2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
58 #include <trace/events/block.h>
65 #define cpu_to_group(cpu) cpu_to_node(cpu)
66 #define ANY_GROUP NUMA_NO_NODE
68 static bool devices_handle_discard_safely = false;
69 module_param(devices_handle_discard_safely, bool, 0644);
70 MODULE_PARM_DESC(devices_handle_discard_safely,
71 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
72 static struct workqueue_struct *raid5_wq;
77 #define NR_STRIPES 256
78 #define STRIPE_SIZE PAGE_SIZE
79 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
80 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
81 #define IO_THRESHOLD 1
82 #define BYPASS_THRESHOLD 1
83 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
84 #define HASH_MASK (NR_HASH - 1)
85 #define MAX_STRIPE_BATCH 8
87 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
89 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
90 return &conf->stripe_hashtbl[hash];
93 static inline int stripe_hash_locks_hash(sector_t sect)
95 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
98 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
100 spin_lock_irq(conf->hash_locks + hash);
101 spin_lock(&conf->device_lock);
104 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
106 spin_unlock(&conf->device_lock);
107 spin_unlock_irq(conf->hash_locks + hash);
110 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
114 spin_lock(conf->hash_locks);
115 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
116 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
117 spin_lock(&conf->device_lock);
120 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
123 spin_unlock(&conf->device_lock);
124 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
125 spin_unlock(conf->hash_locks + i - 1);
129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
130 * order without overlap. There may be several bio's per stripe+device, and
131 * a bio could span several devices.
132 * When walking this list for a particular stripe+device, we must never proceed
133 * beyond a bio that extends past this device, as the next bio might no longer
135 * This function is used to determine the 'next' bio in the list, given the sector
136 * of the current stripe+device
138 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
140 int sectors = bio_sectors(bio);
141 if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
148 * We maintain a biased count of active stripes in the bottom 16 bits of
149 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
151 static inline int raid5_bi_processed_stripes(struct bio *bio)
153 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
154 return (atomic_read(segments) >> 16) & 0xffff;
157 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
159 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
160 return atomic_sub_return(1, segments) & 0xffff;
163 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
165 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
166 atomic_inc(segments);
169 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
172 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
176 old = atomic_read(segments);
177 new = (old & 0xffff) | (cnt << 16);
178 } while (atomic_cmpxchg(segments, old, new) != old);
181 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
183 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
184 atomic_set(segments, cnt);
187 /* Find first data disk in a raid6 stripe */
188 static inline int raid6_d0(struct stripe_head *sh)
191 /* ddf always start from first device */
193 /* md starts just after Q block */
194 if (sh->qd_idx == sh->disks - 1)
197 return sh->qd_idx + 1;
199 static inline int raid6_next_disk(int disk, int raid_disks)
202 return (disk < raid_disks) ? disk : 0;
205 /* When walking through the disks in a raid5, starting at raid6_d0,
206 * We need to map each disk to a 'slot', where the data disks are slot
207 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
208 * is raid_disks-1. This help does that mapping.
210 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
211 int *count, int syndrome_disks)
217 if (idx == sh->pd_idx)
218 return syndrome_disks;
219 if (idx == sh->qd_idx)
220 return syndrome_disks + 1;
226 static void return_io(struct bio_list *return_bi)
229 while ((bi = bio_list_pop(return_bi)) != NULL) {
230 bi->bi_iter.bi_size = 0;
231 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
237 static void print_raid5_conf (struct r5conf *conf);
239 static int stripe_operations_active(struct stripe_head *sh)
241 return sh->check_state || sh->reconstruct_state ||
242 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
243 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
246 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
248 struct r5conf *conf = sh->raid_conf;
249 struct r5worker_group *group;
251 int i, cpu = sh->cpu;
253 if (!cpu_online(cpu)) {
254 cpu = cpumask_any(cpu_online_mask);
258 if (list_empty(&sh->lru)) {
259 struct r5worker_group *group;
260 group = conf->worker_groups + cpu_to_group(cpu);
261 list_add_tail(&sh->lru, &group->handle_list);
262 group->stripes_cnt++;
266 if (conf->worker_cnt_per_group == 0) {
267 md_wakeup_thread(conf->mddev->thread);
271 group = conf->worker_groups + cpu_to_group(sh->cpu);
273 group->workers[0].working = true;
274 /* at least one worker should run to avoid race */
275 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
277 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
278 /* wakeup more workers */
279 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
280 if (group->workers[i].working == false) {
281 group->workers[i].working = true;
282 queue_work_on(sh->cpu, raid5_wq,
283 &group->workers[i].work);
289 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
290 struct list_head *temp_inactive_list)
292 BUG_ON(!list_empty(&sh->lru));
293 BUG_ON(atomic_read(&conf->active_stripes)==0);
294 if (test_bit(STRIPE_HANDLE, &sh->state)) {
295 if (test_bit(STRIPE_DELAYED, &sh->state) &&
296 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
297 list_add_tail(&sh->lru, &conf->delayed_list);
298 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
299 sh->bm_seq - conf->seq_write > 0)
300 list_add_tail(&sh->lru, &conf->bitmap_list);
302 clear_bit(STRIPE_DELAYED, &sh->state);
303 clear_bit(STRIPE_BIT_DELAY, &sh->state);
304 if (conf->worker_cnt_per_group == 0) {
305 list_add_tail(&sh->lru, &conf->handle_list);
307 raid5_wakeup_stripe_thread(sh);
311 md_wakeup_thread(conf->mddev->thread);
313 BUG_ON(stripe_operations_active(sh));
314 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
315 if (atomic_dec_return(&conf->preread_active_stripes)
317 md_wakeup_thread(conf->mddev->thread);
318 atomic_dec(&conf->active_stripes);
319 if (!test_bit(STRIPE_EXPANDING, &sh->state))
320 list_add_tail(&sh->lru, temp_inactive_list);
324 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
325 struct list_head *temp_inactive_list)
327 if (atomic_dec_and_test(&sh->count))
328 do_release_stripe(conf, sh, temp_inactive_list);
332 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
334 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
335 * given time. Adding stripes only takes device lock, while deleting stripes
336 * only takes hash lock.
338 static void release_inactive_stripe_list(struct r5conf *conf,
339 struct list_head *temp_inactive_list,
343 bool do_wakeup = false;
346 if (hash == NR_STRIPE_HASH_LOCKS) {
347 size = NR_STRIPE_HASH_LOCKS;
348 hash = NR_STRIPE_HASH_LOCKS - 1;
352 struct list_head *list = &temp_inactive_list[size - 1];
355 * We don't hold any lock here yet, raid5_get_active_stripe() might
356 * remove stripes from the list
358 if (!list_empty_careful(list)) {
359 spin_lock_irqsave(conf->hash_locks + hash, flags);
360 if (list_empty(conf->inactive_list + hash) &&
362 atomic_dec(&conf->empty_inactive_list_nr);
363 list_splice_tail_init(list, conf->inactive_list + hash);
365 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
372 wake_up(&conf->wait_for_stripe);
373 if (atomic_read(&conf->active_stripes) == 0)
374 wake_up(&conf->wait_for_quiescent);
375 if (conf->retry_read_aligned)
376 md_wakeup_thread(conf->mddev->thread);
380 /* should hold conf->device_lock already */
381 static int release_stripe_list(struct r5conf *conf,
382 struct list_head *temp_inactive_list)
384 struct stripe_head *sh;
386 struct llist_node *head;
388 head = llist_del_all(&conf->released_stripes);
389 head = llist_reverse_order(head);
393 sh = llist_entry(head, struct stripe_head, release_list);
394 head = llist_next(head);
395 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
397 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
399 * Don't worry the bit is set here, because if the bit is set
400 * again, the count is always > 1. This is true for
401 * STRIPE_ON_UNPLUG_LIST bit too.
403 hash = sh->hash_lock_index;
404 __release_stripe(conf, sh, &temp_inactive_list[hash]);
411 void raid5_release_stripe(struct stripe_head *sh)
413 struct r5conf *conf = sh->raid_conf;
415 struct list_head list;
419 /* Avoid release_list until the last reference.
421 if (atomic_add_unless(&sh->count, -1, 1))
424 if (unlikely(!conf->mddev->thread) ||
425 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
427 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
429 md_wakeup_thread(conf->mddev->thread);
432 local_irq_save(flags);
433 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
434 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
435 INIT_LIST_HEAD(&list);
436 hash = sh->hash_lock_index;
437 do_release_stripe(conf, sh, &list);
438 spin_unlock(&conf->device_lock);
439 release_inactive_stripe_list(conf, &list, hash);
441 local_irq_restore(flags);
444 static inline void remove_hash(struct stripe_head *sh)
446 pr_debug("remove_hash(), stripe %llu\n",
447 (unsigned long long)sh->sector);
449 hlist_del_init(&sh->hash);
452 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
454 struct hlist_head *hp = stripe_hash(conf, sh->sector);
456 pr_debug("insert_hash(), stripe %llu\n",
457 (unsigned long long)sh->sector);
459 hlist_add_head(&sh->hash, hp);
462 /* find an idle stripe, make sure it is unhashed, and return it. */
463 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
465 struct stripe_head *sh = NULL;
466 struct list_head *first;
468 if (list_empty(conf->inactive_list + hash))
470 first = (conf->inactive_list + hash)->next;
471 sh = list_entry(first, struct stripe_head, lru);
472 list_del_init(first);
474 atomic_inc(&conf->active_stripes);
475 BUG_ON(hash != sh->hash_lock_index);
476 if (list_empty(conf->inactive_list + hash))
477 atomic_inc(&conf->empty_inactive_list_nr);
482 static void shrink_buffers(struct stripe_head *sh)
486 int num = sh->raid_conf->pool_size;
488 for (i = 0; i < num ; i++) {
489 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
493 sh->dev[i].page = NULL;
498 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
501 int num = sh->raid_conf->pool_size;
503 for (i = 0; i < num; i++) {
506 if (!(page = alloc_page(gfp))) {
509 sh->dev[i].page = page;
510 sh->dev[i].orig_page = page;
515 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
516 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
517 struct stripe_head *sh);
519 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
521 struct r5conf *conf = sh->raid_conf;
524 BUG_ON(atomic_read(&sh->count) != 0);
525 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
526 BUG_ON(stripe_operations_active(sh));
527 BUG_ON(sh->batch_head);
529 pr_debug("init_stripe called, stripe %llu\n",
530 (unsigned long long)sector);
532 seq = read_seqcount_begin(&conf->gen_lock);
533 sh->generation = conf->generation - previous;
534 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
536 stripe_set_idx(sector, conf, previous, sh);
539 for (i = sh->disks; i--; ) {
540 struct r5dev *dev = &sh->dev[i];
542 if (dev->toread || dev->read || dev->towrite || dev->written ||
543 test_bit(R5_LOCKED, &dev->flags)) {
544 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
545 (unsigned long long)sh->sector, i, dev->toread,
546 dev->read, dev->towrite, dev->written,
547 test_bit(R5_LOCKED, &dev->flags));
551 raid5_build_block(sh, i, previous);
553 if (read_seqcount_retry(&conf->gen_lock, seq))
555 sh->overwrite_disks = 0;
556 insert_hash(conf, sh);
557 sh->cpu = smp_processor_id();
558 set_bit(STRIPE_BATCH_READY, &sh->state);
561 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
564 struct stripe_head *sh;
566 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
567 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
568 if (sh->sector == sector && sh->generation == generation)
570 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
575 * Need to check if array has failed when deciding whether to:
577 * - remove non-faulty devices
580 * This determination is simple when no reshape is happening.
581 * However if there is a reshape, we need to carefully check
582 * both the before and after sections.
583 * This is because some failed devices may only affect one
584 * of the two sections, and some non-in_sync devices may
585 * be insync in the section most affected by failed devices.
587 static int calc_degraded(struct r5conf *conf)
589 int degraded, degraded2;
594 for (i = 0; i < conf->previous_raid_disks; i++) {
595 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
596 if (rdev && test_bit(Faulty, &rdev->flags))
597 rdev = rcu_dereference(conf->disks[i].replacement);
598 if (!rdev || test_bit(Faulty, &rdev->flags))
600 else if (test_bit(In_sync, &rdev->flags))
603 /* not in-sync or faulty.
604 * If the reshape increases the number of devices,
605 * this is being recovered by the reshape, so
606 * this 'previous' section is not in_sync.
607 * If the number of devices is being reduced however,
608 * the device can only be part of the array if
609 * we are reverting a reshape, so this section will
612 if (conf->raid_disks >= conf->previous_raid_disks)
616 if (conf->raid_disks == conf->previous_raid_disks)
620 for (i = 0; i < conf->raid_disks; i++) {
621 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
622 if (rdev && test_bit(Faulty, &rdev->flags))
623 rdev = rcu_dereference(conf->disks[i].replacement);
624 if (!rdev || test_bit(Faulty, &rdev->flags))
626 else if (test_bit(In_sync, &rdev->flags))
629 /* not in-sync or faulty.
630 * If reshape increases the number of devices, this
631 * section has already been recovered, else it
632 * almost certainly hasn't.
634 if (conf->raid_disks <= conf->previous_raid_disks)
638 if (degraded2 > degraded)
643 static int has_failed(struct r5conf *conf)
647 if (conf->mddev->reshape_position == MaxSector)
648 return conf->mddev->degraded > conf->max_degraded;
650 degraded = calc_degraded(conf);
651 if (degraded > conf->max_degraded)
657 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
658 int previous, int noblock, int noquiesce)
660 struct stripe_head *sh;
661 int hash = stripe_hash_locks_hash(sector);
662 int inc_empty_inactive_list_flag;
664 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
666 spin_lock_irq(conf->hash_locks + hash);
669 wait_event_lock_irq(conf->wait_for_quiescent,
670 conf->quiesce == 0 || noquiesce,
671 *(conf->hash_locks + hash));
672 sh = __find_stripe(conf, sector, conf->generation - previous);
674 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
675 sh = get_free_stripe(conf, hash);
676 if (!sh && !test_bit(R5_DID_ALLOC,
678 set_bit(R5_ALLOC_MORE,
681 if (noblock && sh == NULL)
684 set_bit(R5_INACTIVE_BLOCKED,
687 conf->wait_for_stripe,
688 !list_empty(conf->inactive_list + hash) &&
689 (atomic_read(&conf->active_stripes)
690 < (conf->max_nr_stripes * 3 / 4)
691 || !test_bit(R5_INACTIVE_BLOCKED,
692 &conf->cache_state)),
693 *(conf->hash_locks + hash));
694 clear_bit(R5_INACTIVE_BLOCKED,
697 init_stripe(sh, sector, previous);
698 atomic_inc(&sh->count);
700 } else if (!atomic_inc_not_zero(&sh->count)) {
701 spin_lock(&conf->device_lock);
702 if (!atomic_read(&sh->count)) {
703 if (!test_bit(STRIPE_HANDLE, &sh->state))
704 atomic_inc(&conf->active_stripes);
705 BUG_ON(list_empty(&sh->lru) &&
706 !test_bit(STRIPE_EXPANDING, &sh->state));
707 inc_empty_inactive_list_flag = 0;
708 if (!list_empty(conf->inactive_list + hash))
709 inc_empty_inactive_list_flag = 1;
710 list_del_init(&sh->lru);
711 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
712 atomic_inc(&conf->empty_inactive_list_nr);
714 sh->group->stripes_cnt--;
718 atomic_inc(&sh->count);
719 spin_unlock(&conf->device_lock);
721 } while (sh == NULL);
723 spin_unlock_irq(conf->hash_locks + hash);
727 static bool is_full_stripe_write(struct stripe_head *sh)
729 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
730 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
733 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
737 spin_lock(&sh2->stripe_lock);
738 spin_lock_nested(&sh1->stripe_lock, 1);
740 spin_lock(&sh1->stripe_lock);
741 spin_lock_nested(&sh2->stripe_lock, 1);
745 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
747 spin_unlock(&sh1->stripe_lock);
748 spin_unlock(&sh2->stripe_lock);
752 /* Only freshly new full stripe normal write stripe can be added to a batch list */
753 static bool stripe_can_batch(struct stripe_head *sh)
755 struct r5conf *conf = sh->raid_conf;
759 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
760 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
761 is_full_stripe_write(sh);
764 /* we only do back search */
765 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
767 struct stripe_head *head;
768 sector_t head_sector, tmp_sec;
771 int inc_empty_inactive_list_flag;
773 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
774 tmp_sec = sh->sector;
775 if (!sector_div(tmp_sec, conf->chunk_sectors))
777 head_sector = sh->sector - STRIPE_SECTORS;
779 hash = stripe_hash_locks_hash(head_sector);
780 spin_lock_irq(conf->hash_locks + hash);
781 head = __find_stripe(conf, head_sector, conf->generation);
782 if (head && !atomic_inc_not_zero(&head->count)) {
783 spin_lock(&conf->device_lock);
784 if (!atomic_read(&head->count)) {
785 if (!test_bit(STRIPE_HANDLE, &head->state))
786 atomic_inc(&conf->active_stripes);
787 BUG_ON(list_empty(&head->lru) &&
788 !test_bit(STRIPE_EXPANDING, &head->state));
789 inc_empty_inactive_list_flag = 0;
790 if (!list_empty(conf->inactive_list + hash))
791 inc_empty_inactive_list_flag = 1;
792 list_del_init(&head->lru);
793 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
794 atomic_inc(&conf->empty_inactive_list_nr);
796 head->group->stripes_cnt--;
800 atomic_inc(&head->count);
801 spin_unlock(&conf->device_lock);
803 spin_unlock_irq(conf->hash_locks + hash);
807 if (!stripe_can_batch(head))
810 lock_two_stripes(head, sh);
811 /* clear_batch_ready clear the flag */
812 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
819 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
821 if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
822 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
825 if (head->batch_head) {
826 spin_lock(&head->batch_head->batch_lock);
827 /* This batch list is already running */
828 if (!stripe_can_batch(head)) {
829 spin_unlock(&head->batch_head->batch_lock);
834 * at this point, head's BATCH_READY could be cleared, but we
835 * can still add the stripe to batch list
837 list_add(&sh->batch_list, &head->batch_list);
838 spin_unlock(&head->batch_head->batch_lock);
840 sh->batch_head = head->batch_head;
842 head->batch_head = head;
843 sh->batch_head = head->batch_head;
844 spin_lock(&head->batch_lock);
845 list_add_tail(&sh->batch_list, &head->batch_list);
846 spin_unlock(&head->batch_lock);
849 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
850 if (atomic_dec_return(&conf->preread_active_stripes)
852 md_wakeup_thread(conf->mddev->thread);
854 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
855 int seq = sh->bm_seq;
856 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
857 sh->batch_head->bm_seq > seq)
858 seq = sh->batch_head->bm_seq;
859 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
860 sh->batch_head->bm_seq = seq;
863 atomic_inc(&sh->count);
865 unlock_two_stripes(head, sh);
867 raid5_release_stripe(head);
870 /* Determine if 'data_offset' or 'new_data_offset' should be used
871 * in this stripe_head.
873 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
875 sector_t progress = conf->reshape_progress;
876 /* Need a memory barrier to make sure we see the value
877 * of conf->generation, or ->data_offset that was set before
878 * reshape_progress was updated.
881 if (progress == MaxSector)
883 if (sh->generation == conf->generation - 1)
885 /* We are in a reshape, and this is a new-generation stripe,
886 * so use new_data_offset.
892 raid5_end_read_request(struct bio *bi);
894 raid5_end_write_request(struct bio *bi);
896 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
898 struct r5conf *conf = sh->raid_conf;
899 int i, disks = sh->disks;
900 struct stripe_head *head_sh = sh;
904 if (r5l_write_stripe(conf->log, sh) == 0)
906 for (i = disks; i--; ) {
907 int op, op_flags = 0;
908 int replace_only = 0;
909 struct bio *bi, *rbi;
910 struct md_rdev *rdev, *rrdev = NULL;
913 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
915 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
916 op_flags = WRITE_FUA;
917 if (test_bit(R5_Discard, &sh->dev[i].flags))
919 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
921 else if (test_and_clear_bit(R5_WantReplace,
922 &sh->dev[i].flags)) {
927 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
928 op_flags |= REQ_SYNC;
931 bi = &sh->dev[i].req;
932 rbi = &sh->dev[i].rreq; /* For writing to replacement */
935 rrdev = rcu_dereference(conf->disks[i].replacement);
936 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
937 rdev = rcu_dereference(conf->disks[i].rdev);
942 if (op_is_write(op)) {
946 /* We raced and saw duplicates */
949 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
954 if (rdev && test_bit(Faulty, &rdev->flags))
957 atomic_inc(&rdev->nr_pending);
958 if (rrdev && test_bit(Faulty, &rrdev->flags))
961 atomic_inc(&rrdev->nr_pending);
964 /* We have already checked bad blocks for reads. Now
965 * need to check for writes. We never accept write errors
966 * on the replacement, so we don't to check rrdev.
968 while (op_is_write(op) && rdev &&
969 test_bit(WriteErrorSeen, &rdev->flags)) {
972 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
973 &first_bad, &bad_sectors);
978 set_bit(BlockedBadBlocks, &rdev->flags);
979 if (!conf->mddev->external &&
980 conf->mddev->flags) {
981 /* It is very unlikely, but we might
982 * still need to write out the
983 * bad block log - better give it
985 md_check_recovery(conf->mddev);
988 * Because md_wait_for_blocked_rdev
989 * will dec nr_pending, we must
990 * increment it first.
992 atomic_inc(&rdev->nr_pending);
993 md_wait_for_blocked_rdev(rdev, conf->mddev);
995 /* Acknowledged bad block - skip the write */
996 rdev_dec_pending(rdev, conf->mddev);
1002 if (s->syncing || s->expanding || s->expanded
1004 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1006 set_bit(STRIPE_IO_STARTED, &sh->state);
1008 bi->bi_bdev = rdev->bdev;
1009 bio_set_op_attrs(bi, op, op_flags);
1010 bi->bi_end_io = op_is_write(op)
1011 ? raid5_end_write_request
1012 : raid5_end_read_request;
1013 bi->bi_private = sh;
1015 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1016 __func__, (unsigned long long)sh->sector,
1018 atomic_inc(&sh->count);
1020 atomic_inc(&head_sh->count);
1021 if (use_new_offset(conf, sh))
1022 bi->bi_iter.bi_sector = (sh->sector
1023 + rdev->new_data_offset);
1025 bi->bi_iter.bi_sector = (sh->sector
1026 + rdev->data_offset);
1027 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1028 bi->bi_opf |= REQ_NOMERGE;
1030 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1031 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1032 sh->dev[i].vec.bv_page = sh->dev[i].page;
1034 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1035 bi->bi_io_vec[0].bv_offset = 0;
1036 bi->bi_iter.bi_size = STRIPE_SIZE;
1038 * If this is discard request, set bi_vcnt 0. We don't
1039 * want to confuse SCSI because SCSI will replace payload
1041 if (op == REQ_OP_DISCARD)
1044 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1046 if (conf->mddev->gendisk)
1047 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1048 bi, disk_devt(conf->mddev->gendisk),
1050 generic_make_request(bi);
1053 if (s->syncing || s->expanding || s->expanded
1055 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1057 set_bit(STRIPE_IO_STARTED, &sh->state);
1059 rbi->bi_bdev = rrdev->bdev;
1060 bio_set_op_attrs(rbi, op, op_flags);
1061 BUG_ON(!op_is_write(op));
1062 rbi->bi_end_io = raid5_end_write_request;
1063 rbi->bi_private = sh;
1065 pr_debug("%s: for %llu schedule op %d on "
1066 "replacement disc %d\n",
1067 __func__, (unsigned long long)sh->sector,
1069 atomic_inc(&sh->count);
1071 atomic_inc(&head_sh->count);
1072 if (use_new_offset(conf, sh))
1073 rbi->bi_iter.bi_sector = (sh->sector
1074 + rrdev->new_data_offset);
1076 rbi->bi_iter.bi_sector = (sh->sector
1077 + rrdev->data_offset);
1078 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1079 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1080 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1082 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1083 rbi->bi_io_vec[0].bv_offset = 0;
1084 rbi->bi_iter.bi_size = STRIPE_SIZE;
1086 * If this is discard request, set bi_vcnt 0. We don't
1087 * want to confuse SCSI because SCSI will replace payload
1089 if (op == REQ_OP_DISCARD)
1091 if (conf->mddev->gendisk)
1092 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1093 rbi, disk_devt(conf->mddev->gendisk),
1095 generic_make_request(rbi);
1097 if (!rdev && !rrdev) {
1098 if (op_is_write(op))
1099 set_bit(STRIPE_DEGRADED, &sh->state);
1100 pr_debug("skip op %d on disc %d for sector %llu\n",
1101 bi->bi_opf, i, (unsigned long long)sh->sector);
1102 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1103 set_bit(STRIPE_HANDLE, &sh->state);
1106 if (!head_sh->batch_head)
1108 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1115 static struct dma_async_tx_descriptor *
1116 async_copy_data(int frombio, struct bio *bio, struct page **page,
1117 sector_t sector, struct dma_async_tx_descriptor *tx,
1118 struct stripe_head *sh)
1121 struct bvec_iter iter;
1122 struct page *bio_page;
1124 struct async_submit_ctl submit;
1125 enum async_tx_flags flags = 0;
1127 if (bio->bi_iter.bi_sector >= sector)
1128 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1130 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1133 flags |= ASYNC_TX_FENCE;
1134 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1136 bio_for_each_segment(bvl, bio, iter) {
1137 int len = bvl.bv_len;
1141 if (page_offset < 0) {
1142 b_offset = -page_offset;
1143 page_offset += b_offset;
1147 if (len > 0 && page_offset + len > STRIPE_SIZE)
1148 clen = STRIPE_SIZE - page_offset;
1153 b_offset += bvl.bv_offset;
1154 bio_page = bvl.bv_page;
1156 if (sh->raid_conf->skip_copy &&
1157 b_offset == 0 && page_offset == 0 &&
1158 clen == STRIPE_SIZE)
1161 tx = async_memcpy(*page, bio_page, page_offset,
1162 b_offset, clen, &submit);
1164 tx = async_memcpy(bio_page, *page, b_offset,
1165 page_offset, clen, &submit);
1167 /* chain the operations */
1168 submit.depend_tx = tx;
1170 if (clen < len) /* hit end of page */
1178 static void ops_complete_biofill(void *stripe_head_ref)
1180 struct stripe_head *sh = stripe_head_ref;
1181 struct bio_list return_bi = BIO_EMPTY_LIST;
1184 pr_debug("%s: stripe %llu\n", __func__,
1185 (unsigned long long)sh->sector);
1187 /* clear completed biofills */
1188 for (i = sh->disks; i--; ) {
1189 struct r5dev *dev = &sh->dev[i];
1191 /* acknowledge completion of a biofill operation */
1192 /* and check if we need to reply to a read request,
1193 * new R5_Wantfill requests are held off until
1194 * !STRIPE_BIOFILL_RUN
1196 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1197 struct bio *rbi, *rbi2;
1202 while (rbi && rbi->bi_iter.bi_sector <
1203 dev->sector + STRIPE_SECTORS) {
1204 rbi2 = r5_next_bio(rbi, dev->sector);
1205 if (!raid5_dec_bi_active_stripes(rbi))
1206 bio_list_add(&return_bi, rbi);
1211 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1213 return_io(&return_bi);
1215 set_bit(STRIPE_HANDLE, &sh->state);
1216 raid5_release_stripe(sh);
1219 static void ops_run_biofill(struct stripe_head *sh)
1221 struct dma_async_tx_descriptor *tx = NULL;
1222 struct async_submit_ctl submit;
1225 BUG_ON(sh->batch_head);
1226 pr_debug("%s: stripe %llu\n", __func__,
1227 (unsigned long long)sh->sector);
1229 for (i = sh->disks; i--; ) {
1230 struct r5dev *dev = &sh->dev[i];
1231 if (test_bit(R5_Wantfill, &dev->flags)) {
1233 spin_lock_irq(&sh->stripe_lock);
1234 dev->read = rbi = dev->toread;
1236 spin_unlock_irq(&sh->stripe_lock);
1237 while (rbi && rbi->bi_iter.bi_sector <
1238 dev->sector + STRIPE_SECTORS) {
1239 tx = async_copy_data(0, rbi, &dev->page,
1240 dev->sector, tx, sh);
1241 rbi = r5_next_bio(rbi, dev->sector);
1246 atomic_inc(&sh->count);
1247 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1248 async_trigger_callback(&submit);
1251 static void mark_target_uptodate(struct stripe_head *sh, int target)
1258 tgt = &sh->dev[target];
1259 set_bit(R5_UPTODATE, &tgt->flags);
1260 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1261 clear_bit(R5_Wantcompute, &tgt->flags);
1264 static void ops_complete_compute(void *stripe_head_ref)
1266 struct stripe_head *sh = stripe_head_ref;
1268 pr_debug("%s: stripe %llu\n", __func__,
1269 (unsigned long long)sh->sector);
1271 /* mark the computed target(s) as uptodate */
1272 mark_target_uptodate(sh, sh->ops.target);
1273 mark_target_uptodate(sh, sh->ops.target2);
1275 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1276 if (sh->check_state == check_state_compute_run)
1277 sh->check_state = check_state_compute_result;
1278 set_bit(STRIPE_HANDLE, &sh->state);
1279 raid5_release_stripe(sh);
1282 /* return a pointer to the address conversion region of the scribble buffer */
1283 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1284 struct raid5_percpu *percpu, int i)
1288 addr = flex_array_get(percpu->scribble, i);
1289 return addr + sizeof(struct page *) * (sh->disks + 2);
1292 /* return a pointer to the address conversion region of the scribble buffer */
1293 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1297 addr = flex_array_get(percpu->scribble, i);
1301 static struct dma_async_tx_descriptor *
1302 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1304 int disks = sh->disks;
1305 struct page **xor_srcs = to_addr_page(percpu, 0);
1306 int target = sh->ops.target;
1307 struct r5dev *tgt = &sh->dev[target];
1308 struct page *xor_dest = tgt->page;
1310 struct dma_async_tx_descriptor *tx;
1311 struct async_submit_ctl submit;
1314 BUG_ON(sh->batch_head);
1316 pr_debug("%s: stripe %llu block: %d\n",
1317 __func__, (unsigned long long)sh->sector, target);
1318 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1320 for (i = disks; i--; )
1322 xor_srcs[count++] = sh->dev[i].page;
1324 atomic_inc(&sh->count);
1326 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1327 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1328 if (unlikely(count == 1))
1329 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1331 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1336 /* set_syndrome_sources - populate source buffers for gen_syndrome
1337 * @srcs - (struct page *) array of size sh->disks
1338 * @sh - stripe_head to parse
1340 * Populates srcs in proper layout order for the stripe and returns the
1341 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1342 * destination buffer is recorded in srcs[count] and the Q destination
1343 * is recorded in srcs[count+1]].
1345 static int set_syndrome_sources(struct page **srcs,
1346 struct stripe_head *sh,
1349 int disks = sh->disks;
1350 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1351 int d0_idx = raid6_d0(sh);
1355 for (i = 0; i < disks; i++)
1361 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1362 struct r5dev *dev = &sh->dev[i];
1364 if (i == sh->qd_idx || i == sh->pd_idx ||
1365 (srctype == SYNDROME_SRC_ALL) ||
1366 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1367 test_bit(R5_Wantdrain, &dev->flags)) ||
1368 (srctype == SYNDROME_SRC_WRITTEN &&
1370 srcs[slot] = sh->dev[i].page;
1371 i = raid6_next_disk(i, disks);
1372 } while (i != d0_idx);
1374 return syndrome_disks;
1377 static struct dma_async_tx_descriptor *
1378 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1380 int disks = sh->disks;
1381 struct page **blocks = to_addr_page(percpu, 0);
1383 int qd_idx = sh->qd_idx;
1384 struct dma_async_tx_descriptor *tx;
1385 struct async_submit_ctl submit;
1391 BUG_ON(sh->batch_head);
1392 if (sh->ops.target < 0)
1393 target = sh->ops.target2;
1394 else if (sh->ops.target2 < 0)
1395 target = sh->ops.target;
1397 /* we should only have one valid target */
1400 pr_debug("%s: stripe %llu block: %d\n",
1401 __func__, (unsigned long long)sh->sector, target);
1403 tgt = &sh->dev[target];
1404 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1407 atomic_inc(&sh->count);
1409 if (target == qd_idx) {
1410 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1411 blocks[count] = NULL; /* regenerating p is not necessary */
1412 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1413 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1414 ops_complete_compute, sh,
1415 to_addr_conv(sh, percpu, 0));
1416 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1418 /* Compute any data- or p-drive using XOR */
1420 for (i = disks; i-- ; ) {
1421 if (i == target || i == qd_idx)
1423 blocks[count++] = sh->dev[i].page;
1426 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1427 NULL, ops_complete_compute, sh,
1428 to_addr_conv(sh, percpu, 0));
1429 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1435 static struct dma_async_tx_descriptor *
1436 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1438 int i, count, disks = sh->disks;
1439 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1440 int d0_idx = raid6_d0(sh);
1441 int faila = -1, failb = -1;
1442 int target = sh->ops.target;
1443 int target2 = sh->ops.target2;
1444 struct r5dev *tgt = &sh->dev[target];
1445 struct r5dev *tgt2 = &sh->dev[target2];
1446 struct dma_async_tx_descriptor *tx;
1447 struct page **blocks = to_addr_page(percpu, 0);
1448 struct async_submit_ctl submit;
1450 BUG_ON(sh->batch_head);
1451 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1452 __func__, (unsigned long long)sh->sector, target, target2);
1453 BUG_ON(target < 0 || target2 < 0);
1454 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1455 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1457 /* we need to open-code set_syndrome_sources to handle the
1458 * slot number conversion for 'faila' and 'failb'
1460 for (i = 0; i < disks ; i++)
1465 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1467 blocks[slot] = sh->dev[i].page;
1473 i = raid6_next_disk(i, disks);
1474 } while (i != d0_idx);
1476 BUG_ON(faila == failb);
1479 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1480 __func__, (unsigned long long)sh->sector, faila, failb);
1482 atomic_inc(&sh->count);
1484 if (failb == syndrome_disks+1) {
1485 /* Q disk is one of the missing disks */
1486 if (faila == syndrome_disks) {
1487 /* Missing P+Q, just recompute */
1488 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1489 ops_complete_compute, sh,
1490 to_addr_conv(sh, percpu, 0));
1491 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1492 STRIPE_SIZE, &submit);
1496 int qd_idx = sh->qd_idx;
1498 /* Missing D+Q: recompute D from P, then recompute Q */
1499 if (target == qd_idx)
1500 data_target = target2;
1502 data_target = target;
1505 for (i = disks; i-- ; ) {
1506 if (i == data_target || i == qd_idx)
1508 blocks[count++] = sh->dev[i].page;
1510 dest = sh->dev[data_target].page;
1511 init_async_submit(&submit,
1512 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1514 to_addr_conv(sh, percpu, 0));
1515 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1518 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1519 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1520 ops_complete_compute, sh,
1521 to_addr_conv(sh, percpu, 0));
1522 return async_gen_syndrome(blocks, 0, count+2,
1523 STRIPE_SIZE, &submit);
1526 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1527 ops_complete_compute, sh,
1528 to_addr_conv(sh, percpu, 0));
1529 if (failb == syndrome_disks) {
1530 /* We're missing D+P. */
1531 return async_raid6_datap_recov(syndrome_disks+2,
1535 /* We're missing D+D. */
1536 return async_raid6_2data_recov(syndrome_disks+2,
1537 STRIPE_SIZE, faila, failb,
1543 static void ops_complete_prexor(void *stripe_head_ref)
1545 struct stripe_head *sh = stripe_head_ref;
1547 pr_debug("%s: stripe %llu\n", __func__,
1548 (unsigned long long)sh->sector);
1551 static struct dma_async_tx_descriptor *
1552 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1553 struct dma_async_tx_descriptor *tx)
1555 int disks = sh->disks;
1556 struct page **xor_srcs = to_addr_page(percpu, 0);
1557 int count = 0, pd_idx = sh->pd_idx, i;
1558 struct async_submit_ctl submit;
1560 /* existing parity data subtracted */
1561 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1563 BUG_ON(sh->batch_head);
1564 pr_debug("%s: stripe %llu\n", __func__,
1565 (unsigned long long)sh->sector);
1567 for (i = disks; i--; ) {
1568 struct r5dev *dev = &sh->dev[i];
1569 /* Only process blocks that are known to be uptodate */
1570 if (test_bit(R5_Wantdrain, &dev->flags))
1571 xor_srcs[count++] = dev->page;
1574 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1575 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1576 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1581 static struct dma_async_tx_descriptor *
1582 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1583 struct dma_async_tx_descriptor *tx)
1585 struct page **blocks = to_addr_page(percpu, 0);
1587 struct async_submit_ctl submit;
1589 pr_debug("%s: stripe %llu\n", __func__,
1590 (unsigned long long)sh->sector);
1592 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1594 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1595 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1596 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1601 static struct dma_async_tx_descriptor *
1602 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1604 int disks = sh->disks;
1606 struct stripe_head *head_sh = sh;
1608 pr_debug("%s: stripe %llu\n", __func__,
1609 (unsigned long long)sh->sector);
1611 for (i = disks; i--; ) {
1616 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1621 spin_lock_irq(&sh->stripe_lock);
1622 chosen = dev->towrite;
1623 dev->towrite = NULL;
1624 sh->overwrite_disks = 0;
1625 BUG_ON(dev->written);
1626 wbi = dev->written = chosen;
1627 spin_unlock_irq(&sh->stripe_lock);
1628 WARN_ON(dev->page != dev->orig_page);
1630 while (wbi && wbi->bi_iter.bi_sector <
1631 dev->sector + STRIPE_SECTORS) {
1632 if (wbi->bi_opf & REQ_FUA)
1633 set_bit(R5_WantFUA, &dev->flags);
1634 if (wbi->bi_opf & REQ_SYNC)
1635 set_bit(R5_SyncIO, &dev->flags);
1636 if (bio_op(wbi) == REQ_OP_DISCARD)
1637 set_bit(R5_Discard, &dev->flags);
1639 tx = async_copy_data(1, wbi, &dev->page,
1640 dev->sector, tx, sh);
1641 if (dev->page != dev->orig_page) {
1642 set_bit(R5_SkipCopy, &dev->flags);
1643 clear_bit(R5_UPTODATE, &dev->flags);
1644 clear_bit(R5_OVERWRITE, &dev->flags);
1647 wbi = r5_next_bio(wbi, dev->sector);
1650 if (head_sh->batch_head) {
1651 sh = list_first_entry(&sh->batch_list,
1664 static void ops_complete_reconstruct(void *stripe_head_ref)
1666 struct stripe_head *sh = stripe_head_ref;
1667 int disks = sh->disks;
1668 int pd_idx = sh->pd_idx;
1669 int qd_idx = sh->qd_idx;
1671 bool fua = false, sync = false, discard = false;
1673 pr_debug("%s: stripe %llu\n", __func__,
1674 (unsigned long long)sh->sector);
1676 for (i = disks; i--; ) {
1677 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1678 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1679 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1682 for (i = disks; i--; ) {
1683 struct r5dev *dev = &sh->dev[i];
1685 if (dev->written || i == pd_idx || i == qd_idx) {
1686 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1687 set_bit(R5_UPTODATE, &dev->flags);
1689 set_bit(R5_WantFUA, &dev->flags);
1691 set_bit(R5_SyncIO, &dev->flags);
1695 if (sh->reconstruct_state == reconstruct_state_drain_run)
1696 sh->reconstruct_state = reconstruct_state_drain_result;
1697 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1698 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1700 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1701 sh->reconstruct_state = reconstruct_state_result;
1704 set_bit(STRIPE_HANDLE, &sh->state);
1705 raid5_release_stripe(sh);
1709 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1710 struct dma_async_tx_descriptor *tx)
1712 int disks = sh->disks;
1713 struct page **xor_srcs;
1714 struct async_submit_ctl submit;
1715 int count, pd_idx = sh->pd_idx, i;
1716 struct page *xor_dest;
1718 unsigned long flags;
1720 struct stripe_head *head_sh = sh;
1723 pr_debug("%s: stripe %llu\n", __func__,
1724 (unsigned long long)sh->sector);
1726 for (i = 0; i < sh->disks; i++) {
1729 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1732 if (i >= sh->disks) {
1733 atomic_inc(&sh->count);
1734 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1735 ops_complete_reconstruct(sh);
1740 xor_srcs = to_addr_page(percpu, j);
1741 /* check if prexor is active which means only process blocks
1742 * that are part of a read-modify-write (written)
1744 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1746 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1747 for (i = disks; i--; ) {
1748 struct r5dev *dev = &sh->dev[i];
1749 if (head_sh->dev[i].written)
1750 xor_srcs[count++] = dev->page;
1753 xor_dest = sh->dev[pd_idx].page;
1754 for (i = disks; i--; ) {
1755 struct r5dev *dev = &sh->dev[i];
1757 xor_srcs[count++] = dev->page;
1761 /* 1/ if we prexor'd then the dest is reused as a source
1762 * 2/ if we did not prexor then we are redoing the parity
1763 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1764 * for the synchronous xor case
1766 last_stripe = !head_sh->batch_head ||
1767 list_first_entry(&sh->batch_list,
1768 struct stripe_head, batch_list) == head_sh;
1770 flags = ASYNC_TX_ACK |
1771 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1773 atomic_inc(&head_sh->count);
1774 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1775 to_addr_conv(sh, percpu, j));
1777 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1778 init_async_submit(&submit, flags, tx, NULL, NULL,
1779 to_addr_conv(sh, percpu, j));
1782 if (unlikely(count == 1))
1783 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1785 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1788 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1795 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1796 struct dma_async_tx_descriptor *tx)
1798 struct async_submit_ctl submit;
1799 struct page **blocks;
1800 int count, i, j = 0;
1801 struct stripe_head *head_sh = sh;
1804 unsigned long txflags;
1806 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1808 for (i = 0; i < sh->disks; i++) {
1809 if (sh->pd_idx == i || sh->qd_idx == i)
1811 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1814 if (i >= sh->disks) {
1815 atomic_inc(&sh->count);
1816 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1817 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1818 ops_complete_reconstruct(sh);
1823 blocks = to_addr_page(percpu, j);
1825 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1826 synflags = SYNDROME_SRC_WRITTEN;
1827 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1829 synflags = SYNDROME_SRC_ALL;
1830 txflags = ASYNC_TX_ACK;
1833 count = set_syndrome_sources(blocks, sh, synflags);
1834 last_stripe = !head_sh->batch_head ||
1835 list_first_entry(&sh->batch_list,
1836 struct stripe_head, batch_list) == head_sh;
1839 atomic_inc(&head_sh->count);
1840 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1841 head_sh, to_addr_conv(sh, percpu, j));
1843 init_async_submit(&submit, 0, tx, NULL, NULL,
1844 to_addr_conv(sh, percpu, j));
1845 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1848 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1854 static void ops_complete_check(void *stripe_head_ref)
1856 struct stripe_head *sh = stripe_head_ref;
1858 pr_debug("%s: stripe %llu\n", __func__,
1859 (unsigned long long)sh->sector);
1861 sh->check_state = check_state_check_result;
1862 set_bit(STRIPE_HANDLE, &sh->state);
1863 raid5_release_stripe(sh);
1866 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1868 int disks = sh->disks;
1869 int pd_idx = sh->pd_idx;
1870 int qd_idx = sh->qd_idx;
1871 struct page *xor_dest;
1872 struct page **xor_srcs = to_addr_page(percpu, 0);
1873 struct dma_async_tx_descriptor *tx;
1874 struct async_submit_ctl submit;
1878 pr_debug("%s: stripe %llu\n", __func__,
1879 (unsigned long long)sh->sector);
1881 BUG_ON(sh->batch_head);
1883 xor_dest = sh->dev[pd_idx].page;
1884 xor_srcs[count++] = xor_dest;
1885 for (i = disks; i--; ) {
1886 if (i == pd_idx || i == qd_idx)
1888 xor_srcs[count++] = sh->dev[i].page;
1891 init_async_submit(&submit, 0, NULL, NULL, NULL,
1892 to_addr_conv(sh, percpu, 0));
1893 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1894 &sh->ops.zero_sum_result, &submit);
1896 atomic_inc(&sh->count);
1897 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1898 tx = async_trigger_callback(&submit);
1901 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1903 struct page **srcs = to_addr_page(percpu, 0);
1904 struct async_submit_ctl submit;
1907 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1908 (unsigned long long)sh->sector, checkp);
1910 BUG_ON(sh->batch_head);
1911 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1915 atomic_inc(&sh->count);
1916 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1917 sh, to_addr_conv(sh, percpu, 0));
1918 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1919 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1922 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1924 int overlap_clear = 0, i, disks = sh->disks;
1925 struct dma_async_tx_descriptor *tx = NULL;
1926 struct r5conf *conf = sh->raid_conf;
1927 int level = conf->level;
1928 struct raid5_percpu *percpu;
1932 percpu = per_cpu_ptr(conf->percpu, cpu);
1933 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1934 ops_run_biofill(sh);
1938 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1940 tx = ops_run_compute5(sh, percpu);
1942 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1943 tx = ops_run_compute6_1(sh, percpu);
1945 tx = ops_run_compute6_2(sh, percpu);
1947 /* terminate the chain if reconstruct is not set to be run */
1948 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1952 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1954 tx = ops_run_prexor5(sh, percpu, tx);
1956 tx = ops_run_prexor6(sh, percpu, tx);
1959 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1960 tx = ops_run_biodrain(sh, tx);
1964 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1966 ops_run_reconstruct5(sh, percpu, tx);
1968 ops_run_reconstruct6(sh, percpu, tx);
1971 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1972 if (sh->check_state == check_state_run)
1973 ops_run_check_p(sh, percpu);
1974 else if (sh->check_state == check_state_run_q)
1975 ops_run_check_pq(sh, percpu, 0);
1976 else if (sh->check_state == check_state_run_pq)
1977 ops_run_check_pq(sh, percpu, 1);
1982 if (overlap_clear && !sh->batch_head)
1983 for (i = disks; i--; ) {
1984 struct r5dev *dev = &sh->dev[i];
1985 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1986 wake_up(&sh->raid_conf->wait_for_overlap);
1991 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
1994 struct stripe_head *sh;
1997 sh = kmem_cache_zalloc(sc, gfp);
1999 spin_lock_init(&sh->stripe_lock);
2000 spin_lock_init(&sh->batch_lock);
2001 INIT_LIST_HEAD(&sh->batch_list);
2002 INIT_LIST_HEAD(&sh->lru);
2003 atomic_set(&sh->count, 1);
2004 for (i = 0; i < disks; i++) {
2005 struct r5dev *dev = &sh->dev[i];
2007 bio_init(&dev->req);
2008 dev->req.bi_io_vec = &dev->vec;
2009 dev->req.bi_max_vecs = 1;
2011 bio_init(&dev->rreq);
2012 dev->rreq.bi_io_vec = &dev->rvec;
2013 dev->rreq.bi_max_vecs = 1;
2018 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2020 struct stripe_head *sh;
2022 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size);
2026 sh->raid_conf = conf;
2028 if (grow_buffers(sh, gfp)) {
2030 kmem_cache_free(conf->slab_cache, sh);
2033 sh->hash_lock_index =
2034 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2035 /* we just created an active stripe so... */
2036 atomic_inc(&conf->active_stripes);
2038 raid5_release_stripe(sh);
2039 conf->max_nr_stripes++;
2043 static int grow_stripes(struct r5conf *conf, int num)
2045 struct kmem_cache *sc;
2046 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2048 if (conf->mddev->gendisk)
2049 sprintf(conf->cache_name[0],
2050 "raid%d-%s", conf->level, mdname(conf->mddev));
2052 sprintf(conf->cache_name[0],
2053 "raid%d-%p", conf->level, conf->mddev);
2054 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2056 conf->active_name = 0;
2057 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2058 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2062 conf->slab_cache = sc;
2063 conf->pool_size = devs;
2065 if (!grow_one_stripe(conf, GFP_KERNEL))
2072 * scribble_len - return the required size of the scribble region
2073 * @num - total number of disks in the array
2075 * The size must be enough to contain:
2076 * 1/ a struct page pointer for each device in the array +2
2077 * 2/ room to convert each entry in (1) to its corresponding dma
2078 * (dma_map_page()) or page (page_address()) address.
2080 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2081 * calculate over all devices (not just the data blocks), using zeros in place
2082 * of the P and Q blocks.
2084 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2086 struct flex_array *ret;
2089 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2090 ret = flex_array_alloc(len, cnt, flags);
2093 /* always prealloc all elements, so no locking is required */
2094 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2095 flex_array_free(ret);
2101 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2107 * Never shrink. And mddev_suspend() could deadlock if this is called
2108 * from raid5d. In that case, scribble_disks and scribble_sectors
2109 * should equal to new_disks and new_sectors
2111 if (conf->scribble_disks >= new_disks &&
2112 conf->scribble_sectors >= new_sectors)
2114 mddev_suspend(conf->mddev);
2116 for_each_present_cpu(cpu) {
2117 struct raid5_percpu *percpu;
2118 struct flex_array *scribble;
2120 percpu = per_cpu_ptr(conf->percpu, cpu);
2121 scribble = scribble_alloc(new_disks,
2122 new_sectors / STRIPE_SECTORS,
2126 flex_array_free(percpu->scribble);
2127 percpu->scribble = scribble;
2134 mddev_resume(conf->mddev);
2136 conf->scribble_disks = new_disks;
2137 conf->scribble_sectors = new_sectors;
2142 static int resize_stripes(struct r5conf *conf, int newsize)
2144 /* Make all the stripes able to hold 'newsize' devices.
2145 * New slots in each stripe get 'page' set to a new page.
2147 * This happens in stages:
2148 * 1/ create a new kmem_cache and allocate the required number of
2150 * 2/ gather all the old stripe_heads and transfer the pages across
2151 * to the new stripe_heads. This will have the side effect of
2152 * freezing the array as once all stripe_heads have been collected,
2153 * no IO will be possible. Old stripe heads are freed once their
2154 * pages have been transferred over, and the old kmem_cache is
2155 * freed when all stripes are done.
2156 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2157 * we simple return a failre status - no need to clean anything up.
2158 * 4/ allocate new pages for the new slots in the new stripe_heads.
2159 * If this fails, we don't bother trying the shrink the
2160 * stripe_heads down again, we just leave them as they are.
2161 * As each stripe_head is processed the new one is released into
2164 * Once step2 is started, we cannot afford to wait for a write,
2165 * so we use GFP_NOIO allocations.
2167 struct stripe_head *osh, *nsh;
2168 LIST_HEAD(newstripes);
2169 struct disk_info *ndisks;
2171 struct kmem_cache *sc;
2175 if (newsize <= conf->pool_size)
2176 return 0; /* never bother to shrink */
2178 err = md_allow_write(conf->mddev);
2183 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2184 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2189 /* Need to ensure auto-resizing doesn't interfere */
2190 mutex_lock(&conf->cache_size_mutex);
2192 for (i = conf->max_nr_stripes; i; i--) {
2193 nsh = alloc_stripe(sc, GFP_KERNEL, newsize);
2197 nsh->raid_conf = conf;
2198 list_add(&nsh->lru, &newstripes);
2201 /* didn't get enough, give up */
2202 while (!list_empty(&newstripes)) {
2203 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2204 list_del(&nsh->lru);
2205 kmem_cache_free(sc, nsh);
2207 kmem_cache_destroy(sc);
2208 mutex_unlock(&conf->cache_size_mutex);
2211 /* Step 2 - Must use GFP_NOIO now.
2212 * OK, we have enough stripes, start collecting inactive
2213 * stripes and copying them over
2217 list_for_each_entry(nsh, &newstripes, lru) {
2218 lock_device_hash_lock(conf, hash);
2219 wait_event_cmd(conf->wait_for_stripe,
2220 !list_empty(conf->inactive_list + hash),
2221 unlock_device_hash_lock(conf, hash),
2222 lock_device_hash_lock(conf, hash));
2223 osh = get_free_stripe(conf, hash);
2224 unlock_device_hash_lock(conf, hash);
2226 for(i=0; i<conf->pool_size; i++) {
2227 nsh->dev[i].page = osh->dev[i].page;
2228 nsh->dev[i].orig_page = osh->dev[i].page;
2230 nsh->hash_lock_index = hash;
2231 kmem_cache_free(conf->slab_cache, osh);
2233 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2234 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2239 kmem_cache_destroy(conf->slab_cache);
2242 * At this point, we are holding all the stripes so the array
2243 * is completely stalled, so now is a good time to resize
2244 * conf->disks and the scribble region
2246 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2248 for (i=0; i<conf->raid_disks; i++)
2249 ndisks[i] = conf->disks[i];
2251 conf->disks = ndisks;
2255 mutex_unlock(&conf->cache_size_mutex);
2257 conf->slab_cache = sc;
2258 conf->active_name = 1-conf->active_name;
2260 /* Step 4, return new stripes to service */
2261 while(!list_empty(&newstripes)) {
2262 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2263 list_del_init(&nsh->lru);
2265 for (i=conf->raid_disks; i < newsize; i++)
2266 if (nsh->dev[i].page == NULL) {
2267 struct page *p = alloc_page(GFP_NOIO);
2268 nsh->dev[i].page = p;
2269 nsh->dev[i].orig_page = p;
2273 raid5_release_stripe(nsh);
2275 /* critical section pass, GFP_NOIO no longer needed */
2278 conf->pool_size = newsize;
2282 static int drop_one_stripe(struct r5conf *conf)
2284 struct stripe_head *sh;
2285 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2287 spin_lock_irq(conf->hash_locks + hash);
2288 sh = get_free_stripe(conf, hash);
2289 spin_unlock_irq(conf->hash_locks + hash);
2292 BUG_ON(atomic_read(&sh->count));
2294 kmem_cache_free(conf->slab_cache, sh);
2295 atomic_dec(&conf->active_stripes);
2296 conf->max_nr_stripes--;
2300 static void shrink_stripes(struct r5conf *conf)
2302 while (conf->max_nr_stripes &&
2303 drop_one_stripe(conf))
2306 kmem_cache_destroy(conf->slab_cache);
2307 conf->slab_cache = NULL;
2310 static void raid5_end_read_request(struct bio * bi)
2312 struct stripe_head *sh = bi->bi_private;
2313 struct r5conf *conf = sh->raid_conf;
2314 int disks = sh->disks, i;
2315 char b[BDEVNAME_SIZE];
2316 struct md_rdev *rdev = NULL;
2319 for (i=0 ; i<disks; i++)
2320 if (bi == &sh->dev[i].req)
2323 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2324 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2331 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2332 /* If replacement finished while this request was outstanding,
2333 * 'replacement' might be NULL already.
2334 * In that case it moved down to 'rdev'.
2335 * rdev is not removed until all requests are finished.
2337 rdev = conf->disks[i].replacement;
2339 rdev = conf->disks[i].rdev;
2341 if (use_new_offset(conf, sh))
2342 s = sh->sector + rdev->new_data_offset;
2344 s = sh->sector + rdev->data_offset;
2345 if (!bi->bi_error) {
2346 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2347 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2348 /* Note that this cannot happen on a
2349 * replacement device. We just fail those on
2354 "md/raid:%s: read error corrected"
2355 " (%lu sectors at %llu on %s)\n",
2356 mdname(conf->mddev), STRIPE_SECTORS,
2357 (unsigned long long)s,
2358 bdevname(rdev->bdev, b));
2359 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2360 clear_bit(R5_ReadError, &sh->dev[i].flags);
2361 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2362 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2363 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2365 if (atomic_read(&rdev->read_errors))
2366 atomic_set(&rdev->read_errors, 0);
2368 const char *bdn = bdevname(rdev->bdev, b);
2372 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2373 atomic_inc(&rdev->read_errors);
2374 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2377 "md/raid:%s: read error on replacement device "
2378 "(sector %llu on %s).\n",
2379 mdname(conf->mddev),
2380 (unsigned long long)s,
2382 else if (conf->mddev->degraded >= conf->max_degraded) {
2386 "md/raid:%s: read error not correctable "
2387 "(sector %llu on %s).\n",
2388 mdname(conf->mddev),
2389 (unsigned long long)s,
2391 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2396 "md/raid:%s: read error NOT corrected!! "
2397 "(sector %llu on %s).\n",
2398 mdname(conf->mddev),
2399 (unsigned long long)s,
2401 } else if (atomic_read(&rdev->read_errors)
2402 > conf->max_nr_stripes)
2404 "md/raid:%s: Too many read errors, failing device %s.\n",
2405 mdname(conf->mddev), bdn);
2408 if (set_bad && test_bit(In_sync, &rdev->flags)
2409 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2412 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2413 set_bit(R5_ReadError, &sh->dev[i].flags);
2414 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2416 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2418 clear_bit(R5_ReadError, &sh->dev[i].flags);
2419 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2421 && test_bit(In_sync, &rdev->flags)
2422 && rdev_set_badblocks(
2423 rdev, sh->sector, STRIPE_SECTORS, 0)))
2424 md_error(conf->mddev, rdev);
2427 rdev_dec_pending(rdev, conf->mddev);
2429 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2430 set_bit(STRIPE_HANDLE, &sh->state);
2431 raid5_release_stripe(sh);
2434 static void raid5_end_write_request(struct bio *bi)
2436 struct stripe_head *sh = bi->bi_private;
2437 struct r5conf *conf = sh->raid_conf;
2438 int disks = sh->disks, i;
2439 struct md_rdev *uninitialized_var(rdev);
2442 int replacement = 0;
2444 for (i = 0 ; i < disks; i++) {
2445 if (bi == &sh->dev[i].req) {
2446 rdev = conf->disks[i].rdev;
2449 if (bi == &sh->dev[i].rreq) {
2450 rdev = conf->disks[i].replacement;
2454 /* rdev was removed and 'replacement'
2455 * replaced it. rdev is not removed
2456 * until all requests are finished.
2458 rdev = conf->disks[i].rdev;
2462 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2463 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2473 md_error(conf->mddev, rdev);
2474 else if (is_badblock(rdev, sh->sector,
2476 &first_bad, &bad_sectors))
2477 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2480 set_bit(STRIPE_DEGRADED, &sh->state);
2481 set_bit(WriteErrorSeen, &rdev->flags);
2482 set_bit(R5_WriteError, &sh->dev[i].flags);
2483 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2484 set_bit(MD_RECOVERY_NEEDED,
2485 &rdev->mddev->recovery);
2486 } else if (is_badblock(rdev, sh->sector,
2488 &first_bad, &bad_sectors)) {
2489 set_bit(R5_MadeGood, &sh->dev[i].flags);
2490 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2491 /* That was a successful write so make
2492 * sure it looks like we already did
2495 set_bit(R5_ReWrite, &sh->dev[i].flags);
2498 rdev_dec_pending(rdev, conf->mddev);
2500 if (sh->batch_head && bi->bi_error && !replacement)
2501 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2504 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2505 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2506 set_bit(STRIPE_HANDLE, &sh->state);
2507 raid5_release_stripe(sh);
2509 if (sh->batch_head && sh != sh->batch_head)
2510 raid5_release_stripe(sh->batch_head);
2513 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2515 struct r5dev *dev = &sh->dev[i];
2518 dev->sector = raid5_compute_blocknr(sh, i, previous);
2521 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2523 char b[BDEVNAME_SIZE];
2524 struct r5conf *conf = mddev->private;
2525 unsigned long flags;
2526 pr_debug("raid456: error called\n");
2528 spin_lock_irqsave(&conf->device_lock, flags);
2529 clear_bit(In_sync, &rdev->flags);
2530 mddev->degraded = calc_degraded(conf);
2531 spin_unlock_irqrestore(&conf->device_lock, flags);
2532 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2534 set_bit(Blocked, &rdev->flags);
2535 set_bit(Faulty, &rdev->flags);
2536 set_mask_bits(&mddev->flags, 0,
2537 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
2539 "md/raid:%s: Disk failure on %s, disabling device.\n"
2540 "md/raid:%s: Operation continuing on %d devices.\n",
2542 bdevname(rdev->bdev, b),
2544 conf->raid_disks - mddev->degraded);
2548 * Input: a 'big' sector number,
2549 * Output: index of the data and parity disk, and the sector # in them.
2551 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2552 int previous, int *dd_idx,
2553 struct stripe_head *sh)
2555 sector_t stripe, stripe2;
2556 sector_t chunk_number;
2557 unsigned int chunk_offset;
2560 sector_t new_sector;
2561 int algorithm = previous ? conf->prev_algo
2563 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2564 : conf->chunk_sectors;
2565 int raid_disks = previous ? conf->previous_raid_disks
2567 int data_disks = raid_disks - conf->max_degraded;
2569 /* First compute the information on this sector */
2572 * Compute the chunk number and the sector offset inside the chunk
2574 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2575 chunk_number = r_sector;
2578 * Compute the stripe number
2580 stripe = chunk_number;
2581 *dd_idx = sector_div(stripe, data_disks);
2584 * Select the parity disk based on the user selected algorithm.
2586 pd_idx = qd_idx = -1;
2587 switch(conf->level) {
2589 pd_idx = data_disks;
2592 switch (algorithm) {
2593 case ALGORITHM_LEFT_ASYMMETRIC:
2594 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2595 if (*dd_idx >= pd_idx)
2598 case ALGORITHM_RIGHT_ASYMMETRIC:
2599 pd_idx = sector_div(stripe2, raid_disks);
2600 if (*dd_idx >= pd_idx)
2603 case ALGORITHM_LEFT_SYMMETRIC:
2604 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2605 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2607 case ALGORITHM_RIGHT_SYMMETRIC:
2608 pd_idx = sector_div(stripe2, raid_disks);
2609 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2611 case ALGORITHM_PARITY_0:
2615 case ALGORITHM_PARITY_N:
2616 pd_idx = data_disks;
2624 switch (algorithm) {
2625 case ALGORITHM_LEFT_ASYMMETRIC:
2626 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2627 qd_idx = pd_idx + 1;
2628 if (pd_idx == raid_disks-1) {
2629 (*dd_idx)++; /* Q D D D P */
2631 } else if (*dd_idx >= pd_idx)
2632 (*dd_idx) += 2; /* D D P Q D */
2634 case ALGORITHM_RIGHT_ASYMMETRIC:
2635 pd_idx = sector_div(stripe2, raid_disks);
2636 qd_idx = pd_idx + 1;
2637 if (pd_idx == raid_disks-1) {
2638 (*dd_idx)++; /* Q D D D P */
2640 } else if (*dd_idx >= pd_idx)
2641 (*dd_idx) += 2; /* D D P Q D */
2643 case ALGORITHM_LEFT_SYMMETRIC:
2644 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2645 qd_idx = (pd_idx + 1) % raid_disks;
2646 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2648 case ALGORITHM_RIGHT_SYMMETRIC:
2649 pd_idx = sector_div(stripe2, raid_disks);
2650 qd_idx = (pd_idx + 1) % raid_disks;
2651 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2654 case ALGORITHM_PARITY_0:
2659 case ALGORITHM_PARITY_N:
2660 pd_idx = data_disks;
2661 qd_idx = data_disks + 1;
2664 case ALGORITHM_ROTATING_ZERO_RESTART:
2665 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2666 * of blocks for computing Q is different.
2668 pd_idx = sector_div(stripe2, raid_disks);
2669 qd_idx = pd_idx + 1;
2670 if (pd_idx == raid_disks-1) {
2671 (*dd_idx)++; /* Q D D D P */
2673 } else if (*dd_idx >= pd_idx)
2674 (*dd_idx) += 2; /* D D P Q D */
2678 case ALGORITHM_ROTATING_N_RESTART:
2679 /* Same a left_asymmetric, by first stripe is
2680 * D D D P Q rather than
2684 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2685 qd_idx = pd_idx + 1;
2686 if (pd_idx == raid_disks-1) {
2687 (*dd_idx)++; /* Q D D D P */
2689 } else if (*dd_idx >= pd_idx)
2690 (*dd_idx) += 2; /* D D P Q D */
2694 case ALGORITHM_ROTATING_N_CONTINUE:
2695 /* Same as left_symmetric but Q is before P */
2696 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2697 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2698 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2702 case ALGORITHM_LEFT_ASYMMETRIC_6:
2703 /* RAID5 left_asymmetric, with Q on last device */
2704 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2705 if (*dd_idx >= pd_idx)
2707 qd_idx = raid_disks - 1;
2710 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2711 pd_idx = sector_div(stripe2, raid_disks-1);
2712 if (*dd_idx >= pd_idx)
2714 qd_idx = raid_disks - 1;
2717 case ALGORITHM_LEFT_SYMMETRIC_6:
2718 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2719 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2720 qd_idx = raid_disks - 1;
2723 case ALGORITHM_RIGHT_SYMMETRIC_6:
2724 pd_idx = sector_div(stripe2, raid_disks-1);
2725 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2726 qd_idx = raid_disks - 1;
2729 case ALGORITHM_PARITY_0_6:
2732 qd_idx = raid_disks - 1;
2742 sh->pd_idx = pd_idx;
2743 sh->qd_idx = qd_idx;
2744 sh->ddf_layout = ddf_layout;
2747 * Finally, compute the new sector number
2749 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2753 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2755 struct r5conf *conf = sh->raid_conf;
2756 int raid_disks = sh->disks;
2757 int data_disks = raid_disks - conf->max_degraded;
2758 sector_t new_sector = sh->sector, check;
2759 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2760 : conf->chunk_sectors;
2761 int algorithm = previous ? conf->prev_algo
2765 sector_t chunk_number;
2766 int dummy1, dd_idx = i;
2768 struct stripe_head sh2;
2770 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2771 stripe = new_sector;
2773 if (i == sh->pd_idx)
2775 switch(conf->level) {
2778 switch (algorithm) {
2779 case ALGORITHM_LEFT_ASYMMETRIC:
2780 case ALGORITHM_RIGHT_ASYMMETRIC:
2784 case ALGORITHM_LEFT_SYMMETRIC:
2785 case ALGORITHM_RIGHT_SYMMETRIC:
2788 i -= (sh->pd_idx + 1);
2790 case ALGORITHM_PARITY_0:
2793 case ALGORITHM_PARITY_N:
2800 if (i == sh->qd_idx)
2801 return 0; /* It is the Q disk */
2802 switch (algorithm) {
2803 case ALGORITHM_LEFT_ASYMMETRIC:
2804 case ALGORITHM_RIGHT_ASYMMETRIC:
2805 case ALGORITHM_ROTATING_ZERO_RESTART:
2806 case ALGORITHM_ROTATING_N_RESTART:
2807 if (sh->pd_idx == raid_disks-1)
2808 i--; /* Q D D D P */
2809 else if (i > sh->pd_idx)
2810 i -= 2; /* D D P Q D */
2812 case ALGORITHM_LEFT_SYMMETRIC:
2813 case ALGORITHM_RIGHT_SYMMETRIC:
2814 if (sh->pd_idx == raid_disks-1)
2815 i--; /* Q D D D P */
2820 i -= (sh->pd_idx + 2);
2823 case ALGORITHM_PARITY_0:
2826 case ALGORITHM_PARITY_N:
2828 case ALGORITHM_ROTATING_N_CONTINUE:
2829 /* Like left_symmetric, but P is before Q */
2830 if (sh->pd_idx == 0)
2831 i--; /* P D D D Q */
2836 i -= (sh->pd_idx + 1);
2839 case ALGORITHM_LEFT_ASYMMETRIC_6:
2840 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2844 case ALGORITHM_LEFT_SYMMETRIC_6:
2845 case ALGORITHM_RIGHT_SYMMETRIC_6:
2847 i += data_disks + 1;
2848 i -= (sh->pd_idx + 1);
2850 case ALGORITHM_PARITY_0_6:
2859 chunk_number = stripe * data_disks + i;
2860 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2862 check = raid5_compute_sector(conf, r_sector,
2863 previous, &dummy1, &sh2);
2864 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2865 || sh2.qd_idx != sh->qd_idx) {
2866 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2867 mdname(conf->mddev));
2874 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2875 int rcw, int expand)
2877 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2878 struct r5conf *conf = sh->raid_conf;
2879 int level = conf->level;
2883 for (i = disks; i--; ) {
2884 struct r5dev *dev = &sh->dev[i];
2887 set_bit(R5_LOCKED, &dev->flags);
2888 set_bit(R5_Wantdrain, &dev->flags);
2890 clear_bit(R5_UPTODATE, &dev->flags);
2894 /* if we are not expanding this is a proper write request, and
2895 * there will be bios with new data to be drained into the
2900 /* False alarm, nothing to do */
2902 sh->reconstruct_state = reconstruct_state_drain_run;
2903 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2905 sh->reconstruct_state = reconstruct_state_run;
2907 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2909 if (s->locked + conf->max_degraded == disks)
2910 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2911 atomic_inc(&conf->pending_full_writes);
2913 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2914 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2915 BUG_ON(level == 6 &&
2916 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2917 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2919 for (i = disks; i--; ) {
2920 struct r5dev *dev = &sh->dev[i];
2921 if (i == pd_idx || i == qd_idx)
2925 (test_bit(R5_UPTODATE, &dev->flags) ||
2926 test_bit(R5_Wantcompute, &dev->flags))) {
2927 set_bit(R5_Wantdrain, &dev->flags);
2928 set_bit(R5_LOCKED, &dev->flags);
2929 clear_bit(R5_UPTODATE, &dev->flags);
2934 /* False alarm - nothing to do */
2936 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2937 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2938 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2939 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2942 /* keep the parity disk(s) locked while asynchronous operations
2945 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2946 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2950 int qd_idx = sh->qd_idx;
2951 struct r5dev *dev = &sh->dev[qd_idx];
2953 set_bit(R5_LOCKED, &dev->flags);
2954 clear_bit(R5_UPTODATE, &dev->flags);
2958 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2959 __func__, (unsigned long long)sh->sector,
2960 s->locked, s->ops_request);
2964 * Each stripe/dev can have one or more bion attached.
2965 * toread/towrite point to the first in a chain.
2966 * The bi_next chain must be in order.
2968 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2969 int forwrite, int previous)
2972 struct r5conf *conf = sh->raid_conf;
2975 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2976 (unsigned long long)bi->bi_iter.bi_sector,
2977 (unsigned long long)sh->sector);
2980 * If several bio share a stripe. The bio bi_phys_segments acts as a
2981 * reference count to avoid race. The reference count should already be
2982 * increased before this function is called (for example, in
2983 * raid5_make_request()), so other bio sharing this stripe will not free the
2984 * stripe. If a stripe is owned by one stripe, the stripe lock will
2987 spin_lock_irq(&sh->stripe_lock);
2988 /* Don't allow new IO added to stripes in batch list */
2992 bip = &sh->dev[dd_idx].towrite;
2996 bip = &sh->dev[dd_idx].toread;
2997 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2998 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3000 bip = & (*bip)->bi_next;
3002 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3005 if (!forwrite || previous)
3006 clear_bit(STRIPE_BATCH_READY, &sh->state);
3008 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3012 raid5_inc_bi_active_stripes(bi);
3015 /* check if page is covered */
3016 sector_t sector = sh->dev[dd_idx].sector;
3017 for (bi=sh->dev[dd_idx].towrite;
3018 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3019 bi && bi->bi_iter.bi_sector <= sector;
3020 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3021 if (bio_end_sector(bi) >= sector)
3022 sector = bio_end_sector(bi);
3024 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3025 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3026 sh->overwrite_disks++;
3029 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3030 (unsigned long long)(*bip)->bi_iter.bi_sector,
3031 (unsigned long long)sh->sector, dd_idx);
3033 if (conf->mddev->bitmap && firstwrite) {
3034 /* Cannot hold spinlock over bitmap_startwrite,
3035 * but must ensure this isn't added to a batch until
3036 * we have added to the bitmap and set bm_seq.
3037 * So set STRIPE_BITMAP_PENDING to prevent
3039 * If multiple add_stripe_bio() calls race here they
3040 * much all set STRIPE_BITMAP_PENDING. So only the first one
3041 * to complete "bitmap_startwrite" gets to set
3042 * STRIPE_BIT_DELAY. This is important as once a stripe
3043 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3046 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3047 spin_unlock_irq(&sh->stripe_lock);
3048 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3050 spin_lock_irq(&sh->stripe_lock);
3051 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3052 if (!sh->batch_head) {
3053 sh->bm_seq = conf->seq_flush+1;
3054 set_bit(STRIPE_BIT_DELAY, &sh->state);
3057 spin_unlock_irq(&sh->stripe_lock);
3059 if (stripe_can_batch(sh))
3060 stripe_add_to_batch_list(conf, sh);
3064 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3065 spin_unlock_irq(&sh->stripe_lock);
3069 static void end_reshape(struct r5conf *conf);
3071 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3072 struct stripe_head *sh)
3074 int sectors_per_chunk =
3075 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3077 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3078 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3080 raid5_compute_sector(conf,
3081 stripe * (disks - conf->max_degraded)
3082 *sectors_per_chunk + chunk_offset,
3088 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3089 struct stripe_head_state *s, int disks,
3090 struct bio_list *return_bi)
3093 BUG_ON(sh->batch_head);
3094 for (i = disks; i--; ) {
3098 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3099 struct md_rdev *rdev;
3101 rdev = rcu_dereference(conf->disks[i].rdev);
3102 if (rdev && test_bit(In_sync, &rdev->flags) &&
3103 !test_bit(Faulty, &rdev->flags))
3104 atomic_inc(&rdev->nr_pending);
3109 if (!rdev_set_badblocks(
3113 md_error(conf->mddev, rdev);
3114 rdev_dec_pending(rdev, conf->mddev);
3117 spin_lock_irq(&sh->stripe_lock);
3118 /* fail all writes first */
3119 bi = sh->dev[i].towrite;
3120 sh->dev[i].towrite = NULL;
3121 sh->overwrite_disks = 0;
3122 spin_unlock_irq(&sh->stripe_lock);
3126 r5l_stripe_write_finished(sh);
3128 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3129 wake_up(&conf->wait_for_overlap);
3131 while (bi && bi->bi_iter.bi_sector <
3132 sh->dev[i].sector + STRIPE_SECTORS) {
3133 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3135 bi->bi_error = -EIO;
3136 if (!raid5_dec_bi_active_stripes(bi)) {
3137 md_write_end(conf->mddev);
3138 bio_list_add(return_bi, bi);
3143 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3144 STRIPE_SECTORS, 0, 0);
3146 /* and fail all 'written' */
3147 bi = sh->dev[i].written;
3148 sh->dev[i].written = NULL;
3149 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3150 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3151 sh->dev[i].page = sh->dev[i].orig_page;
3154 if (bi) bitmap_end = 1;
3155 while (bi && bi->bi_iter.bi_sector <
3156 sh->dev[i].sector + STRIPE_SECTORS) {
3157 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3159 bi->bi_error = -EIO;
3160 if (!raid5_dec_bi_active_stripes(bi)) {
3161 md_write_end(conf->mddev);
3162 bio_list_add(return_bi, bi);
3167 /* fail any reads if this device is non-operational and
3168 * the data has not reached the cache yet.
3170 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3171 s->failed > conf->max_degraded &&
3172 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3173 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3174 spin_lock_irq(&sh->stripe_lock);
3175 bi = sh->dev[i].toread;
3176 sh->dev[i].toread = NULL;
3177 spin_unlock_irq(&sh->stripe_lock);
3178 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3179 wake_up(&conf->wait_for_overlap);
3182 while (bi && bi->bi_iter.bi_sector <
3183 sh->dev[i].sector + STRIPE_SECTORS) {
3184 struct bio *nextbi =
3185 r5_next_bio(bi, sh->dev[i].sector);
3187 bi->bi_error = -EIO;
3188 if (!raid5_dec_bi_active_stripes(bi))
3189 bio_list_add(return_bi, bi);
3194 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3195 STRIPE_SECTORS, 0, 0);
3196 /* If we were in the middle of a write the parity block might
3197 * still be locked - so just clear all R5_LOCKED flags
3199 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3204 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3205 if (atomic_dec_and_test(&conf->pending_full_writes))
3206 md_wakeup_thread(conf->mddev->thread);
3210 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3211 struct stripe_head_state *s)
3216 BUG_ON(sh->batch_head);
3217 clear_bit(STRIPE_SYNCING, &sh->state);
3218 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3219 wake_up(&conf->wait_for_overlap);
3222 /* There is nothing more to do for sync/check/repair.
3223 * Don't even need to abort as that is handled elsewhere
3224 * if needed, and not always wanted e.g. if there is a known
3226 * For recover/replace we need to record a bad block on all
3227 * non-sync devices, or abort the recovery
3229 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3230 /* During recovery devices cannot be removed, so
3231 * locking and refcounting of rdevs is not needed
3234 for (i = 0; i < conf->raid_disks; i++) {
3235 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3237 && !test_bit(Faulty, &rdev->flags)
3238 && !test_bit(In_sync, &rdev->flags)
3239 && !rdev_set_badblocks(rdev, sh->sector,
3242 rdev = rcu_dereference(conf->disks[i].replacement);
3244 && !test_bit(Faulty, &rdev->flags)
3245 && !test_bit(In_sync, &rdev->flags)
3246 && !rdev_set_badblocks(rdev, sh->sector,
3252 conf->recovery_disabled =
3253 conf->mddev->recovery_disabled;
3255 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3258 static int want_replace(struct stripe_head *sh, int disk_idx)
3260 struct md_rdev *rdev;
3264 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3266 && !test_bit(Faulty, &rdev->flags)
3267 && !test_bit(In_sync, &rdev->flags)
3268 && (rdev->recovery_offset <= sh->sector
3269 || rdev->mddev->recovery_cp <= sh->sector))
3275 /* fetch_block - checks the given member device to see if its data needs
3276 * to be read or computed to satisfy a request.
3278 * Returns 1 when no more member devices need to be checked, otherwise returns
3279 * 0 to tell the loop in handle_stripe_fill to continue
3282 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3283 int disk_idx, int disks)
3285 struct r5dev *dev = &sh->dev[disk_idx];
3286 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3287 &sh->dev[s->failed_num[1]] };
3291 if (test_bit(R5_LOCKED, &dev->flags) ||
3292 test_bit(R5_UPTODATE, &dev->flags))
3293 /* No point reading this as we already have it or have
3294 * decided to get it.
3299 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3300 /* We need this block to directly satisfy a request */
3303 if (s->syncing || s->expanding ||
3304 (s->replacing && want_replace(sh, disk_idx)))
3305 /* When syncing, or expanding we read everything.
3306 * When replacing, we need the replaced block.
3310 if ((s->failed >= 1 && fdev[0]->toread) ||
3311 (s->failed >= 2 && fdev[1]->toread))
3312 /* If we want to read from a failed device, then
3313 * we need to actually read every other device.
3317 /* Sometimes neither read-modify-write nor reconstruct-write
3318 * cycles can work. In those cases we read every block we
3319 * can. Then the parity-update is certain to have enough to
3321 * This can only be a problem when we need to write something,
3322 * and some device has failed. If either of those tests
3323 * fail we need look no further.
3325 if (!s->failed || !s->to_write)
3328 if (test_bit(R5_Insync, &dev->flags) &&
3329 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3330 /* Pre-reads at not permitted until after short delay
3331 * to gather multiple requests. However if this
3332 * device is no Insync, the block could only be be computed
3333 * and there is no need to delay that.
3337 for (i = 0; i < s->failed && i < 2; i++) {
3338 if (fdev[i]->towrite &&
3339 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3340 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3341 /* If we have a partial write to a failed
3342 * device, then we will need to reconstruct
3343 * the content of that device, so all other
3344 * devices must be read.
3349 /* If we are forced to do a reconstruct-write, either because
3350 * the current RAID6 implementation only supports that, or
3351 * or because parity cannot be trusted and we are currently
3352 * recovering it, there is extra need to be careful.
3353 * If one of the devices that we would need to read, because
3354 * it is not being overwritten (and maybe not written at all)
3355 * is missing/faulty, then we need to read everything we can.
3357 if (sh->raid_conf->level != 6 &&
3358 sh->sector < sh->raid_conf->mddev->recovery_cp)
3359 /* reconstruct-write isn't being forced */
3361 for (i = 0; i < s->failed && i < 2; i++) {
3362 if (s->failed_num[i] != sh->pd_idx &&
3363 s->failed_num[i] != sh->qd_idx &&
3364 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3365 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3372 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3373 int disk_idx, int disks)
3375 struct r5dev *dev = &sh->dev[disk_idx];
3377 /* is the data in this block needed, and can we get it? */
3378 if (need_this_block(sh, s, disk_idx, disks)) {
3379 /* we would like to get this block, possibly by computing it,
3380 * otherwise read it if the backing disk is insync
3382 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3383 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3384 BUG_ON(sh->batch_head);
3385 if ((s->uptodate == disks - 1) &&
3386 (s->failed && (disk_idx == s->failed_num[0] ||
3387 disk_idx == s->failed_num[1]))) {
3388 /* have disk failed, and we're requested to fetch it;
3391 pr_debug("Computing stripe %llu block %d\n",
3392 (unsigned long long)sh->sector, disk_idx);
3393 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3394 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3395 set_bit(R5_Wantcompute, &dev->flags);
3396 sh->ops.target = disk_idx;
3397 sh->ops.target2 = -1; /* no 2nd target */
3399 /* Careful: from this point on 'uptodate' is in the eye
3400 * of raid_run_ops which services 'compute' operations
3401 * before writes. R5_Wantcompute flags a block that will
3402 * be R5_UPTODATE by the time it is needed for a
3403 * subsequent operation.
3407 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3408 /* Computing 2-failure is *very* expensive; only
3409 * do it if failed >= 2
3412 for (other = disks; other--; ) {
3413 if (other == disk_idx)
3415 if (!test_bit(R5_UPTODATE,
3416 &sh->dev[other].flags))
3420 pr_debug("Computing stripe %llu blocks %d,%d\n",
3421 (unsigned long long)sh->sector,
3423 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3424 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3425 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3426 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3427 sh->ops.target = disk_idx;
3428 sh->ops.target2 = other;
3432 } else if (test_bit(R5_Insync, &dev->flags)) {
3433 set_bit(R5_LOCKED, &dev->flags);
3434 set_bit(R5_Wantread, &dev->flags);
3436 pr_debug("Reading block %d (sync=%d)\n",
3437 disk_idx, s->syncing);
3445 * handle_stripe_fill - read or compute data to satisfy pending requests.
3447 static void handle_stripe_fill(struct stripe_head *sh,
3448 struct stripe_head_state *s,
3453 /* look for blocks to read/compute, skip this if a compute
3454 * is already in flight, or if the stripe contents are in the
3455 * midst of changing due to a write
3457 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3458 !sh->reconstruct_state)
3459 for (i = disks; i--; )
3460 if (fetch_block(sh, s, i, disks))
3462 set_bit(STRIPE_HANDLE, &sh->state);
3465 static void break_stripe_batch_list(struct stripe_head *head_sh,
3466 unsigned long handle_flags);
3467 /* handle_stripe_clean_event
3468 * any written block on an uptodate or failed drive can be returned.
3469 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3470 * never LOCKED, so we don't need to test 'failed' directly.
3472 static void handle_stripe_clean_event(struct r5conf *conf,
3473 struct stripe_head *sh, int disks, struct bio_list *return_bi)
3477 int discard_pending = 0;
3478 struct stripe_head *head_sh = sh;
3479 bool do_endio = false;
3481 for (i = disks; i--; )
3482 if (sh->dev[i].written) {
3484 if (!test_bit(R5_LOCKED, &dev->flags) &&
3485 (test_bit(R5_UPTODATE, &dev->flags) ||
3486 test_bit(R5_Discard, &dev->flags) ||
3487 test_bit(R5_SkipCopy, &dev->flags))) {
3488 /* We can return any write requests */
3489 struct bio *wbi, *wbi2;
3490 pr_debug("Return write for disc %d\n", i);
3491 if (test_and_clear_bit(R5_Discard, &dev->flags))
3492 clear_bit(R5_UPTODATE, &dev->flags);
3493 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3494 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3499 dev->page = dev->orig_page;
3501 dev->written = NULL;
3502 while (wbi && wbi->bi_iter.bi_sector <
3503 dev->sector + STRIPE_SECTORS) {
3504 wbi2 = r5_next_bio(wbi, dev->sector);
3505 if (!raid5_dec_bi_active_stripes(wbi)) {
3506 md_write_end(conf->mddev);
3507 bio_list_add(return_bi, wbi);
3511 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3513 !test_bit(STRIPE_DEGRADED, &sh->state),
3515 if (head_sh->batch_head) {
3516 sh = list_first_entry(&sh->batch_list,
3519 if (sh != head_sh) {
3526 } else if (test_bit(R5_Discard, &dev->flags))
3527 discard_pending = 1;
3530 r5l_stripe_write_finished(sh);
3532 if (!discard_pending &&
3533 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3535 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3536 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3537 if (sh->qd_idx >= 0) {
3538 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3539 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3541 /* now that discard is done we can proceed with any sync */
3542 clear_bit(STRIPE_DISCARD, &sh->state);
3544 * SCSI discard will change some bio fields and the stripe has
3545 * no updated data, so remove it from hash list and the stripe
3546 * will be reinitialized
3549 hash = sh->hash_lock_index;
3550 spin_lock_irq(conf->hash_locks + hash);
3552 spin_unlock_irq(conf->hash_locks + hash);
3553 if (head_sh->batch_head) {
3554 sh = list_first_entry(&sh->batch_list,
3555 struct stripe_head, batch_list);
3561 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3562 set_bit(STRIPE_HANDLE, &sh->state);
3566 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3567 if (atomic_dec_and_test(&conf->pending_full_writes))
3568 md_wakeup_thread(conf->mddev->thread);
3570 if (head_sh->batch_head && do_endio)
3571 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3574 static void handle_stripe_dirtying(struct r5conf *conf,
3575 struct stripe_head *sh,
3576 struct stripe_head_state *s,
3579 int rmw = 0, rcw = 0, i;
3580 sector_t recovery_cp = conf->mddev->recovery_cp;
3582 /* Check whether resync is now happening or should start.
3583 * If yes, then the array is dirty (after unclean shutdown or
3584 * initial creation), so parity in some stripes might be inconsistent.
3585 * In this case, we need to always do reconstruct-write, to ensure
3586 * that in case of drive failure or read-error correction, we
3587 * generate correct data from the parity.
3589 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3590 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3592 /* Calculate the real rcw later - for now make it
3593 * look like rcw is cheaper
3596 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3597 conf->rmw_level, (unsigned long long)recovery_cp,
3598 (unsigned long long)sh->sector);
3599 } else for (i = disks; i--; ) {
3600 /* would I have to read this buffer for read_modify_write */
3601 struct r5dev *dev = &sh->dev[i];
3602 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3603 !test_bit(R5_LOCKED, &dev->flags) &&
3604 !(test_bit(R5_UPTODATE, &dev->flags) ||
3605 test_bit(R5_Wantcompute, &dev->flags))) {
3606 if (test_bit(R5_Insync, &dev->flags))
3609 rmw += 2*disks; /* cannot read it */
3611 /* Would I have to read this buffer for reconstruct_write */
3612 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3613 i != sh->pd_idx && i != sh->qd_idx &&
3614 !test_bit(R5_LOCKED, &dev->flags) &&
3615 !(test_bit(R5_UPTODATE, &dev->flags) ||
3616 test_bit(R5_Wantcompute, &dev->flags))) {
3617 if (test_bit(R5_Insync, &dev->flags))
3623 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3624 (unsigned long long)sh->sector, rmw, rcw);
3625 set_bit(STRIPE_HANDLE, &sh->state);
3626 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3627 /* prefer read-modify-write, but need to get some data */
3628 if (conf->mddev->queue)
3629 blk_add_trace_msg(conf->mddev->queue,
3630 "raid5 rmw %llu %d",
3631 (unsigned long long)sh->sector, rmw);
3632 for (i = disks; i--; ) {
3633 struct r5dev *dev = &sh->dev[i];
3634 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3635 !test_bit(R5_LOCKED, &dev->flags) &&
3636 !(test_bit(R5_UPTODATE, &dev->flags) ||
3637 test_bit(R5_Wantcompute, &dev->flags)) &&
3638 test_bit(R5_Insync, &dev->flags)) {
3639 if (test_bit(STRIPE_PREREAD_ACTIVE,
3641 pr_debug("Read_old block %d for r-m-w\n",
3643 set_bit(R5_LOCKED, &dev->flags);
3644 set_bit(R5_Wantread, &dev->flags);
3647 set_bit(STRIPE_DELAYED, &sh->state);
3648 set_bit(STRIPE_HANDLE, &sh->state);
3653 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3654 /* want reconstruct write, but need to get some data */
3657 for (i = disks; i--; ) {
3658 struct r5dev *dev = &sh->dev[i];
3659 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3660 i != sh->pd_idx && i != sh->qd_idx &&
3661 !test_bit(R5_LOCKED, &dev->flags) &&
3662 !(test_bit(R5_UPTODATE, &dev->flags) ||
3663 test_bit(R5_Wantcompute, &dev->flags))) {
3665 if (test_bit(R5_Insync, &dev->flags) &&
3666 test_bit(STRIPE_PREREAD_ACTIVE,
3668 pr_debug("Read_old block "
3669 "%d for Reconstruct\n", i);
3670 set_bit(R5_LOCKED, &dev->flags);
3671 set_bit(R5_Wantread, &dev->flags);
3675 set_bit(STRIPE_DELAYED, &sh->state);
3676 set_bit(STRIPE_HANDLE, &sh->state);
3680 if (rcw && conf->mddev->queue)
3681 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3682 (unsigned long long)sh->sector,
3683 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3686 if (rcw > disks && rmw > disks &&
3687 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3688 set_bit(STRIPE_DELAYED, &sh->state);
3690 /* now if nothing is locked, and if we have enough data,
3691 * we can start a write request
3693 /* since handle_stripe can be called at any time we need to handle the
3694 * case where a compute block operation has been submitted and then a
3695 * subsequent call wants to start a write request. raid_run_ops only
3696 * handles the case where compute block and reconstruct are requested
3697 * simultaneously. If this is not the case then new writes need to be
3698 * held off until the compute completes.
3700 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3701 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3702 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3703 schedule_reconstruction(sh, s, rcw == 0, 0);
3706 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3707 struct stripe_head_state *s, int disks)
3709 struct r5dev *dev = NULL;
3711 BUG_ON(sh->batch_head);
3712 set_bit(STRIPE_HANDLE, &sh->state);
3714 switch (sh->check_state) {
3715 case check_state_idle:
3716 /* start a new check operation if there are no failures */
3717 if (s->failed == 0) {
3718 BUG_ON(s->uptodate != disks);
3719 sh->check_state = check_state_run;
3720 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3721 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3725 dev = &sh->dev[s->failed_num[0]];
3727 case check_state_compute_result:
3728 sh->check_state = check_state_idle;
3730 dev = &sh->dev[sh->pd_idx];
3732 /* check that a write has not made the stripe insync */
3733 if (test_bit(STRIPE_INSYNC, &sh->state))
3736 /* either failed parity check, or recovery is happening */
3737 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3738 BUG_ON(s->uptodate != disks);
3740 set_bit(R5_LOCKED, &dev->flags);
3742 set_bit(R5_Wantwrite, &dev->flags);
3744 clear_bit(STRIPE_DEGRADED, &sh->state);
3745 set_bit(STRIPE_INSYNC, &sh->state);
3747 case check_state_run:
3748 break; /* we will be called again upon completion */
3749 case check_state_check_result:
3750 sh->check_state = check_state_idle;
3752 /* if a failure occurred during the check operation, leave
3753 * STRIPE_INSYNC not set and let the stripe be handled again
3758 /* handle a successful check operation, if parity is correct
3759 * we are done. Otherwise update the mismatch count and repair
3760 * parity if !MD_RECOVERY_CHECK
3762 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3763 /* parity is correct (on disc,
3764 * not in buffer any more)
3766 set_bit(STRIPE_INSYNC, &sh->state);
3768 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3769 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3770 /* don't try to repair!! */
3771 set_bit(STRIPE_INSYNC, &sh->state);
3773 sh->check_state = check_state_compute_run;
3774 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3775 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3776 set_bit(R5_Wantcompute,
3777 &sh->dev[sh->pd_idx].flags);
3778 sh->ops.target = sh->pd_idx;
3779 sh->ops.target2 = -1;
3784 case check_state_compute_run:
3787 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3788 __func__, sh->check_state,
3789 (unsigned long long) sh->sector);
3794 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3795 struct stripe_head_state *s,
3798 int pd_idx = sh->pd_idx;
3799 int qd_idx = sh->qd_idx;
3802 BUG_ON(sh->batch_head);
3803 set_bit(STRIPE_HANDLE, &sh->state);
3805 BUG_ON(s->failed > 2);
3807 /* Want to check and possibly repair P and Q.
3808 * However there could be one 'failed' device, in which
3809 * case we can only check one of them, possibly using the
3810 * other to generate missing data
3813 switch (sh->check_state) {
3814 case check_state_idle:
3815 /* start a new check operation if there are < 2 failures */
3816 if (s->failed == s->q_failed) {
3817 /* The only possible failed device holds Q, so it
3818 * makes sense to check P (If anything else were failed,
3819 * we would have used P to recreate it).
3821 sh->check_state = check_state_run;
3823 if (!s->q_failed && s->failed < 2) {
3824 /* Q is not failed, and we didn't use it to generate
3825 * anything, so it makes sense to check it
3827 if (sh->check_state == check_state_run)
3828 sh->check_state = check_state_run_pq;
3830 sh->check_state = check_state_run_q;
3833 /* discard potentially stale zero_sum_result */
3834 sh->ops.zero_sum_result = 0;
3836 if (sh->check_state == check_state_run) {
3837 /* async_xor_zero_sum destroys the contents of P */
3838 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3841 if (sh->check_state >= check_state_run &&
3842 sh->check_state <= check_state_run_pq) {
3843 /* async_syndrome_zero_sum preserves P and Q, so
3844 * no need to mark them !uptodate here
3846 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3850 /* we have 2-disk failure */
3851 BUG_ON(s->failed != 2);
3853 case check_state_compute_result:
3854 sh->check_state = check_state_idle;
3856 /* check that a write has not made the stripe insync */
3857 if (test_bit(STRIPE_INSYNC, &sh->state))
3860 /* now write out any block on a failed drive,
3861 * or P or Q if they were recomputed
3863 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3864 if (s->failed == 2) {
3865 dev = &sh->dev[s->failed_num[1]];
3867 set_bit(R5_LOCKED, &dev->flags);
3868 set_bit(R5_Wantwrite, &dev->flags);
3870 if (s->failed >= 1) {
3871 dev = &sh->dev[s->failed_num[0]];
3873 set_bit(R5_LOCKED, &dev->flags);
3874 set_bit(R5_Wantwrite, &dev->flags);
3876 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3877 dev = &sh->dev[pd_idx];
3879 set_bit(R5_LOCKED, &dev->flags);
3880 set_bit(R5_Wantwrite, &dev->flags);
3882 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3883 dev = &sh->dev[qd_idx];
3885 set_bit(R5_LOCKED, &dev->flags);
3886 set_bit(R5_Wantwrite, &dev->flags);
3888 clear_bit(STRIPE_DEGRADED, &sh->state);
3890 set_bit(STRIPE_INSYNC, &sh->state);
3892 case check_state_run:
3893 case check_state_run_q:
3894 case check_state_run_pq:
3895 break; /* we will be called again upon completion */
3896 case check_state_check_result:
3897 sh->check_state = check_state_idle;
3899 /* handle a successful check operation, if parity is correct
3900 * we are done. Otherwise update the mismatch count and repair
3901 * parity if !MD_RECOVERY_CHECK
3903 if (sh->ops.zero_sum_result == 0) {
3904 /* both parities are correct */
3906 set_bit(STRIPE_INSYNC, &sh->state);
3908 /* in contrast to the raid5 case we can validate
3909 * parity, but still have a failure to write
3912 sh->check_state = check_state_compute_result;
3913 /* Returning at this point means that we may go
3914 * off and bring p and/or q uptodate again so
3915 * we make sure to check zero_sum_result again
3916 * to verify if p or q need writeback
3920 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3921 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3922 /* don't try to repair!! */
3923 set_bit(STRIPE_INSYNC, &sh->state);
3925 int *target = &sh->ops.target;
3927 sh->ops.target = -1;
3928 sh->ops.target2 = -1;
3929 sh->check_state = check_state_compute_run;
3930 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3931 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3932 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3933 set_bit(R5_Wantcompute,
3934 &sh->dev[pd_idx].flags);
3936 target = &sh->ops.target2;
3939 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3940 set_bit(R5_Wantcompute,
3941 &sh->dev[qd_idx].flags);
3948 case check_state_compute_run:
3951 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3952 __func__, sh->check_state,
3953 (unsigned long long) sh->sector);
3958 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3962 /* We have read all the blocks in this stripe and now we need to
3963 * copy some of them into a target stripe for expand.
3965 struct dma_async_tx_descriptor *tx = NULL;
3966 BUG_ON(sh->batch_head);
3967 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3968 for (i = 0; i < sh->disks; i++)
3969 if (i != sh->pd_idx && i != sh->qd_idx) {
3971 struct stripe_head *sh2;
3972 struct async_submit_ctl submit;
3974 sector_t bn = raid5_compute_blocknr(sh, i, 1);
3975 sector_t s = raid5_compute_sector(conf, bn, 0,
3977 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
3979 /* so far only the early blocks of this stripe
3980 * have been requested. When later blocks
3981 * get requested, we will try again
3984 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3985 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3986 /* must have already done this block */
3987 raid5_release_stripe(sh2);
3991 /* place all the copies on one channel */
3992 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3993 tx = async_memcpy(sh2->dev[dd_idx].page,
3994 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3997 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3998 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3999 for (j = 0; j < conf->raid_disks; j++)
4000 if (j != sh2->pd_idx &&
4002 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4004 if (j == conf->raid_disks) {
4005 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4006 set_bit(STRIPE_HANDLE, &sh2->state);
4008 raid5_release_stripe(sh2);
4011 /* done submitting copies, wait for them to complete */
4012 async_tx_quiesce(&tx);
4016 * handle_stripe - do things to a stripe.
4018 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4019 * state of various bits to see what needs to be done.
4021 * return some read requests which now have data
4022 * return some write requests which are safely on storage
4023 * schedule a read on some buffers
4024 * schedule a write of some buffers
4025 * return confirmation of parity correctness
4029 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4031 struct r5conf *conf = sh->raid_conf;
4032 int disks = sh->disks;
4035 int do_recovery = 0;
4037 memset(s, 0, sizeof(*s));
4039 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4040 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4041 s->failed_num[0] = -1;
4042 s->failed_num[1] = -1;
4043 s->log_failed = r5l_log_disk_error(conf);
4045 /* Now to look around and see what can be done */
4047 for (i=disks; i--; ) {
4048 struct md_rdev *rdev;
4055 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4057 dev->toread, dev->towrite, dev->written);
4058 /* maybe we can reply to a read
4060 * new wantfill requests are only permitted while
4061 * ops_complete_biofill is guaranteed to be inactive
4063 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4064 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4065 set_bit(R5_Wantfill, &dev->flags);
4067 /* now count some things */
4068 if (test_bit(R5_LOCKED, &dev->flags))
4070 if (test_bit(R5_UPTODATE, &dev->flags))
4072 if (test_bit(R5_Wantcompute, &dev->flags)) {
4074 BUG_ON(s->compute > 2);
4077 if (test_bit(R5_Wantfill, &dev->flags))
4079 else if (dev->toread)
4083 if (!test_bit(R5_OVERWRITE, &dev->flags))
4088 /* Prefer to use the replacement for reads, but only
4089 * if it is recovered enough and has no bad blocks.
4091 rdev = rcu_dereference(conf->disks[i].replacement);
4092 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4093 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4094 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4095 &first_bad, &bad_sectors))
4096 set_bit(R5_ReadRepl, &dev->flags);
4098 if (rdev && !test_bit(Faulty, &rdev->flags))
4099 set_bit(R5_NeedReplace, &dev->flags);
4101 clear_bit(R5_NeedReplace, &dev->flags);
4102 rdev = rcu_dereference(conf->disks[i].rdev);
4103 clear_bit(R5_ReadRepl, &dev->flags);
4105 if (rdev && test_bit(Faulty, &rdev->flags))
4108 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4109 &first_bad, &bad_sectors);
4110 if (s->blocked_rdev == NULL
4111 && (test_bit(Blocked, &rdev->flags)
4114 set_bit(BlockedBadBlocks,
4116 s->blocked_rdev = rdev;
4117 atomic_inc(&rdev->nr_pending);
4120 clear_bit(R5_Insync, &dev->flags);
4124 /* also not in-sync */
4125 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4126 test_bit(R5_UPTODATE, &dev->flags)) {
4127 /* treat as in-sync, but with a read error
4128 * which we can now try to correct
4130 set_bit(R5_Insync, &dev->flags);
4131 set_bit(R5_ReadError, &dev->flags);
4133 } else if (test_bit(In_sync, &rdev->flags))
4134 set_bit(R5_Insync, &dev->flags);
4135 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4136 /* in sync if before recovery_offset */
4137 set_bit(R5_Insync, &dev->flags);
4138 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4139 test_bit(R5_Expanded, &dev->flags))
4140 /* If we've reshaped into here, we assume it is Insync.
4141 * We will shortly update recovery_offset to make
4144 set_bit(R5_Insync, &dev->flags);
4146 if (test_bit(R5_WriteError, &dev->flags)) {
4147 /* This flag does not apply to '.replacement'
4148 * only to .rdev, so make sure to check that*/
4149 struct md_rdev *rdev2 = rcu_dereference(
4150 conf->disks[i].rdev);
4152 clear_bit(R5_Insync, &dev->flags);
4153 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4154 s->handle_bad_blocks = 1;
4155 atomic_inc(&rdev2->nr_pending);
4157 clear_bit(R5_WriteError, &dev->flags);
4159 if (test_bit(R5_MadeGood, &dev->flags)) {
4160 /* This flag does not apply to '.replacement'
4161 * only to .rdev, so make sure to check that*/
4162 struct md_rdev *rdev2 = rcu_dereference(
4163 conf->disks[i].rdev);
4164 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4165 s->handle_bad_blocks = 1;
4166 atomic_inc(&rdev2->nr_pending);
4168 clear_bit(R5_MadeGood, &dev->flags);
4170 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4171 struct md_rdev *rdev2 = rcu_dereference(
4172 conf->disks[i].replacement);
4173 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4174 s->handle_bad_blocks = 1;
4175 atomic_inc(&rdev2->nr_pending);
4177 clear_bit(R5_MadeGoodRepl, &dev->flags);
4179 if (!test_bit(R5_Insync, &dev->flags)) {
4180 /* The ReadError flag will just be confusing now */
4181 clear_bit(R5_ReadError, &dev->flags);
4182 clear_bit(R5_ReWrite, &dev->flags);
4184 if (test_bit(R5_ReadError, &dev->flags))
4185 clear_bit(R5_Insync, &dev->flags);
4186 if (!test_bit(R5_Insync, &dev->flags)) {
4188 s->failed_num[s->failed] = i;
4190 if (rdev && !test_bit(Faulty, &rdev->flags))
4194 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4195 /* If there is a failed device being replaced,
4196 * we must be recovering.
4197 * else if we are after recovery_cp, we must be syncing
4198 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4199 * else we can only be replacing
4200 * sync and recovery both need to read all devices, and so
4201 * use the same flag.
4204 sh->sector >= conf->mddev->recovery_cp ||
4205 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4213 static int clear_batch_ready(struct stripe_head *sh)
4215 /* Return '1' if this is a member of batch, or
4216 * '0' if it is a lone stripe or a head which can now be
4219 struct stripe_head *tmp;
4220 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4221 return (sh->batch_head && sh->batch_head != sh);
4222 spin_lock(&sh->stripe_lock);
4223 if (!sh->batch_head) {
4224 spin_unlock(&sh->stripe_lock);
4229 * this stripe could be added to a batch list before we check
4230 * BATCH_READY, skips it
4232 if (sh->batch_head != sh) {
4233 spin_unlock(&sh->stripe_lock);
4236 spin_lock(&sh->batch_lock);
4237 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4238 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4239 spin_unlock(&sh->batch_lock);
4240 spin_unlock(&sh->stripe_lock);
4243 * BATCH_READY is cleared, no new stripes can be added.
4244 * batch_list can be accessed without lock
4249 static void break_stripe_batch_list(struct stripe_head *head_sh,
4250 unsigned long handle_flags)
4252 struct stripe_head *sh, *next;
4256 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4258 list_del_init(&sh->batch_list);
4260 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4261 (1 << STRIPE_SYNCING) |
4262 (1 << STRIPE_REPLACED) |
4263 (1 << STRIPE_DELAYED) |
4264 (1 << STRIPE_BIT_DELAY) |
4265 (1 << STRIPE_FULL_WRITE) |
4266 (1 << STRIPE_BIOFILL_RUN) |
4267 (1 << STRIPE_COMPUTE_RUN) |
4268 (1 << STRIPE_OPS_REQ_PENDING) |
4269 (1 << STRIPE_DISCARD) |
4270 (1 << STRIPE_BATCH_READY) |
4271 (1 << STRIPE_BATCH_ERR) |
4272 (1 << STRIPE_BITMAP_PENDING)),
4273 "stripe state: %lx\n", sh->state);
4274 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4275 (1 << STRIPE_REPLACED)),
4276 "head stripe state: %lx\n", head_sh->state);
4278 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4279 (1 << STRIPE_PREREAD_ACTIVE) |
4280 (1 << STRIPE_DEGRADED)),
4281 head_sh->state & (1 << STRIPE_INSYNC));
4283 sh->check_state = head_sh->check_state;
4284 sh->reconstruct_state = head_sh->reconstruct_state;
4285 for (i = 0; i < sh->disks; i++) {
4286 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4288 sh->dev[i].flags = head_sh->dev[i].flags &
4289 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4291 spin_lock_irq(&sh->stripe_lock);
4292 sh->batch_head = NULL;
4293 spin_unlock_irq(&sh->stripe_lock);
4294 if (handle_flags == 0 ||
4295 sh->state & handle_flags)
4296 set_bit(STRIPE_HANDLE, &sh->state);
4297 raid5_release_stripe(sh);
4299 spin_lock_irq(&head_sh->stripe_lock);
4300 head_sh->batch_head = NULL;
4301 spin_unlock_irq(&head_sh->stripe_lock);
4302 for (i = 0; i < head_sh->disks; i++)
4303 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4305 if (head_sh->state & handle_flags)
4306 set_bit(STRIPE_HANDLE, &head_sh->state);
4309 wake_up(&head_sh->raid_conf->wait_for_overlap);
4312 static void handle_stripe(struct stripe_head *sh)
4314 struct stripe_head_state s;
4315 struct r5conf *conf = sh->raid_conf;
4318 int disks = sh->disks;
4319 struct r5dev *pdev, *qdev;
4321 clear_bit(STRIPE_HANDLE, &sh->state);
4322 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4323 /* already being handled, ensure it gets handled
4324 * again when current action finishes */
4325 set_bit(STRIPE_HANDLE, &sh->state);
4329 if (clear_batch_ready(sh) ) {
4330 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4334 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4335 break_stripe_batch_list(sh, 0);
4337 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4338 spin_lock(&sh->stripe_lock);
4339 /* Cannot process 'sync' concurrently with 'discard' */
4340 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4341 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4342 set_bit(STRIPE_SYNCING, &sh->state);
4343 clear_bit(STRIPE_INSYNC, &sh->state);
4344 clear_bit(STRIPE_REPLACED, &sh->state);
4346 spin_unlock(&sh->stripe_lock);
4348 clear_bit(STRIPE_DELAYED, &sh->state);
4350 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4351 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4352 (unsigned long long)sh->sector, sh->state,
4353 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4354 sh->check_state, sh->reconstruct_state);
4356 analyse_stripe(sh, &s);
4358 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4361 if (s.handle_bad_blocks) {
4362 set_bit(STRIPE_HANDLE, &sh->state);
4366 if (unlikely(s.blocked_rdev)) {
4367 if (s.syncing || s.expanding || s.expanded ||
4368 s.replacing || s.to_write || s.written) {
4369 set_bit(STRIPE_HANDLE, &sh->state);
4372 /* There is nothing for the blocked_rdev to block */
4373 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4374 s.blocked_rdev = NULL;
4377 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4378 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4379 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4382 pr_debug("locked=%d uptodate=%d to_read=%d"
4383 " to_write=%d failed=%d failed_num=%d,%d\n",
4384 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4385 s.failed_num[0], s.failed_num[1]);
4386 /* check if the array has lost more than max_degraded devices and,
4387 * if so, some requests might need to be failed.
4389 if (s.failed > conf->max_degraded || s.log_failed) {
4390 sh->check_state = 0;
4391 sh->reconstruct_state = 0;
4392 break_stripe_batch_list(sh, 0);
4393 if (s.to_read+s.to_write+s.written)
4394 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4395 if (s.syncing + s.replacing)
4396 handle_failed_sync(conf, sh, &s);
4399 /* Now we check to see if any write operations have recently
4403 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4405 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4406 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4407 sh->reconstruct_state = reconstruct_state_idle;
4409 /* All the 'written' buffers and the parity block are ready to
4410 * be written back to disk
4412 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4413 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4414 BUG_ON(sh->qd_idx >= 0 &&
4415 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4416 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4417 for (i = disks; i--; ) {
4418 struct r5dev *dev = &sh->dev[i];
4419 if (test_bit(R5_LOCKED, &dev->flags) &&
4420 (i == sh->pd_idx || i == sh->qd_idx ||
4422 pr_debug("Writing block %d\n", i);
4423 set_bit(R5_Wantwrite, &dev->flags);
4428 if (!test_bit(R5_Insync, &dev->flags) ||
4429 ((i == sh->pd_idx || i == sh->qd_idx) &&
4431 set_bit(STRIPE_INSYNC, &sh->state);
4434 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4435 s.dec_preread_active = 1;
4439 * might be able to return some write requests if the parity blocks
4440 * are safe, or on a failed drive
4442 pdev = &sh->dev[sh->pd_idx];
4443 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4444 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4445 qdev = &sh->dev[sh->qd_idx];
4446 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4447 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4451 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4452 && !test_bit(R5_LOCKED, &pdev->flags)
4453 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4454 test_bit(R5_Discard, &pdev->flags))))) &&
4455 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4456 && !test_bit(R5_LOCKED, &qdev->flags)
4457 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4458 test_bit(R5_Discard, &qdev->flags))))))
4459 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4461 /* Now we might consider reading some blocks, either to check/generate
4462 * parity, or to satisfy requests
4463 * or to load a block that is being partially written.
4465 if (s.to_read || s.non_overwrite
4466 || (conf->level == 6 && s.to_write && s.failed)
4467 || (s.syncing && (s.uptodate + s.compute < disks))
4470 handle_stripe_fill(sh, &s, disks);
4472 /* Now to consider new write requests and what else, if anything
4473 * should be read. We do not handle new writes when:
4474 * 1/ A 'write' operation (copy+xor) is already in flight.
4475 * 2/ A 'check' operation is in flight, as it may clobber the parity
4478 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4479 handle_stripe_dirtying(conf, sh, &s, disks);
4481 /* maybe we need to check and possibly fix the parity for this stripe
4482 * Any reads will already have been scheduled, so we just see if enough
4483 * data is available. The parity check is held off while parity
4484 * dependent operations are in flight.
4486 if (sh->check_state ||
4487 (s.syncing && s.locked == 0 &&
4488 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4489 !test_bit(STRIPE_INSYNC, &sh->state))) {
4490 if (conf->level == 6)
4491 handle_parity_checks6(conf, sh, &s, disks);
4493 handle_parity_checks5(conf, sh, &s, disks);
4496 if ((s.replacing || s.syncing) && s.locked == 0
4497 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4498 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4499 /* Write out to replacement devices where possible */
4500 for (i = 0; i < conf->raid_disks; i++)
4501 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4502 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4503 set_bit(R5_WantReplace, &sh->dev[i].flags);
4504 set_bit(R5_LOCKED, &sh->dev[i].flags);
4508 set_bit(STRIPE_INSYNC, &sh->state);
4509 set_bit(STRIPE_REPLACED, &sh->state);
4511 if ((s.syncing || s.replacing) && s.locked == 0 &&
4512 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4513 test_bit(STRIPE_INSYNC, &sh->state)) {
4514 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4515 clear_bit(STRIPE_SYNCING, &sh->state);
4516 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4517 wake_up(&conf->wait_for_overlap);
4520 /* If the failed drives are just a ReadError, then we might need
4521 * to progress the repair/check process
4523 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4524 for (i = 0; i < s.failed; i++) {
4525 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4526 if (test_bit(R5_ReadError, &dev->flags)
4527 && !test_bit(R5_LOCKED, &dev->flags)
4528 && test_bit(R5_UPTODATE, &dev->flags)
4530 if (!test_bit(R5_ReWrite, &dev->flags)) {
4531 set_bit(R5_Wantwrite, &dev->flags);
4532 set_bit(R5_ReWrite, &dev->flags);
4533 set_bit(R5_LOCKED, &dev->flags);
4536 /* let's read it back */
4537 set_bit(R5_Wantread, &dev->flags);
4538 set_bit(R5_LOCKED, &dev->flags);
4544 /* Finish reconstruct operations initiated by the expansion process */
4545 if (sh->reconstruct_state == reconstruct_state_result) {
4546 struct stripe_head *sh_src
4547 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4548 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4549 /* sh cannot be written until sh_src has been read.
4550 * so arrange for sh to be delayed a little
4552 set_bit(STRIPE_DELAYED, &sh->state);
4553 set_bit(STRIPE_HANDLE, &sh->state);
4554 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4556 atomic_inc(&conf->preread_active_stripes);
4557 raid5_release_stripe(sh_src);
4561 raid5_release_stripe(sh_src);
4563 sh->reconstruct_state = reconstruct_state_idle;
4564 clear_bit(STRIPE_EXPANDING, &sh->state);
4565 for (i = conf->raid_disks; i--; ) {
4566 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4567 set_bit(R5_LOCKED, &sh->dev[i].flags);
4572 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4573 !sh->reconstruct_state) {
4574 /* Need to write out all blocks after computing parity */
4575 sh->disks = conf->raid_disks;
4576 stripe_set_idx(sh->sector, conf, 0, sh);
4577 schedule_reconstruction(sh, &s, 1, 1);
4578 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4579 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4580 atomic_dec(&conf->reshape_stripes);
4581 wake_up(&conf->wait_for_overlap);
4582 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4585 if (s.expanding && s.locked == 0 &&
4586 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4587 handle_stripe_expansion(conf, sh);
4590 /* wait for this device to become unblocked */
4591 if (unlikely(s.blocked_rdev)) {
4592 if (conf->mddev->external)
4593 md_wait_for_blocked_rdev(s.blocked_rdev,
4596 /* Internal metadata will immediately
4597 * be written by raid5d, so we don't
4598 * need to wait here.
4600 rdev_dec_pending(s.blocked_rdev,
4604 if (s.handle_bad_blocks)
4605 for (i = disks; i--; ) {
4606 struct md_rdev *rdev;
4607 struct r5dev *dev = &sh->dev[i];
4608 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4609 /* We own a safe reference to the rdev */
4610 rdev = conf->disks[i].rdev;
4611 if (!rdev_set_badblocks(rdev, sh->sector,
4613 md_error(conf->mddev, rdev);
4614 rdev_dec_pending(rdev, conf->mddev);
4616 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4617 rdev = conf->disks[i].rdev;
4618 rdev_clear_badblocks(rdev, sh->sector,
4620 rdev_dec_pending(rdev, conf->mddev);
4622 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4623 rdev = conf->disks[i].replacement;
4625 /* rdev have been moved down */
4626 rdev = conf->disks[i].rdev;
4627 rdev_clear_badblocks(rdev, sh->sector,
4629 rdev_dec_pending(rdev, conf->mddev);
4634 raid_run_ops(sh, s.ops_request);
4638 if (s.dec_preread_active) {
4639 /* We delay this until after ops_run_io so that if make_request
4640 * is waiting on a flush, it won't continue until the writes
4641 * have actually been submitted.
4643 atomic_dec(&conf->preread_active_stripes);
4644 if (atomic_read(&conf->preread_active_stripes) <
4646 md_wakeup_thread(conf->mddev->thread);
4649 if (!bio_list_empty(&s.return_bi)) {
4650 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags) &&
4651 (s.failed <= conf->max_degraded ||
4652 conf->mddev->external == 0)) {
4653 spin_lock_irq(&conf->device_lock);
4654 bio_list_merge(&conf->return_bi, &s.return_bi);
4655 spin_unlock_irq(&conf->device_lock);
4656 md_wakeup_thread(conf->mddev->thread);
4658 return_io(&s.return_bi);
4661 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4664 static void raid5_activate_delayed(struct r5conf *conf)
4666 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4667 while (!list_empty(&conf->delayed_list)) {
4668 struct list_head *l = conf->delayed_list.next;
4669 struct stripe_head *sh;
4670 sh = list_entry(l, struct stripe_head, lru);
4672 clear_bit(STRIPE_DELAYED, &sh->state);
4673 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4674 atomic_inc(&conf->preread_active_stripes);
4675 list_add_tail(&sh->lru, &conf->hold_list);
4676 raid5_wakeup_stripe_thread(sh);
4681 static void activate_bit_delay(struct r5conf *conf,
4682 struct list_head *temp_inactive_list)
4684 /* device_lock is held */
4685 struct list_head head;
4686 list_add(&head, &conf->bitmap_list);
4687 list_del_init(&conf->bitmap_list);
4688 while (!list_empty(&head)) {
4689 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4691 list_del_init(&sh->lru);
4692 atomic_inc(&sh->count);
4693 hash = sh->hash_lock_index;
4694 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4698 static int raid5_congested(struct mddev *mddev, int bits)
4700 struct r5conf *conf = mddev->private;
4702 /* No difference between reads and writes. Just check
4703 * how busy the stripe_cache is
4706 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4710 if (atomic_read(&conf->empty_inactive_list_nr))
4716 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4718 struct r5conf *conf = mddev->private;
4719 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4720 unsigned int chunk_sectors;
4721 unsigned int bio_sectors = bio_sectors(bio);
4723 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4724 return chunk_sectors >=
4725 ((sector & (chunk_sectors - 1)) + bio_sectors);
4729 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4730 * later sampled by raid5d.
4732 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4734 unsigned long flags;
4736 spin_lock_irqsave(&conf->device_lock, flags);
4738 bi->bi_next = conf->retry_read_aligned_list;
4739 conf->retry_read_aligned_list = bi;
4741 spin_unlock_irqrestore(&conf->device_lock, flags);
4742 md_wakeup_thread(conf->mddev->thread);
4745 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4749 bi = conf->retry_read_aligned;
4751 conf->retry_read_aligned = NULL;
4754 bi = conf->retry_read_aligned_list;
4756 conf->retry_read_aligned_list = bi->bi_next;
4759 * this sets the active strip count to 1 and the processed
4760 * strip count to zero (upper 8 bits)
4762 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4769 * The "raid5_align_endio" should check if the read succeeded and if it
4770 * did, call bio_endio on the original bio (having bio_put the new bio
4772 * If the read failed..
4774 static void raid5_align_endio(struct bio *bi)
4776 struct bio* raid_bi = bi->bi_private;
4777 struct mddev *mddev;
4778 struct r5conf *conf;
4779 struct md_rdev *rdev;
4780 int error = bi->bi_error;
4784 rdev = (void*)raid_bi->bi_next;
4785 raid_bi->bi_next = NULL;
4786 mddev = rdev->mddev;
4787 conf = mddev->private;
4789 rdev_dec_pending(rdev, conf->mddev);
4792 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4795 if (atomic_dec_and_test(&conf->active_aligned_reads))
4796 wake_up(&conf->wait_for_quiescent);
4800 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4802 add_bio_to_retry(raid_bi, conf);
4805 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4807 struct r5conf *conf = mddev->private;
4809 struct bio* align_bi;
4810 struct md_rdev *rdev;
4811 sector_t end_sector;
4813 if (!in_chunk_boundary(mddev, raid_bio)) {
4814 pr_debug("%s: non aligned\n", __func__);
4818 * use bio_clone_mddev to make a copy of the bio
4820 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4824 * set bi_end_io to a new function, and set bi_private to the
4827 align_bi->bi_end_io = raid5_align_endio;
4828 align_bi->bi_private = raid_bio;
4832 align_bi->bi_iter.bi_sector =
4833 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4836 end_sector = bio_end_sector(align_bi);
4838 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4839 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4840 rdev->recovery_offset < end_sector) {
4841 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4843 (test_bit(Faulty, &rdev->flags) ||
4844 !(test_bit(In_sync, &rdev->flags) ||
4845 rdev->recovery_offset >= end_sector)))
4852 atomic_inc(&rdev->nr_pending);
4854 raid_bio->bi_next = (void*)rdev;
4855 align_bi->bi_bdev = rdev->bdev;
4856 bio_clear_flag(align_bi, BIO_SEG_VALID);
4858 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4859 bio_sectors(align_bi),
4860 &first_bad, &bad_sectors)) {
4862 rdev_dec_pending(rdev, mddev);
4866 /* No reshape active, so we can trust rdev->data_offset */
4867 align_bi->bi_iter.bi_sector += rdev->data_offset;
4869 spin_lock_irq(&conf->device_lock);
4870 wait_event_lock_irq(conf->wait_for_quiescent,
4873 atomic_inc(&conf->active_aligned_reads);
4874 spin_unlock_irq(&conf->device_lock);
4877 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4878 align_bi, disk_devt(mddev->gendisk),
4879 raid_bio->bi_iter.bi_sector);
4880 generic_make_request(align_bi);
4889 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4894 sector_t sector = raid_bio->bi_iter.bi_sector;
4895 unsigned chunk_sects = mddev->chunk_sectors;
4896 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4898 if (sectors < bio_sectors(raid_bio)) {
4899 split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4900 bio_chain(split, raid_bio);
4904 if (!raid5_read_one_chunk(mddev, split)) {
4905 if (split != raid_bio)
4906 generic_make_request(raid_bio);
4909 } while (split != raid_bio);
4914 /* __get_priority_stripe - get the next stripe to process
4916 * Full stripe writes are allowed to pass preread active stripes up until
4917 * the bypass_threshold is exceeded. In general the bypass_count
4918 * increments when the handle_list is handled before the hold_list; however, it
4919 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4920 * stripe with in flight i/o. The bypass_count will be reset when the
4921 * head of the hold_list has changed, i.e. the head was promoted to the
4924 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4926 struct stripe_head *sh = NULL, *tmp;
4927 struct list_head *handle_list = NULL;
4928 struct r5worker_group *wg = NULL;
4930 if (conf->worker_cnt_per_group == 0) {
4931 handle_list = &conf->handle_list;
4932 } else if (group != ANY_GROUP) {
4933 handle_list = &conf->worker_groups[group].handle_list;
4934 wg = &conf->worker_groups[group];
4937 for (i = 0; i < conf->group_cnt; i++) {
4938 handle_list = &conf->worker_groups[i].handle_list;
4939 wg = &conf->worker_groups[i];
4940 if (!list_empty(handle_list))
4945 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4947 list_empty(handle_list) ? "empty" : "busy",
4948 list_empty(&conf->hold_list) ? "empty" : "busy",
4949 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4951 if (!list_empty(handle_list)) {
4952 sh = list_entry(handle_list->next, typeof(*sh), lru);
4954 if (list_empty(&conf->hold_list))
4955 conf->bypass_count = 0;
4956 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4957 if (conf->hold_list.next == conf->last_hold)
4958 conf->bypass_count++;
4960 conf->last_hold = conf->hold_list.next;
4961 conf->bypass_count -= conf->bypass_threshold;
4962 if (conf->bypass_count < 0)
4963 conf->bypass_count = 0;
4966 } else if (!list_empty(&conf->hold_list) &&
4967 ((conf->bypass_threshold &&
4968 conf->bypass_count > conf->bypass_threshold) ||
4969 atomic_read(&conf->pending_full_writes) == 0)) {
4971 list_for_each_entry(tmp, &conf->hold_list, lru) {
4972 if (conf->worker_cnt_per_group == 0 ||
4973 group == ANY_GROUP ||
4974 !cpu_online(tmp->cpu) ||
4975 cpu_to_group(tmp->cpu) == group) {
4982 conf->bypass_count -= conf->bypass_threshold;
4983 if (conf->bypass_count < 0)
4984 conf->bypass_count = 0;
4996 list_del_init(&sh->lru);
4997 BUG_ON(atomic_inc_return(&sh->count) != 1);
5001 struct raid5_plug_cb {
5002 struct blk_plug_cb cb;
5003 struct list_head list;
5004 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5007 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5009 struct raid5_plug_cb *cb = container_of(
5010 blk_cb, struct raid5_plug_cb, cb);
5011 struct stripe_head *sh;
5012 struct mddev *mddev = cb->cb.data;
5013 struct r5conf *conf = mddev->private;
5017 if (cb->list.next && !list_empty(&cb->list)) {
5018 spin_lock_irq(&conf->device_lock);
5019 while (!list_empty(&cb->list)) {
5020 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5021 list_del_init(&sh->lru);
5023 * avoid race release_stripe_plug() sees
5024 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5025 * is still in our list
5027 smp_mb__before_atomic();
5028 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5030 * STRIPE_ON_RELEASE_LIST could be set here. In that
5031 * case, the count is always > 1 here
5033 hash = sh->hash_lock_index;
5034 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5037 spin_unlock_irq(&conf->device_lock);
5039 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5040 NR_STRIPE_HASH_LOCKS);
5042 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5046 static void release_stripe_plug(struct mddev *mddev,
5047 struct stripe_head *sh)
5049 struct blk_plug_cb *blk_cb = blk_check_plugged(
5050 raid5_unplug, mddev,
5051 sizeof(struct raid5_plug_cb));
5052 struct raid5_plug_cb *cb;
5055 raid5_release_stripe(sh);
5059 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5061 if (cb->list.next == NULL) {
5063 INIT_LIST_HEAD(&cb->list);
5064 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5065 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5068 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5069 list_add_tail(&sh->lru, &cb->list);
5071 raid5_release_stripe(sh);
5074 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5076 struct r5conf *conf = mddev->private;
5077 sector_t logical_sector, last_sector;
5078 struct stripe_head *sh;
5082 if (mddev->reshape_position != MaxSector)
5083 /* Skip discard while reshape is happening */
5086 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5087 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5090 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5092 stripe_sectors = conf->chunk_sectors *
5093 (conf->raid_disks - conf->max_degraded);
5094 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5096 sector_div(last_sector, stripe_sectors);
5098 logical_sector *= conf->chunk_sectors;
5099 last_sector *= conf->chunk_sectors;
5101 for (; logical_sector < last_sector;
5102 logical_sector += STRIPE_SECTORS) {
5106 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5107 prepare_to_wait(&conf->wait_for_overlap, &w,
5108 TASK_UNINTERRUPTIBLE);
5109 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5110 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5111 raid5_release_stripe(sh);
5115 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5116 spin_lock_irq(&sh->stripe_lock);
5117 for (d = 0; d < conf->raid_disks; d++) {
5118 if (d == sh->pd_idx || d == sh->qd_idx)
5120 if (sh->dev[d].towrite || sh->dev[d].toread) {
5121 set_bit(R5_Overlap, &sh->dev[d].flags);
5122 spin_unlock_irq(&sh->stripe_lock);
5123 raid5_release_stripe(sh);
5128 set_bit(STRIPE_DISCARD, &sh->state);
5129 finish_wait(&conf->wait_for_overlap, &w);
5130 sh->overwrite_disks = 0;
5131 for (d = 0; d < conf->raid_disks; d++) {
5132 if (d == sh->pd_idx || d == sh->qd_idx)
5134 sh->dev[d].towrite = bi;
5135 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5136 raid5_inc_bi_active_stripes(bi);
5137 sh->overwrite_disks++;
5139 spin_unlock_irq(&sh->stripe_lock);
5140 if (conf->mddev->bitmap) {
5142 d < conf->raid_disks - conf->max_degraded;
5144 bitmap_startwrite(mddev->bitmap,
5148 sh->bm_seq = conf->seq_flush + 1;
5149 set_bit(STRIPE_BIT_DELAY, &sh->state);
5152 set_bit(STRIPE_HANDLE, &sh->state);
5153 clear_bit(STRIPE_DELAYED, &sh->state);
5154 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5155 atomic_inc(&conf->preread_active_stripes);
5156 release_stripe_plug(mddev, sh);
5159 remaining = raid5_dec_bi_active_stripes(bi);
5160 if (remaining == 0) {
5161 md_write_end(mddev);
5166 static void raid5_make_request(struct mddev *mddev, struct bio * bi)
5168 struct r5conf *conf = mddev->private;
5170 sector_t new_sector;
5171 sector_t logical_sector, last_sector;
5172 struct stripe_head *sh;
5173 const int rw = bio_data_dir(bi);
5178 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5179 int ret = r5l_handle_flush_request(conf->log, bi);
5183 if (ret == -ENODEV) {
5184 md_flush_request(mddev, bi);
5187 /* ret == -EAGAIN, fallback */
5190 md_write_start(mddev, bi);
5193 * If array is degraded, better not do chunk aligned read because
5194 * later we might have to read it again in order to reconstruct
5195 * data on failed drives.
5197 if (rw == READ && mddev->degraded == 0 &&
5198 mddev->reshape_position == MaxSector) {
5199 bi = chunk_aligned_read(mddev, bi);
5204 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5205 make_discard_request(mddev, bi);
5209 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5210 last_sector = bio_end_sector(bi);
5212 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5214 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5215 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5221 seq = read_seqcount_begin(&conf->gen_lock);
5224 prepare_to_wait(&conf->wait_for_overlap, &w,
5225 TASK_UNINTERRUPTIBLE);
5226 if (unlikely(conf->reshape_progress != MaxSector)) {
5227 /* spinlock is needed as reshape_progress may be
5228 * 64bit on a 32bit platform, and so it might be
5229 * possible to see a half-updated value
5230 * Of course reshape_progress could change after
5231 * the lock is dropped, so once we get a reference
5232 * to the stripe that we think it is, we will have
5235 spin_lock_irq(&conf->device_lock);
5236 if (mddev->reshape_backwards
5237 ? logical_sector < conf->reshape_progress
5238 : logical_sector >= conf->reshape_progress) {
5241 if (mddev->reshape_backwards
5242 ? logical_sector < conf->reshape_safe
5243 : logical_sector >= conf->reshape_safe) {
5244 spin_unlock_irq(&conf->device_lock);
5250 spin_unlock_irq(&conf->device_lock);
5253 new_sector = raid5_compute_sector(conf, logical_sector,
5256 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5257 (unsigned long long)new_sector,
5258 (unsigned long long)logical_sector);
5260 sh = raid5_get_active_stripe(conf, new_sector, previous,
5261 (bi->bi_opf & REQ_RAHEAD), 0);
5263 if (unlikely(previous)) {
5264 /* expansion might have moved on while waiting for a
5265 * stripe, so we must do the range check again.
5266 * Expansion could still move past after this
5267 * test, but as we are holding a reference to
5268 * 'sh', we know that if that happens,
5269 * STRIPE_EXPANDING will get set and the expansion
5270 * won't proceed until we finish with the stripe.
5273 spin_lock_irq(&conf->device_lock);
5274 if (mddev->reshape_backwards
5275 ? logical_sector >= conf->reshape_progress
5276 : logical_sector < conf->reshape_progress)
5277 /* mismatch, need to try again */
5279 spin_unlock_irq(&conf->device_lock);
5281 raid5_release_stripe(sh);
5287 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5288 /* Might have got the wrong stripe_head
5291 raid5_release_stripe(sh);
5296 logical_sector >= mddev->suspend_lo &&
5297 logical_sector < mddev->suspend_hi) {
5298 raid5_release_stripe(sh);
5299 /* As the suspend_* range is controlled by
5300 * userspace, we want an interruptible
5303 prepare_to_wait(&conf->wait_for_overlap,
5304 &w, TASK_INTERRUPTIBLE);
5305 if (logical_sector >= mddev->suspend_lo &&
5306 logical_sector < mddev->suspend_hi) {
5309 sigprocmask(SIG_BLOCK, &full, &old);
5311 sigprocmask(SIG_SETMASK, &old, NULL);
5317 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5318 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5319 /* Stripe is busy expanding or
5320 * add failed due to overlap. Flush everything
5323 md_wakeup_thread(mddev->thread);
5324 raid5_release_stripe(sh);
5329 set_bit(STRIPE_HANDLE, &sh->state);
5330 clear_bit(STRIPE_DELAYED, &sh->state);
5331 if ((!sh->batch_head || sh == sh->batch_head) &&
5332 (bi->bi_opf & REQ_SYNC) &&
5333 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5334 atomic_inc(&conf->preread_active_stripes);
5335 release_stripe_plug(mddev, sh);
5337 /* cannot get stripe for read-ahead, just give-up */
5338 bi->bi_error = -EIO;
5342 finish_wait(&conf->wait_for_overlap, &w);
5344 remaining = raid5_dec_bi_active_stripes(bi);
5345 if (remaining == 0) {
5348 md_write_end(mddev);
5350 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5356 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5358 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5360 /* reshaping is quite different to recovery/resync so it is
5361 * handled quite separately ... here.
5363 * On each call to sync_request, we gather one chunk worth of
5364 * destination stripes and flag them as expanding.
5365 * Then we find all the source stripes and request reads.
5366 * As the reads complete, handle_stripe will copy the data
5367 * into the destination stripe and release that stripe.
5369 struct r5conf *conf = mddev->private;
5370 struct stripe_head *sh;
5371 sector_t first_sector, last_sector;
5372 int raid_disks = conf->previous_raid_disks;
5373 int data_disks = raid_disks - conf->max_degraded;
5374 int new_data_disks = conf->raid_disks - conf->max_degraded;
5377 sector_t writepos, readpos, safepos;
5378 sector_t stripe_addr;
5379 int reshape_sectors;
5380 struct list_head stripes;
5383 if (sector_nr == 0) {
5384 /* If restarting in the middle, skip the initial sectors */
5385 if (mddev->reshape_backwards &&
5386 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5387 sector_nr = raid5_size(mddev, 0, 0)
5388 - conf->reshape_progress;
5389 } else if (mddev->reshape_backwards &&
5390 conf->reshape_progress == MaxSector) {
5391 /* shouldn't happen, but just in case, finish up.*/
5392 sector_nr = MaxSector;
5393 } else if (!mddev->reshape_backwards &&
5394 conf->reshape_progress > 0)
5395 sector_nr = conf->reshape_progress;
5396 sector_div(sector_nr, new_data_disks);
5398 mddev->curr_resync_completed = sector_nr;
5399 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5406 /* We need to process a full chunk at a time.
5407 * If old and new chunk sizes differ, we need to process the
5411 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5413 /* We update the metadata at least every 10 seconds, or when
5414 * the data about to be copied would over-write the source of
5415 * the data at the front of the range. i.e. one new_stripe
5416 * along from reshape_progress new_maps to after where
5417 * reshape_safe old_maps to
5419 writepos = conf->reshape_progress;
5420 sector_div(writepos, new_data_disks);
5421 readpos = conf->reshape_progress;
5422 sector_div(readpos, data_disks);
5423 safepos = conf->reshape_safe;
5424 sector_div(safepos, data_disks);
5425 if (mddev->reshape_backwards) {
5426 BUG_ON(writepos < reshape_sectors);
5427 writepos -= reshape_sectors;
5428 readpos += reshape_sectors;
5429 safepos += reshape_sectors;
5431 writepos += reshape_sectors;
5432 /* readpos and safepos are worst-case calculations.
5433 * A negative number is overly pessimistic, and causes
5434 * obvious problems for unsigned storage. So clip to 0.
5436 readpos -= min_t(sector_t, reshape_sectors, readpos);
5437 safepos -= min_t(sector_t, reshape_sectors, safepos);
5440 /* Having calculated the 'writepos' possibly use it
5441 * to set 'stripe_addr' which is where we will write to.
5443 if (mddev->reshape_backwards) {
5444 BUG_ON(conf->reshape_progress == 0);
5445 stripe_addr = writepos;
5446 BUG_ON((mddev->dev_sectors &
5447 ~((sector_t)reshape_sectors - 1))
5448 - reshape_sectors - stripe_addr
5451 BUG_ON(writepos != sector_nr + reshape_sectors);
5452 stripe_addr = sector_nr;
5455 /* 'writepos' is the most advanced device address we might write.
5456 * 'readpos' is the least advanced device address we might read.
5457 * 'safepos' is the least address recorded in the metadata as having
5459 * If there is a min_offset_diff, these are adjusted either by
5460 * increasing the safepos/readpos if diff is negative, or
5461 * increasing writepos if diff is positive.
5462 * If 'readpos' is then behind 'writepos', there is no way that we can
5463 * ensure safety in the face of a crash - that must be done by userspace
5464 * making a backup of the data. So in that case there is no particular
5465 * rush to update metadata.
5466 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5467 * update the metadata to advance 'safepos' to match 'readpos' so that
5468 * we can be safe in the event of a crash.
5469 * So we insist on updating metadata if safepos is behind writepos and
5470 * readpos is beyond writepos.
5471 * In any case, update the metadata every 10 seconds.
5472 * Maybe that number should be configurable, but I'm not sure it is
5473 * worth it.... maybe it could be a multiple of safemode_delay???
5475 if (conf->min_offset_diff < 0) {
5476 safepos += -conf->min_offset_diff;
5477 readpos += -conf->min_offset_diff;
5479 writepos += conf->min_offset_diff;
5481 if ((mddev->reshape_backwards
5482 ? (safepos > writepos && readpos < writepos)
5483 : (safepos < writepos && readpos > writepos)) ||
5484 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5485 /* Cannot proceed until we've updated the superblock... */
5486 wait_event(conf->wait_for_overlap,
5487 atomic_read(&conf->reshape_stripes)==0
5488 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5489 if (atomic_read(&conf->reshape_stripes) != 0)
5491 mddev->reshape_position = conf->reshape_progress;
5492 mddev->curr_resync_completed = sector_nr;
5493 conf->reshape_checkpoint = jiffies;
5494 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5495 md_wakeup_thread(mddev->thread);
5496 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5497 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5498 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5500 spin_lock_irq(&conf->device_lock);
5501 conf->reshape_safe = mddev->reshape_position;
5502 spin_unlock_irq(&conf->device_lock);
5503 wake_up(&conf->wait_for_overlap);
5504 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5507 INIT_LIST_HEAD(&stripes);
5508 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5510 int skipped_disk = 0;
5511 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5512 set_bit(STRIPE_EXPANDING, &sh->state);
5513 atomic_inc(&conf->reshape_stripes);
5514 /* If any of this stripe is beyond the end of the old
5515 * array, then we need to zero those blocks
5517 for (j=sh->disks; j--;) {
5519 if (j == sh->pd_idx)
5521 if (conf->level == 6 &&
5524 s = raid5_compute_blocknr(sh, j, 0);
5525 if (s < raid5_size(mddev, 0, 0)) {
5529 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5530 set_bit(R5_Expanded, &sh->dev[j].flags);
5531 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5533 if (!skipped_disk) {
5534 set_bit(STRIPE_EXPAND_READY, &sh->state);
5535 set_bit(STRIPE_HANDLE, &sh->state);
5537 list_add(&sh->lru, &stripes);
5539 spin_lock_irq(&conf->device_lock);
5540 if (mddev->reshape_backwards)
5541 conf->reshape_progress -= reshape_sectors * new_data_disks;
5543 conf->reshape_progress += reshape_sectors * new_data_disks;
5544 spin_unlock_irq(&conf->device_lock);
5545 /* Ok, those stripe are ready. We can start scheduling
5546 * reads on the source stripes.
5547 * The source stripes are determined by mapping the first and last
5548 * block on the destination stripes.
5551 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5554 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5555 * new_data_disks - 1),
5557 if (last_sector >= mddev->dev_sectors)
5558 last_sector = mddev->dev_sectors - 1;
5559 while (first_sector <= last_sector) {
5560 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5561 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5562 set_bit(STRIPE_HANDLE, &sh->state);
5563 raid5_release_stripe(sh);
5564 first_sector += STRIPE_SECTORS;
5566 /* Now that the sources are clearly marked, we can release
5567 * the destination stripes
5569 while (!list_empty(&stripes)) {
5570 sh = list_entry(stripes.next, struct stripe_head, lru);
5571 list_del_init(&sh->lru);
5572 raid5_release_stripe(sh);
5574 /* If this takes us to the resync_max point where we have to pause,
5575 * then we need to write out the superblock.
5577 sector_nr += reshape_sectors;
5578 retn = reshape_sectors;
5580 if (mddev->curr_resync_completed > mddev->resync_max ||
5581 (sector_nr - mddev->curr_resync_completed) * 2
5582 >= mddev->resync_max - mddev->curr_resync_completed) {
5583 /* Cannot proceed until we've updated the superblock... */
5584 wait_event(conf->wait_for_overlap,
5585 atomic_read(&conf->reshape_stripes) == 0
5586 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5587 if (atomic_read(&conf->reshape_stripes) != 0)
5589 mddev->reshape_position = conf->reshape_progress;
5590 mddev->curr_resync_completed = sector_nr;
5591 conf->reshape_checkpoint = jiffies;
5592 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5593 md_wakeup_thread(mddev->thread);
5594 wait_event(mddev->sb_wait,
5595 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5596 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5597 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5599 spin_lock_irq(&conf->device_lock);
5600 conf->reshape_safe = mddev->reshape_position;
5601 spin_unlock_irq(&conf->device_lock);
5602 wake_up(&conf->wait_for_overlap);
5603 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5609 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
5612 struct r5conf *conf = mddev->private;
5613 struct stripe_head *sh;
5614 sector_t max_sector = mddev->dev_sectors;
5615 sector_t sync_blocks;
5616 int still_degraded = 0;
5619 if (sector_nr >= max_sector) {
5620 /* just being told to finish up .. nothing much to do */
5622 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5627 if (mddev->curr_resync < max_sector) /* aborted */
5628 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5630 else /* completed sync */
5632 bitmap_close_sync(mddev->bitmap);
5637 /* Allow raid5_quiesce to complete */
5638 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5640 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5641 return reshape_request(mddev, sector_nr, skipped);
5643 /* No need to check resync_max as we never do more than one
5644 * stripe, and as resync_max will always be on a chunk boundary,
5645 * if the check in md_do_sync didn't fire, there is no chance
5646 * of overstepping resync_max here
5649 /* if there is too many failed drives and we are trying
5650 * to resync, then assert that we are finished, because there is
5651 * nothing we can do.
5653 if (mddev->degraded >= conf->max_degraded &&
5654 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5655 sector_t rv = mddev->dev_sectors - sector_nr;
5659 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5661 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5662 sync_blocks >= STRIPE_SECTORS) {
5663 /* we can skip this block, and probably more */
5664 sync_blocks /= STRIPE_SECTORS;
5666 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5669 bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5671 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5673 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5674 /* make sure we don't swamp the stripe cache if someone else
5675 * is trying to get access
5677 schedule_timeout_uninterruptible(1);
5679 /* Need to check if array will still be degraded after recovery/resync
5680 * Note in case of > 1 drive failures it's possible we're rebuilding
5681 * one drive while leaving another faulty drive in array.
5684 for (i = 0; i < conf->raid_disks; i++) {
5685 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5687 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5692 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5694 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5695 set_bit(STRIPE_HANDLE, &sh->state);
5697 raid5_release_stripe(sh);
5699 return STRIPE_SECTORS;
5702 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5704 /* We may not be able to submit a whole bio at once as there
5705 * may not be enough stripe_heads available.
5706 * We cannot pre-allocate enough stripe_heads as we may need
5707 * more than exist in the cache (if we allow ever large chunks).
5708 * So we do one stripe head at a time and record in
5709 * ->bi_hw_segments how many have been done.
5711 * We *know* that this entire raid_bio is in one chunk, so
5712 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5714 struct stripe_head *sh;
5716 sector_t sector, logical_sector, last_sector;
5721 logical_sector = raid_bio->bi_iter.bi_sector &
5722 ~((sector_t)STRIPE_SECTORS-1);
5723 sector = raid5_compute_sector(conf, logical_sector,
5725 last_sector = bio_end_sector(raid_bio);
5727 for (; logical_sector < last_sector;
5728 logical_sector += STRIPE_SECTORS,
5729 sector += STRIPE_SECTORS,
5732 if (scnt < raid5_bi_processed_stripes(raid_bio))
5733 /* already done this stripe */
5736 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5739 /* failed to get a stripe - must wait */
5740 raid5_set_bi_processed_stripes(raid_bio, scnt);
5741 conf->retry_read_aligned = raid_bio;
5745 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5746 raid5_release_stripe(sh);
5747 raid5_set_bi_processed_stripes(raid_bio, scnt);
5748 conf->retry_read_aligned = raid_bio;
5752 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5754 raid5_release_stripe(sh);
5757 remaining = raid5_dec_bi_active_stripes(raid_bio);
5758 if (remaining == 0) {
5759 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5761 bio_endio(raid_bio);
5763 if (atomic_dec_and_test(&conf->active_aligned_reads))
5764 wake_up(&conf->wait_for_quiescent);
5768 static int handle_active_stripes(struct r5conf *conf, int group,
5769 struct r5worker *worker,
5770 struct list_head *temp_inactive_list)
5772 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5773 int i, batch_size = 0, hash;
5774 bool release_inactive = false;
5776 while (batch_size < MAX_STRIPE_BATCH &&
5777 (sh = __get_priority_stripe(conf, group)) != NULL)
5778 batch[batch_size++] = sh;
5780 if (batch_size == 0) {
5781 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5782 if (!list_empty(temp_inactive_list + i))
5784 if (i == NR_STRIPE_HASH_LOCKS) {
5785 spin_unlock_irq(&conf->device_lock);
5786 r5l_flush_stripe_to_raid(conf->log);
5787 spin_lock_irq(&conf->device_lock);
5790 release_inactive = true;
5792 spin_unlock_irq(&conf->device_lock);
5794 release_inactive_stripe_list(conf, temp_inactive_list,
5795 NR_STRIPE_HASH_LOCKS);
5797 r5l_flush_stripe_to_raid(conf->log);
5798 if (release_inactive) {
5799 spin_lock_irq(&conf->device_lock);
5803 for (i = 0; i < batch_size; i++)
5804 handle_stripe(batch[i]);
5805 r5l_write_stripe_run(conf->log);
5809 spin_lock_irq(&conf->device_lock);
5810 for (i = 0; i < batch_size; i++) {
5811 hash = batch[i]->hash_lock_index;
5812 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5817 static void raid5_do_work(struct work_struct *work)
5819 struct r5worker *worker = container_of(work, struct r5worker, work);
5820 struct r5worker_group *group = worker->group;
5821 struct r5conf *conf = group->conf;
5822 int group_id = group - conf->worker_groups;
5824 struct blk_plug plug;
5826 pr_debug("+++ raid5worker active\n");
5828 blk_start_plug(&plug);
5830 spin_lock_irq(&conf->device_lock);
5832 int batch_size, released;
5834 released = release_stripe_list(conf, worker->temp_inactive_list);
5836 batch_size = handle_active_stripes(conf, group_id, worker,
5837 worker->temp_inactive_list);
5838 worker->working = false;
5839 if (!batch_size && !released)
5841 handled += batch_size;
5843 pr_debug("%d stripes handled\n", handled);
5845 spin_unlock_irq(&conf->device_lock);
5847 async_tx_issue_pending_all();
5848 blk_finish_plug(&plug);
5850 pr_debug("--- raid5worker inactive\n");
5854 * This is our raid5 kernel thread.
5856 * We scan the hash table for stripes which can be handled now.
5857 * During the scan, completed stripes are saved for us by the interrupt
5858 * handler, so that they will not have to wait for our next wakeup.
5860 static void raid5d(struct md_thread *thread)
5862 struct mddev *mddev = thread->mddev;
5863 struct r5conf *conf = mddev->private;
5865 struct blk_plug plug;
5867 pr_debug("+++ raid5d active\n");
5869 md_check_recovery(mddev);
5871 if (!bio_list_empty(&conf->return_bi) &&
5872 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5873 struct bio_list tmp = BIO_EMPTY_LIST;
5874 spin_lock_irq(&conf->device_lock);
5875 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5876 bio_list_merge(&tmp, &conf->return_bi);
5877 bio_list_init(&conf->return_bi);
5879 spin_unlock_irq(&conf->device_lock);
5883 blk_start_plug(&plug);
5885 spin_lock_irq(&conf->device_lock);
5888 int batch_size, released;
5890 released = release_stripe_list(conf, conf->temp_inactive_list);
5892 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5895 !list_empty(&conf->bitmap_list)) {
5896 /* Now is a good time to flush some bitmap updates */
5898 spin_unlock_irq(&conf->device_lock);
5899 bitmap_unplug(mddev->bitmap);
5900 spin_lock_irq(&conf->device_lock);
5901 conf->seq_write = conf->seq_flush;
5902 activate_bit_delay(conf, conf->temp_inactive_list);
5904 raid5_activate_delayed(conf);
5906 while ((bio = remove_bio_from_retry(conf))) {
5908 spin_unlock_irq(&conf->device_lock);
5909 ok = retry_aligned_read(conf, bio);
5910 spin_lock_irq(&conf->device_lock);
5916 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5917 conf->temp_inactive_list);
5918 if (!batch_size && !released)
5920 handled += batch_size;
5922 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5923 spin_unlock_irq(&conf->device_lock);
5924 md_check_recovery(mddev);
5925 spin_lock_irq(&conf->device_lock);
5928 pr_debug("%d stripes handled\n", handled);
5930 spin_unlock_irq(&conf->device_lock);
5931 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5932 mutex_trylock(&conf->cache_size_mutex)) {
5933 grow_one_stripe(conf, __GFP_NOWARN);
5934 /* Set flag even if allocation failed. This helps
5935 * slow down allocation requests when mem is short
5937 set_bit(R5_DID_ALLOC, &conf->cache_state);
5938 mutex_unlock(&conf->cache_size_mutex);
5941 r5l_flush_stripe_to_raid(conf->log);
5943 async_tx_issue_pending_all();
5944 blk_finish_plug(&plug);
5946 pr_debug("--- raid5d inactive\n");
5950 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5952 struct r5conf *conf;
5954 spin_lock(&mddev->lock);
5955 conf = mddev->private;
5957 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5958 spin_unlock(&mddev->lock);
5963 raid5_set_cache_size(struct mddev *mddev, int size)
5965 struct r5conf *conf = mddev->private;
5968 if (size <= 16 || size > 32768)
5971 conf->min_nr_stripes = size;
5972 mutex_lock(&conf->cache_size_mutex);
5973 while (size < conf->max_nr_stripes &&
5974 drop_one_stripe(conf))
5976 mutex_unlock(&conf->cache_size_mutex);
5979 err = md_allow_write(mddev);
5983 mutex_lock(&conf->cache_size_mutex);
5984 while (size > conf->max_nr_stripes)
5985 if (!grow_one_stripe(conf, GFP_KERNEL))
5987 mutex_unlock(&conf->cache_size_mutex);
5991 EXPORT_SYMBOL(raid5_set_cache_size);
5994 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5996 struct r5conf *conf;
6000 if (len >= PAGE_SIZE)
6002 if (kstrtoul(page, 10, &new))
6004 err = mddev_lock(mddev);
6007 conf = mddev->private;
6011 err = raid5_set_cache_size(mddev, new);
6012 mddev_unlock(mddev);
6017 static struct md_sysfs_entry
6018 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6019 raid5_show_stripe_cache_size,
6020 raid5_store_stripe_cache_size);
6023 raid5_show_rmw_level(struct mddev *mddev, char *page)
6025 struct r5conf *conf = mddev->private;
6027 return sprintf(page, "%d\n", conf->rmw_level);
6033 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6035 struct r5conf *conf = mddev->private;
6041 if (len >= PAGE_SIZE)
6044 if (kstrtoul(page, 10, &new))
6047 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6050 if (new != PARITY_DISABLE_RMW &&
6051 new != PARITY_ENABLE_RMW &&
6052 new != PARITY_PREFER_RMW)
6055 conf->rmw_level = new;
6059 static struct md_sysfs_entry
6060 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6061 raid5_show_rmw_level,
6062 raid5_store_rmw_level);
6066 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6068 struct r5conf *conf;
6070 spin_lock(&mddev->lock);
6071 conf = mddev->private;
6073 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6074 spin_unlock(&mddev->lock);
6079 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6081 struct r5conf *conf;
6085 if (len >= PAGE_SIZE)
6087 if (kstrtoul(page, 10, &new))
6090 err = mddev_lock(mddev);
6093 conf = mddev->private;
6096 else if (new > conf->min_nr_stripes)
6099 conf->bypass_threshold = new;
6100 mddev_unlock(mddev);
6104 static struct md_sysfs_entry
6105 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6107 raid5_show_preread_threshold,
6108 raid5_store_preread_threshold);
6111 raid5_show_skip_copy(struct mddev *mddev, char *page)
6113 struct r5conf *conf;
6115 spin_lock(&mddev->lock);
6116 conf = mddev->private;
6118 ret = sprintf(page, "%d\n", conf->skip_copy);
6119 spin_unlock(&mddev->lock);
6124 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6126 struct r5conf *conf;
6130 if (len >= PAGE_SIZE)
6132 if (kstrtoul(page, 10, &new))
6136 err = mddev_lock(mddev);
6139 conf = mddev->private;
6142 else if (new != conf->skip_copy) {
6143 mddev_suspend(mddev);
6144 conf->skip_copy = new;
6146 mddev->queue->backing_dev_info.capabilities |=
6147 BDI_CAP_STABLE_WRITES;
6149 mddev->queue->backing_dev_info.capabilities &=
6150 ~BDI_CAP_STABLE_WRITES;
6151 mddev_resume(mddev);
6153 mddev_unlock(mddev);
6157 static struct md_sysfs_entry
6158 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6159 raid5_show_skip_copy,
6160 raid5_store_skip_copy);
6163 stripe_cache_active_show(struct mddev *mddev, char *page)
6165 struct r5conf *conf = mddev->private;
6167 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6172 static struct md_sysfs_entry
6173 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6176 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6178 struct r5conf *conf;
6180 spin_lock(&mddev->lock);
6181 conf = mddev->private;
6183 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6184 spin_unlock(&mddev->lock);
6188 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6190 int *worker_cnt_per_group,
6191 struct r5worker_group **worker_groups);
6193 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6195 struct r5conf *conf;
6198 struct r5worker_group *new_groups, *old_groups;
6199 int group_cnt, worker_cnt_per_group;
6201 if (len >= PAGE_SIZE)
6203 if (kstrtoul(page, 10, &new))
6206 err = mddev_lock(mddev);
6209 conf = mddev->private;
6212 else if (new != conf->worker_cnt_per_group) {
6213 mddev_suspend(mddev);
6215 old_groups = conf->worker_groups;
6217 flush_workqueue(raid5_wq);
6219 err = alloc_thread_groups(conf, new,
6220 &group_cnt, &worker_cnt_per_group,
6223 spin_lock_irq(&conf->device_lock);
6224 conf->group_cnt = group_cnt;
6225 conf->worker_cnt_per_group = worker_cnt_per_group;
6226 conf->worker_groups = new_groups;
6227 spin_unlock_irq(&conf->device_lock);
6230 kfree(old_groups[0].workers);
6233 mddev_resume(mddev);
6235 mddev_unlock(mddev);
6240 static struct md_sysfs_entry
6241 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6242 raid5_show_group_thread_cnt,
6243 raid5_store_group_thread_cnt);
6245 static struct attribute *raid5_attrs[] = {
6246 &raid5_stripecache_size.attr,
6247 &raid5_stripecache_active.attr,
6248 &raid5_preread_bypass_threshold.attr,
6249 &raid5_group_thread_cnt.attr,
6250 &raid5_skip_copy.attr,
6251 &raid5_rmw_level.attr,
6254 static struct attribute_group raid5_attrs_group = {
6256 .attrs = raid5_attrs,
6259 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6261 int *worker_cnt_per_group,
6262 struct r5worker_group **worker_groups)
6266 struct r5worker *workers;
6268 *worker_cnt_per_group = cnt;
6271 *worker_groups = NULL;
6274 *group_cnt = num_possible_nodes();
6275 size = sizeof(struct r5worker) * cnt;
6276 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6277 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6278 *group_cnt, GFP_NOIO);
6279 if (!*worker_groups || !workers) {
6281 kfree(*worker_groups);
6285 for (i = 0; i < *group_cnt; i++) {
6286 struct r5worker_group *group;
6288 group = &(*worker_groups)[i];
6289 INIT_LIST_HEAD(&group->handle_list);
6291 group->workers = workers + i * cnt;
6293 for (j = 0; j < cnt; j++) {
6294 struct r5worker *worker = group->workers + j;
6295 worker->group = group;
6296 INIT_WORK(&worker->work, raid5_do_work);
6298 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6299 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6306 static void free_thread_groups(struct r5conf *conf)
6308 if (conf->worker_groups)
6309 kfree(conf->worker_groups[0].workers);
6310 kfree(conf->worker_groups);
6311 conf->worker_groups = NULL;
6315 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6317 struct r5conf *conf = mddev->private;
6320 sectors = mddev->dev_sectors;
6322 /* size is defined by the smallest of previous and new size */
6323 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6325 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6326 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6327 return sectors * (raid_disks - conf->max_degraded);
6330 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6332 safe_put_page(percpu->spare_page);
6333 if (percpu->scribble)
6334 flex_array_free(percpu->scribble);
6335 percpu->spare_page = NULL;
6336 percpu->scribble = NULL;
6339 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6341 if (conf->level == 6 && !percpu->spare_page)
6342 percpu->spare_page = alloc_page(GFP_KERNEL);
6343 if (!percpu->scribble)
6344 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6345 conf->previous_raid_disks),
6346 max(conf->chunk_sectors,
6347 conf->prev_chunk_sectors)
6351 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6352 free_scratch_buffer(conf, percpu);
6359 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6361 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6363 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6367 static void raid5_free_percpu(struct r5conf *conf)
6372 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6373 free_percpu(conf->percpu);
6376 static void free_conf(struct r5conf *conf)
6379 r5l_exit_log(conf->log);
6380 if (conf->shrinker.nr_deferred)
6381 unregister_shrinker(&conf->shrinker);
6383 free_thread_groups(conf);
6384 shrink_stripes(conf);
6385 raid5_free_percpu(conf);
6387 kfree(conf->stripe_hashtbl);
6391 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6393 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6394 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6396 if (alloc_scratch_buffer(conf, percpu)) {
6397 pr_err("%s: failed memory allocation for cpu%u\n",
6404 static int raid5_alloc_percpu(struct r5conf *conf)
6408 conf->percpu = alloc_percpu(struct raid5_percpu);
6412 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6414 conf->scribble_disks = max(conf->raid_disks,
6415 conf->previous_raid_disks);
6416 conf->scribble_sectors = max(conf->chunk_sectors,
6417 conf->prev_chunk_sectors);
6422 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6423 struct shrink_control *sc)
6425 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6426 unsigned long ret = SHRINK_STOP;
6428 if (mutex_trylock(&conf->cache_size_mutex)) {
6430 while (ret < sc->nr_to_scan &&
6431 conf->max_nr_stripes > conf->min_nr_stripes) {
6432 if (drop_one_stripe(conf) == 0) {
6438 mutex_unlock(&conf->cache_size_mutex);
6443 static unsigned long raid5_cache_count(struct shrinker *shrink,
6444 struct shrink_control *sc)
6446 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6448 if (conf->max_nr_stripes < conf->min_nr_stripes)
6449 /* unlikely, but not impossible */
6451 return conf->max_nr_stripes - conf->min_nr_stripes;
6454 static struct r5conf *setup_conf(struct mddev *mddev)
6456 struct r5conf *conf;
6457 int raid_disk, memory, max_disks;
6458 struct md_rdev *rdev;
6459 struct disk_info *disk;
6462 int group_cnt, worker_cnt_per_group;
6463 struct r5worker_group *new_group;
6465 if (mddev->new_level != 5
6466 && mddev->new_level != 4
6467 && mddev->new_level != 6) {
6468 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6469 mdname(mddev), mddev->new_level);
6470 return ERR_PTR(-EIO);
6472 if ((mddev->new_level == 5
6473 && !algorithm_valid_raid5(mddev->new_layout)) ||
6474 (mddev->new_level == 6
6475 && !algorithm_valid_raid6(mddev->new_layout))) {
6476 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6477 mdname(mddev), mddev->new_layout);
6478 return ERR_PTR(-EIO);
6480 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6481 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6482 mdname(mddev), mddev->raid_disks);
6483 return ERR_PTR(-EINVAL);
6486 if (!mddev->new_chunk_sectors ||
6487 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6488 !is_power_of_2(mddev->new_chunk_sectors)) {
6489 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6490 mdname(mddev), mddev->new_chunk_sectors << 9);
6491 return ERR_PTR(-EINVAL);
6494 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6497 /* Don't enable multi-threading by default*/
6498 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6500 conf->group_cnt = group_cnt;
6501 conf->worker_cnt_per_group = worker_cnt_per_group;
6502 conf->worker_groups = new_group;
6505 spin_lock_init(&conf->device_lock);
6506 seqcount_init(&conf->gen_lock);
6507 mutex_init(&conf->cache_size_mutex);
6508 init_waitqueue_head(&conf->wait_for_quiescent);
6509 init_waitqueue_head(&conf->wait_for_stripe);
6510 init_waitqueue_head(&conf->wait_for_overlap);
6511 INIT_LIST_HEAD(&conf->handle_list);
6512 INIT_LIST_HEAD(&conf->hold_list);
6513 INIT_LIST_HEAD(&conf->delayed_list);
6514 INIT_LIST_HEAD(&conf->bitmap_list);
6515 bio_list_init(&conf->return_bi);
6516 init_llist_head(&conf->released_stripes);
6517 atomic_set(&conf->active_stripes, 0);
6518 atomic_set(&conf->preread_active_stripes, 0);
6519 atomic_set(&conf->active_aligned_reads, 0);
6520 conf->bypass_threshold = BYPASS_THRESHOLD;
6521 conf->recovery_disabled = mddev->recovery_disabled - 1;
6523 conf->raid_disks = mddev->raid_disks;
6524 if (mddev->reshape_position == MaxSector)
6525 conf->previous_raid_disks = mddev->raid_disks;
6527 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6528 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6530 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6535 conf->mddev = mddev;
6537 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6540 /* We init hash_locks[0] separately to that it can be used
6541 * as the reference lock in the spin_lock_nest_lock() call
6542 * in lock_all_device_hash_locks_irq in order to convince
6543 * lockdep that we know what we are doing.
6545 spin_lock_init(conf->hash_locks);
6546 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6547 spin_lock_init(conf->hash_locks + i);
6549 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6550 INIT_LIST_HEAD(conf->inactive_list + i);
6552 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6553 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6555 conf->level = mddev->new_level;
6556 conf->chunk_sectors = mddev->new_chunk_sectors;
6557 if (raid5_alloc_percpu(conf) != 0)
6560 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6562 rdev_for_each(rdev, mddev) {
6563 raid_disk = rdev->raid_disk;
6564 if (raid_disk >= max_disks
6565 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6567 disk = conf->disks + raid_disk;
6569 if (test_bit(Replacement, &rdev->flags)) {
6570 if (disk->replacement)
6572 disk->replacement = rdev;
6579 if (test_bit(In_sync, &rdev->flags)) {
6580 char b[BDEVNAME_SIZE];
6581 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6583 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6584 } else if (rdev->saved_raid_disk != raid_disk)
6585 /* Cannot rely on bitmap to complete recovery */
6589 conf->level = mddev->new_level;
6590 if (conf->level == 6) {
6591 conf->max_degraded = 2;
6592 if (raid6_call.xor_syndrome)
6593 conf->rmw_level = PARITY_ENABLE_RMW;
6595 conf->rmw_level = PARITY_DISABLE_RMW;
6597 conf->max_degraded = 1;
6598 conf->rmw_level = PARITY_ENABLE_RMW;
6600 conf->algorithm = mddev->new_layout;
6601 conf->reshape_progress = mddev->reshape_position;
6602 if (conf->reshape_progress != MaxSector) {
6603 conf->prev_chunk_sectors = mddev->chunk_sectors;
6604 conf->prev_algo = mddev->layout;
6606 conf->prev_chunk_sectors = conf->chunk_sectors;
6607 conf->prev_algo = conf->algorithm;
6610 conf->min_nr_stripes = NR_STRIPES;
6611 if (mddev->reshape_position != MaxSector) {
6612 int stripes = max_t(int,
6613 ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
6614 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
6615 conf->min_nr_stripes = max(NR_STRIPES, stripes);
6616 if (conf->min_nr_stripes != NR_STRIPES)
6618 "md/raid:%s: force stripe size %d for reshape\n",
6619 mdname(mddev), conf->min_nr_stripes);
6621 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6622 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6623 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6624 if (grow_stripes(conf, conf->min_nr_stripes)) {
6626 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6627 mdname(mddev), memory);
6630 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6631 mdname(mddev), memory);
6633 * Losing a stripe head costs more than the time to refill it,
6634 * it reduces the queue depth and so can hurt throughput.
6635 * So set it rather large, scaled by number of devices.
6637 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6638 conf->shrinker.scan_objects = raid5_cache_scan;
6639 conf->shrinker.count_objects = raid5_cache_count;
6640 conf->shrinker.batch = 128;
6641 conf->shrinker.flags = 0;
6642 if (register_shrinker(&conf->shrinker)) {
6644 "md/raid:%s: couldn't register shrinker.\n",
6649 sprintf(pers_name, "raid%d", mddev->new_level);
6650 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6651 if (!conf->thread) {
6653 "md/raid:%s: couldn't allocate thread.\n",
6663 return ERR_PTR(-EIO);
6665 return ERR_PTR(-ENOMEM);
6668 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6671 case ALGORITHM_PARITY_0:
6672 if (raid_disk < max_degraded)
6675 case ALGORITHM_PARITY_N:
6676 if (raid_disk >= raid_disks - max_degraded)
6679 case ALGORITHM_PARITY_0_6:
6680 if (raid_disk == 0 ||
6681 raid_disk == raid_disks - 1)
6684 case ALGORITHM_LEFT_ASYMMETRIC_6:
6685 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6686 case ALGORITHM_LEFT_SYMMETRIC_6:
6687 case ALGORITHM_RIGHT_SYMMETRIC_6:
6688 if (raid_disk == raid_disks - 1)
6694 static int raid5_run(struct mddev *mddev)
6696 struct r5conf *conf;
6697 int working_disks = 0;
6698 int dirty_parity_disks = 0;
6699 struct md_rdev *rdev;
6700 struct md_rdev *journal_dev = NULL;
6701 sector_t reshape_offset = 0;
6703 long long min_offset_diff = 0;
6706 if (mddev->recovery_cp != MaxSector)
6707 printk(KERN_NOTICE "md/raid:%s: not clean"
6708 " -- starting background reconstruction\n",
6711 rdev_for_each(rdev, mddev) {
6714 if (test_bit(Journal, &rdev->flags)) {
6718 if (rdev->raid_disk < 0)
6720 diff = (rdev->new_data_offset - rdev->data_offset);
6722 min_offset_diff = diff;
6724 } else if (mddev->reshape_backwards &&
6725 diff < min_offset_diff)
6726 min_offset_diff = diff;
6727 else if (!mddev->reshape_backwards &&
6728 diff > min_offset_diff)
6729 min_offset_diff = diff;
6732 if (mddev->reshape_position != MaxSector) {
6733 /* Check that we can continue the reshape.
6734 * Difficulties arise if the stripe we would write to
6735 * next is at or after the stripe we would read from next.
6736 * For a reshape that changes the number of devices, this
6737 * is only possible for a very short time, and mdadm makes
6738 * sure that time appears to have past before assembling
6739 * the array. So we fail if that time hasn't passed.
6740 * For a reshape that keeps the number of devices the same
6741 * mdadm must be monitoring the reshape can keeping the
6742 * critical areas read-only and backed up. It will start
6743 * the array in read-only mode, so we check for that.
6745 sector_t here_new, here_old;
6747 int max_degraded = (mddev->level == 6 ? 2 : 1);
6752 printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6757 if (mddev->new_level != mddev->level) {
6758 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6759 "required - aborting.\n",
6763 old_disks = mddev->raid_disks - mddev->delta_disks;
6764 /* reshape_position must be on a new-stripe boundary, and one
6765 * further up in new geometry must map after here in old
6767 * If the chunk sizes are different, then as we perform reshape
6768 * in units of the largest of the two, reshape_position needs
6769 * be a multiple of the largest chunk size times new data disks.
6771 here_new = mddev->reshape_position;
6772 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6773 new_data_disks = mddev->raid_disks - max_degraded;
6774 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6775 printk(KERN_ERR "md/raid:%s: reshape_position not "
6776 "on a stripe boundary\n", mdname(mddev));
6779 reshape_offset = here_new * chunk_sectors;
6780 /* here_new is the stripe we will write to */
6781 here_old = mddev->reshape_position;
6782 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6783 /* here_old is the first stripe that we might need to read
6785 if (mddev->delta_disks == 0) {
6786 /* We cannot be sure it is safe to start an in-place
6787 * reshape. It is only safe if user-space is monitoring
6788 * and taking constant backups.
6789 * mdadm always starts a situation like this in
6790 * readonly mode so it can take control before
6791 * allowing any writes. So just check for that.
6793 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6794 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6795 /* not really in-place - so OK */;
6796 else if (mddev->ro == 0) {
6797 printk(KERN_ERR "md/raid:%s: in-place reshape "
6798 "must be started in read-only mode "
6803 } else if (mddev->reshape_backwards
6804 ? (here_new * chunk_sectors + min_offset_diff <=
6805 here_old * chunk_sectors)
6806 : (here_new * chunk_sectors >=
6807 here_old * chunk_sectors + (-min_offset_diff))) {
6808 /* Reading from the same stripe as writing to - bad */
6809 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6810 "auto-recovery - aborting.\n",
6814 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6816 /* OK, we should be able to continue; */
6818 BUG_ON(mddev->level != mddev->new_level);
6819 BUG_ON(mddev->layout != mddev->new_layout);
6820 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6821 BUG_ON(mddev->delta_disks != 0);
6824 if (mddev->private == NULL)
6825 conf = setup_conf(mddev);
6827 conf = mddev->private;
6830 return PTR_ERR(conf);
6832 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
6834 pr_err("md/raid:%s: journal disk is missing, force array readonly\n",
6837 set_disk_ro(mddev->gendisk, 1);
6838 } else if (mddev->recovery_cp == MaxSector)
6839 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
6842 conf->min_offset_diff = min_offset_diff;
6843 mddev->thread = conf->thread;
6844 conf->thread = NULL;
6845 mddev->private = conf;
6847 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6849 rdev = conf->disks[i].rdev;
6850 if (!rdev && conf->disks[i].replacement) {
6851 /* The replacement is all we have yet */
6852 rdev = conf->disks[i].replacement;
6853 conf->disks[i].replacement = NULL;
6854 clear_bit(Replacement, &rdev->flags);
6855 conf->disks[i].rdev = rdev;
6859 if (conf->disks[i].replacement &&
6860 conf->reshape_progress != MaxSector) {
6861 /* replacements and reshape simply do not mix. */
6862 printk(KERN_ERR "md: cannot handle concurrent "
6863 "replacement and reshape.\n");
6866 if (test_bit(In_sync, &rdev->flags)) {
6870 /* This disc is not fully in-sync. However if it
6871 * just stored parity (beyond the recovery_offset),
6872 * when we don't need to be concerned about the
6873 * array being dirty.
6874 * When reshape goes 'backwards', we never have
6875 * partially completed devices, so we only need
6876 * to worry about reshape going forwards.
6878 /* Hack because v0.91 doesn't store recovery_offset properly. */
6879 if (mddev->major_version == 0 &&
6880 mddev->minor_version > 90)
6881 rdev->recovery_offset = reshape_offset;
6883 if (rdev->recovery_offset < reshape_offset) {
6884 /* We need to check old and new layout */
6885 if (!only_parity(rdev->raid_disk,
6888 conf->max_degraded))
6891 if (!only_parity(rdev->raid_disk,
6893 conf->previous_raid_disks,
6894 conf->max_degraded))
6896 dirty_parity_disks++;
6900 * 0 for a fully functional array, 1 or 2 for a degraded array.
6902 mddev->degraded = calc_degraded(conf);
6904 if (has_failed(conf)) {
6905 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6906 " (%d/%d failed)\n",
6907 mdname(mddev), mddev->degraded, conf->raid_disks);
6911 /* device size must be a multiple of chunk size */
6912 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6913 mddev->resync_max_sectors = mddev->dev_sectors;
6915 if (mddev->degraded > dirty_parity_disks &&
6916 mddev->recovery_cp != MaxSector) {
6917 if (mddev->ok_start_degraded)
6919 "md/raid:%s: starting dirty degraded array"
6920 " - data corruption possible.\n",
6924 "md/raid:%s: cannot start dirty degraded array.\n",
6930 if (mddev->degraded == 0)
6931 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6932 " devices, algorithm %d\n", mdname(mddev), conf->level,
6933 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6936 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6937 " out of %d devices, algorithm %d\n",
6938 mdname(mddev), conf->level,
6939 mddev->raid_disks - mddev->degraded,
6940 mddev->raid_disks, mddev->new_layout);
6942 print_raid5_conf(conf);
6944 if (conf->reshape_progress != MaxSector) {
6945 conf->reshape_safe = conf->reshape_progress;
6946 atomic_set(&conf->reshape_stripes, 0);
6947 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6948 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6949 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6950 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6951 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6955 /* Ok, everything is just fine now */
6956 if (mddev->to_remove == &raid5_attrs_group)
6957 mddev->to_remove = NULL;
6958 else if (mddev->kobj.sd &&
6959 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6961 "raid5: failed to create sysfs attributes for %s\n",
6963 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6967 bool discard_supported = true;
6968 /* read-ahead size must cover two whole stripes, which
6969 * is 2 * (datadisks) * chunksize where 'n' is the
6970 * number of raid devices
6972 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6973 int stripe = data_disks *
6974 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6975 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6976 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6978 chunk_size = mddev->chunk_sectors << 9;
6979 blk_queue_io_min(mddev->queue, chunk_size);
6980 blk_queue_io_opt(mddev->queue, chunk_size *
6981 (conf->raid_disks - conf->max_degraded));
6982 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6984 * We can only discard a whole stripe. It doesn't make sense to
6985 * discard data disk but write parity disk
6987 stripe = stripe * PAGE_SIZE;
6988 /* Round up to power of 2, as discard handling
6989 * currently assumes that */
6990 while ((stripe-1) & stripe)
6991 stripe = (stripe | (stripe-1)) + 1;
6992 mddev->queue->limits.discard_alignment = stripe;
6993 mddev->queue->limits.discard_granularity = stripe;
6996 * We use 16-bit counter of active stripes in bi_phys_segments
6997 * (minus one for over-loaded initialization)
6999 blk_queue_max_hw_sectors(mddev->queue, 0xfffe * STRIPE_SECTORS);
7000 blk_queue_max_discard_sectors(mddev->queue,
7001 0xfffe * STRIPE_SECTORS);
7004 * unaligned part of discard request will be ignored, so can't
7005 * guarantee discard_zeroes_data
7007 mddev->queue->limits.discard_zeroes_data = 0;
7009 blk_queue_max_write_same_sectors(mddev->queue, 0);
7011 rdev_for_each(rdev, mddev) {
7012 disk_stack_limits(mddev->gendisk, rdev->bdev,
7013 rdev->data_offset << 9);
7014 disk_stack_limits(mddev->gendisk, rdev->bdev,
7015 rdev->new_data_offset << 9);
7017 * discard_zeroes_data is required, otherwise data
7018 * could be lost. Consider a scenario: discard a stripe
7019 * (the stripe could be inconsistent if
7020 * discard_zeroes_data is 0); write one disk of the
7021 * stripe (the stripe could be inconsistent again
7022 * depending on which disks are used to calculate
7023 * parity); the disk is broken; The stripe data of this
7026 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7027 !bdev_get_queue(rdev->bdev)->
7028 limits.discard_zeroes_data)
7029 discard_supported = false;
7030 /* Unfortunately, discard_zeroes_data is not currently
7031 * a guarantee - just a hint. So we only allow DISCARD
7032 * if the sysadmin has confirmed that only safe devices
7033 * are in use by setting a module parameter.
7035 if (!devices_handle_discard_safely) {
7036 if (discard_supported) {
7037 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7038 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7040 discard_supported = false;
7044 if (discard_supported &&
7045 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7046 mddev->queue->limits.discard_granularity >= stripe)
7047 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7050 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7053 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7057 char b[BDEVNAME_SIZE];
7059 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7060 mdname(mddev), bdevname(journal_dev->bdev, b));
7061 r5l_init_log(conf, journal_dev);
7066 md_unregister_thread(&mddev->thread);
7067 print_raid5_conf(conf);
7069 mddev->private = NULL;
7070 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7074 static void raid5_free(struct mddev *mddev, void *priv)
7076 struct r5conf *conf = priv;
7079 mddev->to_remove = &raid5_attrs_group;
7082 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7084 struct r5conf *conf = mddev->private;
7087 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7088 conf->chunk_sectors / 2, mddev->layout);
7089 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7091 for (i = 0; i < conf->raid_disks; i++) {
7092 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7093 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7096 seq_printf (seq, "]");
7099 static void print_raid5_conf (struct r5conf *conf)
7102 struct disk_info *tmp;
7104 printk(KERN_DEBUG "RAID conf printout:\n");
7106 printk("(conf==NULL)\n");
7109 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7111 conf->raid_disks - conf->mddev->degraded);
7113 for (i = 0; i < conf->raid_disks; i++) {
7114 char b[BDEVNAME_SIZE];
7115 tmp = conf->disks + i;
7117 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7118 i, !test_bit(Faulty, &tmp->rdev->flags),
7119 bdevname(tmp->rdev->bdev, b));
7123 static int raid5_spare_active(struct mddev *mddev)
7126 struct r5conf *conf = mddev->private;
7127 struct disk_info *tmp;
7129 unsigned long flags;
7131 for (i = 0; i < conf->raid_disks; i++) {
7132 tmp = conf->disks + i;
7133 if (tmp->replacement
7134 && tmp->replacement->recovery_offset == MaxSector
7135 && !test_bit(Faulty, &tmp->replacement->flags)
7136 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7137 /* Replacement has just become active. */
7139 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7142 /* Replaced device not technically faulty,
7143 * but we need to be sure it gets removed
7144 * and never re-added.
7146 set_bit(Faulty, &tmp->rdev->flags);
7147 sysfs_notify_dirent_safe(
7148 tmp->rdev->sysfs_state);
7150 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7151 } else if (tmp->rdev
7152 && tmp->rdev->recovery_offset == MaxSector
7153 && !test_bit(Faulty, &tmp->rdev->flags)
7154 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7156 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7159 spin_lock_irqsave(&conf->device_lock, flags);
7160 mddev->degraded = calc_degraded(conf);
7161 spin_unlock_irqrestore(&conf->device_lock, flags);
7162 print_raid5_conf(conf);
7166 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7168 struct r5conf *conf = mddev->private;
7170 int number = rdev->raid_disk;
7171 struct md_rdev **rdevp;
7172 struct disk_info *p = conf->disks + number;
7174 print_raid5_conf(conf);
7175 if (test_bit(Journal, &rdev->flags) && conf->log) {
7176 struct r5l_log *log;
7178 * we can't wait pending write here, as this is called in
7179 * raid5d, wait will deadlock.
7181 if (atomic_read(&mddev->writes_pending))
7189 if (rdev == p->rdev)
7191 else if (rdev == p->replacement)
7192 rdevp = &p->replacement;
7196 if (number >= conf->raid_disks &&
7197 conf->reshape_progress == MaxSector)
7198 clear_bit(In_sync, &rdev->flags);
7200 if (test_bit(In_sync, &rdev->flags) ||
7201 atomic_read(&rdev->nr_pending)) {
7205 /* Only remove non-faulty devices if recovery
7208 if (!test_bit(Faulty, &rdev->flags) &&
7209 mddev->recovery_disabled != conf->recovery_disabled &&
7210 !has_failed(conf) &&
7211 (!p->replacement || p->replacement == rdev) &&
7212 number < conf->raid_disks) {
7217 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7219 if (atomic_read(&rdev->nr_pending)) {
7220 /* lost the race, try later */
7225 if (p->replacement) {
7226 /* We must have just cleared 'rdev' */
7227 p->rdev = p->replacement;
7228 clear_bit(Replacement, &p->replacement->flags);
7229 smp_mb(); /* Make sure other CPUs may see both as identical
7230 * but will never see neither - if they are careful
7232 p->replacement = NULL;
7233 clear_bit(WantReplacement, &rdev->flags);
7235 /* We might have just removed the Replacement as faulty-
7236 * clear the bit just in case
7238 clear_bit(WantReplacement, &rdev->flags);
7241 print_raid5_conf(conf);
7245 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7247 struct r5conf *conf = mddev->private;
7250 struct disk_info *p;
7252 int last = conf->raid_disks - 1;
7254 if (test_bit(Journal, &rdev->flags)) {
7255 char b[BDEVNAME_SIZE];
7259 rdev->raid_disk = 0;
7261 * The array is in readonly mode if journal is missing, so no
7262 * write requests running. We should be safe
7264 r5l_init_log(conf, rdev);
7265 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7266 mdname(mddev), bdevname(rdev->bdev, b));
7269 if (mddev->recovery_disabled == conf->recovery_disabled)
7272 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7273 /* no point adding a device */
7276 if (rdev->raid_disk >= 0)
7277 first = last = rdev->raid_disk;
7280 * find the disk ... but prefer rdev->saved_raid_disk
7283 if (rdev->saved_raid_disk >= 0 &&
7284 rdev->saved_raid_disk >= first &&
7285 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7286 first = rdev->saved_raid_disk;
7288 for (disk = first; disk <= last; disk++) {
7289 p = conf->disks + disk;
7290 if (p->rdev == NULL) {
7291 clear_bit(In_sync, &rdev->flags);
7292 rdev->raid_disk = disk;
7294 if (rdev->saved_raid_disk != disk)
7296 rcu_assign_pointer(p->rdev, rdev);
7300 for (disk = first; disk <= last; disk++) {
7301 p = conf->disks + disk;
7302 if (test_bit(WantReplacement, &p->rdev->flags) &&
7303 p->replacement == NULL) {
7304 clear_bit(In_sync, &rdev->flags);
7305 set_bit(Replacement, &rdev->flags);
7306 rdev->raid_disk = disk;
7309 rcu_assign_pointer(p->replacement, rdev);
7314 print_raid5_conf(conf);
7318 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7320 /* no resync is happening, and there is enough space
7321 * on all devices, so we can resize.
7322 * We need to make sure resync covers any new space.
7323 * If the array is shrinking we should possibly wait until
7324 * any io in the removed space completes, but it hardly seems
7328 struct r5conf *conf = mddev->private;
7332 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7333 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7334 if (mddev->external_size &&
7335 mddev->array_sectors > newsize)
7337 if (mddev->bitmap) {
7338 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7342 md_set_array_sectors(mddev, newsize);
7343 set_capacity(mddev->gendisk, mddev->array_sectors);
7344 revalidate_disk(mddev->gendisk);
7345 if (sectors > mddev->dev_sectors &&
7346 mddev->recovery_cp > mddev->dev_sectors) {
7347 mddev->recovery_cp = mddev->dev_sectors;
7348 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7350 mddev->dev_sectors = sectors;
7351 mddev->resync_max_sectors = sectors;
7355 static int check_stripe_cache(struct mddev *mddev)
7357 /* Can only proceed if there are plenty of stripe_heads.
7358 * We need a minimum of one full stripe,, and for sensible progress
7359 * it is best to have about 4 times that.
7360 * If we require 4 times, then the default 256 4K stripe_heads will
7361 * allow for chunk sizes up to 256K, which is probably OK.
7362 * If the chunk size is greater, user-space should request more
7363 * stripe_heads first.
7365 struct r5conf *conf = mddev->private;
7366 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7367 > conf->min_nr_stripes ||
7368 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7369 > conf->min_nr_stripes) {
7370 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7372 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7379 static int check_reshape(struct mddev *mddev)
7381 struct r5conf *conf = mddev->private;
7385 if (mddev->delta_disks == 0 &&
7386 mddev->new_layout == mddev->layout &&
7387 mddev->new_chunk_sectors == mddev->chunk_sectors)
7388 return 0; /* nothing to do */
7389 if (has_failed(conf))
7391 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7392 /* We might be able to shrink, but the devices must
7393 * be made bigger first.
7394 * For raid6, 4 is the minimum size.
7395 * Otherwise 2 is the minimum
7398 if (mddev->level == 6)
7400 if (mddev->raid_disks + mddev->delta_disks < min)
7404 if (!check_stripe_cache(mddev))
7407 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7408 mddev->delta_disks > 0)
7409 if (resize_chunks(conf,
7410 conf->previous_raid_disks
7411 + max(0, mddev->delta_disks),
7412 max(mddev->new_chunk_sectors,
7413 mddev->chunk_sectors)
7416 return resize_stripes(conf, (conf->previous_raid_disks
7417 + mddev->delta_disks));
7420 static int raid5_start_reshape(struct mddev *mddev)
7422 struct r5conf *conf = mddev->private;
7423 struct md_rdev *rdev;
7425 unsigned long flags;
7427 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7430 if (!check_stripe_cache(mddev))
7433 if (has_failed(conf))
7436 rdev_for_each(rdev, mddev) {
7437 if (!test_bit(In_sync, &rdev->flags)
7438 && !test_bit(Faulty, &rdev->flags))
7442 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7443 /* Not enough devices even to make a degraded array
7448 /* Refuse to reduce size of the array. Any reductions in
7449 * array size must be through explicit setting of array_size
7452 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7453 < mddev->array_sectors) {
7454 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7455 "before number of disks\n", mdname(mddev));
7459 atomic_set(&conf->reshape_stripes, 0);
7460 spin_lock_irq(&conf->device_lock);
7461 write_seqcount_begin(&conf->gen_lock);
7462 conf->previous_raid_disks = conf->raid_disks;
7463 conf->raid_disks += mddev->delta_disks;
7464 conf->prev_chunk_sectors = conf->chunk_sectors;
7465 conf->chunk_sectors = mddev->new_chunk_sectors;
7466 conf->prev_algo = conf->algorithm;
7467 conf->algorithm = mddev->new_layout;
7469 /* Code that selects data_offset needs to see the generation update
7470 * if reshape_progress has been set - so a memory barrier needed.
7473 if (mddev->reshape_backwards)
7474 conf->reshape_progress = raid5_size(mddev, 0, 0);
7476 conf->reshape_progress = 0;
7477 conf->reshape_safe = conf->reshape_progress;
7478 write_seqcount_end(&conf->gen_lock);
7479 spin_unlock_irq(&conf->device_lock);
7481 /* Now make sure any requests that proceeded on the assumption
7482 * the reshape wasn't running - like Discard or Read - have
7485 mddev_suspend(mddev);
7486 mddev_resume(mddev);
7488 /* Add some new drives, as many as will fit.
7489 * We know there are enough to make the newly sized array work.
7490 * Don't add devices if we are reducing the number of
7491 * devices in the array. This is because it is not possible
7492 * to correctly record the "partially reconstructed" state of
7493 * such devices during the reshape and confusion could result.
7495 if (mddev->delta_disks >= 0) {
7496 rdev_for_each(rdev, mddev)
7497 if (rdev->raid_disk < 0 &&
7498 !test_bit(Faulty, &rdev->flags)) {
7499 if (raid5_add_disk(mddev, rdev) == 0) {
7501 >= conf->previous_raid_disks)
7502 set_bit(In_sync, &rdev->flags);
7504 rdev->recovery_offset = 0;
7506 if (sysfs_link_rdev(mddev, rdev))
7507 /* Failure here is OK */;
7509 } else if (rdev->raid_disk >= conf->previous_raid_disks
7510 && !test_bit(Faulty, &rdev->flags)) {
7511 /* This is a spare that was manually added */
7512 set_bit(In_sync, &rdev->flags);
7515 /* When a reshape changes the number of devices,
7516 * ->degraded is measured against the larger of the
7517 * pre and post number of devices.
7519 spin_lock_irqsave(&conf->device_lock, flags);
7520 mddev->degraded = calc_degraded(conf);
7521 spin_unlock_irqrestore(&conf->device_lock, flags);
7523 mddev->raid_disks = conf->raid_disks;
7524 mddev->reshape_position = conf->reshape_progress;
7525 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7527 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7528 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7529 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7530 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7531 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7532 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7534 if (!mddev->sync_thread) {
7535 mddev->recovery = 0;
7536 spin_lock_irq(&conf->device_lock);
7537 write_seqcount_begin(&conf->gen_lock);
7538 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7539 mddev->new_chunk_sectors =
7540 conf->chunk_sectors = conf->prev_chunk_sectors;
7541 mddev->new_layout = conf->algorithm = conf->prev_algo;
7542 rdev_for_each(rdev, mddev)
7543 rdev->new_data_offset = rdev->data_offset;
7545 conf->generation --;
7546 conf->reshape_progress = MaxSector;
7547 mddev->reshape_position = MaxSector;
7548 write_seqcount_end(&conf->gen_lock);
7549 spin_unlock_irq(&conf->device_lock);
7552 conf->reshape_checkpoint = jiffies;
7553 md_wakeup_thread(mddev->sync_thread);
7554 md_new_event(mddev);
7558 /* This is called from the reshape thread and should make any
7559 * changes needed in 'conf'
7561 static void end_reshape(struct r5conf *conf)
7564 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7566 spin_lock_irq(&conf->device_lock);
7567 conf->previous_raid_disks = conf->raid_disks;
7568 md_finish_reshape(conf->mddev);
7570 conf->reshape_progress = MaxSector;
7571 conf->mddev->reshape_position = MaxSector;
7572 spin_unlock_irq(&conf->device_lock);
7573 wake_up(&conf->wait_for_overlap);
7575 /* read-ahead size must cover two whole stripes, which is
7576 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7578 if (conf->mddev->queue) {
7579 int data_disks = conf->raid_disks - conf->max_degraded;
7580 int stripe = data_disks * ((conf->chunk_sectors << 9)
7582 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7583 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7588 /* This is called from the raid5d thread with mddev_lock held.
7589 * It makes config changes to the device.
7591 static void raid5_finish_reshape(struct mddev *mddev)
7593 struct r5conf *conf = mddev->private;
7595 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7597 if (mddev->delta_disks > 0) {
7598 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7600 set_capacity(mddev->gendisk, mddev->array_sectors);
7601 revalidate_disk(mddev->gendisk);
7605 spin_lock_irq(&conf->device_lock);
7606 mddev->degraded = calc_degraded(conf);
7607 spin_unlock_irq(&conf->device_lock);
7608 for (d = conf->raid_disks ;
7609 d < conf->raid_disks - mddev->delta_disks;
7611 struct md_rdev *rdev = conf->disks[d].rdev;
7613 clear_bit(In_sync, &rdev->flags);
7614 rdev = conf->disks[d].replacement;
7616 clear_bit(In_sync, &rdev->flags);
7619 mddev->layout = conf->algorithm;
7620 mddev->chunk_sectors = conf->chunk_sectors;
7621 mddev->reshape_position = MaxSector;
7622 mddev->delta_disks = 0;
7623 mddev->reshape_backwards = 0;
7627 static void raid5_quiesce(struct mddev *mddev, int state)
7629 struct r5conf *conf = mddev->private;
7632 case 2: /* resume for a suspend */
7633 wake_up(&conf->wait_for_overlap);
7636 case 1: /* stop all writes */
7637 lock_all_device_hash_locks_irq(conf);
7638 /* '2' tells resync/reshape to pause so that all
7639 * active stripes can drain
7642 wait_event_cmd(conf->wait_for_quiescent,
7643 atomic_read(&conf->active_stripes) == 0 &&
7644 atomic_read(&conf->active_aligned_reads) == 0,
7645 unlock_all_device_hash_locks_irq(conf),
7646 lock_all_device_hash_locks_irq(conf));
7648 unlock_all_device_hash_locks_irq(conf);
7649 /* allow reshape to continue */
7650 wake_up(&conf->wait_for_overlap);
7653 case 0: /* re-enable writes */
7654 lock_all_device_hash_locks_irq(conf);
7656 wake_up(&conf->wait_for_quiescent);
7657 wake_up(&conf->wait_for_overlap);
7658 unlock_all_device_hash_locks_irq(conf);
7661 r5l_quiesce(conf->log, state);
7664 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7666 struct r0conf *raid0_conf = mddev->private;
7669 /* for raid0 takeover only one zone is supported */
7670 if (raid0_conf->nr_strip_zones > 1) {
7671 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7673 return ERR_PTR(-EINVAL);
7676 sectors = raid0_conf->strip_zone[0].zone_end;
7677 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7678 mddev->dev_sectors = sectors;
7679 mddev->new_level = level;
7680 mddev->new_layout = ALGORITHM_PARITY_N;
7681 mddev->new_chunk_sectors = mddev->chunk_sectors;
7682 mddev->raid_disks += 1;
7683 mddev->delta_disks = 1;
7684 /* make sure it will be not marked as dirty */
7685 mddev->recovery_cp = MaxSector;
7687 return setup_conf(mddev);
7690 static void *raid5_takeover_raid1(struct mddev *mddev)
7694 if (mddev->raid_disks != 2 ||
7695 mddev->degraded > 1)
7696 return ERR_PTR(-EINVAL);
7698 /* Should check if there are write-behind devices? */
7700 chunksect = 64*2; /* 64K by default */
7702 /* The array must be an exact multiple of chunksize */
7703 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7706 if ((chunksect<<9) < STRIPE_SIZE)
7707 /* array size does not allow a suitable chunk size */
7708 return ERR_PTR(-EINVAL);
7710 mddev->new_level = 5;
7711 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7712 mddev->new_chunk_sectors = chunksect;
7714 return setup_conf(mddev);
7717 static void *raid5_takeover_raid6(struct mddev *mddev)
7721 switch (mddev->layout) {
7722 case ALGORITHM_LEFT_ASYMMETRIC_6:
7723 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7725 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7726 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7728 case ALGORITHM_LEFT_SYMMETRIC_6:
7729 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7731 case ALGORITHM_RIGHT_SYMMETRIC_6:
7732 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7734 case ALGORITHM_PARITY_0_6:
7735 new_layout = ALGORITHM_PARITY_0;
7737 case ALGORITHM_PARITY_N:
7738 new_layout = ALGORITHM_PARITY_N;
7741 return ERR_PTR(-EINVAL);
7743 mddev->new_level = 5;
7744 mddev->new_layout = new_layout;
7745 mddev->delta_disks = -1;
7746 mddev->raid_disks -= 1;
7747 return setup_conf(mddev);
7750 static int raid5_check_reshape(struct mddev *mddev)
7752 /* For a 2-drive array, the layout and chunk size can be changed
7753 * immediately as not restriping is needed.
7754 * For larger arrays we record the new value - after validation
7755 * to be used by a reshape pass.
7757 struct r5conf *conf = mddev->private;
7758 int new_chunk = mddev->new_chunk_sectors;
7760 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7762 if (new_chunk > 0) {
7763 if (!is_power_of_2(new_chunk))
7765 if (new_chunk < (PAGE_SIZE>>9))
7767 if (mddev->array_sectors & (new_chunk-1))
7768 /* not factor of array size */
7772 /* They look valid */
7774 if (mddev->raid_disks == 2) {
7775 /* can make the change immediately */
7776 if (mddev->new_layout >= 0) {
7777 conf->algorithm = mddev->new_layout;
7778 mddev->layout = mddev->new_layout;
7780 if (new_chunk > 0) {
7781 conf->chunk_sectors = new_chunk ;
7782 mddev->chunk_sectors = new_chunk;
7784 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7785 md_wakeup_thread(mddev->thread);
7787 return check_reshape(mddev);
7790 static int raid6_check_reshape(struct mddev *mddev)
7792 int new_chunk = mddev->new_chunk_sectors;
7794 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7796 if (new_chunk > 0) {
7797 if (!is_power_of_2(new_chunk))
7799 if (new_chunk < (PAGE_SIZE >> 9))
7801 if (mddev->array_sectors & (new_chunk-1))
7802 /* not factor of array size */
7806 /* They look valid */
7807 return check_reshape(mddev);
7810 static void *raid5_takeover(struct mddev *mddev)
7812 /* raid5 can take over:
7813 * raid0 - if there is only one strip zone - make it a raid4 layout
7814 * raid1 - if there are two drives. We need to know the chunk size
7815 * raid4 - trivial - just use a raid4 layout.
7816 * raid6 - Providing it is a *_6 layout
7818 if (mddev->level == 0)
7819 return raid45_takeover_raid0(mddev, 5);
7820 if (mddev->level == 1)
7821 return raid5_takeover_raid1(mddev);
7822 if (mddev->level == 4) {
7823 mddev->new_layout = ALGORITHM_PARITY_N;
7824 mddev->new_level = 5;
7825 return setup_conf(mddev);
7827 if (mddev->level == 6)
7828 return raid5_takeover_raid6(mddev);
7830 return ERR_PTR(-EINVAL);
7833 static void *raid4_takeover(struct mddev *mddev)
7835 /* raid4 can take over:
7836 * raid0 - if there is only one strip zone
7837 * raid5 - if layout is right
7839 if (mddev->level == 0)
7840 return raid45_takeover_raid0(mddev, 4);
7841 if (mddev->level == 5 &&
7842 mddev->layout == ALGORITHM_PARITY_N) {
7843 mddev->new_layout = 0;
7844 mddev->new_level = 4;
7845 return setup_conf(mddev);
7847 return ERR_PTR(-EINVAL);
7850 static struct md_personality raid5_personality;
7852 static void *raid6_takeover(struct mddev *mddev)
7854 /* Currently can only take over a raid5. We map the
7855 * personality to an equivalent raid6 personality
7856 * with the Q block at the end.
7860 if (mddev->pers != &raid5_personality)
7861 return ERR_PTR(-EINVAL);
7862 if (mddev->degraded > 1)
7863 return ERR_PTR(-EINVAL);
7864 if (mddev->raid_disks > 253)
7865 return ERR_PTR(-EINVAL);
7866 if (mddev->raid_disks < 3)
7867 return ERR_PTR(-EINVAL);
7869 switch (mddev->layout) {
7870 case ALGORITHM_LEFT_ASYMMETRIC:
7871 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7873 case ALGORITHM_RIGHT_ASYMMETRIC:
7874 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7876 case ALGORITHM_LEFT_SYMMETRIC:
7877 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7879 case ALGORITHM_RIGHT_SYMMETRIC:
7880 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7882 case ALGORITHM_PARITY_0:
7883 new_layout = ALGORITHM_PARITY_0_6;
7885 case ALGORITHM_PARITY_N:
7886 new_layout = ALGORITHM_PARITY_N;
7889 return ERR_PTR(-EINVAL);
7891 mddev->new_level = 6;
7892 mddev->new_layout = new_layout;
7893 mddev->delta_disks = 1;
7894 mddev->raid_disks += 1;
7895 return setup_conf(mddev);
7898 static struct md_personality raid6_personality =
7902 .owner = THIS_MODULE,
7903 .make_request = raid5_make_request,
7906 .status = raid5_status,
7907 .error_handler = raid5_error,
7908 .hot_add_disk = raid5_add_disk,
7909 .hot_remove_disk= raid5_remove_disk,
7910 .spare_active = raid5_spare_active,
7911 .sync_request = raid5_sync_request,
7912 .resize = raid5_resize,
7914 .check_reshape = raid6_check_reshape,
7915 .start_reshape = raid5_start_reshape,
7916 .finish_reshape = raid5_finish_reshape,
7917 .quiesce = raid5_quiesce,
7918 .takeover = raid6_takeover,
7919 .congested = raid5_congested,
7921 static struct md_personality raid5_personality =
7925 .owner = THIS_MODULE,
7926 .make_request = raid5_make_request,
7929 .status = raid5_status,
7930 .error_handler = raid5_error,
7931 .hot_add_disk = raid5_add_disk,
7932 .hot_remove_disk= raid5_remove_disk,
7933 .spare_active = raid5_spare_active,
7934 .sync_request = raid5_sync_request,
7935 .resize = raid5_resize,
7937 .check_reshape = raid5_check_reshape,
7938 .start_reshape = raid5_start_reshape,
7939 .finish_reshape = raid5_finish_reshape,
7940 .quiesce = raid5_quiesce,
7941 .takeover = raid5_takeover,
7942 .congested = raid5_congested,
7945 static struct md_personality raid4_personality =
7949 .owner = THIS_MODULE,
7950 .make_request = raid5_make_request,
7953 .status = raid5_status,
7954 .error_handler = raid5_error,
7955 .hot_add_disk = raid5_add_disk,
7956 .hot_remove_disk= raid5_remove_disk,
7957 .spare_active = raid5_spare_active,
7958 .sync_request = raid5_sync_request,
7959 .resize = raid5_resize,
7961 .check_reshape = raid5_check_reshape,
7962 .start_reshape = raid5_start_reshape,
7963 .finish_reshape = raid5_finish_reshape,
7964 .quiesce = raid5_quiesce,
7965 .takeover = raid4_takeover,
7966 .congested = raid5_congested,
7969 static int __init raid5_init(void)
7973 raid5_wq = alloc_workqueue("raid5wq",
7974 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7978 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
7980 raid456_cpu_up_prepare,
7983 destroy_workqueue(raid5_wq);
7986 register_md_personality(&raid6_personality);
7987 register_md_personality(&raid5_personality);
7988 register_md_personality(&raid4_personality);
7992 static void raid5_exit(void)
7994 unregister_md_personality(&raid6_personality);
7995 unregister_md_personality(&raid5_personality);
7996 unregister_md_personality(&raid4_personality);
7997 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
7998 destroy_workqueue(raid5_wq);
8001 module_init(raid5_init);
8002 module_exit(raid5_exit);
8003 MODULE_LICENSE("GPL");
8004 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8005 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8006 MODULE_ALIAS("md-raid5");
8007 MODULE_ALIAS("md-raid4");
8008 MODULE_ALIAS("md-level-5");
8009 MODULE_ALIAS("md-level-4");
8010 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8011 MODULE_ALIAS("md-raid6");
8012 MODULE_ALIAS("md-level-6");
8014 /* This used to be two separate modules, they were: */
8015 MODULE_ALIAS("raid5");
8016 MODULE_ALIAS("raid6");