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 >> STRIPE_SHIFT) & HASH_MASK;
73 return &conf->stripe_hashtbl[hash];
76 static inline int stripe_hash_locks_hash(sector_t sect)
78 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
81 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
83 spin_lock_irq(conf->hash_locks + hash);
84 spin_lock(&conf->device_lock);
87 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
89 spin_unlock(&conf->device_lock);
90 spin_unlock_irq(conf->hash_locks + hash);
93 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
96 spin_lock_irq(conf->hash_locks);
97 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
98 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
99 spin_lock(&conf->device_lock);
102 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
105 spin_unlock(&conf->device_lock);
106 for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
107 spin_unlock(conf->hash_locks + i);
108 spin_unlock_irq(conf->hash_locks);
111 /* Find first data disk in a raid6 stripe */
112 static inline int raid6_d0(struct stripe_head *sh)
115 /* ddf always start from first device */
117 /* md starts just after Q block */
118 if (sh->qd_idx == sh->disks - 1)
121 return sh->qd_idx + 1;
123 static inline int raid6_next_disk(int disk, int raid_disks)
126 return (disk < raid_disks) ? disk : 0;
129 /* When walking through the disks in a raid5, starting at raid6_d0,
130 * We need to map each disk to a 'slot', where the data disks are slot
131 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
132 * is raid_disks-1. This help does that mapping.
134 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
135 int *count, int syndrome_disks)
141 if (idx == sh->pd_idx)
142 return syndrome_disks;
143 if (idx == sh->qd_idx)
144 return syndrome_disks + 1;
150 static void print_raid5_conf (struct r5conf *conf);
152 static int stripe_operations_active(struct stripe_head *sh)
154 return sh->check_state || sh->reconstruct_state ||
155 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
156 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
159 static bool stripe_is_lowprio(struct stripe_head *sh)
161 return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
162 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
163 !test_bit(STRIPE_R5C_CACHING, &sh->state);
166 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
168 struct r5conf *conf = sh->raid_conf;
169 struct r5worker_group *group;
171 int i, cpu = sh->cpu;
173 if (!cpu_online(cpu)) {
174 cpu = cpumask_any(cpu_online_mask);
178 if (list_empty(&sh->lru)) {
179 struct r5worker_group *group;
180 group = conf->worker_groups + cpu_to_group(cpu);
181 if (stripe_is_lowprio(sh))
182 list_add_tail(&sh->lru, &group->loprio_list);
184 list_add_tail(&sh->lru, &group->handle_list);
185 group->stripes_cnt++;
189 if (conf->worker_cnt_per_group == 0) {
190 md_wakeup_thread(conf->mddev->thread);
194 group = conf->worker_groups + cpu_to_group(sh->cpu);
196 group->workers[0].working = true;
197 /* at least one worker should run to avoid race */
198 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
200 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
201 /* wakeup more workers */
202 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
203 if (group->workers[i].working == false) {
204 group->workers[i].working = true;
205 queue_work_on(sh->cpu, raid5_wq,
206 &group->workers[i].work);
212 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
213 struct list_head *temp_inactive_list)
216 int injournal = 0; /* number of date pages with R5_InJournal */
218 BUG_ON(!list_empty(&sh->lru));
219 BUG_ON(atomic_read(&conf->active_stripes)==0);
221 if (r5c_is_writeback(conf->log))
222 for (i = sh->disks; i--; )
223 if (test_bit(R5_InJournal, &sh->dev[i].flags))
226 * In the following cases, the stripe cannot be released to cached
227 * lists. Therefore, we make the stripe write out and set
229 * 1. when quiesce in r5c write back;
230 * 2. when resync is requested fot the stripe.
232 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
233 (conf->quiesce && r5c_is_writeback(conf->log) &&
234 !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
235 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
236 r5c_make_stripe_write_out(sh);
237 set_bit(STRIPE_HANDLE, &sh->state);
240 if (test_bit(STRIPE_HANDLE, &sh->state)) {
241 if (test_bit(STRIPE_DELAYED, &sh->state) &&
242 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
243 list_add_tail(&sh->lru, &conf->delayed_list);
244 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
245 sh->bm_seq - conf->seq_write > 0)
246 list_add_tail(&sh->lru, &conf->bitmap_list);
248 clear_bit(STRIPE_DELAYED, &sh->state);
249 clear_bit(STRIPE_BIT_DELAY, &sh->state);
250 if (conf->worker_cnt_per_group == 0) {
251 if (stripe_is_lowprio(sh))
252 list_add_tail(&sh->lru,
255 list_add_tail(&sh->lru,
258 raid5_wakeup_stripe_thread(sh);
262 md_wakeup_thread(conf->mddev->thread);
264 BUG_ON(stripe_operations_active(sh));
265 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
266 if (atomic_dec_return(&conf->preread_active_stripes)
268 md_wakeup_thread(conf->mddev->thread);
269 atomic_dec(&conf->active_stripes);
270 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
271 if (!r5c_is_writeback(conf->log))
272 list_add_tail(&sh->lru, temp_inactive_list);
274 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
276 list_add_tail(&sh->lru, temp_inactive_list);
277 else if (injournal == conf->raid_disks - conf->max_degraded) {
279 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
280 atomic_inc(&conf->r5c_cached_full_stripes);
281 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
282 atomic_dec(&conf->r5c_cached_partial_stripes);
283 list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
284 r5c_check_cached_full_stripe(conf);
287 * STRIPE_R5C_PARTIAL_STRIPE is set in
288 * r5c_try_caching_write(). No need to
291 list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
297 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
298 struct list_head *temp_inactive_list)
300 if (atomic_dec_and_test(&sh->count))
301 do_release_stripe(conf, sh, temp_inactive_list);
305 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
307 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
308 * given time. Adding stripes only takes device lock, while deleting stripes
309 * only takes hash lock.
311 static void release_inactive_stripe_list(struct r5conf *conf,
312 struct list_head *temp_inactive_list,
316 bool do_wakeup = false;
319 if (hash == NR_STRIPE_HASH_LOCKS) {
320 size = NR_STRIPE_HASH_LOCKS;
321 hash = NR_STRIPE_HASH_LOCKS - 1;
325 struct list_head *list = &temp_inactive_list[size - 1];
328 * We don't hold any lock here yet, raid5_get_active_stripe() might
329 * remove stripes from the list
331 if (!list_empty_careful(list)) {
332 spin_lock_irqsave(conf->hash_locks + hash, flags);
333 if (list_empty(conf->inactive_list + hash) &&
335 atomic_dec(&conf->empty_inactive_list_nr);
336 list_splice_tail_init(list, conf->inactive_list + hash);
338 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
345 wake_up(&conf->wait_for_stripe);
346 if (atomic_read(&conf->active_stripes) == 0)
347 wake_up(&conf->wait_for_quiescent);
348 if (conf->retry_read_aligned)
349 md_wakeup_thread(conf->mddev->thread);
353 /* should hold conf->device_lock already */
354 static int release_stripe_list(struct r5conf *conf,
355 struct list_head *temp_inactive_list)
357 struct stripe_head *sh, *t;
359 struct llist_node *head;
361 head = llist_del_all(&conf->released_stripes);
362 head = llist_reverse_order(head);
363 llist_for_each_entry_safe(sh, t, head, release_list) {
366 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
368 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
370 * Don't worry the bit is set here, because if the bit is set
371 * again, the count is always > 1. This is true for
372 * STRIPE_ON_UNPLUG_LIST bit too.
374 hash = sh->hash_lock_index;
375 __release_stripe(conf, sh, &temp_inactive_list[hash]);
382 void raid5_release_stripe(struct stripe_head *sh)
384 struct r5conf *conf = sh->raid_conf;
386 struct list_head list;
390 /* Avoid release_list until the last reference.
392 if (atomic_add_unless(&sh->count, -1, 1))
395 if (unlikely(!conf->mddev->thread) ||
396 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
398 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
400 md_wakeup_thread(conf->mddev->thread);
403 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
404 if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
405 INIT_LIST_HEAD(&list);
406 hash = sh->hash_lock_index;
407 do_release_stripe(conf, sh, &list);
408 spin_unlock_irqrestore(&conf->device_lock, flags);
409 release_inactive_stripe_list(conf, &list, hash);
413 static inline void remove_hash(struct stripe_head *sh)
415 pr_debug("remove_hash(), stripe %llu\n",
416 (unsigned long long)sh->sector);
418 hlist_del_init(&sh->hash);
421 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
423 struct hlist_head *hp = stripe_hash(conf, sh->sector);
425 pr_debug("insert_hash(), stripe %llu\n",
426 (unsigned long long)sh->sector);
428 hlist_add_head(&sh->hash, hp);
431 /* find an idle stripe, make sure it is unhashed, and return it. */
432 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
434 struct stripe_head *sh = NULL;
435 struct list_head *first;
437 if (list_empty(conf->inactive_list + hash))
439 first = (conf->inactive_list + hash)->next;
440 sh = list_entry(first, struct stripe_head, lru);
441 list_del_init(first);
443 atomic_inc(&conf->active_stripes);
444 BUG_ON(hash != sh->hash_lock_index);
445 if (list_empty(conf->inactive_list + hash))
446 atomic_inc(&conf->empty_inactive_list_nr);
451 static void shrink_buffers(struct stripe_head *sh)
455 int num = sh->raid_conf->pool_size;
457 for (i = 0; i < num ; i++) {
458 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
462 sh->dev[i].page = NULL;
467 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
470 int num = sh->raid_conf->pool_size;
472 for (i = 0; i < num; i++) {
475 if (!(page = alloc_page(gfp))) {
478 sh->dev[i].page = page;
479 sh->dev[i].orig_page = page;
485 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
486 struct stripe_head *sh);
488 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
490 struct r5conf *conf = sh->raid_conf;
493 BUG_ON(atomic_read(&sh->count) != 0);
494 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
495 BUG_ON(stripe_operations_active(sh));
496 BUG_ON(sh->batch_head);
498 pr_debug("init_stripe called, stripe %llu\n",
499 (unsigned long long)sector);
501 seq = read_seqcount_begin(&conf->gen_lock);
502 sh->generation = conf->generation - previous;
503 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
505 stripe_set_idx(sector, conf, previous, sh);
508 for (i = sh->disks; i--; ) {
509 struct r5dev *dev = &sh->dev[i];
511 if (dev->toread || dev->read || dev->towrite || dev->written ||
512 test_bit(R5_LOCKED, &dev->flags)) {
513 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
514 (unsigned long long)sh->sector, i, dev->toread,
515 dev->read, dev->towrite, dev->written,
516 test_bit(R5_LOCKED, &dev->flags));
520 dev->sector = raid5_compute_blocknr(sh, i, previous);
522 if (read_seqcount_retry(&conf->gen_lock, seq))
524 sh->overwrite_disks = 0;
525 insert_hash(conf, sh);
526 sh->cpu = smp_processor_id();
527 set_bit(STRIPE_BATCH_READY, &sh->state);
530 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
533 struct stripe_head *sh;
535 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
536 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
537 if (sh->sector == sector && sh->generation == generation)
539 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
544 * Need to check if array has failed when deciding whether to:
546 * - remove non-faulty devices
549 * This determination is simple when no reshape is happening.
550 * However if there is a reshape, we need to carefully check
551 * both the before and after sections.
552 * This is because some failed devices may only affect one
553 * of the two sections, and some non-in_sync devices may
554 * be insync in the section most affected by failed devices.
556 int raid5_calc_degraded(struct r5conf *conf)
558 int degraded, degraded2;
563 for (i = 0; i < conf->previous_raid_disks; i++) {
564 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
565 if (rdev && test_bit(Faulty, &rdev->flags))
566 rdev = rcu_dereference(conf->disks[i].replacement);
567 if (!rdev || test_bit(Faulty, &rdev->flags))
569 else if (test_bit(In_sync, &rdev->flags))
572 /* not in-sync or faulty.
573 * If the reshape increases the number of devices,
574 * this is being recovered by the reshape, so
575 * this 'previous' section is not in_sync.
576 * If the number of devices is being reduced however,
577 * the device can only be part of the array if
578 * we are reverting a reshape, so this section will
581 if (conf->raid_disks >= conf->previous_raid_disks)
585 if (conf->raid_disks == conf->previous_raid_disks)
589 for (i = 0; i < conf->raid_disks; i++) {
590 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
591 if (rdev && test_bit(Faulty, &rdev->flags))
592 rdev = rcu_dereference(conf->disks[i].replacement);
593 if (!rdev || test_bit(Faulty, &rdev->flags))
595 else if (test_bit(In_sync, &rdev->flags))
598 /* not in-sync or faulty.
599 * If reshape increases the number of devices, this
600 * section has already been recovered, else it
601 * almost certainly hasn't.
603 if (conf->raid_disks <= conf->previous_raid_disks)
607 if (degraded2 > degraded)
612 static int has_failed(struct r5conf *conf)
616 if (conf->mddev->reshape_position == MaxSector)
617 return conf->mddev->degraded > conf->max_degraded;
619 degraded = raid5_calc_degraded(conf);
620 if (degraded > conf->max_degraded)
626 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
627 int previous, int noblock, int noquiesce)
629 struct stripe_head *sh;
630 int hash = stripe_hash_locks_hash(sector);
631 int inc_empty_inactive_list_flag;
633 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
635 spin_lock_irq(conf->hash_locks + hash);
638 wait_event_lock_irq(conf->wait_for_quiescent,
639 conf->quiesce == 0 || noquiesce,
640 *(conf->hash_locks + hash));
641 sh = __find_stripe(conf, sector, conf->generation - previous);
643 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
644 sh = get_free_stripe(conf, hash);
645 if (!sh && !test_bit(R5_DID_ALLOC,
647 set_bit(R5_ALLOC_MORE,
650 if (noblock && sh == NULL)
653 r5c_check_stripe_cache_usage(conf);
655 set_bit(R5_INACTIVE_BLOCKED,
657 r5l_wake_reclaim(conf->log, 0);
659 conf->wait_for_stripe,
660 !list_empty(conf->inactive_list + hash) &&
661 (atomic_read(&conf->active_stripes)
662 < (conf->max_nr_stripes * 3 / 4)
663 || !test_bit(R5_INACTIVE_BLOCKED,
664 &conf->cache_state)),
665 *(conf->hash_locks + hash));
666 clear_bit(R5_INACTIVE_BLOCKED,
669 init_stripe(sh, sector, previous);
670 atomic_inc(&sh->count);
672 } else if (!atomic_inc_not_zero(&sh->count)) {
673 spin_lock(&conf->device_lock);
674 if (!atomic_read(&sh->count)) {
675 if (!test_bit(STRIPE_HANDLE, &sh->state))
676 atomic_inc(&conf->active_stripes);
677 BUG_ON(list_empty(&sh->lru) &&
678 !test_bit(STRIPE_EXPANDING, &sh->state));
679 inc_empty_inactive_list_flag = 0;
680 if (!list_empty(conf->inactive_list + hash))
681 inc_empty_inactive_list_flag = 1;
682 list_del_init(&sh->lru);
683 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
684 atomic_inc(&conf->empty_inactive_list_nr);
686 sh->group->stripes_cnt--;
690 atomic_inc(&sh->count);
691 spin_unlock(&conf->device_lock);
693 } while (sh == NULL);
695 spin_unlock_irq(conf->hash_locks + hash);
699 static bool is_full_stripe_write(struct stripe_head *sh)
701 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
702 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
705 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
706 __acquires(&sh1->stripe_lock)
707 __acquires(&sh2->stripe_lock)
710 spin_lock_irq(&sh2->stripe_lock);
711 spin_lock_nested(&sh1->stripe_lock, 1);
713 spin_lock_irq(&sh1->stripe_lock);
714 spin_lock_nested(&sh2->stripe_lock, 1);
718 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
719 __releases(&sh1->stripe_lock)
720 __releases(&sh2->stripe_lock)
722 spin_unlock(&sh1->stripe_lock);
723 spin_unlock_irq(&sh2->stripe_lock);
726 /* Only freshly new full stripe normal write stripe can be added to a batch list */
727 static bool stripe_can_batch(struct stripe_head *sh)
729 struct r5conf *conf = sh->raid_conf;
731 if (raid5_has_log(conf) || raid5_has_ppl(conf))
733 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
734 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
735 is_full_stripe_write(sh);
738 /* we only do back search */
739 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
741 struct stripe_head *head;
742 sector_t head_sector, tmp_sec;
745 int inc_empty_inactive_list_flag;
747 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
748 tmp_sec = sh->sector;
749 if (!sector_div(tmp_sec, conf->chunk_sectors))
751 head_sector = sh->sector - STRIPE_SECTORS;
753 hash = stripe_hash_locks_hash(head_sector);
754 spin_lock_irq(conf->hash_locks + hash);
755 head = __find_stripe(conf, head_sector, conf->generation);
756 if (head && !atomic_inc_not_zero(&head->count)) {
757 spin_lock(&conf->device_lock);
758 if (!atomic_read(&head->count)) {
759 if (!test_bit(STRIPE_HANDLE, &head->state))
760 atomic_inc(&conf->active_stripes);
761 BUG_ON(list_empty(&head->lru) &&
762 !test_bit(STRIPE_EXPANDING, &head->state));
763 inc_empty_inactive_list_flag = 0;
764 if (!list_empty(conf->inactive_list + hash))
765 inc_empty_inactive_list_flag = 1;
766 list_del_init(&head->lru);
767 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
768 atomic_inc(&conf->empty_inactive_list_nr);
770 head->group->stripes_cnt--;
774 atomic_inc(&head->count);
775 spin_unlock(&conf->device_lock);
777 spin_unlock_irq(conf->hash_locks + hash);
781 if (!stripe_can_batch(head))
784 lock_two_stripes(head, sh);
785 /* clear_batch_ready clear the flag */
786 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
793 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
795 if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
796 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
799 if (head->batch_head) {
800 spin_lock(&head->batch_head->batch_lock);
801 /* This batch list is already running */
802 if (!stripe_can_batch(head)) {
803 spin_unlock(&head->batch_head->batch_lock);
807 * We must assign batch_head of this stripe within the
808 * batch_lock, otherwise clear_batch_ready of batch head
809 * stripe could clear BATCH_READY bit of this stripe and
810 * this stripe->batch_head doesn't get assigned, which
811 * could confuse clear_batch_ready for this stripe
813 sh->batch_head = head->batch_head;
816 * at this point, head's BATCH_READY could be cleared, but we
817 * can still add the stripe to batch list
819 list_add(&sh->batch_list, &head->batch_list);
820 spin_unlock(&head->batch_head->batch_lock);
822 head->batch_head = head;
823 sh->batch_head = head->batch_head;
824 spin_lock(&head->batch_lock);
825 list_add_tail(&sh->batch_list, &head->batch_list);
826 spin_unlock(&head->batch_lock);
829 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
830 if (atomic_dec_return(&conf->preread_active_stripes)
832 md_wakeup_thread(conf->mddev->thread);
834 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
835 int seq = sh->bm_seq;
836 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
837 sh->batch_head->bm_seq > seq)
838 seq = sh->batch_head->bm_seq;
839 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
840 sh->batch_head->bm_seq = seq;
843 atomic_inc(&sh->count);
845 unlock_two_stripes(head, sh);
847 raid5_release_stripe(head);
850 /* Determine if 'data_offset' or 'new_data_offset' should be used
851 * in this stripe_head.
853 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
855 sector_t progress = conf->reshape_progress;
856 /* Need a memory barrier to make sure we see the value
857 * of conf->generation, or ->data_offset that was set before
858 * reshape_progress was updated.
861 if (progress == MaxSector)
863 if (sh->generation == conf->generation - 1)
865 /* We are in a reshape, and this is a new-generation stripe,
866 * so use new_data_offset.
871 static void dispatch_bio_list(struct bio_list *tmp)
875 while ((bio = bio_list_pop(tmp)))
876 generic_make_request(bio);
879 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
881 const struct r5pending_data *da = list_entry(a,
882 struct r5pending_data, sibling);
883 const struct r5pending_data *db = list_entry(b,
884 struct r5pending_data, sibling);
885 if (da->sector > db->sector)
887 if (da->sector < db->sector)
892 static void dispatch_defer_bios(struct r5conf *conf, int target,
893 struct bio_list *list)
895 struct r5pending_data *data;
896 struct list_head *first, *next = NULL;
899 if (conf->pending_data_cnt == 0)
902 list_sort(NULL, &conf->pending_list, cmp_stripe);
904 first = conf->pending_list.next;
906 /* temporarily move the head */
907 if (conf->next_pending_data)
908 list_move_tail(&conf->pending_list,
909 &conf->next_pending_data->sibling);
911 while (!list_empty(&conf->pending_list)) {
912 data = list_first_entry(&conf->pending_list,
913 struct r5pending_data, sibling);
914 if (&data->sibling == first)
915 first = data->sibling.next;
916 next = data->sibling.next;
918 bio_list_merge(list, &data->bios);
919 list_move(&data->sibling, &conf->free_list);
924 conf->pending_data_cnt -= cnt;
925 BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
927 if (next != &conf->pending_list)
928 conf->next_pending_data = list_entry(next,
929 struct r5pending_data, sibling);
931 conf->next_pending_data = NULL;
932 /* list isn't empty */
933 if (first != &conf->pending_list)
934 list_move_tail(&conf->pending_list, first);
937 static void flush_deferred_bios(struct r5conf *conf)
939 struct bio_list tmp = BIO_EMPTY_LIST;
941 if (conf->pending_data_cnt == 0)
944 spin_lock(&conf->pending_bios_lock);
945 dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
946 BUG_ON(conf->pending_data_cnt != 0);
947 spin_unlock(&conf->pending_bios_lock);
949 dispatch_bio_list(&tmp);
952 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
953 struct bio_list *bios)
955 struct bio_list tmp = BIO_EMPTY_LIST;
956 struct r5pending_data *ent;
958 spin_lock(&conf->pending_bios_lock);
959 ent = list_first_entry(&conf->free_list, struct r5pending_data,
961 list_move_tail(&ent->sibling, &conf->pending_list);
962 ent->sector = sector;
963 bio_list_init(&ent->bios);
964 bio_list_merge(&ent->bios, bios);
965 conf->pending_data_cnt++;
966 if (conf->pending_data_cnt >= PENDING_IO_MAX)
967 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
969 spin_unlock(&conf->pending_bios_lock);
971 dispatch_bio_list(&tmp);
975 raid5_end_read_request(struct bio *bi);
977 raid5_end_write_request(struct bio *bi);
979 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
981 struct r5conf *conf = sh->raid_conf;
982 int i, disks = sh->disks;
983 struct stripe_head *head_sh = sh;
984 struct bio_list pending_bios = BIO_EMPTY_LIST;
989 if (log_stripe(sh, s) == 0)
992 should_defer = conf->batch_bio_dispatch && conf->group_cnt;
994 for (i = disks; i--; ) {
995 int op, op_flags = 0;
996 int replace_only = 0;
997 struct bio *bi, *rbi;
998 struct md_rdev *rdev, *rrdev = NULL;
1001 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1003 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1005 if (test_bit(R5_Discard, &sh->dev[i].flags))
1006 op = REQ_OP_DISCARD;
1007 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1009 else if (test_and_clear_bit(R5_WantReplace,
1010 &sh->dev[i].flags)) {
1015 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1016 op_flags |= REQ_SYNC;
1019 bi = &sh->dev[i].req;
1020 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1023 rrdev = rcu_dereference(conf->disks[i].replacement);
1024 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1025 rdev = rcu_dereference(conf->disks[i].rdev);
1030 if (op_is_write(op)) {
1034 /* We raced and saw duplicates */
1037 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1042 if (rdev && test_bit(Faulty, &rdev->flags))
1045 atomic_inc(&rdev->nr_pending);
1046 if (rrdev && test_bit(Faulty, &rrdev->flags))
1049 atomic_inc(&rrdev->nr_pending);
1052 /* We have already checked bad blocks for reads. Now
1053 * need to check for writes. We never accept write errors
1054 * on the replacement, so we don't to check rrdev.
1056 while (op_is_write(op) && rdev &&
1057 test_bit(WriteErrorSeen, &rdev->flags)) {
1060 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1061 &first_bad, &bad_sectors);
1066 set_bit(BlockedBadBlocks, &rdev->flags);
1067 if (!conf->mddev->external &&
1068 conf->mddev->sb_flags) {
1069 /* It is very unlikely, but we might
1070 * still need to write out the
1071 * bad block log - better give it
1073 md_check_recovery(conf->mddev);
1076 * Because md_wait_for_blocked_rdev
1077 * will dec nr_pending, we must
1078 * increment it first.
1080 atomic_inc(&rdev->nr_pending);
1081 md_wait_for_blocked_rdev(rdev, conf->mddev);
1083 /* Acknowledged bad block - skip the write */
1084 rdev_dec_pending(rdev, conf->mddev);
1090 if (s->syncing || s->expanding || s->expanded
1092 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1094 set_bit(STRIPE_IO_STARTED, &sh->state);
1096 bio_set_dev(bi, rdev->bdev);
1097 bio_set_op_attrs(bi, op, op_flags);
1098 bi->bi_end_io = op_is_write(op)
1099 ? raid5_end_write_request
1100 : raid5_end_read_request;
1101 bi->bi_private = sh;
1103 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1104 __func__, (unsigned long long)sh->sector,
1106 atomic_inc(&sh->count);
1108 atomic_inc(&head_sh->count);
1109 if (use_new_offset(conf, sh))
1110 bi->bi_iter.bi_sector = (sh->sector
1111 + rdev->new_data_offset);
1113 bi->bi_iter.bi_sector = (sh->sector
1114 + rdev->data_offset);
1115 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1116 bi->bi_opf |= REQ_NOMERGE;
1118 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1119 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1121 if (!op_is_write(op) &&
1122 test_bit(R5_InJournal, &sh->dev[i].flags))
1124 * issuing read for a page in journal, this
1125 * must be preparing for prexor in rmw; read
1126 * the data into orig_page
1128 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1130 sh->dev[i].vec.bv_page = sh->dev[i].page;
1132 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1133 bi->bi_io_vec[0].bv_offset = 0;
1134 bi->bi_iter.bi_size = STRIPE_SIZE;
1135 bi->bi_write_hint = sh->dev[i].write_hint;
1137 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1139 * If this is discard request, set bi_vcnt 0. We don't
1140 * want to confuse SCSI because SCSI will replace payload
1142 if (op == REQ_OP_DISCARD)
1145 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1147 if (conf->mddev->gendisk)
1148 trace_block_bio_remap(bi->bi_disk->queue,
1149 bi, disk_devt(conf->mddev->gendisk),
1151 if (should_defer && op_is_write(op))
1152 bio_list_add(&pending_bios, bi);
1154 generic_make_request(bi);
1157 if (s->syncing || s->expanding || s->expanded
1159 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1161 set_bit(STRIPE_IO_STARTED, &sh->state);
1163 bio_set_dev(rbi, rrdev->bdev);
1164 bio_set_op_attrs(rbi, op, op_flags);
1165 BUG_ON(!op_is_write(op));
1166 rbi->bi_end_io = raid5_end_write_request;
1167 rbi->bi_private = sh;
1169 pr_debug("%s: for %llu schedule op %d on "
1170 "replacement disc %d\n",
1171 __func__, (unsigned long long)sh->sector,
1173 atomic_inc(&sh->count);
1175 atomic_inc(&head_sh->count);
1176 if (use_new_offset(conf, sh))
1177 rbi->bi_iter.bi_sector = (sh->sector
1178 + rrdev->new_data_offset);
1180 rbi->bi_iter.bi_sector = (sh->sector
1181 + rrdev->data_offset);
1182 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1183 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1184 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1186 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1187 rbi->bi_io_vec[0].bv_offset = 0;
1188 rbi->bi_iter.bi_size = STRIPE_SIZE;
1189 rbi->bi_write_hint = sh->dev[i].write_hint;
1190 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1192 * If this is discard request, set bi_vcnt 0. We don't
1193 * want to confuse SCSI because SCSI will replace payload
1195 if (op == REQ_OP_DISCARD)
1197 if (conf->mddev->gendisk)
1198 trace_block_bio_remap(rbi->bi_disk->queue,
1199 rbi, disk_devt(conf->mddev->gendisk),
1201 if (should_defer && op_is_write(op))
1202 bio_list_add(&pending_bios, rbi);
1204 generic_make_request(rbi);
1206 if (!rdev && !rrdev) {
1207 if (op_is_write(op))
1208 set_bit(STRIPE_DEGRADED, &sh->state);
1209 pr_debug("skip op %d on disc %d for sector %llu\n",
1210 bi->bi_opf, i, (unsigned long long)sh->sector);
1211 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1212 set_bit(STRIPE_HANDLE, &sh->state);
1215 if (!head_sh->batch_head)
1217 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1223 if (should_defer && !bio_list_empty(&pending_bios))
1224 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1227 static struct dma_async_tx_descriptor *
1228 async_copy_data(int frombio, struct bio *bio, struct page **page,
1229 sector_t sector, struct dma_async_tx_descriptor *tx,
1230 struct stripe_head *sh, int no_skipcopy)
1233 struct bvec_iter iter;
1234 struct page *bio_page;
1236 struct async_submit_ctl submit;
1237 enum async_tx_flags flags = 0;
1239 if (bio->bi_iter.bi_sector >= sector)
1240 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1242 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1245 flags |= ASYNC_TX_FENCE;
1246 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1248 bio_for_each_segment(bvl, bio, iter) {
1249 int len = bvl.bv_len;
1253 if (page_offset < 0) {
1254 b_offset = -page_offset;
1255 page_offset += b_offset;
1259 if (len > 0 && page_offset + len > STRIPE_SIZE)
1260 clen = STRIPE_SIZE - page_offset;
1265 b_offset += bvl.bv_offset;
1266 bio_page = bvl.bv_page;
1268 if (sh->raid_conf->skip_copy &&
1269 b_offset == 0 && page_offset == 0 &&
1270 clen == STRIPE_SIZE &&
1274 tx = async_memcpy(*page, bio_page, page_offset,
1275 b_offset, clen, &submit);
1277 tx = async_memcpy(bio_page, *page, b_offset,
1278 page_offset, clen, &submit);
1280 /* chain the operations */
1281 submit.depend_tx = tx;
1283 if (clen < len) /* hit end of page */
1291 static void ops_complete_biofill(void *stripe_head_ref)
1293 struct stripe_head *sh = stripe_head_ref;
1296 pr_debug("%s: stripe %llu\n", __func__,
1297 (unsigned long long)sh->sector);
1299 /* clear completed biofills */
1300 for (i = sh->disks; i--; ) {
1301 struct r5dev *dev = &sh->dev[i];
1303 /* acknowledge completion of a biofill operation */
1304 /* and check if we need to reply to a read request,
1305 * new R5_Wantfill requests are held off until
1306 * !STRIPE_BIOFILL_RUN
1308 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1309 struct bio *rbi, *rbi2;
1314 while (rbi && rbi->bi_iter.bi_sector <
1315 dev->sector + STRIPE_SECTORS) {
1316 rbi2 = r5_next_bio(rbi, dev->sector);
1322 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1324 set_bit(STRIPE_HANDLE, &sh->state);
1325 raid5_release_stripe(sh);
1328 static void ops_run_biofill(struct stripe_head *sh)
1330 struct dma_async_tx_descriptor *tx = NULL;
1331 struct async_submit_ctl submit;
1334 BUG_ON(sh->batch_head);
1335 pr_debug("%s: stripe %llu\n", __func__,
1336 (unsigned long long)sh->sector);
1338 for (i = sh->disks; i--; ) {
1339 struct r5dev *dev = &sh->dev[i];
1340 if (test_bit(R5_Wantfill, &dev->flags)) {
1342 spin_lock_irq(&sh->stripe_lock);
1343 dev->read = rbi = dev->toread;
1345 spin_unlock_irq(&sh->stripe_lock);
1346 while (rbi && rbi->bi_iter.bi_sector <
1347 dev->sector + STRIPE_SECTORS) {
1348 tx = async_copy_data(0, rbi, &dev->page,
1349 dev->sector, tx, sh, 0);
1350 rbi = r5_next_bio(rbi, dev->sector);
1355 atomic_inc(&sh->count);
1356 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1357 async_trigger_callback(&submit);
1360 static void mark_target_uptodate(struct stripe_head *sh, int target)
1367 tgt = &sh->dev[target];
1368 set_bit(R5_UPTODATE, &tgt->flags);
1369 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1370 clear_bit(R5_Wantcompute, &tgt->flags);
1373 static void ops_complete_compute(void *stripe_head_ref)
1375 struct stripe_head *sh = stripe_head_ref;
1377 pr_debug("%s: stripe %llu\n", __func__,
1378 (unsigned long long)sh->sector);
1380 /* mark the computed target(s) as uptodate */
1381 mark_target_uptodate(sh, sh->ops.target);
1382 mark_target_uptodate(sh, sh->ops.target2);
1384 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1385 if (sh->check_state == check_state_compute_run)
1386 sh->check_state = check_state_compute_result;
1387 set_bit(STRIPE_HANDLE, &sh->state);
1388 raid5_release_stripe(sh);
1391 /* return a pointer to the address conversion region of the scribble buffer */
1392 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1394 return percpu->scribble + i * percpu->scribble_obj_size;
1397 /* return a pointer to the address conversion region of the scribble buffer */
1398 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1399 struct raid5_percpu *percpu, int i)
1401 return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1404 static struct dma_async_tx_descriptor *
1405 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1407 int disks = sh->disks;
1408 struct page **xor_srcs = to_addr_page(percpu, 0);
1409 int target = sh->ops.target;
1410 struct r5dev *tgt = &sh->dev[target];
1411 struct page *xor_dest = tgt->page;
1413 struct dma_async_tx_descriptor *tx;
1414 struct async_submit_ctl submit;
1417 BUG_ON(sh->batch_head);
1419 pr_debug("%s: stripe %llu block: %d\n",
1420 __func__, (unsigned long long)sh->sector, target);
1421 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1423 for (i = disks; i--; )
1425 xor_srcs[count++] = sh->dev[i].page;
1427 atomic_inc(&sh->count);
1429 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1430 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1431 if (unlikely(count == 1))
1432 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1434 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1439 /* set_syndrome_sources - populate source buffers for gen_syndrome
1440 * @srcs - (struct page *) array of size sh->disks
1441 * @sh - stripe_head to parse
1443 * Populates srcs in proper layout order for the stripe and returns the
1444 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1445 * destination buffer is recorded in srcs[count] and the Q destination
1446 * is recorded in srcs[count+1]].
1448 static int set_syndrome_sources(struct page **srcs,
1449 struct stripe_head *sh,
1452 int disks = sh->disks;
1453 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1454 int d0_idx = raid6_d0(sh);
1458 for (i = 0; i < disks; i++)
1464 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1465 struct r5dev *dev = &sh->dev[i];
1467 if (i == sh->qd_idx || i == sh->pd_idx ||
1468 (srctype == SYNDROME_SRC_ALL) ||
1469 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1470 (test_bit(R5_Wantdrain, &dev->flags) ||
1471 test_bit(R5_InJournal, &dev->flags))) ||
1472 (srctype == SYNDROME_SRC_WRITTEN &&
1474 test_bit(R5_InJournal, &dev->flags)))) {
1475 if (test_bit(R5_InJournal, &dev->flags))
1476 srcs[slot] = sh->dev[i].orig_page;
1478 srcs[slot] = sh->dev[i].page;
1480 i = raid6_next_disk(i, disks);
1481 } while (i != d0_idx);
1483 return syndrome_disks;
1486 static struct dma_async_tx_descriptor *
1487 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1489 int disks = sh->disks;
1490 struct page **blocks = to_addr_page(percpu, 0);
1492 int qd_idx = sh->qd_idx;
1493 struct dma_async_tx_descriptor *tx;
1494 struct async_submit_ctl submit;
1500 BUG_ON(sh->batch_head);
1501 if (sh->ops.target < 0)
1502 target = sh->ops.target2;
1503 else if (sh->ops.target2 < 0)
1504 target = sh->ops.target;
1506 /* we should only have one valid target */
1509 pr_debug("%s: stripe %llu block: %d\n",
1510 __func__, (unsigned long long)sh->sector, target);
1512 tgt = &sh->dev[target];
1513 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1516 atomic_inc(&sh->count);
1518 if (target == qd_idx) {
1519 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1520 blocks[count] = NULL; /* regenerating p is not necessary */
1521 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1522 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1523 ops_complete_compute, sh,
1524 to_addr_conv(sh, percpu, 0));
1525 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1527 /* Compute any data- or p-drive using XOR */
1529 for (i = disks; i-- ; ) {
1530 if (i == target || i == qd_idx)
1532 blocks[count++] = sh->dev[i].page;
1535 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1536 NULL, ops_complete_compute, sh,
1537 to_addr_conv(sh, percpu, 0));
1538 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1544 static struct dma_async_tx_descriptor *
1545 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1547 int i, count, disks = sh->disks;
1548 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1549 int d0_idx = raid6_d0(sh);
1550 int faila = -1, failb = -1;
1551 int target = sh->ops.target;
1552 int target2 = sh->ops.target2;
1553 struct r5dev *tgt = &sh->dev[target];
1554 struct r5dev *tgt2 = &sh->dev[target2];
1555 struct dma_async_tx_descriptor *tx;
1556 struct page **blocks = to_addr_page(percpu, 0);
1557 struct async_submit_ctl submit;
1559 BUG_ON(sh->batch_head);
1560 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1561 __func__, (unsigned long long)sh->sector, target, target2);
1562 BUG_ON(target < 0 || target2 < 0);
1563 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1564 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1566 /* we need to open-code set_syndrome_sources to handle the
1567 * slot number conversion for 'faila' and 'failb'
1569 for (i = 0; i < disks ; i++)
1574 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1576 blocks[slot] = sh->dev[i].page;
1582 i = raid6_next_disk(i, disks);
1583 } while (i != d0_idx);
1585 BUG_ON(faila == failb);
1588 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1589 __func__, (unsigned long long)sh->sector, faila, failb);
1591 atomic_inc(&sh->count);
1593 if (failb == syndrome_disks+1) {
1594 /* Q disk is one of the missing disks */
1595 if (faila == syndrome_disks) {
1596 /* Missing P+Q, just recompute */
1597 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1598 ops_complete_compute, sh,
1599 to_addr_conv(sh, percpu, 0));
1600 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1601 STRIPE_SIZE, &submit);
1605 int qd_idx = sh->qd_idx;
1607 /* Missing D+Q: recompute D from P, then recompute Q */
1608 if (target == qd_idx)
1609 data_target = target2;
1611 data_target = target;
1614 for (i = disks; i-- ; ) {
1615 if (i == data_target || i == qd_idx)
1617 blocks[count++] = sh->dev[i].page;
1619 dest = sh->dev[data_target].page;
1620 init_async_submit(&submit,
1621 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1623 to_addr_conv(sh, percpu, 0));
1624 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1627 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1628 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1629 ops_complete_compute, sh,
1630 to_addr_conv(sh, percpu, 0));
1631 return async_gen_syndrome(blocks, 0, count+2,
1632 STRIPE_SIZE, &submit);
1635 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1636 ops_complete_compute, sh,
1637 to_addr_conv(sh, percpu, 0));
1638 if (failb == syndrome_disks) {
1639 /* We're missing D+P. */
1640 return async_raid6_datap_recov(syndrome_disks+2,
1644 /* We're missing D+D. */
1645 return async_raid6_2data_recov(syndrome_disks+2,
1646 STRIPE_SIZE, faila, failb,
1652 static void ops_complete_prexor(void *stripe_head_ref)
1654 struct stripe_head *sh = stripe_head_ref;
1656 pr_debug("%s: stripe %llu\n", __func__,
1657 (unsigned long long)sh->sector);
1659 if (r5c_is_writeback(sh->raid_conf->log))
1661 * raid5-cache write back uses orig_page during prexor.
1662 * After prexor, it is time to free orig_page
1664 r5c_release_extra_page(sh);
1667 static struct dma_async_tx_descriptor *
1668 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1669 struct dma_async_tx_descriptor *tx)
1671 int disks = sh->disks;
1672 struct page **xor_srcs = to_addr_page(percpu, 0);
1673 int count = 0, pd_idx = sh->pd_idx, i;
1674 struct async_submit_ctl submit;
1676 /* existing parity data subtracted */
1677 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1679 BUG_ON(sh->batch_head);
1680 pr_debug("%s: stripe %llu\n", __func__,
1681 (unsigned long long)sh->sector);
1683 for (i = disks; i--; ) {
1684 struct r5dev *dev = &sh->dev[i];
1685 /* Only process blocks that are known to be uptodate */
1686 if (test_bit(R5_InJournal, &dev->flags))
1687 xor_srcs[count++] = dev->orig_page;
1688 else if (test_bit(R5_Wantdrain, &dev->flags))
1689 xor_srcs[count++] = dev->page;
1692 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1693 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1694 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1699 static struct dma_async_tx_descriptor *
1700 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1701 struct dma_async_tx_descriptor *tx)
1703 struct page **blocks = to_addr_page(percpu, 0);
1705 struct async_submit_ctl submit;
1707 pr_debug("%s: stripe %llu\n", __func__,
1708 (unsigned long long)sh->sector);
1710 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1712 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1713 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1714 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1719 static struct dma_async_tx_descriptor *
1720 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1722 struct r5conf *conf = sh->raid_conf;
1723 int disks = sh->disks;
1725 struct stripe_head *head_sh = sh;
1727 pr_debug("%s: stripe %llu\n", __func__,
1728 (unsigned long long)sh->sector);
1730 for (i = disks; i--; ) {
1735 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1741 * clear R5_InJournal, so when rewriting a page in
1742 * journal, it is not skipped by r5l_log_stripe()
1744 clear_bit(R5_InJournal, &dev->flags);
1745 spin_lock_irq(&sh->stripe_lock);
1746 chosen = dev->towrite;
1747 dev->towrite = NULL;
1748 sh->overwrite_disks = 0;
1749 BUG_ON(dev->written);
1750 wbi = dev->written = chosen;
1751 spin_unlock_irq(&sh->stripe_lock);
1752 WARN_ON(dev->page != dev->orig_page);
1754 while (wbi && wbi->bi_iter.bi_sector <
1755 dev->sector + STRIPE_SECTORS) {
1756 if (wbi->bi_opf & REQ_FUA)
1757 set_bit(R5_WantFUA, &dev->flags);
1758 if (wbi->bi_opf & REQ_SYNC)
1759 set_bit(R5_SyncIO, &dev->flags);
1760 if (bio_op(wbi) == REQ_OP_DISCARD)
1761 set_bit(R5_Discard, &dev->flags);
1763 tx = async_copy_data(1, wbi, &dev->page,
1764 dev->sector, tx, sh,
1765 r5c_is_writeback(conf->log));
1766 if (dev->page != dev->orig_page &&
1767 !r5c_is_writeback(conf->log)) {
1768 set_bit(R5_SkipCopy, &dev->flags);
1769 clear_bit(R5_UPTODATE, &dev->flags);
1770 clear_bit(R5_OVERWRITE, &dev->flags);
1773 wbi = r5_next_bio(wbi, dev->sector);
1776 if (head_sh->batch_head) {
1777 sh = list_first_entry(&sh->batch_list,
1790 static void ops_complete_reconstruct(void *stripe_head_ref)
1792 struct stripe_head *sh = stripe_head_ref;
1793 int disks = sh->disks;
1794 int pd_idx = sh->pd_idx;
1795 int qd_idx = sh->qd_idx;
1797 bool fua = false, sync = false, discard = false;
1799 pr_debug("%s: stripe %llu\n", __func__,
1800 (unsigned long long)sh->sector);
1802 for (i = disks; i--; ) {
1803 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1804 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1805 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1808 for (i = disks; i--; ) {
1809 struct r5dev *dev = &sh->dev[i];
1811 if (dev->written || i == pd_idx || i == qd_idx) {
1812 if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1813 set_bit(R5_UPTODATE, &dev->flags);
1814 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1815 set_bit(R5_Expanded, &dev->flags);
1818 set_bit(R5_WantFUA, &dev->flags);
1820 set_bit(R5_SyncIO, &dev->flags);
1824 if (sh->reconstruct_state == reconstruct_state_drain_run)
1825 sh->reconstruct_state = reconstruct_state_drain_result;
1826 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1827 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1829 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1830 sh->reconstruct_state = reconstruct_state_result;
1833 set_bit(STRIPE_HANDLE, &sh->state);
1834 raid5_release_stripe(sh);
1838 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1839 struct dma_async_tx_descriptor *tx)
1841 int disks = sh->disks;
1842 struct page **xor_srcs;
1843 struct async_submit_ctl submit;
1844 int count, pd_idx = sh->pd_idx, i;
1845 struct page *xor_dest;
1847 unsigned long flags;
1849 struct stripe_head *head_sh = sh;
1852 pr_debug("%s: stripe %llu\n", __func__,
1853 (unsigned long long)sh->sector);
1855 for (i = 0; i < sh->disks; i++) {
1858 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1861 if (i >= sh->disks) {
1862 atomic_inc(&sh->count);
1863 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1864 ops_complete_reconstruct(sh);
1869 xor_srcs = to_addr_page(percpu, j);
1870 /* check if prexor is active which means only process blocks
1871 * that are part of a read-modify-write (written)
1873 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1875 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1876 for (i = disks; i--; ) {
1877 struct r5dev *dev = &sh->dev[i];
1878 if (head_sh->dev[i].written ||
1879 test_bit(R5_InJournal, &head_sh->dev[i].flags))
1880 xor_srcs[count++] = dev->page;
1883 xor_dest = sh->dev[pd_idx].page;
1884 for (i = disks; i--; ) {
1885 struct r5dev *dev = &sh->dev[i];
1887 xor_srcs[count++] = dev->page;
1891 /* 1/ if we prexor'd then the dest is reused as a source
1892 * 2/ if we did not prexor then we are redoing the parity
1893 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1894 * for the synchronous xor case
1896 last_stripe = !head_sh->batch_head ||
1897 list_first_entry(&sh->batch_list,
1898 struct stripe_head, batch_list) == head_sh;
1900 flags = ASYNC_TX_ACK |
1901 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1903 atomic_inc(&head_sh->count);
1904 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1905 to_addr_conv(sh, percpu, j));
1907 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1908 init_async_submit(&submit, flags, tx, NULL, NULL,
1909 to_addr_conv(sh, percpu, j));
1912 if (unlikely(count == 1))
1913 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1915 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1918 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1925 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1926 struct dma_async_tx_descriptor *tx)
1928 struct async_submit_ctl submit;
1929 struct page **blocks;
1930 int count, i, j = 0;
1931 struct stripe_head *head_sh = sh;
1934 unsigned long txflags;
1936 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1938 for (i = 0; i < sh->disks; i++) {
1939 if (sh->pd_idx == i || sh->qd_idx == i)
1941 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1944 if (i >= sh->disks) {
1945 atomic_inc(&sh->count);
1946 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1947 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1948 ops_complete_reconstruct(sh);
1953 blocks = to_addr_page(percpu, j);
1955 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1956 synflags = SYNDROME_SRC_WRITTEN;
1957 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1959 synflags = SYNDROME_SRC_ALL;
1960 txflags = ASYNC_TX_ACK;
1963 count = set_syndrome_sources(blocks, sh, synflags);
1964 last_stripe = !head_sh->batch_head ||
1965 list_first_entry(&sh->batch_list,
1966 struct stripe_head, batch_list) == head_sh;
1969 atomic_inc(&head_sh->count);
1970 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1971 head_sh, to_addr_conv(sh, percpu, j));
1973 init_async_submit(&submit, 0, tx, NULL, NULL,
1974 to_addr_conv(sh, percpu, j));
1975 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1978 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1984 static void ops_complete_check(void *stripe_head_ref)
1986 struct stripe_head *sh = stripe_head_ref;
1988 pr_debug("%s: stripe %llu\n", __func__,
1989 (unsigned long long)sh->sector);
1991 sh->check_state = check_state_check_result;
1992 set_bit(STRIPE_HANDLE, &sh->state);
1993 raid5_release_stripe(sh);
1996 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1998 int disks = sh->disks;
1999 int pd_idx = sh->pd_idx;
2000 int qd_idx = sh->qd_idx;
2001 struct page *xor_dest;
2002 struct page **xor_srcs = to_addr_page(percpu, 0);
2003 struct dma_async_tx_descriptor *tx;
2004 struct async_submit_ctl submit;
2008 pr_debug("%s: stripe %llu\n", __func__,
2009 (unsigned long long)sh->sector);
2011 BUG_ON(sh->batch_head);
2013 xor_dest = sh->dev[pd_idx].page;
2014 xor_srcs[count++] = xor_dest;
2015 for (i = disks; i--; ) {
2016 if (i == pd_idx || i == qd_idx)
2018 xor_srcs[count++] = sh->dev[i].page;
2021 init_async_submit(&submit, 0, NULL, NULL, NULL,
2022 to_addr_conv(sh, percpu, 0));
2023 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2024 &sh->ops.zero_sum_result, &submit);
2026 atomic_inc(&sh->count);
2027 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2028 tx = async_trigger_callback(&submit);
2031 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2033 struct page **srcs = to_addr_page(percpu, 0);
2034 struct async_submit_ctl submit;
2037 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2038 (unsigned long long)sh->sector, checkp);
2040 BUG_ON(sh->batch_head);
2041 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2045 atomic_inc(&sh->count);
2046 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2047 sh, to_addr_conv(sh, percpu, 0));
2048 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2049 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2052 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2054 int overlap_clear = 0, i, disks = sh->disks;
2055 struct dma_async_tx_descriptor *tx = NULL;
2056 struct r5conf *conf = sh->raid_conf;
2057 int level = conf->level;
2058 struct raid5_percpu *percpu;
2062 percpu = per_cpu_ptr(conf->percpu, cpu);
2063 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2064 ops_run_biofill(sh);
2068 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2070 tx = ops_run_compute5(sh, percpu);
2072 if (sh->ops.target2 < 0 || sh->ops.target < 0)
2073 tx = ops_run_compute6_1(sh, percpu);
2075 tx = ops_run_compute6_2(sh, percpu);
2077 /* terminate the chain if reconstruct is not set to be run */
2078 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2082 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2084 tx = ops_run_prexor5(sh, percpu, tx);
2086 tx = ops_run_prexor6(sh, percpu, tx);
2089 if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2090 tx = ops_run_partial_parity(sh, percpu, tx);
2092 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2093 tx = ops_run_biodrain(sh, tx);
2097 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2099 ops_run_reconstruct5(sh, percpu, tx);
2101 ops_run_reconstruct6(sh, percpu, tx);
2104 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2105 if (sh->check_state == check_state_run)
2106 ops_run_check_p(sh, percpu);
2107 else if (sh->check_state == check_state_run_q)
2108 ops_run_check_pq(sh, percpu, 0);
2109 else if (sh->check_state == check_state_run_pq)
2110 ops_run_check_pq(sh, percpu, 1);
2115 if (overlap_clear && !sh->batch_head)
2116 for (i = disks; i--; ) {
2117 struct r5dev *dev = &sh->dev[i];
2118 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2119 wake_up(&sh->raid_conf->wait_for_overlap);
2124 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2127 __free_page(sh->ppl_page);
2128 kmem_cache_free(sc, sh);
2131 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2132 int disks, struct r5conf *conf)
2134 struct stripe_head *sh;
2137 sh = kmem_cache_zalloc(sc, gfp);
2139 spin_lock_init(&sh->stripe_lock);
2140 spin_lock_init(&sh->batch_lock);
2141 INIT_LIST_HEAD(&sh->batch_list);
2142 INIT_LIST_HEAD(&sh->lru);
2143 INIT_LIST_HEAD(&sh->r5c);
2144 INIT_LIST_HEAD(&sh->log_list);
2145 atomic_set(&sh->count, 1);
2146 sh->raid_conf = conf;
2147 sh->log_start = MaxSector;
2148 for (i = 0; i < disks; i++) {
2149 struct r5dev *dev = &sh->dev[i];
2151 bio_init(&dev->req, &dev->vec, 1);
2152 bio_init(&dev->rreq, &dev->rvec, 1);
2155 if (raid5_has_ppl(conf)) {
2156 sh->ppl_page = alloc_page(gfp);
2157 if (!sh->ppl_page) {
2158 free_stripe(sc, sh);
2165 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2167 struct stripe_head *sh;
2169 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2173 if (grow_buffers(sh, gfp)) {
2175 free_stripe(conf->slab_cache, sh);
2178 sh->hash_lock_index =
2179 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2180 /* we just created an active stripe so... */
2181 atomic_inc(&conf->active_stripes);
2183 raid5_release_stripe(sh);
2184 conf->max_nr_stripes++;
2188 static int grow_stripes(struct r5conf *conf, int num)
2190 struct kmem_cache *sc;
2191 size_t namelen = sizeof(conf->cache_name[0]);
2192 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2194 if (conf->mddev->gendisk)
2195 snprintf(conf->cache_name[0], namelen,
2196 "raid%d-%s", conf->level, mdname(conf->mddev));
2198 snprintf(conf->cache_name[0], namelen,
2199 "raid%d-%p", conf->level, conf->mddev);
2200 snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2202 conf->active_name = 0;
2203 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2204 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2208 conf->slab_cache = sc;
2209 conf->pool_size = devs;
2211 if (!grow_one_stripe(conf, GFP_KERNEL))
2218 * scribble_len - return the required size of the scribble region
2219 * @num - total number of disks in the array
2221 * The size must be enough to contain:
2222 * 1/ a struct page pointer for each device in the array +2
2223 * 2/ room to convert each entry in (1) to its corresponding dma
2224 * (dma_map_page()) or page (page_address()) address.
2226 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2227 * calculate over all devices (not just the data blocks), using zeros in place
2228 * of the P and Q blocks.
2230 static int scribble_alloc(struct raid5_percpu *percpu,
2231 int num, int cnt, gfp_t flags)
2234 sizeof(struct page *) * (num+2) +
2235 sizeof(addr_conv_t) * (num+2);
2238 scribble = kvmalloc_array(cnt, obj_size, flags);
2242 kvfree(percpu->scribble);
2244 percpu->scribble = scribble;
2245 percpu->scribble_obj_size = obj_size;
2249 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2255 * Never shrink. And mddev_suspend() could deadlock if this is called
2256 * from raid5d. In that case, scribble_disks and scribble_sectors
2257 * should equal to new_disks and new_sectors
2259 if (conf->scribble_disks >= new_disks &&
2260 conf->scribble_sectors >= new_sectors)
2262 mddev_suspend(conf->mddev);
2265 for_each_present_cpu(cpu) {
2266 struct raid5_percpu *percpu;
2268 percpu = per_cpu_ptr(conf->percpu, cpu);
2269 err = scribble_alloc(percpu, new_disks,
2270 new_sectors / STRIPE_SECTORS,
2277 mddev_resume(conf->mddev);
2279 conf->scribble_disks = new_disks;
2280 conf->scribble_sectors = new_sectors;
2285 static int resize_stripes(struct r5conf *conf, int newsize)
2287 /* Make all the stripes able to hold 'newsize' devices.
2288 * New slots in each stripe get 'page' set to a new page.
2290 * This happens in stages:
2291 * 1/ create a new kmem_cache and allocate the required number of
2293 * 2/ gather all the old stripe_heads and transfer the pages across
2294 * to the new stripe_heads. This will have the side effect of
2295 * freezing the array as once all stripe_heads have been collected,
2296 * no IO will be possible. Old stripe heads are freed once their
2297 * pages have been transferred over, and the old kmem_cache is
2298 * freed when all stripes are done.
2299 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2300 * we simple return a failure status - no need to clean anything up.
2301 * 4/ allocate new pages for the new slots in the new stripe_heads.
2302 * If this fails, we don't bother trying the shrink the
2303 * stripe_heads down again, we just leave them as they are.
2304 * As each stripe_head is processed the new one is released into
2307 * Once step2 is started, we cannot afford to wait for a write,
2308 * so we use GFP_NOIO allocations.
2310 struct stripe_head *osh, *nsh;
2311 LIST_HEAD(newstripes);
2312 struct disk_info *ndisks;
2314 struct kmem_cache *sc;
2318 md_allow_write(conf->mddev);
2321 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2322 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2327 /* Need to ensure auto-resizing doesn't interfere */
2328 mutex_lock(&conf->cache_size_mutex);
2330 for (i = conf->max_nr_stripes; i; i--) {
2331 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2335 list_add(&nsh->lru, &newstripes);
2338 /* didn't get enough, give up */
2339 while (!list_empty(&newstripes)) {
2340 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2341 list_del(&nsh->lru);
2342 free_stripe(sc, nsh);
2344 kmem_cache_destroy(sc);
2345 mutex_unlock(&conf->cache_size_mutex);
2348 /* Step 2 - Must use GFP_NOIO now.
2349 * OK, we have enough stripes, start collecting inactive
2350 * stripes and copying them over
2354 list_for_each_entry(nsh, &newstripes, lru) {
2355 lock_device_hash_lock(conf, hash);
2356 wait_event_cmd(conf->wait_for_stripe,
2357 !list_empty(conf->inactive_list + hash),
2358 unlock_device_hash_lock(conf, hash),
2359 lock_device_hash_lock(conf, hash));
2360 osh = get_free_stripe(conf, hash);
2361 unlock_device_hash_lock(conf, hash);
2363 for(i=0; i<conf->pool_size; i++) {
2364 nsh->dev[i].page = osh->dev[i].page;
2365 nsh->dev[i].orig_page = osh->dev[i].page;
2367 nsh->hash_lock_index = hash;
2368 free_stripe(conf->slab_cache, osh);
2370 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2371 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2376 kmem_cache_destroy(conf->slab_cache);
2379 * At this point, we are holding all the stripes so the array
2380 * is completely stalled, so now is a good time to resize
2381 * conf->disks and the scribble region
2383 ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2385 for (i = 0; i < conf->pool_size; i++)
2386 ndisks[i] = conf->disks[i];
2388 for (i = conf->pool_size; i < newsize; i++) {
2389 ndisks[i].extra_page = alloc_page(GFP_NOIO);
2390 if (!ndisks[i].extra_page)
2395 for (i = conf->pool_size; i < newsize; i++)
2396 if (ndisks[i].extra_page)
2397 put_page(ndisks[i].extra_page);
2401 conf->disks = ndisks;
2406 mutex_unlock(&conf->cache_size_mutex);
2408 conf->slab_cache = sc;
2409 conf->active_name = 1-conf->active_name;
2411 /* Step 4, return new stripes to service */
2412 while(!list_empty(&newstripes)) {
2413 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2414 list_del_init(&nsh->lru);
2416 for (i=conf->raid_disks; i < newsize; i++)
2417 if (nsh->dev[i].page == NULL) {
2418 struct page *p = alloc_page(GFP_NOIO);
2419 nsh->dev[i].page = p;
2420 nsh->dev[i].orig_page = p;
2424 raid5_release_stripe(nsh);
2426 /* critical section pass, GFP_NOIO no longer needed */
2429 conf->pool_size = newsize;
2433 static int drop_one_stripe(struct r5conf *conf)
2435 struct stripe_head *sh;
2436 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2438 spin_lock_irq(conf->hash_locks + hash);
2439 sh = get_free_stripe(conf, hash);
2440 spin_unlock_irq(conf->hash_locks + hash);
2443 BUG_ON(atomic_read(&sh->count));
2445 free_stripe(conf->slab_cache, sh);
2446 atomic_dec(&conf->active_stripes);
2447 conf->max_nr_stripes--;
2451 static void shrink_stripes(struct r5conf *conf)
2453 while (conf->max_nr_stripes &&
2454 drop_one_stripe(conf))
2457 kmem_cache_destroy(conf->slab_cache);
2458 conf->slab_cache = NULL;
2461 static void raid5_end_read_request(struct bio * bi)
2463 struct stripe_head *sh = bi->bi_private;
2464 struct r5conf *conf = sh->raid_conf;
2465 int disks = sh->disks, i;
2466 char b[BDEVNAME_SIZE];
2467 struct md_rdev *rdev = NULL;
2470 for (i=0 ; i<disks; i++)
2471 if (bi == &sh->dev[i].req)
2474 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2475 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2482 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2483 /* If replacement finished while this request was outstanding,
2484 * 'replacement' might be NULL already.
2485 * In that case it moved down to 'rdev'.
2486 * rdev is not removed until all requests are finished.
2488 rdev = conf->disks[i].replacement;
2490 rdev = conf->disks[i].rdev;
2492 if (use_new_offset(conf, sh))
2493 s = sh->sector + rdev->new_data_offset;
2495 s = sh->sector + rdev->data_offset;
2496 if (!bi->bi_status) {
2497 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2498 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2499 /* Note that this cannot happen on a
2500 * replacement device. We just fail those on
2503 pr_info_ratelimited(
2504 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2505 mdname(conf->mddev), STRIPE_SECTORS,
2506 (unsigned long long)s,
2507 bdevname(rdev->bdev, b));
2508 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2509 clear_bit(R5_ReadError, &sh->dev[i].flags);
2510 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2511 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2512 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2514 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2516 * end read for a page in journal, this
2517 * must be preparing for prexor in rmw
2519 set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2521 if (atomic_read(&rdev->read_errors))
2522 atomic_set(&rdev->read_errors, 0);
2524 const char *bdn = bdevname(rdev->bdev, b);
2528 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2529 atomic_inc(&rdev->read_errors);
2530 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2531 pr_warn_ratelimited(
2532 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2533 mdname(conf->mddev),
2534 (unsigned long long)s,
2536 else if (conf->mddev->degraded >= conf->max_degraded) {
2538 pr_warn_ratelimited(
2539 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2540 mdname(conf->mddev),
2541 (unsigned long long)s,
2543 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2546 pr_warn_ratelimited(
2547 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2548 mdname(conf->mddev),
2549 (unsigned long long)s,
2551 } else if (atomic_read(&rdev->read_errors)
2552 > conf->max_nr_stripes)
2553 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2554 mdname(conf->mddev), bdn);
2557 if (set_bad && test_bit(In_sync, &rdev->flags)
2558 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2561 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2562 set_bit(R5_ReadError, &sh->dev[i].flags);
2563 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2565 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2567 clear_bit(R5_ReadError, &sh->dev[i].flags);
2568 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2570 && test_bit(In_sync, &rdev->flags)
2571 && rdev_set_badblocks(
2572 rdev, sh->sector, STRIPE_SECTORS, 0)))
2573 md_error(conf->mddev, rdev);
2576 rdev_dec_pending(rdev, conf->mddev);
2578 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2579 set_bit(STRIPE_HANDLE, &sh->state);
2580 raid5_release_stripe(sh);
2583 static void raid5_end_write_request(struct bio *bi)
2585 struct stripe_head *sh = bi->bi_private;
2586 struct r5conf *conf = sh->raid_conf;
2587 int disks = sh->disks, i;
2588 struct md_rdev *uninitialized_var(rdev);
2591 int replacement = 0;
2593 for (i = 0 ; i < disks; i++) {
2594 if (bi == &sh->dev[i].req) {
2595 rdev = conf->disks[i].rdev;
2598 if (bi == &sh->dev[i].rreq) {
2599 rdev = conf->disks[i].replacement;
2603 /* rdev was removed and 'replacement'
2604 * replaced it. rdev is not removed
2605 * until all requests are finished.
2607 rdev = conf->disks[i].rdev;
2611 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2612 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2622 md_error(conf->mddev, rdev);
2623 else if (is_badblock(rdev, sh->sector,
2625 &first_bad, &bad_sectors))
2626 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2628 if (bi->bi_status) {
2629 set_bit(STRIPE_DEGRADED, &sh->state);
2630 set_bit(WriteErrorSeen, &rdev->flags);
2631 set_bit(R5_WriteError, &sh->dev[i].flags);
2632 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2633 set_bit(MD_RECOVERY_NEEDED,
2634 &rdev->mddev->recovery);
2635 } else if (is_badblock(rdev, sh->sector,
2637 &first_bad, &bad_sectors)) {
2638 set_bit(R5_MadeGood, &sh->dev[i].flags);
2639 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2640 /* That was a successful write so make
2641 * sure it looks like we already did
2644 set_bit(R5_ReWrite, &sh->dev[i].flags);
2647 rdev_dec_pending(rdev, conf->mddev);
2649 if (sh->batch_head && bi->bi_status && !replacement)
2650 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2653 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2654 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2655 set_bit(STRIPE_HANDLE, &sh->state);
2656 raid5_release_stripe(sh);
2658 if (sh->batch_head && sh != sh->batch_head)
2659 raid5_release_stripe(sh->batch_head);
2662 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2664 char b[BDEVNAME_SIZE];
2665 struct r5conf *conf = mddev->private;
2666 unsigned long flags;
2667 pr_debug("raid456: error called\n");
2669 spin_lock_irqsave(&conf->device_lock, flags);
2671 if (test_bit(In_sync, &rdev->flags) &&
2672 mddev->degraded == conf->max_degraded) {
2674 * Don't allow to achieve failed state
2675 * Don't try to recover this device
2677 conf->recovery_disabled = mddev->recovery_disabled;
2678 spin_unlock_irqrestore(&conf->device_lock, flags);
2682 set_bit(Faulty, &rdev->flags);
2683 clear_bit(In_sync, &rdev->flags);
2684 mddev->degraded = raid5_calc_degraded(conf);
2685 spin_unlock_irqrestore(&conf->device_lock, flags);
2686 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2688 set_bit(Blocked, &rdev->flags);
2689 set_mask_bits(&mddev->sb_flags, 0,
2690 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2691 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2692 "md/raid:%s: Operation continuing on %d devices.\n",
2694 bdevname(rdev->bdev, b),
2696 conf->raid_disks - mddev->degraded);
2697 r5c_update_on_rdev_error(mddev, rdev);
2701 * Input: a 'big' sector number,
2702 * Output: index of the data and parity disk, and the sector # in them.
2704 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2705 int previous, int *dd_idx,
2706 struct stripe_head *sh)
2708 sector_t stripe, stripe2;
2709 sector_t chunk_number;
2710 unsigned int chunk_offset;
2713 sector_t new_sector;
2714 int algorithm = previous ? conf->prev_algo
2716 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2717 : conf->chunk_sectors;
2718 int raid_disks = previous ? conf->previous_raid_disks
2720 int data_disks = raid_disks - conf->max_degraded;
2722 /* First compute the information on this sector */
2725 * Compute the chunk number and the sector offset inside the chunk
2727 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2728 chunk_number = r_sector;
2731 * Compute the stripe number
2733 stripe = chunk_number;
2734 *dd_idx = sector_div(stripe, data_disks);
2737 * Select the parity disk based on the user selected algorithm.
2739 pd_idx = qd_idx = -1;
2740 switch(conf->level) {
2742 pd_idx = data_disks;
2745 switch (algorithm) {
2746 case ALGORITHM_LEFT_ASYMMETRIC:
2747 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2748 if (*dd_idx >= pd_idx)
2751 case ALGORITHM_RIGHT_ASYMMETRIC:
2752 pd_idx = sector_div(stripe2, raid_disks);
2753 if (*dd_idx >= pd_idx)
2756 case ALGORITHM_LEFT_SYMMETRIC:
2757 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2758 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2760 case ALGORITHM_RIGHT_SYMMETRIC:
2761 pd_idx = sector_div(stripe2, raid_disks);
2762 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2764 case ALGORITHM_PARITY_0:
2768 case ALGORITHM_PARITY_N:
2769 pd_idx = data_disks;
2777 switch (algorithm) {
2778 case ALGORITHM_LEFT_ASYMMETRIC:
2779 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2780 qd_idx = pd_idx + 1;
2781 if (pd_idx == raid_disks-1) {
2782 (*dd_idx)++; /* Q D D D P */
2784 } else if (*dd_idx >= pd_idx)
2785 (*dd_idx) += 2; /* D D P Q D */
2787 case ALGORITHM_RIGHT_ASYMMETRIC:
2788 pd_idx = sector_div(stripe2, raid_disks);
2789 qd_idx = pd_idx + 1;
2790 if (pd_idx == raid_disks-1) {
2791 (*dd_idx)++; /* Q D D D P */
2793 } else if (*dd_idx >= pd_idx)
2794 (*dd_idx) += 2; /* D D P Q D */
2796 case ALGORITHM_LEFT_SYMMETRIC:
2797 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2798 qd_idx = (pd_idx + 1) % raid_disks;
2799 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2801 case ALGORITHM_RIGHT_SYMMETRIC:
2802 pd_idx = sector_div(stripe2, raid_disks);
2803 qd_idx = (pd_idx + 1) % raid_disks;
2804 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2807 case ALGORITHM_PARITY_0:
2812 case ALGORITHM_PARITY_N:
2813 pd_idx = data_disks;
2814 qd_idx = data_disks + 1;
2817 case ALGORITHM_ROTATING_ZERO_RESTART:
2818 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2819 * of blocks for computing Q is different.
2821 pd_idx = sector_div(stripe2, raid_disks);
2822 qd_idx = pd_idx + 1;
2823 if (pd_idx == raid_disks-1) {
2824 (*dd_idx)++; /* Q D D D P */
2826 } else if (*dd_idx >= pd_idx)
2827 (*dd_idx) += 2; /* D D P Q D */
2831 case ALGORITHM_ROTATING_N_RESTART:
2832 /* Same a left_asymmetric, by first stripe is
2833 * D D D P Q rather than
2837 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2838 qd_idx = pd_idx + 1;
2839 if (pd_idx == raid_disks-1) {
2840 (*dd_idx)++; /* Q D D D P */
2842 } else if (*dd_idx >= pd_idx)
2843 (*dd_idx) += 2; /* D D P Q D */
2847 case ALGORITHM_ROTATING_N_CONTINUE:
2848 /* Same as left_symmetric but Q is before P */
2849 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2850 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2851 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2855 case ALGORITHM_LEFT_ASYMMETRIC_6:
2856 /* RAID5 left_asymmetric, with Q on last device */
2857 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2858 if (*dd_idx >= pd_idx)
2860 qd_idx = raid_disks - 1;
2863 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2864 pd_idx = sector_div(stripe2, raid_disks-1);
2865 if (*dd_idx >= pd_idx)
2867 qd_idx = raid_disks - 1;
2870 case ALGORITHM_LEFT_SYMMETRIC_6:
2871 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2872 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2873 qd_idx = raid_disks - 1;
2876 case ALGORITHM_RIGHT_SYMMETRIC_6:
2877 pd_idx = sector_div(stripe2, raid_disks-1);
2878 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2879 qd_idx = raid_disks - 1;
2882 case ALGORITHM_PARITY_0_6:
2885 qd_idx = raid_disks - 1;
2895 sh->pd_idx = pd_idx;
2896 sh->qd_idx = qd_idx;
2897 sh->ddf_layout = ddf_layout;
2900 * Finally, compute the new sector number
2902 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2906 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2908 struct r5conf *conf = sh->raid_conf;
2909 int raid_disks = sh->disks;
2910 int data_disks = raid_disks - conf->max_degraded;
2911 sector_t new_sector = sh->sector, check;
2912 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2913 : conf->chunk_sectors;
2914 int algorithm = previous ? conf->prev_algo
2918 sector_t chunk_number;
2919 int dummy1, dd_idx = i;
2921 struct stripe_head sh2;
2923 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2924 stripe = new_sector;
2926 if (i == sh->pd_idx)
2928 switch(conf->level) {
2931 switch (algorithm) {
2932 case ALGORITHM_LEFT_ASYMMETRIC:
2933 case ALGORITHM_RIGHT_ASYMMETRIC:
2937 case ALGORITHM_LEFT_SYMMETRIC:
2938 case ALGORITHM_RIGHT_SYMMETRIC:
2941 i -= (sh->pd_idx + 1);
2943 case ALGORITHM_PARITY_0:
2946 case ALGORITHM_PARITY_N:
2953 if (i == sh->qd_idx)
2954 return 0; /* It is the Q disk */
2955 switch (algorithm) {
2956 case ALGORITHM_LEFT_ASYMMETRIC:
2957 case ALGORITHM_RIGHT_ASYMMETRIC:
2958 case ALGORITHM_ROTATING_ZERO_RESTART:
2959 case ALGORITHM_ROTATING_N_RESTART:
2960 if (sh->pd_idx == raid_disks-1)
2961 i--; /* Q D D D P */
2962 else if (i > sh->pd_idx)
2963 i -= 2; /* D D P Q D */
2965 case ALGORITHM_LEFT_SYMMETRIC:
2966 case ALGORITHM_RIGHT_SYMMETRIC:
2967 if (sh->pd_idx == raid_disks-1)
2968 i--; /* Q D D D P */
2973 i -= (sh->pd_idx + 2);
2976 case ALGORITHM_PARITY_0:
2979 case ALGORITHM_PARITY_N:
2981 case ALGORITHM_ROTATING_N_CONTINUE:
2982 /* Like left_symmetric, but P is before Q */
2983 if (sh->pd_idx == 0)
2984 i--; /* P D D D Q */
2989 i -= (sh->pd_idx + 1);
2992 case ALGORITHM_LEFT_ASYMMETRIC_6:
2993 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2997 case ALGORITHM_LEFT_SYMMETRIC_6:
2998 case ALGORITHM_RIGHT_SYMMETRIC_6:
3000 i += data_disks + 1;
3001 i -= (sh->pd_idx + 1);
3003 case ALGORITHM_PARITY_0_6:
3012 chunk_number = stripe * data_disks + i;
3013 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3015 check = raid5_compute_sector(conf, r_sector,
3016 previous, &dummy1, &sh2);
3017 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3018 || sh2.qd_idx != sh->qd_idx) {
3019 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3020 mdname(conf->mddev));
3027 * There are cases where we want handle_stripe_dirtying() and
3028 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3030 * This function checks whether we want to delay the towrite. Specifically,
3031 * we delay the towrite when:
3033 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3034 * stripe has data in journal (for other devices).
3036 * In this case, when reading data for the non-overwrite dev, it is
3037 * necessary to handle complex rmw of write back cache (prexor with
3038 * orig_page, and xor with page). To keep read path simple, we would
3039 * like to flush data in journal to RAID disks first, so complex rmw
3040 * is handled in the write patch (handle_stripe_dirtying).
3042 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3044 * It is important to be able to flush all stripes in raid5-cache.
3045 * Therefore, we need reserve some space on the journal device for
3046 * these flushes. If flush operation includes pending writes to the
3047 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3048 * for the flush out. If we exclude these pending writes from flush
3049 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3050 * Therefore, excluding pending writes in these cases enables more
3051 * efficient use of the journal device.
3053 * Note: To make sure the stripe makes progress, we only delay
3054 * towrite for stripes with data already in journal (injournal > 0).
3055 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3056 * no_space_stripes list.
3058 * 3. during journal failure
3059 * In journal failure, we try to flush all cached data to raid disks
3060 * based on data in stripe cache. The array is read-only to upper
3061 * layers, so we would skip all pending writes.
3064 static inline bool delay_towrite(struct r5conf *conf,
3066 struct stripe_head_state *s)
3069 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3070 !test_bit(R5_Insync, &dev->flags) && s->injournal)
3073 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3077 if (s->log_failed && s->injournal)
3083 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3084 int rcw, int expand)
3086 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3087 struct r5conf *conf = sh->raid_conf;
3088 int level = conf->level;
3092 * In some cases, handle_stripe_dirtying initially decided to
3093 * run rmw and allocates extra page for prexor. However, rcw is
3094 * cheaper later on. We need to free the extra page now,
3095 * because we won't be able to do that in ops_complete_prexor().
3097 r5c_release_extra_page(sh);
3099 for (i = disks; i--; ) {
3100 struct r5dev *dev = &sh->dev[i];
3102 if (dev->towrite && !delay_towrite(conf, dev, s)) {
3103 set_bit(R5_LOCKED, &dev->flags);
3104 set_bit(R5_Wantdrain, &dev->flags);
3106 clear_bit(R5_UPTODATE, &dev->flags);
3108 } else if (test_bit(R5_InJournal, &dev->flags)) {
3109 set_bit(R5_LOCKED, &dev->flags);
3113 /* if we are not expanding this is a proper write request, and
3114 * there will be bios with new data to be drained into the
3119 /* False alarm, nothing to do */
3121 sh->reconstruct_state = reconstruct_state_drain_run;
3122 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3124 sh->reconstruct_state = reconstruct_state_run;
3126 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3128 if (s->locked + conf->max_degraded == disks)
3129 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3130 atomic_inc(&conf->pending_full_writes);
3132 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3133 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3134 BUG_ON(level == 6 &&
3135 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3136 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3138 for (i = disks; i--; ) {
3139 struct r5dev *dev = &sh->dev[i];
3140 if (i == pd_idx || i == qd_idx)
3144 (test_bit(R5_UPTODATE, &dev->flags) ||
3145 test_bit(R5_Wantcompute, &dev->flags))) {
3146 set_bit(R5_Wantdrain, &dev->flags);
3147 set_bit(R5_LOCKED, &dev->flags);
3148 clear_bit(R5_UPTODATE, &dev->flags);
3150 } else if (test_bit(R5_InJournal, &dev->flags)) {
3151 set_bit(R5_LOCKED, &dev->flags);
3156 /* False alarm - nothing to do */
3158 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3159 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3160 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3161 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3164 /* keep the parity disk(s) locked while asynchronous operations
3167 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3168 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3172 int qd_idx = sh->qd_idx;
3173 struct r5dev *dev = &sh->dev[qd_idx];
3175 set_bit(R5_LOCKED, &dev->flags);
3176 clear_bit(R5_UPTODATE, &dev->flags);
3180 if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3181 test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3182 !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3183 test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3184 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3186 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3187 __func__, (unsigned long long)sh->sector,
3188 s->locked, s->ops_request);
3192 * Each stripe/dev can have one or more bion attached.
3193 * toread/towrite point to the first in a chain.
3194 * The bi_next chain must be in order.
3196 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3197 int forwrite, int previous)
3200 struct r5conf *conf = sh->raid_conf;
3203 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3204 (unsigned long long)bi->bi_iter.bi_sector,
3205 (unsigned long long)sh->sector);
3207 spin_lock_irq(&sh->stripe_lock);
3208 sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3209 /* Don't allow new IO added to stripes in batch list */
3213 bip = &sh->dev[dd_idx].towrite;
3217 bip = &sh->dev[dd_idx].toread;
3218 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3219 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3221 bip = & (*bip)->bi_next;
3223 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3226 if (forwrite && raid5_has_ppl(conf)) {
3228 * With PPL only writes to consecutive data chunks within a
3229 * stripe are allowed because for a single stripe_head we can
3230 * only have one PPL entry at a time, which describes one data
3231 * range. Not really an overlap, but wait_for_overlap can be
3232 * used to handle this.
3240 for (i = 0; i < sh->disks; i++) {
3241 if (i != sh->pd_idx &&
3242 (i == dd_idx || sh->dev[i].towrite)) {
3243 sector = sh->dev[i].sector;
3244 if (count == 0 || sector < first)
3252 if (first + conf->chunk_sectors * (count - 1) != last)
3256 if (!forwrite || previous)
3257 clear_bit(STRIPE_BATCH_READY, &sh->state);
3259 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3263 bio_inc_remaining(bi);
3264 md_write_inc(conf->mddev, bi);
3267 /* check if page is covered */
3268 sector_t sector = sh->dev[dd_idx].sector;
3269 for (bi=sh->dev[dd_idx].towrite;
3270 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3271 bi && bi->bi_iter.bi_sector <= sector;
3272 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3273 if (bio_end_sector(bi) >= sector)
3274 sector = bio_end_sector(bi);
3276 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3277 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3278 sh->overwrite_disks++;
3281 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3282 (unsigned long long)(*bip)->bi_iter.bi_sector,
3283 (unsigned long long)sh->sector, dd_idx);
3285 if (conf->mddev->bitmap && firstwrite) {
3286 /* Cannot hold spinlock over bitmap_startwrite,
3287 * but must ensure this isn't added to a batch until
3288 * we have added to the bitmap and set bm_seq.
3289 * So set STRIPE_BITMAP_PENDING to prevent
3291 * If multiple add_stripe_bio() calls race here they
3292 * much all set STRIPE_BITMAP_PENDING. So only the first one
3293 * to complete "bitmap_startwrite" gets to set
3294 * STRIPE_BIT_DELAY. This is important as once a stripe
3295 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3298 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3299 spin_unlock_irq(&sh->stripe_lock);
3300 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3302 spin_lock_irq(&sh->stripe_lock);
3303 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3304 if (!sh->batch_head) {
3305 sh->bm_seq = conf->seq_flush+1;
3306 set_bit(STRIPE_BIT_DELAY, &sh->state);
3309 spin_unlock_irq(&sh->stripe_lock);
3311 if (stripe_can_batch(sh))
3312 stripe_add_to_batch_list(conf, sh);
3316 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3317 spin_unlock_irq(&sh->stripe_lock);
3321 static void end_reshape(struct r5conf *conf);
3323 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3324 struct stripe_head *sh)
3326 int sectors_per_chunk =
3327 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3329 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3330 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3332 raid5_compute_sector(conf,
3333 stripe * (disks - conf->max_degraded)
3334 *sectors_per_chunk + chunk_offset,
3340 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3341 struct stripe_head_state *s, int disks)
3344 BUG_ON(sh->batch_head);
3345 for (i = disks; i--; ) {
3349 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3350 struct md_rdev *rdev;
3352 rdev = rcu_dereference(conf->disks[i].rdev);
3353 if (rdev && test_bit(In_sync, &rdev->flags) &&
3354 !test_bit(Faulty, &rdev->flags))
3355 atomic_inc(&rdev->nr_pending);
3360 if (!rdev_set_badblocks(
3364 md_error(conf->mddev, rdev);
3365 rdev_dec_pending(rdev, conf->mddev);
3368 spin_lock_irq(&sh->stripe_lock);
3369 /* fail all writes first */
3370 bi = sh->dev[i].towrite;
3371 sh->dev[i].towrite = NULL;
3372 sh->overwrite_disks = 0;
3373 spin_unlock_irq(&sh->stripe_lock);
3377 log_stripe_write_finished(sh);
3379 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3380 wake_up(&conf->wait_for_overlap);
3382 while (bi && bi->bi_iter.bi_sector <
3383 sh->dev[i].sector + STRIPE_SECTORS) {
3384 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3386 md_write_end(conf->mddev);
3391 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3392 STRIPE_SECTORS, 0, 0);
3394 /* and fail all 'written' */
3395 bi = sh->dev[i].written;
3396 sh->dev[i].written = NULL;
3397 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3398 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3399 sh->dev[i].page = sh->dev[i].orig_page;
3402 if (bi) bitmap_end = 1;
3403 while (bi && bi->bi_iter.bi_sector <
3404 sh->dev[i].sector + STRIPE_SECTORS) {
3405 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3407 md_write_end(conf->mddev);
3412 /* fail any reads if this device is non-operational and
3413 * the data has not reached the cache yet.
3415 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3416 s->failed > conf->max_degraded &&
3417 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3418 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3419 spin_lock_irq(&sh->stripe_lock);
3420 bi = sh->dev[i].toread;
3421 sh->dev[i].toread = NULL;
3422 spin_unlock_irq(&sh->stripe_lock);
3423 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3424 wake_up(&conf->wait_for_overlap);
3427 while (bi && bi->bi_iter.bi_sector <
3428 sh->dev[i].sector + STRIPE_SECTORS) {
3429 struct bio *nextbi =
3430 r5_next_bio(bi, sh->dev[i].sector);
3437 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3438 STRIPE_SECTORS, 0, 0);
3439 /* If we were in the middle of a write the parity block might
3440 * still be locked - so just clear all R5_LOCKED flags
3442 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3447 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3448 if (atomic_dec_and_test(&conf->pending_full_writes))
3449 md_wakeup_thread(conf->mddev->thread);
3453 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3454 struct stripe_head_state *s)
3459 BUG_ON(sh->batch_head);
3460 clear_bit(STRIPE_SYNCING, &sh->state);
3461 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3462 wake_up(&conf->wait_for_overlap);
3465 /* There is nothing more to do for sync/check/repair.
3466 * Don't even need to abort as that is handled elsewhere
3467 * if needed, and not always wanted e.g. if there is a known
3469 * For recover/replace we need to record a bad block on all
3470 * non-sync devices, or abort the recovery
3472 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3473 /* During recovery devices cannot be removed, so
3474 * locking and refcounting of rdevs is not needed
3477 for (i = 0; i < conf->raid_disks; i++) {
3478 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3480 && !test_bit(Faulty, &rdev->flags)
3481 && !test_bit(In_sync, &rdev->flags)
3482 && !rdev_set_badblocks(rdev, sh->sector,
3485 rdev = rcu_dereference(conf->disks[i].replacement);
3487 && !test_bit(Faulty, &rdev->flags)
3488 && !test_bit(In_sync, &rdev->flags)
3489 && !rdev_set_badblocks(rdev, sh->sector,
3495 conf->recovery_disabled =
3496 conf->mddev->recovery_disabled;
3498 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3501 static int want_replace(struct stripe_head *sh, int disk_idx)
3503 struct md_rdev *rdev;
3507 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3509 && !test_bit(Faulty, &rdev->flags)
3510 && !test_bit(In_sync, &rdev->flags)
3511 && (rdev->recovery_offset <= sh->sector
3512 || rdev->mddev->recovery_cp <= sh->sector))
3518 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3519 int disk_idx, int disks)
3521 struct r5dev *dev = &sh->dev[disk_idx];
3522 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3523 &sh->dev[s->failed_num[1]] };
3527 if (test_bit(R5_LOCKED, &dev->flags) ||
3528 test_bit(R5_UPTODATE, &dev->flags))
3529 /* No point reading this as we already have it or have
3530 * decided to get it.
3535 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3536 /* We need this block to directly satisfy a request */
3539 if (s->syncing || s->expanding ||
3540 (s->replacing && want_replace(sh, disk_idx)))
3541 /* When syncing, or expanding we read everything.
3542 * When replacing, we need the replaced block.
3546 if ((s->failed >= 1 && fdev[0]->toread) ||
3547 (s->failed >= 2 && fdev[1]->toread))
3548 /* If we want to read from a failed device, then
3549 * we need to actually read every other device.
3553 /* Sometimes neither read-modify-write nor reconstruct-write
3554 * cycles can work. In those cases we read every block we
3555 * can. Then the parity-update is certain to have enough to
3557 * This can only be a problem when we need to write something,
3558 * and some device has failed. If either of those tests
3559 * fail we need look no further.
3561 if (!s->failed || !s->to_write)
3564 if (test_bit(R5_Insync, &dev->flags) &&
3565 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3566 /* Pre-reads at not permitted until after short delay
3567 * to gather multiple requests. However if this
3568 * device is no Insync, the block could only be computed
3569 * and there is no need to delay that.
3573 for (i = 0; i < s->failed && i < 2; i++) {
3574 if (fdev[i]->towrite &&
3575 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3576 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3577 /* If we have a partial write to a failed
3578 * device, then we will need to reconstruct
3579 * the content of that device, so all other
3580 * devices must be read.
3585 /* If we are forced to do a reconstruct-write, either because
3586 * the current RAID6 implementation only supports that, or
3587 * because parity cannot be trusted and we are currently
3588 * recovering it, there is extra need to be careful.
3589 * If one of the devices that we would need to read, because
3590 * it is not being overwritten (and maybe not written at all)
3591 * is missing/faulty, then we need to read everything we can.
3593 if (sh->raid_conf->level != 6 &&
3594 sh->sector < sh->raid_conf->mddev->recovery_cp)
3595 /* reconstruct-write isn't being forced */
3597 for (i = 0; i < s->failed && i < 2; i++) {
3598 if (s->failed_num[i] != sh->pd_idx &&
3599 s->failed_num[i] != sh->qd_idx &&
3600 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3601 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3608 /* fetch_block - checks the given member device to see if its data needs
3609 * to be read or computed to satisfy a request.
3611 * Returns 1 when no more member devices need to be checked, otherwise returns
3612 * 0 to tell the loop in handle_stripe_fill to continue
3614 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3615 int disk_idx, int disks)
3617 struct r5dev *dev = &sh->dev[disk_idx];
3619 /* is the data in this block needed, and can we get it? */
3620 if (need_this_block(sh, s, disk_idx, disks)) {
3621 /* we would like to get this block, possibly by computing it,
3622 * otherwise read it if the backing disk is insync
3624 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3625 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3626 BUG_ON(sh->batch_head);
3629 * In the raid6 case if the only non-uptodate disk is P
3630 * then we already trusted P to compute the other failed
3631 * drives. It is safe to compute rather than re-read P.
3632 * In other cases we only compute blocks from failed
3633 * devices, otherwise check/repair might fail to detect
3634 * a real inconsistency.
3637 if ((s->uptodate == disks - 1) &&
3638 ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3639 (s->failed && (disk_idx == s->failed_num[0] ||
3640 disk_idx == s->failed_num[1])))) {
3641 /* have disk failed, and we're requested to fetch it;
3644 pr_debug("Computing stripe %llu block %d\n",
3645 (unsigned long long)sh->sector, disk_idx);
3646 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3647 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3648 set_bit(R5_Wantcompute, &dev->flags);
3649 sh->ops.target = disk_idx;
3650 sh->ops.target2 = -1; /* no 2nd target */
3652 /* Careful: from this point on 'uptodate' is in the eye
3653 * of raid_run_ops which services 'compute' operations
3654 * before writes. R5_Wantcompute flags a block that will
3655 * be R5_UPTODATE by the time it is needed for a
3656 * subsequent operation.
3660 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3661 /* Computing 2-failure is *very* expensive; only
3662 * do it if failed >= 2
3665 for (other = disks; other--; ) {
3666 if (other == disk_idx)
3668 if (!test_bit(R5_UPTODATE,
3669 &sh->dev[other].flags))
3673 pr_debug("Computing stripe %llu blocks %d,%d\n",
3674 (unsigned long long)sh->sector,
3676 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3677 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3678 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3679 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3680 sh->ops.target = disk_idx;
3681 sh->ops.target2 = other;
3685 } else if (test_bit(R5_Insync, &dev->flags)) {
3686 set_bit(R5_LOCKED, &dev->flags);
3687 set_bit(R5_Wantread, &dev->flags);
3689 pr_debug("Reading block %d (sync=%d)\n",
3690 disk_idx, s->syncing);
3698 * handle_stripe_fill - read or compute data to satisfy pending requests.
3700 static void handle_stripe_fill(struct stripe_head *sh,
3701 struct stripe_head_state *s,
3706 /* look for blocks to read/compute, skip this if a compute
3707 * is already in flight, or if the stripe contents are in the
3708 * midst of changing due to a write
3710 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3711 !sh->reconstruct_state) {
3714 * For degraded stripe with data in journal, do not handle
3715 * read requests yet, instead, flush the stripe to raid
3716 * disks first, this avoids handling complex rmw of write
3717 * back cache (prexor with orig_page, and then xor with
3718 * page) in the read path
3720 if (s->injournal && s->failed) {
3721 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3722 r5c_make_stripe_write_out(sh);
3726 for (i = disks; i--; )
3727 if (fetch_block(sh, s, i, disks))
3731 set_bit(STRIPE_HANDLE, &sh->state);
3734 static void break_stripe_batch_list(struct stripe_head *head_sh,
3735 unsigned long handle_flags);
3736 /* handle_stripe_clean_event
3737 * any written block on an uptodate or failed drive can be returned.
3738 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3739 * never LOCKED, so we don't need to test 'failed' directly.
3741 static void handle_stripe_clean_event(struct r5conf *conf,
3742 struct stripe_head *sh, int disks)
3746 int discard_pending = 0;
3747 struct stripe_head *head_sh = sh;
3748 bool do_endio = false;
3750 for (i = disks; i--; )
3751 if (sh->dev[i].written) {
3753 if (!test_bit(R5_LOCKED, &dev->flags) &&
3754 (test_bit(R5_UPTODATE, &dev->flags) ||
3755 test_bit(R5_Discard, &dev->flags) ||
3756 test_bit(R5_SkipCopy, &dev->flags))) {
3757 /* We can return any write requests */
3758 struct bio *wbi, *wbi2;
3759 pr_debug("Return write for disc %d\n", i);
3760 if (test_and_clear_bit(R5_Discard, &dev->flags))
3761 clear_bit(R5_UPTODATE, &dev->flags);
3762 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3763 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3768 dev->page = dev->orig_page;
3770 dev->written = NULL;
3771 while (wbi && wbi->bi_iter.bi_sector <
3772 dev->sector + STRIPE_SECTORS) {
3773 wbi2 = r5_next_bio(wbi, dev->sector);
3774 md_write_end(conf->mddev);
3778 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3780 !test_bit(STRIPE_DEGRADED, &sh->state),
3782 if (head_sh->batch_head) {
3783 sh = list_first_entry(&sh->batch_list,
3786 if (sh != head_sh) {
3793 } else if (test_bit(R5_Discard, &dev->flags))
3794 discard_pending = 1;
3797 log_stripe_write_finished(sh);
3799 if (!discard_pending &&
3800 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3802 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3803 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3804 if (sh->qd_idx >= 0) {
3805 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3806 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3808 /* now that discard is done we can proceed with any sync */
3809 clear_bit(STRIPE_DISCARD, &sh->state);
3811 * SCSI discard will change some bio fields and the stripe has
3812 * no updated data, so remove it from hash list and the stripe
3813 * will be reinitialized
3816 hash = sh->hash_lock_index;
3817 spin_lock_irq(conf->hash_locks + hash);
3819 spin_unlock_irq(conf->hash_locks + hash);
3820 if (head_sh->batch_head) {
3821 sh = list_first_entry(&sh->batch_list,
3822 struct stripe_head, batch_list);
3828 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3829 set_bit(STRIPE_HANDLE, &sh->state);
3833 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3834 if (atomic_dec_and_test(&conf->pending_full_writes))
3835 md_wakeup_thread(conf->mddev->thread);
3837 if (head_sh->batch_head && do_endio)
3838 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3842 * For RMW in write back cache, we need extra page in prexor to store the
3843 * old data. This page is stored in dev->orig_page.
3845 * This function checks whether we have data for prexor. The exact logic
3847 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3849 static inline bool uptodate_for_rmw(struct r5dev *dev)
3851 return (test_bit(R5_UPTODATE, &dev->flags)) &&
3852 (!test_bit(R5_InJournal, &dev->flags) ||
3853 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3856 static int handle_stripe_dirtying(struct r5conf *conf,
3857 struct stripe_head *sh,
3858 struct stripe_head_state *s,
3861 int rmw = 0, rcw = 0, i;
3862 sector_t recovery_cp = conf->mddev->recovery_cp;
3864 /* Check whether resync is now happening or should start.
3865 * If yes, then the array is dirty (after unclean shutdown or
3866 * initial creation), so parity in some stripes might be inconsistent.
3867 * In this case, we need to always do reconstruct-write, to ensure
3868 * that in case of drive failure or read-error correction, we
3869 * generate correct data from the parity.
3871 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3872 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3874 /* Calculate the real rcw later - for now make it
3875 * look like rcw is cheaper
3878 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3879 conf->rmw_level, (unsigned long long)recovery_cp,
3880 (unsigned long long)sh->sector);
3881 } else for (i = disks; i--; ) {
3882 /* would I have to read this buffer for read_modify_write */
3883 struct r5dev *dev = &sh->dev[i];
3884 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3885 i == sh->pd_idx || i == sh->qd_idx ||
3886 test_bit(R5_InJournal, &dev->flags)) &&
3887 !test_bit(R5_LOCKED, &dev->flags) &&
3888 !(uptodate_for_rmw(dev) ||
3889 test_bit(R5_Wantcompute, &dev->flags))) {
3890 if (test_bit(R5_Insync, &dev->flags))
3893 rmw += 2*disks; /* cannot read it */
3895 /* Would I have to read this buffer for reconstruct_write */
3896 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3897 i != sh->pd_idx && i != sh->qd_idx &&
3898 !test_bit(R5_LOCKED, &dev->flags) &&
3899 !(test_bit(R5_UPTODATE, &dev->flags) ||
3900 test_bit(R5_Wantcompute, &dev->flags))) {
3901 if (test_bit(R5_Insync, &dev->flags))
3908 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3909 (unsigned long long)sh->sector, sh->state, rmw, rcw);
3910 set_bit(STRIPE_HANDLE, &sh->state);
3911 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3912 /* prefer read-modify-write, but need to get some data */
3913 if (conf->mddev->queue)
3914 blk_add_trace_msg(conf->mddev->queue,
3915 "raid5 rmw %llu %d",
3916 (unsigned long long)sh->sector, rmw);
3917 for (i = disks; i--; ) {
3918 struct r5dev *dev = &sh->dev[i];
3919 if (test_bit(R5_InJournal, &dev->flags) &&
3920 dev->page == dev->orig_page &&
3921 !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3922 /* alloc page for prexor */
3923 struct page *p = alloc_page(GFP_NOIO);
3931 * alloc_page() failed, try use
3932 * disk_info->extra_page
3934 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3935 &conf->cache_state)) {
3936 r5c_use_extra_page(sh);
3940 /* extra_page in use, add to delayed_list */
3941 set_bit(STRIPE_DELAYED, &sh->state);
3942 s->waiting_extra_page = 1;
3947 for (i = disks; i--; ) {
3948 struct r5dev *dev = &sh->dev[i];
3949 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3950 i == sh->pd_idx || i == sh->qd_idx ||
3951 test_bit(R5_InJournal, &dev->flags)) &&
3952 !test_bit(R5_LOCKED, &dev->flags) &&
3953 !(uptodate_for_rmw(dev) ||
3954 test_bit(R5_Wantcompute, &dev->flags)) &&
3955 test_bit(R5_Insync, &dev->flags)) {
3956 if (test_bit(STRIPE_PREREAD_ACTIVE,
3958 pr_debug("Read_old block %d for r-m-w\n",
3960 set_bit(R5_LOCKED, &dev->flags);
3961 set_bit(R5_Wantread, &dev->flags);
3964 set_bit(STRIPE_DELAYED, &sh->state);
3965 set_bit(STRIPE_HANDLE, &sh->state);
3970 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3971 /* want reconstruct write, but need to get some data */
3974 for (i = disks; i--; ) {
3975 struct r5dev *dev = &sh->dev[i];
3976 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3977 i != sh->pd_idx && i != sh->qd_idx &&
3978 !test_bit(R5_LOCKED, &dev->flags) &&
3979 !(test_bit(R5_UPTODATE, &dev->flags) ||
3980 test_bit(R5_Wantcompute, &dev->flags))) {
3982 if (test_bit(R5_Insync, &dev->flags) &&
3983 test_bit(STRIPE_PREREAD_ACTIVE,
3985 pr_debug("Read_old block "
3986 "%d for Reconstruct\n", i);
3987 set_bit(R5_LOCKED, &dev->flags);
3988 set_bit(R5_Wantread, &dev->flags);
3992 set_bit(STRIPE_DELAYED, &sh->state);
3993 set_bit(STRIPE_HANDLE, &sh->state);
3997 if (rcw && conf->mddev->queue)
3998 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3999 (unsigned long long)sh->sector,
4000 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4003 if (rcw > disks && rmw > disks &&
4004 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4005 set_bit(STRIPE_DELAYED, &sh->state);
4007 /* now if nothing is locked, and if we have enough data,
4008 * we can start a write request
4010 /* since handle_stripe can be called at any time we need to handle the
4011 * case where a compute block operation has been submitted and then a
4012 * subsequent call wants to start a write request. raid_run_ops only
4013 * handles the case where compute block and reconstruct are requested
4014 * simultaneously. If this is not the case then new writes need to be
4015 * held off until the compute completes.
4017 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4018 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4019 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4020 schedule_reconstruction(sh, s, rcw == 0, 0);
4024 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4025 struct stripe_head_state *s, int disks)
4027 struct r5dev *dev = NULL;
4029 BUG_ON(sh->batch_head);
4030 set_bit(STRIPE_HANDLE, &sh->state);
4032 switch (sh->check_state) {
4033 case check_state_idle:
4034 /* start a new check operation if there are no failures */
4035 if (s->failed == 0) {
4036 BUG_ON(s->uptodate != disks);
4037 sh->check_state = check_state_run;
4038 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4039 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4043 dev = &sh->dev[s->failed_num[0]];
4045 case check_state_compute_result:
4046 sh->check_state = check_state_idle;
4048 dev = &sh->dev[sh->pd_idx];
4050 /* check that a write has not made the stripe insync */
4051 if (test_bit(STRIPE_INSYNC, &sh->state))
4054 /* either failed parity check, or recovery is happening */
4055 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4056 BUG_ON(s->uptodate != disks);
4058 set_bit(R5_LOCKED, &dev->flags);
4060 set_bit(R5_Wantwrite, &dev->flags);
4062 clear_bit(STRIPE_DEGRADED, &sh->state);
4063 set_bit(STRIPE_INSYNC, &sh->state);
4065 case check_state_run:
4066 break; /* we will be called again upon completion */
4067 case check_state_check_result:
4068 sh->check_state = check_state_idle;
4070 /* if a failure occurred during the check operation, leave
4071 * STRIPE_INSYNC not set and let the stripe be handled again
4076 /* handle a successful check operation, if parity is correct
4077 * we are done. Otherwise update the mismatch count and repair
4078 * parity if !MD_RECOVERY_CHECK
4080 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4081 /* parity is correct (on disc,
4082 * not in buffer any more)
4084 set_bit(STRIPE_INSYNC, &sh->state);
4086 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4087 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4088 /* don't try to repair!! */
4089 set_bit(STRIPE_INSYNC, &sh->state);
4090 pr_warn_ratelimited("%s: mismatch sector in range "
4091 "%llu-%llu\n", mdname(conf->mddev),
4092 (unsigned long long) sh->sector,
4093 (unsigned long long) sh->sector +
4096 sh->check_state = check_state_compute_run;
4097 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4098 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4099 set_bit(R5_Wantcompute,
4100 &sh->dev[sh->pd_idx].flags);
4101 sh->ops.target = sh->pd_idx;
4102 sh->ops.target2 = -1;
4107 case check_state_compute_run:
4110 pr_err("%s: unknown check_state: %d sector: %llu\n",
4111 __func__, sh->check_state,
4112 (unsigned long long) sh->sector);
4117 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4118 struct stripe_head_state *s,
4121 int pd_idx = sh->pd_idx;
4122 int qd_idx = sh->qd_idx;
4125 BUG_ON(sh->batch_head);
4126 set_bit(STRIPE_HANDLE, &sh->state);
4128 BUG_ON(s->failed > 2);
4130 /* Want to check and possibly repair P and Q.
4131 * However there could be one 'failed' device, in which
4132 * case we can only check one of them, possibly using the
4133 * other to generate missing data
4136 switch (sh->check_state) {
4137 case check_state_idle:
4138 /* start a new check operation if there are < 2 failures */
4139 if (s->failed == s->q_failed) {
4140 /* The only possible failed device holds Q, so it
4141 * makes sense to check P (If anything else were failed,
4142 * we would have used P to recreate it).
4144 sh->check_state = check_state_run;
4146 if (!s->q_failed && s->failed < 2) {
4147 /* Q is not failed, and we didn't use it to generate
4148 * anything, so it makes sense to check it
4150 if (sh->check_state == check_state_run)
4151 sh->check_state = check_state_run_pq;
4153 sh->check_state = check_state_run_q;
4156 /* discard potentially stale zero_sum_result */
4157 sh->ops.zero_sum_result = 0;
4159 if (sh->check_state == check_state_run) {
4160 /* async_xor_zero_sum destroys the contents of P */
4161 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4164 if (sh->check_state >= check_state_run &&
4165 sh->check_state <= check_state_run_pq) {
4166 /* async_syndrome_zero_sum preserves P and Q, so
4167 * no need to mark them !uptodate here
4169 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4173 /* we have 2-disk failure */
4174 BUG_ON(s->failed != 2);
4176 case check_state_compute_result:
4177 sh->check_state = check_state_idle;
4179 /* check that a write has not made the stripe insync */
4180 if (test_bit(STRIPE_INSYNC, &sh->state))
4183 /* now write out any block on a failed drive,
4184 * or P or Q if they were recomputed
4187 if (s->failed == 2) {
4188 dev = &sh->dev[s->failed_num[1]];
4190 set_bit(R5_LOCKED, &dev->flags);
4191 set_bit(R5_Wantwrite, &dev->flags);
4193 if (s->failed >= 1) {
4194 dev = &sh->dev[s->failed_num[0]];
4196 set_bit(R5_LOCKED, &dev->flags);
4197 set_bit(R5_Wantwrite, &dev->flags);
4199 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4200 dev = &sh->dev[pd_idx];
4202 set_bit(R5_LOCKED, &dev->flags);
4203 set_bit(R5_Wantwrite, &dev->flags);
4205 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4206 dev = &sh->dev[qd_idx];
4208 set_bit(R5_LOCKED, &dev->flags);
4209 set_bit(R5_Wantwrite, &dev->flags);
4211 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4212 "%s: disk%td not up to date\n",
4213 mdname(conf->mddev),
4214 dev - (struct r5dev *) &sh->dev)) {
4215 clear_bit(R5_LOCKED, &dev->flags);
4216 clear_bit(R5_Wantwrite, &dev->flags);
4219 clear_bit(STRIPE_DEGRADED, &sh->state);
4221 set_bit(STRIPE_INSYNC, &sh->state);
4223 case check_state_run:
4224 case check_state_run_q:
4225 case check_state_run_pq:
4226 break; /* we will be called again upon completion */
4227 case check_state_check_result:
4228 sh->check_state = check_state_idle;
4230 /* handle a successful check operation, if parity is correct
4231 * we are done. Otherwise update the mismatch count and repair
4232 * parity if !MD_RECOVERY_CHECK
4234 if (sh->ops.zero_sum_result == 0) {
4235 /* both parities are correct */
4237 set_bit(STRIPE_INSYNC, &sh->state);
4239 /* in contrast to the raid5 case we can validate
4240 * parity, but still have a failure to write
4243 sh->check_state = check_state_compute_result;
4244 /* Returning at this point means that we may go
4245 * off and bring p and/or q uptodate again so
4246 * we make sure to check zero_sum_result again
4247 * to verify if p or q need writeback
4251 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4252 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4253 /* don't try to repair!! */
4254 set_bit(STRIPE_INSYNC, &sh->state);
4255 pr_warn_ratelimited("%s: mismatch sector in range "
4256 "%llu-%llu\n", mdname(conf->mddev),
4257 (unsigned long long) sh->sector,
4258 (unsigned long long) sh->sector +
4261 int *target = &sh->ops.target;
4263 sh->ops.target = -1;
4264 sh->ops.target2 = -1;
4265 sh->check_state = check_state_compute_run;
4266 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4267 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4268 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4269 set_bit(R5_Wantcompute,
4270 &sh->dev[pd_idx].flags);
4272 target = &sh->ops.target2;
4275 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4276 set_bit(R5_Wantcompute,
4277 &sh->dev[qd_idx].flags);
4284 case check_state_compute_run:
4287 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4288 __func__, sh->check_state,
4289 (unsigned long long) sh->sector);
4294 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4298 /* We have read all the blocks in this stripe and now we need to
4299 * copy some of them into a target stripe for expand.
4301 struct dma_async_tx_descriptor *tx = NULL;
4302 BUG_ON(sh->batch_head);
4303 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4304 for (i = 0; i < sh->disks; i++)
4305 if (i != sh->pd_idx && i != sh->qd_idx) {
4307 struct stripe_head *sh2;
4308 struct async_submit_ctl submit;
4310 sector_t bn = raid5_compute_blocknr(sh, i, 1);
4311 sector_t s = raid5_compute_sector(conf, bn, 0,
4313 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4315 /* so far only the early blocks of this stripe
4316 * have been requested. When later blocks
4317 * get requested, we will try again
4320 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4321 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4322 /* must have already done this block */
4323 raid5_release_stripe(sh2);
4327 /* place all the copies on one channel */
4328 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4329 tx = async_memcpy(sh2->dev[dd_idx].page,
4330 sh->dev[i].page, 0, 0, STRIPE_SIZE,
4333 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4334 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4335 for (j = 0; j < conf->raid_disks; j++)
4336 if (j != sh2->pd_idx &&
4338 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4340 if (j == conf->raid_disks) {
4341 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4342 set_bit(STRIPE_HANDLE, &sh2->state);
4344 raid5_release_stripe(sh2);
4347 /* done submitting copies, wait for them to complete */
4348 async_tx_quiesce(&tx);
4352 * handle_stripe - do things to a stripe.
4354 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4355 * state of various bits to see what needs to be done.
4357 * return some read requests which now have data
4358 * return some write requests which are safely on storage
4359 * schedule a read on some buffers
4360 * schedule a write of some buffers
4361 * return confirmation of parity correctness
4365 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4367 struct r5conf *conf = sh->raid_conf;
4368 int disks = sh->disks;
4371 int do_recovery = 0;
4373 memset(s, 0, sizeof(*s));
4375 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4376 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4377 s->failed_num[0] = -1;
4378 s->failed_num[1] = -1;
4379 s->log_failed = r5l_log_disk_error(conf);
4381 /* Now to look around and see what can be done */
4383 for (i=disks; i--; ) {
4384 struct md_rdev *rdev;
4391 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4393 dev->toread, dev->towrite, dev->written);
4394 /* maybe we can reply to a read
4396 * new wantfill requests are only permitted while
4397 * ops_complete_biofill is guaranteed to be inactive
4399 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4400 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4401 set_bit(R5_Wantfill, &dev->flags);
4403 /* now count some things */
4404 if (test_bit(R5_LOCKED, &dev->flags))
4406 if (test_bit(R5_UPTODATE, &dev->flags))
4408 if (test_bit(R5_Wantcompute, &dev->flags)) {
4410 BUG_ON(s->compute > 2);
4413 if (test_bit(R5_Wantfill, &dev->flags))
4415 else if (dev->toread)
4419 if (!test_bit(R5_OVERWRITE, &dev->flags))
4424 /* Prefer to use the replacement for reads, but only
4425 * if it is recovered enough and has no bad blocks.
4427 rdev = rcu_dereference(conf->disks[i].replacement);
4428 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4429 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4430 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4431 &first_bad, &bad_sectors))
4432 set_bit(R5_ReadRepl, &dev->flags);
4434 if (rdev && !test_bit(Faulty, &rdev->flags))
4435 set_bit(R5_NeedReplace, &dev->flags);
4437 clear_bit(R5_NeedReplace, &dev->flags);
4438 rdev = rcu_dereference(conf->disks[i].rdev);
4439 clear_bit(R5_ReadRepl, &dev->flags);
4441 if (rdev && test_bit(Faulty, &rdev->flags))
4444 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4445 &first_bad, &bad_sectors);
4446 if (s->blocked_rdev == NULL
4447 && (test_bit(Blocked, &rdev->flags)
4450 set_bit(BlockedBadBlocks,
4452 s->blocked_rdev = rdev;
4453 atomic_inc(&rdev->nr_pending);
4456 clear_bit(R5_Insync, &dev->flags);
4460 /* also not in-sync */
4461 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4462 test_bit(R5_UPTODATE, &dev->flags)) {
4463 /* treat as in-sync, but with a read error
4464 * which we can now try to correct
4466 set_bit(R5_Insync, &dev->flags);
4467 set_bit(R5_ReadError, &dev->flags);
4469 } else if (test_bit(In_sync, &rdev->flags))
4470 set_bit(R5_Insync, &dev->flags);
4471 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4472 /* in sync if before recovery_offset */
4473 set_bit(R5_Insync, &dev->flags);
4474 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4475 test_bit(R5_Expanded, &dev->flags))
4476 /* If we've reshaped into here, we assume it is Insync.
4477 * We will shortly update recovery_offset to make
4480 set_bit(R5_Insync, &dev->flags);
4482 if (test_bit(R5_WriteError, &dev->flags)) {
4483 /* This flag does not apply to '.replacement'
4484 * only to .rdev, so make sure to check that*/
4485 struct md_rdev *rdev2 = rcu_dereference(
4486 conf->disks[i].rdev);
4488 clear_bit(R5_Insync, &dev->flags);
4489 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4490 s->handle_bad_blocks = 1;
4491 atomic_inc(&rdev2->nr_pending);
4493 clear_bit(R5_WriteError, &dev->flags);
4495 if (test_bit(R5_MadeGood, &dev->flags)) {
4496 /* This flag does not apply to '.replacement'
4497 * only to .rdev, so make sure to check that*/
4498 struct md_rdev *rdev2 = rcu_dereference(
4499 conf->disks[i].rdev);
4500 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4501 s->handle_bad_blocks = 1;
4502 atomic_inc(&rdev2->nr_pending);
4504 clear_bit(R5_MadeGood, &dev->flags);
4506 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4507 struct md_rdev *rdev2 = rcu_dereference(
4508 conf->disks[i].replacement);
4509 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4510 s->handle_bad_blocks = 1;
4511 atomic_inc(&rdev2->nr_pending);
4513 clear_bit(R5_MadeGoodRepl, &dev->flags);
4515 if (!test_bit(R5_Insync, &dev->flags)) {
4516 /* The ReadError flag will just be confusing now */
4517 clear_bit(R5_ReadError, &dev->flags);
4518 clear_bit(R5_ReWrite, &dev->flags);
4520 if (test_bit(R5_ReadError, &dev->flags))
4521 clear_bit(R5_Insync, &dev->flags);
4522 if (!test_bit(R5_Insync, &dev->flags)) {
4524 s->failed_num[s->failed] = i;
4526 if (rdev && !test_bit(Faulty, &rdev->flags))
4529 rdev = rcu_dereference(
4530 conf->disks[i].replacement);
4531 if (rdev && !test_bit(Faulty, &rdev->flags))
4536 if (test_bit(R5_InJournal, &dev->flags))
4538 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4541 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4542 /* If there is a failed device being replaced,
4543 * we must be recovering.
4544 * else if we are after recovery_cp, we must be syncing
4545 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4546 * else we can only be replacing
4547 * sync and recovery both need to read all devices, and so
4548 * use the same flag.
4551 sh->sector >= conf->mddev->recovery_cp ||
4552 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4560 static int clear_batch_ready(struct stripe_head *sh)
4562 /* Return '1' if this is a member of batch, or
4563 * '0' if it is a lone stripe or a head which can now be
4566 struct stripe_head *tmp;
4567 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4568 return (sh->batch_head && sh->batch_head != sh);
4569 spin_lock(&sh->stripe_lock);
4570 if (!sh->batch_head) {
4571 spin_unlock(&sh->stripe_lock);
4576 * this stripe could be added to a batch list before we check
4577 * BATCH_READY, skips it
4579 if (sh->batch_head != sh) {
4580 spin_unlock(&sh->stripe_lock);
4583 spin_lock(&sh->batch_lock);
4584 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4585 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4586 spin_unlock(&sh->batch_lock);
4587 spin_unlock(&sh->stripe_lock);
4590 * BATCH_READY is cleared, no new stripes can be added.
4591 * batch_list can be accessed without lock
4596 static void break_stripe_batch_list(struct stripe_head *head_sh,
4597 unsigned long handle_flags)
4599 struct stripe_head *sh, *next;
4603 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4605 list_del_init(&sh->batch_list);
4607 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4608 (1 << STRIPE_SYNCING) |
4609 (1 << STRIPE_REPLACED) |
4610 (1 << STRIPE_DELAYED) |
4611 (1 << STRIPE_BIT_DELAY) |
4612 (1 << STRIPE_FULL_WRITE) |
4613 (1 << STRIPE_BIOFILL_RUN) |
4614 (1 << STRIPE_COMPUTE_RUN) |
4615 (1 << STRIPE_OPS_REQ_PENDING) |
4616 (1 << STRIPE_DISCARD) |
4617 (1 << STRIPE_BATCH_READY) |
4618 (1 << STRIPE_BATCH_ERR) |
4619 (1 << STRIPE_BITMAP_PENDING)),
4620 "stripe state: %lx\n", sh->state);
4621 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4622 (1 << STRIPE_REPLACED)),
4623 "head stripe state: %lx\n", head_sh->state);
4625 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4626 (1 << STRIPE_PREREAD_ACTIVE) |
4627 (1 << STRIPE_DEGRADED) |
4628 (1 << STRIPE_ON_UNPLUG_LIST)),
4629 head_sh->state & (1 << STRIPE_INSYNC));
4631 sh->check_state = head_sh->check_state;
4632 sh->reconstruct_state = head_sh->reconstruct_state;
4633 spin_lock_irq(&sh->stripe_lock);
4634 sh->batch_head = NULL;
4635 spin_unlock_irq(&sh->stripe_lock);
4636 for (i = 0; i < sh->disks; i++) {
4637 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4639 sh->dev[i].flags = head_sh->dev[i].flags &
4640 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4642 if (handle_flags == 0 ||
4643 sh->state & handle_flags)
4644 set_bit(STRIPE_HANDLE, &sh->state);
4645 raid5_release_stripe(sh);
4647 spin_lock_irq(&head_sh->stripe_lock);
4648 head_sh->batch_head = NULL;
4649 spin_unlock_irq(&head_sh->stripe_lock);
4650 for (i = 0; i < head_sh->disks; i++)
4651 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4653 if (head_sh->state & handle_flags)
4654 set_bit(STRIPE_HANDLE, &head_sh->state);
4657 wake_up(&head_sh->raid_conf->wait_for_overlap);
4660 static void handle_stripe(struct stripe_head *sh)
4662 struct stripe_head_state s;
4663 struct r5conf *conf = sh->raid_conf;
4666 int disks = sh->disks;
4667 struct r5dev *pdev, *qdev;
4669 clear_bit(STRIPE_HANDLE, &sh->state);
4670 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4671 /* already being handled, ensure it gets handled
4672 * again when current action finishes */
4673 set_bit(STRIPE_HANDLE, &sh->state);
4677 if (clear_batch_ready(sh) ) {
4678 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4682 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4683 break_stripe_batch_list(sh, 0);
4685 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4686 spin_lock(&sh->stripe_lock);
4688 * Cannot process 'sync' concurrently with 'discard'.
4689 * Flush data in r5cache before 'sync'.
4691 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4692 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4693 !test_bit(STRIPE_DISCARD, &sh->state) &&
4694 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4695 set_bit(STRIPE_SYNCING, &sh->state);
4696 clear_bit(STRIPE_INSYNC, &sh->state);
4697 clear_bit(STRIPE_REPLACED, &sh->state);
4699 spin_unlock(&sh->stripe_lock);
4701 clear_bit(STRIPE_DELAYED, &sh->state);
4703 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4704 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4705 (unsigned long long)sh->sector, sh->state,
4706 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4707 sh->check_state, sh->reconstruct_state);
4709 analyse_stripe(sh, &s);
4711 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4714 if (s.handle_bad_blocks ||
4715 test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4716 set_bit(STRIPE_HANDLE, &sh->state);
4720 if (unlikely(s.blocked_rdev)) {
4721 if (s.syncing || s.expanding || s.expanded ||
4722 s.replacing || s.to_write || s.written) {
4723 set_bit(STRIPE_HANDLE, &sh->state);
4726 /* There is nothing for the blocked_rdev to block */
4727 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4728 s.blocked_rdev = NULL;
4731 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4732 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4733 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4736 pr_debug("locked=%d uptodate=%d to_read=%d"
4737 " to_write=%d failed=%d failed_num=%d,%d\n",
4738 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4739 s.failed_num[0], s.failed_num[1]);
4741 * check if the array has lost more than max_degraded devices and,
4742 * if so, some requests might need to be failed.
4744 * When journal device failed (log_failed), we will only process
4745 * the stripe if there is data need write to raid disks
4747 if (s.failed > conf->max_degraded ||
4748 (s.log_failed && s.injournal == 0)) {
4749 sh->check_state = 0;
4750 sh->reconstruct_state = 0;
4751 break_stripe_batch_list(sh, 0);
4752 if (s.to_read+s.to_write+s.written)
4753 handle_failed_stripe(conf, sh, &s, disks);
4754 if (s.syncing + s.replacing)
4755 handle_failed_sync(conf, sh, &s);
4758 /* Now we check to see if any write operations have recently
4762 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4764 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4765 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4766 sh->reconstruct_state = reconstruct_state_idle;
4768 /* All the 'written' buffers and the parity block are ready to
4769 * be written back to disk
4771 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4772 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4773 BUG_ON(sh->qd_idx >= 0 &&
4774 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4775 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4776 for (i = disks; i--; ) {
4777 struct r5dev *dev = &sh->dev[i];
4778 if (test_bit(R5_LOCKED, &dev->flags) &&
4779 (i == sh->pd_idx || i == sh->qd_idx ||
4780 dev->written || test_bit(R5_InJournal,
4782 pr_debug("Writing block %d\n", i);
4783 set_bit(R5_Wantwrite, &dev->flags);
4788 if (!test_bit(R5_Insync, &dev->flags) ||
4789 ((i == sh->pd_idx || i == sh->qd_idx) &&
4791 set_bit(STRIPE_INSYNC, &sh->state);
4794 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4795 s.dec_preread_active = 1;
4799 * might be able to return some write requests if the parity blocks
4800 * are safe, or on a failed drive
4802 pdev = &sh->dev[sh->pd_idx];
4803 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4804 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4805 qdev = &sh->dev[sh->qd_idx];
4806 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4807 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4811 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4812 && !test_bit(R5_LOCKED, &pdev->flags)
4813 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4814 test_bit(R5_Discard, &pdev->flags))))) &&
4815 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4816 && !test_bit(R5_LOCKED, &qdev->flags)
4817 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4818 test_bit(R5_Discard, &qdev->flags))))))
4819 handle_stripe_clean_event(conf, sh, disks);
4822 r5c_handle_cached_data_endio(conf, sh, disks);
4823 log_stripe_write_finished(sh);
4825 /* Now we might consider reading some blocks, either to check/generate
4826 * parity, or to satisfy requests
4827 * or to load a block that is being partially written.
4829 if (s.to_read || s.non_overwrite
4830 || (conf->level == 6 && s.to_write && s.failed)
4831 || (s.syncing && (s.uptodate + s.compute < disks))
4834 handle_stripe_fill(sh, &s, disks);
4837 * When the stripe finishes full journal write cycle (write to journal
4838 * and raid disk), this is the clean up procedure so it is ready for
4841 r5c_finish_stripe_write_out(conf, sh, &s);
4844 * Now to consider new write requests, cache write back and what else,
4845 * if anything should be read. We do not handle new writes when:
4846 * 1/ A 'write' operation (copy+xor) is already in flight.
4847 * 2/ A 'check' operation is in flight, as it may clobber the parity
4849 * 3/ A r5c cache log write is in flight.
4852 if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4853 if (!r5c_is_writeback(conf->log)) {
4855 handle_stripe_dirtying(conf, sh, &s, disks);
4856 } else { /* write back cache */
4859 /* First, try handle writes in caching phase */
4861 ret = r5c_try_caching_write(conf, sh, &s,
4864 * If caching phase failed: ret == -EAGAIN
4866 * stripe under reclaim: !caching && injournal
4868 * fall back to handle_stripe_dirtying()
4870 if (ret == -EAGAIN ||
4871 /* stripe under reclaim: !caching && injournal */
4872 (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4874 ret = handle_stripe_dirtying(conf, sh, &s,
4882 /* maybe we need to check and possibly fix the parity for this stripe
4883 * Any reads will already have been scheduled, so we just see if enough
4884 * data is available. The parity check is held off while parity
4885 * dependent operations are in flight.
4887 if (sh->check_state ||
4888 (s.syncing && s.locked == 0 &&
4889 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4890 !test_bit(STRIPE_INSYNC, &sh->state))) {
4891 if (conf->level == 6)
4892 handle_parity_checks6(conf, sh, &s, disks);
4894 handle_parity_checks5(conf, sh, &s, disks);
4897 if ((s.replacing || s.syncing) && s.locked == 0
4898 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4899 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4900 /* Write out to replacement devices where possible */
4901 for (i = 0; i < conf->raid_disks; i++)
4902 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4903 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4904 set_bit(R5_WantReplace, &sh->dev[i].flags);
4905 set_bit(R5_LOCKED, &sh->dev[i].flags);
4909 set_bit(STRIPE_INSYNC, &sh->state);
4910 set_bit(STRIPE_REPLACED, &sh->state);
4912 if ((s.syncing || s.replacing) && s.locked == 0 &&
4913 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4914 test_bit(STRIPE_INSYNC, &sh->state)) {
4915 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4916 clear_bit(STRIPE_SYNCING, &sh->state);
4917 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4918 wake_up(&conf->wait_for_overlap);
4921 /* If the failed drives are just a ReadError, then we might need
4922 * to progress the repair/check process
4924 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4925 for (i = 0; i < s.failed; i++) {
4926 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4927 if (test_bit(R5_ReadError, &dev->flags)
4928 && !test_bit(R5_LOCKED, &dev->flags)
4929 && test_bit(R5_UPTODATE, &dev->flags)
4931 if (!test_bit(R5_ReWrite, &dev->flags)) {
4932 set_bit(R5_Wantwrite, &dev->flags);
4933 set_bit(R5_ReWrite, &dev->flags);
4934 set_bit(R5_LOCKED, &dev->flags);
4937 /* let's read it back */
4938 set_bit(R5_Wantread, &dev->flags);
4939 set_bit(R5_LOCKED, &dev->flags);
4945 /* Finish reconstruct operations initiated by the expansion process */
4946 if (sh->reconstruct_state == reconstruct_state_result) {
4947 struct stripe_head *sh_src
4948 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4949 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4950 /* sh cannot be written until sh_src has been read.
4951 * so arrange for sh to be delayed a little
4953 set_bit(STRIPE_DELAYED, &sh->state);
4954 set_bit(STRIPE_HANDLE, &sh->state);
4955 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4957 atomic_inc(&conf->preread_active_stripes);
4958 raid5_release_stripe(sh_src);
4962 raid5_release_stripe(sh_src);
4964 sh->reconstruct_state = reconstruct_state_idle;
4965 clear_bit(STRIPE_EXPANDING, &sh->state);
4966 for (i = conf->raid_disks; i--; ) {
4967 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4968 set_bit(R5_LOCKED, &sh->dev[i].flags);
4973 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4974 !sh->reconstruct_state) {
4975 /* Need to write out all blocks after computing parity */
4976 sh->disks = conf->raid_disks;
4977 stripe_set_idx(sh->sector, conf, 0, sh);
4978 schedule_reconstruction(sh, &s, 1, 1);
4979 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4980 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4981 atomic_dec(&conf->reshape_stripes);
4982 wake_up(&conf->wait_for_overlap);
4983 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4986 if (s.expanding && s.locked == 0 &&
4987 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4988 handle_stripe_expansion(conf, sh);
4991 /* wait for this device to become unblocked */
4992 if (unlikely(s.blocked_rdev)) {
4993 if (conf->mddev->external)
4994 md_wait_for_blocked_rdev(s.blocked_rdev,
4997 /* Internal metadata will immediately
4998 * be written by raid5d, so we don't
4999 * need to wait here.
5001 rdev_dec_pending(s.blocked_rdev,
5005 if (s.handle_bad_blocks)
5006 for (i = disks; i--; ) {
5007 struct md_rdev *rdev;
5008 struct r5dev *dev = &sh->dev[i];
5009 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5010 /* We own a safe reference to the rdev */
5011 rdev = conf->disks[i].rdev;
5012 if (!rdev_set_badblocks(rdev, sh->sector,
5014 md_error(conf->mddev, rdev);
5015 rdev_dec_pending(rdev, conf->mddev);
5017 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5018 rdev = conf->disks[i].rdev;
5019 rdev_clear_badblocks(rdev, sh->sector,
5021 rdev_dec_pending(rdev, conf->mddev);
5023 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5024 rdev = conf->disks[i].replacement;
5026 /* rdev have been moved down */
5027 rdev = conf->disks[i].rdev;
5028 rdev_clear_badblocks(rdev, sh->sector,
5030 rdev_dec_pending(rdev, conf->mddev);
5035 raid_run_ops(sh, s.ops_request);
5039 if (s.dec_preread_active) {
5040 /* We delay this until after ops_run_io so that if make_request
5041 * is waiting on a flush, it won't continue until the writes
5042 * have actually been submitted.
5044 atomic_dec(&conf->preread_active_stripes);
5045 if (atomic_read(&conf->preread_active_stripes) <
5047 md_wakeup_thread(conf->mddev->thread);
5050 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5053 static void raid5_activate_delayed(struct r5conf *conf)
5055 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5056 while (!list_empty(&conf->delayed_list)) {
5057 struct list_head *l = conf->delayed_list.next;
5058 struct stripe_head *sh;
5059 sh = list_entry(l, struct stripe_head, lru);
5061 clear_bit(STRIPE_DELAYED, &sh->state);
5062 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5063 atomic_inc(&conf->preread_active_stripes);
5064 list_add_tail(&sh->lru, &conf->hold_list);
5065 raid5_wakeup_stripe_thread(sh);
5070 static void activate_bit_delay(struct r5conf *conf,
5071 struct list_head *temp_inactive_list)
5073 /* device_lock is held */
5074 struct list_head head;
5075 list_add(&head, &conf->bitmap_list);
5076 list_del_init(&conf->bitmap_list);
5077 while (!list_empty(&head)) {
5078 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5080 list_del_init(&sh->lru);
5081 atomic_inc(&sh->count);
5082 hash = sh->hash_lock_index;
5083 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5087 static int raid5_congested(struct mddev *mddev, int bits)
5089 struct r5conf *conf = mddev->private;
5091 /* No difference between reads and writes. Just check
5092 * how busy the stripe_cache is
5095 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5098 /* Also checks whether there is pressure on r5cache log space */
5099 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5103 if (atomic_read(&conf->empty_inactive_list_nr))
5109 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5111 struct r5conf *conf = mddev->private;
5112 sector_t sector = bio->bi_iter.bi_sector;
5113 unsigned int chunk_sectors;
5114 unsigned int bio_sectors = bio_sectors(bio);
5116 WARN_ON_ONCE(bio->bi_partno);
5118 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5119 return chunk_sectors >=
5120 ((sector & (chunk_sectors - 1)) + bio_sectors);
5124 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5125 * later sampled by raid5d.
5127 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5129 unsigned long flags;
5131 spin_lock_irqsave(&conf->device_lock, flags);
5133 bi->bi_next = conf->retry_read_aligned_list;
5134 conf->retry_read_aligned_list = bi;
5136 spin_unlock_irqrestore(&conf->device_lock, flags);
5137 md_wakeup_thread(conf->mddev->thread);
5140 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5141 unsigned int *offset)
5145 bi = conf->retry_read_aligned;
5147 *offset = conf->retry_read_offset;
5148 conf->retry_read_aligned = NULL;
5151 bi = conf->retry_read_aligned_list;
5153 conf->retry_read_aligned_list = bi->bi_next;
5162 * The "raid5_align_endio" should check if the read succeeded and if it
5163 * did, call bio_endio on the original bio (having bio_put the new bio
5165 * If the read failed..
5167 static void raid5_align_endio(struct bio *bi)
5169 struct bio* raid_bi = bi->bi_private;
5170 struct mddev *mddev;
5171 struct r5conf *conf;
5172 struct md_rdev *rdev;
5173 blk_status_t error = bi->bi_status;
5177 rdev = (void*)raid_bi->bi_next;
5178 raid_bi->bi_next = NULL;
5179 mddev = rdev->mddev;
5180 conf = mddev->private;
5182 rdev_dec_pending(rdev, conf->mddev);
5186 if (atomic_dec_and_test(&conf->active_aligned_reads))
5187 wake_up(&conf->wait_for_quiescent);
5191 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5193 add_bio_to_retry(raid_bi, conf);
5196 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5198 struct r5conf *conf = mddev->private;
5200 struct bio* align_bi;
5201 struct md_rdev *rdev;
5202 sector_t end_sector;
5204 if (!in_chunk_boundary(mddev, raid_bio)) {
5205 pr_debug("%s: non aligned\n", __func__);
5209 * use bio_clone_fast to make a copy of the bio
5211 align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5215 * set bi_end_io to a new function, and set bi_private to the
5218 align_bi->bi_end_io = raid5_align_endio;
5219 align_bi->bi_private = raid_bio;
5223 align_bi->bi_iter.bi_sector =
5224 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5227 end_sector = bio_end_sector(align_bi);
5229 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5230 if (!rdev || test_bit(Faulty, &rdev->flags) ||
5231 rdev->recovery_offset < end_sector) {
5232 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5234 (test_bit(Faulty, &rdev->flags) ||
5235 !(test_bit(In_sync, &rdev->flags) ||
5236 rdev->recovery_offset >= end_sector)))
5240 if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5250 atomic_inc(&rdev->nr_pending);
5252 raid_bio->bi_next = (void*)rdev;
5253 bio_set_dev(align_bi, rdev->bdev);
5255 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5256 bio_sectors(align_bi),
5257 &first_bad, &bad_sectors)) {
5259 rdev_dec_pending(rdev, mddev);
5263 /* No reshape active, so we can trust rdev->data_offset */
5264 align_bi->bi_iter.bi_sector += rdev->data_offset;
5266 spin_lock_irq(&conf->device_lock);
5267 wait_event_lock_irq(conf->wait_for_quiescent,
5270 atomic_inc(&conf->active_aligned_reads);
5271 spin_unlock_irq(&conf->device_lock);
5274 trace_block_bio_remap(align_bi->bi_disk->queue,
5275 align_bi, disk_devt(mddev->gendisk),
5276 raid_bio->bi_iter.bi_sector);
5277 generic_make_request(align_bi);
5286 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5289 sector_t sector = raid_bio->bi_iter.bi_sector;
5290 unsigned chunk_sects = mddev->chunk_sectors;
5291 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5293 if (sectors < bio_sectors(raid_bio)) {
5294 struct r5conf *conf = mddev->private;
5295 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5296 bio_chain(split, raid_bio);
5297 generic_make_request(raid_bio);
5301 if (!raid5_read_one_chunk(mddev, raid_bio))
5307 /* __get_priority_stripe - get the next stripe to process
5309 * Full stripe writes are allowed to pass preread active stripes up until
5310 * the bypass_threshold is exceeded. In general the bypass_count
5311 * increments when the handle_list is handled before the hold_list; however, it
5312 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5313 * stripe with in flight i/o. The bypass_count will be reset when the
5314 * head of the hold_list has changed, i.e. the head was promoted to the
5317 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5319 struct stripe_head *sh, *tmp;
5320 struct list_head *handle_list = NULL;
5321 struct r5worker_group *wg;
5322 bool second_try = !r5c_is_writeback(conf->log) &&
5323 !r5l_log_disk_error(conf);
5324 bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5325 r5l_log_disk_error(conf);
5330 if (conf->worker_cnt_per_group == 0) {
5331 handle_list = try_loprio ? &conf->loprio_list :
5333 } else if (group != ANY_GROUP) {
5334 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5335 &conf->worker_groups[group].handle_list;
5336 wg = &conf->worker_groups[group];
5339 for (i = 0; i < conf->group_cnt; i++) {
5340 handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5341 &conf->worker_groups[i].handle_list;
5342 wg = &conf->worker_groups[i];
5343 if (!list_empty(handle_list))
5348 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5350 list_empty(handle_list) ? "empty" : "busy",
5351 list_empty(&conf->hold_list) ? "empty" : "busy",
5352 atomic_read(&conf->pending_full_writes), conf->bypass_count);
5354 if (!list_empty(handle_list)) {
5355 sh = list_entry(handle_list->next, typeof(*sh), lru);
5357 if (list_empty(&conf->hold_list))
5358 conf->bypass_count = 0;
5359 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5360 if (conf->hold_list.next == conf->last_hold)
5361 conf->bypass_count++;
5363 conf->last_hold = conf->hold_list.next;
5364 conf->bypass_count -= conf->bypass_threshold;
5365 if (conf->bypass_count < 0)
5366 conf->bypass_count = 0;
5369 } else if (!list_empty(&conf->hold_list) &&
5370 ((conf->bypass_threshold &&
5371 conf->bypass_count > conf->bypass_threshold) ||
5372 atomic_read(&conf->pending_full_writes) == 0)) {
5374 list_for_each_entry(tmp, &conf->hold_list, lru) {
5375 if (conf->worker_cnt_per_group == 0 ||
5376 group == ANY_GROUP ||
5377 !cpu_online(tmp->cpu) ||
5378 cpu_to_group(tmp->cpu) == group) {
5385 conf->bypass_count -= conf->bypass_threshold;
5386 if (conf->bypass_count < 0)
5387 conf->bypass_count = 0;
5396 try_loprio = !try_loprio;
5404 list_del_init(&sh->lru);
5405 BUG_ON(atomic_inc_return(&sh->count) != 1);
5409 struct raid5_plug_cb {
5410 struct blk_plug_cb cb;
5411 struct list_head list;
5412 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5415 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5417 struct raid5_plug_cb *cb = container_of(
5418 blk_cb, struct raid5_plug_cb, cb);
5419 struct stripe_head *sh;
5420 struct mddev *mddev = cb->cb.data;
5421 struct r5conf *conf = mddev->private;
5425 if (cb->list.next && !list_empty(&cb->list)) {
5426 spin_lock_irq(&conf->device_lock);
5427 while (!list_empty(&cb->list)) {
5428 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5429 list_del_init(&sh->lru);
5431 * avoid race release_stripe_plug() sees
5432 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5433 * is still in our list
5435 smp_mb__before_atomic();
5436 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5438 * STRIPE_ON_RELEASE_LIST could be set here. In that
5439 * case, the count is always > 1 here
5441 hash = sh->hash_lock_index;
5442 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5445 spin_unlock_irq(&conf->device_lock);
5447 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5448 NR_STRIPE_HASH_LOCKS);
5450 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5454 static void release_stripe_plug(struct mddev *mddev,
5455 struct stripe_head *sh)
5457 struct blk_plug_cb *blk_cb = blk_check_plugged(
5458 raid5_unplug, mddev,
5459 sizeof(struct raid5_plug_cb));
5460 struct raid5_plug_cb *cb;
5463 raid5_release_stripe(sh);
5467 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5469 if (cb->list.next == NULL) {
5471 INIT_LIST_HEAD(&cb->list);
5472 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5473 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5476 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5477 list_add_tail(&sh->lru, &cb->list);
5479 raid5_release_stripe(sh);
5482 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5484 struct r5conf *conf = mddev->private;
5485 sector_t logical_sector, last_sector;
5486 struct stripe_head *sh;
5489 if (mddev->reshape_position != MaxSector)
5490 /* Skip discard while reshape is happening */
5493 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5494 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5498 stripe_sectors = conf->chunk_sectors *
5499 (conf->raid_disks - conf->max_degraded);
5500 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5502 sector_div(last_sector, stripe_sectors);
5504 logical_sector *= conf->chunk_sectors;
5505 last_sector *= conf->chunk_sectors;
5507 for (; logical_sector < last_sector;
5508 logical_sector += STRIPE_SECTORS) {
5512 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5513 prepare_to_wait(&conf->wait_for_overlap, &w,
5514 TASK_UNINTERRUPTIBLE);
5515 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5516 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5517 raid5_release_stripe(sh);
5521 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5522 spin_lock_irq(&sh->stripe_lock);
5523 for (d = 0; d < conf->raid_disks; d++) {
5524 if (d == sh->pd_idx || d == sh->qd_idx)
5526 if (sh->dev[d].towrite || sh->dev[d].toread) {
5527 set_bit(R5_Overlap, &sh->dev[d].flags);
5528 spin_unlock_irq(&sh->stripe_lock);
5529 raid5_release_stripe(sh);
5534 set_bit(STRIPE_DISCARD, &sh->state);
5535 finish_wait(&conf->wait_for_overlap, &w);
5536 sh->overwrite_disks = 0;
5537 for (d = 0; d < conf->raid_disks; d++) {
5538 if (d == sh->pd_idx || d == sh->qd_idx)
5540 sh->dev[d].towrite = bi;
5541 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5542 bio_inc_remaining(bi);
5543 md_write_inc(mddev, bi);
5544 sh->overwrite_disks++;
5546 spin_unlock_irq(&sh->stripe_lock);
5547 if (conf->mddev->bitmap) {
5549 d < conf->raid_disks - conf->max_degraded;
5551 md_bitmap_startwrite(mddev->bitmap,
5555 sh->bm_seq = conf->seq_flush + 1;
5556 set_bit(STRIPE_BIT_DELAY, &sh->state);
5559 set_bit(STRIPE_HANDLE, &sh->state);
5560 clear_bit(STRIPE_DELAYED, &sh->state);
5561 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5562 atomic_inc(&conf->preread_active_stripes);
5563 release_stripe_plug(mddev, sh);
5569 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5571 struct r5conf *conf = mddev->private;
5573 sector_t new_sector;
5574 sector_t logical_sector, last_sector;
5575 struct stripe_head *sh;
5576 const int rw = bio_data_dir(bi);
5579 bool do_flush = false;
5581 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5582 int ret = log_handle_flush_request(conf, bi);
5586 if (ret == -ENODEV) {
5587 md_flush_request(mddev, bi);
5590 /* ret == -EAGAIN, fallback */
5592 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5593 * we need to flush journal device
5595 do_flush = bi->bi_opf & REQ_PREFLUSH;
5598 if (!md_write_start(mddev, bi))
5601 * If array is degraded, better not do chunk aligned read because
5602 * later we might have to read it again in order to reconstruct
5603 * data on failed drives.
5605 if (rw == READ && mddev->degraded == 0 &&
5606 mddev->reshape_position == MaxSector) {
5607 bi = chunk_aligned_read(mddev, bi);
5612 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5613 make_discard_request(mddev, bi);
5614 md_write_end(mddev);
5618 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5619 last_sector = bio_end_sector(bi);
5622 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5623 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5629 seq = read_seqcount_begin(&conf->gen_lock);
5632 prepare_to_wait(&conf->wait_for_overlap, &w,
5633 TASK_UNINTERRUPTIBLE);
5634 if (unlikely(conf->reshape_progress != MaxSector)) {
5635 /* spinlock is needed as reshape_progress may be
5636 * 64bit on a 32bit platform, and so it might be
5637 * possible to see a half-updated value
5638 * Of course reshape_progress could change after
5639 * the lock is dropped, so once we get a reference
5640 * to the stripe that we think it is, we will have
5643 spin_lock_irq(&conf->device_lock);
5644 if (mddev->reshape_backwards
5645 ? logical_sector < conf->reshape_progress
5646 : logical_sector >= conf->reshape_progress) {
5649 if (mddev->reshape_backwards
5650 ? logical_sector < conf->reshape_safe
5651 : logical_sector >= conf->reshape_safe) {
5652 spin_unlock_irq(&conf->device_lock);
5658 spin_unlock_irq(&conf->device_lock);
5661 new_sector = raid5_compute_sector(conf, logical_sector,
5664 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5665 (unsigned long long)new_sector,
5666 (unsigned long long)logical_sector);
5668 sh = raid5_get_active_stripe(conf, new_sector, previous,
5669 (bi->bi_opf & REQ_RAHEAD), 0);
5671 if (unlikely(previous)) {
5672 /* expansion might have moved on while waiting for a
5673 * stripe, so we must do the range check again.
5674 * Expansion could still move past after this
5675 * test, but as we are holding a reference to
5676 * 'sh', we know that if that happens,
5677 * STRIPE_EXPANDING will get set and the expansion
5678 * won't proceed until we finish with the stripe.
5681 spin_lock_irq(&conf->device_lock);
5682 if (mddev->reshape_backwards
5683 ? logical_sector >= conf->reshape_progress
5684 : logical_sector < conf->reshape_progress)
5685 /* mismatch, need to try again */
5687 spin_unlock_irq(&conf->device_lock);
5689 raid5_release_stripe(sh);
5695 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5696 /* Might have got the wrong stripe_head
5699 raid5_release_stripe(sh);
5703 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5704 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5705 /* Stripe is busy expanding or
5706 * add failed due to overlap. Flush everything
5709 md_wakeup_thread(mddev->thread);
5710 raid5_release_stripe(sh);
5716 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5717 /* we only need flush for one stripe */
5721 set_bit(STRIPE_HANDLE, &sh->state);
5722 clear_bit(STRIPE_DELAYED, &sh->state);
5723 if ((!sh->batch_head || sh == sh->batch_head) &&
5724 (bi->bi_opf & REQ_SYNC) &&
5725 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5726 atomic_inc(&conf->preread_active_stripes);
5727 release_stripe_plug(mddev, sh);
5729 /* cannot get stripe for read-ahead, just give-up */
5730 bi->bi_status = BLK_STS_IOERR;
5734 finish_wait(&conf->wait_for_overlap, &w);
5737 md_write_end(mddev);
5742 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5744 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5746 /* reshaping is quite different to recovery/resync so it is
5747 * handled quite separately ... here.
5749 * On each call to sync_request, we gather one chunk worth of
5750 * destination stripes and flag them as expanding.
5751 * Then we find all the source stripes and request reads.
5752 * As the reads complete, handle_stripe will copy the data
5753 * into the destination stripe and release that stripe.
5755 struct r5conf *conf = mddev->private;
5756 struct stripe_head *sh;
5757 struct md_rdev *rdev;
5758 sector_t first_sector, last_sector;
5759 int raid_disks = conf->previous_raid_disks;
5760 int data_disks = raid_disks - conf->max_degraded;
5761 int new_data_disks = conf->raid_disks - conf->max_degraded;
5764 sector_t writepos, readpos, safepos;
5765 sector_t stripe_addr;
5766 int reshape_sectors;
5767 struct list_head stripes;
5770 if (sector_nr == 0) {
5771 /* If restarting in the middle, skip the initial sectors */
5772 if (mddev->reshape_backwards &&
5773 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5774 sector_nr = raid5_size(mddev, 0, 0)
5775 - conf->reshape_progress;
5776 } else if (mddev->reshape_backwards &&
5777 conf->reshape_progress == MaxSector) {
5778 /* shouldn't happen, but just in case, finish up.*/
5779 sector_nr = MaxSector;
5780 } else if (!mddev->reshape_backwards &&
5781 conf->reshape_progress > 0)
5782 sector_nr = conf->reshape_progress;
5783 sector_div(sector_nr, new_data_disks);
5785 mddev->curr_resync_completed = sector_nr;
5786 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5793 /* We need to process a full chunk at a time.
5794 * If old and new chunk sizes differ, we need to process the
5798 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5800 /* We update the metadata at least every 10 seconds, or when
5801 * the data about to be copied would over-write the source of
5802 * the data at the front of the range. i.e. one new_stripe
5803 * along from reshape_progress new_maps to after where
5804 * reshape_safe old_maps to
5806 writepos = conf->reshape_progress;
5807 sector_div(writepos, new_data_disks);
5808 readpos = conf->reshape_progress;
5809 sector_div(readpos, data_disks);
5810 safepos = conf->reshape_safe;
5811 sector_div(safepos, data_disks);
5812 if (mddev->reshape_backwards) {
5813 BUG_ON(writepos < reshape_sectors);
5814 writepos -= reshape_sectors;
5815 readpos += reshape_sectors;
5816 safepos += reshape_sectors;
5818 writepos += reshape_sectors;
5819 /* readpos and safepos are worst-case calculations.
5820 * A negative number is overly pessimistic, and causes
5821 * obvious problems for unsigned storage. So clip to 0.
5823 readpos -= min_t(sector_t, reshape_sectors, readpos);
5824 safepos -= min_t(sector_t, reshape_sectors, safepos);
5827 /* Having calculated the 'writepos' possibly use it
5828 * to set 'stripe_addr' which is where we will write to.
5830 if (mddev->reshape_backwards) {
5831 BUG_ON(conf->reshape_progress == 0);
5832 stripe_addr = writepos;
5833 BUG_ON((mddev->dev_sectors &
5834 ~((sector_t)reshape_sectors - 1))
5835 - reshape_sectors - stripe_addr
5838 BUG_ON(writepos != sector_nr + reshape_sectors);
5839 stripe_addr = sector_nr;
5842 /* 'writepos' is the most advanced device address we might write.
5843 * 'readpos' is the least advanced device address we might read.
5844 * 'safepos' is the least address recorded in the metadata as having
5846 * If there is a min_offset_diff, these are adjusted either by
5847 * increasing the safepos/readpos if diff is negative, or
5848 * increasing writepos if diff is positive.
5849 * If 'readpos' is then behind 'writepos', there is no way that we can
5850 * ensure safety in the face of a crash - that must be done by userspace
5851 * making a backup of the data. So in that case there is no particular
5852 * rush to update metadata.
5853 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5854 * update the metadata to advance 'safepos' to match 'readpos' so that
5855 * we can be safe in the event of a crash.
5856 * So we insist on updating metadata if safepos is behind writepos and
5857 * readpos is beyond writepos.
5858 * In any case, update the metadata every 10 seconds.
5859 * Maybe that number should be configurable, but I'm not sure it is
5860 * worth it.... maybe it could be a multiple of safemode_delay???
5862 if (conf->min_offset_diff < 0) {
5863 safepos += -conf->min_offset_diff;
5864 readpos += -conf->min_offset_diff;
5866 writepos += conf->min_offset_diff;
5868 if ((mddev->reshape_backwards
5869 ? (safepos > writepos && readpos < writepos)
5870 : (safepos < writepos && readpos > writepos)) ||
5871 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5872 /* Cannot proceed until we've updated the superblock... */
5873 wait_event(conf->wait_for_overlap,
5874 atomic_read(&conf->reshape_stripes)==0
5875 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5876 if (atomic_read(&conf->reshape_stripes) != 0)
5878 mddev->reshape_position = conf->reshape_progress;
5879 mddev->curr_resync_completed = sector_nr;
5880 if (!mddev->reshape_backwards)
5881 /* Can update recovery_offset */
5882 rdev_for_each(rdev, mddev)
5883 if (rdev->raid_disk >= 0 &&
5884 !test_bit(Journal, &rdev->flags) &&
5885 !test_bit(In_sync, &rdev->flags) &&
5886 rdev->recovery_offset < sector_nr)
5887 rdev->recovery_offset = sector_nr;
5889 conf->reshape_checkpoint = jiffies;
5890 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5891 md_wakeup_thread(mddev->thread);
5892 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5893 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5894 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5896 spin_lock_irq(&conf->device_lock);
5897 conf->reshape_safe = mddev->reshape_position;
5898 spin_unlock_irq(&conf->device_lock);
5899 wake_up(&conf->wait_for_overlap);
5900 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5903 INIT_LIST_HEAD(&stripes);
5904 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5906 int skipped_disk = 0;
5907 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5908 set_bit(STRIPE_EXPANDING, &sh->state);
5909 atomic_inc(&conf->reshape_stripes);
5910 /* If any of this stripe is beyond the end of the old
5911 * array, then we need to zero those blocks
5913 for (j=sh->disks; j--;) {
5915 if (j == sh->pd_idx)
5917 if (conf->level == 6 &&
5920 s = raid5_compute_blocknr(sh, j, 0);
5921 if (s < raid5_size(mddev, 0, 0)) {
5925 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5926 set_bit(R5_Expanded, &sh->dev[j].flags);
5927 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5929 if (!skipped_disk) {
5930 set_bit(STRIPE_EXPAND_READY, &sh->state);
5931 set_bit(STRIPE_HANDLE, &sh->state);
5933 list_add(&sh->lru, &stripes);
5935 spin_lock_irq(&conf->device_lock);
5936 if (mddev->reshape_backwards)
5937 conf->reshape_progress -= reshape_sectors * new_data_disks;
5939 conf->reshape_progress += reshape_sectors * new_data_disks;
5940 spin_unlock_irq(&conf->device_lock);
5941 /* Ok, those stripe are ready. We can start scheduling
5942 * reads on the source stripes.
5943 * The source stripes are determined by mapping the first and last
5944 * block on the destination stripes.
5947 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5950 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5951 * new_data_disks - 1),
5953 if (last_sector >= mddev->dev_sectors)
5954 last_sector = mddev->dev_sectors - 1;
5955 while (first_sector <= last_sector) {
5956 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5957 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5958 set_bit(STRIPE_HANDLE, &sh->state);
5959 raid5_release_stripe(sh);
5960 first_sector += STRIPE_SECTORS;
5962 /* Now that the sources are clearly marked, we can release
5963 * the destination stripes
5965 while (!list_empty(&stripes)) {
5966 sh = list_entry(stripes.next, struct stripe_head, lru);
5967 list_del_init(&sh->lru);
5968 raid5_release_stripe(sh);
5970 /* If this takes us to the resync_max point where we have to pause,
5971 * then we need to write out the superblock.
5973 sector_nr += reshape_sectors;
5974 retn = reshape_sectors;
5976 if (mddev->curr_resync_completed > mddev->resync_max ||
5977 (sector_nr - mddev->curr_resync_completed) * 2
5978 >= mddev->resync_max - mddev->curr_resync_completed) {
5979 /* Cannot proceed until we've updated the superblock... */
5980 wait_event(conf->wait_for_overlap,
5981 atomic_read(&conf->reshape_stripes) == 0
5982 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5983 if (atomic_read(&conf->reshape_stripes) != 0)
5985 mddev->reshape_position = conf->reshape_progress;
5986 mddev->curr_resync_completed = sector_nr;
5987 if (!mddev->reshape_backwards)
5988 /* Can update recovery_offset */
5989 rdev_for_each(rdev, mddev)
5990 if (rdev->raid_disk >= 0 &&
5991 !test_bit(Journal, &rdev->flags) &&
5992 !test_bit(In_sync, &rdev->flags) &&
5993 rdev->recovery_offset < sector_nr)
5994 rdev->recovery_offset = sector_nr;
5995 conf->reshape_checkpoint = jiffies;
5996 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5997 md_wakeup_thread(mddev->thread);
5998 wait_event(mddev->sb_wait,
5999 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6000 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6001 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6003 spin_lock_irq(&conf->device_lock);
6004 conf->reshape_safe = mddev->reshape_position;
6005 spin_unlock_irq(&conf->device_lock);
6006 wake_up(&conf->wait_for_overlap);
6007 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6013 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6016 struct r5conf *conf = mddev->private;
6017 struct stripe_head *sh;
6018 sector_t max_sector = mddev->dev_sectors;
6019 sector_t sync_blocks;
6020 int still_degraded = 0;
6023 if (sector_nr >= max_sector) {
6024 /* just being told to finish up .. nothing much to do */
6026 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6031 if (mddev->curr_resync < max_sector) /* aborted */
6032 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6034 else /* completed sync */
6036 md_bitmap_close_sync(mddev->bitmap);
6041 /* Allow raid5_quiesce to complete */
6042 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6044 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6045 return reshape_request(mddev, sector_nr, skipped);
6047 /* No need to check resync_max as we never do more than one
6048 * stripe, and as resync_max will always be on a chunk boundary,
6049 * if the check in md_do_sync didn't fire, there is no chance
6050 * of overstepping resync_max here
6053 /* if there is too many failed drives and we are trying
6054 * to resync, then assert that we are finished, because there is
6055 * nothing we can do.
6057 if (mddev->degraded >= conf->max_degraded &&
6058 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6059 sector_t rv = mddev->dev_sectors - sector_nr;
6063 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6065 !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6066 sync_blocks >= STRIPE_SECTORS) {
6067 /* we can skip this block, and probably more */
6068 sync_blocks /= STRIPE_SECTORS;
6070 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6073 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6075 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6077 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6078 /* make sure we don't swamp the stripe cache if someone else
6079 * is trying to get access
6081 schedule_timeout_uninterruptible(1);
6083 /* Need to check if array will still be degraded after recovery/resync
6084 * Note in case of > 1 drive failures it's possible we're rebuilding
6085 * one drive while leaving another faulty drive in array.
6088 for (i = 0; i < conf->raid_disks; i++) {
6089 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6091 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6096 md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6098 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6099 set_bit(STRIPE_HANDLE, &sh->state);
6101 raid5_release_stripe(sh);
6103 return STRIPE_SECTORS;
6106 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6107 unsigned int offset)
6109 /* We may not be able to submit a whole bio at once as there
6110 * may not be enough stripe_heads available.
6111 * We cannot pre-allocate enough stripe_heads as we may need
6112 * more than exist in the cache (if we allow ever large chunks).
6113 * So we do one stripe head at a time and record in
6114 * ->bi_hw_segments how many have been done.
6116 * We *know* that this entire raid_bio is in one chunk, so
6117 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6119 struct stripe_head *sh;
6121 sector_t sector, logical_sector, last_sector;
6125 logical_sector = raid_bio->bi_iter.bi_sector &
6126 ~((sector_t)STRIPE_SECTORS-1);
6127 sector = raid5_compute_sector(conf, logical_sector,
6129 last_sector = bio_end_sector(raid_bio);
6131 for (; logical_sector < last_sector;
6132 logical_sector += STRIPE_SECTORS,
6133 sector += STRIPE_SECTORS,
6137 /* already done this stripe */
6140 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6143 /* failed to get a stripe - must wait */
6144 conf->retry_read_aligned = raid_bio;
6145 conf->retry_read_offset = scnt;
6149 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6150 raid5_release_stripe(sh);
6151 conf->retry_read_aligned = raid_bio;
6152 conf->retry_read_offset = scnt;
6156 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6158 raid5_release_stripe(sh);
6162 bio_endio(raid_bio);
6164 if (atomic_dec_and_test(&conf->active_aligned_reads))
6165 wake_up(&conf->wait_for_quiescent);
6169 static int handle_active_stripes(struct r5conf *conf, int group,
6170 struct r5worker *worker,
6171 struct list_head *temp_inactive_list)
6172 __releases(&conf->device_lock)
6173 __acquires(&conf->device_lock)
6175 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6176 int i, batch_size = 0, hash;
6177 bool release_inactive = false;
6179 while (batch_size < MAX_STRIPE_BATCH &&
6180 (sh = __get_priority_stripe(conf, group)) != NULL)
6181 batch[batch_size++] = sh;
6183 if (batch_size == 0) {
6184 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6185 if (!list_empty(temp_inactive_list + i))
6187 if (i == NR_STRIPE_HASH_LOCKS) {
6188 spin_unlock_irq(&conf->device_lock);
6189 log_flush_stripe_to_raid(conf);
6190 spin_lock_irq(&conf->device_lock);
6193 release_inactive = true;
6195 spin_unlock_irq(&conf->device_lock);
6197 release_inactive_stripe_list(conf, temp_inactive_list,
6198 NR_STRIPE_HASH_LOCKS);
6200 r5l_flush_stripe_to_raid(conf->log);
6201 if (release_inactive) {
6202 spin_lock_irq(&conf->device_lock);
6206 for (i = 0; i < batch_size; i++)
6207 handle_stripe(batch[i]);
6208 log_write_stripe_run(conf);
6212 spin_lock_irq(&conf->device_lock);
6213 for (i = 0; i < batch_size; i++) {
6214 hash = batch[i]->hash_lock_index;
6215 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6220 static void raid5_do_work(struct work_struct *work)
6222 struct r5worker *worker = container_of(work, struct r5worker, work);
6223 struct r5worker_group *group = worker->group;
6224 struct r5conf *conf = group->conf;
6225 struct mddev *mddev = conf->mddev;
6226 int group_id = group - conf->worker_groups;
6228 struct blk_plug plug;
6230 pr_debug("+++ raid5worker active\n");
6232 blk_start_plug(&plug);
6234 spin_lock_irq(&conf->device_lock);
6236 int batch_size, released;
6238 released = release_stripe_list(conf, worker->temp_inactive_list);
6240 batch_size = handle_active_stripes(conf, group_id, worker,
6241 worker->temp_inactive_list);
6242 worker->working = false;
6243 if (!batch_size && !released)
6245 handled += batch_size;
6246 wait_event_lock_irq(mddev->sb_wait,
6247 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6250 pr_debug("%d stripes handled\n", handled);
6252 spin_unlock_irq(&conf->device_lock);
6254 flush_deferred_bios(conf);
6256 r5l_flush_stripe_to_raid(conf->log);
6258 async_tx_issue_pending_all();
6259 blk_finish_plug(&plug);
6261 pr_debug("--- raid5worker inactive\n");
6265 * This is our raid5 kernel thread.
6267 * We scan the hash table for stripes which can be handled now.
6268 * During the scan, completed stripes are saved for us by the interrupt
6269 * handler, so that they will not have to wait for our next wakeup.
6271 static void raid5d(struct md_thread *thread)
6273 struct mddev *mddev = thread->mddev;
6274 struct r5conf *conf = mddev->private;
6276 struct blk_plug plug;
6278 pr_debug("+++ raid5d active\n");
6280 md_check_recovery(mddev);
6282 blk_start_plug(&plug);
6284 spin_lock_irq(&conf->device_lock);
6287 int batch_size, released;
6288 unsigned int offset;
6290 released = release_stripe_list(conf, conf->temp_inactive_list);
6292 clear_bit(R5_DID_ALLOC, &conf->cache_state);
6295 !list_empty(&conf->bitmap_list)) {
6296 /* Now is a good time to flush some bitmap updates */
6298 spin_unlock_irq(&conf->device_lock);
6299 md_bitmap_unplug(mddev->bitmap);
6300 spin_lock_irq(&conf->device_lock);
6301 conf->seq_write = conf->seq_flush;
6302 activate_bit_delay(conf, conf->temp_inactive_list);
6304 raid5_activate_delayed(conf);
6306 while ((bio = remove_bio_from_retry(conf, &offset))) {
6308 spin_unlock_irq(&conf->device_lock);
6309 ok = retry_aligned_read(conf, bio, offset);
6310 spin_lock_irq(&conf->device_lock);
6316 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6317 conf->temp_inactive_list);
6318 if (!batch_size && !released)
6320 handled += batch_size;
6322 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6323 spin_unlock_irq(&conf->device_lock);
6324 md_check_recovery(mddev);
6325 spin_lock_irq(&conf->device_lock);
6328 pr_debug("%d stripes handled\n", handled);
6330 spin_unlock_irq(&conf->device_lock);
6331 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6332 mutex_trylock(&conf->cache_size_mutex)) {
6333 grow_one_stripe(conf, __GFP_NOWARN);
6334 /* Set flag even if allocation failed. This helps
6335 * slow down allocation requests when mem is short
6337 set_bit(R5_DID_ALLOC, &conf->cache_state);
6338 mutex_unlock(&conf->cache_size_mutex);
6341 flush_deferred_bios(conf);
6343 r5l_flush_stripe_to_raid(conf->log);
6345 async_tx_issue_pending_all();
6346 blk_finish_plug(&plug);
6348 pr_debug("--- raid5d inactive\n");
6352 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6354 struct r5conf *conf;
6356 spin_lock(&mddev->lock);
6357 conf = mddev->private;
6359 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6360 spin_unlock(&mddev->lock);
6365 raid5_set_cache_size(struct mddev *mddev, int size)
6368 struct r5conf *conf = mddev->private;
6370 if (size <= 16 || size > 32768)
6373 conf->min_nr_stripes = size;
6374 mutex_lock(&conf->cache_size_mutex);
6375 while (size < conf->max_nr_stripes &&
6376 drop_one_stripe(conf))
6378 mutex_unlock(&conf->cache_size_mutex);
6380 md_allow_write(mddev);
6382 mutex_lock(&conf->cache_size_mutex);
6383 while (size > conf->max_nr_stripes)
6384 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6385 conf->min_nr_stripes = conf->max_nr_stripes;
6389 mutex_unlock(&conf->cache_size_mutex);
6393 EXPORT_SYMBOL(raid5_set_cache_size);
6396 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6398 struct r5conf *conf;
6402 if (len >= PAGE_SIZE)
6404 if (kstrtoul(page, 10, &new))
6406 err = mddev_lock(mddev);
6409 conf = mddev->private;
6413 err = raid5_set_cache_size(mddev, new);
6414 mddev_unlock(mddev);
6419 static struct md_sysfs_entry
6420 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6421 raid5_show_stripe_cache_size,
6422 raid5_store_stripe_cache_size);
6425 raid5_show_rmw_level(struct mddev *mddev, char *page)
6427 struct r5conf *conf = mddev->private;
6429 return sprintf(page, "%d\n", conf->rmw_level);
6435 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6437 struct r5conf *conf = mddev->private;
6443 if (len >= PAGE_SIZE)
6446 if (kstrtoul(page, 10, &new))
6449 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6452 if (new != PARITY_DISABLE_RMW &&
6453 new != PARITY_ENABLE_RMW &&
6454 new != PARITY_PREFER_RMW)
6457 conf->rmw_level = new;
6461 static struct md_sysfs_entry
6462 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6463 raid5_show_rmw_level,
6464 raid5_store_rmw_level);
6468 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6470 struct r5conf *conf;
6472 spin_lock(&mddev->lock);
6473 conf = mddev->private;
6475 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6476 spin_unlock(&mddev->lock);
6481 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6483 struct r5conf *conf;
6487 if (len >= PAGE_SIZE)
6489 if (kstrtoul(page, 10, &new))
6492 err = mddev_lock(mddev);
6495 conf = mddev->private;
6498 else if (new > conf->min_nr_stripes)
6501 conf->bypass_threshold = new;
6502 mddev_unlock(mddev);
6506 static struct md_sysfs_entry
6507 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6509 raid5_show_preread_threshold,
6510 raid5_store_preread_threshold);
6513 raid5_show_skip_copy(struct mddev *mddev, char *page)
6515 struct r5conf *conf;
6517 spin_lock(&mddev->lock);
6518 conf = mddev->private;
6520 ret = sprintf(page, "%d\n", conf->skip_copy);
6521 spin_unlock(&mddev->lock);
6526 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6528 struct r5conf *conf;
6532 if (len >= PAGE_SIZE)
6534 if (kstrtoul(page, 10, &new))
6538 err = mddev_lock(mddev);
6541 conf = mddev->private;
6544 else if (new != conf->skip_copy) {
6545 mddev_suspend(mddev);
6546 conf->skip_copy = new;
6548 mddev->queue->backing_dev_info->capabilities |=
6549 BDI_CAP_STABLE_WRITES;
6551 mddev->queue->backing_dev_info->capabilities &=
6552 ~BDI_CAP_STABLE_WRITES;
6553 mddev_resume(mddev);
6555 mddev_unlock(mddev);
6559 static struct md_sysfs_entry
6560 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6561 raid5_show_skip_copy,
6562 raid5_store_skip_copy);
6565 stripe_cache_active_show(struct mddev *mddev, char *page)
6567 struct r5conf *conf = mddev->private;
6569 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6574 static struct md_sysfs_entry
6575 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6578 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6580 struct r5conf *conf;
6582 spin_lock(&mddev->lock);
6583 conf = mddev->private;
6585 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6586 spin_unlock(&mddev->lock);
6590 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6592 int *worker_cnt_per_group,
6593 struct r5worker_group **worker_groups);
6595 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6597 struct r5conf *conf;
6600 struct r5worker_group *new_groups, *old_groups;
6601 int group_cnt, worker_cnt_per_group;
6603 if (len >= PAGE_SIZE)
6605 if (kstrtouint(page, 10, &new))
6607 /* 8192 should be big enough */
6611 err = mddev_lock(mddev);
6614 conf = mddev->private;
6617 else if (new != conf->worker_cnt_per_group) {
6618 mddev_suspend(mddev);
6620 old_groups = conf->worker_groups;
6622 flush_workqueue(raid5_wq);
6624 err = alloc_thread_groups(conf, new,
6625 &group_cnt, &worker_cnt_per_group,
6628 spin_lock_irq(&conf->device_lock);
6629 conf->group_cnt = group_cnt;
6630 conf->worker_cnt_per_group = worker_cnt_per_group;
6631 conf->worker_groups = new_groups;
6632 spin_unlock_irq(&conf->device_lock);
6635 kfree(old_groups[0].workers);
6638 mddev_resume(mddev);
6640 mddev_unlock(mddev);
6645 static struct md_sysfs_entry
6646 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6647 raid5_show_group_thread_cnt,
6648 raid5_store_group_thread_cnt);
6650 static struct attribute *raid5_attrs[] = {
6651 &raid5_stripecache_size.attr,
6652 &raid5_stripecache_active.attr,
6653 &raid5_preread_bypass_threshold.attr,
6654 &raid5_group_thread_cnt.attr,
6655 &raid5_skip_copy.attr,
6656 &raid5_rmw_level.attr,
6657 &r5c_journal_mode.attr,
6658 &ppl_write_hint.attr,
6661 static struct attribute_group raid5_attrs_group = {
6663 .attrs = raid5_attrs,
6666 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6668 int *worker_cnt_per_group,
6669 struct r5worker_group **worker_groups)
6673 struct r5worker *workers;
6675 *worker_cnt_per_group = cnt;
6678 *worker_groups = NULL;
6681 *group_cnt = num_possible_nodes();
6682 size = sizeof(struct r5worker) * cnt;
6683 workers = kcalloc(size, *group_cnt, GFP_NOIO);
6684 *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6686 if (!*worker_groups || !workers) {
6688 kfree(*worker_groups);
6692 for (i = 0; i < *group_cnt; i++) {
6693 struct r5worker_group *group;
6695 group = &(*worker_groups)[i];
6696 INIT_LIST_HEAD(&group->handle_list);
6697 INIT_LIST_HEAD(&group->loprio_list);
6699 group->workers = workers + i * cnt;
6701 for (j = 0; j < cnt; j++) {
6702 struct r5worker *worker = group->workers + j;
6703 worker->group = group;
6704 INIT_WORK(&worker->work, raid5_do_work);
6706 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6707 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6714 static void free_thread_groups(struct r5conf *conf)
6716 if (conf->worker_groups)
6717 kfree(conf->worker_groups[0].workers);
6718 kfree(conf->worker_groups);
6719 conf->worker_groups = NULL;
6723 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6725 struct r5conf *conf = mddev->private;
6728 sectors = mddev->dev_sectors;
6730 /* size is defined by the smallest of previous and new size */
6731 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6733 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6734 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6735 return sectors * (raid_disks - conf->max_degraded);
6738 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6740 safe_put_page(percpu->spare_page);
6741 percpu->spare_page = NULL;
6742 kvfree(percpu->scribble);
6743 percpu->scribble = NULL;
6746 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6748 if (conf->level == 6 && !percpu->spare_page) {
6749 percpu->spare_page = alloc_page(GFP_KERNEL);
6750 if (!percpu->spare_page)
6754 if (scribble_alloc(percpu,
6755 max(conf->raid_disks,
6756 conf->previous_raid_disks),
6757 max(conf->chunk_sectors,
6758 conf->prev_chunk_sectors)
6761 free_scratch_buffer(conf, percpu);
6768 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6770 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6772 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6776 static void raid5_free_percpu(struct r5conf *conf)
6781 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6782 free_percpu(conf->percpu);
6785 static void free_conf(struct r5conf *conf)
6791 unregister_shrinker(&conf->shrinker);
6792 free_thread_groups(conf);
6793 shrink_stripes(conf);
6794 raid5_free_percpu(conf);
6795 for (i = 0; i < conf->pool_size; i++)
6796 if (conf->disks[i].extra_page)
6797 put_page(conf->disks[i].extra_page);
6799 bioset_exit(&conf->bio_split);
6800 kfree(conf->stripe_hashtbl);
6801 kfree(conf->pending_data);
6805 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6807 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6808 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6810 if (alloc_scratch_buffer(conf, percpu)) {
6811 pr_warn("%s: failed memory allocation for cpu%u\n",
6818 static int raid5_alloc_percpu(struct r5conf *conf)
6822 conf->percpu = alloc_percpu(struct raid5_percpu);
6826 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6828 conf->scribble_disks = max(conf->raid_disks,
6829 conf->previous_raid_disks);
6830 conf->scribble_sectors = max(conf->chunk_sectors,
6831 conf->prev_chunk_sectors);
6836 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6837 struct shrink_control *sc)
6839 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6840 unsigned long ret = SHRINK_STOP;
6842 if (mutex_trylock(&conf->cache_size_mutex)) {
6844 while (ret < sc->nr_to_scan &&
6845 conf->max_nr_stripes > conf->min_nr_stripes) {
6846 if (drop_one_stripe(conf) == 0) {
6852 mutex_unlock(&conf->cache_size_mutex);
6857 static unsigned long raid5_cache_count(struct shrinker *shrink,
6858 struct shrink_control *sc)
6860 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6862 if (conf->max_nr_stripes < conf->min_nr_stripes)
6863 /* unlikely, but not impossible */
6865 return conf->max_nr_stripes - conf->min_nr_stripes;
6868 static struct r5conf *setup_conf(struct mddev *mddev)
6870 struct r5conf *conf;
6871 int raid_disk, memory, max_disks;
6872 struct md_rdev *rdev;
6873 struct disk_info *disk;
6876 int group_cnt, worker_cnt_per_group;
6877 struct r5worker_group *new_group;
6880 if (mddev->new_level != 5
6881 && mddev->new_level != 4
6882 && mddev->new_level != 6) {
6883 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6884 mdname(mddev), mddev->new_level);
6885 return ERR_PTR(-EIO);
6887 if ((mddev->new_level == 5
6888 && !algorithm_valid_raid5(mddev->new_layout)) ||
6889 (mddev->new_level == 6
6890 && !algorithm_valid_raid6(mddev->new_layout))) {
6891 pr_warn("md/raid:%s: layout %d not supported\n",
6892 mdname(mddev), mddev->new_layout);
6893 return ERR_PTR(-EIO);
6895 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6896 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6897 mdname(mddev), mddev->raid_disks);
6898 return ERR_PTR(-EINVAL);
6901 if (!mddev->new_chunk_sectors ||
6902 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6903 !is_power_of_2(mddev->new_chunk_sectors)) {
6904 pr_warn("md/raid:%s: invalid chunk size %d\n",
6905 mdname(mddev), mddev->new_chunk_sectors << 9);
6906 return ERR_PTR(-EINVAL);
6909 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6912 INIT_LIST_HEAD(&conf->free_list);
6913 INIT_LIST_HEAD(&conf->pending_list);
6914 conf->pending_data = kcalloc(PENDING_IO_MAX,
6915 sizeof(struct r5pending_data),
6917 if (!conf->pending_data)
6919 for (i = 0; i < PENDING_IO_MAX; i++)
6920 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6921 /* Don't enable multi-threading by default*/
6922 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6924 conf->group_cnt = group_cnt;
6925 conf->worker_cnt_per_group = worker_cnt_per_group;
6926 conf->worker_groups = new_group;
6929 spin_lock_init(&conf->device_lock);
6930 seqcount_init(&conf->gen_lock);
6931 mutex_init(&conf->cache_size_mutex);
6932 init_waitqueue_head(&conf->wait_for_quiescent);
6933 init_waitqueue_head(&conf->wait_for_stripe);
6934 init_waitqueue_head(&conf->wait_for_overlap);
6935 INIT_LIST_HEAD(&conf->handle_list);
6936 INIT_LIST_HEAD(&conf->loprio_list);
6937 INIT_LIST_HEAD(&conf->hold_list);
6938 INIT_LIST_HEAD(&conf->delayed_list);
6939 INIT_LIST_HEAD(&conf->bitmap_list);
6940 init_llist_head(&conf->released_stripes);
6941 atomic_set(&conf->active_stripes, 0);
6942 atomic_set(&conf->preread_active_stripes, 0);
6943 atomic_set(&conf->active_aligned_reads, 0);
6944 spin_lock_init(&conf->pending_bios_lock);
6945 conf->batch_bio_dispatch = true;
6946 rdev_for_each(rdev, mddev) {
6947 if (test_bit(Journal, &rdev->flags))
6949 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6950 conf->batch_bio_dispatch = false;
6955 conf->bypass_threshold = BYPASS_THRESHOLD;
6956 conf->recovery_disabled = mddev->recovery_disabled - 1;
6958 conf->raid_disks = mddev->raid_disks;
6959 if (mddev->reshape_position == MaxSector)
6960 conf->previous_raid_disks = mddev->raid_disks;
6962 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6963 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6965 conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
6971 for (i = 0; i < max_disks; i++) {
6972 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6973 if (!conf->disks[i].extra_page)
6977 ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
6980 conf->mddev = mddev;
6982 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6985 /* We init hash_locks[0] separately to that it can be used
6986 * as the reference lock in the spin_lock_nest_lock() call
6987 * in lock_all_device_hash_locks_irq in order to convince
6988 * lockdep that we know what we are doing.
6990 spin_lock_init(conf->hash_locks);
6991 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6992 spin_lock_init(conf->hash_locks + i);
6994 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6995 INIT_LIST_HEAD(conf->inactive_list + i);
6997 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6998 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7000 atomic_set(&conf->r5c_cached_full_stripes, 0);
7001 INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7002 atomic_set(&conf->r5c_cached_partial_stripes, 0);
7003 INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7004 atomic_set(&conf->r5c_flushing_full_stripes, 0);
7005 atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7007 conf->level = mddev->new_level;
7008 conf->chunk_sectors = mddev->new_chunk_sectors;
7009 if (raid5_alloc_percpu(conf) != 0)
7012 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7014 rdev_for_each(rdev, mddev) {
7015 raid_disk = rdev->raid_disk;
7016 if (raid_disk >= max_disks
7017 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7019 disk = conf->disks + raid_disk;
7021 if (test_bit(Replacement, &rdev->flags)) {
7022 if (disk->replacement)
7024 disk->replacement = rdev;
7031 if (test_bit(In_sync, &rdev->flags)) {
7032 char b[BDEVNAME_SIZE];
7033 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7034 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7035 } else if (rdev->saved_raid_disk != raid_disk)
7036 /* Cannot rely on bitmap to complete recovery */
7040 conf->level = mddev->new_level;
7041 if (conf->level == 6) {
7042 conf->max_degraded = 2;
7043 if (raid6_call.xor_syndrome)
7044 conf->rmw_level = PARITY_ENABLE_RMW;
7046 conf->rmw_level = PARITY_DISABLE_RMW;
7048 conf->max_degraded = 1;
7049 conf->rmw_level = PARITY_ENABLE_RMW;
7051 conf->algorithm = mddev->new_layout;
7052 conf->reshape_progress = mddev->reshape_position;
7053 if (conf->reshape_progress != MaxSector) {
7054 conf->prev_chunk_sectors = mddev->chunk_sectors;
7055 conf->prev_algo = mddev->layout;
7057 conf->prev_chunk_sectors = conf->chunk_sectors;
7058 conf->prev_algo = conf->algorithm;
7061 conf->min_nr_stripes = NR_STRIPES;
7062 if (mddev->reshape_position != MaxSector) {
7063 int stripes = max_t(int,
7064 ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7065 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7066 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7067 if (conf->min_nr_stripes != NR_STRIPES)
7068 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7069 mdname(mddev), conf->min_nr_stripes);
7071 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7072 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7073 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7074 if (grow_stripes(conf, conf->min_nr_stripes)) {
7075 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7076 mdname(mddev), memory);
7079 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7081 * Losing a stripe head costs more than the time to refill it,
7082 * it reduces the queue depth and so can hurt throughput.
7083 * So set it rather large, scaled by number of devices.
7085 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7086 conf->shrinker.scan_objects = raid5_cache_scan;
7087 conf->shrinker.count_objects = raid5_cache_count;
7088 conf->shrinker.batch = 128;
7089 conf->shrinker.flags = 0;
7090 if (register_shrinker(&conf->shrinker)) {
7091 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7096 sprintf(pers_name, "raid%d", mddev->new_level);
7097 conf->thread = md_register_thread(raid5d, mddev, pers_name);
7098 if (!conf->thread) {
7099 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7109 return ERR_PTR(-EIO);
7111 return ERR_PTR(-ENOMEM);
7114 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7117 case ALGORITHM_PARITY_0:
7118 if (raid_disk < max_degraded)
7121 case ALGORITHM_PARITY_N:
7122 if (raid_disk >= raid_disks - max_degraded)
7125 case ALGORITHM_PARITY_0_6:
7126 if (raid_disk == 0 ||
7127 raid_disk == raid_disks - 1)
7130 case ALGORITHM_LEFT_ASYMMETRIC_6:
7131 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7132 case ALGORITHM_LEFT_SYMMETRIC_6:
7133 case ALGORITHM_RIGHT_SYMMETRIC_6:
7134 if (raid_disk == raid_disks - 1)
7140 static int raid5_run(struct mddev *mddev)
7142 struct r5conf *conf;
7143 int working_disks = 0;
7144 int dirty_parity_disks = 0;
7145 struct md_rdev *rdev;
7146 struct md_rdev *journal_dev = NULL;
7147 sector_t reshape_offset = 0;
7149 long long min_offset_diff = 0;
7152 if (mddev_init_writes_pending(mddev) < 0)
7155 if (mddev->recovery_cp != MaxSector)
7156 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7159 rdev_for_each(rdev, mddev) {
7162 if (test_bit(Journal, &rdev->flags)) {
7166 if (rdev->raid_disk < 0)
7168 diff = (rdev->new_data_offset - rdev->data_offset);
7170 min_offset_diff = diff;
7172 } else if (mddev->reshape_backwards &&
7173 diff < min_offset_diff)
7174 min_offset_diff = diff;
7175 else if (!mddev->reshape_backwards &&
7176 diff > min_offset_diff)
7177 min_offset_diff = diff;
7180 if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7181 (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7182 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7187 if (mddev->reshape_position != MaxSector) {
7188 /* Check that we can continue the reshape.
7189 * Difficulties arise if the stripe we would write to
7190 * next is at or after the stripe we would read from next.
7191 * For a reshape that changes the number of devices, this
7192 * is only possible for a very short time, and mdadm makes
7193 * sure that time appears to have past before assembling
7194 * the array. So we fail if that time hasn't passed.
7195 * For a reshape that keeps the number of devices the same
7196 * mdadm must be monitoring the reshape can keeping the
7197 * critical areas read-only and backed up. It will start
7198 * the array in read-only mode, so we check for that.
7200 sector_t here_new, here_old;
7202 int max_degraded = (mddev->level == 6 ? 2 : 1);
7207 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7212 if (mddev->new_level != mddev->level) {
7213 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7217 old_disks = mddev->raid_disks - mddev->delta_disks;
7218 /* reshape_position must be on a new-stripe boundary, and one
7219 * further up in new geometry must map after here in old
7221 * If the chunk sizes are different, then as we perform reshape
7222 * in units of the largest of the two, reshape_position needs
7223 * be a multiple of the largest chunk size times new data disks.
7225 here_new = mddev->reshape_position;
7226 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7227 new_data_disks = mddev->raid_disks - max_degraded;
7228 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7229 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7233 reshape_offset = here_new * chunk_sectors;
7234 /* here_new is the stripe we will write to */
7235 here_old = mddev->reshape_position;
7236 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7237 /* here_old is the first stripe that we might need to read
7239 if (mddev->delta_disks == 0) {
7240 /* We cannot be sure it is safe to start an in-place
7241 * reshape. It is only safe if user-space is monitoring
7242 * and taking constant backups.
7243 * mdadm always starts a situation like this in
7244 * readonly mode so it can take control before
7245 * allowing any writes. So just check for that.
7247 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7248 abs(min_offset_diff) >= mddev->new_chunk_sectors)
7249 /* not really in-place - so OK */;
7250 else if (mddev->ro == 0) {
7251 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7255 } else if (mddev->reshape_backwards
7256 ? (here_new * chunk_sectors + min_offset_diff <=
7257 here_old * chunk_sectors)
7258 : (here_new * chunk_sectors >=
7259 here_old * chunk_sectors + (-min_offset_diff))) {
7260 /* Reading from the same stripe as writing to - bad */
7261 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7265 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7266 /* OK, we should be able to continue; */
7268 BUG_ON(mddev->level != mddev->new_level);
7269 BUG_ON(mddev->layout != mddev->new_layout);
7270 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7271 BUG_ON(mddev->delta_disks != 0);
7274 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7275 test_bit(MD_HAS_PPL, &mddev->flags)) {
7276 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7278 clear_bit(MD_HAS_PPL, &mddev->flags);
7279 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7282 if (mddev->private == NULL)
7283 conf = setup_conf(mddev);
7285 conf = mddev->private;
7288 return PTR_ERR(conf);
7290 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7292 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7295 set_disk_ro(mddev->gendisk, 1);
7296 } else if (mddev->recovery_cp == MaxSector)
7297 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7300 conf->min_offset_diff = min_offset_diff;
7301 mddev->thread = conf->thread;
7302 conf->thread = NULL;
7303 mddev->private = conf;
7305 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7307 rdev = conf->disks[i].rdev;
7308 if (!rdev && conf->disks[i].replacement) {
7309 /* The replacement is all we have yet */
7310 rdev = conf->disks[i].replacement;
7311 conf->disks[i].replacement = NULL;
7312 clear_bit(Replacement, &rdev->flags);
7313 conf->disks[i].rdev = rdev;
7317 if (conf->disks[i].replacement &&
7318 conf->reshape_progress != MaxSector) {
7319 /* replacements and reshape simply do not mix. */
7320 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7323 if (test_bit(In_sync, &rdev->flags)) {
7327 /* This disc is not fully in-sync. However if it
7328 * just stored parity (beyond the recovery_offset),
7329 * when we don't need to be concerned about the
7330 * array being dirty.
7331 * When reshape goes 'backwards', we never have
7332 * partially completed devices, so we only need
7333 * to worry about reshape going forwards.
7335 /* Hack because v0.91 doesn't store recovery_offset properly. */
7336 if (mddev->major_version == 0 &&
7337 mddev->minor_version > 90)
7338 rdev->recovery_offset = reshape_offset;
7340 if (rdev->recovery_offset < reshape_offset) {
7341 /* We need to check old and new layout */
7342 if (!only_parity(rdev->raid_disk,
7345 conf->max_degraded))
7348 if (!only_parity(rdev->raid_disk,
7350 conf->previous_raid_disks,
7351 conf->max_degraded))
7353 dirty_parity_disks++;
7357 * 0 for a fully functional array, 1 or 2 for a degraded array.
7359 mddev->degraded = raid5_calc_degraded(conf);
7361 if (has_failed(conf)) {
7362 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7363 mdname(mddev), mddev->degraded, conf->raid_disks);
7367 /* device size must be a multiple of chunk size */
7368 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7369 mddev->resync_max_sectors = mddev->dev_sectors;
7371 if (mddev->degraded > dirty_parity_disks &&
7372 mddev->recovery_cp != MaxSector) {
7373 if (test_bit(MD_HAS_PPL, &mddev->flags))
7374 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7376 else if (mddev->ok_start_degraded)
7377 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7380 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7386 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7387 mdname(mddev), conf->level,
7388 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7391 print_raid5_conf(conf);
7393 if (conf->reshape_progress != MaxSector) {
7394 conf->reshape_safe = conf->reshape_progress;
7395 atomic_set(&conf->reshape_stripes, 0);
7396 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7397 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7398 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7399 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7400 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7402 if (!mddev->sync_thread)
7406 /* Ok, everything is just fine now */
7407 if (mddev->to_remove == &raid5_attrs_group)
7408 mddev->to_remove = NULL;
7409 else if (mddev->kobj.sd &&
7410 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7411 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7413 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7417 /* read-ahead size must cover two whole stripes, which
7418 * is 2 * (datadisks) * chunksize where 'n' is the
7419 * number of raid devices
7421 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7422 int stripe = data_disks *
7423 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7424 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7425 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7427 chunk_size = mddev->chunk_sectors << 9;
7428 blk_queue_io_min(mddev->queue, chunk_size);
7429 blk_queue_io_opt(mddev->queue, chunk_size *
7430 (conf->raid_disks - conf->max_degraded));
7431 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7433 * We can only discard a whole stripe. It doesn't make sense to
7434 * discard data disk but write parity disk
7436 stripe = stripe * PAGE_SIZE;
7437 /* Round up to power of 2, as discard handling
7438 * currently assumes that */
7439 while ((stripe-1) & stripe)
7440 stripe = (stripe | (stripe-1)) + 1;
7441 mddev->queue->limits.discard_alignment = stripe;
7442 mddev->queue->limits.discard_granularity = stripe;
7444 blk_queue_max_write_same_sectors(mddev->queue, 0);
7445 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7447 rdev_for_each(rdev, mddev) {
7448 disk_stack_limits(mddev->gendisk, rdev->bdev,
7449 rdev->data_offset << 9);
7450 disk_stack_limits(mddev->gendisk, rdev->bdev,
7451 rdev->new_data_offset << 9);
7455 * zeroing is required, otherwise data
7456 * could be lost. Consider a scenario: discard a stripe
7457 * (the stripe could be inconsistent if
7458 * discard_zeroes_data is 0); write one disk of the
7459 * stripe (the stripe could be inconsistent again
7460 * depending on which disks are used to calculate
7461 * parity); the disk is broken; The stripe data of this
7464 * We only allow DISCARD if the sysadmin has confirmed that
7465 * only safe devices are in use by setting a module parameter.
7466 * A better idea might be to turn DISCARD into WRITE_ZEROES
7467 * requests, as that is required to be safe.
7469 if (devices_handle_discard_safely &&
7470 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7471 mddev->queue->limits.discard_granularity >= stripe)
7472 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7475 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7478 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7481 if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7486 md_unregister_thread(&mddev->thread);
7487 print_raid5_conf(conf);
7489 mddev->private = NULL;
7490 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7494 static void raid5_free(struct mddev *mddev, void *priv)
7496 struct r5conf *conf = priv;
7499 mddev->to_remove = &raid5_attrs_group;
7502 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7504 struct r5conf *conf = mddev->private;
7507 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7508 conf->chunk_sectors / 2, mddev->layout);
7509 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7511 for (i = 0; i < conf->raid_disks; i++) {
7512 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7513 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7516 seq_printf (seq, "]");
7519 static void print_raid5_conf (struct r5conf *conf)
7522 struct disk_info *tmp;
7524 pr_debug("RAID conf printout:\n");
7526 pr_debug("(conf==NULL)\n");
7529 pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7531 conf->raid_disks - conf->mddev->degraded);
7533 for (i = 0; i < conf->raid_disks; i++) {
7534 char b[BDEVNAME_SIZE];
7535 tmp = conf->disks + i;
7537 pr_debug(" disk %d, o:%d, dev:%s\n",
7538 i, !test_bit(Faulty, &tmp->rdev->flags),
7539 bdevname(tmp->rdev->bdev, b));
7543 static int raid5_spare_active(struct mddev *mddev)
7546 struct r5conf *conf = mddev->private;
7547 struct disk_info *tmp;
7549 unsigned long flags;
7551 for (i = 0; i < conf->raid_disks; i++) {
7552 tmp = conf->disks + i;
7553 if (tmp->replacement
7554 && tmp->replacement->recovery_offset == MaxSector
7555 && !test_bit(Faulty, &tmp->replacement->flags)
7556 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7557 /* Replacement has just become active. */
7559 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7562 /* Replaced device not technically faulty,
7563 * but we need to be sure it gets removed
7564 * and never re-added.
7566 set_bit(Faulty, &tmp->rdev->flags);
7567 sysfs_notify_dirent_safe(
7568 tmp->rdev->sysfs_state);
7570 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7571 } else if (tmp->rdev
7572 && tmp->rdev->recovery_offset == MaxSector
7573 && !test_bit(Faulty, &tmp->rdev->flags)
7574 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7576 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7579 spin_lock_irqsave(&conf->device_lock, flags);
7580 mddev->degraded = raid5_calc_degraded(conf);
7581 spin_unlock_irqrestore(&conf->device_lock, flags);
7582 print_raid5_conf(conf);
7586 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7588 struct r5conf *conf = mddev->private;
7590 int number = rdev->raid_disk;
7591 struct md_rdev **rdevp;
7592 struct disk_info *p = conf->disks + number;
7594 print_raid5_conf(conf);
7595 if (test_bit(Journal, &rdev->flags) && conf->log) {
7597 * we can't wait pending write here, as this is called in
7598 * raid5d, wait will deadlock.
7599 * neilb: there is no locking about new writes here,
7600 * so this cannot be safe.
7602 if (atomic_read(&conf->active_stripes) ||
7603 atomic_read(&conf->r5c_cached_full_stripes) ||
7604 atomic_read(&conf->r5c_cached_partial_stripes)) {
7610 if (rdev == p->rdev)
7612 else if (rdev == p->replacement)
7613 rdevp = &p->replacement;
7617 if (number >= conf->raid_disks &&
7618 conf->reshape_progress == MaxSector)
7619 clear_bit(In_sync, &rdev->flags);
7621 if (test_bit(In_sync, &rdev->flags) ||
7622 atomic_read(&rdev->nr_pending)) {
7626 /* Only remove non-faulty devices if recovery
7629 if (!test_bit(Faulty, &rdev->flags) &&
7630 mddev->recovery_disabled != conf->recovery_disabled &&
7631 !has_failed(conf) &&
7632 (!p->replacement || p->replacement == rdev) &&
7633 number < conf->raid_disks) {
7638 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7640 if (atomic_read(&rdev->nr_pending)) {
7641 /* lost the race, try later */
7647 err = log_modify(conf, rdev, false);
7651 if (p->replacement) {
7652 /* We must have just cleared 'rdev' */
7653 p->rdev = p->replacement;
7654 clear_bit(Replacement, &p->replacement->flags);
7655 smp_mb(); /* Make sure other CPUs may see both as identical
7656 * but will never see neither - if they are careful
7658 p->replacement = NULL;
7661 err = log_modify(conf, p->rdev, true);
7664 clear_bit(WantReplacement, &rdev->flags);
7667 print_raid5_conf(conf);
7671 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7673 struct r5conf *conf = mddev->private;
7674 int ret, err = -EEXIST;
7676 struct disk_info *p;
7678 int last = conf->raid_disks - 1;
7680 if (test_bit(Journal, &rdev->flags)) {
7684 rdev->raid_disk = 0;
7686 * The array is in readonly mode if journal is missing, so no
7687 * write requests running. We should be safe
7689 ret = log_init(conf, rdev, false);
7693 ret = r5l_start(conf->log);
7699 if (mddev->recovery_disabled == conf->recovery_disabled)
7702 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7703 /* no point adding a device */
7706 if (rdev->raid_disk >= 0)
7707 first = last = rdev->raid_disk;
7710 * find the disk ... but prefer rdev->saved_raid_disk
7713 if (rdev->saved_raid_disk >= 0 &&
7714 rdev->saved_raid_disk >= first &&
7715 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7716 first = rdev->saved_raid_disk;
7718 for (disk = first; disk <= last; disk++) {
7719 p = conf->disks + disk;
7720 if (p->rdev == NULL) {
7721 clear_bit(In_sync, &rdev->flags);
7722 rdev->raid_disk = disk;
7723 if (rdev->saved_raid_disk != disk)
7725 rcu_assign_pointer(p->rdev, rdev);
7727 err = log_modify(conf, rdev, true);
7732 for (disk = first; disk <= last; disk++) {
7733 p = conf->disks + disk;
7734 if (test_bit(WantReplacement, &p->rdev->flags) &&
7735 p->replacement == NULL) {
7736 clear_bit(In_sync, &rdev->flags);
7737 set_bit(Replacement, &rdev->flags);
7738 rdev->raid_disk = disk;
7741 rcu_assign_pointer(p->replacement, rdev);
7746 print_raid5_conf(conf);
7750 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7752 /* no resync is happening, and there is enough space
7753 * on all devices, so we can resize.
7754 * We need to make sure resync covers any new space.
7755 * If the array is shrinking we should possibly wait until
7756 * any io in the removed space completes, but it hardly seems
7760 struct r5conf *conf = mddev->private;
7762 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7764 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7765 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7766 if (mddev->external_size &&
7767 mddev->array_sectors > newsize)
7769 if (mddev->bitmap) {
7770 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
7774 md_set_array_sectors(mddev, newsize);
7775 if (sectors > mddev->dev_sectors &&
7776 mddev->recovery_cp > mddev->dev_sectors) {
7777 mddev->recovery_cp = mddev->dev_sectors;
7778 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7780 mddev->dev_sectors = sectors;
7781 mddev->resync_max_sectors = sectors;
7785 static int check_stripe_cache(struct mddev *mddev)
7787 /* Can only proceed if there are plenty of stripe_heads.
7788 * We need a minimum of one full stripe,, and for sensible progress
7789 * it is best to have about 4 times that.
7790 * If we require 4 times, then the default 256 4K stripe_heads will
7791 * allow for chunk sizes up to 256K, which is probably OK.
7792 * If the chunk size is greater, user-space should request more
7793 * stripe_heads first.
7795 struct r5conf *conf = mddev->private;
7796 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7797 > conf->min_nr_stripes ||
7798 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7799 > conf->min_nr_stripes) {
7800 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7802 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7809 static int check_reshape(struct mddev *mddev)
7811 struct r5conf *conf = mddev->private;
7813 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7815 if (mddev->delta_disks == 0 &&
7816 mddev->new_layout == mddev->layout &&
7817 mddev->new_chunk_sectors == mddev->chunk_sectors)
7818 return 0; /* nothing to do */
7819 if (has_failed(conf))
7821 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7822 /* We might be able to shrink, but the devices must
7823 * be made bigger first.
7824 * For raid6, 4 is the minimum size.
7825 * Otherwise 2 is the minimum
7828 if (mddev->level == 6)
7830 if (mddev->raid_disks + mddev->delta_disks < min)
7834 if (!check_stripe_cache(mddev))
7837 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7838 mddev->delta_disks > 0)
7839 if (resize_chunks(conf,
7840 conf->previous_raid_disks
7841 + max(0, mddev->delta_disks),
7842 max(mddev->new_chunk_sectors,
7843 mddev->chunk_sectors)
7847 if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7848 return 0; /* never bother to shrink */
7849 return resize_stripes(conf, (conf->previous_raid_disks
7850 + mddev->delta_disks));
7853 static int raid5_start_reshape(struct mddev *mddev)
7855 struct r5conf *conf = mddev->private;
7856 struct md_rdev *rdev;
7858 unsigned long flags;
7860 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7863 if (!check_stripe_cache(mddev))
7866 if (has_failed(conf))
7869 rdev_for_each(rdev, mddev) {
7870 if (!test_bit(In_sync, &rdev->flags)
7871 && !test_bit(Faulty, &rdev->flags))
7875 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7876 /* Not enough devices even to make a degraded array
7881 /* Refuse to reduce size of the array. Any reductions in
7882 * array size must be through explicit setting of array_size
7885 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7886 < mddev->array_sectors) {
7887 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7892 atomic_set(&conf->reshape_stripes, 0);
7893 spin_lock_irq(&conf->device_lock);
7894 write_seqcount_begin(&conf->gen_lock);
7895 conf->previous_raid_disks = conf->raid_disks;
7896 conf->raid_disks += mddev->delta_disks;
7897 conf->prev_chunk_sectors = conf->chunk_sectors;
7898 conf->chunk_sectors = mddev->new_chunk_sectors;
7899 conf->prev_algo = conf->algorithm;
7900 conf->algorithm = mddev->new_layout;
7902 /* Code that selects data_offset needs to see the generation update
7903 * if reshape_progress has been set - so a memory barrier needed.
7906 if (mddev->reshape_backwards)
7907 conf->reshape_progress = raid5_size(mddev, 0, 0);
7909 conf->reshape_progress = 0;
7910 conf->reshape_safe = conf->reshape_progress;
7911 write_seqcount_end(&conf->gen_lock);
7912 spin_unlock_irq(&conf->device_lock);
7914 /* Now make sure any requests that proceeded on the assumption
7915 * the reshape wasn't running - like Discard or Read - have
7918 mddev_suspend(mddev);
7919 mddev_resume(mddev);
7921 /* Add some new drives, as many as will fit.
7922 * We know there are enough to make the newly sized array work.
7923 * Don't add devices if we are reducing the number of
7924 * devices in the array. This is because it is not possible
7925 * to correctly record the "partially reconstructed" state of
7926 * such devices during the reshape and confusion could result.
7928 if (mddev->delta_disks >= 0) {
7929 rdev_for_each(rdev, mddev)
7930 if (rdev->raid_disk < 0 &&
7931 !test_bit(Faulty, &rdev->flags)) {
7932 if (raid5_add_disk(mddev, rdev) == 0) {
7934 >= conf->previous_raid_disks)
7935 set_bit(In_sync, &rdev->flags);
7937 rdev->recovery_offset = 0;
7939 if (sysfs_link_rdev(mddev, rdev))
7940 /* Failure here is OK */;
7942 } else if (rdev->raid_disk >= conf->previous_raid_disks
7943 && !test_bit(Faulty, &rdev->flags)) {
7944 /* This is a spare that was manually added */
7945 set_bit(In_sync, &rdev->flags);
7948 /* When a reshape changes the number of devices,
7949 * ->degraded is measured against the larger of the
7950 * pre and post number of devices.
7952 spin_lock_irqsave(&conf->device_lock, flags);
7953 mddev->degraded = raid5_calc_degraded(conf);
7954 spin_unlock_irqrestore(&conf->device_lock, flags);
7956 mddev->raid_disks = conf->raid_disks;
7957 mddev->reshape_position = conf->reshape_progress;
7958 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7960 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7961 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7962 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7963 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7964 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7965 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7967 if (!mddev->sync_thread) {
7968 mddev->recovery = 0;
7969 spin_lock_irq(&conf->device_lock);
7970 write_seqcount_begin(&conf->gen_lock);
7971 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7972 mddev->new_chunk_sectors =
7973 conf->chunk_sectors = conf->prev_chunk_sectors;
7974 mddev->new_layout = conf->algorithm = conf->prev_algo;
7975 rdev_for_each(rdev, mddev)
7976 rdev->new_data_offset = rdev->data_offset;
7978 conf->generation --;
7979 conf->reshape_progress = MaxSector;
7980 mddev->reshape_position = MaxSector;
7981 write_seqcount_end(&conf->gen_lock);
7982 spin_unlock_irq(&conf->device_lock);
7985 conf->reshape_checkpoint = jiffies;
7986 md_wakeup_thread(mddev->sync_thread);
7987 md_new_event(mddev);
7991 /* This is called from the reshape thread and should make any
7992 * changes needed in 'conf'
7994 static void end_reshape(struct r5conf *conf)
7997 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7998 struct md_rdev *rdev;
8000 spin_lock_irq(&conf->device_lock);
8001 conf->previous_raid_disks = conf->raid_disks;
8002 md_finish_reshape(conf->mddev);
8004 conf->reshape_progress = MaxSector;
8005 conf->mddev->reshape_position = MaxSector;
8006 rdev_for_each(rdev, conf->mddev)
8007 if (rdev->raid_disk >= 0 &&
8008 !test_bit(Journal, &rdev->flags) &&
8009 !test_bit(In_sync, &rdev->flags))
8010 rdev->recovery_offset = MaxSector;
8011 spin_unlock_irq(&conf->device_lock);
8012 wake_up(&conf->wait_for_overlap);
8014 /* read-ahead size must cover two whole stripes, which is
8015 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
8017 if (conf->mddev->queue) {
8018 int data_disks = conf->raid_disks - conf->max_degraded;
8019 int stripe = data_disks * ((conf->chunk_sectors << 9)
8021 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
8022 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
8027 /* This is called from the raid5d thread with mddev_lock held.
8028 * It makes config changes to the device.
8030 static void raid5_finish_reshape(struct mddev *mddev)
8032 struct r5conf *conf = mddev->private;
8034 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8036 if (mddev->delta_disks <= 0) {
8038 spin_lock_irq(&conf->device_lock);
8039 mddev->degraded = raid5_calc_degraded(conf);
8040 spin_unlock_irq(&conf->device_lock);
8041 for (d = conf->raid_disks ;
8042 d < conf->raid_disks - mddev->delta_disks;
8044 struct md_rdev *rdev = conf->disks[d].rdev;
8046 clear_bit(In_sync, &rdev->flags);
8047 rdev = conf->disks[d].replacement;
8049 clear_bit(In_sync, &rdev->flags);
8052 mddev->layout = conf->algorithm;
8053 mddev->chunk_sectors = conf->chunk_sectors;
8054 mddev->reshape_position = MaxSector;
8055 mddev->delta_disks = 0;
8056 mddev->reshape_backwards = 0;
8060 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8062 struct r5conf *conf = mddev->private;
8065 /* stop all writes */
8066 lock_all_device_hash_locks_irq(conf);
8067 /* '2' tells resync/reshape to pause so that all
8068 * active stripes can drain
8070 r5c_flush_cache(conf, INT_MAX);
8072 wait_event_cmd(conf->wait_for_quiescent,
8073 atomic_read(&conf->active_stripes) == 0 &&
8074 atomic_read(&conf->active_aligned_reads) == 0,
8075 unlock_all_device_hash_locks_irq(conf),
8076 lock_all_device_hash_locks_irq(conf));
8078 unlock_all_device_hash_locks_irq(conf);
8079 /* allow reshape to continue */
8080 wake_up(&conf->wait_for_overlap);
8082 /* re-enable writes */
8083 lock_all_device_hash_locks_irq(conf);
8085 wake_up(&conf->wait_for_quiescent);
8086 wake_up(&conf->wait_for_overlap);
8087 unlock_all_device_hash_locks_irq(conf);
8089 log_quiesce(conf, quiesce);
8092 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8094 struct r0conf *raid0_conf = mddev->private;
8097 /* for raid0 takeover only one zone is supported */
8098 if (raid0_conf->nr_strip_zones > 1) {
8099 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8101 return ERR_PTR(-EINVAL);
8104 sectors = raid0_conf->strip_zone[0].zone_end;
8105 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8106 mddev->dev_sectors = sectors;
8107 mddev->new_level = level;
8108 mddev->new_layout = ALGORITHM_PARITY_N;
8109 mddev->new_chunk_sectors = mddev->chunk_sectors;
8110 mddev->raid_disks += 1;
8111 mddev->delta_disks = 1;
8112 /* make sure it will be not marked as dirty */
8113 mddev->recovery_cp = MaxSector;
8115 return setup_conf(mddev);
8118 static void *raid5_takeover_raid1(struct mddev *mddev)
8123 if (mddev->raid_disks != 2 ||
8124 mddev->degraded > 1)
8125 return ERR_PTR(-EINVAL);
8127 /* Should check if there are write-behind devices? */
8129 chunksect = 64*2; /* 64K by default */
8131 /* The array must be an exact multiple of chunksize */
8132 while (chunksect && (mddev->array_sectors & (chunksect-1)))
8135 if ((chunksect<<9) < STRIPE_SIZE)
8136 /* array size does not allow a suitable chunk size */
8137 return ERR_PTR(-EINVAL);
8139 mddev->new_level = 5;
8140 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8141 mddev->new_chunk_sectors = chunksect;
8143 ret = setup_conf(mddev);
8145 mddev_clear_unsupported_flags(mddev,
8146 UNSUPPORTED_MDDEV_FLAGS);
8150 static void *raid5_takeover_raid6(struct mddev *mddev)
8154 switch (mddev->layout) {
8155 case ALGORITHM_LEFT_ASYMMETRIC_6:
8156 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8158 case ALGORITHM_RIGHT_ASYMMETRIC_6:
8159 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8161 case ALGORITHM_LEFT_SYMMETRIC_6:
8162 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8164 case ALGORITHM_RIGHT_SYMMETRIC_6:
8165 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8167 case ALGORITHM_PARITY_0_6:
8168 new_layout = ALGORITHM_PARITY_0;
8170 case ALGORITHM_PARITY_N:
8171 new_layout = ALGORITHM_PARITY_N;
8174 return ERR_PTR(-EINVAL);
8176 mddev->new_level = 5;
8177 mddev->new_layout = new_layout;
8178 mddev->delta_disks = -1;
8179 mddev->raid_disks -= 1;
8180 return setup_conf(mddev);
8183 static int raid5_check_reshape(struct mddev *mddev)
8185 /* For a 2-drive array, the layout and chunk size can be changed
8186 * immediately as not restriping is needed.
8187 * For larger arrays we record the new value - after validation
8188 * to be used by a reshape pass.
8190 struct r5conf *conf = mddev->private;
8191 int new_chunk = mddev->new_chunk_sectors;
8193 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8195 if (new_chunk > 0) {
8196 if (!is_power_of_2(new_chunk))
8198 if (new_chunk < (PAGE_SIZE>>9))
8200 if (mddev->array_sectors & (new_chunk-1))
8201 /* not factor of array size */
8205 /* They look valid */
8207 if (mddev->raid_disks == 2) {
8208 /* can make the change immediately */
8209 if (mddev->new_layout >= 0) {
8210 conf->algorithm = mddev->new_layout;
8211 mddev->layout = mddev->new_layout;
8213 if (new_chunk > 0) {
8214 conf->chunk_sectors = new_chunk ;
8215 mddev->chunk_sectors = new_chunk;
8217 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8218 md_wakeup_thread(mddev->thread);
8220 return check_reshape(mddev);
8223 static int raid6_check_reshape(struct mddev *mddev)
8225 int new_chunk = mddev->new_chunk_sectors;
8227 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8229 if (new_chunk > 0) {
8230 if (!is_power_of_2(new_chunk))
8232 if (new_chunk < (PAGE_SIZE >> 9))
8234 if (mddev->array_sectors & (new_chunk-1))
8235 /* not factor of array size */
8239 /* They look valid */
8240 return check_reshape(mddev);
8243 static void *raid5_takeover(struct mddev *mddev)
8245 /* raid5 can take over:
8246 * raid0 - if there is only one strip zone - make it a raid4 layout
8247 * raid1 - if there are two drives. We need to know the chunk size
8248 * raid4 - trivial - just use a raid4 layout.
8249 * raid6 - Providing it is a *_6 layout
8251 if (mddev->level == 0)
8252 return raid45_takeover_raid0(mddev, 5);
8253 if (mddev->level == 1)
8254 return raid5_takeover_raid1(mddev);
8255 if (mddev->level == 4) {
8256 mddev->new_layout = ALGORITHM_PARITY_N;
8257 mddev->new_level = 5;
8258 return setup_conf(mddev);
8260 if (mddev->level == 6)
8261 return raid5_takeover_raid6(mddev);
8263 return ERR_PTR(-EINVAL);
8266 static void *raid4_takeover(struct mddev *mddev)
8268 /* raid4 can take over:
8269 * raid0 - if there is only one strip zone
8270 * raid5 - if layout is right
8272 if (mddev->level == 0)
8273 return raid45_takeover_raid0(mddev, 4);
8274 if (mddev->level == 5 &&
8275 mddev->layout == ALGORITHM_PARITY_N) {
8276 mddev->new_layout = 0;
8277 mddev->new_level = 4;
8278 return setup_conf(mddev);
8280 return ERR_PTR(-EINVAL);
8283 static struct md_personality raid5_personality;
8285 static void *raid6_takeover(struct mddev *mddev)
8287 /* Currently can only take over a raid5. We map the
8288 * personality to an equivalent raid6 personality
8289 * with the Q block at the end.
8293 if (mddev->pers != &raid5_personality)
8294 return ERR_PTR(-EINVAL);
8295 if (mddev->degraded > 1)
8296 return ERR_PTR(-EINVAL);
8297 if (mddev->raid_disks > 253)
8298 return ERR_PTR(-EINVAL);
8299 if (mddev->raid_disks < 3)
8300 return ERR_PTR(-EINVAL);
8302 switch (mddev->layout) {
8303 case ALGORITHM_LEFT_ASYMMETRIC:
8304 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8306 case ALGORITHM_RIGHT_ASYMMETRIC:
8307 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8309 case ALGORITHM_LEFT_SYMMETRIC:
8310 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8312 case ALGORITHM_RIGHT_SYMMETRIC:
8313 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8315 case ALGORITHM_PARITY_0:
8316 new_layout = ALGORITHM_PARITY_0_6;
8318 case ALGORITHM_PARITY_N:
8319 new_layout = ALGORITHM_PARITY_N;
8322 return ERR_PTR(-EINVAL);
8324 mddev->new_level = 6;
8325 mddev->new_layout = new_layout;
8326 mddev->delta_disks = 1;
8327 mddev->raid_disks += 1;
8328 return setup_conf(mddev);
8331 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8333 struct r5conf *conf;
8336 err = mddev_lock(mddev);
8339 conf = mddev->private;
8341 mddev_unlock(mddev);
8345 if (strncmp(buf, "ppl", 3) == 0) {
8346 /* ppl only works with RAID 5 */
8347 if (!raid5_has_ppl(conf) && conf->level == 5) {
8348 err = log_init(conf, NULL, true);
8350 err = resize_stripes(conf, conf->pool_size);
8356 } else if (strncmp(buf, "resync", 6) == 0) {
8357 if (raid5_has_ppl(conf)) {
8358 mddev_suspend(mddev);
8360 mddev_resume(mddev);
8361 err = resize_stripes(conf, conf->pool_size);
8362 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8363 r5l_log_disk_error(conf)) {
8364 bool journal_dev_exists = false;
8365 struct md_rdev *rdev;
8367 rdev_for_each(rdev, mddev)
8368 if (test_bit(Journal, &rdev->flags)) {
8369 journal_dev_exists = true;
8373 if (!journal_dev_exists) {
8374 mddev_suspend(mddev);
8375 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8376 mddev_resume(mddev);
8377 } else /* need remove journal device first */
8386 md_update_sb(mddev, 1);
8388 mddev_unlock(mddev);
8393 static int raid5_start(struct mddev *mddev)
8395 struct r5conf *conf = mddev->private;
8397 return r5l_start(conf->log);
8400 static struct md_personality raid6_personality =
8404 .owner = THIS_MODULE,
8405 .make_request = raid5_make_request,
8407 .start = raid5_start,
8409 .status = raid5_status,
8410 .error_handler = raid5_error,
8411 .hot_add_disk = raid5_add_disk,
8412 .hot_remove_disk= raid5_remove_disk,
8413 .spare_active = raid5_spare_active,
8414 .sync_request = raid5_sync_request,
8415 .resize = raid5_resize,
8417 .check_reshape = raid6_check_reshape,
8418 .start_reshape = raid5_start_reshape,
8419 .finish_reshape = raid5_finish_reshape,
8420 .quiesce = raid5_quiesce,
8421 .takeover = raid6_takeover,
8422 .congested = raid5_congested,
8423 .change_consistency_policy = raid5_change_consistency_policy,
8425 static struct md_personality raid5_personality =
8429 .owner = THIS_MODULE,
8430 .make_request = raid5_make_request,
8432 .start = raid5_start,
8434 .status = raid5_status,
8435 .error_handler = raid5_error,
8436 .hot_add_disk = raid5_add_disk,
8437 .hot_remove_disk= raid5_remove_disk,
8438 .spare_active = raid5_spare_active,
8439 .sync_request = raid5_sync_request,
8440 .resize = raid5_resize,
8442 .check_reshape = raid5_check_reshape,
8443 .start_reshape = raid5_start_reshape,
8444 .finish_reshape = raid5_finish_reshape,
8445 .quiesce = raid5_quiesce,
8446 .takeover = raid5_takeover,
8447 .congested = raid5_congested,
8448 .change_consistency_policy = raid5_change_consistency_policy,
8451 static struct md_personality raid4_personality =
8455 .owner = THIS_MODULE,
8456 .make_request = raid5_make_request,
8458 .start = raid5_start,
8460 .status = raid5_status,
8461 .error_handler = raid5_error,
8462 .hot_add_disk = raid5_add_disk,
8463 .hot_remove_disk= raid5_remove_disk,
8464 .spare_active = raid5_spare_active,
8465 .sync_request = raid5_sync_request,
8466 .resize = raid5_resize,
8468 .check_reshape = raid5_check_reshape,
8469 .start_reshape = raid5_start_reshape,
8470 .finish_reshape = raid5_finish_reshape,
8471 .quiesce = raid5_quiesce,
8472 .takeover = raid4_takeover,
8473 .congested = raid5_congested,
8474 .change_consistency_policy = raid5_change_consistency_policy,
8477 static int __init raid5_init(void)
8481 raid5_wq = alloc_workqueue("raid5wq",
8482 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8486 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8488 raid456_cpu_up_prepare,
8491 destroy_workqueue(raid5_wq);
8494 register_md_personality(&raid6_personality);
8495 register_md_personality(&raid5_personality);
8496 register_md_personality(&raid4_personality);
8500 static void raid5_exit(void)
8502 unregister_md_personality(&raid6_personality);
8503 unregister_md_personality(&raid5_personality);
8504 unregister_md_personality(&raid4_personality);
8505 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8506 destroy_workqueue(raid5_wq);
8509 module_init(raid5_init);
8510 module_exit(raid5_exit);
8511 MODULE_LICENSE("GPL");
8512 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8513 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8514 MODULE_ALIAS("md-raid5");
8515 MODULE_ALIAS("md-raid4");
8516 MODULE_ALIAS("md-level-5");
8517 MODULE_ALIAS("md-level-4");
8518 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8519 MODULE_ALIAS("md-raid6");
8520 MODULE_ALIAS("md-level-6");
8522 /* This used to be two separate modules, they were: */
8523 MODULE_ALIAS("raid5");
8524 MODULE_ALIAS("raid6");