1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * raid5.c : Multiple Devices driver for Linux
4 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
5 * Copyright (C) 1999, 2000 Ingo Molnar
6 * Copyright (C) 2002, 2003 H. Peter Anvin
8 * RAID-4/5/6 management functions.
9 * Thanks to Penguin Computing for making the RAID-6 development possible
10 * by donating a test server!
16 * The sequencing for updating the bitmap reliably is a little
17 * subtle (and I got it wrong the first time) so it deserves some
20 * We group bitmap updates into batches. Each batch has a number.
21 * We may write out several batches at once, but that isn't very important.
22 * conf->seq_write is the number of the last batch successfully written.
23 * conf->seq_flush is the number of the last batch that was closed to
25 * When we discover that we will need to write to any block in a stripe
26 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
27 * the number of the batch it will be in. This is seq_flush+1.
28 * When we are ready to do a write, if that batch hasn't been written yet,
29 * we plug the array and queue the stripe for later.
30 * When an unplug happens, we increment bm_flush, thus closing the current
32 * When we notice that bm_flush > bm_write, we write out all pending updates
33 * to the bitmap, and advance bm_write to where bm_flush was.
34 * This may occasionally write a bit out twice, but is sure never to
38 #include <linux/blkdev.h>
39 #include <linux/kthread.h>
40 #include <linux/raid/pq.h>
41 #include <linux/async_tx.h>
42 #include <linux/module.h>
43 #include <linux/async.h>
44 #include <linux/seq_file.h>
45 #include <linux/cpu.h>
46 #include <linux/slab.h>
47 #include <linux/ratelimit.h>
48 #include <linux/nodemask.h>
50 #include <trace/events/block.h>
51 #include <linux/list_sort.h>
56 #include "md-bitmap.h"
57 #include "raid5-log.h"
59 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
61 #define cpu_to_group(cpu) cpu_to_node(cpu)
62 #define ANY_GROUP NUMA_NO_NODE
64 static bool devices_handle_discard_safely = false;
65 module_param(devices_handle_discard_safely, bool, 0644);
66 MODULE_PARM_DESC(devices_handle_discard_safely,
67 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
68 static struct workqueue_struct *raid5_wq;
70 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
72 int hash = (sect >> RAID5_STRIPE_SHIFT(conf)) & HASH_MASK;
73 return &conf->stripe_hashtbl[hash];
76 static inline int stripe_hash_locks_hash(struct r5conf *conf, sector_t sect)
78 return (sect >> RAID5_STRIPE_SHIFT(conf)) & STRIPE_HASH_LOCKS_MASK;
81 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
83 spin_lock_irq(conf->hash_locks + hash);
84 spin_lock(&conf->device_lock);
87 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
89 spin_unlock(&conf->device_lock);
90 spin_unlock_irq(conf->hash_locks + hash);
93 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
96 spin_lock_irq(conf->hash_locks);
97 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
98 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
99 spin_lock(&conf->device_lock);
102 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
105 spin_unlock(&conf->device_lock);
106 for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
107 spin_unlock(conf->hash_locks + i);
108 spin_unlock_irq(conf->hash_locks);
111 /* Find first data disk in a raid6 stripe */
112 static inline int raid6_d0(struct stripe_head *sh)
115 /* ddf always start from first device */
117 /* md starts just after Q block */
118 if (sh->qd_idx == sh->disks - 1)
121 return sh->qd_idx + 1;
123 static inline int raid6_next_disk(int disk, int raid_disks)
126 return (disk < raid_disks) ? disk : 0;
129 /* When walking through the disks in a raid5, starting at raid6_d0,
130 * We need to map each disk to a 'slot', where the data disks are slot
131 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
132 * is raid_disks-1. This help does that mapping.
134 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
135 int *count, int syndrome_disks)
141 if (idx == sh->pd_idx)
142 return syndrome_disks;
143 if (idx == sh->qd_idx)
144 return syndrome_disks + 1;
150 static void print_raid5_conf (struct r5conf *conf);
152 static int stripe_operations_active(struct stripe_head *sh)
154 return sh->check_state || sh->reconstruct_state ||
155 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
156 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
159 static bool stripe_is_lowprio(struct stripe_head *sh)
161 return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
162 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
163 !test_bit(STRIPE_R5C_CACHING, &sh->state);
166 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
168 struct r5conf *conf = sh->raid_conf;
169 struct r5worker_group *group;
171 int i, cpu = sh->cpu;
173 if (!cpu_online(cpu)) {
174 cpu = cpumask_any(cpu_online_mask);
178 if (list_empty(&sh->lru)) {
179 struct r5worker_group *group;
180 group = conf->worker_groups + cpu_to_group(cpu);
181 if (stripe_is_lowprio(sh))
182 list_add_tail(&sh->lru, &group->loprio_list);
184 list_add_tail(&sh->lru, &group->handle_list);
185 group->stripes_cnt++;
189 if (conf->worker_cnt_per_group == 0) {
190 md_wakeup_thread(conf->mddev->thread);
194 group = conf->worker_groups + cpu_to_group(sh->cpu);
196 group->workers[0].working = true;
197 /* at least one worker should run to avoid race */
198 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
200 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
201 /* wakeup more workers */
202 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
203 if (group->workers[i].working == false) {
204 group->workers[i].working = true;
205 queue_work_on(sh->cpu, raid5_wq,
206 &group->workers[i].work);
212 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
213 struct list_head *temp_inactive_list)
216 int injournal = 0; /* number of date pages with R5_InJournal */
218 BUG_ON(!list_empty(&sh->lru));
219 BUG_ON(atomic_read(&conf->active_stripes)==0);
221 if (r5c_is_writeback(conf->log))
222 for (i = sh->disks; i--; )
223 if (test_bit(R5_InJournal, &sh->dev[i].flags))
226 * In the following cases, the stripe cannot be released to cached
227 * lists. Therefore, we make the stripe write out and set
229 * 1. when quiesce in r5c write back;
230 * 2. when resync is requested fot the stripe.
232 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
233 (conf->quiesce && r5c_is_writeback(conf->log) &&
234 !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
235 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
236 r5c_make_stripe_write_out(sh);
237 set_bit(STRIPE_HANDLE, &sh->state);
240 if (test_bit(STRIPE_HANDLE, &sh->state)) {
241 if (test_bit(STRIPE_DELAYED, &sh->state) &&
242 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
243 list_add_tail(&sh->lru, &conf->delayed_list);
244 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
245 sh->bm_seq - conf->seq_write > 0)
246 list_add_tail(&sh->lru, &conf->bitmap_list);
248 clear_bit(STRIPE_DELAYED, &sh->state);
249 clear_bit(STRIPE_BIT_DELAY, &sh->state);
250 if (conf->worker_cnt_per_group == 0) {
251 if (stripe_is_lowprio(sh))
252 list_add_tail(&sh->lru,
255 list_add_tail(&sh->lru,
258 raid5_wakeup_stripe_thread(sh);
262 md_wakeup_thread(conf->mddev->thread);
264 BUG_ON(stripe_operations_active(sh));
265 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
266 if (atomic_dec_return(&conf->preread_active_stripes)
268 md_wakeup_thread(conf->mddev->thread);
269 atomic_dec(&conf->active_stripes);
270 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
271 if (!r5c_is_writeback(conf->log))
272 list_add_tail(&sh->lru, temp_inactive_list);
274 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
276 list_add_tail(&sh->lru, temp_inactive_list);
277 else if (injournal == conf->raid_disks - conf->max_degraded) {
279 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
280 atomic_inc(&conf->r5c_cached_full_stripes);
281 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
282 atomic_dec(&conf->r5c_cached_partial_stripes);
283 list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
284 r5c_check_cached_full_stripe(conf);
287 * STRIPE_R5C_PARTIAL_STRIPE is set in
288 * r5c_try_caching_write(). No need to
291 list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
297 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
298 struct list_head *temp_inactive_list)
300 if (atomic_dec_and_test(&sh->count))
301 do_release_stripe(conf, sh, temp_inactive_list);
305 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
307 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
308 * given time. Adding stripes only takes device lock, while deleting stripes
309 * only takes hash lock.
311 static void release_inactive_stripe_list(struct r5conf *conf,
312 struct list_head *temp_inactive_list,
316 bool do_wakeup = false;
319 if (hash == NR_STRIPE_HASH_LOCKS) {
320 size = NR_STRIPE_HASH_LOCKS;
321 hash = NR_STRIPE_HASH_LOCKS - 1;
325 struct list_head *list = &temp_inactive_list[size - 1];
328 * We don't hold any lock here yet, raid5_get_active_stripe() might
329 * remove stripes from the list
331 if (!list_empty_careful(list)) {
332 spin_lock_irqsave(conf->hash_locks + hash, flags);
333 if (list_empty(conf->inactive_list + hash) &&
335 atomic_dec(&conf->empty_inactive_list_nr);
336 list_splice_tail_init(list, conf->inactive_list + hash);
338 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
345 wake_up(&conf->wait_for_stripe);
346 if (atomic_read(&conf->active_stripes) == 0)
347 wake_up(&conf->wait_for_quiescent);
348 if (conf->retry_read_aligned)
349 md_wakeup_thread(conf->mddev->thread);
353 /* should hold conf->device_lock already */
354 static int release_stripe_list(struct r5conf *conf,
355 struct list_head *temp_inactive_list)
357 struct stripe_head *sh, *t;
359 struct llist_node *head;
361 head = llist_del_all(&conf->released_stripes);
362 head = llist_reverse_order(head);
363 llist_for_each_entry_safe(sh, t, head, release_list) {
366 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
368 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
370 * Don't worry the bit is set here, because if the bit is set
371 * again, the count is always > 1. This is true for
372 * STRIPE_ON_UNPLUG_LIST bit too.
374 hash = sh->hash_lock_index;
375 __release_stripe(conf, sh, &temp_inactive_list[hash]);
382 void raid5_release_stripe(struct stripe_head *sh)
384 struct r5conf *conf = sh->raid_conf;
386 struct list_head list;
390 /* Avoid release_list until the last reference.
392 if (atomic_add_unless(&sh->count, -1, 1))
395 if (unlikely(!conf->mddev->thread) ||
396 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
398 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
400 md_wakeup_thread(conf->mddev->thread);
403 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
404 if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
405 INIT_LIST_HEAD(&list);
406 hash = sh->hash_lock_index;
407 do_release_stripe(conf, sh, &list);
408 spin_unlock_irqrestore(&conf->device_lock, flags);
409 release_inactive_stripe_list(conf, &list, hash);
413 static inline void remove_hash(struct stripe_head *sh)
415 pr_debug("remove_hash(), stripe %llu\n",
416 (unsigned long long)sh->sector);
418 hlist_del_init(&sh->hash);
421 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
423 struct hlist_head *hp = stripe_hash(conf, sh->sector);
425 pr_debug("insert_hash(), stripe %llu\n",
426 (unsigned long long)sh->sector);
428 hlist_add_head(&sh->hash, hp);
431 /* find an idle stripe, make sure it is unhashed, and return it. */
432 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
434 struct stripe_head *sh = NULL;
435 struct list_head *first;
437 if (list_empty(conf->inactive_list + hash))
439 first = (conf->inactive_list + hash)->next;
440 sh = list_entry(first, struct stripe_head, lru);
441 list_del_init(first);
443 atomic_inc(&conf->active_stripes);
444 BUG_ON(hash != sh->hash_lock_index);
445 if (list_empty(conf->inactive_list + hash))
446 atomic_inc(&conf->empty_inactive_list_nr);
451 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
452 static void free_stripe_pages(struct stripe_head *sh)
457 /* Have not allocate page pool */
461 for (i = 0; i < sh->nr_pages; i++) {
469 static int alloc_stripe_pages(struct stripe_head *sh, gfp_t gfp)
474 for (i = 0; i < sh->nr_pages; i++) {
475 /* The page have allocated. */
481 free_stripe_pages(sh);
490 init_stripe_shared_pages(struct stripe_head *sh, struct r5conf *conf, int disks)
497 /* Each of the sh->dev[i] need one conf->stripe_size */
498 cnt = PAGE_SIZE / conf->stripe_size;
499 nr_pages = (disks + cnt - 1) / cnt;
501 sh->pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
504 sh->nr_pages = nr_pages;
505 sh->stripes_per_page = cnt;
510 static void shrink_buffers(struct stripe_head *sh)
513 int num = sh->raid_conf->pool_size;
515 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
516 for (i = 0; i < num ; i++) {
519 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
523 sh->dev[i].page = NULL;
527 for (i = 0; i < num; i++)
528 sh->dev[i].page = NULL;
529 free_stripe_pages(sh); /* Free pages */
533 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
536 int num = sh->raid_conf->pool_size;
538 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
539 for (i = 0; i < num; i++) {
542 if (!(page = alloc_page(gfp))) {
545 sh->dev[i].page = page;
546 sh->dev[i].orig_page = page;
547 sh->dev[i].offset = 0;
550 if (alloc_stripe_pages(sh, gfp))
553 for (i = 0; i < num; i++) {
554 sh->dev[i].page = raid5_get_dev_page(sh, i);
555 sh->dev[i].orig_page = sh->dev[i].page;
556 sh->dev[i].offset = raid5_get_page_offset(sh, i);
562 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
563 struct stripe_head *sh);
565 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
567 struct r5conf *conf = sh->raid_conf;
570 BUG_ON(atomic_read(&sh->count) != 0);
571 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
572 BUG_ON(stripe_operations_active(sh));
573 BUG_ON(sh->batch_head);
575 pr_debug("init_stripe called, stripe %llu\n",
576 (unsigned long long)sector);
578 seq = read_seqcount_begin(&conf->gen_lock);
579 sh->generation = conf->generation - previous;
580 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
582 stripe_set_idx(sector, conf, previous, sh);
585 for (i = sh->disks; i--; ) {
586 struct r5dev *dev = &sh->dev[i];
588 if (dev->toread || dev->read || dev->towrite || dev->written ||
589 test_bit(R5_LOCKED, &dev->flags)) {
590 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
591 (unsigned long long)sh->sector, i, dev->toread,
592 dev->read, dev->towrite, dev->written,
593 test_bit(R5_LOCKED, &dev->flags));
597 dev->sector = raid5_compute_blocknr(sh, i, previous);
599 if (read_seqcount_retry(&conf->gen_lock, seq))
601 sh->overwrite_disks = 0;
602 insert_hash(conf, sh);
603 sh->cpu = smp_processor_id();
604 set_bit(STRIPE_BATCH_READY, &sh->state);
607 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
610 struct stripe_head *sh;
612 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
613 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
614 if (sh->sector == sector && sh->generation == generation)
616 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
621 * Need to check if array has failed when deciding whether to:
623 * - remove non-faulty devices
626 * This determination is simple when no reshape is happening.
627 * However if there is a reshape, we need to carefully check
628 * both the before and after sections.
629 * This is because some failed devices may only affect one
630 * of the two sections, and some non-in_sync devices may
631 * be insync in the section most affected by failed devices.
633 int raid5_calc_degraded(struct r5conf *conf)
635 int degraded, degraded2;
640 for (i = 0; i < conf->previous_raid_disks; i++) {
641 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
642 if (rdev && test_bit(Faulty, &rdev->flags))
643 rdev = rcu_dereference(conf->disks[i].replacement);
644 if (!rdev || test_bit(Faulty, &rdev->flags))
646 else if (test_bit(In_sync, &rdev->flags))
649 /* not in-sync or faulty.
650 * If the reshape increases the number of devices,
651 * this is being recovered by the reshape, so
652 * this 'previous' section is not in_sync.
653 * If the number of devices is being reduced however,
654 * the device can only be part of the array if
655 * we are reverting a reshape, so this section will
658 if (conf->raid_disks >= conf->previous_raid_disks)
662 if (conf->raid_disks == conf->previous_raid_disks)
666 for (i = 0; i < conf->raid_disks; i++) {
667 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
668 if (rdev && test_bit(Faulty, &rdev->flags))
669 rdev = rcu_dereference(conf->disks[i].replacement);
670 if (!rdev || test_bit(Faulty, &rdev->flags))
672 else if (test_bit(In_sync, &rdev->flags))
675 /* not in-sync or faulty.
676 * If reshape increases the number of devices, this
677 * section has already been recovered, else it
678 * almost certainly hasn't.
680 if (conf->raid_disks <= conf->previous_raid_disks)
684 if (degraded2 > degraded)
689 static bool has_failed(struct r5conf *conf)
691 int degraded = conf->mddev->degraded;
693 if (test_bit(MD_BROKEN, &conf->mddev->flags))
696 if (conf->mddev->reshape_position != MaxSector)
697 degraded = raid5_calc_degraded(conf);
699 return degraded > conf->max_degraded;
703 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
704 int previous, int noblock, int noquiesce)
706 struct stripe_head *sh;
707 int hash = stripe_hash_locks_hash(conf, sector);
708 int inc_empty_inactive_list_flag;
710 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
712 spin_lock_irq(conf->hash_locks + hash);
715 wait_event_lock_irq(conf->wait_for_quiescent,
716 conf->quiesce == 0 || noquiesce,
717 *(conf->hash_locks + hash));
718 sh = __find_stripe(conf, sector, conf->generation - previous);
720 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
721 sh = get_free_stripe(conf, hash);
722 if (!sh && !test_bit(R5_DID_ALLOC,
724 set_bit(R5_ALLOC_MORE,
727 if (noblock && sh == NULL)
730 r5c_check_stripe_cache_usage(conf);
732 set_bit(R5_INACTIVE_BLOCKED,
734 r5l_wake_reclaim(conf->log, 0);
736 conf->wait_for_stripe,
737 !list_empty(conf->inactive_list + hash) &&
738 (atomic_read(&conf->active_stripes)
739 < (conf->max_nr_stripes * 3 / 4)
740 || !test_bit(R5_INACTIVE_BLOCKED,
741 &conf->cache_state)),
742 *(conf->hash_locks + hash));
743 clear_bit(R5_INACTIVE_BLOCKED,
746 init_stripe(sh, sector, previous);
747 atomic_inc(&sh->count);
749 } else if (!atomic_inc_not_zero(&sh->count)) {
750 spin_lock(&conf->device_lock);
751 if (!atomic_read(&sh->count)) {
752 if (!test_bit(STRIPE_HANDLE, &sh->state))
753 atomic_inc(&conf->active_stripes);
754 BUG_ON(list_empty(&sh->lru) &&
755 !test_bit(STRIPE_EXPANDING, &sh->state));
756 inc_empty_inactive_list_flag = 0;
757 if (!list_empty(conf->inactive_list + hash))
758 inc_empty_inactive_list_flag = 1;
759 list_del_init(&sh->lru);
760 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
761 atomic_inc(&conf->empty_inactive_list_nr);
763 sh->group->stripes_cnt--;
767 atomic_inc(&sh->count);
768 spin_unlock(&conf->device_lock);
770 } while (sh == NULL);
772 spin_unlock_irq(conf->hash_locks + hash);
776 static bool is_full_stripe_write(struct stripe_head *sh)
778 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
779 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
782 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
783 __acquires(&sh1->stripe_lock)
784 __acquires(&sh2->stripe_lock)
787 spin_lock_irq(&sh2->stripe_lock);
788 spin_lock_nested(&sh1->stripe_lock, 1);
790 spin_lock_irq(&sh1->stripe_lock);
791 spin_lock_nested(&sh2->stripe_lock, 1);
795 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
796 __releases(&sh1->stripe_lock)
797 __releases(&sh2->stripe_lock)
799 spin_unlock(&sh1->stripe_lock);
800 spin_unlock_irq(&sh2->stripe_lock);
803 /* Only freshly new full stripe normal write stripe can be added to a batch list */
804 static bool stripe_can_batch(struct stripe_head *sh)
806 struct r5conf *conf = sh->raid_conf;
808 if (raid5_has_log(conf) || raid5_has_ppl(conf))
810 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
811 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
812 is_full_stripe_write(sh);
815 /* we only do back search */
816 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
818 struct stripe_head *head;
819 sector_t head_sector, tmp_sec;
822 int inc_empty_inactive_list_flag;
824 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
825 tmp_sec = sh->sector;
826 if (!sector_div(tmp_sec, conf->chunk_sectors))
828 head_sector = sh->sector - RAID5_STRIPE_SECTORS(conf);
830 hash = stripe_hash_locks_hash(conf, head_sector);
831 spin_lock_irq(conf->hash_locks + hash);
832 head = __find_stripe(conf, head_sector, conf->generation);
833 if (head && !atomic_inc_not_zero(&head->count)) {
834 spin_lock(&conf->device_lock);
835 if (!atomic_read(&head->count)) {
836 if (!test_bit(STRIPE_HANDLE, &head->state))
837 atomic_inc(&conf->active_stripes);
838 BUG_ON(list_empty(&head->lru) &&
839 !test_bit(STRIPE_EXPANDING, &head->state));
840 inc_empty_inactive_list_flag = 0;
841 if (!list_empty(conf->inactive_list + hash))
842 inc_empty_inactive_list_flag = 1;
843 list_del_init(&head->lru);
844 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
845 atomic_inc(&conf->empty_inactive_list_nr);
847 head->group->stripes_cnt--;
851 atomic_inc(&head->count);
852 spin_unlock(&conf->device_lock);
854 spin_unlock_irq(conf->hash_locks + hash);
858 if (!stripe_can_batch(head))
861 lock_two_stripes(head, sh);
862 /* clear_batch_ready clear the flag */
863 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
870 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
872 if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
873 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
876 if (head->batch_head) {
877 spin_lock(&head->batch_head->batch_lock);
878 /* This batch list is already running */
879 if (!stripe_can_batch(head)) {
880 spin_unlock(&head->batch_head->batch_lock);
884 * We must assign batch_head of this stripe within the
885 * batch_lock, otherwise clear_batch_ready of batch head
886 * stripe could clear BATCH_READY bit of this stripe and
887 * this stripe->batch_head doesn't get assigned, which
888 * could confuse clear_batch_ready for this stripe
890 sh->batch_head = head->batch_head;
893 * at this point, head's BATCH_READY could be cleared, but we
894 * can still add the stripe to batch list
896 list_add(&sh->batch_list, &head->batch_list);
897 spin_unlock(&head->batch_head->batch_lock);
899 head->batch_head = head;
900 sh->batch_head = head->batch_head;
901 spin_lock(&head->batch_lock);
902 list_add_tail(&sh->batch_list, &head->batch_list);
903 spin_unlock(&head->batch_lock);
906 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
907 if (atomic_dec_return(&conf->preread_active_stripes)
909 md_wakeup_thread(conf->mddev->thread);
911 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
912 int seq = sh->bm_seq;
913 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
914 sh->batch_head->bm_seq > seq)
915 seq = sh->batch_head->bm_seq;
916 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
917 sh->batch_head->bm_seq = seq;
920 atomic_inc(&sh->count);
922 unlock_two_stripes(head, sh);
924 raid5_release_stripe(head);
927 /* Determine if 'data_offset' or 'new_data_offset' should be used
928 * in this stripe_head.
930 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
932 sector_t progress = conf->reshape_progress;
933 /* Need a memory barrier to make sure we see the value
934 * of conf->generation, or ->data_offset that was set before
935 * reshape_progress was updated.
938 if (progress == MaxSector)
940 if (sh->generation == conf->generation - 1)
942 /* We are in a reshape, and this is a new-generation stripe,
943 * so use new_data_offset.
948 static void dispatch_bio_list(struct bio_list *tmp)
952 while ((bio = bio_list_pop(tmp)))
953 submit_bio_noacct(bio);
956 static int cmp_stripe(void *priv, const struct list_head *a,
957 const struct list_head *b)
959 const struct r5pending_data *da = list_entry(a,
960 struct r5pending_data, sibling);
961 const struct r5pending_data *db = list_entry(b,
962 struct r5pending_data, sibling);
963 if (da->sector > db->sector)
965 if (da->sector < db->sector)
970 static void dispatch_defer_bios(struct r5conf *conf, int target,
971 struct bio_list *list)
973 struct r5pending_data *data;
974 struct list_head *first, *next = NULL;
977 if (conf->pending_data_cnt == 0)
980 list_sort(NULL, &conf->pending_list, cmp_stripe);
982 first = conf->pending_list.next;
984 /* temporarily move the head */
985 if (conf->next_pending_data)
986 list_move_tail(&conf->pending_list,
987 &conf->next_pending_data->sibling);
989 while (!list_empty(&conf->pending_list)) {
990 data = list_first_entry(&conf->pending_list,
991 struct r5pending_data, sibling);
992 if (&data->sibling == first)
993 first = data->sibling.next;
994 next = data->sibling.next;
996 bio_list_merge(list, &data->bios);
997 list_move(&data->sibling, &conf->free_list);
1002 conf->pending_data_cnt -= cnt;
1003 BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
1005 if (next != &conf->pending_list)
1006 conf->next_pending_data = list_entry(next,
1007 struct r5pending_data, sibling);
1009 conf->next_pending_data = NULL;
1010 /* list isn't empty */
1011 if (first != &conf->pending_list)
1012 list_move_tail(&conf->pending_list, first);
1015 static void flush_deferred_bios(struct r5conf *conf)
1017 struct bio_list tmp = BIO_EMPTY_LIST;
1019 if (conf->pending_data_cnt == 0)
1022 spin_lock(&conf->pending_bios_lock);
1023 dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
1024 BUG_ON(conf->pending_data_cnt != 0);
1025 spin_unlock(&conf->pending_bios_lock);
1027 dispatch_bio_list(&tmp);
1030 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
1031 struct bio_list *bios)
1033 struct bio_list tmp = BIO_EMPTY_LIST;
1034 struct r5pending_data *ent;
1036 spin_lock(&conf->pending_bios_lock);
1037 ent = list_first_entry(&conf->free_list, struct r5pending_data,
1039 list_move_tail(&ent->sibling, &conf->pending_list);
1040 ent->sector = sector;
1041 bio_list_init(&ent->bios);
1042 bio_list_merge(&ent->bios, bios);
1043 conf->pending_data_cnt++;
1044 if (conf->pending_data_cnt >= PENDING_IO_MAX)
1045 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
1047 spin_unlock(&conf->pending_bios_lock);
1049 dispatch_bio_list(&tmp);
1053 raid5_end_read_request(struct bio *bi);
1055 raid5_end_write_request(struct bio *bi);
1057 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
1059 struct r5conf *conf = sh->raid_conf;
1060 int i, disks = sh->disks;
1061 struct stripe_head *head_sh = sh;
1062 struct bio_list pending_bios = BIO_EMPTY_LIST;
1067 if (log_stripe(sh, s) == 0)
1070 should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1072 for (i = disks; i--; ) {
1073 int op, op_flags = 0;
1074 int replace_only = 0;
1075 struct bio *bi, *rbi;
1076 struct md_rdev *rdev, *rrdev = NULL;
1079 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1081 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1083 if (test_bit(R5_Discard, &sh->dev[i].flags))
1084 op = REQ_OP_DISCARD;
1085 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1087 else if (test_and_clear_bit(R5_WantReplace,
1088 &sh->dev[i].flags)) {
1093 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1094 op_flags |= REQ_SYNC;
1097 bi = &sh->dev[i].req;
1098 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1101 rrdev = rcu_dereference(conf->disks[i].replacement);
1102 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1103 rdev = rcu_dereference(conf->disks[i].rdev);
1108 if (op_is_write(op)) {
1112 /* We raced and saw duplicates */
1115 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1120 if (rdev && test_bit(Faulty, &rdev->flags))
1123 atomic_inc(&rdev->nr_pending);
1124 if (rrdev && test_bit(Faulty, &rrdev->flags))
1127 atomic_inc(&rrdev->nr_pending);
1130 /* We have already checked bad blocks for reads. Now
1131 * need to check for writes. We never accept write errors
1132 * on the replacement, so we don't to check rrdev.
1134 while (op_is_write(op) && rdev &&
1135 test_bit(WriteErrorSeen, &rdev->flags)) {
1138 int bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
1139 &first_bad, &bad_sectors);
1144 set_bit(BlockedBadBlocks, &rdev->flags);
1145 if (!conf->mddev->external &&
1146 conf->mddev->sb_flags) {
1147 /* It is very unlikely, but we might
1148 * still need to write out the
1149 * bad block log - better give it
1151 md_check_recovery(conf->mddev);
1154 * Because md_wait_for_blocked_rdev
1155 * will dec nr_pending, we must
1156 * increment it first.
1158 atomic_inc(&rdev->nr_pending);
1159 md_wait_for_blocked_rdev(rdev, conf->mddev);
1161 /* Acknowledged bad block - skip the write */
1162 rdev_dec_pending(rdev, conf->mddev);
1168 if (s->syncing || s->expanding || s->expanded
1170 md_sync_acct(rdev->bdev, RAID5_STRIPE_SECTORS(conf));
1172 set_bit(STRIPE_IO_STARTED, &sh->state);
1174 bio_set_dev(bi, rdev->bdev);
1175 bio_set_op_attrs(bi, op, op_flags);
1176 bi->bi_end_io = op_is_write(op)
1177 ? raid5_end_write_request
1178 : raid5_end_read_request;
1179 bi->bi_private = sh;
1181 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1182 __func__, (unsigned long long)sh->sector,
1184 atomic_inc(&sh->count);
1186 atomic_inc(&head_sh->count);
1187 if (use_new_offset(conf, sh))
1188 bi->bi_iter.bi_sector = (sh->sector
1189 + rdev->new_data_offset);
1191 bi->bi_iter.bi_sector = (sh->sector
1192 + rdev->data_offset);
1193 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1194 bi->bi_opf |= REQ_NOMERGE;
1196 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1197 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1199 if (!op_is_write(op) &&
1200 test_bit(R5_InJournal, &sh->dev[i].flags))
1202 * issuing read for a page in journal, this
1203 * must be preparing for prexor in rmw; read
1204 * the data into orig_page
1206 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1208 sh->dev[i].vec.bv_page = sh->dev[i].page;
1210 bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1211 bi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1212 bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1213 bi->bi_write_hint = sh->dev[i].write_hint;
1215 sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1217 * If this is discard request, set bi_vcnt 0. We don't
1218 * want to confuse SCSI because SCSI will replace payload
1220 if (op == REQ_OP_DISCARD)
1223 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1225 if (conf->mddev->gendisk)
1226 trace_block_bio_remap(bi,
1227 disk_devt(conf->mddev->gendisk),
1229 if (should_defer && op_is_write(op))
1230 bio_list_add(&pending_bios, bi);
1232 submit_bio_noacct(bi);
1235 if (s->syncing || s->expanding || s->expanded
1237 md_sync_acct(rrdev->bdev, RAID5_STRIPE_SECTORS(conf));
1239 set_bit(STRIPE_IO_STARTED, &sh->state);
1241 bio_set_dev(rbi, rrdev->bdev);
1242 bio_set_op_attrs(rbi, op, op_flags);
1243 BUG_ON(!op_is_write(op));
1244 rbi->bi_end_io = raid5_end_write_request;
1245 rbi->bi_private = sh;
1247 pr_debug("%s: for %llu schedule op %d on "
1248 "replacement disc %d\n",
1249 __func__, (unsigned long long)sh->sector,
1251 atomic_inc(&sh->count);
1253 atomic_inc(&head_sh->count);
1254 if (use_new_offset(conf, sh))
1255 rbi->bi_iter.bi_sector = (sh->sector
1256 + rrdev->new_data_offset);
1258 rbi->bi_iter.bi_sector = (sh->sector
1259 + rrdev->data_offset);
1260 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1261 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1262 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1264 rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1265 rbi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1266 rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1267 rbi->bi_write_hint = sh->dev[i].write_hint;
1268 sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1270 * If this is discard request, set bi_vcnt 0. We don't
1271 * want to confuse SCSI because SCSI will replace payload
1273 if (op == REQ_OP_DISCARD)
1275 if (conf->mddev->gendisk)
1276 trace_block_bio_remap(rbi,
1277 disk_devt(conf->mddev->gendisk),
1279 if (should_defer && op_is_write(op))
1280 bio_list_add(&pending_bios, rbi);
1282 submit_bio_noacct(rbi);
1284 if (!rdev && !rrdev) {
1285 if (op_is_write(op))
1286 set_bit(STRIPE_DEGRADED, &sh->state);
1287 pr_debug("skip op %d on disc %d for sector %llu\n",
1288 bi->bi_opf, i, (unsigned long long)sh->sector);
1289 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1290 set_bit(STRIPE_HANDLE, &sh->state);
1293 if (!head_sh->batch_head)
1295 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1301 if (should_defer && !bio_list_empty(&pending_bios))
1302 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1305 static struct dma_async_tx_descriptor *
1306 async_copy_data(int frombio, struct bio *bio, struct page **page,
1307 unsigned int poff, sector_t sector, struct dma_async_tx_descriptor *tx,
1308 struct stripe_head *sh, int no_skipcopy)
1311 struct bvec_iter iter;
1312 struct page *bio_page;
1314 struct async_submit_ctl submit;
1315 enum async_tx_flags flags = 0;
1316 struct r5conf *conf = sh->raid_conf;
1318 if (bio->bi_iter.bi_sector >= sector)
1319 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1321 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1324 flags |= ASYNC_TX_FENCE;
1325 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1327 bio_for_each_segment(bvl, bio, iter) {
1328 int len = bvl.bv_len;
1332 if (page_offset < 0) {
1333 b_offset = -page_offset;
1334 page_offset += b_offset;
1338 if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
1339 clen = RAID5_STRIPE_SIZE(conf) - page_offset;
1344 b_offset += bvl.bv_offset;
1345 bio_page = bvl.bv_page;
1347 if (conf->skip_copy &&
1348 b_offset == 0 && page_offset == 0 &&
1349 clen == RAID5_STRIPE_SIZE(conf) &&
1353 tx = async_memcpy(*page, bio_page, page_offset + poff,
1354 b_offset, clen, &submit);
1356 tx = async_memcpy(bio_page, *page, b_offset,
1357 page_offset + poff, clen, &submit);
1359 /* chain the operations */
1360 submit.depend_tx = tx;
1362 if (clen < len) /* hit end of page */
1370 static void ops_complete_biofill(void *stripe_head_ref)
1372 struct stripe_head *sh = stripe_head_ref;
1374 struct r5conf *conf = sh->raid_conf;
1376 pr_debug("%s: stripe %llu\n", __func__,
1377 (unsigned long long)sh->sector);
1379 /* clear completed biofills */
1380 for (i = sh->disks; i--; ) {
1381 struct r5dev *dev = &sh->dev[i];
1383 /* acknowledge completion of a biofill operation */
1384 /* and check if we need to reply to a read request,
1385 * new R5_Wantfill requests are held off until
1386 * !STRIPE_BIOFILL_RUN
1388 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1389 struct bio *rbi, *rbi2;
1394 while (rbi && rbi->bi_iter.bi_sector <
1395 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1396 rbi2 = r5_next_bio(conf, rbi, dev->sector);
1402 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1404 set_bit(STRIPE_HANDLE, &sh->state);
1405 raid5_release_stripe(sh);
1408 static void ops_run_biofill(struct stripe_head *sh)
1410 struct dma_async_tx_descriptor *tx = NULL;
1411 struct async_submit_ctl submit;
1413 struct r5conf *conf = sh->raid_conf;
1415 BUG_ON(sh->batch_head);
1416 pr_debug("%s: stripe %llu\n", __func__,
1417 (unsigned long long)sh->sector);
1419 for (i = sh->disks; i--; ) {
1420 struct r5dev *dev = &sh->dev[i];
1421 if (test_bit(R5_Wantfill, &dev->flags)) {
1423 spin_lock_irq(&sh->stripe_lock);
1424 dev->read = rbi = dev->toread;
1426 spin_unlock_irq(&sh->stripe_lock);
1427 while (rbi && rbi->bi_iter.bi_sector <
1428 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1429 tx = async_copy_data(0, rbi, &dev->page,
1431 dev->sector, tx, sh, 0);
1432 rbi = r5_next_bio(conf, rbi, dev->sector);
1437 atomic_inc(&sh->count);
1438 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1439 async_trigger_callback(&submit);
1442 static void mark_target_uptodate(struct stripe_head *sh, int target)
1449 tgt = &sh->dev[target];
1450 set_bit(R5_UPTODATE, &tgt->flags);
1451 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1452 clear_bit(R5_Wantcompute, &tgt->flags);
1455 static void ops_complete_compute(void *stripe_head_ref)
1457 struct stripe_head *sh = stripe_head_ref;
1459 pr_debug("%s: stripe %llu\n", __func__,
1460 (unsigned long long)sh->sector);
1462 /* mark the computed target(s) as uptodate */
1463 mark_target_uptodate(sh, sh->ops.target);
1464 mark_target_uptodate(sh, sh->ops.target2);
1466 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1467 if (sh->check_state == check_state_compute_run)
1468 sh->check_state = check_state_compute_result;
1469 set_bit(STRIPE_HANDLE, &sh->state);
1470 raid5_release_stripe(sh);
1473 /* return a pointer to the address conversion region of the scribble buffer */
1474 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1476 return percpu->scribble + i * percpu->scribble_obj_size;
1479 /* return a pointer to the address conversion region of the scribble buffer */
1480 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1481 struct raid5_percpu *percpu, int i)
1483 return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1487 * Return a pointer to record offset address.
1489 static unsigned int *
1490 to_addr_offs(struct stripe_head *sh, struct raid5_percpu *percpu)
1492 return (unsigned int *) (to_addr_conv(sh, percpu, 0) + sh->disks + 2);
1495 static struct dma_async_tx_descriptor *
1496 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1498 int disks = sh->disks;
1499 struct page **xor_srcs = to_addr_page(percpu, 0);
1500 unsigned int *off_srcs = to_addr_offs(sh, percpu);
1501 int target = sh->ops.target;
1502 struct r5dev *tgt = &sh->dev[target];
1503 struct page *xor_dest = tgt->page;
1504 unsigned int off_dest = tgt->offset;
1506 struct dma_async_tx_descriptor *tx;
1507 struct async_submit_ctl submit;
1510 BUG_ON(sh->batch_head);
1512 pr_debug("%s: stripe %llu block: %d\n",
1513 __func__, (unsigned long long)sh->sector, target);
1514 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1516 for (i = disks; i--; ) {
1518 off_srcs[count] = sh->dev[i].offset;
1519 xor_srcs[count++] = sh->dev[i].page;
1523 atomic_inc(&sh->count);
1525 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1526 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1527 if (unlikely(count == 1))
1528 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
1529 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1531 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1532 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1537 /* set_syndrome_sources - populate source buffers for gen_syndrome
1538 * @srcs - (struct page *) array of size sh->disks
1539 * @offs - (unsigned int) array of offset for each page
1540 * @sh - stripe_head to parse
1542 * Populates srcs in proper layout order for the stripe and returns the
1543 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1544 * destination buffer is recorded in srcs[count] and the Q destination
1545 * is recorded in srcs[count+1]].
1547 static int set_syndrome_sources(struct page **srcs,
1549 struct stripe_head *sh,
1552 int disks = sh->disks;
1553 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1554 int d0_idx = raid6_d0(sh);
1558 for (i = 0; i < disks; i++)
1564 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1565 struct r5dev *dev = &sh->dev[i];
1567 if (i == sh->qd_idx || i == sh->pd_idx ||
1568 (srctype == SYNDROME_SRC_ALL) ||
1569 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1570 (test_bit(R5_Wantdrain, &dev->flags) ||
1571 test_bit(R5_InJournal, &dev->flags))) ||
1572 (srctype == SYNDROME_SRC_WRITTEN &&
1574 test_bit(R5_InJournal, &dev->flags)))) {
1575 if (test_bit(R5_InJournal, &dev->flags))
1576 srcs[slot] = sh->dev[i].orig_page;
1578 srcs[slot] = sh->dev[i].page;
1580 * For R5_InJournal, PAGE_SIZE must be 4KB and will
1581 * not shared page. In that case, dev[i].offset
1584 offs[slot] = sh->dev[i].offset;
1586 i = raid6_next_disk(i, disks);
1587 } while (i != d0_idx);
1589 return syndrome_disks;
1592 static struct dma_async_tx_descriptor *
1593 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1595 int disks = sh->disks;
1596 struct page **blocks = to_addr_page(percpu, 0);
1597 unsigned int *offs = to_addr_offs(sh, percpu);
1599 int qd_idx = sh->qd_idx;
1600 struct dma_async_tx_descriptor *tx;
1601 struct async_submit_ctl submit;
1604 unsigned int dest_off;
1608 BUG_ON(sh->batch_head);
1609 if (sh->ops.target < 0)
1610 target = sh->ops.target2;
1611 else if (sh->ops.target2 < 0)
1612 target = sh->ops.target;
1614 /* we should only have one valid target */
1617 pr_debug("%s: stripe %llu block: %d\n",
1618 __func__, (unsigned long long)sh->sector, target);
1620 tgt = &sh->dev[target];
1621 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1623 dest_off = tgt->offset;
1625 atomic_inc(&sh->count);
1627 if (target == qd_idx) {
1628 count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1629 blocks[count] = NULL; /* regenerating p is not necessary */
1630 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1631 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1632 ops_complete_compute, sh,
1633 to_addr_conv(sh, percpu, 0));
1634 tx = async_gen_syndrome(blocks, offs, count+2,
1635 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1637 /* Compute any data- or p-drive using XOR */
1639 for (i = disks; i-- ; ) {
1640 if (i == target || i == qd_idx)
1642 offs[count] = sh->dev[i].offset;
1643 blocks[count++] = sh->dev[i].page;
1646 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1647 NULL, ops_complete_compute, sh,
1648 to_addr_conv(sh, percpu, 0));
1649 tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1650 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1656 static struct dma_async_tx_descriptor *
1657 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1659 int i, count, disks = sh->disks;
1660 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1661 int d0_idx = raid6_d0(sh);
1662 int faila = -1, failb = -1;
1663 int target = sh->ops.target;
1664 int target2 = sh->ops.target2;
1665 struct r5dev *tgt = &sh->dev[target];
1666 struct r5dev *tgt2 = &sh->dev[target2];
1667 struct dma_async_tx_descriptor *tx;
1668 struct page **blocks = to_addr_page(percpu, 0);
1669 unsigned int *offs = to_addr_offs(sh, percpu);
1670 struct async_submit_ctl submit;
1672 BUG_ON(sh->batch_head);
1673 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1674 __func__, (unsigned long long)sh->sector, target, target2);
1675 BUG_ON(target < 0 || target2 < 0);
1676 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1677 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1679 /* we need to open-code set_syndrome_sources to handle the
1680 * slot number conversion for 'faila' and 'failb'
1682 for (i = 0; i < disks ; i++) {
1689 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1691 offs[slot] = sh->dev[i].offset;
1692 blocks[slot] = sh->dev[i].page;
1698 i = raid6_next_disk(i, disks);
1699 } while (i != d0_idx);
1701 BUG_ON(faila == failb);
1704 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1705 __func__, (unsigned long long)sh->sector, faila, failb);
1707 atomic_inc(&sh->count);
1709 if (failb == syndrome_disks+1) {
1710 /* Q disk is one of the missing disks */
1711 if (faila == syndrome_disks) {
1712 /* Missing P+Q, just recompute */
1713 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1714 ops_complete_compute, sh,
1715 to_addr_conv(sh, percpu, 0));
1716 return async_gen_syndrome(blocks, offs, syndrome_disks+2,
1717 RAID5_STRIPE_SIZE(sh->raid_conf),
1721 unsigned int dest_off;
1723 int qd_idx = sh->qd_idx;
1725 /* Missing D+Q: recompute D from P, then recompute Q */
1726 if (target == qd_idx)
1727 data_target = target2;
1729 data_target = target;
1732 for (i = disks; i-- ; ) {
1733 if (i == data_target || i == qd_idx)
1735 offs[count] = sh->dev[i].offset;
1736 blocks[count++] = sh->dev[i].page;
1738 dest = sh->dev[data_target].page;
1739 dest_off = sh->dev[data_target].offset;
1740 init_async_submit(&submit,
1741 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1743 to_addr_conv(sh, percpu, 0));
1744 tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1745 RAID5_STRIPE_SIZE(sh->raid_conf),
1748 count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1749 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1750 ops_complete_compute, sh,
1751 to_addr_conv(sh, percpu, 0));
1752 return async_gen_syndrome(blocks, offs, count+2,
1753 RAID5_STRIPE_SIZE(sh->raid_conf),
1757 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1758 ops_complete_compute, sh,
1759 to_addr_conv(sh, percpu, 0));
1760 if (failb == syndrome_disks) {
1761 /* We're missing D+P. */
1762 return async_raid6_datap_recov(syndrome_disks+2,
1763 RAID5_STRIPE_SIZE(sh->raid_conf),
1765 blocks, offs, &submit);
1767 /* We're missing D+D. */
1768 return async_raid6_2data_recov(syndrome_disks+2,
1769 RAID5_STRIPE_SIZE(sh->raid_conf),
1771 blocks, offs, &submit);
1776 static void ops_complete_prexor(void *stripe_head_ref)
1778 struct stripe_head *sh = stripe_head_ref;
1780 pr_debug("%s: stripe %llu\n", __func__,
1781 (unsigned long long)sh->sector);
1783 if (r5c_is_writeback(sh->raid_conf->log))
1785 * raid5-cache write back uses orig_page during prexor.
1786 * After prexor, it is time to free orig_page
1788 r5c_release_extra_page(sh);
1791 static struct dma_async_tx_descriptor *
1792 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1793 struct dma_async_tx_descriptor *tx)
1795 int disks = sh->disks;
1796 struct page **xor_srcs = to_addr_page(percpu, 0);
1797 unsigned int *off_srcs = to_addr_offs(sh, percpu);
1798 int count = 0, pd_idx = sh->pd_idx, i;
1799 struct async_submit_ctl submit;
1801 /* existing parity data subtracted */
1802 unsigned int off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
1803 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1805 BUG_ON(sh->batch_head);
1806 pr_debug("%s: stripe %llu\n", __func__,
1807 (unsigned long long)sh->sector);
1809 for (i = disks; i--; ) {
1810 struct r5dev *dev = &sh->dev[i];
1811 /* Only process blocks that are known to be uptodate */
1812 if (test_bit(R5_InJournal, &dev->flags)) {
1814 * For this case, PAGE_SIZE must be equal to 4KB and
1815 * page offset is zero.
1817 off_srcs[count] = dev->offset;
1818 xor_srcs[count++] = dev->orig_page;
1819 } else if (test_bit(R5_Wantdrain, &dev->flags)) {
1820 off_srcs[count] = dev->offset;
1821 xor_srcs[count++] = dev->page;
1825 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1826 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1827 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1828 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1833 static struct dma_async_tx_descriptor *
1834 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1835 struct dma_async_tx_descriptor *tx)
1837 struct page **blocks = to_addr_page(percpu, 0);
1838 unsigned int *offs = to_addr_offs(sh, percpu);
1840 struct async_submit_ctl submit;
1842 pr_debug("%s: stripe %llu\n", __func__,
1843 (unsigned long long)sh->sector);
1845 count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_WANT_DRAIN);
1847 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1848 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1849 tx = async_gen_syndrome(blocks, offs, count+2,
1850 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1855 static struct dma_async_tx_descriptor *
1856 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1858 struct r5conf *conf = sh->raid_conf;
1859 int disks = sh->disks;
1861 struct stripe_head *head_sh = sh;
1863 pr_debug("%s: stripe %llu\n", __func__,
1864 (unsigned long long)sh->sector);
1866 for (i = disks; i--; ) {
1871 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1877 * clear R5_InJournal, so when rewriting a page in
1878 * journal, it is not skipped by r5l_log_stripe()
1880 clear_bit(R5_InJournal, &dev->flags);
1881 spin_lock_irq(&sh->stripe_lock);
1882 chosen = dev->towrite;
1883 dev->towrite = NULL;
1884 sh->overwrite_disks = 0;
1885 BUG_ON(dev->written);
1886 wbi = dev->written = chosen;
1887 spin_unlock_irq(&sh->stripe_lock);
1888 WARN_ON(dev->page != dev->orig_page);
1890 while (wbi && wbi->bi_iter.bi_sector <
1891 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1892 if (wbi->bi_opf & REQ_FUA)
1893 set_bit(R5_WantFUA, &dev->flags);
1894 if (wbi->bi_opf & REQ_SYNC)
1895 set_bit(R5_SyncIO, &dev->flags);
1896 if (bio_op(wbi) == REQ_OP_DISCARD)
1897 set_bit(R5_Discard, &dev->flags);
1899 tx = async_copy_data(1, wbi, &dev->page,
1901 dev->sector, tx, sh,
1902 r5c_is_writeback(conf->log));
1903 if (dev->page != dev->orig_page &&
1904 !r5c_is_writeback(conf->log)) {
1905 set_bit(R5_SkipCopy, &dev->flags);
1906 clear_bit(R5_UPTODATE, &dev->flags);
1907 clear_bit(R5_OVERWRITE, &dev->flags);
1910 wbi = r5_next_bio(conf, wbi, dev->sector);
1913 if (head_sh->batch_head) {
1914 sh = list_first_entry(&sh->batch_list,
1927 static void ops_complete_reconstruct(void *stripe_head_ref)
1929 struct stripe_head *sh = stripe_head_ref;
1930 int disks = sh->disks;
1931 int pd_idx = sh->pd_idx;
1932 int qd_idx = sh->qd_idx;
1934 bool fua = false, sync = false, discard = false;
1936 pr_debug("%s: stripe %llu\n", __func__,
1937 (unsigned long long)sh->sector);
1939 for (i = disks; i--; ) {
1940 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1941 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1942 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1945 for (i = disks; i--; ) {
1946 struct r5dev *dev = &sh->dev[i];
1948 if (dev->written || i == pd_idx || i == qd_idx) {
1949 if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1950 set_bit(R5_UPTODATE, &dev->flags);
1951 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1952 set_bit(R5_Expanded, &dev->flags);
1955 set_bit(R5_WantFUA, &dev->flags);
1957 set_bit(R5_SyncIO, &dev->flags);
1961 if (sh->reconstruct_state == reconstruct_state_drain_run)
1962 sh->reconstruct_state = reconstruct_state_drain_result;
1963 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1964 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1966 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1967 sh->reconstruct_state = reconstruct_state_result;
1970 set_bit(STRIPE_HANDLE, &sh->state);
1971 raid5_release_stripe(sh);
1975 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1976 struct dma_async_tx_descriptor *tx)
1978 int disks = sh->disks;
1979 struct page **xor_srcs;
1980 unsigned int *off_srcs;
1981 struct async_submit_ctl submit;
1982 int count, pd_idx = sh->pd_idx, i;
1983 struct page *xor_dest;
1984 unsigned int off_dest;
1986 unsigned long flags;
1988 struct stripe_head *head_sh = sh;
1991 pr_debug("%s: stripe %llu\n", __func__,
1992 (unsigned long long)sh->sector);
1994 for (i = 0; i < sh->disks; i++) {
1997 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2000 if (i >= sh->disks) {
2001 atomic_inc(&sh->count);
2002 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
2003 ops_complete_reconstruct(sh);
2008 xor_srcs = to_addr_page(percpu, j);
2009 off_srcs = to_addr_offs(sh, percpu);
2010 /* check if prexor is active which means only process blocks
2011 * that are part of a read-modify-write (written)
2013 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2015 off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
2016 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
2017 for (i = disks; i--; ) {
2018 struct r5dev *dev = &sh->dev[i];
2019 if (head_sh->dev[i].written ||
2020 test_bit(R5_InJournal, &head_sh->dev[i].flags)) {
2021 off_srcs[count] = dev->offset;
2022 xor_srcs[count++] = dev->page;
2026 xor_dest = sh->dev[pd_idx].page;
2027 off_dest = sh->dev[pd_idx].offset;
2028 for (i = disks; i--; ) {
2029 struct r5dev *dev = &sh->dev[i];
2031 off_srcs[count] = dev->offset;
2032 xor_srcs[count++] = dev->page;
2037 /* 1/ if we prexor'd then the dest is reused as a source
2038 * 2/ if we did not prexor then we are redoing the parity
2039 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2040 * for the synchronous xor case
2042 last_stripe = !head_sh->batch_head ||
2043 list_first_entry(&sh->batch_list,
2044 struct stripe_head, batch_list) == head_sh;
2046 flags = ASYNC_TX_ACK |
2047 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
2049 atomic_inc(&head_sh->count);
2050 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
2051 to_addr_conv(sh, percpu, j));
2053 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
2054 init_async_submit(&submit, flags, tx, NULL, NULL,
2055 to_addr_conv(sh, percpu, j));
2058 if (unlikely(count == 1))
2059 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
2060 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2062 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2063 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2066 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2073 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
2074 struct dma_async_tx_descriptor *tx)
2076 struct async_submit_ctl submit;
2077 struct page **blocks;
2079 int count, i, j = 0;
2080 struct stripe_head *head_sh = sh;
2083 unsigned long txflags;
2085 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
2087 for (i = 0; i < sh->disks; i++) {
2088 if (sh->pd_idx == i || sh->qd_idx == i)
2090 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2093 if (i >= sh->disks) {
2094 atomic_inc(&sh->count);
2095 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2096 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2097 ops_complete_reconstruct(sh);
2102 blocks = to_addr_page(percpu, j);
2103 offs = to_addr_offs(sh, percpu);
2105 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2106 synflags = SYNDROME_SRC_WRITTEN;
2107 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
2109 synflags = SYNDROME_SRC_ALL;
2110 txflags = ASYNC_TX_ACK;
2113 count = set_syndrome_sources(blocks, offs, sh, synflags);
2114 last_stripe = !head_sh->batch_head ||
2115 list_first_entry(&sh->batch_list,
2116 struct stripe_head, batch_list) == head_sh;
2119 atomic_inc(&head_sh->count);
2120 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
2121 head_sh, to_addr_conv(sh, percpu, j));
2123 init_async_submit(&submit, 0, tx, NULL, NULL,
2124 to_addr_conv(sh, percpu, j));
2125 tx = async_gen_syndrome(blocks, offs, count+2,
2126 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2129 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2135 static void ops_complete_check(void *stripe_head_ref)
2137 struct stripe_head *sh = stripe_head_ref;
2139 pr_debug("%s: stripe %llu\n", __func__,
2140 (unsigned long long)sh->sector);
2142 sh->check_state = check_state_check_result;
2143 set_bit(STRIPE_HANDLE, &sh->state);
2144 raid5_release_stripe(sh);
2147 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2149 int disks = sh->disks;
2150 int pd_idx = sh->pd_idx;
2151 int qd_idx = sh->qd_idx;
2152 struct page *xor_dest;
2153 unsigned int off_dest;
2154 struct page **xor_srcs = to_addr_page(percpu, 0);
2155 unsigned int *off_srcs = to_addr_offs(sh, percpu);
2156 struct dma_async_tx_descriptor *tx;
2157 struct async_submit_ctl submit;
2161 pr_debug("%s: stripe %llu\n", __func__,
2162 (unsigned long long)sh->sector);
2164 BUG_ON(sh->batch_head);
2166 xor_dest = sh->dev[pd_idx].page;
2167 off_dest = sh->dev[pd_idx].offset;
2168 off_srcs[count] = off_dest;
2169 xor_srcs[count++] = xor_dest;
2170 for (i = disks; i--; ) {
2171 if (i == pd_idx || i == qd_idx)
2173 off_srcs[count] = sh->dev[i].offset;
2174 xor_srcs[count++] = sh->dev[i].page;
2177 init_async_submit(&submit, 0, NULL, NULL, NULL,
2178 to_addr_conv(sh, percpu, 0));
2179 tx = async_xor_val_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2180 RAID5_STRIPE_SIZE(sh->raid_conf),
2181 &sh->ops.zero_sum_result, &submit);
2183 atomic_inc(&sh->count);
2184 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2185 tx = async_trigger_callback(&submit);
2188 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2190 struct page **srcs = to_addr_page(percpu, 0);
2191 unsigned int *offs = to_addr_offs(sh, percpu);
2192 struct async_submit_ctl submit;
2195 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2196 (unsigned long long)sh->sector, checkp);
2198 BUG_ON(sh->batch_head);
2199 count = set_syndrome_sources(srcs, offs, sh, SYNDROME_SRC_ALL);
2203 atomic_inc(&sh->count);
2204 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2205 sh, to_addr_conv(sh, percpu, 0));
2206 async_syndrome_val(srcs, offs, count+2,
2207 RAID5_STRIPE_SIZE(sh->raid_conf),
2208 &sh->ops.zero_sum_result, percpu->spare_page, 0, &submit);
2211 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2213 int overlap_clear = 0, i, disks = sh->disks;
2214 struct dma_async_tx_descriptor *tx = NULL;
2215 struct r5conf *conf = sh->raid_conf;
2216 int level = conf->level;
2217 struct raid5_percpu *percpu;
2221 percpu = per_cpu_ptr(conf->percpu, cpu);
2222 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2223 ops_run_biofill(sh);
2227 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2229 tx = ops_run_compute5(sh, percpu);
2231 if (sh->ops.target2 < 0 || sh->ops.target < 0)
2232 tx = ops_run_compute6_1(sh, percpu);
2234 tx = ops_run_compute6_2(sh, percpu);
2236 /* terminate the chain if reconstruct is not set to be run */
2237 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2241 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2243 tx = ops_run_prexor5(sh, percpu, tx);
2245 tx = ops_run_prexor6(sh, percpu, tx);
2248 if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2249 tx = ops_run_partial_parity(sh, percpu, tx);
2251 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2252 tx = ops_run_biodrain(sh, tx);
2256 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2258 ops_run_reconstruct5(sh, percpu, tx);
2260 ops_run_reconstruct6(sh, percpu, tx);
2263 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2264 if (sh->check_state == check_state_run)
2265 ops_run_check_p(sh, percpu);
2266 else if (sh->check_state == check_state_run_q)
2267 ops_run_check_pq(sh, percpu, 0);
2268 else if (sh->check_state == check_state_run_pq)
2269 ops_run_check_pq(sh, percpu, 1);
2274 if (overlap_clear && !sh->batch_head)
2275 for (i = disks; i--; ) {
2276 struct r5dev *dev = &sh->dev[i];
2277 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2278 wake_up(&sh->raid_conf->wait_for_overlap);
2283 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2285 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2289 __free_page(sh->ppl_page);
2290 kmem_cache_free(sc, sh);
2293 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2294 int disks, struct r5conf *conf)
2296 struct stripe_head *sh;
2299 sh = kmem_cache_zalloc(sc, gfp);
2301 spin_lock_init(&sh->stripe_lock);
2302 spin_lock_init(&sh->batch_lock);
2303 INIT_LIST_HEAD(&sh->batch_list);
2304 INIT_LIST_HEAD(&sh->lru);
2305 INIT_LIST_HEAD(&sh->r5c);
2306 INIT_LIST_HEAD(&sh->log_list);
2307 atomic_set(&sh->count, 1);
2308 sh->raid_conf = conf;
2309 sh->log_start = MaxSector;
2310 for (i = 0; i < disks; i++) {
2311 struct r5dev *dev = &sh->dev[i];
2313 bio_init(&dev->req, &dev->vec, 1);
2314 bio_init(&dev->rreq, &dev->rvec, 1);
2317 if (raid5_has_ppl(conf)) {
2318 sh->ppl_page = alloc_page(gfp);
2319 if (!sh->ppl_page) {
2320 free_stripe(sc, sh);
2324 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2325 if (init_stripe_shared_pages(sh, conf, disks)) {
2326 free_stripe(sc, sh);
2333 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2335 struct stripe_head *sh;
2337 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2341 if (grow_buffers(sh, gfp)) {
2343 free_stripe(conf->slab_cache, sh);
2346 sh->hash_lock_index =
2347 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2348 /* we just created an active stripe so... */
2349 atomic_inc(&conf->active_stripes);
2351 raid5_release_stripe(sh);
2352 conf->max_nr_stripes++;
2356 static int grow_stripes(struct r5conf *conf, int num)
2358 struct kmem_cache *sc;
2359 size_t namelen = sizeof(conf->cache_name[0]);
2360 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2362 if (conf->mddev->gendisk)
2363 snprintf(conf->cache_name[0], namelen,
2364 "raid%d-%s", conf->level, mdname(conf->mddev));
2366 snprintf(conf->cache_name[0], namelen,
2367 "raid%d-%p", conf->level, conf->mddev);
2368 snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2370 conf->active_name = 0;
2371 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2372 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2376 conf->slab_cache = sc;
2377 conf->pool_size = devs;
2379 if (!grow_one_stripe(conf, GFP_KERNEL))
2386 * scribble_alloc - allocate percpu scribble buffer for required size
2387 * of the scribble region
2388 * @percpu: from for_each_present_cpu() of the caller
2389 * @num: total number of disks in the array
2390 * @cnt: scribble objs count for required size of the scribble region
2392 * The scribble buffer size must be enough to contain:
2393 * 1/ a struct page pointer for each device in the array +2
2394 * 2/ room to convert each entry in (1) to its corresponding dma
2395 * (dma_map_page()) or page (page_address()) address.
2397 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2398 * calculate over all devices (not just the data blocks), using zeros in place
2399 * of the P and Q blocks.
2401 static int scribble_alloc(struct raid5_percpu *percpu,
2405 sizeof(struct page *) * (num + 2) +
2406 sizeof(addr_conv_t) * (num + 2) +
2407 sizeof(unsigned int) * (num + 2);
2411 * If here is in raid array suspend context, it is in memalloc noio
2412 * context as well, there is no potential recursive memory reclaim
2413 * I/Os with the GFP_KERNEL flag.
2415 scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2419 kvfree(percpu->scribble);
2421 percpu->scribble = scribble;
2422 percpu->scribble_obj_size = obj_size;
2426 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2432 * Never shrink. And mddev_suspend() could deadlock if this is called
2433 * from raid5d. In that case, scribble_disks and scribble_sectors
2434 * should equal to new_disks and new_sectors
2436 if (conf->scribble_disks >= new_disks &&
2437 conf->scribble_sectors >= new_sectors)
2439 mddev_suspend(conf->mddev);
2442 for_each_present_cpu(cpu) {
2443 struct raid5_percpu *percpu;
2445 percpu = per_cpu_ptr(conf->percpu, cpu);
2446 err = scribble_alloc(percpu, new_disks,
2447 new_sectors / RAID5_STRIPE_SECTORS(conf));
2453 mddev_resume(conf->mddev);
2455 conf->scribble_disks = new_disks;
2456 conf->scribble_sectors = new_sectors;
2461 static int resize_stripes(struct r5conf *conf, int newsize)
2463 /* Make all the stripes able to hold 'newsize' devices.
2464 * New slots in each stripe get 'page' set to a new page.
2466 * This happens in stages:
2467 * 1/ create a new kmem_cache and allocate the required number of
2469 * 2/ gather all the old stripe_heads and transfer the pages across
2470 * to the new stripe_heads. This will have the side effect of
2471 * freezing the array as once all stripe_heads have been collected,
2472 * no IO will be possible. Old stripe heads are freed once their
2473 * pages have been transferred over, and the old kmem_cache is
2474 * freed when all stripes are done.
2475 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2476 * we simple return a failure status - no need to clean anything up.
2477 * 4/ allocate new pages for the new slots in the new stripe_heads.
2478 * If this fails, we don't bother trying the shrink the
2479 * stripe_heads down again, we just leave them as they are.
2480 * As each stripe_head is processed the new one is released into
2483 * Once step2 is started, we cannot afford to wait for a write,
2484 * so we use GFP_NOIO allocations.
2486 struct stripe_head *osh, *nsh;
2487 LIST_HEAD(newstripes);
2488 struct disk_info *ndisks;
2490 struct kmem_cache *sc;
2494 md_allow_write(conf->mddev);
2497 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2498 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2503 /* Need to ensure auto-resizing doesn't interfere */
2504 mutex_lock(&conf->cache_size_mutex);
2506 for (i = conf->max_nr_stripes; i; i--) {
2507 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2511 list_add(&nsh->lru, &newstripes);
2514 /* didn't get enough, give up */
2515 while (!list_empty(&newstripes)) {
2516 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2517 list_del(&nsh->lru);
2518 free_stripe(sc, nsh);
2520 kmem_cache_destroy(sc);
2521 mutex_unlock(&conf->cache_size_mutex);
2524 /* Step 2 - Must use GFP_NOIO now.
2525 * OK, we have enough stripes, start collecting inactive
2526 * stripes and copying them over
2530 list_for_each_entry(nsh, &newstripes, lru) {
2531 lock_device_hash_lock(conf, hash);
2532 wait_event_cmd(conf->wait_for_stripe,
2533 !list_empty(conf->inactive_list + hash),
2534 unlock_device_hash_lock(conf, hash),
2535 lock_device_hash_lock(conf, hash));
2536 osh = get_free_stripe(conf, hash);
2537 unlock_device_hash_lock(conf, hash);
2539 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2540 for (i = 0; i < osh->nr_pages; i++) {
2541 nsh->pages[i] = osh->pages[i];
2542 osh->pages[i] = NULL;
2545 for(i=0; i<conf->pool_size; i++) {
2546 nsh->dev[i].page = osh->dev[i].page;
2547 nsh->dev[i].orig_page = osh->dev[i].page;
2548 nsh->dev[i].offset = osh->dev[i].offset;
2550 nsh->hash_lock_index = hash;
2551 free_stripe(conf->slab_cache, osh);
2553 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2554 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2559 kmem_cache_destroy(conf->slab_cache);
2562 * At this point, we are holding all the stripes so the array
2563 * is completely stalled, so now is a good time to resize
2564 * conf->disks and the scribble region
2566 ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2568 for (i = 0; i < conf->pool_size; i++)
2569 ndisks[i] = conf->disks[i];
2571 for (i = conf->pool_size; i < newsize; i++) {
2572 ndisks[i].extra_page = alloc_page(GFP_NOIO);
2573 if (!ndisks[i].extra_page)
2578 for (i = conf->pool_size; i < newsize; i++)
2579 if (ndisks[i].extra_page)
2580 put_page(ndisks[i].extra_page);
2584 conf->disks = ndisks;
2589 conf->slab_cache = sc;
2590 conf->active_name = 1-conf->active_name;
2592 /* Step 4, return new stripes to service */
2593 while(!list_empty(&newstripes)) {
2594 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2595 list_del_init(&nsh->lru);
2597 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2598 for (i = 0; i < nsh->nr_pages; i++) {
2601 nsh->pages[i] = alloc_page(GFP_NOIO);
2606 for (i = conf->raid_disks; i < newsize; i++) {
2607 if (nsh->dev[i].page)
2609 nsh->dev[i].page = raid5_get_dev_page(nsh, i);
2610 nsh->dev[i].orig_page = nsh->dev[i].page;
2611 nsh->dev[i].offset = raid5_get_page_offset(nsh, i);
2614 for (i=conf->raid_disks; i < newsize; i++)
2615 if (nsh->dev[i].page == NULL) {
2616 struct page *p = alloc_page(GFP_NOIO);
2617 nsh->dev[i].page = p;
2618 nsh->dev[i].orig_page = p;
2619 nsh->dev[i].offset = 0;
2624 raid5_release_stripe(nsh);
2626 /* critical section pass, GFP_NOIO no longer needed */
2629 conf->pool_size = newsize;
2630 mutex_unlock(&conf->cache_size_mutex);
2635 static int drop_one_stripe(struct r5conf *conf)
2637 struct stripe_head *sh;
2638 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2640 spin_lock_irq(conf->hash_locks + hash);
2641 sh = get_free_stripe(conf, hash);
2642 spin_unlock_irq(conf->hash_locks + hash);
2645 BUG_ON(atomic_read(&sh->count));
2647 free_stripe(conf->slab_cache, sh);
2648 atomic_dec(&conf->active_stripes);
2649 conf->max_nr_stripes--;
2653 static void shrink_stripes(struct r5conf *conf)
2655 while (conf->max_nr_stripes &&
2656 drop_one_stripe(conf))
2659 kmem_cache_destroy(conf->slab_cache);
2660 conf->slab_cache = NULL;
2663 static void raid5_end_read_request(struct bio * bi)
2665 struct stripe_head *sh = bi->bi_private;
2666 struct r5conf *conf = sh->raid_conf;
2667 int disks = sh->disks, i;
2668 char b[BDEVNAME_SIZE];
2669 struct md_rdev *rdev = NULL;
2672 for (i=0 ; i<disks; i++)
2673 if (bi == &sh->dev[i].req)
2676 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2677 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2684 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2685 /* If replacement finished while this request was outstanding,
2686 * 'replacement' might be NULL already.
2687 * In that case it moved down to 'rdev'.
2688 * rdev is not removed until all requests are finished.
2690 rdev = conf->disks[i].replacement;
2692 rdev = conf->disks[i].rdev;
2694 if (use_new_offset(conf, sh))
2695 s = sh->sector + rdev->new_data_offset;
2697 s = sh->sector + rdev->data_offset;
2698 if (!bi->bi_status) {
2699 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2700 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2701 /* Note that this cannot happen on a
2702 * replacement device. We just fail those on
2705 pr_info_ratelimited(
2706 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2707 mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
2708 (unsigned long long)s,
2709 bdevname(rdev->bdev, b));
2710 atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
2711 clear_bit(R5_ReadError, &sh->dev[i].flags);
2712 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2713 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2714 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2716 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2718 * end read for a page in journal, this
2719 * must be preparing for prexor in rmw
2721 set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2723 if (atomic_read(&rdev->read_errors))
2724 atomic_set(&rdev->read_errors, 0);
2726 const char *bdn = bdevname(rdev->bdev, b);
2730 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2731 if (!(bi->bi_status == BLK_STS_PROTECTION))
2732 atomic_inc(&rdev->read_errors);
2733 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2734 pr_warn_ratelimited(
2735 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2736 mdname(conf->mddev),
2737 (unsigned long long)s,
2739 else if (conf->mddev->degraded >= conf->max_degraded) {
2741 pr_warn_ratelimited(
2742 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2743 mdname(conf->mddev),
2744 (unsigned long long)s,
2746 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2749 pr_warn_ratelimited(
2750 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2751 mdname(conf->mddev),
2752 (unsigned long long)s,
2754 } else if (atomic_read(&rdev->read_errors)
2755 > conf->max_nr_stripes) {
2756 if (!test_bit(Faulty, &rdev->flags)) {
2757 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2758 mdname(conf->mddev),
2759 atomic_read(&rdev->read_errors),
2760 conf->max_nr_stripes);
2761 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2762 mdname(conf->mddev), bdn);
2766 if (set_bad && test_bit(In_sync, &rdev->flags)
2767 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2770 if (sh->qd_idx >= 0 && sh->pd_idx == i)
2771 set_bit(R5_ReadError, &sh->dev[i].flags);
2772 else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2773 set_bit(R5_ReadError, &sh->dev[i].flags);
2774 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2776 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2778 clear_bit(R5_ReadError, &sh->dev[i].flags);
2779 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2781 && test_bit(In_sync, &rdev->flags)
2782 && rdev_set_badblocks(
2783 rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
2784 md_error(conf->mddev, rdev);
2787 rdev_dec_pending(rdev, conf->mddev);
2789 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2790 set_bit(STRIPE_HANDLE, &sh->state);
2791 raid5_release_stripe(sh);
2794 static void raid5_end_write_request(struct bio *bi)
2796 struct stripe_head *sh = bi->bi_private;
2797 struct r5conf *conf = sh->raid_conf;
2798 int disks = sh->disks, i;
2799 struct md_rdev *rdev;
2802 int replacement = 0;
2804 for (i = 0 ; i < disks; i++) {
2805 if (bi == &sh->dev[i].req) {
2806 rdev = conf->disks[i].rdev;
2809 if (bi == &sh->dev[i].rreq) {
2810 rdev = conf->disks[i].replacement;
2814 /* rdev was removed and 'replacement'
2815 * replaced it. rdev is not removed
2816 * until all requests are finished.
2818 rdev = conf->disks[i].rdev;
2822 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2823 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2833 md_error(conf->mddev, rdev);
2834 else if (is_badblock(rdev, sh->sector,
2835 RAID5_STRIPE_SECTORS(conf),
2836 &first_bad, &bad_sectors))
2837 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2839 if (bi->bi_status) {
2840 set_bit(STRIPE_DEGRADED, &sh->state);
2841 set_bit(WriteErrorSeen, &rdev->flags);
2842 set_bit(R5_WriteError, &sh->dev[i].flags);
2843 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2844 set_bit(MD_RECOVERY_NEEDED,
2845 &rdev->mddev->recovery);
2846 } else if (is_badblock(rdev, sh->sector,
2847 RAID5_STRIPE_SECTORS(conf),
2848 &first_bad, &bad_sectors)) {
2849 set_bit(R5_MadeGood, &sh->dev[i].flags);
2850 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2851 /* That was a successful write so make
2852 * sure it looks like we already did
2855 set_bit(R5_ReWrite, &sh->dev[i].flags);
2858 rdev_dec_pending(rdev, conf->mddev);
2860 if (sh->batch_head && bi->bi_status && !replacement)
2861 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2864 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2865 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2866 set_bit(STRIPE_HANDLE, &sh->state);
2867 raid5_release_stripe(sh);
2869 if (sh->batch_head && sh != sh->batch_head)
2870 raid5_release_stripe(sh->batch_head);
2873 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2875 char b[BDEVNAME_SIZE];
2876 struct r5conf *conf = mddev->private;
2877 unsigned long flags;
2878 pr_debug("raid456: error called\n");
2880 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n",
2881 mdname(mddev), bdevname(rdev->bdev, b));
2883 spin_lock_irqsave(&conf->device_lock, flags);
2884 set_bit(Faulty, &rdev->flags);
2885 clear_bit(In_sync, &rdev->flags);
2886 mddev->degraded = raid5_calc_degraded(conf);
2888 if (has_failed(conf)) {
2889 set_bit(MD_BROKEN, &conf->mddev->flags);
2890 conf->recovery_disabled = mddev->recovery_disabled;
2892 pr_crit("md/raid:%s: Cannot continue operation (%d/%d failed).\n",
2893 mdname(mddev), mddev->degraded, conf->raid_disks);
2895 pr_crit("md/raid:%s: Operation continuing on %d devices.\n",
2896 mdname(mddev), conf->raid_disks - mddev->degraded);
2899 spin_unlock_irqrestore(&conf->device_lock, flags);
2900 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2902 set_bit(Blocked, &rdev->flags);
2903 set_mask_bits(&mddev->sb_flags, 0,
2904 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2905 r5c_update_on_rdev_error(mddev, rdev);
2909 * Input: a 'big' sector number,
2910 * Output: index of the data and parity disk, and the sector # in them.
2912 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2913 int previous, int *dd_idx,
2914 struct stripe_head *sh)
2916 sector_t stripe, stripe2;
2917 sector_t chunk_number;
2918 unsigned int chunk_offset;
2921 sector_t new_sector;
2922 int algorithm = previous ? conf->prev_algo
2924 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2925 : conf->chunk_sectors;
2926 int raid_disks = previous ? conf->previous_raid_disks
2928 int data_disks = raid_disks - conf->max_degraded;
2930 /* First compute the information on this sector */
2933 * Compute the chunk number and the sector offset inside the chunk
2935 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2936 chunk_number = r_sector;
2939 * Compute the stripe number
2941 stripe = chunk_number;
2942 *dd_idx = sector_div(stripe, data_disks);
2945 * Select the parity disk based on the user selected algorithm.
2947 pd_idx = qd_idx = -1;
2948 switch(conf->level) {
2950 pd_idx = data_disks;
2953 switch (algorithm) {
2954 case ALGORITHM_LEFT_ASYMMETRIC:
2955 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2956 if (*dd_idx >= pd_idx)
2959 case ALGORITHM_RIGHT_ASYMMETRIC:
2960 pd_idx = sector_div(stripe2, raid_disks);
2961 if (*dd_idx >= pd_idx)
2964 case ALGORITHM_LEFT_SYMMETRIC:
2965 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2966 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2968 case ALGORITHM_RIGHT_SYMMETRIC:
2969 pd_idx = sector_div(stripe2, raid_disks);
2970 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2972 case ALGORITHM_PARITY_0:
2976 case ALGORITHM_PARITY_N:
2977 pd_idx = data_disks;
2985 switch (algorithm) {
2986 case ALGORITHM_LEFT_ASYMMETRIC:
2987 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2988 qd_idx = pd_idx + 1;
2989 if (pd_idx == raid_disks-1) {
2990 (*dd_idx)++; /* Q D D D P */
2992 } else if (*dd_idx >= pd_idx)
2993 (*dd_idx) += 2; /* D D P Q D */
2995 case ALGORITHM_RIGHT_ASYMMETRIC:
2996 pd_idx = sector_div(stripe2, raid_disks);
2997 qd_idx = pd_idx + 1;
2998 if (pd_idx == raid_disks-1) {
2999 (*dd_idx)++; /* Q D D D P */
3001 } else if (*dd_idx >= pd_idx)
3002 (*dd_idx) += 2; /* D D P Q D */
3004 case ALGORITHM_LEFT_SYMMETRIC:
3005 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3006 qd_idx = (pd_idx + 1) % raid_disks;
3007 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3009 case ALGORITHM_RIGHT_SYMMETRIC:
3010 pd_idx = sector_div(stripe2, raid_disks);
3011 qd_idx = (pd_idx + 1) % raid_disks;
3012 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3015 case ALGORITHM_PARITY_0:
3020 case ALGORITHM_PARITY_N:
3021 pd_idx = data_disks;
3022 qd_idx = data_disks + 1;
3025 case ALGORITHM_ROTATING_ZERO_RESTART:
3026 /* Exactly the same as RIGHT_ASYMMETRIC, but or
3027 * of blocks for computing Q is different.
3029 pd_idx = sector_div(stripe2, raid_disks);
3030 qd_idx = pd_idx + 1;
3031 if (pd_idx == raid_disks-1) {
3032 (*dd_idx)++; /* Q D D D P */
3034 } else if (*dd_idx >= pd_idx)
3035 (*dd_idx) += 2; /* D D P Q D */
3039 case ALGORITHM_ROTATING_N_RESTART:
3040 /* Same a left_asymmetric, by first stripe is
3041 * D D D P Q rather than
3045 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3046 qd_idx = pd_idx + 1;
3047 if (pd_idx == raid_disks-1) {
3048 (*dd_idx)++; /* Q D D D P */
3050 } else if (*dd_idx >= pd_idx)
3051 (*dd_idx) += 2; /* D D P Q D */
3055 case ALGORITHM_ROTATING_N_CONTINUE:
3056 /* Same as left_symmetric but Q is before P */
3057 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3058 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
3059 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3063 case ALGORITHM_LEFT_ASYMMETRIC_6:
3064 /* RAID5 left_asymmetric, with Q on last device */
3065 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3066 if (*dd_idx >= pd_idx)
3068 qd_idx = raid_disks - 1;
3071 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3072 pd_idx = sector_div(stripe2, raid_disks-1);
3073 if (*dd_idx >= pd_idx)
3075 qd_idx = raid_disks - 1;
3078 case ALGORITHM_LEFT_SYMMETRIC_6:
3079 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3080 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3081 qd_idx = raid_disks - 1;
3084 case ALGORITHM_RIGHT_SYMMETRIC_6:
3085 pd_idx = sector_div(stripe2, raid_disks-1);
3086 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3087 qd_idx = raid_disks - 1;
3090 case ALGORITHM_PARITY_0_6:
3093 qd_idx = raid_disks - 1;
3103 sh->pd_idx = pd_idx;
3104 sh->qd_idx = qd_idx;
3105 sh->ddf_layout = ddf_layout;
3108 * Finally, compute the new sector number
3110 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
3114 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
3116 struct r5conf *conf = sh->raid_conf;
3117 int raid_disks = sh->disks;
3118 int data_disks = raid_disks - conf->max_degraded;
3119 sector_t new_sector = sh->sector, check;
3120 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
3121 : conf->chunk_sectors;
3122 int algorithm = previous ? conf->prev_algo
3126 sector_t chunk_number;
3127 int dummy1, dd_idx = i;
3129 struct stripe_head sh2;
3131 chunk_offset = sector_div(new_sector, sectors_per_chunk);
3132 stripe = new_sector;
3134 if (i == sh->pd_idx)
3136 switch(conf->level) {
3139 switch (algorithm) {
3140 case ALGORITHM_LEFT_ASYMMETRIC:
3141 case ALGORITHM_RIGHT_ASYMMETRIC:
3145 case ALGORITHM_LEFT_SYMMETRIC:
3146 case ALGORITHM_RIGHT_SYMMETRIC:
3149 i -= (sh->pd_idx + 1);
3151 case ALGORITHM_PARITY_0:
3154 case ALGORITHM_PARITY_N:
3161 if (i == sh->qd_idx)
3162 return 0; /* It is the Q disk */
3163 switch (algorithm) {
3164 case ALGORITHM_LEFT_ASYMMETRIC:
3165 case ALGORITHM_RIGHT_ASYMMETRIC:
3166 case ALGORITHM_ROTATING_ZERO_RESTART:
3167 case ALGORITHM_ROTATING_N_RESTART:
3168 if (sh->pd_idx == raid_disks-1)
3169 i--; /* Q D D D P */
3170 else if (i > sh->pd_idx)
3171 i -= 2; /* D D P Q D */
3173 case ALGORITHM_LEFT_SYMMETRIC:
3174 case ALGORITHM_RIGHT_SYMMETRIC:
3175 if (sh->pd_idx == raid_disks-1)
3176 i--; /* Q D D D P */
3181 i -= (sh->pd_idx + 2);
3184 case ALGORITHM_PARITY_0:
3187 case ALGORITHM_PARITY_N:
3189 case ALGORITHM_ROTATING_N_CONTINUE:
3190 /* Like left_symmetric, but P is before Q */
3191 if (sh->pd_idx == 0)
3192 i--; /* P D D D Q */
3197 i -= (sh->pd_idx + 1);
3200 case ALGORITHM_LEFT_ASYMMETRIC_6:
3201 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3205 case ALGORITHM_LEFT_SYMMETRIC_6:
3206 case ALGORITHM_RIGHT_SYMMETRIC_6:
3208 i += data_disks + 1;
3209 i -= (sh->pd_idx + 1);
3211 case ALGORITHM_PARITY_0_6:
3220 chunk_number = stripe * data_disks + i;
3221 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3223 check = raid5_compute_sector(conf, r_sector,
3224 previous, &dummy1, &sh2);
3225 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3226 || sh2.qd_idx != sh->qd_idx) {
3227 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3228 mdname(conf->mddev));
3235 * There are cases where we want handle_stripe_dirtying() and
3236 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3238 * This function checks whether we want to delay the towrite. Specifically,
3239 * we delay the towrite when:
3241 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3242 * stripe has data in journal (for other devices).
3244 * In this case, when reading data for the non-overwrite dev, it is
3245 * necessary to handle complex rmw of write back cache (prexor with
3246 * orig_page, and xor with page). To keep read path simple, we would
3247 * like to flush data in journal to RAID disks first, so complex rmw
3248 * is handled in the write patch (handle_stripe_dirtying).
3250 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3252 * It is important to be able to flush all stripes in raid5-cache.
3253 * Therefore, we need reserve some space on the journal device for
3254 * these flushes. If flush operation includes pending writes to the
3255 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3256 * for the flush out. If we exclude these pending writes from flush
3257 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3258 * Therefore, excluding pending writes in these cases enables more
3259 * efficient use of the journal device.
3261 * Note: To make sure the stripe makes progress, we only delay
3262 * towrite for stripes with data already in journal (injournal > 0).
3263 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3264 * no_space_stripes list.
3266 * 3. during journal failure
3267 * In journal failure, we try to flush all cached data to raid disks
3268 * based on data in stripe cache. The array is read-only to upper
3269 * layers, so we would skip all pending writes.
3272 static inline bool delay_towrite(struct r5conf *conf,
3274 struct stripe_head_state *s)
3277 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3278 !test_bit(R5_Insync, &dev->flags) && s->injournal)
3281 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3285 if (s->log_failed && s->injournal)
3291 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3292 int rcw, int expand)
3294 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3295 struct r5conf *conf = sh->raid_conf;
3296 int level = conf->level;
3300 * In some cases, handle_stripe_dirtying initially decided to
3301 * run rmw and allocates extra page for prexor. However, rcw is
3302 * cheaper later on. We need to free the extra page now,
3303 * because we won't be able to do that in ops_complete_prexor().
3305 r5c_release_extra_page(sh);
3307 for (i = disks; i--; ) {
3308 struct r5dev *dev = &sh->dev[i];
3310 if (dev->towrite && !delay_towrite(conf, dev, s)) {
3311 set_bit(R5_LOCKED, &dev->flags);
3312 set_bit(R5_Wantdrain, &dev->flags);
3314 clear_bit(R5_UPTODATE, &dev->flags);
3316 } else if (test_bit(R5_InJournal, &dev->flags)) {
3317 set_bit(R5_LOCKED, &dev->flags);
3321 /* if we are not expanding this is a proper write request, and
3322 * there will be bios with new data to be drained into the
3327 /* False alarm, nothing to do */
3329 sh->reconstruct_state = reconstruct_state_drain_run;
3330 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3332 sh->reconstruct_state = reconstruct_state_run;
3334 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3336 if (s->locked + conf->max_degraded == disks)
3337 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3338 atomic_inc(&conf->pending_full_writes);
3340 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3341 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3342 BUG_ON(level == 6 &&
3343 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3344 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3346 for (i = disks; i--; ) {
3347 struct r5dev *dev = &sh->dev[i];
3348 if (i == pd_idx || i == qd_idx)
3352 (test_bit(R5_UPTODATE, &dev->flags) ||
3353 test_bit(R5_Wantcompute, &dev->flags))) {
3354 set_bit(R5_Wantdrain, &dev->flags);
3355 set_bit(R5_LOCKED, &dev->flags);
3356 clear_bit(R5_UPTODATE, &dev->flags);
3358 } else if (test_bit(R5_InJournal, &dev->flags)) {
3359 set_bit(R5_LOCKED, &dev->flags);
3364 /* False alarm - nothing to do */
3366 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3367 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3368 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3369 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3372 /* keep the parity disk(s) locked while asynchronous operations
3375 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3376 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3380 int qd_idx = sh->qd_idx;
3381 struct r5dev *dev = &sh->dev[qd_idx];
3383 set_bit(R5_LOCKED, &dev->flags);
3384 clear_bit(R5_UPTODATE, &dev->flags);
3388 if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3389 test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3390 !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3391 test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3392 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3394 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3395 __func__, (unsigned long long)sh->sector,
3396 s->locked, s->ops_request);
3400 * Each stripe/dev can have one or more bion attached.
3401 * toread/towrite point to the first in a chain.
3402 * The bi_next chain must be in order.
3404 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3405 int forwrite, int previous)
3408 struct r5conf *conf = sh->raid_conf;
3411 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3412 (unsigned long long)bi->bi_iter.bi_sector,
3413 (unsigned long long)sh->sector);
3415 spin_lock_irq(&sh->stripe_lock);
3416 sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3417 /* Don't allow new IO added to stripes in batch list */
3421 bip = &sh->dev[dd_idx].towrite;
3425 bip = &sh->dev[dd_idx].toread;
3426 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3427 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3429 bip = & (*bip)->bi_next;
3431 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3434 if (forwrite && raid5_has_ppl(conf)) {
3436 * With PPL only writes to consecutive data chunks within a
3437 * stripe are allowed because for a single stripe_head we can
3438 * only have one PPL entry at a time, which describes one data
3439 * range. Not really an overlap, but wait_for_overlap can be
3440 * used to handle this.
3448 for (i = 0; i < sh->disks; i++) {
3449 if (i != sh->pd_idx &&
3450 (i == dd_idx || sh->dev[i].towrite)) {
3451 sector = sh->dev[i].sector;
3452 if (count == 0 || sector < first)
3460 if (first + conf->chunk_sectors * (count - 1) != last)
3464 if (!forwrite || previous)
3465 clear_bit(STRIPE_BATCH_READY, &sh->state);
3467 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3471 bio_inc_remaining(bi);
3472 md_write_inc(conf->mddev, bi);
3475 /* check if page is covered */
3476 sector_t sector = sh->dev[dd_idx].sector;
3477 for (bi=sh->dev[dd_idx].towrite;
3478 sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
3479 bi && bi->bi_iter.bi_sector <= sector;
3480 bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
3481 if (bio_end_sector(bi) >= sector)
3482 sector = bio_end_sector(bi);
3484 if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
3485 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3486 sh->overwrite_disks++;
3489 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3490 (unsigned long long)(*bip)->bi_iter.bi_sector,
3491 (unsigned long long)sh->sector, dd_idx);
3493 if (conf->mddev->bitmap && firstwrite) {
3494 /* Cannot hold spinlock over bitmap_startwrite,
3495 * but must ensure this isn't added to a batch until
3496 * we have added to the bitmap and set bm_seq.
3497 * So set STRIPE_BITMAP_PENDING to prevent
3499 * If multiple add_stripe_bio() calls race here they
3500 * much all set STRIPE_BITMAP_PENDING. So only the first one
3501 * to complete "bitmap_startwrite" gets to set
3502 * STRIPE_BIT_DELAY. This is important as once a stripe
3503 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3506 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3507 spin_unlock_irq(&sh->stripe_lock);
3508 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3509 RAID5_STRIPE_SECTORS(conf), 0);
3510 spin_lock_irq(&sh->stripe_lock);
3511 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3512 if (!sh->batch_head) {
3513 sh->bm_seq = conf->seq_flush+1;
3514 set_bit(STRIPE_BIT_DELAY, &sh->state);
3517 spin_unlock_irq(&sh->stripe_lock);
3519 if (stripe_can_batch(sh))
3520 stripe_add_to_batch_list(conf, sh);
3524 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3525 spin_unlock_irq(&sh->stripe_lock);
3529 static void end_reshape(struct r5conf *conf);
3531 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3532 struct stripe_head *sh)
3534 int sectors_per_chunk =
3535 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3537 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3538 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3540 raid5_compute_sector(conf,
3541 stripe * (disks - conf->max_degraded)
3542 *sectors_per_chunk + chunk_offset,
3548 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3549 struct stripe_head_state *s, int disks)
3552 BUG_ON(sh->batch_head);
3553 for (i = disks; i--; ) {
3557 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3558 struct md_rdev *rdev;
3560 rdev = rcu_dereference(conf->disks[i].rdev);
3561 if (rdev && test_bit(In_sync, &rdev->flags) &&
3562 !test_bit(Faulty, &rdev->flags))
3563 atomic_inc(&rdev->nr_pending);
3568 if (!rdev_set_badblocks(
3571 RAID5_STRIPE_SECTORS(conf), 0))
3572 md_error(conf->mddev, rdev);
3573 rdev_dec_pending(rdev, conf->mddev);
3576 spin_lock_irq(&sh->stripe_lock);
3577 /* fail all writes first */
3578 bi = sh->dev[i].towrite;
3579 sh->dev[i].towrite = NULL;
3580 sh->overwrite_disks = 0;
3581 spin_unlock_irq(&sh->stripe_lock);
3585 log_stripe_write_finished(sh);
3587 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3588 wake_up(&conf->wait_for_overlap);
3590 while (bi && bi->bi_iter.bi_sector <
3591 sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3592 struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
3594 md_write_end(conf->mddev);
3599 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3600 RAID5_STRIPE_SECTORS(conf), 0, 0);
3602 /* and fail all 'written' */
3603 bi = sh->dev[i].written;
3604 sh->dev[i].written = NULL;
3605 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3606 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3607 sh->dev[i].page = sh->dev[i].orig_page;
3610 if (bi) bitmap_end = 1;
3611 while (bi && bi->bi_iter.bi_sector <
3612 sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3613 struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
3615 md_write_end(conf->mddev);
3620 /* fail any reads if this device is non-operational and
3621 * the data has not reached the cache yet.
3623 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3624 s->failed > conf->max_degraded &&
3625 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3626 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3627 spin_lock_irq(&sh->stripe_lock);
3628 bi = sh->dev[i].toread;
3629 sh->dev[i].toread = NULL;
3630 spin_unlock_irq(&sh->stripe_lock);
3631 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3632 wake_up(&conf->wait_for_overlap);
3635 while (bi && bi->bi_iter.bi_sector <
3636 sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3637 struct bio *nextbi =
3638 r5_next_bio(conf, bi, sh->dev[i].sector);
3645 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3646 RAID5_STRIPE_SECTORS(conf), 0, 0);
3647 /* If we were in the middle of a write the parity block might
3648 * still be locked - so just clear all R5_LOCKED flags
3650 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3655 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3656 if (atomic_dec_and_test(&conf->pending_full_writes))
3657 md_wakeup_thread(conf->mddev->thread);
3661 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3662 struct stripe_head_state *s)
3667 BUG_ON(sh->batch_head);
3668 clear_bit(STRIPE_SYNCING, &sh->state);
3669 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3670 wake_up(&conf->wait_for_overlap);
3673 /* There is nothing more to do for sync/check/repair.
3674 * Don't even need to abort as that is handled elsewhere
3675 * if needed, and not always wanted e.g. if there is a known
3677 * For recover/replace we need to record a bad block on all
3678 * non-sync devices, or abort the recovery
3680 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3681 /* During recovery devices cannot be removed, so
3682 * locking and refcounting of rdevs is not needed
3685 for (i = 0; i < conf->raid_disks; i++) {
3686 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3688 && !test_bit(Faulty, &rdev->flags)
3689 && !test_bit(In_sync, &rdev->flags)
3690 && !rdev_set_badblocks(rdev, sh->sector,
3691 RAID5_STRIPE_SECTORS(conf), 0))
3693 rdev = rcu_dereference(conf->disks[i].replacement);
3695 && !test_bit(Faulty, &rdev->flags)
3696 && !test_bit(In_sync, &rdev->flags)
3697 && !rdev_set_badblocks(rdev, sh->sector,
3698 RAID5_STRIPE_SECTORS(conf), 0))
3703 conf->recovery_disabled =
3704 conf->mddev->recovery_disabled;
3706 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
3709 static int want_replace(struct stripe_head *sh, int disk_idx)
3711 struct md_rdev *rdev;
3715 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3717 && !test_bit(Faulty, &rdev->flags)
3718 && !test_bit(In_sync, &rdev->flags)
3719 && (rdev->recovery_offset <= sh->sector
3720 || rdev->mddev->recovery_cp <= sh->sector))
3726 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3727 int disk_idx, int disks)
3729 struct r5dev *dev = &sh->dev[disk_idx];
3730 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3731 &sh->dev[s->failed_num[1]] };
3733 bool force_rcw = (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW);
3736 if (test_bit(R5_LOCKED, &dev->flags) ||
3737 test_bit(R5_UPTODATE, &dev->flags))
3738 /* No point reading this as we already have it or have
3739 * decided to get it.
3744 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3745 /* We need this block to directly satisfy a request */
3748 if (s->syncing || s->expanding ||
3749 (s->replacing && want_replace(sh, disk_idx)))
3750 /* When syncing, or expanding we read everything.
3751 * When replacing, we need the replaced block.
3755 if ((s->failed >= 1 && fdev[0]->toread) ||
3756 (s->failed >= 2 && fdev[1]->toread))
3757 /* If we want to read from a failed device, then
3758 * we need to actually read every other device.
3762 /* Sometimes neither read-modify-write nor reconstruct-write
3763 * cycles can work. In those cases we read every block we
3764 * can. Then the parity-update is certain to have enough to
3766 * This can only be a problem when we need to write something,
3767 * and some device has failed. If either of those tests
3768 * fail we need look no further.
3770 if (!s->failed || !s->to_write)
3773 if (test_bit(R5_Insync, &dev->flags) &&
3774 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3775 /* Pre-reads at not permitted until after short delay
3776 * to gather multiple requests. However if this
3777 * device is no Insync, the block could only be computed
3778 * and there is no need to delay that.
3782 for (i = 0; i < s->failed && i < 2; i++) {
3783 if (fdev[i]->towrite &&
3784 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3785 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3786 /* If we have a partial write to a failed
3787 * device, then we will need to reconstruct
3788 * the content of that device, so all other
3789 * devices must be read.
3793 if (s->failed >= 2 &&
3794 (fdev[i]->towrite ||
3795 s->failed_num[i] == sh->pd_idx ||
3796 s->failed_num[i] == sh->qd_idx) &&
3797 !test_bit(R5_UPTODATE, &fdev[i]->flags))
3798 /* In max degraded raid6, If the failed disk is P, Q,
3799 * or we want to read the failed disk, we need to do
3800 * reconstruct-write.
3805 /* If we are forced to do a reconstruct-write, because parity
3806 * cannot be trusted and we are currently recovering it, there
3807 * is extra need to be careful.
3808 * If one of the devices that we would need to read, because
3809 * it is not being overwritten (and maybe not written at all)
3810 * is missing/faulty, then we need to read everything we can.
3813 sh->sector < sh->raid_conf->mddev->recovery_cp)
3814 /* reconstruct-write isn't being forced */
3816 for (i = 0; i < s->failed && i < 2; i++) {
3817 if (s->failed_num[i] != sh->pd_idx &&
3818 s->failed_num[i] != sh->qd_idx &&
3819 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3820 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3827 /* fetch_block - checks the given member device to see if its data needs
3828 * to be read or computed to satisfy a request.
3830 * Returns 1 when no more member devices need to be checked, otherwise returns
3831 * 0 to tell the loop in handle_stripe_fill to continue
3833 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3834 int disk_idx, int disks)
3836 struct r5dev *dev = &sh->dev[disk_idx];
3838 /* is the data in this block needed, and can we get it? */
3839 if (need_this_block(sh, s, disk_idx, disks)) {
3840 /* we would like to get this block, possibly by computing it,
3841 * otherwise read it if the backing disk is insync
3843 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3844 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3845 BUG_ON(sh->batch_head);
3848 * In the raid6 case if the only non-uptodate disk is P
3849 * then we already trusted P to compute the other failed
3850 * drives. It is safe to compute rather than re-read P.
3851 * In other cases we only compute blocks from failed
3852 * devices, otherwise check/repair might fail to detect
3853 * a real inconsistency.
3856 if ((s->uptodate == disks - 1) &&
3857 ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3858 (s->failed && (disk_idx == s->failed_num[0] ||
3859 disk_idx == s->failed_num[1])))) {
3860 /* have disk failed, and we're requested to fetch it;
3863 pr_debug("Computing stripe %llu block %d\n",
3864 (unsigned long long)sh->sector, disk_idx);
3865 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3866 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3867 set_bit(R5_Wantcompute, &dev->flags);
3868 sh->ops.target = disk_idx;
3869 sh->ops.target2 = -1; /* no 2nd target */
3871 /* Careful: from this point on 'uptodate' is in the eye
3872 * of raid_run_ops which services 'compute' operations
3873 * before writes. R5_Wantcompute flags a block that will
3874 * be R5_UPTODATE by the time it is needed for a
3875 * subsequent operation.
3879 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3880 /* Computing 2-failure is *very* expensive; only
3881 * do it if failed >= 2
3884 for (other = disks; other--; ) {
3885 if (other == disk_idx)
3887 if (!test_bit(R5_UPTODATE,
3888 &sh->dev[other].flags))
3892 pr_debug("Computing stripe %llu blocks %d,%d\n",
3893 (unsigned long long)sh->sector,
3895 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3896 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3897 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3898 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3899 sh->ops.target = disk_idx;
3900 sh->ops.target2 = other;
3904 } else if (test_bit(R5_Insync, &dev->flags)) {
3905 set_bit(R5_LOCKED, &dev->flags);
3906 set_bit(R5_Wantread, &dev->flags);
3908 pr_debug("Reading block %d (sync=%d)\n",
3909 disk_idx, s->syncing);
3917 * handle_stripe_fill - read or compute data to satisfy pending requests.
3919 static void handle_stripe_fill(struct stripe_head *sh,
3920 struct stripe_head_state *s,
3925 /* look for blocks to read/compute, skip this if a compute
3926 * is already in flight, or if the stripe contents are in the
3927 * midst of changing due to a write
3929 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3930 !sh->reconstruct_state) {
3933 * For degraded stripe with data in journal, do not handle
3934 * read requests yet, instead, flush the stripe to raid
3935 * disks first, this avoids handling complex rmw of write
3936 * back cache (prexor with orig_page, and then xor with
3937 * page) in the read path
3939 if (s->injournal && s->failed) {
3940 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3941 r5c_make_stripe_write_out(sh);
3945 for (i = disks; i--; )
3946 if (fetch_block(sh, s, i, disks))
3950 set_bit(STRIPE_HANDLE, &sh->state);
3953 static void break_stripe_batch_list(struct stripe_head *head_sh,
3954 unsigned long handle_flags);
3955 /* handle_stripe_clean_event
3956 * any written block on an uptodate or failed drive can be returned.
3957 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3958 * never LOCKED, so we don't need to test 'failed' directly.
3960 static void handle_stripe_clean_event(struct r5conf *conf,
3961 struct stripe_head *sh, int disks)
3965 int discard_pending = 0;
3966 struct stripe_head *head_sh = sh;
3967 bool do_endio = false;
3969 for (i = disks; i--; )
3970 if (sh->dev[i].written) {
3972 if (!test_bit(R5_LOCKED, &dev->flags) &&
3973 (test_bit(R5_UPTODATE, &dev->flags) ||
3974 test_bit(R5_Discard, &dev->flags) ||
3975 test_bit(R5_SkipCopy, &dev->flags))) {
3976 /* We can return any write requests */
3977 struct bio *wbi, *wbi2;
3978 pr_debug("Return write for disc %d\n", i);
3979 if (test_and_clear_bit(R5_Discard, &dev->flags))
3980 clear_bit(R5_UPTODATE, &dev->flags);
3981 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3982 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3987 dev->page = dev->orig_page;
3989 dev->written = NULL;
3990 while (wbi && wbi->bi_iter.bi_sector <
3991 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
3992 wbi2 = r5_next_bio(conf, wbi, dev->sector);
3993 md_write_end(conf->mddev);
3997 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3998 RAID5_STRIPE_SECTORS(conf),
3999 !test_bit(STRIPE_DEGRADED, &sh->state),
4001 if (head_sh->batch_head) {
4002 sh = list_first_entry(&sh->batch_list,
4005 if (sh != head_sh) {
4012 } else if (test_bit(R5_Discard, &dev->flags))
4013 discard_pending = 1;
4016 log_stripe_write_finished(sh);
4018 if (!discard_pending &&
4019 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
4021 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
4022 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4023 if (sh->qd_idx >= 0) {
4024 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
4025 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
4027 /* now that discard is done we can proceed with any sync */
4028 clear_bit(STRIPE_DISCARD, &sh->state);
4030 * SCSI discard will change some bio fields and the stripe has
4031 * no updated data, so remove it from hash list and the stripe
4032 * will be reinitialized
4035 hash = sh->hash_lock_index;
4036 spin_lock_irq(conf->hash_locks + hash);
4038 spin_unlock_irq(conf->hash_locks + hash);
4039 if (head_sh->batch_head) {
4040 sh = list_first_entry(&sh->batch_list,
4041 struct stripe_head, batch_list);
4047 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
4048 set_bit(STRIPE_HANDLE, &sh->state);
4052 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
4053 if (atomic_dec_and_test(&conf->pending_full_writes))
4054 md_wakeup_thread(conf->mddev->thread);
4056 if (head_sh->batch_head && do_endio)
4057 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
4061 * For RMW in write back cache, we need extra page in prexor to store the
4062 * old data. This page is stored in dev->orig_page.
4064 * This function checks whether we have data for prexor. The exact logic
4066 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
4068 static inline bool uptodate_for_rmw(struct r5dev *dev)
4070 return (test_bit(R5_UPTODATE, &dev->flags)) &&
4071 (!test_bit(R5_InJournal, &dev->flags) ||
4072 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
4075 static int handle_stripe_dirtying(struct r5conf *conf,
4076 struct stripe_head *sh,
4077 struct stripe_head_state *s,
4080 int rmw = 0, rcw = 0, i;
4081 sector_t recovery_cp = conf->mddev->recovery_cp;
4083 /* Check whether resync is now happening or should start.
4084 * If yes, then the array is dirty (after unclean shutdown or
4085 * initial creation), so parity in some stripes might be inconsistent.
4086 * In this case, we need to always do reconstruct-write, to ensure
4087 * that in case of drive failure or read-error correction, we
4088 * generate correct data from the parity.
4090 if (conf->rmw_level == PARITY_DISABLE_RMW ||
4091 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
4093 /* Calculate the real rcw later - for now make it
4094 * look like rcw is cheaper
4097 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
4098 conf->rmw_level, (unsigned long long)recovery_cp,
4099 (unsigned long long)sh->sector);
4100 } else for (i = disks; i--; ) {
4101 /* would I have to read this buffer for read_modify_write */
4102 struct r5dev *dev = &sh->dev[i];
4103 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4104 i == sh->pd_idx || i == sh->qd_idx ||
4105 test_bit(R5_InJournal, &dev->flags)) &&
4106 !test_bit(R5_LOCKED, &dev->flags) &&
4107 !(uptodate_for_rmw(dev) ||
4108 test_bit(R5_Wantcompute, &dev->flags))) {
4109 if (test_bit(R5_Insync, &dev->flags))
4112 rmw += 2*disks; /* cannot read it */
4114 /* Would I have to read this buffer for reconstruct_write */
4115 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4116 i != sh->pd_idx && i != sh->qd_idx &&
4117 !test_bit(R5_LOCKED, &dev->flags) &&
4118 !(test_bit(R5_UPTODATE, &dev->flags) ||
4119 test_bit(R5_Wantcompute, &dev->flags))) {
4120 if (test_bit(R5_Insync, &dev->flags))
4127 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
4128 (unsigned long long)sh->sector, sh->state, rmw, rcw);
4129 set_bit(STRIPE_HANDLE, &sh->state);
4130 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
4131 /* prefer read-modify-write, but need to get some data */
4132 if (conf->mddev->queue)
4133 blk_add_trace_msg(conf->mddev->queue,
4134 "raid5 rmw %llu %d",
4135 (unsigned long long)sh->sector, rmw);
4136 for (i = disks; i--; ) {
4137 struct r5dev *dev = &sh->dev[i];
4138 if (test_bit(R5_InJournal, &dev->flags) &&
4139 dev->page == dev->orig_page &&
4140 !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
4141 /* alloc page for prexor */
4142 struct page *p = alloc_page(GFP_NOIO);
4150 * alloc_page() failed, try use
4151 * disk_info->extra_page
4153 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
4154 &conf->cache_state)) {
4155 r5c_use_extra_page(sh);
4159 /* extra_page in use, add to delayed_list */
4160 set_bit(STRIPE_DELAYED, &sh->state);
4161 s->waiting_extra_page = 1;
4166 for (i = disks; i--; ) {
4167 struct r5dev *dev = &sh->dev[i];
4168 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4169 i == sh->pd_idx || i == sh->qd_idx ||
4170 test_bit(R5_InJournal, &dev->flags)) &&
4171 !test_bit(R5_LOCKED, &dev->flags) &&
4172 !(uptodate_for_rmw(dev) ||
4173 test_bit(R5_Wantcompute, &dev->flags)) &&
4174 test_bit(R5_Insync, &dev->flags)) {
4175 if (test_bit(STRIPE_PREREAD_ACTIVE,
4177 pr_debug("Read_old block %d for r-m-w\n",
4179 set_bit(R5_LOCKED, &dev->flags);
4180 set_bit(R5_Wantread, &dev->flags);
4183 set_bit(STRIPE_DELAYED, &sh->state);
4187 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
4188 /* want reconstruct write, but need to get some data */
4191 for (i = disks; i--; ) {
4192 struct r5dev *dev = &sh->dev[i];
4193 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4194 i != sh->pd_idx && i != sh->qd_idx &&
4195 !test_bit(R5_LOCKED, &dev->flags) &&
4196 !(test_bit(R5_UPTODATE, &dev->flags) ||
4197 test_bit(R5_Wantcompute, &dev->flags))) {
4199 if (test_bit(R5_Insync, &dev->flags) &&
4200 test_bit(STRIPE_PREREAD_ACTIVE,
4202 pr_debug("Read_old block "
4203 "%d for Reconstruct\n", i);
4204 set_bit(R5_LOCKED, &dev->flags);
4205 set_bit(R5_Wantread, &dev->flags);
4209 set_bit(STRIPE_DELAYED, &sh->state);
4212 if (rcw && conf->mddev->queue)
4213 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4214 (unsigned long long)sh->sector,
4215 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4218 if (rcw > disks && rmw > disks &&
4219 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4220 set_bit(STRIPE_DELAYED, &sh->state);
4222 /* now if nothing is locked, and if we have enough data,
4223 * we can start a write request
4225 /* since handle_stripe can be called at any time we need to handle the
4226 * case where a compute block operation has been submitted and then a
4227 * subsequent call wants to start a write request. raid_run_ops only
4228 * handles the case where compute block and reconstruct are requested
4229 * simultaneously. If this is not the case then new writes need to be
4230 * held off until the compute completes.
4232 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4233 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4234 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4235 schedule_reconstruction(sh, s, rcw == 0, 0);
4239 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4240 struct stripe_head_state *s, int disks)
4242 struct r5dev *dev = NULL;
4244 BUG_ON(sh->batch_head);
4245 set_bit(STRIPE_HANDLE, &sh->state);
4247 switch (sh->check_state) {
4248 case check_state_idle:
4249 /* start a new check operation if there are no failures */
4250 if (s->failed == 0) {
4251 BUG_ON(s->uptodate != disks);
4252 sh->check_state = check_state_run;
4253 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4254 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4258 dev = &sh->dev[s->failed_num[0]];
4260 case check_state_compute_result:
4261 sh->check_state = check_state_idle;
4263 dev = &sh->dev[sh->pd_idx];
4265 /* check that a write has not made the stripe insync */
4266 if (test_bit(STRIPE_INSYNC, &sh->state))
4269 /* either failed parity check, or recovery is happening */
4270 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4271 BUG_ON(s->uptodate != disks);
4273 set_bit(R5_LOCKED, &dev->flags);
4275 set_bit(R5_Wantwrite, &dev->flags);
4277 clear_bit(STRIPE_DEGRADED, &sh->state);
4278 set_bit(STRIPE_INSYNC, &sh->state);
4280 case check_state_run:
4281 break; /* we will be called again upon completion */
4282 case check_state_check_result:
4283 sh->check_state = check_state_idle;
4285 /* if a failure occurred during the check operation, leave
4286 * STRIPE_INSYNC not set and let the stripe be handled again
4291 /* handle a successful check operation, if parity is correct
4292 * we are done. Otherwise update the mismatch count and repair
4293 * parity if !MD_RECOVERY_CHECK
4295 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4296 /* parity is correct (on disc,
4297 * not in buffer any more)
4299 set_bit(STRIPE_INSYNC, &sh->state);
4301 atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4302 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4303 /* don't try to repair!! */
4304 set_bit(STRIPE_INSYNC, &sh->state);
4305 pr_warn_ratelimited("%s: mismatch sector in range "
4306 "%llu-%llu\n", mdname(conf->mddev),
4307 (unsigned long long) sh->sector,
4308 (unsigned long long) sh->sector +
4309 RAID5_STRIPE_SECTORS(conf));
4311 sh->check_state = check_state_compute_run;
4312 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4313 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4314 set_bit(R5_Wantcompute,
4315 &sh->dev[sh->pd_idx].flags);
4316 sh->ops.target = sh->pd_idx;
4317 sh->ops.target2 = -1;
4322 case check_state_compute_run:
4325 pr_err("%s: unknown check_state: %d sector: %llu\n",
4326 __func__, sh->check_state,
4327 (unsigned long long) sh->sector);
4332 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4333 struct stripe_head_state *s,
4336 int pd_idx = sh->pd_idx;
4337 int qd_idx = sh->qd_idx;
4340 BUG_ON(sh->batch_head);
4341 set_bit(STRIPE_HANDLE, &sh->state);
4343 BUG_ON(s->failed > 2);
4345 /* Want to check and possibly repair P and Q.
4346 * However there could be one 'failed' device, in which
4347 * case we can only check one of them, possibly using the
4348 * other to generate missing data
4351 switch (sh->check_state) {
4352 case check_state_idle:
4353 /* start a new check operation if there are < 2 failures */
4354 if (s->failed == s->q_failed) {
4355 /* The only possible failed device holds Q, so it
4356 * makes sense to check P (If anything else were failed,
4357 * we would have used P to recreate it).
4359 sh->check_state = check_state_run;
4361 if (!s->q_failed && s->failed < 2) {
4362 /* Q is not failed, and we didn't use it to generate
4363 * anything, so it makes sense to check it
4365 if (sh->check_state == check_state_run)
4366 sh->check_state = check_state_run_pq;
4368 sh->check_state = check_state_run_q;
4371 /* discard potentially stale zero_sum_result */
4372 sh->ops.zero_sum_result = 0;
4374 if (sh->check_state == check_state_run) {
4375 /* async_xor_zero_sum destroys the contents of P */
4376 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4379 if (sh->check_state >= check_state_run &&
4380 sh->check_state <= check_state_run_pq) {
4381 /* async_syndrome_zero_sum preserves P and Q, so
4382 * no need to mark them !uptodate here
4384 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4388 /* we have 2-disk failure */
4389 BUG_ON(s->failed != 2);
4391 case check_state_compute_result:
4392 sh->check_state = check_state_idle;
4394 /* check that a write has not made the stripe insync */
4395 if (test_bit(STRIPE_INSYNC, &sh->state))
4398 /* now write out any block on a failed drive,
4399 * or P or Q if they were recomputed
4402 if (s->failed == 2) {
4403 dev = &sh->dev[s->failed_num[1]];
4405 set_bit(R5_LOCKED, &dev->flags);
4406 set_bit(R5_Wantwrite, &dev->flags);
4408 if (s->failed >= 1) {
4409 dev = &sh->dev[s->failed_num[0]];
4411 set_bit(R5_LOCKED, &dev->flags);
4412 set_bit(R5_Wantwrite, &dev->flags);
4414 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4415 dev = &sh->dev[pd_idx];
4417 set_bit(R5_LOCKED, &dev->flags);
4418 set_bit(R5_Wantwrite, &dev->flags);
4420 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4421 dev = &sh->dev[qd_idx];
4423 set_bit(R5_LOCKED, &dev->flags);
4424 set_bit(R5_Wantwrite, &dev->flags);
4426 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4427 "%s: disk%td not up to date\n",
4428 mdname(conf->mddev),
4429 dev - (struct r5dev *) &sh->dev)) {
4430 clear_bit(R5_LOCKED, &dev->flags);
4431 clear_bit(R5_Wantwrite, &dev->flags);
4434 clear_bit(STRIPE_DEGRADED, &sh->state);
4436 set_bit(STRIPE_INSYNC, &sh->state);
4438 case check_state_run:
4439 case check_state_run_q:
4440 case check_state_run_pq:
4441 break; /* we will be called again upon completion */
4442 case check_state_check_result:
4443 sh->check_state = check_state_idle;
4445 /* handle a successful check operation, if parity is correct
4446 * we are done. Otherwise update the mismatch count and repair
4447 * parity if !MD_RECOVERY_CHECK
4449 if (sh->ops.zero_sum_result == 0) {
4450 /* both parities are correct */
4452 set_bit(STRIPE_INSYNC, &sh->state);
4454 /* in contrast to the raid5 case we can validate
4455 * parity, but still have a failure to write
4458 sh->check_state = check_state_compute_result;
4459 /* Returning at this point means that we may go
4460 * off and bring p and/or q uptodate again so
4461 * we make sure to check zero_sum_result again
4462 * to verify if p or q need writeback
4466 atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4467 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4468 /* don't try to repair!! */
4469 set_bit(STRIPE_INSYNC, &sh->state);
4470 pr_warn_ratelimited("%s: mismatch sector in range "
4471 "%llu-%llu\n", mdname(conf->mddev),
4472 (unsigned long long) sh->sector,
4473 (unsigned long long) sh->sector +
4474 RAID5_STRIPE_SECTORS(conf));
4476 int *target = &sh->ops.target;
4478 sh->ops.target = -1;
4479 sh->ops.target2 = -1;
4480 sh->check_state = check_state_compute_run;
4481 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4482 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4483 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4484 set_bit(R5_Wantcompute,
4485 &sh->dev[pd_idx].flags);
4487 target = &sh->ops.target2;
4490 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4491 set_bit(R5_Wantcompute,
4492 &sh->dev[qd_idx].flags);
4499 case check_state_compute_run:
4502 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4503 __func__, sh->check_state,
4504 (unsigned long long) sh->sector);
4509 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4513 /* We have read all the blocks in this stripe and now we need to
4514 * copy some of them into a target stripe for expand.
4516 struct dma_async_tx_descriptor *tx = NULL;
4517 BUG_ON(sh->batch_head);
4518 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4519 for (i = 0; i < sh->disks; i++)
4520 if (i != sh->pd_idx && i != sh->qd_idx) {
4522 struct stripe_head *sh2;
4523 struct async_submit_ctl submit;
4525 sector_t bn = raid5_compute_blocknr(sh, i, 1);
4526 sector_t s = raid5_compute_sector(conf, bn, 0,
4528 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4530 /* so far only the early blocks of this stripe
4531 * have been requested. When later blocks
4532 * get requested, we will try again
4535 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4536 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4537 /* must have already done this block */
4538 raid5_release_stripe(sh2);
4542 /* place all the copies on one channel */
4543 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4544 tx = async_memcpy(sh2->dev[dd_idx].page,
4545 sh->dev[i].page, sh2->dev[dd_idx].offset,
4546 sh->dev[i].offset, RAID5_STRIPE_SIZE(conf),
4549 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4550 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4551 for (j = 0; j < conf->raid_disks; j++)
4552 if (j != sh2->pd_idx &&
4554 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4556 if (j == conf->raid_disks) {
4557 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4558 set_bit(STRIPE_HANDLE, &sh2->state);
4560 raid5_release_stripe(sh2);
4563 /* done submitting copies, wait for them to complete */
4564 async_tx_quiesce(&tx);
4568 * handle_stripe - do things to a stripe.
4570 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4571 * state of various bits to see what needs to be done.
4573 * return some read requests which now have data
4574 * return some write requests which are safely on storage
4575 * schedule a read on some buffers
4576 * schedule a write of some buffers
4577 * return confirmation of parity correctness
4581 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4583 struct r5conf *conf = sh->raid_conf;
4584 int disks = sh->disks;
4587 int do_recovery = 0;
4589 memset(s, 0, sizeof(*s));
4591 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4592 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4593 s->failed_num[0] = -1;
4594 s->failed_num[1] = -1;
4595 s->log_failed = r5l_log_disk_error(conf);
4597 /* Now to look around and see what can be done */
4599 for (i=disks; i--; ) {
4600 struct md_rdev *rdev;
4607 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4609 dev->toread, dev->towrite, dev->written);
4610 /* maybe we can reply to a read
4612 * new wantfill requests are only permitted while
4613 * ops_complete_biofill is guaranteed to be inactive
4615 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4616 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4617 set_bit(R5_Wantfill, &dev->flags);
4619 /* now count some things */
4620 if (test_bit(R5_LOCKED, &dev->flags))
4622 if (test_bit(R5_UPTODATE, &dev->flags))
4624 if (test_bit(R5_Wantcompute, &dev->flags)) {
4626 BUG_ON(s->compute > 2);
4629 if (test_bit(R5_Wantfill, &dev->flags))
4631 else if (dev->toread)
4635 if (!test_bit(R5_OVERWRITE, &dev->flags))
4640 /* Prefer to use the replacement for reads, but only
4641 * if it is recovered enough and has no bad blocks.
4643 rdev = rcu_dereference(conf->disks[i].replacement);
4644 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4645 rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
4646 !is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4647 &first_bad, &bad_sectors))
4648 set_bit(R5_ReadRepl, &dev->flags);
4650 if (rdev && !test_bit(Faulty, &rdev->flags))
4651 set_bit(R5_NeedReplace, &dev->flags);
4653 clear_bit(R5_NeedReplace, &dev->flags);
4654 rdev = rcu_dereference(conf->disks[i].rdev);
4655 clear_bit(R5_ReadRepl, &dev->flags);
4657 if (rdev && test_bit(Faulty, &rdev->flags))
4660 is_bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4661 &first_bad, &bad_sectors);
4662 if (s->blocked_rdev == NULL
4663 && (test_bit(Blocked, &rdev->flags)
4666 set_bit(BlockedBadBlocks,
4668 s->blocked_rdev = rdev;
4669 atomic_inc(&rdev->nr_pending);
4672 clear_bit(R5_Insync, &dev->flags);
4676 /* also not in-sync */
4677 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4678 test_bit(R5_UPTODATE, &dev->flags)) {
4679 /* treat as in-sync, but with a read error
4680 * which we can now try to correct
4682 set_bit(R5_Insync, &dev->flags);
4683 set_bit(R5_ReadError, &dev->flags);
4685 } else if (test_bit(In_sync, &rdev->flags))
4686 set_bit(R5_Insync, &dev->flags);
4687 else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <= rdev->recovery_offset)
4688 /* in sync if before recovery_offset */
4689 set_bit(R5_Insync, &dev->flags);
4690 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4691 test_bit(R5_Expanded, &dev->flags))
4692 /* If we've reshaped into here, we assume it is Insync.
4693 * We will shortly update recovery_offset to make
4696 set_bit(R5_Insync, &dev->flags);
4698 if (test_bit(R5_WriteError, &dev->flags)) {
4699 /* This flag does not apply to '.replacement'
4700 * only to .rdev, so make sure to check that*/
4701 struct md_rdev *rdev2 = rcu_dereference(
4702 conf->disks[i].rdev);
4704 clear_bit(R5_Insync, &dev->flags);
4705 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4706 s->handle_bad_blocks = 1;
4707 atomic_inc(&rdev2->nr_pending);
4709 clear_bit(R5_WriteError, &dev->flags);
4711 if (test_bit(R5_MadeGood, &dev->flags)) {
4712 /* This flag does not apply to '.replacement'
4713 * only to .rdev, so make sure to check that*/
4714 struct md_rdev *rdev2 = rcu_dereference(
4715 conf->disks[i].rdev);
4716 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4717 s->handle_bad_blocks = 1;
4718 atomic_inc(&rdev2->nr_pending);
4720 clear_bit(R5_MadeGood, &dev->flags);
4722 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4723 struct md_rdev *rdev2 = rcu_dereference(
4724 conf->disks[i].replacement);
4725 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4726 s->handle_bad_blocks = 1;
4727 atomic_inc(&rdev2->nr_pending);
4729 clear_bit(R5_MadeGoodRepl, &dev->flags);
4731 if (!test_bit(R5_Insync, &dev->flags)) {
4732 /* The ReadError flag will just be confusing now */
4733 clear_bit(R5_ReadError, &dev->flags);
4734 clear_bit(R5_ReWrite, &dev->flags);
4736 if (test_bit(R5_ReadError, &dev->flags))
4737 clear_bit(R5_Insync, &dev->flags);
4738 if (!test_bit(R5_Insync, &dev->flags)) {
4740 s->failed_num[s->failed] = i;
4742 if (rdev && !test_bit(Faulty, &rdev->flags))
4745 rdev = rcu_dereference(
4746 conf->disks[i].replacement);
4747 if (rdev && !test_bit(Faulty, &rdev->flags))
4752 if (test_bit(R5_InJournal, &dev->flags))
4754 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4757 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4758 /* If there is a failed device being replaced,
4759 * we must be recovering.
4760 * else if we are after recovery_cp, we must be syncing
4761 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4762 * else we can only be replacing
4763 * sync and recovery both need to read all devices, and so
4764 * use the same flag.
4767 sh->sector >= conf->mddev->recovery_cp ||
4768 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4777 * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4778 * a head which can now be handled.
4780 static int clear_batch_ready(struct stripe_head *sh)
4782 struct stripe_head *tmp;
4783 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4784 return (sh->batch_head && sh->batch_head != sh);
4785 spin_lock(&sh->stripe_lock);
4786 if (!sh->batch_head) {
4787 spin_unlock(&sh->stripe_lock);
4792 * this stripe could be added to a batch list before we check
4793 * BATCH_READY, skips it
4795 if (sh->batch_head != sh) {
4796 spin_unlock(&sh->stripe_lock);
4799 spin_lock(&sh->batch_lock);
4800 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4801 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4802 spin_unlock(&sh->batch_lock);
4803 spin_unlock(&sh->stripe_lock);
4806 * BATCH_READY is cleared, no new stripes can be added.
4807 * batch_list can be accessed without lock
4812 static void break_stripe_batch_list(struct stripe_head *head_sh,
4813 unsigned long handle_flags)
4815 struct stripe_head *sh, *next;
4819 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4821 list_del_init(&sh->batch_list);
4823 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4824 (1 << STRIPE_SYNCING) |
4825 (1 << STRIPE_REPLACED) |
4826 (1 << STRIPE_DELAYED) |
4827 (1 << STRIPE_BIT_DELAY) |
4828 (1 << STRIPE_FULL_WRITE) |
4829 (1 << STRIPE_BIOFILL_RUN) |
4830 (1 << STRIPE_COMPUTE_RUN) |
4831 (1 << STRIPE_DISCARD) |
4832 (1 << STRIPE_BATCH_READY) |
4833 (1 << STRIPE_BATCH_ERR) |
4834 (1 << STRIPE_BITMAP_PENDING)),
4835 "stripe state: %lx\n", sh->state);
4836 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4837 (1 << STRIPE_REPLACED)),
4838 "head stripe state: %lx\n", head_sh->state);
4840 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4841 (1 << STRIPE_PREREAD_ACTIVE) |
4842 (1 << STRIPE_DEGRADED) |
4843 (1 << STRIPE_ON_UNPLUG_LIST)),
4844 head_sh->state & (1 << STRIPE_INSYNC));
4846 sh->check_state = head_sh->check_state;
4847 sh->reconstruct_state = head_sh->reconstruct_state;
4848 spin_lock_irq(&sh->stripe_lock);
4849 sh->batch_head = NULL;
4850 spin_unlock_irq(&sh->stripe_lock);
4851 for (i = 0; i < sh->disks; i++) {
4852 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4854 sh->dev[i].flags = head_sh->dev[i].flags &
4855 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4857 if (handle_flags == 0 ||
4858 sh->state & handle_flags)
4859 set_bit(STRIPE_HANDLE, &sh->state);
4860 raid5_release_stripe(sh);
4862 spin_lock_irq(&head_sh->stripe_lock);
4863 head_sh->batch_head = NULL;
4864 spin_unlock_irq(&head_sh->stripe_lock);
4865 for (i = 0; i < head_sh->disks; i++)
4866 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4868 if (head_sh->state & handle_flags)
4869 set_bit(STRIPE_HANDLE, &head_sh->state);
4872 wake_up(&head_sh->raid_conf->wait_for_overlap);
4875 static void handle_stripe(struct stripe_head *sh)
4877 struct stripe_head_state s;
4878 struct r5conf *conf = sh->raid_conf;
4881 int disks = sh->disks;
4882 struct r5dev *pdev, *qdev;
4884 clear_bit(STRIPE_HANDLE, &sh->state);
4887 * handle_stripe should not continue handle the batched stripe, only
4888 * the head of batch list or lone stripe can continue. Otherwise we
4889 * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
4890 * is set for the batched stripe.
4892 if (clear_batch_ready(sh))
4895 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4896 /* already being handled, ensure it gets handled
4897 * again when current action finishes */
4898 set_bit(STRIPE_HANDLE, &sh->state);
4902 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4903 break_stripe_batch_list(sh, 0);
4905 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4906 spin_lock(&sh->stripe_lock);
4908 * Cannot process 'sync' concurrently with 'discard'.
4909 * Flush data in r5cache before 'sync'.
4911 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4912 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4913 !test_bit(STRIPE_DISCARD, &sh->state) &&
4914 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4915 set_bit(STRIPE_SYNCING, &sh->state);
4916 clear_bit(STRIPE_INSYNC, &sh->state);
4917 clear_bit(STRIPE_REPLACED, &sh->state);
4919 spin_unlock(&sh->stripe_lock);
4921 clear_bit(STRIPE_DELAYED, &sh->state);
4923 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4924 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4925 (unsigned long long)sh->sector, sh->state,
4926 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4927 sh->check_state, sh->reconstruct_state);
4929 analyse_stripe(sh, &s);
4931 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4934 if (s.handle_bad_blocks ||
4935 test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4936 set_bit(STRIPE_HANDLE, &sh->state);
4940 if (unlikely(s.blocked_rdev)) {
4941 if (s.syncing || s.expanding || s.expanded ||
4942 s.replacing || s.to_write || s.written) {
4943 set_bit(STRIPE_HANDLE, &sh->state);
4946 /* There is nothing for the blocked_rdev to block */
4947 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4948 s.blocked_rdev = NULL;
4951 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4952 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4953 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4956 pr_debug("locked=%d uptodate=%d to_read=%d"
4957 " to_write=%d failed=%d failed_num=%d,%d\n",
4958 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4959 s.failed_num[0], s.failed_num[1]);
4961 * check if the array has lost more than max_degraded devices and,
4962 * if so, some requests might need to be failed.
4964 * When journal device failed (log_failed), we will only process
4965 * the stripe if there is data need write to raid disks
4967 if (s.failed > conf->max_degraded ||
4968 (s.log_failed && s.injournal == 0)) {
4969 sh->check_state = 0;
4970 sh->reconstruct_state = 0;
4971 break_stripe_batch_list(sh, 0);
4972 if (s.to_read+s.to_write+s.written)
4973 handle_failed_stripe(conf, sh, &s, disks);
4974 if (s.syncing + s.replacing)
4975 handle_failed_sync(conf, sh, &s);
4978 /* Now we check to see if any write operations have recently
4982 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4984 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4985 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4986 sh->reconstruct_state = reconstruct_state_idle;
4988 /* All the 'written' buffers and the parity block are ready to
4989 * be written back to disk
4991 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4992 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4993 BUG_ON(sh->qd_idx >= 0 &&
4994 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4995 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4996 for (i = disks; i--; ) {
4997 struct r5dev *dev = &sh->dev[i];
4998 if (test_bit(R5_LOCKED, &dev->flags) &&
4999 (i == sh->pd_idx || i == sh->qd_idx ||
5000 dev->written || test_bit(R5_InJournal,
5002 pr_debug("Writing block %d\n", i);
5003 set_bit(R5_Wantwrite, &dev->flags);
5008 if (!test_bit(R5_Insync, &dev->flags) ||
5009 ((i == sh->pd_idx || i == sh->qd_idx) &&
5011 set_bit(STRIPE_INSYNC, &sh->state);
5014 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5015 s.dec_preread_active = 1;
5019 * might be able to return some write requests if the parity blocks
5020 * are safe, or on a failed drive
5022 pdev = &sh->dev[sh->pd_idx];
5023 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
5024 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
5025 qdev = &sh->dev[sh->qd_idx];
5026 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
5027 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
5031 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
5032 && !test_bit(R5_LOCKED, &pdev->flags)
5033 && (test_bit(R5_UPTODATE, &pdev->flags) ||
5034 test_bit(R5_Discard, &pdev->flags))))) &&
5035 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
5036 && !test_bit(R5_LOCKED, &qdev->flags)
5037 && (test_bit(R5_UPTODATE, &qdev->flags) ||
5038 test_bit(R5_Discard, &qdev->flags))))))
5039 handle_stripe_clean_event(conf, sh, disks);
5042 r5c_handle_cached_data_endio(conf, sh, disks);
5043 log_stripe_write_finished(sh);
5045 /* Now we might consider reading some blocks, either to check/generate
5046 * parity, or to satisfy requests
5047 * or to load a block that is being partially written.
5049 if (s.to_read || s.non_overwrite
5050 || (s.to_write && s.failed)
5051 || (s.syncing && (s.uptodate + s.compute < disks))
5054 handle_stripe_fill(sh, &s, disks);
5057 * When the stripe finishes full journal write cycle (write to journal
5058 * and raid disk), this is the clean up procedure so it is ready for
5061 r5c_finish_stripe_write_out(conf, sh, &s);
5064 * Now to consider new write requests, cache write back and what else,
5065 * if anything should be read. We do not handle new writes when:
5066 * 1/ A 'write' operation (copy+xor) is already in flight.
5067 * 2/ A 'check' operation is in flight, as it may clobber the parity
5069 * 3/ A r5c cache log write is in flight.
5072 if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
5073 if (!r5c_is_writeback(conf->log)) {
5075 handle_stripe_dirtying(conf, sh, &s, disks);
5076 } else { /* write back cache */
5079 /* First, try handle writes in caching phase */
5081 ret = r5c_try_caching_write(conf, sh, &s,
5084 * If caching phase failed: ret == -EAGAIN
5086 * stripe under reclaim: !caching && injournal
5088 * fall back to handle_stripe_dirtying()
5090 if (ret == -EAGAIN ||
5091 /* stripe under reclaim: !caching && injournal */
5092 (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
5094 ret = handle_stripe_dirtying(conf, sh, &s,
5102 /* maybe we need to check and possibly fix the parity for this stripe
5103 * Any reads will already have been scheduled, so we just see if enough
5104 * data is available. The parity check is held off while parity
5105 * dependent operations are in flight.
5107 if (sh->check_state ||
5108 (s.syncing && s.locked == 0 &&
5109 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5110 !test_bit(STRIPE_INSYNC, &sh->state))) {
5111 if (conf->level == 6)
5112 handle_parity_checks6(conf, sh, &s, disks);
5114 handle_parity_checks5(conf, sh, &s, disks);
5117 if ((s.replacing || s.syncing) && s.locked == 0
5118 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
5119 && !test_bit(STRIPE_REPLACED, &sh->state)) {
5120 /* Write out to replacement devices where possible */
5121 for (i = 0; i < conf->raid_disks; i++)
5122 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
5123 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
5124 set_bit(R5_WantReplace, &sh->dev[i].flags);
5125 set_bit(R5_LOCKED, &sh->dev[i].flags);
5129 set_bit(STRIPE_INSYNC, &sh->state);
5130 set_bit(STRIPE_REPLACED, &sh->state);
5132 if ((s.syncing || s.replacing) && s.locked == 0 &&
5133 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5134 test_bit(STRIPE_INSYNC, &sh->state)) {
5135 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5136 clear_bit(STRIPE_SYNCING, &sh->state);
5137 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
5138 wake_up(&conf->wait_for_overlap);
5141 /* If the failed drives are just a ReadError, then we might need
5142 * to progress the repair/check process
5144 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
5145 for (i = 0; i < s.failed; i++) {
5146 struct r5dev *dev = &sh->dev[s.failed_num[i]];
5147 if (test_bit(R5_ReadError, &dev->flags)
5148 && !test_bit(R5_LOCKED, &dev->flags)
5149 && test_bit(R5_UPTODATE, &dev->flags)
5151 if (!test_bit(R5_ReWrite, &dev->flags)) {
5152 set_bit(R5_Wantwrite, &dev->flags);
5153 set_bit(R5_ReWrite, &dev->flags);
5155 /* let's read it back */
5156 set_bit(R5_Wantread, &dev->flags);
5157 set_bit(R5_LOCKED, &dev->flags);
5162 /* Finish reconstruct operations initiated by the expansion process */
5163 if (sh->reconstruct_state == reconstruct_state_result) {
5164 struct stripe_head *sh_src
5165 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
5166 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
5167 /* sh cannot be written until sh_src has been read.
5168 * so arrange for sh to be delayed a little
5170 set_bit(STRIPE_DELAYED, &sh->state);
5171 set_bit(STRIPE_HANDLE, &sh->state);
5172 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
5174 atomic_inc(&conf->preread_active_stripes);
5175 raid5_release_stripe(sh_src);
5179 raid5_release_stripe(sh_src);
5181 sh->reconstruct_state = reconstruct_state_idle;
5182 clear_bit(STRIPE_EXPANDING, &sh->state);
5183 for (i = conf->raid_disks; i--; ) {
5184 set_bit(R5_Wantwrite, &sh->dev[i].flags);
5185 set_bit(R5_LOCKED, &sh->dev[i].flags);
5190 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
5191 !sh->reconstruct_state) {
5192 /* Need to write out all blocks after computing parity */
5193 sh->disks = conf->raid_disks;
5194 stripe_set_idx(sh->sector, conf, 0, sh);
5195 schedule_reconstruction(sh, &s, 1, 1);
5196 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
5197 clear_bit(STRIPE_EXPAND_READY, &sh->state);
5198 atomic_dec(&conf->reshape_stripes);
5199 wake_up(&conf->wait_for_overlap);
5200 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5203 if (s.expanding && s.locked == 0 &&
5204 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
5205 handle_stripe_expansion(conf, sh);
5208 /* wait for this device to become unblocked */
5209 if (unlikely(s.blocked_rdev)) {
5210 if (conf->mddev->external)
5211 md_wait_for_blocked_rdev(s.blocked_rdev,
5214 /* Internal metadata will immediately
5215 * be written by raid5d, so we don't
5216 * need to wait here.
5218 rdev_dec_pending(s.blocked_rdev,
5222 if (s.handle_bad_blocks)
5223 for (i = disks; i--; ) {
5224 struct md_rdev *rdev;
5225 struct r5dev *dev = &sh->dev[i];
5226 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5227 /* We own a safe reference to the rdev */
5228 rdev = conf->disks[i].rdev;
5229 if (!rdev_set_badblocks(rdev, sh->sector,
5230 RAID5_STRIPE_SECTORS(conf), 0))
5231 md_error(conf->mddev, rdev);
5232 rdev_dec_pending(rdev, conf->mddev);
5234 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5235 rdev = conf->disks[i].rdev;
5236 rdev_clear_badblocks(rdev, sh->sector,
5237 RAID5_STRIPE_SECTORS(conf), 0);
5238 rdev_dec_pending(rdev, conf->mddev);
5240 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5241 rdev = conf->disks[i].replacement;
5243 /* rdev have been moved down */
5244 rdev = conf->disks[i].rdev;
5245 rdev_clear_badblocks(rdev, sh->sector,
5246 RAID5_STRIPE_SECTORS(conf), 0);
5247 rdev_dec_pending(rdev, conf->mddev);
5252 raid_run_ops(sh, s.ops_request);
5256 if (s.dec_preread_active) {
5257 /* We delay this until after ops_run_io so that if make_request
5258 * is waiting on a flush, it won't continue until the writes
5259 * have actually been submitted.
5261 atomic_dec(&conf->preread_active_stripes);
5262 if (atomic_read(&conf->preread_active_stripes) <
5264 md_wakeup_thread(conf->mddev->thread);
5267 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5270 static void raid5_activate_delayed(struct r5conf *conf)
5272 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5273 while (!list_empty(&conf->delayed_list)) {
5274 struct list_head *l = conf->delayed_list.next;
5275 struct stripe_head *sh;
5276 sh = list_entry(l, struct stripe_head, lru);
5278 clear_bit(STRIPE_DELAYED, &sh->state);
5279 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5280 atomic_inc(&conf->preread_active_stripes);
5281 list_add_tail(&sh->lru, &conf->hold_list);
5282 raid5_wakeup_stripe_thread(sh);
5287 static void activate_bit_delay(struct r5conf *conf,
5288 struct list_head *temp_inactive_list)
5290 /* device_lock is held */
5291 struct list_head head;
5292 list_add(&head, &conf->bitmap_list);
5293 list_del_init(&conf->bitmap_list);
5294 while (!list_empty(&head)) {
5295 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5297 list_del_init(&sh->lru);
5298 atomic_inc(&sh->count);
5299 hash = sh->hash_lock_index;
5300 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5304 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5306 struct r5conf *conf = mddev->private;
5307 sector_t sector = bio->bi_iter.bi_sector;
5308 unsigned int chunk_sectors;
5309 unsigned int bio_sectors = bio_sectors(bio);
5311 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5312 return chunk_sectors >=
5313 ((sector & (chunk_sectors - 1)) + bio_sectors);
5317 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5318 * later sampled by raid5d.
5320 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5322 unsigned long flags;
5324 spin_lock_irqsave(&conf->device_lock, flags);
5326 bi->bi_next = conf->retry_read_aligned_list;
5327 conf->retry_read_aligned_list = bi;
5329 spin_unlock_irqrestore(&conf->device_lock, flags);
5330 md_wakeup_thread(conf->mddev->thread);
5333 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5334 unsigned int *offset)
5338 bi = conf->retry_read_aligned;
5340 *offset = conf->retry_read_offset;
5341 conf->retry_read_aligned = NULL;
5344 bi = conf->retry_read_aligned_list;
5346 conf->retry_read_aligned_list = bi->bi_next;
5355 * The "raid5_align_endio" should check if the read succeeded and if it
5356 * did, call bio_endio on the original bio (having bio_put the new bio
5358 * If the read failed..
5360 static void raid5_align_endio(struct bio *bi)
5362 struct md_io_acct *md_io_acct = bi->bi_private;
5363 struct bio *raid_bi = md_io_acct->orig_bio;
5364 struct mddev *mddev;
5365 struct r5conf *conf;
5366 struct md_rdev *rdev;
5367 blk_status_t error = bi->bi_status;
5368 unsigned long start_time = md_io_acct->start_time;
5372 rdev = (void*)raid_bi->bi_next;
5373 raid_bi->bi_next = NULL;
5374 mddev = rdev->mddev;
5375 conf = mddev->private;
5377 rdev_dec_pending(rdev, conf->mddev);
5380 if (blk_queue_io_stat(raid_bi->bi_bdev->bd_disk->queue))
5381 bio_end_io_acct(raid_bi, start_time);
5383 if (atomic_dec_and_test(&conf->active_aligned_reads))
5384 wake_up(&conf->wait_for_quiescent);
5388 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5390 add_bio_to_retry(raid_bi, conf);
5393 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5395 struct r5conf *conf = mddev->private;
5396 struct bio *align_bio;
5397 struct md_rdev *rdev;
5398 sector_t sector, end_sector, first_bad;
5399 int bad_sectors, dd_idx;
5400 struct md_io_acct *md_io_acct;
5403 if (!in_chunk_boundary(mddev, raid_bio)) {
5404 pr_debug("%s: non aligned\n", __func__);
5408 sector = raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector, 0,
5410 end_sector = bio_end_sector(raid_bio);
5413 if (r5c_big_stripe_cached(conf, sector))
5414 goto out_rcu_unlock;
5416 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5417 if (!rdev || test_bit(Faulty, &rdev->flags) ||
5418 rdev->recovery_offset < end_sector) {
5419 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5421 goto out_rcu_unlock;
5422 if (test_bit(Faulty, &rdev->flags) ||
5423 !(test_bit(In_sync, &rdev->flags) ||
5424 rdev->recovery_offset >= end_sector))
5425 goto out_rcu_unlock;
5428 atomic_inc(&rdev->nr_pending);
5431 if (is_badblock(rdev, sector, bio_sectors(raid_bio), &first_bad,
5434 rdev_dec_pending(rdev, mddev);
5438 align_bio = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->io_acct_set);
5439 md_io_acct = container_of(align_bio, struct md_io_acct, bio_clone);
5440 raid_bio->bi_next = (void *)rdev;
5441 if (blk_queue_io_stat(raid_bio->bi_bdev->bd_disk->queue))
5442 md_io_acct->start_time = bio_start_io_acct(raid_bio);
5443 md_io_acct->orig_bio = raid_bio;
5445 bio_set_dev(align_bio, rdev->bdev);
5446 align_bio->bi_end_io = raid5_align_endio;
5447 align_bio->bi_private = md_io_acct;
5448 align_bio->bi_iter.bi_sector = sector;
5450 /* No reshape active, so we can trust rdev->data_offset */
5451 align_bio->bi_iter.bi_sector += rdev->data_offset;
5454 if (conf->quiesce == 0) {
5455 atomic_inc(&conf->active_aligned_reads);
5458 /* need a memory barrier to detect the race with raid5_quiesce() */
5459 if (!did_inc || smp_load_acquire(&conf->quiesce) != 0) {
5460 /* quiesce is in progress, so we need to undo io activation and wait
5463 if (did_inc && atomic_dec_and_test(&conf->active_aligned_reads))
5464 wake_up(&conf->wait_for_quiescent);
5465 spin_lock_irq(&conf->device_lock);
5466 wait_event_lock_irq(conf->wait_for_quiescent, conf->quiesce == 0,
5468 atomic_inc(&conf->active_aligned_reads);
5469 spin_unlock_irq(&conf->device_lock);
5473 trace_block_bio_remap(align_bio, disk_devt(mddev->gendisk),
5474 raid_bio->bi_iter.bi_sector);
5475 submit_bio_noacct(align_bio);
5483 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5486 sector_t sector = raid_bio->bi_iter.bi_sector;
5487 unsigned chunk_sects = mddev->chunk_sectors;
5488 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5490 if (sectors < bio_sectors(raid_bio)) {
5491 struct r5conf *conf = mddev->private;
5492 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5493 bio_chain(split, raid_bio);
5494 submit_bio_noacct(raid_bio);
5498 if (!raid5_read_one_chunk(mddev, raid_bio))
5504 /* __get_priority_stripe - get the next stripe to process
5506 * Full stripe writes are allowed to pass preread active stripes up until
5507 * the bypass_threshold is exceeded. In general the bypass_count
5508 * increments when the handle_list is handled before the hold_list; however, it
5509 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5510 * stripe with in flight i/o. The bypass_count will be reset when the
5511 * head of the hold_list has changed, i.e. the head was promoted to the
5514 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5516 struct stripe_head *sh, *tmp;
5517 struct list_head *handle_list = NULL;
5518 struct r5worker_group *wg;
5519 bool second_try = !r5c_is_writeback(conf->log) &&
5520 !r5l_log_disk_error(conf);
5521 bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5522 r5l_log_disk_error(conf);
5527 if (conf->worker_cnt_per_group == 0) {
5528 handle_list = try_loprio ? &conf->loprio_list :
5530 } else if (group != ANY_GROUP) {
5531 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5532 &conf->worker_groups[group].handle_list;
5533 wg = &conf->worker_groups[group];
5536 for (i = 0; i < conf->group_cnt; i++) {
5537 handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5538 &conf->worker_groups[i].handle_list;
5539 wg = &conf->worker_groups[i];
5540 if (!list_empty(handle_list))
5545 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5547 list_empty(handle_list) ? "empty" : "busy",
5548 list_empty(&conf->hold_list) ? "empty" : "busy",
5549 atomic_read(&conf->pending_full_writes), conf->bypass_count);
5551 if (!list_empty(handle_list)) {
5552 sh = list_entry(handle_list->next, typeof(*sh), lru);
5554 if (list_empty(&conf->hold_list))
5555 conf->bypass_count = 0;
5556 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5557 if (conf->hold_list.next == conf->last_hold)
5558 conf->bypass_count++;
5560 conf->last_hold = conf->hold_list.next;
5561 conf->bypass_count -= conf->bypass_threshold;
5562 if (conf->bypass_count < 0)
5563 conf->bypass_count = 0;
5566 } else if (!list_empty(&conf->hold_list) &&
5567 ((conf->bypass_threshold &&
5568 conf->bypass_count > conf->bypass_threshold) ||
5569 atomic_read(&conf->pending_full_writes) == 0)) {
5571 list_for_each_entry(tmp, &conf->hold_list, lru) {
5572 if (conf->worker_cnt_per_group == 0 ||
5573 group == ANY_GROUP ||
5574 !cpu_online(tmp->cpu) ||
5575 cpu_to_group(tmp->cpu) == group) {
5582 conf->bypass_count -= conf->bypass_threshold;
5583 if (conf->bypass_count < 0)
5584 conf->bypass_count = 0;
5593 try_loprio = !try_loprio;
5601 list_del_init(&sh->lru);
5602 BUG_ON(atomic_inc_return(&sh->count) != 1);
5606 struct raid5_plug_cb {
5607 struct blk_plug_cb cb;
5608 struct list_head list;
5609 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5612 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5614 struct raid5_plug_cb *cb = container_of(
5615 blk_cb, struct raid5_plug_cb, cb);
5616 struct stripe_head *sh;
5617 struct mddev *mddev = cb->cb.data;
5618 struct r5conf *conf = mddev->private;
5622 if (cb->list.next && !list_empty(&cb->list)) {
5623 spin_lock_irq(&conf->device_lock);
5624 while (!list_empty(&cb->list)) {
5625 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5626 list_del_init(&sh->lru);
5628 * avoid race release_stripe_plug() sees
5629 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5630 * is still in our list
5632 smp_mb__before_atomic();
5633 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5635 * STRIPE_ON_RELEASE_LIST could be set here. In that
5636 * case, the count is always > 1 here
5638 hash = sh->hash_lock_index;
5639 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5642 spin_unlock_irq(&conf->device_lock);
5644 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5645 NR_STRIPE_HASH_LOCKS);
5647 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5651 static void release_stripe_plug(struct mddev *mddev,
5652 struct stripe_head *sh)
5654 struct blk_plug_cb *blk_cb = blk_check_plugged(
5655 raid5_unplug, mddev,
5656 sizeof(struct raid5_plug_cb));
5657 struct raid5_plug_cb *cb;
5660 raid5_release_stripe(sh);
5664 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5666 if (cb->list.next == NULL) {
5668 INIT_LIST_HEAD(&cb->list);
5669 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5670 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5673 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5674 list_add_tail(&sh->lru, &cb->list);
5676 raid5_release_stripe(sh);
5679 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5681 struct r5conf *conf = mddev->private;
5682 sector_t logical_sector, last_sector;
5683 struct stripe_head *sh;
5686 if (mddev->reshape_position != MaxSector)
5687 /* Skip discard while reshape is happening */
5690 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5691 last_sector = bio_end_sector(bi);
5695 stripe_sectors = conf->chunk_sectors *
5696 (conf->raid_disks - conf->max_degraded);
5697 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5699 sector_div(last_sector, stripe_sectors);
5701 logical_sector *= conf->chunk_sectors;
5702 last_sector *= conf->chunk_sectors;
5704 for (; logical_sector < last_sector;
5705 logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5709 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5710 prepare_to_wait(&conf->wait_for_overlap, &w,
5711 TASK_UNINTERRUPTIBLE);
5712 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5713 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5714 raid5_release_stripe(sh);
5718 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5719 spin_lock_irq(&sh->stripe_lock);
5720 for (d = 0; d < conf->raid_disks; d++) {
5721 if (d == sh->pd_idx || d == sh->qd_idx)
5723 if (sh->dev[d].towrite || sh->dev[d].toread) {
5724 set_bit(R5_Overlap, &sh->dev[d].flags);
5725 spin_unlock_irq(&sh->stripe_lock);
5726 raid5_release_stripe(sh);
5731 set_bit(STRIPE_DISCARD, &sh->state);
5732 finish_wait(&conf->wait_for_overlap, &w);
5733 sh->overwrite_disks = 0;
5734 for (d = 0; d < conf->raid_disks; d++) {
5735 if (d == sh->pd_idx || d == sh->qd_idx)
5737 sh->dev[d].towrite = bi;
5738 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5739 bio_inc_remaining(bi);
5740 md_write_inc(mddev, bi);
5741 sh->overwrite_disks++;
5743 spin_unlock_irq(&sh->stripe_lock);
5744 if (conf->mddev->bitmap) {
5746 d < conf->raid_disks - conf->max_degraded;
5748 md_bitmap_startwrite(mddev->bitmap,
5750 RAID5_STRIPE_SECTORS(conf),
5752 sh->bm_seq = conf->seq_flush + 1;
5753 set_bit(STRIPE_BIT_DELAY, &sh->state);
5756 set_bit(STRIPE_HANDLE, &sh->state);
5757 clear_bit(STRIPE_DELAYED, &sh->state);
5758 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5759 atomic_inc(&conf->preread_active_stripes);
5760 release_stripe_plug(mddev, sh);
5766 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5768 struct r5conf *conf = mddev->private;
5770 sector_t new_sector;
5771 sector_t logical_sector, last_sector;
5772 struct stripe_head *sh;
5773 const int rw = bio_data_dir(bi);
5776 bool do_flush = false;
5778 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5779 int ret = log_handle_flush_request(conf, bi);
5783 if (ret == -ENODEV) {
5784 if (md_flush_request(mddev, bi))
5787 /* ret == -EAGAIN, fallback */
5789 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5790 * we need to flush journal device
5792 do_flush = bi->bi_opf & REQ_PREFLUSH;
5795 if (!md_write_start(mddev, bi))
5798 * If array is degraded, better not do chunk aligned read because
5799 * later we might have to read it again in order to reconstruct
5800 * data on failed drives.
5802 if (rw == READ && mddev->degraded == 0 &&
5803 mddev->reshape_position == MaxSector) {
5804 bi = chunk_aligned_read(mddev, bi);
5809 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5810 make_discard_request(mddev, bi);
5811 md_write_end(mddev);
5815 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5816 last_sector = bio_end_sector(bi);
5819 md_account_bio(mddev, &bi);
5820 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5821 for (; logical_sector < last_sector; logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5827 seq = read_seqcount_begin(&conf->gen_lock);
5830 prepare_to_wait(&conf->wait_for_overlap, &w,
5831 TASK_UNINTERRUPTIBLE);
5832 if (unlikely(conf->reshape_progress != MaxSector)) {
5833 /* spinlock is needed as reshape_progress may be
5834 * 64bit on a 32bit platform, and so it might be
5835 * possible to see a half-updated value
5836 * Of course reshape_progress could change after
5837 * the lock is dropped, so once we get a reference
5838 * to the stripe that we think it is, we will have
5841 spin_lock_irq(&conf->device_lock);
5842 if (mddev->reshape_backwards
5843 ? logical_sector < conf->reshape_progress
5844 : logical_sector >= conf->reshape_progress) {
5847 if (mddev->reshape_backwards
5848 ? logical_sector < conf->reshape_safe
5849 : logical_sector >= conf->reshape_safe) {
5850 spin_unlock_irq(&conf->device_lock);
5856 spin_unlock_irq(&conf->device_lock);
5859 new_sector = raid5_compute_sector(conf, logical_sector,
5862 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5863 (unsigned long long)new_sector,
5864 (unsigned long long)logical_sector);
5866 sh = raid5_get_active_stripe(conf, new_sector, previous,
5867 (bi->bi_opf & REQ_RAHEAD), 0);
5869 if (unlikely(previous)) {
5870 /* expansion might have moved on while waiting for a
5871 * stripe, so we must do the range check again.
5872 * Expansion could still move past after this
5873 * test, but as we are holding a reference to
5874 * 'sh', we know that if that happens,
5875 * STRIPE_EXPANDING will get set and the expansion
5876 * won't proceed until we finish with the stripe.
5879 spin_lock_irq(&conf->device_lock);
5880 if (mddev->reshape_backwards
5881 ? logical_sector >= conf->reshape_progress
5882 : logical_sector < conf->reshape_progress)
5883 /* mismatch, need to try again */
5885 spin_unlock_irq(&conf->device_lock);
5887 raid5_release_stripe(sh);
5893 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5894 /* Might have got the wrong stripe_head
5897 raid5_release_stripe(sh);
5901 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5902 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5903 /* Stripe is busy expanding or
5904 * add failed due to overlap. Flush everything
5907 md_wakeup_thread(mddev->thread);
5908 raid5_release_stripe(sh);
5914 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5915 /* we only need flush for one stripe */
5919 set_bit(STRIPE_HANDLE, &sh->state);
5920 clear_bit(STRIPE_DELAYED, &sh->state);
5921 if ((!sh->batch_head || sh == sh->batch_head) &&
5922 (bi->bi_opf & REQ_SYNC) &&
5923 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5924 atomic_inc(&conf->preread_active_stripes);
5925 release_stripe_plug(mddev, sh);
5927 /* cannot get stripe for read-ahead, just give-up */
5928 bi->bi_status = BLK_STS_IOERR;
5932 finish_wait(&conf->wait_for_overlap, &w);
5935 md_write_end(mddev);
5940 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5942 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5944 /* reshaping is quite different to recovery/resync so it is
5945 * handled quite separately ... here.
5947 * On each call to sync_request, we gather one chunk worth of
5948 * destination stripes and flag them as expanding.
5949 * Then we find all the source stripes and request reads.
5950 * As the reads complete, handle_stripe will copy the data
5951 * into the destination stripe and release that stripe.
5953 struct r5conf *conf = mddev->private;
5954 struct stripe_head *sh;
5955 struct md_rdev *rdev;
5956 sector_t first_sector, last_sector;
5957 int raid_disks = conf->previous_raid_disks;
5958 int data_disks = raid_disks - conf->max_degraded;
5959 int new_data_disks = conf->raid_disks - conf->max_degraded;
5962 sector_t writepos, readpos, safepos;
5963 sector_t stripe_addr;
5964 int reshape_sectors;
5965 struct list_head stripes;
5968 if (sector_nr == 0) {
5969 /* If restarting in the middle, skip the initial sectors */
5970 if (mddev->reshape_backwards &&
5971 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5972 sector_nr = raid5_size(mddev, 0, 0)
5973 - conf->reshape_progress;
5974 } else if (mddev->reshape_backwards &&
5975 conf->reshape_progress == MaxSector) {
5976 /* shouldn't happen, but just in case, finish up.*/
5977 sector_nr = MaxSector;
5978 } else if (!mddev->reshape_backwards &&
5979 conf->reshape_progress > 0)
5980 sector_nr = conf->reshape_progress;
5981 sector_div(sector_nr, new_data_disks);
5983 mddev->curr_resync_completed = sector_nr;
5984 sysfs_notify_dirent_safe(mddev->sysfs_completed);
5991 /* We need to process a full chunk at a time.
5992 * If old and new chunk sizes differ, we need to process the
5996 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5998 /* We update the metadata at least every 10 seconds, or when
5999 * the data about to be copied would over-write the source of
6000 * the data at the front of the range. i.e. one new_stripe
6001 * along from reshape_progress new_maps to after where
6002 * reshape_safe old_maps to
6004 writepos = conf->reshape_progress;
6005 sector_div(writepos, new_data_disks);
6006 readpos = conf->reshape_progress;
6007 sector_div(readpos, data_disks);
6008 safepos = conf->reshape_safe;
6009 sector_div(safepos, data_disks);
6010 if (mddev->reshape_backwards) {
6011 BUG_ON(writepos < reshape_sectors);
6012 writepos -= reshape_sectors;
6013 readpos += reshape_sectors;
6014 safepos += reshape_sectors;
6016 writepos += reshape_sectors;
6017 /* readpos and safepos are worst-case calculations.
6018 * A negative number is overly pessimistic, and causes
6019 * obvious problems for unsigned storage. So clip to 0.
6021 readpos -= min_t(sector_t, reshape_sectors, readpos);
6022 safepos -= min_t(sector_t, reshape_sectors, safepos);
6025 /* Having calculated the 'writepos' possibly use it
6026 * to set 'stripe_addr' which is where we will write to.
6028 if (mddev->reshape_backwards) {
6029 BUG_ON(conf->reshape_progress == 0);
6030 stripe_addr = writepos;
6031 BUG_ON((mddev->dev_sectors &
6032 ~((sector_t)reshape_sectors - 1))
6033 - reshape_sectors - stripe_addr
6036 BUG_ON(writepos != sector_nr + reshape_sectors);
6037 stripe_addr = sector_nr;
6040 /* 'writepos' is the most advanced device address we might write.
6041 * 'readpos' is the least advanced device address we might read.
6042 * 'safepos' is the least address recorded in the metadata as having
6044 * If there is a min_offset_diff, these are adjusted either by
6045 * increasing the safepos/readpos if diff is negative, or
6046 * increasing writepos if diff is positive.
6047 * If 'readpos' is then behind 'writepos', there is no way that we can
6048 * ensure safety in the face of a crash - that must be done by userspace
6049 * making a backup of the data. So in that case there is no particular
6050 * rush to update metadata.
6051 * Otherwise if 'safepos' is behind 'writepos', then we really need to
6052 * update the metadata to advance 'safepos' to match 'readpos' so that
6053 * we can be safe in the event of a crash.
6054 * So we insist on updating metadata if safepos is behind writepos and
6055 * readpos is beyond writepos.
6056 * In any case, update the metadata every 10 seconds.
6057 * Maybe that number should be configurable, but I'm not sure it is
6058 * worth it.... maybe it could be a multiple of safemode_delay???
6060 if (conf->min_offset_diff < 0) {
6061 safepos += -conf->min_offset_diff;
6062 readpos += -conf->min_offset_diff;
6064 writepos += conf->min_offset_diff;
6066 if ((mddev->reshape_backwards
6067 ? (safepos > writepos && readpos < writepos)
6068 : (safepos < writepos && readpos > writepos)) ||
6069 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
6070 /* Cannot proceed until we've updated the superblock... */
6071 wait_event(conf->wait_for_overlap,
6072 atomic_read(&conf->reshape_stripes)==0
6073 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6074 if (atomic_read(&conf->reshape_stripes) != 0)
6076 mddev->reshape_position = conf->reshape_progress;
6077 mddev->curr_resync_completed = sector_nr;
6078 if (!mddev->reshape_backwards)
6079 /* Can update recovery_offset */
6080 rdev_for_each(rdev, mddev)
6081 if (rdev->raid_disk >= 0 &&
6082 !test_bit(Journal, &rdev->flags) &&
6083 !test_bit(In_sync, &rdev->flags) &&
6084 rdev->recovery_offset < sector_nr)
6085 rdev->recovery_offset = sector_nr;
6087 conf->reshape_checkpoint = jiffies;
6088 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6089 md_wakeup_thread(mddev->thread);
6090 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
6091 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6092 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6094 spin_lock_irq(&conf->device_lock);
6095 conf->reshape_safe = mddev->reshape_position;
6096 spin_unlock_irq(&conf->device_lock);
6097 wake_up(&conf->wait_for_overlap);
6098 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6101 INIT_LIST_HEAD(&stripes);
6102 for (i = 0; i < reshape_sectors; i += RAID5_STRIPE_SECTORS(conf)) {
6104 int skipped_disk = 0;
6105 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
6106 set_bit(STRIPE_EXPANDING, &sh->state);
6107 atomic_inc(&conf->reshape_stripes);
6108 /* If any of this stripe is beyond the end of the old
6109 * array, then we need to zero those blocks
6111 for (j=sh->disks; j--;) {
6113 if (j == sh->pd_idx)
6115 if (conf->level == 6 &&
6118 s = raid5_compute_blocknr(sh, j, 0);
6119 if (s < raid5_size(mddev, 0, 0)) {
6123 memset(page_address(sh->dev[j].page), 0, RAID5_STRIPE_SIZE(conf));
6124 set_bit(R5_Expanded, &sh->dev[j].flags);
6125 set_bit(R5_UPTODATE, &sh->dev[j].flags);
6127 if (!skipped_disk) {
6128 set_bit(STRIPE_EXPAND_READY, &sh->state);
6129 set_bit(STRIPE_HANDLE, &sh->state);
6131 list_add(&sh->lru, &stripes);
6133 spin_lock_irq(&conf->device_lock);
6134 if (mddev->reshape_backwards)
6135 conf->reshape_progress -= reshape_sectors * new_data_disks;
6137 conf->reshape_progress += reshape_sectors * new_data_disks;
6138 spin_unlock_irq(&conf->device_lock);
6139 /* Ok, those stripe are ready. We can start scheduling
6140 * reads on the source stripes.
6141 * The source stripes are determined by mapping the first and last
6142 * block on the destination stripes.
6145 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
6148 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
6149 * new_data_disks - 1),
6151 if (last_sector >= mddev->dev_sectors)
6152 last_sector = mddev->dev_sectors - 1;
6153 while (first_sector <= last_sector) {
6154 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
6155 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
6156 set_bit(STRIPE_HANDLE, &sh->state);
6157 raid5_release_stripe(sh);
6158 first_sector += RAID5_STRIPE_SECTORS(conf);
6160 /* Now that the sources are clearly marked, we can release
6161 * the destination stripes
6163 while (!list_empty(&stripes)) {
6164 sh = list_entry(stripes.next, struct stripe_head, lru);
6165 list_del_init(&sh->lru);
6166 raid5_release_stripe(sh);
6168 /* If this takes us to the resync_max point where we have to pause,
6169 * then we need to write out the superblock.
6171 sector_nr += reshape_sectors;
6172 retn = reshape_sectors;
6174 if (mddev->curr_resync_completed > mddev->resync_max ||
6175 (sector_nr - mddev->curr_resync_completed) * 2
6176 >= mddev->resync_max - mddev->curr_resync_completed) {
6177 /* Cannot proceed until we've updated the superblock... */
6178 wait_event(conf->wait_for_overlap,
6179 atomic_read(&conf->reshape_stripes) == 0
6180 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6181 if (atomic_read(&conf->reshape_stripes) != 0)
6183 mddev->reshape_position = conf->reshape_progress;
6184 mddev->curr_resync_completed = sector_nr;
6185 if (!mddev->reshape_backwards)
6186 /* Can update recovery_offset */
6187 rdev_for_each(rdev, mddev)
6188 if (rdev->raid_disk >= 0 &&
6189 !test_bit(Journal, &rdev->flags) &&
6190 !test_bit(In_sync, &rdev->flags) &&
6191 rdev->recovery_offset < sector_nr)
6192 rdev->recovery_offset = sector_nr;
6193 conf->reshape_checkpoint = jiffies;
6194 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6195 md_wakeup_thread(mddev->thread);
6196 wait_event(mddev->sb_wait,
6197 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6198 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6199 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6201 spin_lock_irq(&conf->device_lock);
6202 conf->reshape_safe = mddev->reshape_position;
6203 spin_unlock_irq(&conf->device_lock);
6204 wake_up(&conf->wait_for_overlap);
6205 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6211 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6214 struct r5conf *conf = mddev->private;
6215 struct stripe_head *sh;
6216 sector_t max_sector = mddev->dev_sectors;
6217 sector_t sync_blocks;
6218 int still_degraded = 0;
6221 if (sector_nr >= max_sector) {
6222 /* just being told to finish up .. nothing much to do */
6224 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6229 if (mddev->curr_resync < max_sector) /* aborted */
6230 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6232 else /* completed sync */
6234 md_bitmap_close_sync(mddev->bitmap);
6239 /* Allow raid5_quiesce to complete */
6240 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6242 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6243 return reshape_request(mddev, sector_nr, skipped);
6245 /* No need to check resync_max as we never do more than one
6246 * stripe, and as resync_max will always be on a chunk boundary,
6247 * if the check in md_do_sync didn't fire, there is no chance
6248 * of overstepping resync_max here
6251 /* if there is too many failed drives and we are trying
6252 * to resync, then assert that we are finished, because there is
6253 * nothing we can do.
6255 if (mddev->degraded >= conf->max_degraded &&
6256 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6257 sector_t rv = mddev->dev_sectors - sector_nr;
6261 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6263 !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6264 sync_blocks >= RAID5_STRIPE_SECTORS(conf)) {
6265 /* we can skip this block, and probably more */
6266 do_div(sync_blocks, RAID5_STRIPE_SECTORS(conf));
6268 /* keep things rounded to whole stripes */
6269 return sync_blocks * RAID5_STRIPE_SECTORS(conf);
6272 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6274 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6276 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6277 /* make sure we don't swamp the stripe cache if someone else
6278 * is trying to get access
6280 schedule_timeout_uninterruptible(1);
6282 /* Need to check if array will still be degraded after recovery/resync
6283 * Note in case of > 1 drive failures it's possible we're rebuilding
6284 * one drive while leaving another faulty drive in array.
6287 for (i = 0; i < conf->raid_disks; i++) {
6288 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6290 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6295 md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6297 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6298 set_bit(STRIPE_HANDLE, &sh->state);
6300 raid5_release_stripe(sh);
6302 return RAID5_STRIPE_SECTORS(conf);
6305 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6306 unsigned int offset)
6308 /* We may not be able to submit a whole bio at once as there
6309 * may not be enough stripe_heads available.
6310 * We cannot pre-allocate enough stripe_heads as we may need
6311 * more than exist in the cache (if we allow ever large chunks).
6312 * So we do one stripe head at a time and record in
6313 * ->bi_hw_segments how many have been done.
6315 * We *know* that this entire raid_bio is in one chunk, so
6316 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6318 struct stripe_head *sh;
6320 sector_t sector, logical_sector, last_sector;
6324 logical_sector = raid_bio->bi_iter.bi_sector &
6325 ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6326 sector = raid5_compute_sector(conf, logical_sector,
6328 last_sector = bio_end_sector(raid_bio);
6330 for (; logical_sector < last_sector;
6331 logical_sector += RAID5_STRIPE_SECTORS(conf),
6332 sector += RAID5_STRIPE_SECTORS(conf),
6336 /* already done this stripe */
6339 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6342 /* failed to get a stripe - must wait */
6343 conf->retry_read_aligned = raid_bio;
6344 conf->retry_read_offset = scnt;
6348 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6349 raid5_release_stripe(sh);
6350 conf->retry_read_aligned = raid_bio;
6351 conf->retry_read_offset = scnt;
6355 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6357 raid5_release_stripe(sh);
6361 bio_endio(raid_bio);
6363 if (atomic_dec_and_test(&conf->active_aligned_reads))
6364 wake_up(&conf->wait_for_quiescent);
6368 static int handle_active_stripes(struct r5conf *conf, int group,
6369 struct r5worker *worker,
6370 struct list_head *temp_inactive_list)
6371 __releases(&conf->device_lock)
6372 __acquires(&conf->device_lock)
6374 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6375 int i, batch_size = 0, hash;
6376 bool release_inactive = false;
6378 while (batch_size < MAX_STRIPE_BATCH &&
6379 (sh = __get_priority_stripe(conf, group)) != NULL)
6380 batch[batch_size++] = sh;
6382 if (batch_size == 0) {
6383 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6384 if (!list_empty(temp_inactive_list + i))
6386 if (i == NR_STRIPE_HASH_LOCKS) {
6387 spin_unlock_irq(&conf->device_lock);
6388 log_flush_stripe_to_raid(conf);
6389 spin_lock_irq(&conf->device_lock);
6392 release_inactive = true;
6394 spin_unlock_irq(&conf->device_lock);
6396 release_inactive_stripe_list(conf, temp_inactive_list,
6397 NR_STRIPE_HASH_LOCKS);
6399 r5l_flush_stripe_to_raid(conf->log);
6400 if (release_inactive) {
6401 spin_lock_irq(&conf->device_lock);
6405 for (i = 0; i < batch_size; i++)
6406 handle_stripe(batch[i]);
6407 log_write_stripe_run(conf);
6411 spin_lock_irq(&conf->device_lock);
6412 for (i = 0; i < batch_size; i++) {
6413 hash = batch[i]->hash_lock_index;
6414 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6419 static void raid5_do_work(struct work_struct *work)
6421 struct r5worker *worker = container_of(work, struct r5worker, work);
6422 struct r5worker_group *group = worker->group;
6423 struct r5conf *conf = group->conf;
6424 struct mddev *mddev = conf->mddev;
6425 int group_id = group - conf->worker_groups;
6427 struct blk_plug plug;
6429 pr_debug("+++ raid5worker active\n");
6431 blk_start_plug(&plug);
6433 spin_lock_irq(&conf->device_lock);
6435 int batch_size, released;
6437 released = release_stripe_list(conf, worker->temp_inactive_list);
6439 batch_size = handle_active_stripes(conf, group_id, worker,
6440 worker->temp_inactive_list);
6441 worker->working = false;
6442 if (!batch_size && !released)
6444 handled += batch_size;
6445 wait_event_lock_irq(mddev->sb_wait,
6446 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6449 pr_debug("%d stripes handled\n", handled);
6451 spin_unlock_irq(&conf->device_lock);
6453 flush_deferred_bios(conf);
6455 r5l_flush_stripe_to_raid(conf->log);
6457 async_tx_issue_pending_all();
6458 blk_finish_plug(&plug);
6460 pr_debug("--- raid5worker inactive\n");
6464 * This is our raid5 kernel thread.
6466 * We scan the hash table for stripes which can be handled now.
6467 * During the scan, completed stripes are saved for us by the interrupt
6468 * handler, so that they will not have to wait for our next wakeup.
6470 static void raid5d(struct md_thread *thread)
6472 struct mddev *mddev = thread->mddev;
6473 struct r5conf *conf = mddev->private;
6475 struct blk_plug plug;
6477 pr_debug("+++ raid5d active\n");
6479 md_check_recovery(mddev);
6481 blk_start_plug(&plug);
6483 spin_lock_irq(&conf->device_lock);
6486 int batch_size, released;
6487 unsigned int offset;
6489 released = release_stripe_list(conf, conf->temp_inactive_list);
6491 clear_bit(R5_DID_ALLOC, &conf->cache_state);
6494 !list_empty(&conf->bitmap_list)) {
6495 /* Now is a good time to flush some bitmap updates */
6497 spin_unlock_irq(&conf->device_lock);
6498 md_bitmap_unplug(mddev->bitmap);
6499 spin_lock_irq(&conf->device_lock);
6500 conf->seq_write = conf->seq_flush;
6501 activate_bit_delay(conf, conf->temp_inactive_list);
6503 raid5_activate_delayed(conf);
6505 while ((bio = remove_bio_from_retry(conf, &offset))) {
6507 spin_unlock_irq(&conf->device_lock);
6508 ok = retry_aligned_read(conf, bio, offset);
6509 spin_lock_irq(&conf->device_lock);
6515 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6516 conf->temp_inactive_list);
6517 if (!batch_size && !released)
6519 handled += batch_size;
6521 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6522 spin_unlock_irq(&conf->device_lock);
6523 md_check_recovery(mddev);
6524 spin_lock_irq(&conf->device_lock);
6527 pr_debug("%d stripes handled\n", handled);
6529 spin_unlock_irq(&conf->device_lock);
6530 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6531 mutex_trylock(&conf->cache_size_mutex)) {
6532 grow_one_stripe(conf, __GFP_NOWARN);
6533 /* Set flag even if allocation failed. This helps
6534 * slow down allocation requests when mem is short
6536 set_bit(R5_DID_ALLOC, &conf->cache_state);
6537 mutex_unlock(&conf->cache_size_mutex);
6540 flush_deferred_bios(conf);
6542 r5l_flush_stripe_to_raid(conf->log);
6544 async_tx_issue_pending_all();
6545 blk_finish_plug(&plug);
6547 pr_debug("--- raid5d inactive\n");
6551 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6553 struct r5conf *conf;
6555 spin_lock(&mddev->lock);
6556 conf = mddev->private;
6558 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6559 spin_unlock(&mddev->lock);
6564 raid5_set_cache_size(struct mddev *mddev, int size)
6567 struct r5conf *conf = mddev->private;
6569 if (size <= 16 || size > 32768)
6572 conf->min_nr_stripes = size;
6573 mutex_lock(&conf->cache_size_mutex);
6574 while (size < conf->max_nr_stripes &&
6575 drop_one_stripe(conf))
6577 mutex_unlock(&conf->cache_size_mutex);
6579 md_allow_write(mddev);
6581 mutex_lock(&conf->cache_size_mutex);
6582 while (size > conf->max_nr_stripes)
6583 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6584 conf->min_nr_stripes = conf->max_nr_stripes;
6588 mutex_unlock(&conf->cache_size_mutex);
6592 EXPORT_SYMBOL(raid5_set_cache_size);
6595 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6597 struct r5conf *conf;
6601 if (len >= PAGE_SIZE)
6603 if (kstrtoul(page, 10, &new))
6605 err = mddev_lock(mddev);
6608 conf = mddev->private;
6612 err = raid5_set_cache_size(mddev, new);
6613 mddev_unlock(mddev);
6618 static struct md_sysfs_entry
6619 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6620 raid5_show_stripe_cache_size,
6621 raid5_store_stripe_cache_size);
6624 raid5_show_rmw_level(struct mddev *mddev, char *page)
6626 struct r5conf *conf = mddev->private;
6628 return sprintf(page, "%d\n", conf->rmw_level);
6634 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6636 struct r5conf *conf = mddev->private;
6642 if (len >= PAGE_SIZE)
6645 if (kstrtoul(page, 10, &new))
6648 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6651 if (new != PARITY_DISABLE_RMW &&
6652 new != PARITY_ENABLE_RMW &&
6653 new != PARITY_PREFER_RMW)
6656 conf->rmw_level = new;
6660 static struct md_sysfs_entry
6661 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6662 raid5_show_rmw_level,
6663 raid5_store_rmw_level);
6666 raid5_show_stripe_size(struct mddev *mddev, char *page)
6668 struct r5conf *conf;
6671 spin_lock(&mddev->lock);
6672 conf = mddev->private;
6674 ret = sprintf(page, "%lu\n", RAID5_STRIPE_SIZE(conf));
6675 spin_unlock(&mddev->lock);
6679 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
6681 raid5_store_stripe_size(struct mddev *mddev, const char *page, size_t len)
6683 struct r5conf *conf;
6688 if (len >= PAGE_SIZE)
6690 if (kstrtoul(page, 10, &new))
6694 * The value should not be bigger than PAGE_SIZE. It requires to
6695 * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
6698 if (new % DEFAULT_STRIPE_SIZE != 0 ||
6699 new > PAGE_SIZE || new == 0 ||
6700 new != roundup_pow_of_two(new))
6703 err = mddev_lock(mddev);
6707 conf = mddev->private;
6713 if (new == conf->stripe_size)
6716 pr_debug("md/raid: change stripe_size from %lu to %lu\n",
6717 conf->stripe_size, new);
6719 if (mddev->sync_thread ||
6720 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
6721 mddev->reshape_position != MaxSector ||
6722 mddev->sysfs_active) {
6727 mddev_suspend(mddev);
6728 mutex_lock(&conf->cache_size_mutex);
6729 size = conf->max_nr_stripes;
6731 shrink_stripes(conf);
6733 conf->stripe_size = new;
6734 conf->stripe_shift = ilog2(new) - 9;
6735 conf->stripe_sectors = new >> 9;
6736 if (grow_stripes(conf, size)) {
6737 pr_warn("md/raid:%s: couldn't allocate buffers\n",
6741 mutex_unlock(&conf->cache_size_mutex);
6742 mddev_resume(mddev);
6745 mddev_unlock(mddev);
6749 static struct md_sysfs_entry
6750 raid5_stripe_size = __ATTR(stripe_size, 0644,
6751 raid5_show_stripe_size,
6752 raid5_store_stripe_size);
6754 static struct md_sysfs_entry
6755 raid5_stripe_size = __ATTR(stripe_size, 0444,
6756 raid5_show_stripe_size,
6761 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6763 struct r5conf *conf;
6765 spin_lock(&mddev->lock);
6766 conf = mddev->private;
6768 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6769 spin_unlock(&mddev->lock);
6774 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6776 struct r5conf *conf;
6780 if (len >= PAGE_SIZE)
6782 if (kstrtoul(page, 10, &new))
6785 err = mddev_lock(mddev);
6788 conf = mddev->private;
6791 else if (new > conf->min_nr_stripes)
6794 conf->bypass_threshold = new;
6795 mddev_unlock(mddev);
6799 static struct md_sysfs_entry
6800 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6802 raid5_show_preread_threshold,
6803 raid5_store_preread_threshold);
6806 raid5_show_skip_copy(struct mddev *mddev, char *page)
6808 struct r5conf *conf;
6810 spin_lock(&mddev->lock);
6811 conf = mddev->private;
6813 ret = sprintf(page, "%d\n", conf->skip_copy);
6814 spin_unlock(&mddev->lock);
6819 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6821 struct r5conf *conf;
6825 if (len >= PAGE_SIZE)
6827 if (kstrtoul(page, 10, &new))
6831 err = mddev_lock(mddev);
6834 conf = mddev->private;
6837 else if (new != conf->skip_copy) {
6838 struct request_queue *q = mddev->queue;
6840 mddev_suspend(mddev);
6841 conf->skip_copy = new;
6843 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, q);
6845 blk_queue_flag_clear(QUEUE_FLAG_STABLE_WRITES, q);
6846 mddev_resume(mddev);
6848 mddev_unlock(mddev);
6852 static struct md_sysfs_entry
6853 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6854 raid5_show_skip_copy,
6855 raid5_store_skip_copy);
6858 stripe_cache_active_show(struct mddev *mddev, char *page)
6860 struct r5conf *conf = mddev->private;
6862 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6867 static struct md_sysfs_entry
6868 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6871 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6873 struct r5conf *conf;
6875 spin_lock(&mddev->lock);
6876 conf = mddev->private;
6878 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6879 spin_unlock(&mddev->lock);
6883 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6885 struct r5worker_group **worker_groups);
6887 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6889 struct r5conf *conf;
6892 struct r5worker_group *new_groups, *old_groups;
6895 if (len >= PAGE_SIZE)
6897 if (kstrtouint(page, 10, &new))
6899 /* 8192 should be big enough */
6903 err = mddev_lock(mddev);
6906 conf = mddev->private;
6909 else if (new != conf->worker_cnt_per_group) {
6910 mddev_suspend(mddev);
6912 old_groups = conf->worker_groups;
6914 flush_workqueue(raid5_wq);
6916 err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
6918 spin_lock_irq(&conf->device_lock);
6919 conf->group_cnt = group_cnt;
6920 conf->worker_cnt_per_group = new;
6921 conf->worker_groups = new_groups;
6922 spin_unlock_irq(&conf->device_lock);
6925 kfree(old_groups[0].workers);
6928 mddev_resume(mddev);
6930 mddev_unlock(mddev);
6935 static struct md_sysfs_entry
6936 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6937 raid5_show_group_thread_cnt,
6938 raid5_store_group_thread_cnt);
6940 static struct attribute *raid5_attrs[] = {
6941 &raid5_stripecache_size.attr,
6942 &raid5_stripecache_active.attr,
6943 &raid5_preread_bypass_threshold.attr,
6944 &raid5_group_thread_cnt.attr,
6945 &raid5_skip_copy.attr,
6946 &raid5_rmw_level.attr,
6947 &raid5_stripe_size.attr,
6948 &r5c_journal_mode.attr,
6949 &ppl_write_hint.attr,
6952 static const struct attribute_group raid5_attrs_group = {
6954 .attrs = raid5_attrs,
6957 static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
6958 struct r5worker_group **worker_groups)
6962 struct r5worker *workers;
6966 *worker_groups = NULL;
6969 *group_cnt = num_possible_nodes();
6970 size = sizeof(struct r5worker) * cnt;
6971 workers = kcalloc(size, *group_cnt, GFP_NOIO);
6972 *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6974 if (!*worker_groups || !workers) {
6976 kfree(*worker_groups);
6980 for (i = 0; i < *group_cnt; i++) {
6981 struct r5worker_group *group;
6983 group = &(*worker_groups)[i];
6984 INIT_LIST_HEAD(&group->handle_list);
6985 INIT_LIST_HEAD(&group->loprio_list);
6987 group->workers = workers + i * cnt;
6989 for (j = 0; j < cnt; j++) {
6990 struct r5worker *worker = group->workers + j;
6991 worker->group = group;
6992 INIT_WORK(&worker->work, raid5_do_work);
6994 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6995 INIT_LIST_HEAD(worker->temp_inactive_list + k);
7002 static void free_thread_groups(struct r5conf *conf)
7004 if (conf->worker_groups)
7005 kfree(conf->worker_groups[0].workers);
7006 kfree(conf->worker_groups);
7007 conf->worker_groups = NULL;
7011 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
7013 struct r5conf *conf = mddev->private;
7016 sectors = mddev->dev_sectors;
7018 /* size is defined by the smallest of previous and new size */
7019 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
7021 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7022 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
7023 return sectors * (raid_disks - conf->max_degraded);
7026 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7028 safe_put_page(percpu->spare_page);
7029 percpu->spare_page = NULL;
7030 kvfree(percpu->scribble);
7031 percpu->scribble = NULL;
7034 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7036 if (conf->level == 6 && !percpu->spare_page) {
7037 percpu->spare_page = alloc_page(GFP_KERNEL);
7038 if (!percpu->spare_page)
7042 if (scribble_alloc(percpu,
7043 max(conf->raid_disks,
7044 conf->previous_raid_disks),
7045 max(conf->chunk_sectors,
7046 conf->prev_chunk_sectors)
7047 / RAID5_STRIPE_SECTORS(conf))) {
7048 free_scratch_buffer(conf, percpu);
7055 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
7057 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7059 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
7063 static void raid5_free_percpu(struct r5conf *conf)
7068 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7069 free_percpu(conf->percpu);
7072 static void free_conf(struct r5conf *conf)
7078 unregister_shrinker(&conf->shrinker);
7079 free_thread_groups(conf);
7080 shrink_stripes(conf);
7081 raid5_free_percpu(conf);
7082 for (i = 0; i < conf->pool_size; i++)
7083 if (conf->disks[i].extra_page)
7084 put_page(conf->disks[i].extra_page);
7086 bioset_exit(&conf->bio_split);
7087 kfree(conf->stripe_hashtbl);
7088 kfree(conf->pending_data);
7092 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
7094 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7095 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
7097 if (alloc_scratch_buffer(conf, percpu)) {
7098 pr_warn("%s: failed memory allocation for cpu%u\n",
7105 static int raid5_alloc_percpu(struct r5conf *conf)
7109 conf->percpu = alloc_percpu(struct raid5_percpu);
7113 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7115 conf->scribble_disks = max(conf->raid_disks,
7116 conf->previous_raid_disks);
7117 conf->scribble_sectors = max(conf->chunk_sectors,
7118 conf->prev_chunk_sectors);
7123 static unsigned long raid5_cache_scan(struct shrinker *shrink,
7124 struct shrink_control *sc)
7126 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
7127 unsigned long ret = SHRINK_STOP;
7129 if (mutex_trylock(&conf->cache_size_mutex)) {
7131 while (ret < sc->nr_to_scan &&
7132 conf->max_nr_stripes > conf->min_nr_stripes) {
7133 if (drop_one_stripe(conf) == 0) {
7139 mutex_unlock(&conf->cache_size_mutex);
7144 static unsigned long raid5_cache_count(struct shrinker *shrink,
7145 struct shrink_control *sc)
7147 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
7149 if (conf->max_nr_stripes < conf->min_nr_stripes)
7150 /* unlikely, but not impossible */
7152 return conf->max_nr_stripes - conf->min_nr_stripes;
7155 static struct r5conf *setup_conf(struct mddev *mddev)
7157 struct r5conf *conf;
7158 int raid_disk, memory, max_disks;
7159 struct md_rdev *rdev;
7160 struct disk_info *disk;
7164 struct r5worker_group *new_group;
7167 if (mddev->new_level != 5
7168 && mddev->new_level != 4
7169 && mddev->new_level != 6) {
7170 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
7171 mdname(mddev), mddev->new_level);
7172 return ERR_PTR(-EIO);
7174 if ((mddev->new_level == 5
7175 && !algorithm_valid_raid5(mddev->new_layout)) ||
7176 (mddev->new_level == 6
7177 && !algorithm_valid_raid6(mddev->new_layout))) {
7178 pr_warn("md/raid:%s: layout %d not supported\n",
7179 mdname(mddev), mddev->new_layout);
7180 return ERR_PTR(-EIO);
7182 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
7183 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
7184 mdname(mddev), mddev->raid_disks);
7185 return ERR_PTR(-EINVAL);
7188 if (!mddev->new_chunk_sectors ||
7189 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
7190 !is_power_of_2(mddev->new_chunk_sectors)) {
7191 pr_warn("md/raid:%s: invalid chunk size %d\n",
7192 mdname(mddev), mddev->new_chunk_sectors << 9);
7193 return ERR_PTR(-EINVAL);
7196 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
7200 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7201 conf->stripe_size = DEFAULT_STRIPE_SIZE;
7202 conf->stripe_shift = ilog2(DEFAULT_STRIPE_SIZE) - 9;
7203 conf->stripe_sectors = DEFAULT_STRIPE_SIZE >> 9;
7205 INIT_LIST_HEAD(&conf->free_list);
7206 INIT_LIST_HEAD(&conf->pending_list);
7207 conf->pending_data = kcalloc(PENDING_IO_MAX,
7208 sizeof(struct r5pending_data),
7210 if (!conf->pending_data)
7212 for (i = 0; i < PENDING_IO_MAX; i++)
7213 list_add(&conf->pending_data[i].sibling, &conf->free_list);
7214 /* Don't enable multi-threading by default*/
7215 if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
7216 conf->group_cnt = group_cnt;
7217 conf->worker_cnt_per_group = 0;
7218 conf->worker_groups = new_group;
7221 spin_lock_init(&conf->device_lock);
7222 seqcount_spinlock_init(&conf->gen_lock, &conf->device_lock);
7223 mutex_init(&conf->cache_size_mutex);
7224 init_waitqueue_head(&conf->wait_for_quiescent);
7225 init_waitqueue_head(&conf->wait_for_stripe);
7226 init_waitqueue_head(&conf->wait_for_overlap);
7227 INIT_LIST_HEAD(&conf->handle_list);
7228 INIT_LIST_HEAD(&conf->loprio_list);
7229 INIT_LIST_HEAD(&conf->hold_list);
7230 INIT_LIST_HEAD(&conf->delayed_list);
7231 INIT_LIST_HEAD(&conf->bitmap_list);
7232 init_llist_head(&conf->released_stripes);
7233 atomic_set(&conf->active_stripes, 0);
7234 atomic_set(&conf->preread_active_stripes, 0);
7235 atomic_set(&conf->active_aligned_reads, 0);
7236 spin_lock_init(&conf->pending_bios_lock);
7237 conf->batch_bio_dispatch = true;
7238 rdev_for_each(rdev, mddev) {
7239 if (test_bit(Journal, &rdev->flags))
7241 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
7242 conf->batch_bio_dispatch = false;
7247 conf->bypass_threshold = BYPASS_THRESHOLD;
7248 conf->recovery_disabled = mddev->recovery_disabled - 1;
7250 conf->raid_disks = mddev->raid_disks;
7251 if (mddev->reshape_position == MaxSector)
7252 conf->previous_raid_disks = mddev->raid_disks;
7254 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
7255 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
7257 conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
7263 for (i = 0; i < max_disks; i++) {
7264 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
7265 if (!conf->disks[i].extra_page)
7269 ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
7272 conf->mddev = mddev;
7274 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
7277 /* We init hash_locks[0] separately to that it can be used
7278 * as the reference lock in the spin_lock_nest_lock() call
7279 * in lock_all_device_hash_locks_irq in order to convince
7280 * lockdep that we know what we are doing.
7282 spin_lock_init(conf->hash_locks);
7283 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7284 spin_lock_init(conf->hash_locks + i);
7286 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7287 INIT_LIST_HEAD(conf->inactive_list + i);
7289 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7290 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7292 atomic_set(&conf->r5c_cached_full_stripes, 0);
7293 INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7294 atomic_set(&conf->r5c_cached_partial_stripes, 0);
7295 INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7296 atomic_set(&conf->r5c_flushing_full_stripes, 0);
7297 atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7299 conf->level = mddev->new_level;
7300 conf->chunk_sectors = mddev->new_chunk_sectors;
7301 if (raid5_alloc_percpu(conf) != 0)
7304 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7306 rdev_for_each(rdev, mddev) {
7307 raid_disk = rdev->raid_disk;
7308 if (raid_disk >= max_disks
7309 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7311 disk = conf->disks + raid_disk;
7313 if (test_bit(Replacement, &rdev->flags)) {
7314 if (disk->replacement)
7316 disk->replacement = rdev;
7323 if (test_bit(In_sync, &rdev->flags)) {
7324 char b[BDEVNAME_SIZE];
7325 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7326 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7327 } else if (rdev->saved_raid_disk != raid_disk)
7328 /* Cannot rely on bitmap to complete recovery */
7332 conf->level = mddev->new_level;
7333 if (conf->level == 6) {
7334 conf->max_degraded = 2;
7335 if (raid6_call.xor_syndrome)
7336 conf->rmw_level = PARITY_ENABLE_RMW;
7338 conf->rmw_level = PARITY_DISABLE_RMW;
7340 conf->max_degraded = 1;
7341 conf->rmw_level = PARITY_ENABLE_RMW;
7343 conf->algorithm = mddev->new_layout;
7344 conf->reshape_progress = mddev->reshape_position;
7345 if (conf->reshape_progress != MaxSector) {
7346 conf->prev_chunk_sectors = mddev->chunk_sectors;
7347 conf->prev_algo = mddev->layout;
7349 conf->prev_chunk_sectors = conf->chunk_sectors;
7350 conf->prev_algo = conf->algorithm;
7353 conf->min_nr_stripes = NR_STRIPES;
7354 if (mddev->reshape_position != MaxSector) {
7355 int stripes = max_t(int,
7356 ((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4,
7357 ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4);
7358 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7359 if (conf->min_nr_stripes != NR_STRIPES)
7360 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7361 mdname(mddev), conf->min_nr_stripes);
7363 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7364 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7365 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7366 if (grow_stripes(conf, conf->min_nr_stripes)) {
7367 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7368 mdname(mddev), memory);
7371 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7373 * Losing a stripe head costs more than the time to refill it,
7374 * it reduces the queue depth and so can hurt throughput.
7375 * So set it rather large, scaled by number of devices.
7377 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7378 conf->shrinker.scan_objects = raid5_cache_scan;
7379 conf->shrinker.count_objects = raid5_cache_count;
7380 conf->shrinker.batch = 128;
7381 conf->shrinker.flags = 0;
7382 if (register_shrinker(&conf->shrinker)) {
7383 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7388 sprintf(pers_name, "raid%d", mddev->new_level);
7389 conf->thread = md_register_thread(raid5d, mddev, pers_name);
7390 if (!conf->thread) {
7391 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7401 return ERR_PTR(-EIO);
7403 return ERR_PTR(-ENOMEM);
7406 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7409 case ALGORITHM_PARITY_0:
7410 if (raid_disk < max_degraded)
7413 case ALGORITHM_PARITY_N:
7414 if (raid_disk >= raid_disks - max_degraded)
7417 case ALGORITHM_PARITY_0_6:
7418 if (raid_disk == 0 ||
7419 raid_disk == raid_disks - 1)
7422 case ALGORITHM_LEFT_ASYMMETRIC_6:
7423 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7424 case ALGORITHM_LEFT_SYMMETRIC_6:
7425 case ALGORITHM_RIGHT_SYMMETRIC_6:
7426 if (raid_disk == raid_disks - 1)
7432 static void raid5_set_io_opt(struct r5conf *conf)
7434 blk_queue_io_opt(conf->mddev->queue, (conf->chunk_sectors << 9) *
7435 (conf->raid_disks - conf->max_degraded));
7438 static int raid5_run(struct mddev *mddev)
7440 struct r5conf *conf;
7441 int working_disks = 0;
7442 int dirty_parity_disks = 0;
7443 struct md_rdev *rdev;
7444 struct md_rdev *journal_dev = NULL;
7445 sector_t reshape_offset = 0;
7447 long long min_offset_diff = 0;
7450 if (acct_bioset_init(mddev)) {
7451 pr_err("md/raid456:%s: alloc acct bioset failed.\n", mdname(mddev));
7455 if (mddev_init_writes_pending(mddev) < 0) {
7460 if (mddev->recovery_cp != MaxSector)
7461 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7464 rdev_for_each(rdev, mddev) {
7467 if (test_bit(Journal, &rdev->flags)) {
7471 if (rdev->raid_disk < 0)
7473 diff = (rdev->new_data_offset - rdev->data_offset);
7475 min_offset_diff = diff;
7477 } else if (mddev->reshape_backwards &&
7478 diff < min_offset_diff)
7479 min_offset_diff = diff;
7480 else if (!mddev->reshape_backwards &&
7481 diff > min_offset_diff)
7482 min_offset_diff = diff;
7485 if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7486 (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7487 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7493 if (mddev->reshape_position != MaxSector) {
7494 /* Check that we can continue the reshape.
7495 * Difficulties arise if the stripe we would write to
7496 * next is at or after the stripe we would read from next.
7497 * For a reshape that changes the number of devices, this
7498 * is only possible for a very short time, and mdadm makes
7499 * sure that time appears to have past before assembling
7500 * the array. So we fail if that time hasn't passed.
7501 * For a reshape that keeps the number of devices the same
7502 * mdadm must be monitoring the reshape can keeping the
7503 * critical areas read-only and backed up. It will start
7504 * the array in read-only mode, so we check for that.
7506 sector_t here_new, here_old;
7508 int max_degraded = (mddev->level == 6 ? 2 : 1);
7513 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7519 if (mddev->new_level != mddev->level) {
7520 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7525 old_disks = mddev->raid_disks - mddev->delta_disks;
7526 /* reshape_position must be on a new-stripe boundary, and one
7527 * further up in new geometry must map after here in old
7529 * If the chunk sizes are different, then as we perform reshape
7530 * in units of the largest of the two, reshape_position needs
7531 * be a multiple of the largest chunk size times new data disks.
7533 here_new = mddev->reshape_position;
7534 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7535 new_data_disks = mddev->raid_disks - max_degraded;
7536 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7537 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7542 reshape_offset = here_new * chunk_sectors;
7543 /* here_new is the stripe we will write to */
7544 here_old = mddev->reshape_position;
7545 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7546 /* here_old is the first stripe that we might need to read
7548 if (mddev->delta_disks == 0) {
7549 /* We cannot be sure it is safe to start an in-place
7550 * reshape. It is only safe if user-space is monitoring
7551 * and taking constant backups.
7552 * mdadm always starts a situation like this in
7553 * readonly mode so it can take control before
7554 * allowing any writes. So just check for that.
7556 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7557 abs(min_offset_diff) >= mddev->new_chunk_sectors)
7558 /* not really in-place - so OK */;
7559 else if (mddev->ro == 0) {
7560 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7565 } else if (mddev->reshape_backwards
7566 ? (here_new * chunk_sectors + min_offset_diff <=
7567 here_old * chunk_sectors)
7568 : (here_new * chunk_sectors >=
7569 here_old * chunk_sectors + (-min_offset_diff))) {
7570 /* Reading from the same stripe as writing to - bad */
7571 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7576 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7577 /* OK, we should be able to continue; */
7579 BUG_ON(mddev->level != mddev->new_level);
7580 BUG_ON(mddev->layout != mddev->new_layout);
7581 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7582 BUG_ON(mddev->delta_disks != 0);
7585 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7586 test_bit(MD_HAS_PPL, &mddev->flags)) {
7587 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7589 clear_bit(MD_HAS_PPL, &mddev->flags);
7590 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7593 if (mddev->private == NULL)
7594 conf = setup_conf(mddev);
7596 conf = mddev->private;
7599 ret = PTR_ERR(conf);
7603 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7605 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7608 set_disk_ro(mddev->gendisk, 1);
7609 } else if (mddev->recovery_cp == MaxSector)
7610 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7613 conf->min_offset_diff = min_offset_diff;
7614 mddev->thread = conf->thread;
7615 conf->thread = NULL;
7616 mddev->private = conf;
7618 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7620 rdev = conf->disks[i].rdev;
7621 if (!rdev && conf->disks[i].replacement) {
7622 /* The replacement is all we have yet */
7623 rdev = conf->disks[i].replacement;
7624 conf->disks[i].replacement = NULL;
7625 clear_bit(Replacement, &rdev->flags);
7626 conf->disks[i].rdev = rdev;
7630 if (conf->disks[i].replacement &&
7631 conf->reshape_progress != MaxSector) {
7632 /* replacements and reshape simply do not mix. */
7633 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7636 if (test_bit(In_sync, &rdev->flags)) {
7640 /* This disc is not fully in-sync. However if it
7641 * just stored parity (beyond the recovery_offset),
7642 * when we don't need to be concerned about the
7643 * array being dirty.
7644 * When reshape goes 'backwards', we never have
7645 * partially completed devices, so we only need
7646 * to worry about reshape going forwards.
7648 /* Hack because v0.91 doesn't store recovery_offset properly. */
7649 if (mddev->major_version == 0 &&
7650 mddev->minor_version > 90)
7651 rdev->recovery_offset = reshape_offset;
7653 if (rdev->recovery_offset < reshape_offset) {
7654 /* We need to check old and new layout */
7655 if (!only_parity(rdev->raid_disk,
7658 conf->max_degraded))
7661 if (!only_parity(rdev->raid_disk,
7663 conf->previous_raid_disks,
7664 conf->max_degraded))
7666 dirty_parity_disks++;
7670 * 0 for a fully functional array, 1 or 2 for a degraded array.
7672 mddev->degraded = raid5_calc_degraded(conf);
7674 if (has_failed(conf)) {
7675 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7676 mdname(mddev), mddev->degraded, conf->raid_disks);
7680 /* device size must be a multiple of chunk size */
7681 mddev->dev_sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7682 mddev->resync_max_sectors = mddev->dev_sectors;
7684 if (mddev->degraded > dirty_parity_disks &&
7685 mddev->recovery_cp != MaxSector) {
7686 if (test_bit(MD_HAS_PPL, &mddev->flags))
7687 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7689 else if (mddev->ok_start_degraded)
7690 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7693 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7699 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7700 mdname(mddev), conf->level,
7701 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7704 print_raid5_conf(conf);
7706 if (conf->reshape_progress != MaxSector) {
7707 conf->reshape_safe = conf->reshape_progress;
7708 atomic_set(&conf->reshape_stripes, 0);
7709 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7710 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7711 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7712 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7713 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7715 if (!mddev->sync_thread)
7719 /* Ok, everything is just fine now */
7720 if (mddev->to_remove == &raid5_attrs_group)
7721 mddev->to_remove = NULL;
7722 else if (mddev->kobj.sd &&
7723 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7724 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7726 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7730 /* read-ahead size must cover two whole stripes, which
7731 * is 2 * (datadisks) * chunksize where 'n' is the
7732 * number of raid devices
7734 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7735 int stripe = data_disks *
7736 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7738 chunk_size = mddev->chunk_sectors << 9;
7739 blk_queue_io_min(mddev->queue, chunk_size);
7740 raid5_set_io_opt(conf);
7741 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7743 * We can only discard a whole stripe. It doesn't make sense to
7744 * discard data disk but write parity disk
7746 stripe = stripe * PAGE_SIZE;
7747 /* Round up to power of 2, as discard handling
7748 * currently assumes that */
7749 while ((stripe-1) & stripe)
7750 stripe = (stripe | (stripe-1)) + 1;
7751 mddev->queue->limits.discard_alignment = stripe;
7752 mddev->queue->limits.discard_granularity = stripe;
7754 blk_queue_max_write_same_sectors(mddev->queue, 0);
7755 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7757 rdev_for_each(rdev, mddev) {
7758 disk_stack_limits(mddev->gendisk, rdev->bdev,
7759 rdev->data_offset << 9);
7760 disk_stack_limits(mddev->gendisk, rdev->bdev,
7761 rdev->new_data_offset << 9);
7765 * zeroing is required, otherwise data
7766 * could be lost. Consider a scenario: discard a stripe
7767 * (the stripe could be inconsistent if
7768 * discard_zeroes_data is 0); write one disk of the
7769 * stripe (the stripe could be inconsistent again
7770 * depending on which disks are used to calculate
7771 * parity); the disk is broken; The stripe data of this
7774 * We only allow DISCARD if the sysadmin has confirmed that
7775 * only safe devices are in use by setting a module parameter.
7776 * A better idea might be to turn DISCARD into WRITE_ZEROES
7777 * requests, as that is required to be safe.
7779 if (devices_handle_discard_safely &&
7780 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7781 mddev->queue->limits.discard_granularity >= stripe)
7782 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7785 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7788 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7791 if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7796 md_unregister_thread(&mddev->thread);
7797 print_raid5_conf(conf);
7799 mddev->private = NULL;
7800 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7803 acct_bioset_exit(mddev);
7807 static void raid5_free(struct mddev *mddev, void *priv)
7809 struct r5conf *conf = priv;
7812 acct_bioset_exit(mddev);
7813 mddev->to_remove = &raid5_attrs_group;
7816 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7818 struct r5conf *conf = mddev->private;
7821 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7822 conf->chunk_sectors / 2, mddev->layout);
7823 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7825 for (i = 0; i < conf->raid_disks; i++) {
7826 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7827 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7830 seq_printf (seq, "]");
7833 static void print_raid5_conf (struct r5conf *conf)
7836 struct disk_info *tmp;
7838 pr_debug("RAID conf printout:\n");
7840 pr_debug("(conf==NULL)\n");
7843 pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7845 conf->raid_disks - conf->mddev->degraded);
7847 for (i = 0; i < conf->raid_disks; i++) {
7848 char b[BDEVNAME_SIZE];
7849 tmp = conf->disks + i;
7851 pr_debug(" disk %d, o:%d, dev:%s\n",
7852 i, !test_bit(Faulty, &tmp->rdev->flags),
7853 bdevname(tmp->rdev->bdev, b));
7857 static int raid5_spare_active(struct mddev *mddev)
7860 struct r5conf *conf = mddev->private;
7861 struct disk_info *tmp;
7863 unsigned long flags;
7865 for (i = 0; i < conf->raid_disks; i++) {
7866 tmp = conf->disks + i;
7867 if (tmp->replacement
7868 && tmp->replacement->recovery_offset == MaxSector
7869 && !test_bit(Faulty, &tmp->replacement->flags)
7870 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7871 /* Replacement has just become active. */
7873 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7876 /* Replaced device not technically faulty,
7877 * but we need to be sure it gets removed
7878 * and never re-added.
7880 set_bit(Faulty, &tmp->rdev->flags);
7881 sysfs_notify_dirent_safe(
7882 tmp->rdev->sysfs_state);
7884 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7885 } else if (tmp->rdev
7886 && tmp->rdev->recovery_offset == MaxSector
7887 && !test_bit(Faulty, &tmp->rdev->flags)
7888 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7890 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7893 spin_lock_irqsave(&conf->device_lock, flags);
7894 mddev->degraded = raid5_calc_degraded(conf);
7895 spin_unlock_irqrestore(&conf->device_lock, flags);
7896 print_raid5_conf(conf);
7900 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7902 struct r5conf *conf = mddev->private;
7904 int number = rdev->raid_disk;
7905 struct md_rdev **rdevp;
7906 struct disk_info *p = conf->disks + number;
7908 print_raid5_conf(conf);
7909 if (test_bit(Journal, &rdev->flags) && conf->log) {
7911 * we can't wait pending write here, as this is called in
7912 * raid5d, wait will deadlock.
7913 * neilb: there is no locking about new writes here,
7914 * so this cannot be safe.
7916 if (atomic_read(&conf->active_stripes) ||
7917 atomic_read(&conf->r5c_cached_full_stripes) ||
7918 atomic_read(&conf->r5c_cached_partial_stripes)) {
7924 if (rdev == p->rdev)
7926 else if (rdev == p->replacement)
7927 rdevp = &p->replacement;
7931 if (number >= conf->raid_disks &&
7932 conf->reshape_progress == MaxSector)
7933 clear_bit(In_sync, &rdev->flags);
7935 if (test_bit(In_sync, &rdev->flags) ||
7936 atomic_read(&rdev->nr_pending)) {
7940 /* Only remove non-faulty devices if recovery
7943 if (!test_bit(Faulty, &rdev->flags) &&
7944 mddev->recovery_disabled != conf->recovery_disabled &&
7945 !has_failed(conf) &&
7946 (!p->replacement || p->replacement == rdev) &&
7947 number < conf->raid_disks) {
7952 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7954 if (atomic_read(&rdev->nr_pending)) {
7955 /* lost the race, try later */
7961 err = log_modify(conf, rdev, false);
7965 if (p->replacement) {
7966 /* We must have just cleared 'rdev' */
7967 p->rdev = p->replacement;
7968 clear_bit(Replacement, &p->replacement->flags);
7969 smp_mb(); /* Make sure other CPUs may see both as identical
7970 * but will never see neither - if they are careful
7972 p->replacement = NULL;
7975 err = log_modify(conf, p->rdev, true);
7978 clear_bit(WantReplacement, &rdev->flags);
7981 print_raid5_conf(conf);
7985 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7987 struct r5conf *conf = mddev->private;
7988 int ret, err = -EEXIST;
7990 struct disk_info *p;
7992 int last = conf->raid_disks - 1;
7994 if (test_bit(Journal, &rdev->flags)) {
7998 rdev->raid_disk = 0;
8000 * The array is in readonly mode if journal is missing, so no
8001 * write requests running. We should be safe
8003 ret = log_init(conf, rdev, false);
8007 ret = r5l_start(conf->log);
8013 if (mddev->recovery_disabled == conf->recovery_disabled)
8016 if (rdev->saved_raid_disk < 0 && has_failed(conf))
8017 /* no point adding a device */
8020 if (rdev->raid_disk >= 0)
8021 first = last = rdev->raid_disk;
8024 * find the disk ... but prefer rdev->saved_raid_disk
8027 if (rdev->saved_raid_disk >= 0 &&
8028 rdev->saved_raid_disk >= first &&
8029 conf->disks[rdev->saved_raid_disk].rdev == NULL)
8030 first = rdev->saved_raid_disk;
8032 for (disk = first; disk <= last; disk++) {
8033 p = conf->disks + disk;
8034 if (p->rdev == NULL) {
8035 clear_bit(In_sync, &rdev->flags);
8036 rdev->raid_disk = disk;
8037 if (rdev->saved_raid_disk != disk)
8039 rcu_assign_pointer(p->rdev, rdev);
8041 err = log_modify(conf, rdev, true);
8046 for (disk = first; disk <= last; disk++) {
8047 p = conf->disks + disk;
8048 if (test_bit(WantReplacement, &p->rdev->flags) &&
8049 p->replacement == NULL) {
8050 clear_bit(In_sync, &rdev->flags);
8051 set_bit(Replacement, &rdev->flags);
8052 rdev->raid_disk = disk;
8055 rcu_assign_pointer(p->replacement, rdev);
8060 print_raid5_conf(conf);
8064 static int raid5_resize(struct mddev *mddev, sector_t sectors)
8066 /* no resync is happening, and there is enough space
8067 * on all devices, so we can resize.
8068 * We need to make sure resync covers any new space.
8069 * If the array is shrinking we should possibly wait until
8070 * any io in the removed space completes, but it hardly seems
8074 struct r5conf *conf = mddev->private;
8076 if (raid5_has_log(conf) || raid5_has_ppl(conf))
8078 sectors &= ~((sector_t)conf->chunk_sectors - 1);
8079 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
8080 if (mddev->external_size &&
8081 mddev->array_sectors > newsize)
8083 if (mddev->bitmap) {
8084 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
8088 md_set_array_sectors(mddev, newsize);
8089 if (sectors > mddev->dev_sectors &&
8090 mddev->recovery_cp > mddev->dev_sectors) {
8091 mddev->recovery_cp = mddev->dev_sectors;
8092 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8094 mddev->dev_sectors = sectors;
8095 mddev->resync_max_sectors = sectors;
8099 static int check_stripe_cache(struct mddev *mddev)
8101 /* Can only proceed if there are plenty of stripe_heads.
8102 * We need a minimum of one full stripe,, and for sensible progress
8103 * it is best to have about 4 times that.
8104 * If we require 4 times, then the default 256 4K stripe_heads will
8105 * allow for chunk sizes up to 256K, which is probably OK.
8106 * If the chunk size is greater, user-space should request more
8107 * stripe_heads first.
8109 struct r5conf *conf = mddev->private;
8110 if (((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8111 > conf->min_nr_stripes ||
8112 ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8113 > conf->min_nr_stripes) {
8114 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
8116 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
8117 / RAID5_STRIPE_SIZE(conf))*4);
8123 static int check_reshape(struct mddev *mddev)
8125 struct r5conf *conf = mddev->private;
8127 if (raid5_has_log(conf) || raid5_has_ppl(conf))
8129 if (mddev->delta_disks == 0 &&
8130 mddev->new_layout == mddev->layout &&
8131 mddev->new_chunk_sectors == mddev->chunk_sectors)
8132 return 0; /* nothing to do */
8133 if (has_failed(conf))
8135 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
8136 /* We might be able to shrink, but the devices must
8137 * be made bigger first.
8138 * For raid6, 4 is the minimum size.
8139 * Otherwise 2 is the minimum
8142 if (mddev->level == 6)
8144 if (mddev->raid_disks + mddev->delta_disks < min)
8148 if (!check_stripe_cache(mddev))
8151 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
8152 mddev->delta_disks > 0)
8153 if (resize_chunks(conf,
8154 conf->previous_raid_disks
8155 + max(0, mddev->delta_disks),
8156 max(mddev->new_chunk_sectors,
8157 mddev->chunk_sectors)
8161 if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
8162 return 0; /* never bother to shrink */
8163 return resize_stripes(conf, (conf->previous_raid_disks
8164 + mddev->delta_disks));
8167 static int raid5_start_reshape(struct mddev *mddev)
8169 struct r5conf *conf = mddev->private;
8170 struct md_rdev *rdev;
8172 unsigned long flags;
8174 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
8177 if (!check_stripe_cache(mddev))
8180 if (has_failed(conf))
8183 rdev_for_each(rdev, mddev) {
8184 if (!test_bit(In_sync, &rdev->flags)
8185 && !test_bit(Faulty, &rdev->flags))
8189 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
8190 /* Not enough devices even to make a degraded array
8195 /* Refuse to reduce size of the array. Any reductions in
8196 * array size must be through explicit setting of array_size
8199 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
8200 < mddev->array_sectors) {
8201 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
8206 atomic_set(&conf->reshape_stripes, 0);
8207 spin_lock_irq(&conf->device_lock);
8208 write_seqcount_begin(&conf->gen_lock);
8209 conf->previous_raid_disks = conf->raid_disks;
8210 conf->raid_disks += mddev->delta_disks;
8211 conf->prev_chunk_sectors = conf->chunk_sectors;
8212 conf->chunk_sectors = mddev->new_chunk_sectors;
8213 conf->prev_algo = conf->algorithm;
8214 conf->algorithm = mddev->new_layout;
8216 /* Code that selects data_offset needs to see the generation update
8217 * if reshape_progress has been set - so a memory barrier needed.
8220 if (mddev->reshape_backwards)
8221 conf->reshape_progress = raid5_size(mddev, 0, 0);
8223 conf->reshape_progress = 0;
8224 conf->reshape_safe = conf->reshape_progress;
8225 write_seqcount_end(&conf->gen_lock);
8226 spin_unlock_irq(&conf->device_lock);
8228 /* Now make sure any requests that proceeded on the assumption
8229 * the reshape wasn't running - like Discard or Read - have
8232 mddev_suspend(mddev);
8233 mddev_resume(mddev);
8235 /* Add some new drives, as many as will fit.
8236 * We know there are enough to make the newly sized array work.
8237 * Don't add devices if we are reducing the number of
8238 * devices in the array. This is because it is not possible
8239 * to correctly record the "partially reconstructed" state of
8240 * such devices during the reshape and confusion could result.
8242 if (mddev->delta_disks >= 0) {
8243 rdev_for_each(rdev, mddev)
8244 if (rdev->raid_disk < 0 &&
8245 !test_bit(Faulty, &rdev->flags)) {
8246 if (raid5_add_disk(mddev, rdev) == 0) {
8248 >= conf->previous_raid_disks)
8249 set_bit(In_sync, &rdev->flags);
8251 rdev->recovery_offset = 0;
8253 /* Failure here is OK */
8254 sysfs_link_rdev(mddev, rdev);
8256 } else if (rdev->raid_disk >= conf->previous_raid_disks
8257 && !test_bit(Faulty, &rdev->flags)) {
8258 /* This is a spare that was manually added */
8259 set_bit(In_sync, &rdev->flags);
8262 /* When a reshape changes the number of devices,
8263 * ->degraded is measured against the larger of the
8264 * pre and post number of devices.
8266 spin_lock_irqsave(&conf->device_lock, flags);
8267 mddev->degraded = raid5_calc_degraded(conf);
8268 spin_unlock_irqrestore(&conf->device_lock, flags);
8270 mddev->raid_disks = conf->raid_disks;
8271 mddev->reshape_position = conf->reshape_progress;
8272 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8274 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8275 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8276 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
8277 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8278 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
8279 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
8281 if (!mddev->sync_thread) {
8282 mddev->recovery = 0;
8283 spin_lock_irq(&conf->device_lock);
8284 write_seqcount_begin(&conf->gen_lock);
8285 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
8286 mddev->new_chunk_sectors =
8287 conf->chunk_sectors = conf->prev_chunk_sectors;
8288 mddev->new_layout = conf->algorithm = conf->prev_algo;
8289 rdev_for_each(rdev, mddev)
8290 rdev->new_data_offset = rdev->data_offset;
8292 conf->generation --;
8293 conf->reshape_progress = MaxSector;
8294 mddev->reshape_position = MaxSector;
8295 write_seqcount_end(&conf->gen_lock);
8296 spin_unlock_irq(&conf->device_lock);
8299 conf->reshape_checkpoint = jiffies;
8300 md_wakeup_thread(mddev->sync_thread);
8301 md_new_event(mddev);
8305 /* This is called from the reshape thread and should make any
8306 * changes needed in 'conf'
8308 static void end_reshape(struct r5conf *conf)
8311 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8312 struct md_rdev *rdev;
8314 spin_lock_irq(&conf->device_lock);
8315 conf->previous_raid_disks = conf->raid_disks;
8316 md_finish_reshape(conf->mddev);
8318 conf->reshape_progress = MaxSector;
8319 conf->mddev->reshape_position = MaxSector;
8320 rdev_for_each(rdev, conf->mddev)
8321 if (rdev->raid_disk >= 0 &&
8322 !test_bit(Journal, &rdev->flags) &&
8323 !test_bit(In_sync, &rdev->flags))
8324 rdev->recovery_offset = MaxSector;
8325 spin_unlock_irq(&conf->device_lock);
8326 wake_up(&conf->wait_for_overlap);
8328 if (conf->mddev->queue)
8329 raid5_set_io_opt(conf);
8333 /* This is called from the raid5d thread with mddev_lock held.
8334 * It makes config changes to the device.
8336 static void raid5_finish_reshape(struct mddev *mddev)
8338 struct r5conf *conf = mddev->private;
8340 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8342 if (mddev->delta_disks <= 0) {
8344 spin_lock_irq(&conf->device_lock);
8345 mddev->degraded = raid5_calc_degraded(conf);
8346 spin_unlock_irq(&conf->device_lock);
8347 for (d = conf->raid_disks ;
8348 d < conf->raid_disks - mddev->delta_disks;
8350 struct md_rdev *rdev = conf->disks[d].rdev;
8352 clear_bit(In_sync, &rdev->flags);
8353 rdev = conf->disks[d].replacement;
8355 clear_bit(In_sync, &rdev->flags);
8358 mddev->layout = conf->algorithm;
8359 mddev->chunk_sectors = conf->chunk_sectors;
8360 mddev->reshape_position = MaxSector;
8361 mddev->delta_disks = 0;
8362 mddev->reshape_backwards = 0;
8366 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8368 struct r5conf *conf = mddev->private;
8371 /* stop all writes */
8372 lock_all_device_hash_locks_irq(conf);
8373 /* '2' tells resync/reshape to pause so that all
8374 * active stripes can drain
8376 r5c_flush_cache(conf, INT_MAX);
8377 /* need a memory barrier to make sure read_one_chunk() sees
8378 * quiesce started and reverts to slow (locked) path.
8380 smp_store_release(&conf->quiesce, 2);
8381 wait_event_cmd(conf->wait_for_quiescent,
8382 atomic_read(&conf->active_stripes) == 0 &&
8383 atomic_read(&conf->active_aligned_reads) == 0,
8384 unlock_all_device_hash_locks_irq(conf),
8385 lock_all_device_hash_locks_irq(conf));
8387 unlock_all_device_hash_locks_irq(conf);
8388 /* allow reshape to continue */
8389 wake_up(&conf->wait_for_overlap);
8391 /* re-enable writes */
8392 lock_all_device_hash_locks_irq(conf);
8394 wake_up(&conf->wait_for_quiescent);
8395 wake_up(&conf->wait_for_overlap);
8396 unlock_all_device_hash_locks_irq(conf);
8398 log_quiesce(conf, quiesce);
8401 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8403 struct r0conf *raid0_conf = mddev->private;
8406 /* for raid0 takeover only one zone is supported */
8407 if (raid0_conf->nr_strip_zones > 1) {
8408 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8410 return ERR_PTR(-EINVAL);
8413 sectors = raid0_conf->strip_zone[0].zone_end;
8414 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8415 mddev->dev_sectors = sectors;
8416 mddev->new_level = level;
8417 mddev->new_layout = ALGORITHM_PARITY_N;
8418 mddev->new_chunk_sectors = mddev->chunk_sectors;
8419 mddev->raid_disks += 1;
8420 mddev->delta_disks = 1;
8421 /* make sure it will be not marked as dirty */
8422 mddev->recovery_cp = MaxSector;
8424 return setup_conf(mddev);
8427 static void *raid5_takeover_raid1(struct mddev *mddev)
8432 if (mddev->raid_disks != 2 ||
8433 mddev->degraded > 1)
8434 return ERR_PTR(-EINVAL);
8436 /* Should check if there are write-behind devices? */
8438 chunksect = 64*2; /* 64K by default */
8440 /* The array must be an exact multiple of chunksize */
8441 while (chunksect && (mddev->array_sectors & (chunksect-1)))
8444 if ((chunksect<<9) < RAID5_STRIPE_SIZE((struct r5conf *)mddev->private))
8445 /* array size does not allow a suitable chunk size */
8446 return ERR_PTR(-EINVAL);
8448 mddev->new_level = 5;
8449 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8450 mddev->new_chunk_sectors = chunksect;
8452 ret = setup_conf(mddev);
8454 mddev_clear_unsupported_flags(mddev,
8455 UNSUPPORTED_MDDEV_FLAGS);
8459 static void *raid5_takeover_raid6(struct mddev *mddev)
8463 switch (mddev->layout) {
8464 case ALGORITHM_LEFT_ASYMMETRIC_6:
8465 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8467 case ALGORITHM_RIGHT_ASYMMETRIC_6:
8468 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8470 case ALGORITHM_LEFT_SYMMETRIC_6:
8471 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8473 case ALGORITHM_RIGHT_SYMMETRIC_6:
8474 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8476 case ALGORITHM_PARITY_0_6:
8477 new_layout = ALGORITHM_PARITY_0;
8479 case ALGORITHM_PARITY_N:
8480 new_layout = ALGORITHM_PARITY_N;
8483 return ERR_PTR(-EINVAL);
8485 mddev->new_level = 5;
8486 mddev->new_layout = new_layout;
8487 mddev->delta_disks = -1;
8488 mddev->raid_disks -= 1;
8489 return setup_conf(mddev);
8492 static int raid5_check_reshape(struct mddev *mddev)
8494 /* For a 2-drive array, the layout and chunk size can be changed
8495 * immediately as not restriping is needed.
8496 * For larger arrays we record the new value - after validation
8497 * to be used by a reshape pass.
8499 struct r5conf *conf = mddev->private;
8500 int new_chunk = mddev->new_chunk_sectors;
8502 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8504 if (new_chunk > 0) {
8505 if (!is_power_of_2(new_chunk))
8507 if (new_chunk < (PAGE_SIZE>>9))
8509 if (mddev->array_sectors & (new_chunk-1))
8510 /* not factor of array size */
8514 /* They look valid */
8516 if (mddev->raid_disks == 2) {
8517 /* can make the change immediately */
8518 if (mddev->new_layout >= 0) {
8519 conf->algorithm = mddev->new_layout;
8520 mddev->layout = mddev->new_layout;
8522 if (new_chunk > 0) {
8523 conf->chunk_sectors = new_chunk ;
8524 mddev->chunk_sectors = new_chunk;
8526 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8527 md_wakeup_thread(mddev->thread);
8529 return check_reshape(mddev);
8532 static int raid6_check_reshape(struct mddev *mddev)
8534 int new_chunk = mddev->new_chunk_sectors;
8536 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8538 if (new_chunk > 0) {
8539 if (!is_power_of_2(new_chunk))
8541 if (new_chunk < (PAGE_SIZE >> 9))
8543 if (mddev->array_sectors & (new_chunk-1))
8544 /* not factor of array size */
8548 /* They look valid */
8549 return check_reshape(mddev);
8552 static void *raid5_takeover(struct mddev *mddev)
8554 /* raid5 can take over:
8555 * raid0 - if there is only one strip zone - make it a raid4 layout
8556 * raid1 - if there are two drives. We need to know the chunk size
8557 * raid4 - trivial - just use a raid4 layout.
8558 * raid6 - Providing it is a *_6 layout
8560 if (mddev->level == 0)
8561 return raid45_takeover_raid0(mddev, 5);
8562 if (mddev->level == 1)
8563 return raid5_takeover_raid1(mddev);
8564 if (mddev->level == 4) {
8565 mddev->new_layout = ALGORITHM_PARITY_N;
8566 mddev->new_level = 5;
8567 return setup_conf(mddev);
8569 if (mddev->level == 6)
8570 return raid5_takeover_raid6(mddev);
8572 return ERR_PTR(-EINVAL);
8575 static void *raid4_takeover(struct mddev *mddev)
8577 /* raid4 can take over:
8578 * raid0 - if there is only one strip zone
8579 * raid5 - if layout is right
8581 if (mddev->level == 0)
8582 return raid45_takeover_raid0(mddev, 4);
8583 if (mddev->level == 5 &&
8584 mddev->layout == ALGORITHM_PARITY_N) {
8585 mddev->new_layout = 0;
8586 mddev->new_level = 4;
8587 return setup_conf(mddev);
8589 return ERR_PTR(-EINVAL);
8592 static struct md_personality raid5_personality;
8594 static void *raid6_takeover(struct mddev *mddev)
8596 /* Currently can only take over a raid5. We map the
8597 * personality to an equivalent raid6 personality
8598 * with the Q block at the end.
8602 if (mddev->pers != &raid5_personality)
8603 return ERR_PTR(-EINVAL);
8604 if (mddev->degraded > 1)
8605 return ERR_PTR(-EINVAL);
8606 if (mddev->raid_disks > 253)
8607 return ERR_PTR(-EINVAL);
8608 if (mddev->raid_disks < 3)
8609 return ERR_PTR(-EINVAL);
8611 switch (mddev->layout) {
8612 case ALGORITHM_LEFT_ASYMMETRIC:
8613 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8615 case ALGORITHM_RIGHT_ASYMMETRIC:
8616 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8618 case ALGORITHM_LEFT_SYMMETRIC:
8619 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8621 case ALGORITHM_RIGHT_SYMMETRIC:
8622 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8624 case ALGORITHM_PARITY_0:
8625 new_layout = ALGORITHM_PARITY_0_6;
8627 case ALGORITHM_PARITY_N:
8628 new_layout = ALGORITHM_PARITY_N;
8631 return ERR_PTR(-EINVAL);
8633 mddev->new_level = 6;
8634 mddev->new_layout = new_layout;
8635 mddev->delta_disks = 1;
8636 mddev->raid_disks += 1;
8637 return setup_conf(mddev);
8640 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8642 struct r5conf *conf;
8645 err = mddev_lock(mddev);
8648 conf = mddev->private;
8650 mddev_unlock(mddev);
8654 if (strncmp(buf, "ppl", 3) == 0) {
8655 /* ppl only works with RAID 5 */
8656 if (!raid5_has_ppl(conf) && conf->level == 5) {
8657 err = log_init(conf, NULL, true);
8659 err = resize_stripes(conf, conf->pool_size);
8665 } else if (strncmp(buf, "resync", 6) == 0) {
8666 if (raid5_has_ppl(conf)) {
8667 mddev_suspend(mddev);
8669 mddev_resume(mddev);
8670 err = resize_stripes(conf, conf->pool_size);
8671 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8672 r5l_log_disk_error(conf)) {
8673 bool journal_dev_exists = false;
8674 struct md_rdev *rdev;
8676 rdev_for_each(rdev, mddev)
8677 if (test_bit(Journal, &rdev->flags)) {
8678 journal_dev_exists = true;
8682 if (!journal_dev_exists) {
8683 mddev_suspend(mddev);
8684 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8685 mddev_resume(mddev);
8686 } else /* need remove journal device first */
8695 md_update_sb(mddev, 1);
8697 mddev_unlock(mddev);
8702 static int raid5_start(struct mddev *mddev)
8704 struct r5conf *conf = mddev->private;
8706 return r5l_start(conf->log);
8709 static struct md_personality raid6_personality =
8713 .owner = THIS_MODULE,
8714 .make_request = raid5_make_request,
8716 .start = raid5_start,
8718 .status = raid5_status,
8719 .error_handler = raid5_error,
8720 .hot_add_disk = raid5_add_disk,
8721 .hot_remove_disk= raid5_remove_disk,
8722 .spare_active = raid5_spare_active,
8723 .sync_request = raid5_sync_request,
8724 .resize = raid5_resize,
8726 .check_reshape = raid6_check_reshape,
8727 .start_reshape = raid5_start_reshape,
8728 .finish_reshape = raid5_finish_reshape,
8729 .quiesce = raid5_quiesce,
8730 .takeover = raid6_takeover,
8731 .change_consistency_policy = raid5_change_consistency_policy,
8733 static struct md_personality raid5_personality =
8737 .owner = THIS_MODULE,
8738 .make_request = raid5_make_request,
8740 .start = raid5_start,
8742 .status = raid5_status,
8743 .error_handler = raid5_error,
8744 .hot_add_disk = raid5_add_disk,
8745 .hot_remove_disk= raid5_remove_disk,
8746 .spare_active = raid5_spare_active,
8747 .sync_request = raid5_sync_request,
8748 .resize = raid5_resize,
8750 .check_reshape = raid5_check_reshape,
8751 .start_reshape = raid5_start_reshape,
8752 .finish_reshape = raid5_finish_reshape,
8753 .quiesce = raid5_quiesce,
8754 .takeover = raid5_takeover,
8755 .change_consistency_policy = raid5_change_consistency_policy,
8758 static struct md_personality raid4_personality =
8762 .owner = THIS_MODULE,
8763 .make_request = raid5_make_request,
8765 .start = raid5_start,
8767 .status = raid5_status,
8768 .error_handler = raid5_error,
8769 .hot_add_disk = raid5_add_disk,
8770 .hot_remove_disk= raid5_remove_disk,
8771 .spare_active = raid5_spare_active,
8772 .sync_request = raid5_sync_request,
8773 .resize = raid5_resize,
8775 .check_reshape = raid5_check_reshape,
8776 .start_reshape = raid5_start_reshape,
8777 .finish_reshape = raid5_finish_reshape,
8778 .quiesce = raid5_quiesce,
8779 .takeover = raid4_takeover,
8780 .change_consistency_policy = raid5_change_consistency_policy,
8783 static int __init raid5_init(void)
8787 raid5_wq = alloc_workqueue("raid5wq",
8788 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8792 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8794 raid456_cpu_up_prepare,
8797 destroy_workqueue(raid5_wq);
8800 register_md_personality(&raid6_personality);
8801 register_md_personality(&raid5_personality);
8802 register_md_personality(&raid4_personality);
8806 static void raid5_exit(void)
8808 unregister_md_personality(&raid6_personality);
8809 unregister_md_personality(&raid5_personality);
8810 unregister_md_personality(&raid4_personality);
8811 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8812 destroy_workqueue(raid5_wq);
8815 module_init(raid5_init);
8816 module_exit(raid5_exit);
8817 MODULE_LICENSE("GPL");
8818 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8819 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8820 MODULE_ALIAS("md-raid5");
8821 MODULE_ALIAS("md-raid4");
8822 MODULE_ALIAS("md-level-5");
8823 MODULE_ALIAS("md-level-4");
8824 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8825 MODULE_ALIAS("md-raid6");
8826 MODULE_ALIAS("md-level-6");
8828 /* This used to be two separate modules, they were: */
8829 MODULE_ALIAS("raid5");
8830 MODULE_ALIAS("raid6");