, so make sure to check that*/
4491 struct md_rdev *rdev2 = rcu_dereference(
4492 conf->disks[i].rdev);
4494 clear_bit(R5_Insync, &dev->flags);
4495 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4496 s->handle_bad_blocks = 1;
4497 atomic_inc(&rdev2->nr_pending);
4499 clear_bit(R5_WriteError, &dev->flags);
4501 if (test_bit(R5_MadeGood, &dev->flags)) {
4502 /* This flag does not apply to '.replacement'
4503 * only to .rdev, so make sure to check that*/
4504 struct md_rdev *rdev2 = rcu_dereference(
4505 conf->disks[i].rdev);
4506 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4507 s->handle_bad_blocks = 1;
4508 atomic_inc(&rdev2->nr_pending);
4510 clear_bit(R5_MadeGood, &dev->flags);
4512 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4513 struct md_rdev *rdev2 = rcu_dereference(
4514 conf->disks[i].replacement);
4515 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4516 s->handle_bad_blocks = 1;
4517 atomic_inc(&rdev2->nr_pending);
4519 clear_bit(R5_MadeGoodRepl, &dev->flags);
4521 if (!test_bit(R5_Insync, &dev->flags)) {
4522 /* The ReadError flag will just be confusing now */
4523 clear_bit(R5_ReadError, &dev->flags);
4524 clear_bit(R5_ReWrite, &dev->flags);
4526 if (test_bit(R5_ReadError, &dev->flags))
4527 clear_bit(R5_Insync, &dev->flags);
4528 if (!test_bit(R5_Insync, &dev->flags)) {
4530 s->failed_num[s->failed] = i;
4532 if (rdev && !test_bit(Faulty, &rdev->flags))
4535 rdev = rcu_dereference(
4536 conf->disks[i].replacement);
4537 if (rdev && !test_bit(Faulty, &rdev->flags))
4542 if (test_bit(R5_InJournal, &dev->flags))
4544 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4547 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4548 /* If there is a failed device being replaced,
4549 * we must be recovering.
4550 * else if we are after recovery_cp, we must be syncing
4551 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4552 * else we can only be replacing
4553 * sync and recovery both need to read all devices, and so
4554 * use the same flag.
4557 sh->sector >= conf->mddev->recovery_cp ||
4558 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4566 static int clear_batch_ready(struct stripe_head *sh)
4568 /* Return '1' if this is a member of batch, or
4569 * '0' if it is a lone stripe or a head which can now be
4572 struct stripe_head *tmp;
4573 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4574 return (sh->batch_head && sh->batch_head != sh);
4575 spin_lock(&sh->stripe_lock);
4576 if (!sh->batch_head) {
4577 spin_unlock(&sh->stripe_lock);
4582 * this stripe could be added to a batch list before we check
4583 * BATCH_READY, skips it
4585 if (sh->batch_head != sh) {
4586 spin_unlock(&sh->stripe_lock);
4589 spin_lock(&sh->batch_lock);
4590 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4591 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4592 spin_unlock(&sh->batch_lock);
4593 spin_unlock(&sh->stripe_lock);
4596 * BATCH_READY is cleared, no new stripes can be added.
4597 * batch_list can be accessed without lock
4602 static void break_stripe_batch_list(struct stripe_head *head_sh,
4603 unsigned long handle_flags)
4605 struct stripe_head *sh, *next;
4609 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4611 list_del_init(&sh->batch_list);
4613 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4614 (1 << STRIPE_SYNCING) |
4615 (1 << STRIPE_REPLACED) |
4616 (1 << STRIPE_DELAYED) |
4617 (1 << STRIPE_BIT_DELAY) |
4618 (1 << STRIPE_FULL_WRITE) |
4619 (1 << STRIPE_BIOFILL_RUN) |
4620 (1 << STRIPE_COMPUTE_RUN) |
4621 (1 << STRIPE_OPS_REQ_PENDING) |
4622 (1 << STRIPE_DISCARD) |
4623 (1 << STRIPE_BATCH_READY) |
4624 (1 << STRIPE_BATCH_ERR) |
4625 (1 << STRIPE_BITMAP_PENDING)),
4626 "stripe state: %lx\n", sh->state);
4627 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4628 (1 << STRIPE_REPLACED)),
4629 "head stripe state: %lx\n", head_sh->state);
4631 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4632 (1 << STRIPE_PREREAD_ACTIVE) |
4633 (1 << STRIPE_DEGRADED) |
4634 (1 << STRIPE_ON_UNPLUG_LIST)),
4635 head_sh->state & (1 << STRIPE_INSYNC));
4637 sh->check_state = head_sh->check_state;
4638 sh->reconstruct_state = head_sh->reconstruct_state;
4639 spin_lock_irq(&sh->stripe_lock);
4640 sh->batch_head = NULL;
4641 spin_unlock_irq(&sh->stripe_lock);
4642 for (i = 0; i < sh->disks; i++) {
4643 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4645 sh->dev[i].flags = head_sh->dev[i].flags &
4646 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4648 if (handle_flags == 0 ||
4649 sh->state & handle_flags)
4650 set_bit(STRIPE_HANDLE, &sh->state);
4651 raid5_release_stripe(sh);
4653 spin_lock_irq(&head_sh->stripe_lock);
4654 head_sh->batch_head = NULL;
4655 spin_unlock_irq(&head_sh->stripe_lock);
4656 for (i = 0; i < head_sh->disks; i++)
4657 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4659 if (head_sh->state & handle_flags)
4660 set_bit(STRIPE_HANDLE, &head_sh->state);
4663 wake_up(&head_sh->raid_conf->wait_for_overlap);
4666 static void handle_stripe(struct stripe_head *sh)
4668 struct stripe_head_state s;
4669 struct r5conf *conf = sh->raid_conf;
4672 int disks = sh->disks;
4673 struct r5dev *pdev, *qdev;
4675 clear_bit(STRIPE_HANDLE, &sh->state);
4676 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4677 /* already being handled, ensure it gets handled
4678 * again when current action finishes */
4679 set_bit(STRIPE_HANDLE, &sh->state);
4683 if (clear_batch_ready(sh) ) {
4684 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4688 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4689 break_stripe_batch_list(sh, 0);
4691 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4692 spin_lock(&sh->stripe_lock);
4694 * Cannot process 'sync' concurrently with 'discard'.
4695 * Flush data in r5cache before 'sync'.
4697 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4698 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4699 !test_bit(STRIPE_DISCARD, &sh->state) &&
4700 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4701 set_bit(STRIPE_SYNCING, &sh->state);
4702 clear_bit(STRIPE_INSYNC, &sh->state);
4703 clear_bit(STRIPE_REPLACED, &sh->state);
4705 spin_unlock(&sh->stripe_lock);
4707 clear_bit(STRIPE_DELAYED, &sh->state);
4709 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4710 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4711 (unsigned long long)sh->sector, sh->state,
4712 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4713 sh->check_state, sh->reconstruct_state);
4715 analyse_stripe(sh, &s);
4717 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4720 if (s.handle_bad_blocks ||
4721 test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4722 set_bit(STRIPE_HANDLE, &sh->state);
4726 if (unlikely(s.blocked_rdev)) {
4727 if (s.syncing || s.expanding || s.expanded ||
4728 s.replacing || s.to_write || s.written) {
4729 set_bit(STRIPE_HANDLE, &sh->state);
4732 /* There is nothing for the blocked_rdev to block */
4733 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4734 s.blocked_rdev = NULL;
4737 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4738 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4739 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4742 pr_debug("locked=%d uptodate=%d to_read=%d"
4743 " to_write=%d failed=%d failed_num=%d,%d\n",
4744 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4745 s.failed_num[0], s.failed_num[1]);
4747 * check if the array has lost more than max_degraded devices and,
4748 * if so, some requests might need to be failed.
4750 * When journal device failed (log_failed), we will only process
4751 * the stripe if there is data need write to raid disks
4753 if (s.failed > conf->max_degraded ||
4754 (s.log_failed && s.injournal == 0)) {
4755 sh->check_state = 0;
4756 sh->reconstruct_state = 0;
4757 break_stripe_batch_list(sh, 0);
4758 if (s.to_read+s.to_write+s.written)
4759 handle_failed_stripe(conf, sh, &s, disks);
4760 if (s.syncing + s.replacing)
4761 handle_failed_sync(conf, sh, &s);
4764 /* Now we check to see if any write operations have recently
4768 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4770 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4771 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4772 sh->reconstruct_state = reconstruct_state_idle;
4774 /* All the 'written' buffers and the parity block are ready to
4775 * be written back to disk
4777 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4778 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4779 BUG_ON(sh->qd_idx >= 0 &&
4780 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4781 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4782 for (i = disks; i--; ) {
4783 struct r5dev *dev = &sh->dev[i];
4784 if (test_bit(R5_LOCKED, &dev->flags) &&
4785 (i == sh->pd_idx || i == sh->qd_idx ||
4786 dev->written || test_bit(R5_InJournal,
4788 pr_debug("Writing block %d\n", i);
4789 set_bit(R5_Wantwrite, &dev->flags);
4794 if (!test_bit(R5_Insync, &dev->flags) ||
4795 ((i == sh->pd_idx || i == sh->qd_idx) &&
4797 set_bit(STRIPE_INSYNC, &sh->state);
4800 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4801 s.dec_preread_active = 1;
4805 * might be able to return some write requests if the parity blocks
4806 * are safe, or on a failed drive
4808 pdev = &sh->dev[sh->pd_idx];
4809 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4810 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4811 qdev = &sh->dev[sh->qd_idx];
4812 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4813 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4817 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4818 && !test_bit(R5_LOCKED, &pdev->flags)
4819 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4820 test_bit(R5_Discard, &pdev->flags))))) &&
4821 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4822 && !test_bit(R5_LOCKED, &qdev->flags)
4823 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4824 test_bit(R5_Discard, &qdev->flags))))))
4825 handle_stripe_clean_event(conf, sh, disks);
4828 r5c_handle_cached_data_endio(conf, sh, disks);
4829 log_stripe_write_finished(sh);
4831 /* Now we might consider reading some blocks, either to check/generate
4832 * parity, or to satisfy requests
4833 * or to load a block that is being partially written.
4835 if (s.to_read || s.non_overwrite
4836 || (s.to_write && s.failed)
4837 || (s.syncing && (s.uptodate + s.compute < disks))
4840 handle_stripe_fill(sh, &s, disks);
4843 * When the stripe finishes full journal write cycle (write to journal
4844 * and raid disk), this is the clean up procedure so it is ready for
4847 r5c_finish_stripe_write_out(conf, sh, &s);
4850 * Now to consider new write requests, cache write back and what else,
4851 * if anything should be read. We do not handle new writes when:
4852 * 1/ A 'write' operation (copy+xor) is already in flight.
4853 * 2/ A 'check' operation is in flight, as it may clobber the parity
4855 * 3/ A r5c cache log write is in flight.
4858 if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4859 if (!r5c_is_writeback(conf->log)) {
4861 handle_stripe_dirtying(conf, sh, &s, disks);
4862 } else { /* write back cache */
4865 /* First, try handle writes in caching phase */
4867 ret = r5c_try_caching_write(conf, sh, &s,
4870 * If caching phase failed: ret == -EAGAIN
4872 * stripe under reclaim: !caching && injournal
4874 * fall back to handle_stripe_dirtying()
4876 if (ret == -EAGAIN ||
4877 /* stripe under reclaim: !caching && injournal */
4878 (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4880 ret = handle_stripe_dirtying(conf, sh, &s,
4888 /* maybe we need to check and possibly fix the parity for this stripe
4889 * Any reads will already have been scheduled, so we just see if enough
4890 * data is available. The parity check is held off while parity
4891 * dependent operations are in flight.
4893 if (sh->check_state ||
4894 (s.syncing && s.locked == 0 &&
4895 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4896 !test_bit(STRIPE_INSYNC, &sh->state))) {
4897 if (conf->level == 6)
4898 handle_parity_checks6(conf, sh, &s, disks);
4900 handle_parity_checks5(conf, sh, &s, disks);
4903 if ((s.replacing || s.syncing) && s.locked == 0
4904 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4905 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4906 /* Write out to replacement devices where possible */
4907 for (i = 0; i < conf->raid_disks; i++)
4908 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4909 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4910 set_bit(R5_WantReplace, &sh->dev[i].flags);
4911 set_bit(R5_LOCKED, &sh->dev[i].flags);
4915 set_bit(STRIPE_INSYNC, &sh->state);
4916 set_bit(STRIPE_REPLACED, &sh->state);
4918 if ((s.syncing || s.replacing) && s.locked == 0 &&
4919 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4920 test_bit(STRIPE_INSYNC, &sh->state)) {
4921 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4922 clear_bit(STRIPE_SYNCING, &sh->state);
4923 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4924 wake_up(&conf->wait_for_overlap);
4927 /* If the failed drives are just a ReadError, then we might need
4928 * to progress the repair/check process
4930 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4931 for (i = 0; i < s.failed; i++) {
4932 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4933 if (test_bit(R5_ReadError, &dev->flags)
4934 && !test_bit(R5_LOCKED, &dev->flags)
4935 && test_bit(R5_UPTODATE, &dev->flags)
4937 if (!test_bit(R5_ReWrite, &dev->flags)) {
4938 set_bit(R5_Wantwrite, &dev->flags);
4939 set_bit(R5_ReWrite, &dev->flags);
4940 set_bit(R5_LOCKED, &dev->flags);
4943 /* let's read it back */
4944 set_bit(R5_Wantread, &dev->flags);
4945 set_bit(R5_LOCKED, &dev->flags);
4951 /* Finish reconstruct operations initiated by the expansion process */
4952 if (sh->reconstruct_state == reconstruct_state_result) {
4953 struct stripe_head *sh_src
4954 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4955 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4956 /* sh cannot be written until sh_src has been read.
4957 * so arrange for sh to be delayed a little
4959 set_bit(STRIPE_DELAYED, &sh->state);
4960 set_bit(STRIPE_HANDLE, &sh->state);
4961 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4963 atomic_inc(&conf->preread_active_stripes);
4964 raid5_release_stripe(sh_src);
4968 raid5_release_stripe(sh_src);
4970 sh->reconstruct_state = reconstruct_state_idle;
4971 clear_bit(STRIPE_EXPANDING, &sh->state);
4972 for (i = conf->raid_disks; i--; ) {
4973 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4974 set_bit(R5_LOCKED, &sh->dev[i].flags);
4979 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4980 !sh->reconstruct_state) {
4981 /* Need to write out all blocks after computing parity */
4982 sh->disks = conf->raid_disks;
4983 stripe_set_idx(sh->sector, conf, 0, sh);
4984 schedule_reconstruction(sh, &s, 1, 1);
4985 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4986 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4987 atomic_dec(&conf->reshape_stripes);
4988 wake_up(&conf->wait_for_overlap);
4989 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4992 if (s.expanding && s.locked == 0 &&
4993 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4994 handle_stripe_expansion(conf, sh);
4997 /* wait for this device to become unblocked */
4998 if (unlikely(s.blocked_rdev)) {
4999 if (conf->mddev->external)
5000 md_wait_for_blocked_rdev(s.blocked_rdev,
5003 /* Internal metadata will immediately
5004 * be written by raid5d, so we don't
5005 * need to wait here.
5007 rdev_dec_pending(s.blocked_rdev,
5011 if (s.handle_bad_blocks)
5012 for (i = disks; i--; ) {
5013 struct md_rdev *rdev;
5014 struct r5dev *dev = &sh->dev[i];
5015 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5016 /* We own a safe reference to the rdev */
5017 rdev = conf->disks[i].rdev;
5018 if (!rdev_set_badblocks(rdev, sh->sector,
5020 md_error(conf->mddev, rdev);
5021 rdev_dec_pending(rdev, conf->mddev);
5023 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5024 rdev = conf->disks[i].rdev;
5025 rdev_clear_badblocks(rdev, sh->sector,
5027 rdev_dec_pending(rdev, conf->mddev);
5029 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5030 rdev = conf->disks[i].replacement;
5032 /* rdev have been moved down */
5033 rdev = conf->disks[i].rdev;
5034 rdev_clear_badblocks(rdev, sh->sector,
5036 rdev_dec_pending(rdev, conf->mddev);
5041 raid_run_ops(sh, s.ops_request);
5045 if (s.dec_preread_active) {
5046 /* We delay this until after ops_run_io so that if make_request
5047 * is waiting on a flush, it won't continue until the writes
5048 * have actually been submitted.
5050 atomic_dec(&conf->preread_active_stripes);
5051 if (atomic_read(&conf->preread_active_stripes) <
5053 md_wakeup_thread(conf->mddev->thread);
5056 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5059 static void raid5_activate_delayed(struct r5conf *conf)
5061 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5062 while (!list_empty(&conf->delayed_list)) {
5063 struct list_head *l = conf->delayed_list.next;
5064 struct stripe_head *sh;
5065 sh = list_entry(l, struct stripe_head, lru);
5067 clear_bit(STRIPE_DELAYED, &sh->state);
5068 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5069 atomic_inc(&conf->preread_active_stripes);
5070 list_add_tail(&sh->lru, &conf->hold_list);
5071 raid5_wakeup_stripe_thread(sh);
5076 static void activate_bit_delay(struct r5conf *conf,
5077 struct list_head *temp_inactive_list)
5079 /* device_lock is held */
5080 struct list_head head;
5081 list_add(&head, &conf->bitmap_list);
5082 list_del_init(&conf->bitmap_list);
5083 while (!list_empty(&head)) {
5084 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5086 list_del_init(&sh->lru);
5087 atomic_inc(&sh->count);
5088 hash = sh->hash_lock_index;
5089 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5093 static int raid5_congested(struct mddev *mddev, int bits)
5095 struct r5conf *conf = mddev->private;
5097 /* No difference between reads and writes. Just check
5098 * how busy the stripe_cache is
5101 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5104 /* Also checks whether there is pressure on r5cache log space */
5105 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5109 if (atomic_read(&conf->empty_inactive_list_nr))
5115 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5117 struct r5conf *conf = mddev->private;
5118 sector_t sector = bio->bi_iter.bi_sector;
5119 unsigned int chunk_sectors;
5120 unsigned int bio_sectors = bio_sectors(bio);
5122 WARN_ON_ONCE(bio->bi_partno);
5124 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5125 return chunk_sectors >=
5126 ((sector & (chunk_sectors - 1)) + bio_sectors);
5130 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5131 * later sampled by raid5d.
5133 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5135 unsigned long flags;
5137 spin_lock_irqsave(&conf->device_lock, flags);
5139 bi->bi_next = conf->retry_read_aligned_list;
5140 conf->retry_read_aligned_list = bi;
5142 spin_unlock_irqrestore(&conf->device_lock, flags);
5143 md_wakeup_thread(conf->mddev->thread);
5146 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5147 unsigned int *offset)
5151 bi = conf->retry_read_aligned;
5153 *offset = conf->retry_read_offset;
5154 conf->retry_read_aligned = NULL;
5157 bi = conf->retry_read_aligned_list;
5159 conf->retry_read_aligned_list = bi->bi_next;
5168 * The "raid5_align_endio" should check if the read succeeded and if it
5169 * did, call bio_endio on the original bio (having bio_put the new bio
5171 * If the read failed..
5173 static void raid5_align_endio(struct bio *bi)
5175 struct bio* raid_bi = bi->bi_private;
5176 struct mddev *mddev;
5177 struct r5conf *conf;
5178 struct md_rdev *rdev;
5179 blk_status_t error = bi->bi_status;
5183 rdev = (void*)raid_bi->bi_next;
5184 raid_bi->bi_next = NULL;
5185 mddev = rdev->mddev;
5186 conf = mddev->private;
5188 rdev_dec_pending(rdev, conf->mddev);
5192 if (atomic_dec_and_test(&conf->active_aligned_reads))
5193 wake_up(&conf->wait_for_quiescent);
5197 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5199 add_bio_to_retry(raid_bi, conf);
5202 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5204 struct r5conf *conf = mddev->private;
5206 struct bio* align_bi;
5207 struct md_rdev *rdev;
5208 sector_t end_sector;
5210 if (!in_chunk_boundary(mddev, raid_bio)) {
5211 pr_debug("%s: non aligned\n", __func__);
5215 * use bio_clone_fast to make a copy of the bio
5217 align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5221 * set bi_end_io to a new function, and set bi_private to the
5224 align_bi->bi_end_io = raid5_align_endio;
5225 align_bi->bi_private = raid_bio;
5229 align_bi->bi_iter.bi_sector =
5230 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5233 end_sector = bio_end_sector(align_bi);
5235 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5236 if (!rdev || test_bit(Faulty, &rdev->flags) ||
5237 rdev->recovery_offset < end_sector) {
5238 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5240 (test_bit(Faulty, &rdev->flags) ||
5241 !(test_bit(In_sync, &rdev->flags) ||
5242 rdev->recovery_offset >= end_sector)))
5246 if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5256 atomic_inc(&rdev->nr_pending);
5258 raid_bio->bi_next = (void*)rdev;
5259 bio_set_dev(align_bi, rdev->bdev);
5260 bio_clear_flag(align_bi, BIO_SEG_VALID);
5262 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5263 bio_sectors(align_bi),
5264 &first_bad, &bad_sectors)) {
5266 rdev_dec_pending(rdev, mddev);
5270 /* No reshape active, so we can trust rdev->data_offset */
5271 align_bi->bi_iter.bi_sector += rdev->data_offset;
5273 spin_lock_irq(&conf->device_lock);
5274 wait_event_lock_irq(conf->wait_for_quiescent,
5277 atomic_inc(&conf->active_aligned_reads);
5278 spin_unlock_irq(&conf->device_lock);
5281 trace_block_bio_remap(align_bi->bi_disk->queue,
5282 align_bi, disk_devt(mddev->gendisk),
5283 raid_bio->bi_iter.bi_sector);
5284 generic_make_request(align_bi);
5293 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5296 sector_t sector = raid_bio->bi_iter.bi_sector;
5297 unsigned chunk_sects = mddev->chunk_sectors;
5298 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5300 if (sectors < bio_sectors(raid_bio)) {
5301 struct r5conf *conf = mddev->private;
5302 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5303 bio_chain(split, raid_bio);
5304 generic_make_request(raid_bio);
5308 if (!raid5_read_one_chunk(mddev, raid_bio))
5314 /* __get_priority_stripe - get the next stripe to process
5316 * Full stripe writes are allowed to pass preread active stripes up until
5317 * the bypass_threshold is exceeded. In general the bypass_count
5318 * increments when the handle_list is handled before the hold_list; however, it
5319 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5320 * stripe with in flight i/o. The bypass_count will be reset when the
5321 * head of the hold_list has changed, i.e. the head was promoted to the
5324 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5326 struct stripe_head *sh, *tmp;
5327 struct list_head *handle_list = NULL;
5328 struct r5worker_group *wg;
5329 bool second_try = !r5c_is_writeback(conf->log) &&
5330 !r5l_log_disk_error(conf);
5331 bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5332 r5l_log_disk_error(conf);
5337 if (conf->worker_cnt_per_group == 0) {
5338 handle_list = try_loprio ? &conf->loprio_list :
5340 } else if (group != ANY_GROUP) {
5341 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5342 &conf->worker_groups[group].handle_list;
5343 wg = &conf->worker_groups[group];
5346 for (i = 0; i < conf->group_cnt; i++) {
5347 handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5348 &conf->worker_groups[i].handle_list;
5349 wg = &conf->worker_groups[i];
5350 if (!list_empty(handle_list))
5355 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5357 list_empty(handle_list) ? "empty" : "busy",
5358 list_empty(&conf->hold_list) ? "empty" : "busy",
5359 atomic_read(&conf->pending_full_writes), conf->bypass_count);
5361 if (!list_empty(handle_list)) {
5362 sh = list_entry(handle_list->next, typeof(*sh), lru);
5364 if (list_empty(&conf->hold_list))
5365 conf->bypass_count = 0;
5366 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5367 if (conf->hold_list.next == conf->last_hold)
5368 conf->bypass_count++;
5370 conf->last_hold = conf->hold_list.next;
5371 conf->bypass_count -= conf->bypass_threshold;
5372 if (conf->bypass_count < 0)
5373 conf->bypass_count = 0;
5376 } else if (!list_empty(&conf->hold_list) &&
5377 ((conf->bypass_threshold &&
5378 conf->bypass_count > conf->bypass_threshold) ||
5379 atomic_read(&conf->pending_full_writes) == 0)) {
5381 list_for_each_entry(tmp, &conf->hold_list, lru) {
5382 if (conf->worker_cnt_per_group == 0 ||
5383 group == ANY_GROUP ||
5384 !cpu_online(tmp->cpu) ||
5385 cpu_to_group(tmp->cpu) == group) {
5392 conf->bypass_count -= conf->bypass_threshold;
5393 if (conf->bypass_count < 0)
5394 conf->bypass_count = 0;
5403 try_loprio = !try_loprio;
5411 list_del_init(&sh->lru);
5412 BUG_ON(atomic_inc_return(&sh->count) != 1);
5416 struct raid5_plug_cb {
5417 struct blk_plug_cb cb;
5418 struct list_head list;
5419 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5422 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5424 struct raid5_plug_cb *cb = container_of(
5425 blk_cb, struct raid5_plug_cb, cb);
5426 struct stripe_head *sh;
5427 struct mddev *mddev = cb->cb.data;
5428 struct r5conf *conf = mddev->private;
5432 if (cb->list.next && !list_empty(&cb->list)) {
5433 spin_lock_irq(&conf->device_lock);
5434 while (!list_empty(&cb->list)) {
5435 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5436 list_del_init(&sh->lru);
5438 * avoid race release_stripe_plug() sees
5439 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5440 * is still in our list
5442 smp_mb__before_atomic();
5443 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5445 * STRIPE_ON_RELEASE_LIST could be set here. In that
5446 * case, the count is always > 1 here
5448 hash = sh->hash_lock_index;
5449 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5452 spin_unlock_irq(&conf->device_lock);
5454 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5455 NR_STRIPE_HASH_LOCKS);
5457 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5461 static void release_stripe_plug(struct mddev *mddev,
5462 struct stripe_head *sh)
5464 struct blk_plug_cb *blk_cb = blk_check_plugged(
5465 raid5_unplug, mddev,
5466 sizeof(struct raid5_plug_cb));
5467 struct raid5_plug_cb *cb;
5470 raid5_release_stripe(sh);
5474 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5476 if (cb->list.next == NULL) {
5478 INIT_LIST_HEAD(&cb->list);
5479 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5480 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5483 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5484 list_add_tail(&sh->lru, &cb->list);
5486 raid5_release_stripe(sh);
5489 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5491 struct r5conf *conf = mddev->private;
5492 sector_t logical_sector, last_sector;
5493 struct stripe_head *sh;
5496 if (mddev->reshape_position != MaxSector)
5497 /* Skip discard while reshape is happening */
5500 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5501 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5505 stripe_sectors = conf->chunk_sectors *
5506 (conf->raid_disks - conf->max_degraded);
5507 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5509 sector_div(last_sector, stripe_sectors);
5511 logical_sector *= conf->chunk_sectors;
5512 last_sector *= conf->chunk_sectors;
5514 for (; logical_sector < last_sector;
5515 logical_sector += STRIPE_SECTORS) {
5519 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5520 prepare_to_wait(&conf->wait_for_overlap, &w,
5521 TASK_UNINTERRUPTIBLE);
5522 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5523 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5524 raid5_release_stripe(sh);
5528 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5529 spin_lock_irq(&sh->stripe_lock);
5530 for (d = 0; d < conf->raid_disks; d++) {
5531 if (d == sh->pd_idx || d == sh->qd_idx)
5533 if (sh->dev[d].towrite || sh->dev[d].toread) {
5534 set_bit(R5_Overlap, &sh->dev[d].flags);
5535 spin_unlock_irq(&sh->stripe_lock);
5536 raid5_release_stripe(sh);
5541 set_bit(STRIPE_DISCARD, &sh->state);
5542 finish_wait(&conf->wait_for_overlap, &w);
5543 sh->overwrite_disks = 0;
5544 for (d = 0; d < conf->raid_disks; d++) {
5545 if (d == sh->pd_idx || d == sh->qd_idx)
5547 sh->dev[d].towrite = bi;
5548 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5549 bio_inc_remaining(bi);
5550 md_write_inc(mddev, bi);
5551 sh->overwrite_disks++;
5553 spin_unlock_irq(&sh->stripe_lock);
5554 if (conf->mddev->bitmap) {
5556 d < conf->raid_disks - conf->max_degraded;
5558 md_bitmap_startwrite(mddev->bitmap,
5562 sh->bm_seq = conf->seq_flush + 1;
5563 set_bit(STRIPE_BIT_DELAY, &sh->state);
5566 set_bit(STRIPE_HANDLE, &sh->state);
5567 clear_bit(STRIPE_DELAYED, &sh->state);
5568 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5569 atomic_inc(&conf->preread_active_stripes);
5570 release_stripe_plug(mddev, sh);
5576 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5578 struct r5conf *conf = mddev->private;
5580 sector_t new_sector;
5581 sector_t logical_sector, last_sector;
5582 struct stripe_head *sh;
5583 const int rw = bio_data_dir(bi);
5586 bool do_flush = false;
5588 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5589 int ret = log_handle_flush_request(conf, bi);
5593 if (ret == -ENODEV) {
5594 if (md_flush_request(mddev, bi))
5597 /* ret == -EAGAIN, fallback */
5599 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5600 * we need to flush journal device
5602 do_flush = bi->bi_opf & REQ_PREFLUSH;
5605 if (!md_write_start(mddev, bi))
5608 * If array is degraded, better not do chunk aligned read because
5609 * later we might have to read it again in order to reconstruct
5610 * data on failed drives.
5612 if (rw == READ && mddev->degraded == 0 &&
5613 mddev->reshape_position == MaxSector) {
5614 bi = chunk_aligned_read(mddev, bi);
5619 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5620 make_discard_request(mddev, bi);
5621 md_write_end(mddev);
5625 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5626 last_sector = bio_end_sector(bi);
5629 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5630 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5636 seq = read_seqcount_begin(&conf->gen_lock);
5639 prepare_to_wait(&conf->wait_for_overlap, &w,
5640 TASK_UNINTERRUPTIBLE);
5641 if (unlikely(conf->reshape_progress != MaxSector)) {
5642 /* spinlock is needed as reshape_progress may be
5643 * 64bit on a 32bit platform, and so it might be
5644 * possible to see a half-updated value
5645 * Of course reshape_progress could change after
5646 * the lock is dropped, so once we get a reference
5647 * to the stripe that we think it is, we will have
5650 spin_lock_irq(&conf->device_lock);
5651 if (mddev->reshape_backwards
5652 ? logical_sector < conf->reshape_progress
5653 : logical_sector >= conf->reshape_progress) {
5656 if (mddev->reshape_backwards
5657 ? logical_sector < conf->reshape_safe
5658 : logical_sector >= conf->reshape_safe) {
5659 spin_unlock_irq(&conf->device_lock);
5665 spin_unlock_irq(&conf->device_lock);
5668 new_sector = raid5_compute_sector(conf, logical_sector,
5671 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5672 (unsigned long long)new_sector,
5673 (unsigned long long)logical_sector);
5675 sh = raid5_get_active_stripe(conf, new_sector, previous,
5676 (bi->bi_opf & REQ_RAHEAD), 0);
5678 if (unlikely(previous)) {
5679 /* expansion might have moved on while waiting for a
5680 * stripe, so we must do the range check again.
5681 * Expansion could still move past after this
5682 * test, but as we are holding a reference to
5683 * 'sh', we know that if that happens,
5684 * STRIPE_EXPANDING will get set and the expansion
5685 * won't proceed until we finish with the stripe.
5688 spin_lock_irq(&conf->device_lock);
5689 if (mddev->reshape_backwards
5690 ? logical_sector >= conf->reshape_progress
5691 : logical_sector < conf->reshape_progress)
5692 /* mismatch, need to try again */
5694 spin_unlock_irq(&conf->device_lock);
5696 raid5_release_stripe(sh);
5702 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5703 /* Might have got the wrong stripe_head
5706 raid5_release_stripe(sh);
5710 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5711 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5712 /* Stripe is busy expanding or
5713 * add failed due to overlap. Flush everything
5716 md_wakeup_thread(mddev->thread);
5717 raid5_release_stripe(sh);
5723 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5724 /* we only need flush for one stripe */
5728 if (!sh->batch_head || sh == sh->batch_head)
5729 set_bit(STRIPE_HANDLE, &sh->state);
5730 clear_bit(STRIPE_DELAYED, &sh->state);
5731 if ((!sh->batch_head || sh == sh->batch_head) &&
5732 (bi->bi_opf & REQ_SYNC) &&
5733 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5734 atomic_inc(&conf->preread_active_stripes);
5735 release_stripe_plug(mddev, sh);
5737 /* cannot get stripe for read-ahead, just give-up */
5738 bi->bi_status = BLK_STS_IOERR;
5742 finish_wait(&conf->wait_for_overlap, &w);
5745 md_write_end(mddev);
5750 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5752 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5754 /* reshaping is quite different to recovery/resync so it is
5755 * handled quite separately ... here.
5757 * On each call to sync_request, we gather one chunk worth of
5758 * destination stripes and flag them as expanding.
5759 * Then we find all the source stripes and request reads.
5760 * As the reads complete, handle_stripe will copy the data
5761 * into the destination stripe and release that stripe.
5763 struct r5conf *conf = mddev->private;
5764 struct stripe_head *sh;
5765 struct md_rdev *rdev;
5766 sector_t first_sector, last_sector;
5767 int raid_disks = conf->previous_raid_disks;
5768 int data_disks = raid_disks - conf->max_degraded;
5769 int new_data_disks = conf->raid_disks - conf->max_degraded;
5772 sector_t writepos, readpos, safepos;
5773 sector_t stripe_addr;
5774 int reshape_sectors;
5775 struct list_head stripes;
5778 if (sector_nr == 0) {
5779 /* If restarting in the middle, skip the initial sectors */
5780 if (mddev->reshape_backwards &&
5781 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5782 sector_nr = raid5_size(mddev, 0, 0)
5783 - conf->reshape_progress;
5784 } else if (mddev->reshape_backwards &&
5785 conf->reshape_progress == MaxSector) {
5786 /* shouldn't happen, but just in case, finish up.*/
5787 sector_nr = MaxSector;
5788 } else if (!mddev->reshape_backwards &&
5789 conf->reshape_progress > 0)
5790 sector_nr = conf->reshape_progress;
5791 sector_div(sector_nr, new_data_disks);
5793 mddev->curr_resync_completed = sector_nr;
5794 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5801 /* We need to process a full chunk at a time.
5802 * If old and new chunk sizes differ, we need to process the
5806 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5808 /* We update the metadata at least every 10 seconds, or when
5809 * the data about to be copied would over-write the source of
5810 * the data at the front of the range. i.e. one new_stripe
5811 * along from reshape_progress new_maps to after where
5812 * reshape_safe old_maps to
5814 writepos = conf->reshape_progress;
5815 sector_div(writepos, new_data_disks);
5816 readpos = conf->reshape_progress;
5817 sector_div(readpos, data_disks);
5818 safepos = conf->reshape_safe;
5819 sector_div(safepos, data_disks);
5820 if (mddev->reshape_backwards) {
5821 BUG_ON(writepos < reshape_sectors);
5822 writepos -= reshape_sectors;
5823 readpos += reshape_sectors;
5824 safepos += reshape_sectors;
5826 writepos += reshape_sectors;
5827 /* readpos and safepos are worst-case calculations.
5828 * A negative number is overly pessimistic, and causes
5829 * obvious problems for unsigned storage. So clip to 0.
5831 readpos -= min_t(sector_t, reshape_sectors, readpos);
5832 safepos -= min_t(sector_t, reshape_sectors, safepos);
5835 /* Having calculated the 'writepos' possibly use it
5836 * to set 'stripe_addr' which is where we will write to.
5838 if (mddev->reshape_backwards) {
5839 BUG_ON(conf->reshape_progress == 0);
5840 stripe_addr = writepos;
5841 BUG_ON((mddev->dev_sectors &
5842 ~((sector_t)reshape_sectors - 1))
5843 - reshape_sectors - stripe_addr
5846 BUG_ON(writepos != sector_nr + reshape_sectors);
5847 stripe_addr = sector_nr;
5850 /* 'writepos' is the most advanced device address we might write.
5851 * 'readpos' is the least advanced device address we might read.
5852 * 'safepos' is the least address recorded in the metadata as having
5854 * If there is a min_offset_diff, these are adjusted either by
5855 * increasing the safepos/readpos if diff is negative, or
5856 * increasing writepos if diff is positive.
5857 * If 'readpos' is then behind 'writepos', there is no way that we can
5858 * ensure safety in the face of a crash - that must be done by userspace
5859 * making a backup of the data. So in that case there is no particular
5860 * rush to update metadata.
5861 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5862 * update the metadata to advance 'safepos' to match 'readpos' so that
5863 * we can be safe in the event of a crash.
5864 * So we insist on updating metadata if safepos is behind writepos and
5865 * readpos is beyond writepos.
5866 * In any case, update the metadata every 10 seconds.
5867 * Maybe that number should be configurable, but I'm not sure it is
5868 * worth it.... maybe it could be a multiple of safemode_delay???
5870 if (conf->min_offset_diff < 0) {
5871 safepos += -conf->min_offset_diff;
5872 readpos += -conf->min_offset_diff;
5874 writepos += conf->min_offset_diff;
5876 if ((mddev->reshape_backwards
5877 ? (safepos > writepos && readpos < writepos)
5878 : (safepos < writepos && readpos > writepos)) ||
5879 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5880 /* Cannot proceed until we've updated the superblock... */
5881 wait_event(conf->wait_for_overlap,
5882 atomic_read(&conf->reshape_stripes)==0
5883 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5884 if (atomic_read(&conf->reshape_stripes) != 0)
5886 mddev->reshape_position = conf->reshape_progress;
5887 mddev->curr_resync_completed = sector_nr;
5888 if (!mddev->reshape_backwards)
5889 /* Can update recovery_offset */
5890 rdev_for_each(rdev, mddev)
5891 if (rdev->raid_disk >= 0 &&
5892 !test_bit(Journal, &rdev->flags) &&
5893 !test_bit(In_sync, &rdev->flags) &&
5894 rdev->recovery_offset < sector_nr)
5895 rdev->recovery_offset = sector_nr;
5897 conf->reshape_checkpoint = jiffies;
5898 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5899 md_wakeup_thread(mddev->thread);
5900 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5901 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5902 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5904 spin_lock_irq(&conf->device_lock);
5905 conf->reshape_safe = mddev->reshape_position;
5906 spin_unlock_irq(&conf->device_lock);
5907 wake_up(&conf->wait_for_overlap);
5908 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5911 INIT_LIST_HEAD(&stripes);
5912 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5914 int skipped_disk = 0;
5915 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5916 set_bit(STRIPE_EXPANDING, &sh->state);
5917 atomic_inc(&conf->reshape_stripes);
5918 /* If any of this stripe is beyond the end of the old
5919 * array, then we need to zero those blocks
5921 for (j=sh->disks; j--;) {
5923 if (j == sh->pd_idx)
5925 if (conf->level == 6 &&
5928 s = raid5_compute_blocknr(sh, j, 0);
5929 if (s < raid5_size(mddev, 0, 0)) {
5933 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5934 set_bit(R5_Expanded, &sh->dev[j].flags);
5935 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5937 if (!skipped_disk) {
5938 set_bit(STRIPE_EXPAND_READY, &sh->state);
5939 set_bit(STRIPE_HANDLE, &sh->state);
5941 list_add(&sh->lru, &stripes);
5943 spin_lock_irq(&conf->device_lock);
5944 if (mddev->reshape_backwards)
5945 conf->reshape_progress -= reshape_sectors * new_data_disks;
5947 conf->reshape_progress += reshape_sectors * new_data_disks;
5948 spin_unlock_irq(&conf->device_lock);
5949 /* Ok, those stripe are ready. We can start scheduling
5950 * reads on the source stripes.
5951 * The source stripes are determined by mapping the first and last
5952 * block on the destination stripes.
5955 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5958 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5959 * new_data_disks - 1),
5961 if (last_sector >= mddev->dev_sectors)
5962 last_sector = mddev->dev_sectors - 1;
5963 while (first_sector <= last_sector) {
5964 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5965 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5966 set_bit(STRIPE_HANDLE, &sh->state);
5967 raid5_release_stripe(sh);
5968 first_sector += STRIPE_SECTORS;
5970 /* Now that the sources are clearly marked, we can release
5971 * the destination stripes
5973 while (!list_empty(&stripes)) {
5974 sh = list_entry(stripes.next, struct stripe_head, lru);
5975 list_del_init(&sh->lru);
5976 raid5_release_stripe(sh);
5978 /* If this takes us to the resync_max point where we have to pause,
5979 * then we need to write out the superblock.
5981 sector_nr += reshape_sectors;
5982 retn = reshape_sectors;
5984 if (mddev->curr_resync_completed > mddev->resync_max ||
5985 (sector_nr - mddev->curr_resync_completed) * 2
5986 >= mddev->resync_max - mddev->curr_resync_completed) {
5987 /* Cannot proceed until we've updated the superblock... */
5988 wait_event(conf->wait_for_overlap,
5989 atomic_read(&conf->reshape_stripes) == 0
5990 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5991 if (atomic_read(&conf->reshape_stripes) != 0)
5993 mddev->reshape_position = conf->reshape_progress;
5994 mddev->curr_resync_completed = sector_nr;
5995 if (!mddev->reshape_backwards)
5996 /* Can update recovery_offset */
5997 rdev_for_each(rdev, mddev)
5998 if (rdev->raid_disk >= 0 &&
5999 !test_bit(Journal, &rdev->flags) &&
6000 !test_bit(In_sync, &rdev->flags) &&
6001 rdev->recovery_offset < sector_nr)
6002 rdev->recovery_offset = sector_nr;
6003 conf->reshape_checkpoint = jiffies;
6004 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6005 md_wakeup_thread(mddev->thread);
6006 wait_event(mddev->sb_wait,
6007 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6008 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6009 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6011 spin_lock_irq(&conf->device_lock);
6012 conf->reshape_safe = mddev->reshape_position;
6013 spin_unlock_irq(&conf->device_lock);
6014 wake_up(&conf->wait_for_overlap);
6015 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6021 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6024 struct r5conf *conf = mddev->private;
6025 struct stripe_head *sh;
6026 sector_t max_sector = mddev->dev_sectors;
6027 sector_t sync_blocks;
6028 int still_degraded = 0;
6031 if (sector_nr >= max_sector) {
6032 /* just being told to finish up .. nothing much to do */
6034 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6039 if (mddev->curr_resync < max_sector) /* aborted */
6040 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6042 else /* completed sync */
6044 md_bitmap_close_sync(mddev->bitmap);
6049 /* Allow raid5_quiesce to complete */
6050 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6052 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6053 return reshape_request(mddev, sector_nr, skipped);
6055 /* No need to check resync_max as we never do more than one
6056 * stripe, and as resync_max will always be on a chunk boundary,
6057 * if the check in md_do_sync didn't fire, there is no chance
6058 * of overstepping resync_max here
6061 /* if there is too many failed drives and we are trying
6062 * to resync, then assert that we are finished, because there is
6063 * nothing we can do.
6065 if (mddev->degraded >= conf->max_degraded &&
6066 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6067 sector_t rv = mddev->dev_sectors - sector_nr;
6071 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6073 !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6074 sync_blocks >= STRIPE_SECTORS) {
6075 /* we can skip this block, and probably more */
6076 sync_blocks /= STRIPE_SECTORS;
6078 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6081 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6083 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6085 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6086 /* make sure we don't swamp the stripe cache if someone else
6087 * is trying to get access
6089 schedule_timeout_uninterruptible(1);
6091 /* Need to check if array will still be degraded after recovery/resync
6092 * Note in case of > 1 drive failures it's possible we're rebuilding
6093 * one drive while leaving another faulty drive in array.
6096 for (i = 0; i < conf->raid_disks; i++) {
6097 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6099 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6104 md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6106 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6107 set_bit(STRIPE_HANDLE, &sh->state);
6109 raid5_release_stripe(sh);
6111 return STRIPE_SECTORS;
6114 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6115 unsigned int offset)
6117 /* We may not be able to submit a whole bio at once as there
6118 * may not be enough stripe_heads available.
6119 * We cannot pre-allocate enough stripe_heads as we may need
6120 * more than exist in the cache (if we allow ever large chunks).
6121 * So we do one stripe head at a time and record in
6122 * ->bi_hw_segments how many have been done.
6124 * We *know* that this entire raid_bio is in one chunk, so
6125 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6127 struct stripe_head *sh;
6129 sector_t sector, logical_sector, last_sector;
6133 logical_sector = raid_bio->bi_iter.bi_sector &
6134 ~((sector_t)STRIPE_SECTORS-1);
6135 sector = raid5_compute_sector(conf, logical_sector,
6137 last_sector = bio_end_sector(raid_bio);
6139 for (; logical_sector < last_sector;
6140 logical_sector += STRIPE_SECTORS,
6141 sector += STRIPE_SECTORS,
6145 /* already done this stripe */
6148 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6151 /* failed to get a stripe - must wait */
6152 conf->retry_read_aligned = raid_bio;
6153 conf->retry_read_offset = scnt;
6157 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6158 raid5_release_stripe(sh);
6159 conf->retry_read_aligned = raid_bio;
6160 conf->retry_read_offset = scnt;
6164 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6166 raid5_release_stripe(sh);
6170 bio_endio(raid_bio);
6172 if (atomic_dec_and_test(&conf->active_aligned_reads))
6173 wake_up(&conf->wait_for_quiescent);
6177 static int handle_active_stripes(struct r5conf *conf, int group,
6178 struct r5worker *worker,
6179 struct list_head *temp_inactive_list)
6181 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6182 int i, batch_size = 0, hash;
6183 bool release_inactive = false;
6185 while (batch_size < MAX_STRIPE_BATCH &&
6186 (sh = __get_priority_stripe(conf, group)) != NULL)
6187 batch[batch_size++] = sh;
6189 if (batch_size == 0) {
6190 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6191 if (!list_empty(temp_inactive_list + i))
6193 if (i == NR_STRIPE_HASH_LOCKS) {
6194 spin_unlock_irq(&conf->device_lock);
6195 log_flush_stripe_to_raid(conf);
6196 spin_lock_irq(&conf->device_lock);
6199 release_inactive = true;
6201 spin_unlock_irq(&conf->device_lock);
6203 release_inactive_stripe_list(conf, temp_inactive_list,
6204 NR_STRIPE_HASH_LOCKS);
6206 r5l_flush_stripe_to_raid(conf->log);
6207 if (release_inactive) {
6208 spin_lock_irq(&conf->device_lock);
6212 for (i = 0; i < batch_size; i++)
6213 handle_stripe(batch[i]);
6214 log_write_stripe_run(conf);
6218 spin_lock_irq(&conf->device_lock);
6219 for (i = 0; i < batch_size; i++) {
6220 hash = batch[i]->hash_lock_index;
6221 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6226 static void raid5_do_work(struct work_struct *work)
6228 struct r5worker *worker = container_of(work, struct r5worker, work);
6229 struct r5worker_group *group = worker->group;
6230 struct r5conf *conf = group->conf;
6231 struct mddev *mddev = conf->mddev;
6232 int group_id = group - conf->worker_groups;
6234 struct blk_plug plug;
6236 pr_debug("+++ raid5worker active\n");
6238 blk_start_plug(&plug);
6240 spin_lock_irq(&conf->device_lock);
6242 int batch_size, released;
6244 released = release_stripe_list(conf, worker->temp_inactive_list);
6246 batch_size = handle_active_stripes(conf, group_id, worker,
6247 worker->temp_inactive_list);
6248 worker->working = false;
6249 if (!batch_size && !released)
6251 handled += batch_size;
6252 wait_event_lock_irq(mddev->sb_wait,
6253 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6256 pr_debug("%d stripes handled\n", handled);
6258 spin_unlock_irq(&conf->device_lock);
6260 flush_deferred_bios(conf);
6262 r5l_flush_stripe_to_raid(conf->log);
6264 async_tx_issue_pending_all();
6265 blk_finish_plug(&plug);
6267 pr_debug("--- raid5worker inactive\n");
6271 * This is our raid5 kernel thread.
6273 * We scan the hash table for stripes which can be handled now.
6274 * During the scan, completed stripes are saved for us by the interrupt
6275 * handler, so that they will not have to wait for our next wakeup.
6277 static void raid5d(struct md_thread *thread)
6279 struct mddev *mddev = thread->mddev;
6280 struct r5conf *conf = mddev->private;
6282 struct blk_plug plug;
6284 pr_debug("+++ raid5d active\n");
6286 md_check_recovery(mddev);
6288 blk_start_plug(&plug);
6290 spin_lock_irq(&conf->device_lock);
6293 int batch_size, released;
6294 unsigned int offset;
6296 released = release_stripe_list(conf, conf->temp_inactive_list);
6298 clear_bit(R5_DID_ALLOC, &conf->cache_state);
6301 !list_empty(&conf->bitmap_list)) {
6302 /* Now is a good time to flush some bitmap updates */
6304 spin_unlock_irq(&conf->device_lock);
6305 md_bitmap_unplug(mddev->bitmap);
6306 spin_lock_irq(&conf->device_lock);
6307 conf->seq_write = conf->seq_flush;
6308 activate_bit_delay(conf, conf->temp_inactive_list);
6310 raid5_activate_delayed(conf);
6312 while ((bio = remove_bio_from_retry(conf, &offset))) {
6314 spin_unlock_irq(&conf->device_lock);
6315 ok = retry_aligned_read(conf, bio, offset);
6316 spin_lock_irq(&conf->device_lock);
6322 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6323 conf->temp_inactive_list);
6324 if (!batch_size && !released)
6326 handled += batch_size;
6328 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6329 spin_unlock_irq(&conf->device_lock);
6330 md_check_recovery(mddev);
6331 spin_lock_irq(&conf->device_lock);
6334 pr_debug("%d stripes handled\n", handled);
6336 spin_unlock_irq(&conf->device_lock);
6337 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6338 mutex_trylock(&conf->cache_size_mutex)) {
6339 grow_one_stripe(conf, __GFP_NOWARN);
6340 /* Set flag even if allocation failed. This helps
6341 * slow down allocation requests when mem is short
6343 set_bit(R5_DID_ALLOC, &conf->cache_state);
6344 mutex_unlock(&conf->cache_size_mutex);
6347 flush_deferred_bios(conf);
6349 r5l_flush_stripe_to_raid(conf->log);
6351 async_tx_issue_pending_all();
6352 blk_finish_plug(&plug);
6354 pr_debug("--- raid5d inactive\n");
6358 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6360 struct r5conf *conf;
6362 spin_lock(&mddev->lock);
6363 conf = mddev->private;
6365 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6366 spin_unlock(&mddev->lock);
6371 raid5_set_cache_size(struct mddev *mddev, int size)
6374 struct r5conf *conf = mddev->private;
6376 if (size <= 16 || size > 32768)
6379 conf->min_nr_stripes = size;
6380 mutex_lock(&conf->cache_size_mutex);
6381 while (size < conf->max_nr_stripes &&
6382 drop_one_stripe(conf))
6384 mutex_unlock(&conf->cache_size_mutex);
6386 md_allow_write(mddev);
6388 mutex_lock(&conf->cache_size_mutex);
6389 while (size > conf->max_nr_stripes)
6390 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6391 conf->min_nr_stripes = conf->max_nr_stripes;
6395 mutex_unlock(&conf->cache_size_mutex);
6399 EXPORT_SYMBOL(raid5_set_cache_size);
6402 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6404 struct r5conf *conf;
6408 if (len >= PAGE_SIZE)
6410 if (kstrtoul(page, 10, &new))
6412 err = mddev_lock(mddev);
6415 conf = mddev->private;
6419 err = raid5_set_cache_size(mddev, new);
6420 mddev_unlock(mddev);
6425 static struct md_sysfs_entry
6426 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6427 raid5_show_stripe_cache_size,
6428 raid5_store_stripe_cache_size);
6431 raid5_show_rmw_level(struct mddev *mddev, char *page)
6433 struct r5conf *conf = mddev->private;
6435 return sprintf(page, "%d\n", conf->rmw_level);
6441 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6443 struct r5conf *conf = mddev->private;
6449 if (len >= PAGE_SIZE)
6452 if (kstrtoul(page, 10, &new))
6455 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6458 if (new != PARITY_DISABLE_RMW &&
6459 new != PARITY_ENABLE_RMW &&
6460 new != PARITY_PREFER_RMW)
6463 conf->rmw_level = new;
6467 static struct md_sysfs_entry
6468 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6469 raid5_show_rmw_level,
6470 raid5_store_rmw_level);
6474 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6476 struct r5conf *conf;
6478 spin_lock(&mddev->lock);
6479 conf = mddev->private;
6481 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6482 spin_unlock(&mddev->lock);
6487 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6489 struct r5conf *conf;
6493 if (len >= PAGE_SIZE)
6495 if (kstrtoul(page, 10, &new))
6498 err = mddev_lock(mddev);
6501 conf = mddev->private;
6504 else if (new > conf->min_nr_stripes)
6507 conf->bypass_threshold = new;
6508 mddev_unlock(mddev);
6512 static struct md_sysfs_entry
6513 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6515 raid5_show_preread_threshold,
6516 raid5_store_preread_threshold);
6519 raid5_show_skip_copy(struct mddev *mddev, char *page)
6521 struct r5conf *conf;
6523 spin_lock(&mddev->lock);
6524 conf = mddev->private;
6526 ret = sprintf(page, "%d\n", conf->skip_copy);
6527 spin_unlock(&mddev->lock);
6532 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6534 struct r5conf *conf;
6538 if (len >= PAGE_SIZE)
6540 if (kstrtoul(page, 10, &new))
6544 err = mddev_lock(mddev);
6547 conf = mddev->private;
6550 else if (new != conf->skip_copy) {
6551 mddev_suspend(mddev);
6552 conf->skip_copy = new;
6554 mddev->queue->backing_dev_info->capabilities |=
6555 BDI_CAP_STABLE_WRITES;
6557 mddev->queue->backing_dev_info->capabilities &=
6558 ~BDI_CAP_STABLE_WRITES;
6559 mddev_resume(mddev);
6561 mddev_unlock(mddev);
6565 static struct md_sysfs_entry
6566 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6567 raid5_show_skip_copy,
6568 raid5_store_skip_copy);
6571 stripe_cache_active_show(struct mddev *mddev, char *page)
6573 struct r5conf *conf = mddev->private;
6575 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6580 static struct md_sysfs_entry
6581 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6584 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6586 struct r5conf *conf;
6588 spin_lock(&mddev->lock);
6589 conf = mddev->private;
6591 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6592 spin_unlock(&mddev->lock);
6596 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6598 int *worker_cnt_per_group,
6599 struct r5worker_group **worker_groups);
6601 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6603 struct r5conf *conf;
6606 struct r5worker_group *new_groups, *old_groups;
6607 int group_cnt, worker_cnt_per_group;
6609 if (len >= PAGE_SIZE)
6611 if (kstrtouint(page, 10, &new))
6613 /* 8192 should be big enough */
6617 err = mddev_lock(mddev);
6620 conf = mddev->private;
6623 else if (new != conf->worker_cnt_per_group) {
6624 mddev_suspend(mddev);
6626 old_groups = conf->worker_groups;
6628 flush_workqueue(raid5_wq);
6630 err = alloc_thread_groups(conf, new,
6631 &group_cnt, &worker_cnt_per_group,
6634 spin_lock_irq(&conf->device_lock);
6635 conf->group_cnt = group_cnt;
6636 conf->worker_cnt_per_group = worker_cnt_per_group;
6637 conf->worker_groups = new_groups;
6638 spin_unlock_irq(&conf->device_lock);
6641 kfree(old_groups[0].workers);
6644 mddev_resume(mddev);
6646 mddev_unlock(mddev);
6651 static struct md_sysfs_entry
6652 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6653 raid5_show_group_thread_cnt,
6654 raid5_store_group_thread_cnt);
6656 static struct attribute *raid5_attrs[] = {
6657 &raid5_stripecache_size.attr,
6658 &raid5_stripecache_active.attr,
6659 &raid5_preread_bypass_threshold.attr,
6660 &raid5_group_thread_cnt.attr,
6661 &raid5_skip_copy.attr,
6662 &raid5_rmw_level.attr,
6663 &r5c_journal_mode.attr,
6666 static struct attribute_group raid5_attrs_group = {
6668 .attrs = raid5_attrs,
6671 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6673 int *worker_cnt_per_group,
6674 struct r5worker_group **worker_groups)
6678 struct r5worker *workers;
6680 *worker_cnt_per_group = cnt;
6683 *worker_groups = NULL;
6686 *group_cnt = num_possible_nodes();
6687 size = sizeof(struct r5worker) * cnt;
6688 workers = kcalloc(size, *group_cnt, GFP_NOIO);
6689 *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6691 if (!*worker_groups || !workers) {
6693 kfree(*worker_groups);
6697 for (i = 0; i < *group_cnt; i++) {
6698 struct r5worker_group *group;
6700 group = &(*worker_groups)[i];
6701 INIT_LIST_HEAD(&group->handle_list);
6702 INIT_LIST_HEAD(&group->loprio_list);
6704 group->workers = workers + i * cnt;
6706 for (j = 0; j < cnt; j++) {
6707 struct r5worker *worker = group->workers + j;
6708 worker->group = group;
6709 INIT_WORK(&worker->work, raid5_do_work);
6711 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6712 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6719 static void free_thread_groups(struct r5conf *conf)
6721 if (conf->worker_groups)
6722 kfree(conf->worker_groups[0].workers);
6723 kfree(conf->worker_groups);
6724 conf->worker_groups = NULL;
6728 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6730 struct r5conf *conf = mddev->private;
6733 sectors = mddev->dev_sectors;
6735 /* size is defined by the smallest of previous and new size */
6736 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6738 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6739 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6740 return sectors * (raid_disks - conf->max_degraded);
6743 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6745 safe_put_page(percpu->spare_page);
6746 if (percpu->scribble)
6747 flex_array_free(percpu->scribble);
6748 percpu->spare_page = NULL;
6749 percpu->scribble = NULL;
6752 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6754 if (conf->level == 6 && !percpu->spare_page)
6755 percpu->spare_page = alloc_page(GFP_KERNEL);
6756 if (!percpu->scribble)
6757 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6758 conf->previous_raid_disks),
6759 max(conf->chunk_sectors,
6760 conf->prev_chunk_sectors)
6764 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6765 free_scratch_buffer(conf, percpu);
6772 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6774 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6776 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6780 static void raid5_free_percpu(struct r5conf *conf)
6785 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6786 free_percpu(conf->percpu);
6789 static void free_conf(struct r5conf *conf)
6795 unregister_shrinker(&conf->shrinker);
6796 free_thread_groups(conf);
6797 shrink_stripes(conf);
6798 raid5_free_percpu(conf);
6799 for (i = 0; i < conf->pool_size; i++)
6800 if (conf->disks[i].extra_page)
6801 put_page(conf->disks[i].extra_page);
6803 bioset_exit(&conf->bio_split);
6804 kfree(conf->stripe_hashtbl);
6805 kfree(conf->pending_data);
6809 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6811 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6812 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6814 if (alloc_scratch_buffer(conf, percpu)) {
6815 pr_warn("%s: failed memory allocation for cpu%u\n",
6822 static int raid5_alloc_percpu(struct r5conf *conf)
6826 conf->percpu = alloc_percpu(struct raid5_percpu);
6830 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6832 conf->scribble_disks = max(conf->raid_disks,
6833 conf->previous_raid_disks);
6834 conf->scribble_sectors = max(conf->chunk_sectors,
6835 conf->prev_chunk_sectors);
6840 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6841 struct shrink_control *sc)
6843 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6844 unsigned long ret = SHRINK_STOP;
6846 if (mutex_trylock(&conf->cache_size_mutex)) {
6848 while (ret < sc->nr_to_scan &&
6849 conf->max_nr_stripes > conf->min_nr_stripes) {
6850 if (drop_one_stripe(conf) == 0) {
6856 mutex_unlock(&conf->cache_size_mutex);
6861 static unsigned long raid5_cache_count(struct shrinker *shrink,
6862 struct shrink_control *sc)
6864 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6866 if (conf->max_nr_stripes < conf->min_nr_stripes)
6867 /* unlikely, but not impossible */
6869 return conf->max_nr_stripes - conf->min_nr_stripes;
6872 static struct r5conf *setup_conf(struct mddev *mddev)
6874 struct r5conf *conf;
6875 int raid_disk, memory, max_disks;
6876 struct md_rdev *rdev;
6877 struct disk_info *disk;
6880 int group_cnt, worker_cnt_per_group;
6881 struct r5worker_group *new_group;
6884 if (mddev->new_level != 5
6885 && mddev->new_level != 4
6886 && mddev->new_level != 6) {
6887 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6888 mdname(mddev), mddev->new_level);
6889 return ERR_PTR(-EIO);
6891 if ((mddev->new_level == 5
6892 && !algorithm_valid_raid5(mddev->new_layout)) ||
6893 (mddev->new_level == 6
6894 && !algorithm_valid_raid6(mddev->new_layout))) {
6895 pr_warn("md/raid:%s: layout %d not supported\n",
6896 mdname(mddev), mddev->new_layout);
6897 return ERR_PTR(-EIO);
6899 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6900 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6901 mdname(mddev), mddev->raid_disks);
6902 return ERR_PTR(-EINVAL);
6905 if (!mddev->new_chunk_sectors ||
6906 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6907 !is_power_of_2(mddev->new_chunk_sectors)) {
6908 pr_warn("md/raid:%s: invalid chunk size %d\n",
6909 mdname(mddev), mddev->new_chunk_sectors << 9);
6910 return ERR_PTR(-EINVAL);
6913 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6916 INIT_LIST_HEAD(&conf->free_list);
6917 INIT_LIST_HEAD(&conf->pending_list);
6918 conf->pending_data = kcalloc(PENDING_IO_MAX,
6919 sizeof(struct r5pending_data),
6921 if (!conf->pending_data)
6923 for (i = 0; i < PENDING_IO_MAX; i++)
6924 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6925 /* Don't enable multi-threading by default*/
6926 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6928 conf->group_cnt = group_cnt;
6929 conf->worker_cnt_per_group = worker_cnt_per_group;
6930 conf->worker_groups = new_group;
6933 spin_lock_init(&conf->device_lock);
6934 seqcount_init(&conf->gen_lock);
6935 mutex_init(&conf->cache_size_mutex);
6936 init_waitqueue_head(&conf->wait_for_quiescent);
6937 init_waitqueue_head(&conf->wait_for_stripe);
6938 init_waitqueue_head(&conf->wait_for_overlap);
6939 INIT_LIST_HEAD(&conf->handle_list);
6940 INIT_LIST_HEAD(&conf->loprio_list);
6941 INIT_LIST_HEAD(&conf->hold_list);
6942 INIT_LIST_HEAD(&conf->delayed_list);
6943 INIT_LIST_HEAD(&conf->bitmap_list);
6944 init_llist_head(&conf->released_stripes);
6945 atomic_set(&conf->active_stripes, 0);
6946 atomic_set(&conf->preread_active_stripes, 0);
6947 atomic_set(&conf->active_aligned_reads, 0);
6948 spin_lock_init(&conf->pending_bios_lock);
6949 conf->batch_bio_dispatch = true;
6950 rdev_for_each(rdev, mddev) {
6951 if (test_bit(Journal, &rdev->flags))
6953 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6954 conf->batch_bio_dispatch = false;
6959 conf->bypass_threshold = BYPASS_THRESHOLD;
6960 conf->recovery_disabled = mddev->recovery_disabled - 1;
6962 conf->raid_disks = mddev->raid_disks;
6963 if (mddev->reshape_position == MaxSector)
6964 conf->previous_raid_disks = mddev->raid_disks;
6966 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6967 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6969 conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
6975 for (i = 0; i < max_disks; i++) {
6976 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6977 if (!conf->disks[i].extra_page)
6981 ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
6984 conf->mddev = mddev;
6986 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6989 /* We init hash_locks[0] separately to that it can be used
6990 * as the reference lock in the spin_lock_nest_lock() call
6991 * in lock_all_device_hash_locks_irq in order to convince
6992 * lockdep that we know what we are doing.
6994 spin_lock_init(conf->hash_locks);
6995 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6996 spin_lock_init(conf->hash_locks + i);
6998 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6999 INIT_LIST_HEAD(conf->inactive_list + i);
7001 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7002 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7004 atomic_set(&conf->r5c_cached_full_stripes, 0);
7005 INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7006 atomic_set(&conf->r5c_cached_partial_stripes, 0);
7007 INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7008 atomic_set(&conf->r5c_flushing_full_stripes, 0);
7009 atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7011 conf->level = mddev->new_level;
7012 conf->chunk_sectors = mddev->new_chunk_sectors;
7013 if (raid5_alloc_percpu(conf) != 0)
7016 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7018 rdev_for_each(rdev, mddev) {
7019 raid_disk = rdev->raid_disk;
7020 if (raid_disk >= max_disks
7021 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7023 disk = conf->disks + raid_disk;
7025 if (test_bit(Replacement, &rdev->flags)) {
7026 if (disk->replacement)
7028 disk->replacement = rdev;
7035 if (test_bit(In_sync, &rdev->flags)) {
7036 char b[BDEVNAME_SIZE];
7037 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7038 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7039 } else if (rdev->saved_raid_disk != raid_disk)
7040 /* Cannot rely on bitmap to complete recovery */
7044 conf->level = mddev->new_level;
7045 if (conf->level == 6) {
7046 conf->max_degraded = 2;
7047 if (raid6_call.xor_syndrome)
7048 conf->rmw_level = PARITY_ENABLE_RMW;
7050 conf->rmw_level = PARITY_DISABLE_RMW;
7052 conf->max_degraded = 1;
7053 conf->rmw_level = PARITY_ENABLE_RMW;
7055 conf->algorithm = mddev->new_layout;
7056 conf->reshape_progress = mddev->reshape_position;
7057 if (conf->reshape_progress != MaxSector) {
7058 conf->prev_chunk_sectors = mddev->chunk_sectors;
7059 conf->prev_algo = mddev->layout;
7061 conf->prev_chunk_sectors = conf->chunk_sectors;
7062 conf->prev_algo = conf->algorithm;
7065 conf->min_nr_stripes = NR_STRIPES;
7066 if (mddev->reshape_position != MaxSector) {
7067 int stripes = max_t(int,
7068 ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7069 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7070 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7071 if (conf->min_nr_stripes != NR_STRIPES)
7072 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7073 mdname(mddev), conf->min_nr_stripes);
7075 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7076 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7077 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7078 if (grow_stripes(conf, conf->min_nr_stripes)) {
7079 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7080 mdname(mddev), memory);
7083 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7085 * Losing a stripe head costs more than the time to refill it,
7086 * it reduces the queue depth and so can hurt throughput.
7087 * So set it rather large, scaled by number of devices.
7089 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7090 conf->shrinker.scan_objects = raid5_cache_scan;
7091 conf->shrinker.count_objects = raid5_cache_count;
7092 conf->shrinker.batch = 128;
7093 conf->shrinker.flags = 0;
7094 if (register_shrinker(&conf->shrinker)) {
7095 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7100 sprintf(pers_name, "raid%d", mddev->new_level);
7101 conf->thread = md_register_thread(raid5d, mddev, pers_name);
7102 if (!conf->thread) {
7103 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7113 return ERR_PTR(-EIO);
7115 return ERR_PTR(-ENOMEM);
7118 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7121 case ALGORITHM_PARITY_0:
7122 if (raid_disk < max_degraded)
7125 case ALGORITHM_PARITY_N:
7126 if (raid_disk >= raid_disks - max_degraded)
7129 case ALGORITHM_PARITY_0_6:
7130 if (raid_disk == 0 ||
7131 raid_disk == raid_disks - 1)
7134 case ALGORITHM_LEFT_ASYMMETRIC_6:
7135 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7136 case ALGORITHM_LEFT_SYMMETRIC_6:
7137 case ALGORITHM_RIGHT_SYMMETRIC_6:
7138 if (raid_disk == raid_disks - 1)
7144 static int raid5_run(struct mddev *mddev)
7146 struct r5conf *conf;
7147 int working_disks = 0;
7148 int dirty_parity_disks = 0;
7149 struct md_rdev *rdev;
7150 struct md_rdev *journal_dev = NULL;
7151 sector_t reshape_offset = 0;
7153 long long min_offset_diff = 0;
7156 if (mddev_init_writes_pending(mddev) < 0)
7159 if (mddev->recovery_cp != MaxSector)
7160 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7163 rdev_for_each(rdev, mddev) {
7166 if (test_bit(Journal, &rdev->flags)) {
7170 if (rdev->raid_disk < 0)
7172 diff = (rdev->new_data_offset - rdev->data_offset);
7174 min_offset_diff = diff;
7176 } else if (mddev->reshape_backwards &&
7177 diff < min_offset_diff)
7178 min_offset_diff = diff;
7179 else if (!mddev->reshape_backwards &&
7180 diff > min_offset_diff)
7181 min_offset_diff = diff;
7184 if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7185 (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7186 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7191 if (mddev->reshape_position != MaxSector) {
7192 /* Check that we can continue the reshape.
7193 * Difficulties arise if the stripe we would write to
7194 * next is at or after the stripe we would read from next.
7195 * For a reshape that changes the number of devices, this
7196 * is only possible for a very short time, and mdadm makes
7197 * sure that time appears to have past before assembling
7198 * the array. So we fail if that time hasn't passed.
7199 * For a reshape that keeps the number of devices the same
7200 * mdadm must be monitoring the reshape can keeping the
7201 * critical areas read-only and backed up. It will start
7202 * the array in read-only mode, so we check for that.
7204 sector_t here_new, here_old;
7206 int max_degraded = (mddev->level == 6 ? 2 : 1);
7211 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7216 if (mddev->new_level != mddev->level) {
7217 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7221 old_disks = mddev->raid_disks - mddev->delta_disks;
7222 /* reshape_position must be on a new-stripe boundary, and one
7223 * further up in new geometry must map after here in old
7225 * If the chunk sizes are different, then as we perform reshape
7226 * in units of the largest of the two, reshape_position needs
7227 * be a multiple of the largest chunk size times new data disks.
7229 here_new = mddev->reshape_position;
7230 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7231 new_data_disks = mddev->raid_disks - max_degraded;
7232 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7233 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7237 reshape_offset = here_new * chunk_sectors;
7238 /* here_new is the stripe we will write to */
7239 here_old = mddev->reshape_position;
7240 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7241 /* here_old is the first stripe that we might need to read
7243 if (mddev->delta_disks == 0) {
7244 /* We cannot be sure it is safe to start an in-place
7245 * reshape. It is only safe if user-space is monitoring
7246 * and taking constant backups.
7247 * mdadm always starts a situation like this in
7248 * readonly mode so it can take control before
7249 * allowing any writes. So just check for that.
7251 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7252 abs(min_offset_diff) >= mddev->new_chunk_sectors)
7253 /* not really in-place - so OK */;
7254 else if (mddev->ro == 0) {
7255 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7259 } else if (mddev->reshape_backwards
7260 ? (here_new * chunk_sectors + min_offset_diff <=
7261 here_old * chunk_sectors)
7262 : (here_new * chunk_sectors >=
7263 here_old * chunk_sectors + (-min_offset_diff))) {
7264 /* Reading from the same stripe as writing to - bad */
7265 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7269 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7270 /* OK, we should be able to continue; */
7272 BUG_ON(mddev->level != mddev->new_level);
7273 BUG_ON(mddev->layout != mddev->new_layout);
7274 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7275 BUG_ON(mddev->delta_disks != 0);
7278 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7279 test_bit(MD_HAS_PPL, &mddev->flags)) {
7280 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7282 clear_bit(MD_HAS_PPL, &mddev->flags);
7283 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7286 if (mddev->private == NULL)
7287 conf = setup_conf(mddev);
7289 conf = mddev->private;
7292 return PTR_ERR(conf);
7294 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7296 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7299 set_disk_ro(mddev->gendisk, 1);
7300 } else if (mddev->recovery_cp == MaxSector)
7301 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7304 conf->min_offset_diff = min_offset_diff;
7305 mddev->thread = conf->thread;
7306 conf->thread = NULL;
7307 mddev->private = conf;
7309 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7311 rdev = conf->disks[i].rdev;
7312 if (!rdev && conf->disks[i].replacement) {
7313 /* The replacement is all we have yet */
7314 rdev = conf->disks[i].replacement;
7315 conf->disks[i].replacement = NULL;
7316 clear_bit(Replacement, &rdev->flags);
7317 conf->disks[i].rdev = rdev;
7321 if (conf->disks[i].replacement &&
7322 conf->reshape_progress != MaxSector) {
7323 /* replacements and reshape simply do not mix. */
7324 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7327 if (test_bit(In_sync, &rdev->flags)) {
7331 /* This disc is not fully in-sync. However if it
7332 * just stored parity (beyond the recovery_offset),
7333 * when we don't need to be concerned about the
7334 * array being dirty.
7335 * When reshape goes 'backwards', we never have
7336 * partially completed devices, so we only need
7337 * to worry about reshape going forwards.
7339 /* Hack because v0.91 doesn't store recovery_offset properly. */
7340 if (mddev->major_version == 0 &&
7341 mddev->minor_version > 90)
7342 rdev->recovery_offset = reshape_offset;
7344 if (rdev->recovery_offset < reshape_offset) {
7345 /* We need to check old and new layout */
7346 if (!only_parity(rdev->raid_disk,
7349 conf->max_degraded))
7352 if (!only_parity(rdev->raid_disk,
7354 conf->previous_raid_disks,
7355 conf->max_degraded))
7357 dirty_parity_disks++;
7361 * 0 for a fully functional array, 1 or 2 for a degraded array.
7363 mddev->degraded = raid5_calc_degraded(conf);
7365 if (has_failed(conf)) {
7366 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7367 mdname(mddev), mddev->degraded, conf->raid_disks);
7371 /* device size must be a multiple of chunk size */
7372 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7373 mddev->resync_max_sectors = mddev->dev_sectors;
7375 if (mddev->degraded > dirty_parity_disks &&
7376 mddev->recovery_cp != MaxSector) {
7377 if (test_bit(MD_HAS_PPL, &mddev->flags))
7378 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7380 else if (mddev->ok_start_degraded)
7381 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7384 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7390 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7391 mdname(mddev), conf->level,
7392 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7395 print_raid5_conf(conf);
7397 if (conf->reshape_progress != MaxSector) {
7398 conf->reshape_safe = conf->reshape_progress;
7399 atomic_set(&conf->reshape_stripes, 0);
7400 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7401 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7402 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7403 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7404 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7406 if (!mddev->sync_thread)
7410 /* Ok, everything is just fine now */
7411 if (mddev->to_remove == &raid5_attrs_group)
7412 mddev->to_remove = NULL;
7413 else if (mddev->kobj.sd &&
7414 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7415 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7417 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7421 /* read-ahead size must cover two whole stripes, which
7422 * is 2 * (datadisks) * chunksize where 'n' is the
7423 * number of raid devices
7425 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7426 int stripe = data_disks *
7427 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7428 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7429 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7431 chunk_size = mddev->chunk_sectors << 9;
7432 blk_queue_io_min(mddev->queue, chunk_size);
7433 blk_queue_io_opt(mddev->queue, chunk_size *
7434 (conf->raid_disks - conf->max_degraded));
7435 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7437 * We can only discard a whole stripe. It doesn't make sense to
7438 * discard data disk but write parity disk
7440 stripe = stripe * PAGE_SIZE;
7441 /* Round up to power of 2, as discard handling
7442 * currently assumes that */
7443 while ((stripe-1) & stripe)
7444 stripe = (stripe | (stripe-1)) + 1;
7445 mddev->queue->limits.discard_alignment = stripe;
7446 mddev->queue->limits.discard_granularity = stripe;
7448 blk_queue_max_write_same_sectors(mddev->queue, 0);
7449 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7451 rdev_for_each(rdev, mddev) {
7452 disk_stack_limits(mddev->gendisk, rdev->bdev,
7453 rdev->data_offset << 9);
7454 disk_stack_limits(mddev->gendisk, rdev->bdev,
7455 rdev->new_data_offset << 9);
7459 * zeroing is required, otherwise data
7460 * could be lost. Consider a scenario: discard a stripe
7461 * (the stripe could be inconsistent if
7462 * discard_zeroes_data is 0); write one disk of the
7463 * stripe (the stripe could be inconsistent again
7464 * depending on which disks are used to calculate
7465 * parity); the disk is broken; The stripe data of this
7468 * We only allow DISCARD if the sysadmin has confirmed that
7469 * only safe devices are in use by setting a module parameter.
7470 * A better idea might be to turn DISCARD into WRITE_ZEROES
7471 * requests, as that is required to be safe.
7473 if (devices_handle_discard_safely &&
7474 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7475 mddev->queue->limits.discard_granularity >= stripe)
7476 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7479 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7482 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7485 if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7490 md_unregister_thread(&mddev->thread);
7491 print_raid5_conf(conf);
7493 mddev->private = NULL;
7494 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7498 static void raid5_free(struct mddev *mddev, void *priv)
7500 struct r5conf *conf = priv;
7503 mddev->to_remove = &raid5_attrs_group;
7506 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7508 struct r5conf *conf = mddev->private;
7511 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7512 conf->chunk_sectors / 2, mddev->layout);
7513 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7515 for (i = 0; i < conf->raid_disks; i++) {
7516 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7517 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7520 seq_printf (seq, "]");
7523 static void print_raid5_conf (struct r5conf *conf)
7526 struct disk_info *tmp;
7528 pr_debug("RAID conf printout:\n");
7530 pr_debug("(conf==NULL)\n");
7533 pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7535 conf->raid_disks - conf->mddev->degraded);
7537 for (i = 0; i < conf->raid_disks; i++) {
7538 char b[BDEVNAME_SIZE];
7539 tmp = conf->disks + i;
7541 pr_debug(" disk %d, o:%d, dev:%s\n",
7542 i, !test_bit(Faulty, &tmp->rdev->flags),
7543 bdevname(tmp->rdev->bdev, b));
7547 static int raid5_spare_active(struct mddev *mddev)
7550 struct r5conf *conf = mddev->private;
7551 struct disk_info *tmp;
7553 unsigned long flags;
7555 for (i = 0; i < conf->raid_disks; i++) {
7556 tmp = conf->disks + i;
7557 if (tmp->replacement
7558 && tmp->replacement->recovery_offset == MaxSector
7559 && !test_bit(Faulty, &tmp->replacement->flags)
7560 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7561 /* Replacement has just become active. */
7563 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7566 /* Replaced device not technically faulty,
7567 * but we need to be sure it gets removed
7568 * and never re-added.
7570 set_bit(Faulty, &tmp->rdev->flags);
7571 sysfs_notify_dirent_safe(
7572 tmp->rdev->sysfs_state);
7574 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7575 } else if (tmp->rdev
7576 && tmp->rdev->recovery_offset == MaxSector
7577 && !test_bit(Faulty, &tmp->rdev->flags)
7578 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7580 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7583 spin_lock_irqsave(&conf->device_lock, flags);
7584 mddev->degraded = raid5_calc_degraded(conf);
7585 spin_unlock_irqrestore(&conf->device_lock, flags);
7586 print_raid5_conf(conf);
7590 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7592 struct r5conf *conf = mddev->private;
7594 int number = rdev->raid_disk;
7595 struct md_rdev **rdevp;
7596 struct disk_info *p = conf->disks + number;
7598 print_raid5_conf(conf);
7599 if (test_bit(Journal, &rdev->flags) && conf->log) {
7601 * we can't wait pending write here, as this is called in
7602 * raid5d, wait will deadlock.
7603 * neilb: there is no locking about new writes here,
7604 * so this cannot be safe.
7606 if (atomic_read(&conf->active_stripes) ||
7607 atomic_read(&conf->r5c_cached_full_stripes) ||
7608 atomic_read(&conf->r5c_cached_partial_stripes)) {
7614 if (rdev == p->rdev)
7616 else if (rdev == p->replacement)
7617 rdevp = &p->replacement;
7621 if (number >= conf->raid_disks &&
7622 conf->reshape_progress == MaxSector)
7623 clear_bit(In_sync, &rdev->flags);
7625 if (test_bit(In_sync, &rdev->flags) ||
7626 atomic_read(&rdev->nr_pending)) {
7630 /* Only remove non-faulty devices if recovery
7633 if (!test_bit(Faulty, &rdev->flags) &&
7634 mddev->recovery_disabled != conf->recovery_disabled &&
7635 !has_failed(conf) &&
7636 (!p->replacement || p->replacement == rdev) &&
7637 number < conf->raid_disks) {
7642 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7644 if (atomic_read(&rdev->nr_pending)) {
7645 /* lost the race, try later */
7651 err = log_modify(conf, rdev, false);
7655 if (p->replacement) {
7656 /* We must have just cleared 'rdev' */
7657 p->rdev = p->replacement;
7658 clear_bit(Replacement, &p->replacement->flags);
7659 smp_mb(); /* Make sure other CPUs may see both as identical
7660 * but will never see neither - if they are careful
7662 p->replacement = NULL;
7665 err = log_modify(conf, p->rdev, true);
7668 clear_bit(WantReplacement, &rdev->flags);
7671 print_raid5_conf(conf);
7675 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7677 struct r5conf *conf = mddev->private;
7678 int ret, err = -EEXIST;
7680 struct disk_info *p;
7682 int last = conf->raid_disks - 1;
7684 if (test_bit(Journal, &rdev->flags)) {
7688 rdev->raid_disk = 0;
7690 * The array is in readonly mode if journal is missing, so no
7691 * write requests running. We should be safe
7693 ret = log_init(conf, rdev, false);
7697 ret = r5l_start(conf->log);
7703 if (mddev->recovery_disabled == conf->recovery_disabled)
7706 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7707 /* no point adding a device */
7710 if (rdev->raid_disk >= 0)
7711 first = last = rdev->raid_disk;
7714 * find the disk ... but prefer rdev->saved_raid_disk
7717 if (rdev->saved_raid_disk >= 0 &&
7718 rdev->saved_raid_disk >= first &&
7719 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7720 first = rdev->saved_raid_disk;
7722 for (disk = first; disk <= last; disk++) {
7723 p = conf->disks + disk;
7724 if (p->rdev == NULL) {
7725 clear_bit(In_sync, &rdev->flags);
7726 rdev->raid_disk = disk;
7727 if (rdev->saved_raid_disk != disk)
7729 rcu_assign_pointer(p->rdev, rdev);
7731 err = log_modify(conf, rdev, true);
7736 for (disk = first; disk <= last; disk++) {
7737 p = conf->disks + disk;
7738 if (test_bit(WantReplacement, &p->rdev->flags) &&
7739 p->replacement == NULL) {
7740 clear_bit(In_sync, &rdev->flags);
7741 set_bit(Replacement, &rdev->flags);
7742 rdev->raid_disk = disk;
7745 rcu_assign_pointer(p->replacement, rdev);
7750 print_raid5_conf(conf);
7754 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7756 /* no resync is happening, and there is enough space
7757 * on all devices, so we can resize.
7758 * We need to make sure resync covers any new space.
7759 * If the array is shrinking we should possibly wait until
7760 * any io in the removed space completes, but it hardly seems
7764 struct r5conf *conf = mddev->private;
7766 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7768 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7769 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7770 if (mddev->external_size &&
7771 mddev->array_sectors > newsize)
7773 if (mddev->bitmap) {
7774 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
7778 md_set_array_sectors(mddev, newsize);
7779 if (sectors > mddev->dev_sectors &&
7780 mddev->recovery_cp > mddev->dev_sectors) {
7781 mddev->recovery_cp = mddev->dev_sectors;
7782 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7784 mddev->dev_sectors = sectors;
7785 mddev->resync_max_sectors = sectors;
7789 static int check_stripe_cache(struct mddev *mddev)
7791 /* Can only proceed if there are plenty of stripe_heads.
7792 * We need a minimum of one full stripe,, and for sensible progress
7793 * it is best to have about 4 times that.
7794 * If we require 4 times, then the default 256 4K stripe_heads will
7795 * allow for chunk sizes up to 256K, which is probably OK.
7796 * If the chunk size is greater, user-space should request more
7797 * stripe_heads first.
7799 struct r5conf *conf = mddev->private;
7800 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7801 > conf->min_nr_stripes ||
7802 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7803 > conf->min_nr_stripes) {
7804 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7806 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7813 static int check_reshape(struct mddev *mddev)
7815 struct r5conf *conf = mddev->private;
7817 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7819 if (mddev->delta_disks == 0 &&
7820 mddev->new_layout == mddev->layout &&
7821 mddev->new_chunk_sectors == mddev->chunk_sectors)
7822 return 0; /* nothing to do */
7823 if (has_failed(conf))
7825 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7826 /* We might be able to shrink, but the devices must
7827 * be made bigger first.
7828 * For raid6, 4 is the minimum size.
7829 * Otherwise 2 is the minimum
7832 if (mddev->level == 6)
7834 if (mddev->raid_disks + mddev->delta_disks < min)
7838 if (!check_stripe_cache(mddev))
7841 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7842 mddev->delta_disks > 0)
7843 if (resize_chunks(conf,
7844 conf->previous_raid_disks
7845 + max(0, mddev->delta_disks),
7846 max(mddev->new_chunk_sectors,
7847 mddev->chunk_sectors)
7851 if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7852 return 0; /* never bother to shrink */
7853 return resize_stripes(conf, (conf->previous_raid_disks
7854 + mddev->delta_disks));
7857 static int raid5_start_reshape(struct mddev *mddev)
7859 struct r5conf *conf = mddev->private;
7860 struct md_rdev *rdev;
7862 unsigned long flags;
7864 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7867 if (!check_stripe_cache(mddev))
7870 if (has_failed(conf))
7873 rdev_for_each(rdev, mddev) {
7874 if (!test_bit(In_sync, &rdev->flags)
7875 && !test_bit(Faulty, &rdev->flags))
7879 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7880 /* Not enough devices even to make a degraded array
7885 /* Refuse to reduce size of the array. Any reductions in
7886 * array size must be through explicit setting of array_size
7889 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7890 < mddev->array_sectors) {
7891 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7896 atomic_set(&conf->reshape_stripes, 0);
7897 spin_lock_irq(&conf->device_lock);
7898 write_seqcount_begin(&conf->gen_lock);
7899 conf->previous_raid_disks = conf->raid_disks;
7900 conf->raid_disks += mddev->delta_disks;
7901 conf->prev_chunk_sectors = conf->chunk_sectors;
7902 conf->chunk_sectors = mddev->new_chunk_sectors;
7903 conf->prev_algo = conf->algorithm;
7904 conf->algorithm = mddev->new_layout;
7906 /* Code that selects data_offset needs to see the generation update
7907 * if reshape_progress has been set - so a memory barrier needed.
7910 if (mddev->reshape_backwards)
7911 conf->reshape_progress = raid5_size(mddev, 0, 0);
7913 conf->reshape_progress = 0;
7914 conf->reshape_safe = conf->reshape_progress;
7915 write_seqcount_end(&conf->gen_lock);
7916 spin_unlock_irq(&conf->device_lock);
7918 /* Now make sure any requests that proceeded on the assumption
7919 * the reshape wasn't running - like Discard or Read - have
7922 mddev_suspend(mddev);
7923 mddev_resume(mddev);
7925 /* Add some new drives, as many as will fit.
7926 * We know there are enough to make the newly sized array work.
7927 * Don't add devices if we are reducing the number of
7928 * devices in the array. This is because it is not possible
7929 * to correctly record the "partially reconstructed" state of
7930 * such devices during the reshape and confusion could result.
7932 if (mddev->delta_disks >= 0) {
7933 rdev_for_each(rdev, mddev)
7934 if (rdev->raid_disk < 0 &&
7935 !test_bit(Faulty, &rdev->flags)) {
7936 if (raid5_add_disk(mddev, rdev) == 0) {
7938 >= conf->previous_raid_disks)
7939 set_bit(In_sync, &rdev->flags);
7941 rdev->recovery_offset = 0;
7943 if (sysfs_link_rdev(mddev, rdev))
7944 /* Failure here is OK */;
7946 } else if (rdev->raid_disk >= conf->previous_raid_disks
7947 && !test_bit(Faulty, &rdev->flags)) {
7948 /* This is a spare that was manually added */
7949 set_bit(In_sync, &rdev->flags);
7952 /* When a reshape changes the number of devices,
7953 * ->degraded is measured against the larger of the
7954 * pre and post number of devices.
7956 spin_lock_irqsave(&conf->device_lock, flags);
7957 mddev->degraded = raid5_calc_degraded(conf);
7958 spin_unlock_irqrestore(&conf->device_lock, flags);
7960 mddev->raid_disks = conf->raid_disks;
7961 mddev->reshape_position = conf->reshape_progress;
7962 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7964 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7965 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7966 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7967 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7968 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7969 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7971 if (!mddev->sync_thread) {
7972 mddev->recovery = 0;
7973 spin_lock_irq(&conf->device_lock);
7974 write_seqcount_begin(&conf->gen_lock);
7975 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7976 mddev->new_chunk_sectors =
7977 conf->chunk_sectors = conf->prev_chunk_sectors;
7978 mddev->new_layout = conf->algorithm = conf->prev_algo;
7979 rdev_for_each(rdev, mddev)
7980 rdev->new_data_offset = rdev->data_offset;
7982 conf->generation --;
7983 conf->reshape_progress = MaxSector;
7984 mddev->reshape_position = MaxSector;
7985 write_seqcount_end(&conf->gen_lock);
7986 spin_unlock_irq(&conf->device_lock);
7989 conf->reshape_checkpoint = jiffies;
7990 md_wakeup_thread(mddev->sync_thread);
7991 md_new_event(mddev);
7995 /* This is called from the reshape thread and should make any
7996 * changes needed in 'conf'
7998 static void end_reshape(struct r5conf *conf)
8001 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8002 struct md_rdev *rdev;
8004 spin_lock_irq(&conf->device_lock);
8005 conf->previous_raid_disks = conf->raid_disks;
8006 md_finish_reshape(conf->mddev);
8008 conf->reshape_progress = MaxSector;
8009 conf->mddev->reshape_position = MaxSector;
8010 rdev_for_each(rdev, conf->mddev)
8011 if (rdev->raid_disk >= 0 &&
8012 !test_bit(Journal, &rdev->flags) &&
8013 !test_bit(In_sync, &rdev->flags))
8014 rdev->recovery_offset = MaxSector;
8015 spin_unlock_irq(&conf->device_lock);
8016 wake_up(&conf->wait_for_overlap);
8018 /* read-ahead size must cover two whole stripes, which is
8019 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
8021 if (conf->mddev->queue) {
8022 int data_disks = conf->raid_disks - conf->max_degraded;
8023 int stripe = data_disks * ((conf->chunk_sectors << 9)
8025 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
8026 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
8031 /* This is called from the raid5d thread with mddev_lock held.
8032 * It makes config changes to the device.
8034 static void raid5_finish_reshape(struct mddev *mddev)
8036 struct r5conf *conf = mddev->private;
8038 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8040 if (mddev->delta_disks <= 0) {
8042 spin_lock_irq(&conf->device_lock);
8043 mddev->degraded = raid5_calc_degraded(conf);
8044 spin_unlock_irq(&conf->device_lock);
8045 for (d = conf->raid_disks ;
8046 d < conf->raid_disks - mddev->delta_disks;
8048 struct md_rdev *rdev = conf->disks[d].rdev;
8050 clear_bit(In_sync, &rdev->flags);
8051 rdev = conf->disks[d].replacement;
8053 clear_bit(In_sync, &rdev->flags);
8056 mddev->layout = conf->algorithm;
8057 mddev->chunk_sectors = conf->chunk_sectors;
8058 mddev->reshape_position = MaxSector;
8059 mddev->delta_disks = 0;
8060 mddev->reshape_backwards = 0;
8064 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8066 struct r5conf *conf = mddev->private;
8069 /* stop all writes */
8070 lock_all_device_hash_locks_irq(conf);
8071 /* '2' tells resync/reshape to pause so that all
8072 * active stripes can drain
8074 r5c_flush_cache(conf, INT_MAX);
8076 wait_event_cmd(conf->wait_for_quiescent,
8077 atomic_read(&conf->active_stripes) == 0 &&
8078 atomic_read(&conf->active_aligned_reads) == 0,
8079 unlock_all_device_hash_locks_irq(conf),
8080 lock_all_device_hash_locks_irq(conf));
8082 unlock_all_device_hash_locks_irq(conf);
8083 /* allow reshape to continue */
8084 wake_up(&conf->wait_for_overlap);
8086 /* re-enable writes */
8087 lock_all_device_hash_locks_irq(conf);
8089 wake_up(&conf->wait_for_quiescent);
8090 wake_up(&conf->wait_for_overlap);
8091 unlock_all_device_hash_locks_irq(conf);
8093 log_quiesce(conf, quiesce);
8096 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8098 struct r0conf *raid0_conf = mddev->private;
8101 /* for raid0 takeover only one zone is supported */
8102 if (raid0_conf->nr_strip_zones > 1) {
8103 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8105 return ERR_PTR(-EINVAL);
8108 sectors = raid0_conf->strip_zone[0].zone_end;
8109 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8110 mddev->dev_sectors = sectors;
8111 mddev->new_level = level;
8112 mddev->new_layout = ALGORITHM_PARITY_N;
8113 mddev->new_chunk_sectors = mddev->chunk_sectors;
8114 mddev->raid_disks += 1;
8115 mddev->delta_disks = 1;
8116 /* make sure it will be not marked as dirty */
8117 mddev->recovery_cp = MaxSector;
8119 return setup_conf(mddev);
8122 static void *raid5_takeover_raid1(struct mddev *mddev)
8127 if (mddev->raid_disks != 2 ||
8128 mddev->degraded > 1)
8129 return ERR_PTR(-EINVAL);
8131 /* Should check if there are write-behind devices? */
8133 chunksect = 64*2; /* 64K by default */
8135 /* The array must be an exact multiple of chunksize */
8136 while (chunksect && (mddev->array_sectors & (chunksect-1)))
8139 if ((chunksect<<9) < STRIPE_SIZE)
8140 /* array size does not allow a suitable chunk size */
8141 return ERR_PTR(-EINVAL);
8143 mddev->new_level = 5;
8144 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8145 mddev->new_chunk_sectors = chunksect;
8147 ret = setup_conf(mddev);
8149 mddev_clear_unsupported_flags(mddev,
8150 UNSUPPORTED_MDDEV_FLAGS);
8154 static void *raid5_takeover_raid6(struct mddev *mddev)
8158 switch (mddev->layout) {
8159 case ALGORITHM_LEFT_ASYMMETRIC_6:
8160 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8162 case ALGORITHM_RIGHT_ASYMMETRIC_6:
8163 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8165 case ALGORITHM_LEFT_SYMMETRIC_6:
8166 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8168 case ALGORITHM_RIGHT_SYMMETRIC_6:
8169 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8171 case ALGORITHM_PARITY_0_6:
8172 new_layout = ALGORITHM_PARITY_0;
8174 case ALGORITHM_PARITY_N:
8175 new_layout = ALGORITHM_PARITY_N;
8178 return ERR_PTR(-EINVAL);
8180 mddev->new_level = 5;
8181 mddev->new_layout = new_layout;
8182 mddev->delta_disks = -1;
8183 mddev->raid_disks -= 1;
8184 return setup_conf(mddev);
8187 static int raid5_check_reshape(struct mddev *mddev)
8189 /* For a 2-drive array, the layout and chunk size can be changed
8190 * immediately as not restriping is needed.
8191 * For larger arrays we record the new value - after validation
8192 * to be used by a reshape pass.
8194 struct r5conf *conf = mddev->private;
8195 int new_chunk = mddev->new_chunk_sectors;
8197 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8199 if (new_chunk > 0) {
8200 if (!is_power_of_2(new_chunk))
8202 if (new_chunk < (PAGE_SIZE>>9))
8204 if (mddev->array_sectors & (new_chunk-1))
8205 /* not factor of array size */
8209 /* They look valid */
8211 if (mddev->raid_disks == 2) {
8212 /* can make the change immediately */
8213 if (mddev->new_layout >= 0) {
8214 conf->algorithm = mddev->new_layout;
8215 mddev->layout = mddev->new_layout;
8217 if (new_chunk > 0) {
8218 conf->chunk_sectors = new_chunk ;
8219 mddev->chunk_sectors = new_chunk;
8221 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8222 md_wakeup_thread(mddev->thread);
8224 return check_reshape(mddev);
8227 static int raid6_check_reshape(struct mddev *mddev)
8229 int new_chunk = mddev->new_chunk_sectors;
8231 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8233 if (new_chunk > 0) {
8234 if (!is_power_of_2(new_chunk))
8236 if (new_chunk < (PAGE_SIZE >> 9))
8238 if (mddev->array_sectors & (new_chunk-1))
8239 /* not factor of array size */
8243 /* They look valid */
8244 return check_reshape(mddev);
8247 static void *raid5_takeover(struct mddev *mddev)
8249 /* raid5 can take over:
8250 * raid0 - if there is only one strip zone - make it a raid4 layout
8251 * raid1 - if there are two drives. We need to know the chunk size
8252 * raid4 - trivial - just use a raid4 layout.
8253 * raid6 - Providing it is a *_6 layout
8255 if (mddev->level == 0)
8256 return raid45_takeover_raid0(mddev, 5);
8257 if (mddev->level == 1)
8258 return raid5_takeover_raid1(mddev);
8259 if (mddev->level == 4) {
8260 mddev->new_layout = ALGORITHM_PARITY_N;
8261 mddev->new_level = 5;
8262 return setup_conf(mddev);
8264 if (mddev->level == 6)
8265 return raid5_takeover_raid6(mddev);
8267 return ERR_PTR(-EINVAL);
8270 static void *raid4_takeover(struct mddev *mddev)
8272 /* raid4 can take over:
8273 * raid0 - if there is only one strip zone
8274 * raid5 - if layout is right
8276 if (mddev->level == 0)
8277 return raid45_takeover_raid0(mddev, 4);
8278 if (mddev->level == 5 &&
8279 mddev->layout == ALGORITHM_PARITY_N) {
8280 mddev->new_layout = 0;
8281 mddev->new_level = 4;
8282 return setup_conf(mddev);
8284 return ERR_PTR(-EINVAL);
8287 static struct md_personality raid5_personality;
8289 static void *raid6_takeover(struct mddev *mddev)
8291 /* Currently can only take over a raid5. We map the
8292 * personality to an equivalent raid6 personality
8293 * with the Q block at the end.
8297 if (mddev->pers != &raid5_personality)
8298 return ERR_PTR(-EINVAL);
8299 if (mddev->degraded > 1)
8300 return ERR_PTR(-EINVAL);
8301 if (mddev->raid_disks > 253)
8302 return ERR_PTR(-EINVAL);
8303 if (mddev->raid_disks < 3)
8304 return ERR_PTR(-EINVAL);
8306 switch (mddev->layout) {
8307 case ALGORITHM_LEFT_ASYMMETRIC:
8308 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8310 case ALGORITHM_RIGHT_ASYMMETRIC:
8311 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8313 case ALGORITHM_LEFT_SYMMETRIC:
8314 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8316 case ALGORITHM_RIGHT_SYMMETRIC:
8317 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8319 case ALGORITHM_PARITY_0:
8320 new_layout = ALGORITHM_PARITY_0_6;
8322 case ALGORITHM_PARITY_N:
8323 new_layout = ALGORITHM_PARITY_N;
8326 return ERR_PTR(-EINVAL);
8328 mddev->new_level = 6;
8329 mddev->new_layout = new_layout;
8330 mddev->delta_disks = 1;
8331 mddev->raid_disks += 1;
8332 return setup_conf(mddev);
8335 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8337 struct r5conf *conf;
8340 err = mddev_lock(mddev);
8343 conf = mddev->private;
8345 mddev_unlock(mddev);
8349 if (strncmp(buf, "ppl", 3) == 0) {
8350 /* ppl only works with RAID 5 */
8351 if (!raid5_has_ppl(conf) && conf->level == 5) {
8352 err = log_init(conf, NULL, true);
8354 err = resize_stripes(conf, conf->pool_size);
8360 } else if (strncmp(buf, "resync", 6) == 0) {
8361 if (raid5_has_ppl(conf)) {
8362 mddev_suspend(mddev);
8364 mddev_resume(mddev);
8365 err = resize_stripes(conf, conf->pool_size);
8366 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8367 r5l_log_disk_error(conf)) {
8368 bool journal_dev_exists = false;
8369 struct md_rdev *rdev;
8371 rdev_for_each(rdev, mddev)
8372 if (test_bit(Journal, &rdev->flags)) {
8373 journal_dev_exists = true;
8377 if (!journal_dev_exists) {
8378 mddev_suspend(mddev);
8379 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8380 mddev_resume(mddev);
8381 } else /* need remove journal device first */
8390 md_update_sb(mddev, 1);
8392 mddev_unlock(mddev);
8397 static int raid5_start(struct mddev *mddev)
8399 struct r5conf *conf = mddev->private;
8401 return r5l_start(conf->log);
8404 static struct md_personality raid6_personality =
8408 .owner = THIS_MODULE,
8409 .make_request = raid5_make_request,
8411 .start = raid5_start,
8413 .status = raid5_status,
8414 .error_handler = raid5_error,
8415 .hot_add_disk = raid5_add_disk,
8416 .hot_remove_disk= raid5_remove_disk,
8417 .spare_active = raid5_spare_active,
8418 .sync_request = raid5_sync_request,
8419 .resize = raid5_resize,
8421 .check_reshape = raid6_check_reshape,
8422 .start_reshape = raid5_start_reshape,
8423 .finish_reshape = raid5_finish_reshape,
8424 .quiesce = raid5_quiesce,
8425 .takeover = raid6_takeover,
8426 .congested = raid5_congested,
8427 .change_consistency_policy = raid5_change_consistency_policy,
8429 static struct md_personality raid5_personality =
8433 .owner = THIS_MODULE,
8434 .make_request = raid5_make_request,
8436 .start = raid5_start,
8438 .status = raid5_status,
8439 .error_handler = raid5_error,
8440 .hot_add_disk = raid5_add_disk,
8441 .hot_remove_disk= raid5_remove_disk,
8442 .spare_active = raid5_spare_active,
8443 .sync_request = raid5_sync_request,
8444 .resize = raid5_resize,
8446 .check_reshape = raid5_check_reshape,
8447 .start_reshape = raid5_start_reshape,
8448 .finish_reshape = raid5_finish_reshape,
8449 .quiesce = raid5_quiesce,
8450 .takeover = raid5_takeover,
8451 .congested = raid5_congested,
8452 .change_consistency_policy = raid5_change_consistency_policy,
8455 static struct md_personality raid4_personality =
8459 .owner = THIS_MODULE,
8460 .make_request = raid5_make_request,
8462 .start = raid5_start,
8464 .status = raid5_status,
8465 .error_handler = raid5_error,
8466 .hot_add_disk = raid5_add_disk,
8467 .hot_remove_disk= raid5_remove_disk,
8468 .spare_active = raid5_spare_active,
8469 .sync_request = raid5_sync_request,
8470 .resize = raid5_resize,
8472 .check_reshape = raid5_check_reshape,
8473 .start_reshape = raid5_start_reshape,
8474 .finish_reshape = raid5_finish_reshape,
8475 .quiesce = raid5_quiesce,
8476 .takeover = raid4_takeover,
8477 .congested = raid5_congested,
8478 .change_consistency_policy = raid5_change_consistency_policy,
8481 static int __init raid5_init(void)
8485 raid5_wq = alloc_workqueue("raid5wq",
8486 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8490 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8492 raid456_cpu_up_prepare,
8495 destroy_workqueue(raid5_wq);
8498 register_md_personality(&raid6_personality);
8499 register_md_personality(&raid5_personality);
8500 register_md_personality(&raid4_personality);
8504 static void raid5_exit(void)
8506 unregister_md_personality(&raid6_personality);
8507 unregister_md_personality(&raid5_personality);
8508 unregister_md_personality(&raid4_personality);
8509 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8510 destroy_workqueue(raid5_wq);
8513 module_init(raid5_init);
8514 module_exit(raid5_exit);
8515 MODULE_LICENSE("GPL");
8516 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8517 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8518 MODULE_ALIAS("md-raid5");
8519 MODULE_ALIAS("md-raid4");
8520 MODULE_ALIAS("md-level-5");
8521 MODULE_ALIAS("md-level-4");
8522 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8523 MODULE_ALIAS("md-raid6");
8524 MODULE_ALIAS("md-level-6");
8526 /* This used to be two separate modules, they were: */
8527 MODULE_ALIAS("raid5");
8528 MODULE_ALIAS("raid6");