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 <trace/events/block.h>
64 #define cpu_to_group(cpu) cpu_to_node(cpu)
65 #define ANY_GROUP NUMA_NO_NODE
67 static struct workqueue_struct *raid5_wq;
72 #define NR_STRIPES 256
73 #define STRIPE_SIZE PAGE_SIZE
74 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
75 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
76 #define IO_THRESHOLD 1
77 #define BYPASS_THRESHOLD 1
78 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
79 #define HASH_MASK (NR_HASH - 1)
80 #define MAX_STRIPE_BATCH 8
82 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
84 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
85 return &conf->stripe_hashtbl[hash];
88 static inline int stripe_hash_locks_hash(sector_t sect)
90 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
93 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
95 spin_lock_irq(conf->hash_locks + hash);
96 spin_lock(&conf->device_lock);
99 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
101 spin_unlock(&conf->device_lock);
102 spin_unlock_irq(conf->hash_locks + hash);
105 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
109 spin_lock(conf->hash_locks);
110 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
111 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
112 spin_lock(&conf->device_lock);
115 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
118 spin_unlock(&conf->device_lock);
119 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
120 spin_unlock(conf->hash_locks + i - 1);
124 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
125 * order without overlap. There may be several bio's per stripe+device, and
126 * a bio could span several devices.
127 * When walking this list for a particular stripe+device, we must never proceed
128 * beyond a bio that extends past this device, as the next bio might no longer
130 * This function is used to determine the 'next' bio in the list, given the sector
131 * of the current stripe+device
133 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
135 int sectors = bio_sectors(bio);
136 if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
143 * We maintain a biased count of active stripes in the bottom 16 bits of
144 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
146 static inline int raid5_bi_processed_stripes(struct bio *bio)
148 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
149 return (atomic_read(segments) >> 16) & 0xffff;
152 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
154 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
155 return atomic_sub_return(1, segments) & 0xffff;
158 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
160 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
161 atomic_inc(segments);
164 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
167 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
171 old = atomic_read(segments);
172 new = (old & 0xffff) | (cnt << 16);
173 } while (atomic_cmpxchg(segments, old, new) != old);
176 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
178 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
179 atomic_set(segments, cnt);
182 /* Find first data disk in a raid6 stripe */
183 static inline int raid6_d0(struct stripe_head *sh)
186 /* ddf always start from first device */
188 /* md starts just after Q block */
189 if (sh->qd_idx == sh->disks - 1)
192 return sh->qd_idx + 1;
194 static inline int raid6_next_disk(int disk, int raid_disks)
197 return (disk < raid_disks) ? disk : 0;
200 /* When walking through the disks in a raid5, starting at raid6_d0,
201 * We need to map each disk to a 'slot', where the data disks are slot
202 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
203 * is raid_disks-1. This help does that mapping.
205 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
206 int *count, int syndrome_disks)
212 if (idx == sh->pd_idx)
213 return syndrome_disks;
214 if (idx == sh->qd_idx)
215 return syndrome_disks + 1;
221 static void return_io(struct bio *return_bi)
223 struct bio *bi = return_bi;
226 return_bi = bi->bi_next;
228 bi->bi_iter.bi_size = 0;
229 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
236 static void print_raid5_conf (struct r5conf *conf);
238 static int stripe_operations_active(struct stripe_head *sh)
240 return sh->check_state || sh->reconstruct_state ||
241 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
242 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
245 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
247 struct r5conf *conf = sh->raid_conf;
248 struct r5worker_group *group;
250 int i, cpu = sh->cpu;
252 if (!cpu_online(cpu)) {
253 cpu = cpumask_any(cpu_online_mask);
257 if (list_empty(&sh->lru)) {
258 struct r5worker_group *group;
259 group = conf->worker_groups + cpu_to_group(cpu);
260 list_add_tail(&sh->lru, &group->handle_list);
261 group->stripes_cnt++;
265 if (conf->worker_cnt_per_group == 0) {
266 md_wakeup_thread(conf->mddev->thread);
270 group = conf->worker_groups + cpu_to_group(sh->cpu);
272 group->workers[0].working = true;
273 /* at least one worker should run to avoid race */
274 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
276 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
277 /* wakeup more workers */
278 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
279 if (group->workers[i].working == false) {
280 group->workers[i].working = true;
281 queue_work_on(sh->cpu, raid5_wq,
282 &group->workers[i].work);
288 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
289 struct list_head *temp_inactive_list)
291 BUG_ON(!list_empty(&sh->lru));
292 BUG_ON(atomic_read(&conf->active_stripes)==0);
293 if (test_bit(STRIPE_HANDLE, &sh->state)) {
294 if (test_bit(STRIPE_DELAYED, &sh->state) &&
295 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
296 list_add_tail(&sh->lru, &conf->delayed_list);
297 if (atomic_read(&conf->preread_active_stripes)
299 md_wakeup_thread(conf->mddev->thread);
300 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
301 sh->bm_seq - conf->seq_write > 0)
302 list_add_tail(&sh->lru, &conf->bitmap_list);
304 clear_bit(STRIPE_DELAYED, &sh->state);
305 clear_bit(STRIPE_BIT_DELAY, &sh->state);
306 if (conf->worker_cnt_per_group == 0) {
307 list_add_tail(&sh->lru, &conf->handle_list);
309 raid5_wakeup_stripe_thread(sh);
313 md_wakeup_thread(conf->mddev->thread);
315 BUG_ON(stripe_operations_active(sh));
316 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
317 if (atomic_dec_return(&conf->preread_active_stripes)
319 md_wakeup_thread(conf->mddev->thread);
320 atomic_dec(&conf->active_stripes);
321 if (!test_bit(STRIPE_EXPANDING, &sh->state))
322 list_add_tail(&sh->lru, temp_inactive_list);
326 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
327 struct list_head *temp_inactive_list)
329 if (atomic_dec_and_test(&sh->count))
330 do_release_stripe(conf, sh, temp_inactive_list);
334 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
336 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
337 * given time. Adding stripes only takes device lock, while deleting stripes
338 * only takes hash lock.
340 static void release_inactive_stripe_list(struct r5conf *conf,
341 struct list_head *temp_inactive_list,
345 bool do_wakeup = false;
348 if (hash == NR_STRIPE_HASH_LOCKS) {
349 size = NR_STRIPE_HASH_LOCKS;
350 hash = NR_STRIPE_HASH_LOCKS - 1;
354 struct list_head *list = &temp_inactive_list[size - 1];
357 * We don't hold any lock here yet, get_active_stripe() might
358 * remove stripes from the list
360 if (!list_empty_careful(list)) {
361 spin_lock_irqsave(conf->hash_locks + hash, flags);
362 if (list_empty(conf->inactive_list + hash) &&
364 atomic_dec(&conf->empty_inactive_list_nr);
365 list_splice_tail_init(list, conf->inactive_list + hash);
367 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
374 wake_up(&conf->wait_for_stripe);
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 static void 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);
463 /* find an idle stripe, make sure it is unhashed, and return it. */
464 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
466 struct stripe_head *sh = NULL;
467 struct list_head *first;
469 if (list_empty(conf->inactive_list + hash))
471 first = (conf->inactive_list + hash)->next;
472 sh = list_entry(first, struct stripe_head, lru);
473 list_del_init(first);
475 atomic_inc(&conf->active_stripes);
476 BUG_ON(hash != sh->hash_lock_index);
477 if (list_empty(conf->inactive_list + hash))
478 atomic_inc(&conf->empty_inactive_list_nr);
483 static void shrink_buffers(struct stripe_head *sh)
487 int num = sh->raid_conf->pool_size;
489 for (i = 0; i < num ; i++) {
490 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
494 sh->dev[i].page = NULL;
499 static int grow_buffers(struct stripe_head *sh)
502 int num = sh->raid_conf->pool_size;
504 for (i = 0; i < num; i++) {
507 if (!(page = alloc_page(GFP_KERNEL))) {
510 sh->dev[i].page = page;
511 sh->dev[i].orig_page = page;
516 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
517 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
518 struct stripe_head *sh);
520 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
522 struct r5conf *conf = sh->raid_conf;
525 BUG_ON(atomic_read(&sh->count) != 0);
526 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
527 BUG_ON(stripe_operations_active(sh));
529 pr_debug("init_stripe called, stripe %llu\n",
530 (unsigned long long)sh->sector);
534 seq = read_seqcount_begin(&conf->gen_lock);
535 sh->generation = conf->generation - previous;
536 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
538 stripe_set_idx(sector, conf, previous, sh);
542 for (i = sh->disks; i--; ) {
543 struct r5dev *dev = &sh->dev[i];
545 if (dev->toread || dev->read || dev->towrite || dev->written ||
546 test_bit(R5_LOCKED, &dev->flags)) {
547 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
548 (unsigned long long)sh->sector, i, dev->toread,
549 dev->read, dev->towrite, dev->written,
550 test_bit(R5_LOCKED, &dev->flags));
554 raid5_build_block(sh, i, previous);
556 if (read_seqcount_retry(&conf->gen_lock, seq))
558 insert_hash(conf, sh);
559 sh->cpu = smp_processor_id();
562 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
565 struct stripe_head *sh;
567 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
568 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
569 if (sh->sector == sector && sh->generation == generation)
571 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
576 * Need to check if array has failed when deciding whether to:
578 * - remove non-faulty devices
581 * This determination is simple when no reshape is happening.
582 * However if there is a reshape, we need to carefully check
583 * both the before and after sections.
584 * This is because some failed devices may only affect one
585 * of the two sections, and some non-in_sync devices may
586 * be insync in the section most affected by failed devices.
588 static int calc_degraded(struct r5conf *conf)
590 int degraded, degraded2;
595 for (i = 0; i < conf->previous_raid_disks; i++) {
596 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
597 if (rdev && test_bit(Faulty, &rdev->flags))
598 rdev = rcu_dereference(conf->disks[i].replacement);
599 if (!rdev || test_bit(Faulty, &rdev->flags))
601 else if (test_bit(In_sync, &rdev->flags))
604 /* not in-sync or faulty.
605 * If the reshape increases the number of devices,
606 * this is being recovered by the reshape, so
607 * this 'previous' section is not in_sync.
608 * If the number of devices is being reduced however,
609 * the device can only be part of the array if
610 * we are reverting a reshape, so this section will
613 if (conf->raid_disks >= conf->previous_raid_disks)
617 if (conf->raid_disks == conf->previous_raid_disks)
621 for (i = 0; i < conf->raid_disks; i++) {
622 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
623 if (rdev && test_bit(Faulty, &rdev->flags))
624 rdev = rcu_dereference(conf->disks[i].replacement);
625 if (!rdev || test_bit(Faulty, &rdev->flags))
627 else if (test_bit(In_sync, &rdev->flags))
630 /* not in-sync or faulty.
631 * If reshape increases the number of devices, this
632 * section has already been recovered, else it
633 * almost certainly hasn't.
635 if (conf->raid_disks <= conf->previous_raid_disks)
639 if (degraded2 > degraded)
644 static int has_failed(struct r5conf *conf)
648 if (conf->mddev->reshape_position == MaxSector)
649 return conf->mddev->degraded > conf->max_degraded;
651 degraded = calc_degraded(conf);
652 if (degraded > conf->max_degraded)
657 static struct stripe_head *
658 get_active_stripe(struct r5conf *conf, sector_t sector,
659 int previous, int noblock, int noquiesce)
661 struct stripe_head *sh;
662 int hash = stripe_hash_locks_hash(sector);
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_stripe,
670 conf->quiesce == 0 || noquiesce,
671 *(conf->hash_locks + hash));
672 sh = __find_stripe(conf, sector, conf->generation - previous);
674 if (!conf->inactive_blocked)
675 sh = get_free_stripe(conf, hash);
676 if (noblock && sh == NULL)
679 conf->inactive_blocked = 1;
681 conf->wait_for_stripe,
682 !list_empty(conf->inactive_list + hash) &&
683 (atomic_read(&conf->active_stripes)
684 < (conf->max_nr_stripes * 3 / 4)
685 || !conf->inactive_blocked),
686 *(conf->hash_locks + hash));
687 conf->inactive_blocked = 0;
689 init_stripe(sh, sector, previous);
690 atomic_inc(&sh->count);
692 } else if (!atomic_inc_not_zero(&sh->count)) {
693 spin_lock(&conf->device_lock);
694 if (!atomic_read(&sh->count)) {
695 if (!test_bit(STRIPE_HANDLE, &sh->state))
696 atomic_inc(&conf->active_stripes);
697 BUG_ON(list_empty(&sh->lru) &&
698 !test_bit(STRIPE_EXPANDING, &sh->state));
699 list_del_init(&sh->lru);
701 sh->group->stripes_cnt--;
705 atomic_inc(&sh->count);
706 spin_unlock(&conf->device_lock);
708 } while (sh == NULL);
710 spin_unlock_irq(conf->hash_locks + hash);
714 /* Determine if 'data_offset' or 'new_data_offset' should be used
715 * in this stripe_head.
717 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
719 sector_t progress = conf->reshape_progress;
720 /* Need a memory barrier to make sure we see the value
721 * of conf->generation, or ->data_offset that was set before
722 * reshape_progress was updated.
725 if (progress == MaxSector)
727 if (sh->generation == conf->generation - 1)
729 /* We are in a reshape, and this is a new-generation stripe,
730 * so use new_data_offset.
736 raid5_end_read_request(struct bio *bi, int error);
738 raid5_end_write_request(struct bio *bi, int error);
740 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
742 struct r5conf *conf = sh->raid_conf;
743 int i, disks = sh->disks;
747 for (i = disks; i--; ) {
749 int replace_only = 0;
750 struct bio *bi, *rbi;
751 struct md_rdev *rdev, *rrdev = NULL;
752 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
753 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
757 if (test_bit(R5_Discard, &sh->dev[i].flags))
759 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
761 else if (test_and_clear_bit(R5_WantReplace,
762 &sh->dev[i].flags)) {
767 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
770 bi = &sh->dev[i].req;
771 rbi = &sh->dev[i].rreq; /* For writing to replacement */
774 rrdev = rcu_dereference(conf->disks[i].replacement);
775 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
776 rdev = rcu_dereference(conf->disks[i].rdev);
785 /* We raced and saw duplicates */
788 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
793 if (rdev && test_bit(Faulty, &rdev->flags))
796 atomic_inc(&rdev->nr_pending);
797 if (rrdev && test_bit(Faulty, &rrdev->flags))
800 atomic_inc(&rrdev->nr_pending);
803 /* We have already checked bad blocks for reads. Now
804 * need to check for writes. We never accept write errors
805 * on the replacement, so we don't to check rrdev.
807 while ((rw & WRITE) && rdev &&
808 test_bit(WriteErrorSeen, &rdev->flags)) {
811 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
812 &first_bad, &bad_sectors);
817 set_bit(BlockedBadBlocks, &rdev->flags);
818 if (!conf->mddev->external &&
819 conf->mddev->flags) {
820 /* It is very unlikely, but we might
821 * still need to write out the
822 * bad block log - better give it
824 md_check_recovery(conf->mddev);
827 * Because md_wait_for_blocked_rdev
828 * will dec nr_pending, we must
829 * increment it first.
831 atomic_inc(&rdev->nr_pending);
832 md_wait_for_blocked_rdev(rdev, conf->mddev);
834 /* Acknowledged bad block - skip the write */
835 rdev_dec_pending(rdev, conf->mddev);
841 if (s->syncing || s->expanding || s->expanded
843 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
845 set_bit(STRIPE_IO_STARTED, &sh->state);
848 bi->bi_bdev = rdev->bdev;
850 bi->bi_end_io = (rw & WRITE)
851 ? raid5_end_write_request
852 : raid5_end_read_request;
855 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
856 __func__, (unsigned long long)sh->sector,
858 atomic_inc(&sh->count);
859 if (use_new_offset(conf, sh))
860 bi->bi_iter.bi_sector = (sh->sector
861 + rdev->new_data_offset);
863 bi->bi_iter.bi_sector = (sh->sector
864 + rdev->data_offset);
865 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
866 bi->bi_rw |= REQ_NOMERGE;
868 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
869 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
870 sh->dev[i].vec.bv_page = sh->dev[i].page;
872 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
873 bi->bi_io_vec[0].bv_offset = 0;
874 bi->bi_iter.bi_size = STRIPE_SIZE;
876 * If this is discard request, set bi_vcnt 0. We don't
877 * want to confuse SCSI because SCSI will replace payload
879 if (rw & REQ_DISCARD)
882 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
884 if (conf->mddev->gendisk)
885 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
886 bi, disk_devt(conf->mddev->gendisk),
888 generic_make_request(bi);
891 if (s->syncing || s->expanding || s->expanded
893 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
895 set_bit(STRIPE_IO_STARTED, &sh->state);
898 rbi->bi_bdev = rrdev->bdev;
900 BUG_ON(!(rw & WRITE));
901 rbi->bi_end_io = raid5_end_write_request;
902 rbi->bi_private = sh;
904 pr_debug("%s: for %llu schedule op %ld on "
905 "replacement disc %d\n",
906 __func__, (unsigned long long)sh->sector,
908 atomic_inc(&sh->count);
909 if (use_new_offset(conf, sh))
910 rbi->bi_iter.bi_sector = (sh->sector
911 + rrdev->new_data_offset);
913 rbi->bi_iter.bi_sector = (sh->sector
914 + rrdev->data_offset);
915 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
916 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
917 sh->dev[i].rvec.bv_page = sh->dev[i].page;
919 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
920 rbi->bi_io_vec[0].bv_offset = 0;
921 rbi->bi_iter.bi_size = STRIPE_SIZE;
923 * If this is discard request, set bi_vcnt 0. We don't
924 * want to confuse SCSI because SCSI will replace payload
926 if (rw & REQ_DISCARD)
928 if (conf->mddev->gendisk)
929 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
930 rbi, disk_devt(conf->mddev->gendisk),
932 generic_make_request(rbi);
934 if (!rdev && !rrdev) {
936 set_bit(STRIPE_DEGRADED, &sh->state);
937 pr_debug("skip op %ld on disc %d for sector %llu\n",
938 bi->bi_rw, i, (unsigned long long)sh->sector);
939 clear_bit(R5_LOCKED, &sh->dev[i].flags);
940 set_bit(STRIPE_HANDLE, &sh->state);
945 static struct dma_async_tx_descriptor *
946 async_copy_data(int frombio, struct bio *bio, struct page **page,
947 sector_t sector, struct dma_async_tx_descriptor *tx,
948 struct stripe_head *sh)
951 struct bvec_iter iter;
952 struct page *bio_page;
954 struct async_submit_ctl submit;
955 enum async_tx_flags flags = 0;
957 if (bio->bi_iter.bi_sector >= sector)
958 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
960 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
963 flags |= ASYNC_TX_FENCE;
964 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
966 bio_for_each_segment(bvl, bio, iter) {
967 int len = bvl.bv_len;
971 if (page_offset < 0) {
972 b_offset = -page_offset;
973 page_offset += b_offset;
977 if (len > 0 && page_offset + len > STRIPE_SIZE)
978 clen = STRIPE_SIZE - page_offset;
983 b_offset += bvl.bv_offset;
984 bio_page = bvl.bv_page;
986 if (sh->raid_conf->skip_copy &&
987 b_offset == 0 && page_offset == 0 &&
991 tx = async_memcpy(*page, bio_page, page_offset,
992 b_offset, clen, &submit);
994 tx = async_memcpy(bio_page, *page, b_offset,
995 page_offset, clen, &submit);
997 /* chain the operations */
998 submit.depend_tx = tx;
1000 if (clen < len) /* hit end of page */
1008 static void ops_complete_biofill(void *stripe_head_ref)
1010 struct stripe_head *sh = stripe_head_ref;
1011 struct bio *return_bi = NULL;
1014 pr_debug("%s: stripe %llu\n", __func__,
1015 (unsigned long long)sh->sector);
1017 /* clear completed biofills */
1018 for (i = sh->disks; i--; ) {
1019 struct r5dev *dev = &sh->dev[i];
1021 /* acknowledge completion of a biofill operation */
1022 /* and check if we need to reply to a read request,
1023 * new R5_Wantfill requests are held off until
1024 * !STRIPE_BIOFILL_RUN
1026 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1027 struct bio *rbi, *rbi2;
1032 while (rbi && rbi->bi_iter.bi_sector <
1033 dev->sector + STRIPE_SECTORS) {
1034 rbi2 = r5_next_bio(rbi, dev->sector);
1035 if (!raid5_dec_bi_active_stripes(rbi)) {
1036 rbi->bi_next = return_bi;
1043 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1045 return_io(return_bi);
1047 set_bit(STRIPE_HANDLE, &sh->state);
1051 static void ops_run_biofill(struct stripe_head *sh)
1053 struct dma_async_tx_descriptor *tx = NULL;
1054 struct async_submit_ctl submit;
1057 pr_debug("%s: stripe %llu\n", __func__,
1058 (unsigned long long)sh->sector);
1060 for (i = sh->disks; i--; ) {
1061 struct r5dev *dev = &sh->dev[i];
1062 if (test_bit(R5_Wantfill, &dev->flags)) {
1064 spin_lock_irq(&sh->stripe_lock);
1065 dev->read = rbi = dev->toread;
1067 spin_unlock_irq(&sh->stripe_lock);
1068 while (rbi && rbi->bi_iter.bi_sector <
1069 dev->sector + STRIPE_SECTORS) {
1070 tx = async_copy_data(0, rbi, &dev->page,
1071 dev->sector, tx, sh);
1072 rbi = r5_next_bio(rbi, dev->sector);
1077 atomic_inc(&sh->count);
1078 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1079 async_trigger_callback(&submit);
1082 static void mark_target_uptodate(struct stripe_head *sh, int target)
1089 tgt = &sh->dev[target];
1090 set_bit(R5_UPTODATE, &tgt->flags);
1091 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1092 clear_bit(R5_Wantcompute, &tgt->flags);
1095 static void ops_complete_compute(void *stripe_head_ref)
1097 struct stripe_head *sh = stripe_head_ref;
1099 pr_debug("%s: stripe %llu\n", __func__,
1100 (unsigned long long)sh->sector);
1102 /* mark the computed target(s) as uptodate */
1103 mark_target_uptodate(sh, sh->ops.target);
1104 mark_target_uptodate(sh, sh->ops.target2);
1106 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1107 if (sh->check_state == check_state_compute_run)
1108 sh->check_state = check_state_compute_result;
1109 set_bit(STRIPE_HANDLE, &sh->state);
1113 /* return a pointer to the address conversion region of the scribble buffer */
1114 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1115 struct raid5_percpu *percpu)
1117 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
1120 static struct dma_async_tx_descriptor *
1121 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1123 int disks = sh->disks;
1124 struct page **xor_srcs = percpu->scribble;
1125 int target = sh->ops.target;
1126 struct r5dev *tgt = &sh->dev[target];
1127 struct page *xor_dest = tgt->page;
1129 struct dma_async_tx_descriptor *tx;
1130 struct async_submit_ctl submit;
1133 pr_debug("%s: stripe %llu block: %d\n",
1134 __func__, (unsigned long long)sh->sector, target);
1135 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1137 for (i = disks; i--; )
1139 xor_srcs[count++] = sh->dev[i].page;
1141 atomic_inc(&sh->count);
1143 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1144 ops_complete_compute, sh, to_addr_conv(sh, percpu));
1145 if (unlikely(count == 1))
1146 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1148 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1153 /* set_syndrome_sources - populate source buffers for gen_syndrome
1154 * @srcs - (struct page *) array of size sh->disks
1155 * @sh - stripe_head to parse
1157 * Populates srcs in proper layout order for the stripe and returns the
1158 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1159 * destination buffer is recorded in srcs[count] and the Q destination
1160 * is recorded in srcs[count+1]].
1162 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
1164 int disks = sh->disks;
1165 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1166 int d0_idx = raid6_d0(sh);
1170 for (i = 0; i < disks; i++)
1176 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1178 srcs[slot] = sh->dev[i].page;
1179 i = raid6_next_disk(i, disks);
1180 } while (i != d0_idx);
1182 return syndrome_disks;
1185 static struct dma_async_tx_descriptor *
1186 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1188 int disks = sh->disks;
1189 struct page **blocks = percpu->scribble;
1191 int qd_idx = sh->qd_idx;
1192 struct dma_async_tx_descriptor *tx;
1193 struct async_submit_ctl submit;
1199 if (sh->ops.target < 0)
1200 target = sh->ops.target2;
1201 else if (sh->ops.target2 < 0)
1202 target = sh->ops.target;
1204 /* we should only have one valid target */
1207 pr_debug("%s: stripe %llu block: %d\n",
1208 __func__, (unsigned long long)sh->sector, target);
1210 tgt = &sh->dev[target];
1211 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1214 atomic_inc(&sh->count);
1216 if (target == qd_idx) {
1217 count = set_syndrome_sources(blocks, sh);
1218 blocks[count] = NULL; /* regenerating p is not necessary */
1219 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1220 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1221 ops_complete_compute, sh,
1222 to_addr_conv(sh, percpu));
1223 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1225 /* Compute any data- or p-drive using XOR */
1227 for (i = disks; i-- ; ) {
1228 if (i == target || i == qd_idx)
1230 blocks[count++] = sh->dev[i].page;
1233 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1234 NULL, ops_complete_compute, sh,
1235 to_addr_conv(sh, percpu));
1236 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1242 static struct dma_async_tx_descriptor *
1243 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1245 int i, count, disks = sh->disks;
1246 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1247 int d0_idx = raid6_d0(sh);
1248 int faila = -1, failb = -1;
1249 int target = sh->ops.target;
1250 int target2 = sh->ops.target2;
1251 struct r5dev *tgt = &sh->dev[target];
1252 struct r5dev *tgt2 = &sh->dev[target2];
1253 struct dma_async_tx_descriptor *tx;
1254 struct page **blocks = percpu->scribble;
1255 struct async_submit_ctl submit;
1257 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1258 __func__, (unsigned long long)sh->sector, target, target2);
1259 BUG_ON(target < 0 || target2 < 0);
1260 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1261 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1263 /* we need to open-code set_syndrome_sources to handle the
1264 * slot number conversion for 'faila' and 'failb'
1266 for (i = 0; i < disks ; i++)
1271 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1273 blocks[slot] = sh->dev[i].page;
1279 i = raid6_next_disk(i, disks);
1280 } while (i != d0_idx);
1282 BUG_ON(faila == failb);
1285 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1286 __func__, (unsigned long long)sh->sector, faila, failb);
1288 atomic_inc(&sh->count);
1290 if (failb == syndrome_disks+1) {
1291 /* Q disk is one of the missing disks */
1292 if (faila == syndrome_disks) {
1293 /* Missing P+Q, just recompute */
1294 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1295 ops_complete_compute, sh,
1296 to_addr_conv(sh, percpu));
1297 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1298 STRIPE_SIZE, &submit);
1302 int qd_idx = sh->qd_idx;
1304 /* Missing D+Q: recompute D from P, then recompute Q */
1305 if (target == qd_idx)
1306 data_target = target2;
1308 data_target = target;
1311 for (i = disks; i-- ; ) {
1312 if (i == data_target || i == qd_idx)
1314 blocks[count++] = sh->dev[i].page;
1316 dest = sh->dev[data_target].page;
1317 init_async_submit(&submit,
1318 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1320 to_addr_conv(sh, percpu));
1321 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1324 count = set_syndrome_sources(blocks, sh);
1325 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1326 ops_complete_compute, sh,
1327 to_addr_conv(sh, percpu));
1328 return async_gen_syndrome(blocks, 0, count+2,
1329 STRIPE_SIZE, &submit);
1332 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1333 ops_complete_compute, sh,
1334 to_addr_conv(sh, percpu));
1335 if (failb == syndrome_disks) {
1336 /* We're missing D+P. */
1337 return async_raid6_datap_recov(syndrome_disks+2,
1341 /* We're missing D+D. */
1342 return async_raid6_2data_recov(syndrome_disks+2,
1343 STRIPE_SIZE, faila, failb,
1350 static void ops_complete_prexor(void *stripe_head_ref)
1352 struct stripe_head *sh = stripe_head_ref;
1354 pr_debug("%s: stripe %llu\n", __func__,
1355 (unsigned long long)sh->sector);
1358 static struct dma_async_tx_descriptor *
1359 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1360 struct dma_async_tx_descriptor *tx)
1362 int disks = sh->disks;
1363 struct page **xor_srcs = percpu->scribble;
1364 int count = 0, pd_idx = sh->pd_idx, i;
1365 struct async_submit_ctl submit;
1367 /* existing parity data subtracted */
1368 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1370 pr_debug("%s: stripe %llu\n", __func__,
1371 (unsigned long long)sh->sector);
1373 for (i = disks; i--; ) {
1374 struct r5dev *dev = &sh->dev[i];
1375 /* Only process blocks that are known to be uptodate */
1376 if (test_bit(R5_Wantdrain, &dev->flags))
1377 xor_srcs[count++] = dev->page;
1380 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1381 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1382 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1387 static struct dma_async_tx_descriptor *
1388 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1390 int disks = sh->disks;
1393 pr_debug("%s: stripe %llu\n", __func__,
1394 (unsigned long long)sh->sector);
1396 for (i = disks; i--; ) {
1397 struct r5dev *dev = &sh->dev[i];
1400 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1403 spin_lock_irq(&sh->stripe_lock);
1404 chosen = dev->towrite;
1405 dev->towrite = NULL;
1406 BUG_ON(dev->written);
1407 wbi = dev->written = chosen;
1408 spin_unlock_irq(&sh->stripe_lock);
1409 WARN_ON(dev->page != dev->orig_page);
1411 while (wbi && wbi->bi_iter.bi_sector <
1412 dev->sector + STRIPE_SECTORS) {
1413 if (wbi->bi_rw & REQ_FUA)
1414 set_bit(R5_WantFUA, &dev->flags);
1415 if (wbi->bi_rw & REQ_SYNC)
1416 set_bit(R5_SyncIO, &dev->flags);
1417 if (wbi->bi_rw & REQ_DISCARD)
1418 set_bit(R5_Discard, &dev->flags);
1420 tx = async_copy_data(1, wbi, &dev->page,
1421 dev->sector, tx, sh);
1422 if (dev->page != dev->orig_page) {
1423 set_bit(R5_SkipCopy, &dev->flags);
1424 clear_bit(R5_UPTODATE, &dev->flags);
1425 clear_bit(R5_OVERWRITE, &dev->flags);
1428 wbi = r5_next_bio(wbi, dev->sector);
1436 static void ops_complete_reconstruct(void *stripe_head_ref)
1438 struct stripe_head *sh = stripe_head_ref;
1439 int disks = sh->disks;
1440 int pd_idx = sh->pd_idx;
1441 int qd_idx = sh->qd_idx;
1443 bool fua = false, sync = false, discard = false;
1445 pr_debug("%s: stripe %llu\n", __func__,
1446 (unsigned long long)sh->sector);
1448 for (i = disks; i--; ) {
1449 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1450 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1451 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1454 for (i = disks; i--; ) {
1455 struct r5dev *dev = &sh->dev[i];
1457 if (dev->written || i == pd_idx || i == qd_idx) {
1458 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1459 set_bit(R5_UPTODATE, &dev->flags);
1461 set_bit(R5_WantFUA, &dev->flags);
1463 set_bit(R5_SyncIO, &dev->flags);
1467 if (sh->reconstruct_state == reconstruct_state_drain_run)
1468 sh->reconstruct_state = reconstruct_state_drain_result;
1469 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1470 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1472 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1473 sh->reconstruct_state = reconstruct_state_result;
1476 set_bit(STRIPE_HANDLE, &sh->state);
1481 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1482 struct dma_async_tx_descriptor *tx)
1484 int disks = sh->disks;
1485 struct page **xor_srcs = percpu->scribble;
1486 struct async_submit_ctl submit;
1487 int count = 0, pd_idx = sh->pd_idx, i;
1488 struct page *xor_dest;
1490 unsigned long flags;
1492 pr_debug("%s: stripe %llu\n", __func__,
1493 (unsigned long long)sh->sector);
1495 for (i = 0; i < sh->disks; i++) {
1498 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1501 if (i >= sh->disks) {
1502 atomic_inc(&sh->count);
1503 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1504 ops_complete_reconstruct(sh);
1507 /* check if prexor is active which means only process blocks
1508 * that are part of a read-modify-write (written)
1510 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1512 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1513 for (i = disks; i--; ) {
1514 struct r5dev *dev = &sh->dev[i];
1516 xor_srcs[count++] = dev->page;
1519 xor_dest = sh->dev[pd_idx].page;
1520 for (i = disks; i--; ) {
1521 struct r5dev *dev = &sh->dev[i];
1523 xor_srcs[count++] = dev->page;
1527 /* 1/ if we prexor'd then the dest is reused as a source
1528 * 2/ if we did not prexor then we are redoing the parity
1529 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1530 * for the synchronous xor case
1532 flags = ASYNC_TX_ACK |
1533 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1535 atomic_inc(&sh->count);
1537 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1538 to_addr_conv(sh, percpu));
1539 if (unlikely(count == 1))
1540 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1542 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1546 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1547 struct dma_async_tx_descriptor *tx)
1549 struct async_submit_ctl submit;
1550 struct page **blocks = percpu->scribble;
1553 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1555 for (i = 0; i < sh->disks; i++) {
1556 if (sh->pd_idx == i || sh->qd_idx == i)
1558 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1561 if (i >= sh->disks) {
1562 atomic_inc(&sh->count);
1563 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1564 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1565 ops_complete_reconstruct(sh);
1569 count = set_syndrome_sources(blocks, sh);
1571 atomic_inc(&sh->count);
1573 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1574 sh, to_addr_conv(sh, percpu));
1575 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1578 static void ops_complete_check(void *stripe_head_ref)
1580 struct stripe_head *sh = stripe_head_ref;
1582 pr_debug("%s: stripe %llu\n", __func__,
1583 (unsigned long long)sh->sector);
1585 sh->check_state = check_state_check_result;
1586 set_bit(STRIPE_HANDLE, &sh->state);
1590 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1592 int disks = sh->disks;
1593 int pd_idx = sh->pd_idx;
1594 int qd_idx = sh->qd_idx;
1595 struct page *xor_dest;
1596 struct page **xor_srcs = percpu->scribble;
1597 struct dma_async_tx_descriptor *tx;
1598 struct async_submit_ctl submit;
1602 pr_debug("%s: stripe %llu\n", __func__,
1603 (unsigned long long)sh->sector);
1606 xor_dest = sh->dev[pd_idx].page;
1607 xor_srcs[count++] = xor_dest;
1608 for (i = disks; i--; ) {
1609 if (i == pd_idx || i == qd_idx)
1611 xor_srcs[count++] = sh->dev[i].page;
1614 init_async_submit(&submit, 0, NULL, NULL, NULL,
1615 to_addr_conv(sh, percpu));
1616 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1617 &sh->ops.zero_sum_result, &submit);
1619 atomic_inc(&sh->count);
1620 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1621 tx = async_trigger_callback(&submit);
1624 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1626 struct page **srcs = percpu->scribble;
1627 struct async_submit_ctl submit;
1630 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1631 (unsigned long long)sh->sector, checkp);
1633 count = set_syndrome_sources(srcs, sh);
1637 atomic_inc(&sh->count);
1638 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1639 sh, to_addr_conv(sh, percpu));
1640 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1641 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1644 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1646 int overlap_clear = 0, i, disks = sh->disks;
1647 struct dma_async_tx_descriptor *tx = NULL;
1648 struct r5conf *conf = sh->raid_conf;
1649 int level = conf->level;
1650 struct raid5_percpu *percpu;
1654 percpu = per_cpu_ptr(conf->percpu, cpu);
1655 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1656 ops_run_biofill(sh);
1660 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1662 tx = ops_run_compute5(sh, percpu);
1664 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1665 tx = ops_run_compute6_1(sh, percpu);
1667 tx = ops_run_compute6_2(sh, percpu);
1669 /* terminate the chain if reconstruct is not set to be run */
1670 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1674 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1675 tx = ops_run_prexor(sh, percpu, tx);
1677 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1678 tx = ops_run_biodrain(sh, tx);
1682 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1684 ops_run_reconstruct5(sh, percpu, tx);
1686 ops_run_reconstruct6(sh, percpu, tx);
1689 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1690 if (sh->check_state == check_state_run)
1691 ops_run_check_p(sh, percpu);
1692 else if (sh->check_state == check_state_run_q)
1693 ops_run_check_pq(sh, percpu, 0);
1694 else if (sh->check_state == check_state_run_pq)
1695 ops_run_check_pq(sh, percpu, 1);
1701 for (i = disks; i--; ) {
1702 struct r5dev *dev = &sh->dev[i];
1703 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1704 wake_up(&sh->raid_conf->wait_for_overlap);
1709 static int grow_one_stripe(struct r5conf *conf, int hash)
1711 struct stripe_head *sh;
1712 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1716 sh->raid_conf = conf;
1718 spin_lock_init(&sh->stripe_lock);
1720 if (grow_buffers(sh)) {
1722 kmem_cache_free(conf->slab_cache, sh);
1725 sh->hash_lock_index = hash;
1726 /* we just created an active stripe so... */
1727 atomic_set(&sh->count, 1);
1728 atomic_inc(&conf->active_stripes);
1729 INIT_LIST_HEAD(&sh->lru);
1734 static int grow_stripes(struct r5conf *conf, int num)
1736 struct kmem_cache *sc;
1737 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1740 if (conf->mddev->gendisk)
1741 sprintf(conf->cache_name[0],
1742 "raid%d-%s", conf->level, mdname(conf->mddev));
1744 sprintf(conf->cache_name[0],
1745 "raid%d-%p", conf->level, conf->mddev);
1746 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1748 conf->active_name = 0;
1749 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1750 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1754 conf->slab_cache = sc;
1755 conf->pool_size = devs;
1756 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
1758 if (!grow_one_stripe(conf, hash))
1760 conf->max_nr_stripes++;
1761 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
1767 * scribble_len - return the required size of the scribble region
1768 * @num - total number of disks in the array
1770 * The size must be enough to contain:
1771 * 1/ a struct page pointer for each device in the array +2
1772 * 2/ room to convert each entry in (1) to its corresponding dma
1773 * (dma_map_page()) or page (page_address()) address.
1775 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1776 * calculate over all devices (not just the data blocks), using zeros in place
1777 * of the P and Q blocks.
1779 static size_t scribble_len(int num)
1783 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1788 static int resize_stripes(struct r5conf *conf, int newsize)
1790 /* Make all the stripes able to hold 'newsize' devices.
1791 * New slots in each stripe get 'page' set to a new page.
1793 * This happens in stages:
1794 * 1/ create a new kmem_cache and allocate the required number of
1796 * 2/ gather all the old stripe_heads and transfer the pages across
1797 * to the new stripe_heads. This will have the side effect of
1798 * freezing the array as once all stripe_heads have been collected,
1799 * no IO will be possible. Old stripe heads are freed once their
1800 * pages have been transferred over, and the old kmem_cache is
1801 * freed when all stripes are done.
1802 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1803 * we simple return a failre status - no need to clean anything up.
1804 * 4/ allocate new pages for the new slots in the new stripe_heads.
1805 * If this fails, we don't bother trying the shrink the
1806 * stripe_heads down again, we just leave them as they are.
1807 * As each stripe_head is processed the new one is released into
1810 * Once step2 is started, we cannot afford to wait for a write,
1811 * so we use GFP_NOIO allocations.
1813 struct stripe_head *osh, *nsh;
1814 LIST_HEAD(newstripes);
1815 struct disk_info *ndisks;
1818 struct kmem_cache *sc;
1822 if (newsize <= conf->pool_size)
1823 return 0; /* never bother to shrink */
1825 err = md_allow_write(conf->mddev);
1830 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1831 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1836 for (i = conf->max_nr_stripes; i; i--) {
1837 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1841 nsh->raid_conf = conf;
1842 spin_lock_init(&nsh->stripe_lock);
1844 list_add(&nsh->lru, &newstripes);
1847 /* didn't get enough, give up */
1848 while (!list_empty(&newstripes)) {
1849 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1850 list_del(&nsh->lru);
1851 kmem_cache_free(sc, nsh);
1853 kmem_cache_destroy(sc);
1856 /* Step 2 - Must use GFP_NOIO now.
1857 * OK, we have enough stripes, start collecting inactive
1858 * stripes and copying them over
1862 list_for_each_entry(nsh, &newstripes, lru) {
1863 lock_device_hash_lock(conf, hash);
1864 wait_event_cmd(conf->wait_for_stripe,
1865 !list_empty(conf->inactive_list + hash),
1866 unlock_device_hash_lock(conf, hash),
1867 lock_device_hash_lock(conf, hash));
1868 osh = get_free_stripe(conf, hash);
1869 unlock_device_hash_lock(conf, hash);
1870 atomic_set(&nsh->count, 1);
1871 for(i=0; i<conf->pool_size; i++) {
1872 nsh->dev[i].page = osh->dev[i].page;
1873 nsh->dev[i].orig_page = osh->dev[i].page;
1875 for( ; i<newsize; i++)
1876 nsh->dev[i].page = NULL;
1877 nsh->hash_lock_index = hash;
1878 kmem_cache_free(conf->slab_cache, osh);
1880 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
1881 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
1886 kmem_cache_destroy(conf->slab_cache);
1889 * At this point, we are holding all the stripes so the array
1890 * is completely stalled, so now is a good time to resize
1891 * conf->disks and the scribble region
1893 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1895 for (i=0; i<conf->raid_disks; i++)
1896 ndisks[i] = conf->disks[i];
1898 conf->disks = ndisks;
1903 conf->scribble_len = scribble_len(newsize);
1904 for_each_present_cpu(cpu) {
1905 struct raid5_percpu *percpu;
1908 percpu = per_cpu_ptr(conf->percpu, cpu);
1909 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1912 kfree(percpu->scribble);
1913 percpu->scribble = scribble;
1921 /* Step 4, return new stripes to service */
1922 while(!list_empty(&newstripes)) {
1923 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1924 list_del_init(&nsh->lru);
1926 for (i=conf->raid_disks; i < newsize; i++)
1927 if (nsh->dev[i].page == NULL) {
1928 struct page *p = alloc_page(GFP_NOIO);
1929 nsh->dev[i].page = p;
1930 nsh->dev[i].orig_page = p;
1934 release_stripe(nsh);
1936 /* critical section pass, GFP_NOIO no longer needed */
1938 conf->slab_cache = sc;
1939 conf->active_name = 1-conf->active_name;
1940 conf->pool_size = newsize;
1944 static int drop_one_stripe(struct r5conf *conf, int hash)
1946 struct stripe_head *sh;
1948 spin_lock_irq(conf->hash_locks + hash);
1949 sh = get_free_stripe(conf, hash);
1950 spin_unlock_irq(conf->hash_locks + hash);
1953 BUG_ON(atomic_read(&sh->count));
1955 kmem_cache_free(conf->slab_cache, sh);
1956 atomic_dec(&conf->active_stripes);
1960 static void shrink_stripes(struct r5conf *conf)
1963 for (hash = 0; hash < NR_STRIPE_HASH_LOCKS; hash++)
1964 while (drop_one_stripe(conf, hash))
1967 if (conf->slab_cache)
1968 kmem_cache_destroy(conf->slab_cache);
1969 conf->slab_cache = NULL;
1972 static void raid5_end_read_request(struct bio * bi, int error)
1974 struct stripe_head *sh = bi->bi_private;
1975 struct r5conf *conf = sh->raid_conf;
1976 int disks = sh->disks, i;
1977 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1978 char b[BDEVNAME_SIZE];
1979 struct md_rdev *rdev = NULL;
1982 for (i=0 ; i<disks; i++)
1983 if (bi == &sh->dev[i].req)
1986 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1987 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1993 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1994 /* If replacement finished while this request was outstanding,
1995 * 'replacement' might be NULL already.
1996 * In that case it moved down to 'rdev'.
1997 * rdev is not removed until all requests are finished.
1999 rdev = conf->disks[i].replacement;
2001 rdev = conf->disks[i].rdev;
2003 if (use_new_offset(conf, sh))
2004 s = sh->sector + rdev->new_data_offset;
2006 s = sh->sector + rdev->data_offset;
2008 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2009 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2010 /* Note that this cannot happen on a
2011 * replacement device. We just fail those on
2016 "md/raid:%s: read error corrected"
2017 " (%lu sectors at %llu on %s)\n",
2018 mdname(conf->mddev), STRIPE_SECTORS,
2019 (unsigned long long)s,
2020 bdevname(rdev->bdev, b));
2021 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2022 clear_bit(R5_ReadError, &sh->dev[i].flags);
2023 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2024 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2025 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2027 if (atomic_read(&rdev->read_errors))
2028 atomic_set(&rdev->read_errors, 0);
2030 const char *bdn = bdevname(rdev->bdev, b);
2034 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2035 atomic_inc(&rdev->read_errors);
2036 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2039 "md/raid:%s: read error on replacement device "
2040 "(sector %llu on %s).\n",
2041 mdname(conf->mddev),
2042 (unsigned long long)s,
2044 else if (conf->mddev->degraded >= conf->max_degraded) {
2048 "md/raid:%s: read error not correctable "
2049 "(sector %llu on %s).\n",
2050 mdname(conf->mddev),
2051 (unsigned long long)s,
2053 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2058 "md/raid:%s: read error NOT corrected!! "
2059 "(sector %llu on %s).\n",
2060 mdname(conf->mddev),
2061 (unsigned long long)s,
2063 } else if (atomic_read(&rdev->read_errors)
2064 > conf->max_nr_stripes)
2066 "md/raid:%s: Too many read errors, failing device %s.\n",
2067 mdname(conf->mddev), bdn);
2070 if (set_bad && test_bit(In_sync, &rdev->flags)
2071 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2074 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2075 set_bit(R5_ReadError, &sh->dev[i].flags);
2076 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2078 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2080 clear_bit(R5_ReadError, &sh->dev[i].flags);
2081 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2083 && test_bit(In_sync, &rdev->flags)
2084 && rdev_set_badblocks(
2085 rdev, sh->sector, STRIPE_SECTORS, 0)))
2086 md_error(conf->mddev, rdev);
2089 rdev_dec_pending(rdev, conf->mddev);
2090 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2091 set_bit(STRIPE_HANDLE, &sh->state);
2095 static void raid5_end_write_request(struct bio *bi, int error)
2097 struct stripe_head *sh = bi->bi_private;
2098 struct r5conf *conf = sh->raid_conf;
2099 int disks = sh->disks, i;
2100 struct md_rdev *uninitialized_var(rdev);
2101 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2104 int replacement = 0;
2106 for (i = 0 ; i < disks; i++) {
2107 if (bi == &sh->dev[i].req) {
2108 rdev = conf->disks[i].rdev;
2111 if (bi == &sh->dev[i].rreq) {
2112 rdev = conf->disks[i].replacement;
2116 /* rdev was removed and 'replacement'
2117 * replaced it. rdev is not removed
2118 * until all requests are finished.
2120 rdev = conf->disks[i].rdev;
2124 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2125 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2134 md_error(conf->mddev, rdev);
2135 else if (is_badblock(rdev, sh->sector,
2137 &first_bad, &bad_sectors))
2138 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2141 set_bit(STRIPE_DEGRADED, &sh->state);
2142 set_bit(WriteErrorSeen, &rdev->flags);
2143 set_bit(R5_WriteError, &sh->dev[i].flags);
2144 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2145 set_bit(MD_RECOVERY_NEEDED,
2146 &rdev->mddev->recovery);
2147 } else if (is_badblock(rdev, sh->sector,
2149 &first_bad, &bad_sectors)) {
2150 set_bit(R5_MadeGood, &sh->dev[i].flags);
2151 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2152 /* That was a successful write so make
2153 * sure it looks like we already did
2156 set_bit(R5_ReWrite, &sh->dev[i].flags);
2159 rdev_dec_pending(rdev, conf->mddev);
2161 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2162 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2163 set_bit(STRIPE_HANDLE, &sh->state);
2167 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2169 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2171 struct r5dev *dev = &sh->dev[i];
2173 bio_init(&dev->req);
2174 dev->req.bi_io_vec = &dev->vec;
2175 dev->req.bi_max_vecs = 1;
2176 dev->req.bi_private = sh;
2178 bio_init(&dev->rreq);
2179 dev->rreq.bi_io_vec = &dev->rvec;
2180 dev->rreq.bi_max_vecs = 1;
2181 dev->rreq.bi_private = sh;
2184 dev->sector = compute_blocknr(sh, i, previous);
2187 static void error(struct mddev *mddev, struct md_rdev *rdev)
2189 char b[BDEVNAME_SIZE];
2190 struct r5conf *conf = mddev->private;
2191 unsigned long flags;
2192 pr_debug("raid456: error called\n");
2194 spin_lock_irqsave(&conf->device_lock, flags);
2195 clear_bit(In_sync, &rdev->flags);
2196 mddev->degraded = calc_degraded(conf);
2197 spin_unlock_irqrestore(&conf->device_lock, flags);
2198 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2200 set_bit(Blocked, &rdev->flags);
2201 set_bit(Faulty, &rdev->flags);
2202 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2204 "md/raid:%s: Disk failure on %s, disabling device.\n"
2205 "md/raid:%s: Operation continuing on %d devices.\n",
2207 bdevname(rdev->bdev, b),
2209 conf->raid_disks - mddev->degraded);
2213 * Input: a 'big' sector number,
2214 * Output: index of the data and parity disk, and the sector # in them.
2216 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2217 int previous, int *dd_idx,
2218 struct stripe_head *sh)
2220 sector_t stripe, stripe2;
2221 sector_t chunk_number;
2222 unsigned int chunk_offset;
2225 sector_t new_sector;
2226 int algorithm = previous ? conf->prev_algo
2228 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2229 : conf->chunk_sectors;
2230 int raid_disks = previous ? conf->previous_raid_disks
2232 int data_disks = raid_disks - conf->max_degraded;
2234 /* First compute the information on this sector */
2237 * Compute the chunk number and the sector offset inside the chunk
2239 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2240 chunk_number = r_sector;
2243 * Compute the stripe number
2245 stripe = chunk_number;
2246 *dd_idx = sector_div(stripe, data_disks);
2249 * Select the parity disk based on the user selected algorithm.
2251 pd_idx = qd_idx = -1;
2252 switch(conf->level) {
2254 pd_idx = data_disks;
2257 switch (algorithm) {
2258 case ALGORITHM_LEFT_ASYMMETRIC:
2259 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2260 if (*dd_idx >= pd_idx)
2263 case ALGORITHM_RIGHT_ASYMMETRIC:
2264 pd_idx = sector_div(stripe2, raid_disks);
2265 if (*dd_idx >= pd_idx)
2268 case ALGORITHM_LEFT_SYMMETRIC:
2269 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2270 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2272 case ALGORITHM_RIGHT_SYMMETRIC:
2273 pd_idx = sector_div(stripe2, raid_disks);
2274 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2276 case ALGORITHM_PARITY_0:
2280 case ALGORITHM_PARITY_N:
2281 pd_idx = data_disks;
2289 switch (algorithm) {
2290 case ALGORITHM_LEFT_ASYMMETRIC:
2291 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2292 qd_idx = pd_idx + 1;
2293 if (pd_idx == raid_disks-1) {
2294 (*dd_idx)++; /* Q D D D P */
2296 } else if (*dd_idx >= pd_idx)
2297 (*dd_idx) += 2; /* D D P Q D */
2299 case ALGORITHM_RIGHT_ASYMMETRIC:
2300 pd_idx = sector_div(stripe2, raid_disks);
2301 qd_idx = pd_idx + 1;
2302 if (pd_idx == raid_disks-1) {
2303 (*dd_idx)++; /* Q D D D P */
2305 } else if (*dd_idx >= pd_idx)
2306 (*dd_idx) += 2; /* D D P Q D */
2308 case ALGORITHM_LEFT_SYMMETRIC:
2309 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2310 qd_idx = (pd_idx + 1) % raid_disks;
2311 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2313 case ALGORITHM_RIGHT_SYMMETRIC:
2314 pd_idx = sector_div(stripe2, raid_disks);
2315 qd_idx = (pd_idx + 1) % raid_disks;
2316 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2319 case ALGORITHM_PARITY_0:
2324 case ALGORITHM_PARITY_N:
2325 pd_idx = data_disks;
2326 qd_idx = data_disks + 1;
2329 case ALGORITHM_ROTATING_ZERO_RESTART:
2330 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2331 * of blocks for computing Q is different.
2333 pd_idx = sector_div(stripe2, raid_disks);
2334 qd_idx = pd_idx + 1;
2335 if (pd_idx == raid_disks-1) {
2336 (*dd_idx)++; /* Q D D D P */
2338 } else if (*dd_idx >= pd_idx)
2339 (*dd_idx) += 2; /* D D P Q D */
2343 case ALGORITHM_ROTATING_N_RESTART:
2344 /* Same a left_asymmetric, by first stripe is
2345 * D D D P Q rather than
2349 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2350 qd_idx = pd_idx + 1;
2351 if (pd_idx == raid_disks-1) {
2352 (*dd_idx)++; /* Q D D D P */
2354 } else if (*dd_idx >= pd_idx)
2355 (*dd_idx) += 2; /* D D P Q D */
2359 case ALGORITHM_ROTATING_N_CONTINUE:
2360 /* Same as left_symmetric but Q is before P */
2361 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2362 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2363 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2367 case ALGORITHM_LEFT_ASYMMETRIC_6:
2368 /* RAID5 left_asymmetric, with Q on last device */
2369 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2370 if (*dd_idx >= pd_idx)
2372 qd_idx = raid_disks - 1;
2375 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2376 pd_idx = sector_div(stripe2, raid_disks-1);
2377 if (*dd_idx >= pd_idx)
2379 qd_idx = raid_disks - 1;
2382 case ALGORITHM_LEFT_SYMMETRIC_6:
2383 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2384 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2385 qd_idx = raid_disks - 1;
2388 case ALGORITHM_RIGHT_SYMMETRIC_6:
2389 pd_idx = sector_div(stripe2, raid_disks-1);
2390 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2391 qd_idx = raid_disks - 1;
2394 case ALGORITHM_PARITY_0_6:
2397 qd_idx = raid_disks - 1;
2407 sh->pd_idx = pd_idx;
2408 sh->qd_idx = qd_idx;
2409 sh->ddf_layout = ddf_layout;
2412 * Finally, compute the new sector number
2414 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2419 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2421 struct r5conf *conf = sh->raid_conf;
2422 int raid_disks = sh->disks;
2423 int data_disks = raid_disks - conf->max_degraded;
2424 sector_t new_sector = sh->sector, check;
2425 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2426 : conf->chunk_sectors;
2427 int algorithm = previous ? conf->prev_algo
2431 sector_t chunk_number;
2432 int dummy1, dd_idx = i;
2434 struct stripe_head sh2;
2437 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2438 stripe = new_sector;
2440 if (i == sh->pd_idx)
2442 switch(conf->level) {
2445 switch (algorithm) {
2446 case ALGORITHM_LEFT_ASYMMETRIC:
2447 case ALGORITHM_RIGHT_ASYMMETRIC:
2451 case ALGORITHM_LEFT_SYMMETRIC:
2452 case ALGORITHM_RIGHT_SYMMETRIC:
2455 i -= (sh->pd_idx + 1);
2457 case ALGORITHM_PARITY_0:
2460 case ALGORITHM_PARITY_N:
2467 if (i == sh->qd_idx)
2468 return 0; /* It is the Q disk */
2469 switch (algorithm) {
2470 case ALGORITHM_LEFT_ASYMMETRIC:
2471 case ALGORITHM_RIGHT_ASYMMETRIC:
2472 case ALGORITHM_ROTATING_ZERO_RESTART:
2473 case ALGORITHM_ROTATING_N_RESTART:
2474 if (sh->pd_idx == raid_disks-1)
2475 i--; /* Q D D D P */
2476 else if (i > sh->pd_idx)
2477 i -= 2; /* D D P Q D */
2479 case ALGORITHM_LEFT_SYMMETRIC:
2480 case ALGORITHM_RIGHT_SYMMETRIC:
2481 if (sh->pd_idx == raid_disks-1)
2482 i--; /* Q D D D P */
2487 i -= (sh->pd_idx + 2);
2490 case ALGORITHM_PARITY_0:
2493 case ALGORITHM_PARITY_N:
2495 case ALGORITHM_ROTATING_N_CONTINUE:
2496 /* Like left_symmetric, but P is before Q */
2497 if (sh->pd_idx == 0)
2498 i--; /* P D D D Q */
2503 i -= (sh->pd_idx + 1);
2506 case ALGORITHM_LEFT_ASYMMETRIC_6:
2507 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2511 case ALGORITHM_LEFT_SYMMETRIC_6:
2512 case ALGORITHM_RIGHT_SYMMETRIC_6:
2514 i += data_disks + 1;
2515 i -= (sh->pd_idx + 1);
2517 case ALGORITHM_PARITY_0_6:
2526 chunk_number = stripe * data_disks + i;
2527 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2529 check = raid5_compute_sector(conf, r_sector,
2530 previous, &dummy1, &sh2);
2531 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2532 || sh2.qd_idx != sh->qd_idx) {
2533 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2534 mdname(conf->mddev));
2542 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2543 int rcw, int expand)
2545 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2546 struct r5conf *conf = sh->raid_conf;
2547 int level = conf->level;
2551 for (i = disks; i--; ) {
2552 struct r5dev *dev = &sh->dev[i];
2555 set_bit(R5_LOCKED, &dev->flags);
2556 set_bit(R5_Wantdrain, &dev->flags);
2558 clear_bit(R5_UPTODATE, &dev->flags);
2562 /* if we are not expanding this is a proper write request, and
2563 * there will be bios with new data to be drained into the
2568 /* False alarm, nothing to do */
2570 sh->reconstruct_state = reconstruct_state_drain_run;
2571 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2573 sh->reconstruct_state = reconstruct_state_run;
2575 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2577 if (s->locked + conf->max_degraded == disks)
2578 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2579 atomic_inc(&conf->pending_full_writes);
2582 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2583 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2585 for (i = disks; i--; ) {
2586 struct r5dev *dev = &sh->dev[i];
2591 (test_bit(R5_UPTODATE, &dev->flags) ||
2592 test_bit(R5_Wantcompute, &dev->flags))) {
2593 set_bit(R5_Wantdrain, &dev->flags);
2594 set_bit(R5_LOCKED, &dev->flags);
2595 clear_bit(R5_UPTODATE, &dev->flags);
2600 /* False alarm - nothing to do */
2602 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2603 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2604 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2605 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2608 /* keep the parity disk(s) locked while asynchronous operations
2611 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2612 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2616 int qd_idx = sh->qd_idx;
2617 struct r5dev *dev = &sh->dev[qd_idx];
2619 set_bit(R5_LOCKED, &dev->flags);
2620 clear_bit(R5_UPTODATE, &dev->flags);
2624 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2625 __func__, (unsigned long long)sh->sector,
2626 s->locked, s->ops_request);
2630 * Each stripe/dev can have one or more bion attached.
2631 * toread/towrite point to the first in a chain.
2632 * The bi_next chain must be in order.
2634 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2637 struct r5conf *conf = sh->raid_conf;
2640 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2641 (unsigned long long)bi->bi_iter.bi_sector,
2642 (unsigned long long)sh->sector);
2645 * If several bio share a stripe. The bio bi_phys_segments acts as a
2646 * reference count to avoid race. The reference count should already be
2647 * increased before this function is called (for example, in
2648 * make_request()), so other bio sharing this stripe will not free the
2649 * stripe. If a stripe is owned by one stripe, the stripe lock will
2652 spin_lock_irq(&sh->stripe_lock);
2654 bip = &sh->dev[dd_idx].towrite;
2658 bip = &sh->dev[dd_idx].toread;
2659 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2660 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2662 bip = & (*bip)->bi_next;
2664 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2667 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2671 raid5_inc_bi_active_stripes(bi);
2674 /* check if page is covered */
2675 sector_t sector = sh->dev[dd_idx].sector;
2676 for (bi=sh->dev[dd_idx].towrite;
2677 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2678 bi && bi->bi_iter.bi_sector <= sector;
2679 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2680 if (bio_end_sector(bi) >= sector)
2681 sector = bio_end_sector(bi);
2683 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2684 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2687 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2688 (unsigned long long)(*bip)->bi_iter.bi_sector,
2689 (unsigned long long)sh->sector, dd_idx);
2690 spin_unlock_irq(&sh->stripe_lock);
2692 if (conf->mddev->bitmap && firstwrite) {
2693 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2695 sh->bm_seq = conf->seq_flush+1;
2696 set_bit(STRIPE_BIT_DELAY, &sh->state);
2701 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2702 spin_unlock_irq(&sh->stripe_lock);
2706 static void end_reshape(struct r5conf *conf);
2708 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2709 struct stripe_head *sh)
2711 int sectors_per_chunk =
2712 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2714 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2715 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2717 raid5_compute_sector(conf,
2718 stripe * (disks - conf->max_degraded)
2719 *sectors_per_chunk + chunk_offset,
2725 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2726 struct stripe_head_state *s, int disks,
2727 struct bio **return_bi)
2730 for (i = disks; i--; ) {
2734 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2735 struct md_rdev *rdev;
2737 rdev = rcu_dereference(conf->disks[i].rdev);
2738 if (rdev && test_bit(In_sync, &rdev->flags))
2739 atomic_inc(&rdev->nr_pending);
2744 if (!rdev_set_badblocks(
2748 md_error(conf->mddev, rdev);
2749 rdev_dec_pending(rdev, conf->mddev);
2752 spin_lock_irq(&sh->stripe_lock);
2753 /* fail all writes first */
2754 bi = sh->dev[i].towrite;
2755 sh->dev[i].towrite = NULL;
2756 spin_unlock_irq(&sh->stripe_lock);
2760 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2761 wake_up(&conf->wait_for_overlap);
2763 while (bi && bi->bi_iter.bi_sector <
2764 sh->dev[i].sector + STRIPE_SECTORS) {
2765 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2766 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2767 if (!raid5_dec_bi_active_stripes(bi)) {
2768 md_write_end(conf->mddev);
2769 bi->bi_next = *return_bi;
2775 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2776 STRIPE_SECTORS, 0, 0);
2778 /* and fail all 'written' */
2779 bi = sh->dev[i].written;
2780 sh->dev[i].written = NULL;
2781 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
2782 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
2783 sh->dev[i].page = sh->dev[i].orig_page;
2786 if (bi) bitmap_end = 1;
2787 while (bi && bi->bi_iter.bi_sector <
2788 sh->dev[i].sector + STRIPE_SECTORS) {
2789 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2790 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2791 if (!raid5_dec_bi_active_stripes(bi)) {
2792 md_write_end(conf->mddev);
2793 bi->bi_next = *return_bi;
2799 /* fail any reads if this device is non-operational and
2800 * the data has not reached the cache yet.
2802 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2803 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2804 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2805 spin_lock_irq(&sh->stripe_lock);
2806 bi = sh->dev[i].toread;
2807 sh->dev[i].toread = NULL;
2808 spin_unlock_irq(&sh->stripe_lock);
2809 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2810 wake_up(&conf->wait_for_overlap);
2811 while (bi && bi->bi_iter.bi_sector <
2812 sh->dev[i].sector + STRIPE_SECTORS) {
2813 struct bio *nextbi =
2814 r5_next_bio(bi, sh->dev[i].sector);
2815 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2816 if (!raid5_dec_bi_active_stripes(bi)) {
2817 bi->bi_next = *return_bi;
2824 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2825 STRIPE_SECTORS, 0, 0);
2826 /* If we were in the middle of a write the parity block might
2827 * still be locked - so just clear all R5_LOCKED flags
2829 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2832 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2833 if (atomic_dec_and_test(&conf->pending_full_writes))
2834 md_wakeup_thread(conf->mddev->thread);
2838 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2839 struct stripe_head_state *s)
2844 clear_bit(STRIPE_SYNCING, &sh->state);
2845 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2846 wake_up(&conf->wait_for_overlap);
2849 /* There is nothing more to do for sync/check/repair.
2850 * Don't even need to abort as that is handled elsewhere
2851 * if needed, and not always wanted e.g. if there is a known
2853 * For recover/replace we need to record a bad block on all
2854 * non-sync devices, or abort the recovery
2856 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2857 /* During recovery devices cannot be removed, so
2858 * locking and refcounting of rdevs is not needed
2860 for (i = 0; i < conf->raid_disks; i++) {
2861 struct md_rdev *rdev = conf->disks[i].rdev;
2863 && !test_bit(Faulty, &rdev->flags)
2864 && !test_bit(In_sync, &rdev->flags)
2865 && !rdev_set_badblocks(rdev, sh->sector,
2868 rdev = conf->disks[i].replacement;
2870 && !test_bit(Faulty, &rdev->flags)
2871 && !test_bit(In_sync, &rdev->flags)
2872 && !rdev_set_badblocks(rdev, sh->sector,
2877 conf->recovery_disabled =
2878 conf->mddev->recovery_disabled;
2880 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2883 static int want_replace(struct stripe_head *sh, int disk_idx)
2885 struct md_rdev *rdev;
2887 /* Doing recovery so rcu locking not required */
2888 rdev = sh->raid_conf->disks[disk_idx].replacement;
2890 && !test_bit(Faulty, &rdev->flags)
2891 && !test_bit(In_sync, &rdev->flags)
2892 && (rdev->recovery_offset <= sh->sector
2893 || rdev->mddev->recovery_cp <= sh->sector))
2899 /* fetch_block - checks the given member device to see if its data needs
2900 * to be read or computed to satisfy a request.
2902 * Returns 1 when no more member devices need to be checked, otherwise returns
2903 * 0 to tell the loop in handle_stripe_fill to continue
2905 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2906 int disk_idx, int disks)
2908 struct r5dev *dev = &sh->dev[disk_idx];
2909 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2910 &sh->dev[s->failed_num[1]] };
2912 /* is the data in this block needed, and can we get it? */
2913 if (!test_bit(R5_LOCKED, &dev->flags) &&
2914 !test_bit(R5_UPTODATE, &dev->flags) &&
2916 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2917 s->syncing || s->expanding ||
2918 (s->replacing && want_replace(sh, disk_idx)) ||
2919 (s->failed >= 1 && fdev[0]->toread) ||
2920 (s->failed >= 2 && fdev[1]->toread) ||
2921 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2922 (!test_bit(R5_Insync, &dev->flags) || test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) &&
2923 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2924 (sh->raid_conf->level == 6 && s->failed && s->to_write &&
2925 s->to_write - s->non_overwrite < sh->raid_conf->raid_disks - 2 &&
2926 (!test_bit(R5_Insync, &dev->flags) || test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))))) {
2927 /* we would like to get this block, possibly by computing it,
2928 * otherwise read it if the backing disk is insync
2930 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2931 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2932 if ((s->uptodate == disks - 1) &&
2933 (s->failed && (disk_idx == s->failed_num[0] ||
2934 disk_idx == s->failed_num[1]))) {
2935 /* have disk failed, and we're requested to fetch it;
2938 pr_debug("Computing stripe %llu block %d\n",
2939 (unsigned long long)sh->sector, disk_idx);
2940 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2941 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2942 set_bit(R5_Wantcompute, &dev->flags);
2943 sh->ops.target = disk_idx;
2944 sh->ops.target2 = -1; /* no 2nd target */
2946 /* Careful: from this point on 'uptodate' is in the eye
2947 * of raid_run_ops which services 'compute' operations
2948 * before writes. R5_Wantcompute flags a block that will
2949 * be R5_UPTODATE by the time it is needed for a
2950 * subsequent operation.
2954 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2955 /* Computing 2-failure is *very* expensive; only
2956 * do it if failed >= 2
2959 for (other = disks; other--; ) {
2960 if (other == disk_idx)
2962 if (!test_bit(R5_UPTODATE,
2963 &sh->dev[other].flags))
2967 pr_debug("Computing stripe %llu blocks %d,%d\n",
2968 (unsigned long long)sh->sector,
2970 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2971 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2972 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2973 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2974 sh->ops.target = disk_idx;
2975 sh->ops.target2 = other;
2979 } else if (test_bit(R5_Insync, &dev->flags)) {
2980 set_bit(R5_LOCKED, &dev->flags);
2981 set_bit(R5_Wantread, &dev->flags);
2983 pr_debug("Reading block %d (sync=%d)\n",
2984 disk_idx, s->syncing);
2992 * handle_stripe_fill - read or compute data to satisfy pending requests.
2994 static void handle_stripe_fill(struct stripe_head *sh,
2995 struct stripe_head_state *s,
3000 /* look for blocks to read/compute, skip this if a compute
3001 * is already in flight, or if the stripe contents are in the
3002 * midst of changing due to a write
3004 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3005 !sh->reconstruct_state)
3006 for (i = disks; i--; )
3007 if (fetch_block(sh, s, i, disks))
3009 set_bit(STRIPE_HANDLE, &sh->state);
3013 /* handle_stripe_clean_event
3014 * any written block on an uptodate or failed drive can be returned.
3015 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3016 * never LOCKED, so we don't need to test 'failed' directly.
3018 static void handle_stripe_clean_event(struct r5conf *conf,
3019 struct stripe_head *sh, int disks, struct bio **return_bi)
3023 int discard_pending = 0;
3025 for (i = disks; i--; )
3026 if (sh->dev[i].written) {
3028 if (!test_bit(R5_LOCKED, &dev->flags) &&
3029 (test_bit(R5_UPTODATE, &dev->flags) ||
3030 test_bit(R5_Discard, &dev->flags) ||
3031 test_bit(R5_SkipCopy, &dev->flags))) {
3032 /* We can return any write requests */
3033 struct bio *wbi, *wbi2;
3034 pr_debug("Return write for disc %d\n", i);
3035 if (test_and_clear_bit(R5_Discard, &dev->flags))
3036 clear_bit(R5_UPTODATE, &dev->flags);
3037 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3038 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3039 dev->page = dev->orig_page;
3042 dev->written = NULL;
3043 while (wbi && wbi->bi_iter.bi_sector <
3044 dev->sector + STRIPE_SECTORS) {
3045 wbi2 = r5_next_bio(wbi, dev->sector);
3046 if (!raid5_dec_bi_active_stripes(wbi)) {
3047 md_write_end(conf->mddev);
3048 wbi->bi_next = *return_bi;
3053 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3055 !test_bit(STRIPE_DEGRADED, &sh->state),
3057 } else if (test_bit(R5_Discard, &dev->flags))
3058 discard_pending = 1;
3059 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3060 WARN_ON(dev->page != dev->orig_page);
3062 if (!discard_pending &&
3063 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3064 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3065 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3066 if (sh->qd_idx >= 0) {
3067 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3068 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3070 /* now that discard is done we can proceed with any sync */
3071 clear_bit(STRIPE_DISCARD, &sh->state);
3073 * SCSI discard will change some bio fields and the stripe has
3074 * no updated data, so remove it from hash list and the stripe
3075 * will be reinitialized
3077 spin_lock_irq(&conf->device_lock);
3079 spin_unlock_irq(&conf->device_lock);
3080 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3081 set_bit(STRIPE_HANDLE, &sh->state);
3085 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3086 if (atomic_dec_and_test(&conf->pending_full_writes))
3087 md_wakeup_thread(conf->mddev->thread);
3090 static void handle_stripe_dirtying(struct r5conf *conf,
3091 struct stripe_head *sh,
3092 struct stripe_head_state *s,
3095 int rmw = 0, rcw = 0, i;
3096 sector_t recovery_cp = conf->mddev->recovery_cp;
3098 /* RAID6 requires 'rcw' in current implementation.
3099 * Otherwise, check whether resync is now happening or should start.
3100 * If yes, then the array is dirty (after unclean shutdown or
3101 * initial creation), so parity in some stripes might be inconsistent.
3102 * In this case, we need to always do reconstruct-write, to ensure
3103 * that in case of drive failure or read-error correction, we
3104 * generate correct data from the parity.
3106 if (conf->max_degraded == 2 ||
3107 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
3108 /* Calculate the real rcw later - for now make it
3109 * look like rcw is cheaper
3112 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
3113 conf->max_degraded, (unsigned long long)recovery_cp,
3114 (unsigned long long)sh->sector);
3115 } else for (i = disks; i--; ) {
3116 /* would I have to read this buffer for read_modify_write */
3117 struct r5dev *dev = &sh->dev[i];
3118 if ((dev->towrite || i == sh->pd_idx) &&
3119 !test_bit(R5_LOCKED, &dev->flags) &&
3120 !(test_bit(R5_UPTODATE, &dev->flags) ||
3121 test_bit(R5_Wantcompute, &dev->flags))) {
3122 if (test_bit(R5_Insync, &dev->flags))
3125 rmw += 2*disks; /* cannot read it */
3127 /* Would I have to read this buffer for reconstruct_write */
3128 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
3129 !test_bit(R5_LOCKED, &dev->flags) &&
3130 !(test_bit(R5_UPTODATE, &dev->flags) ||
3131 test_bit(R5_Wantcompute, &dev->flags))) {
3132 if (test_bit(R5_Insync, &dev->flags))
3138 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3139 (unsigned long long)sh->sector, rmw, rcw);
3140 set_bit(STRIPE_HANDLE, &sh->state);
3141 if (rmw < rcw && rmw > 0) {
3142 /* prefer read-modify-write, but need to get some data */
3143 if (conf->mddev->queue)
3144 blk_add_trace_msg(conf->mddev->queue,
3145 "raid5 rmw %llu %d",
3146 (unsigned long long)sh->sector, rmw);
3147 for (i = disks; i--; ) {
3148 struct r5dev *dev = &sh->dev[i];
3149 if ((dev->towrite || i == sh->pd_idx) &&
3150 !test_bit(R5_LOCKED, &dev->flags) &&
3151 !(test_bit(R5_UPTODATE, &dev->flags) ||
3152 test_bit(R5_Wantcompute, &dev->flags)) &&
3153 test_bit(R5_Insync, &dev->flags)) {
3154 if (test_bit(STRIPE_PREREAD_ACTIVE,
3156 pr_debug("Read_old block %d for r-m-w\n",
3158 set_bit(R5_LOCKED, &dev->flags);
3159 set_bit(R5_Wantread, &dev->flags);
3162 set_bit(STRIPE_DELAYED, &sh->state);
3163 set_bit(STRIPE_HANDLE, &sh->state);
3168 if (rcw <= rmw && rcw > 0) {
3169 /* want reconstruct write, but need to get some data */
3172 for (i = disks; i--; ) {
3173 struct r5dev *dev = &sh->dev[i];
3174 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3175 i != sh->pd_idx && i != sh->qd_idx &&
3176 !test_bit(R5_LOCKED, &dev->flags) &&
3177 !(test_bit(R5_UPTODATE, &dev->flags) ||
3178 test_bit(R5_Wantcompute, &dev->flags))) {
3180 if (test_bit(R5_Insync, &dev->flags) &&
3181 test_bit(STRIPE_PREREAD_ACTIVE,
3183 pr_debug("Read_old block "
3184 "%d for Reconstruct\n", i);
3185 set_bit(R5_LOCKED, &dev->flags);
3186 set_bit(R5_Wantread, &dev->flags);
3190 set_bit(STRIPE_DELAYED, &sh->state);
3191 set_bit(STRIPE_HANDLE, &sh->state);
3195 if (rcw && conf->mddev->queue)
3196 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3197 (unsigned long long)sh->sector,
3198 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3200 /* now if nothing is locked, and if we have enough data,
3201 * we can start a write request
3203 /* since handle_stripe can be called at any time we need to handle the
3204 * case where a compute block operation has been submitted and then a
3205 * subsequent call wants to start a write request. raid_run_ops only
3206 * handles the case where compute block and reconstruct are requested
3207 * simultaneously. If this is not the case then new writes need to be
3208 * held off until the compute completes.
3210 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3211 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3212 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3213 schedule_reconstruction(sh, s, rcw == 0, 0);
3216 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3217 struct stripe_head_state *s, int disks)
3219 struct r5dev *dev = NULL;
3221 set_bit(STRIPE_HANDLE, &sh->state);
3223 switch (sh->check_state) {
3224 case check_state_idle:
3225 /* start a new check operation if there are no failures */
3226 if (s->failed == 0) {
3227 BUG_ON(s->uptodate != disks);
3228 sh->check_state = check_state_run;
3229 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3230 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3234 dev = &sh->dev[s->failed_num[0]];
3236 case check_state_compute_result:
3237 sh->check_state = check_state_idle;
3239 dev = &sh->dev[sh->pd_idx];
3241 /* check that a write has not made the stripe insync */
3242 if (test_bit(STRIPE_INSYNC, &sh->state))
3245 /* either failed parity check, or recovery is happening */
3246 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3247 BUG_ON(s->uptodate != disks);
3249 set_bit(R5_LOCKED, &dev->flags);
3251 set_bit(R5_Wantwrite, &dev->flags);
3253 clear_bit(STRIPE_DEGRADED, &sh->state);
3254 set_bit(STRIPE_INSYNC, &sh->state);
3256 case check_state_run:
3257 break; /* we will be called again upon completion */
3258 case check_state_check_result:
3259 sh->check_state = check_state_idle;
3261 /* if a failure occurred during the check operation, leave
3262 * STRIPE_INSYNC not set and let the stripe be handled again
3267 /* handle a successful check operation, if parity is correct
3268 * we are done. Otherwise update the mismatch count and repair
3269 * parity if !MD_RECOVERY_CHECK
3271 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3272 /* parity is correct (on disc,
3273 * not in buffer any more)
3275 set_bit(STRIPE_INSYNC, &sh->state);
3277 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3278 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3279 /* don't try to repair!! */
3280 set_bit(STRIPE_INSYNC, &sh->state);
3282 sh->check_state = check_state_compute_run;
3283 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3284 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3285 set_bit(R5_Wantcompute,
3286 &sh->dev[sh->pd_idx].flags);
3287 sh->ops.target = sh->pd_idx;
3288 sh->ops.target2 = -1;
3293 case check_state_compute_run:
3296 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3297 __func__, sh->check_state,
3298 (unsigned long long) sh->sector);
3304 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3305 struct stripe_head_state *s,
3308 int pd_idx = sh->pd_idx;
3309 int qd_idx = sh->qd_idx;
3312 set_bit(STRIPE_HANDLE, &sh->state);
3314 BUG_ON(s->failed > 2);
3316 /* Want to check and possibly repair P and Q.
3317 * However there could be one 'failed' device, in which
3318 * case we can only check one of them, possibly using the
3319 * other to generate missing data
3322 switch (sh->check_state) {
3323 case check_state_idle:
3324 /* start a new check operation if there are < 2 failures */
3325 if (s->failed == s->q_failed) {
3326 /* The only possible failed device holds Q, so it
3327 * makes sense to check P (If anything else were failed,
3328 * we would have used P to recreate it).
3330 sh->check_state = check_state_run;
3332 if (!s->q_failed && s->failed < 2) {
3333 /* Q is not failed, and we didn't use it to generate
3334 * anything, so it makes sense to check it
3336 if (sh->check_state == check_state_run)
3337 sh->check_state = check_state_run_pq;
3339 sh->check_state = check_state_run_q;
3342 /* discard potentially stale zero_sum_result */
3343 sh->ops.zero_sum_result = 0;
3345 if (sh->check_state == check_state_run) {
3346 /* async_xor_zero_sum destroys the contents of P */
3347 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3350 if (sh->check_state >= check_state_run &&
3351 sh->check_state <= check_state_run_pq) {
3352 /* async_syndrome_zero_sum preserves P and Q, so
3353 * no need to mark them !uptodate here
3355 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3359 /* we have 2-disk failure */
3360 BUG_ON(s->failed != 2);
3362 case check_state_compute_result:
3363 sh->check_state = check_state_idle;
3365 /* check that a write has not made the stripe insync */
3366 if (test_bit(STRIPE_INSYNC, &sh->state))
3369 /* now write out any block on a failed drive,
3370 * or P or Q if they were recomputed
3372 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3373 if (s->failed == 2) {
3374 dev = &sh->dev[s->failed_num[1]];
3376 set_bit(R5_LOCKED, &dev->flags);
3377 set_bit(R5_Wantwrite, &dev->flags);
3379 if (s->failed >= 1) {
3380 dev = &sh->dev[s->failed_num[0]];
3382 set_bit(R5_LOCKED, &dev->flags);
3383 set_bit(R5_Wantwrite, &dev->flags);
3385 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3386 dev = &sh->dev[pd_idx];
3388 set_bit(R5_LOCKED, &dev->flags);
3389 set_bit(R5_Wantwrite, &dev->flags);
3391 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3392 dev = &sh->dev[qd_idx];
3394 set_bit(R5_LOCKED, &dev->flags);
3395 set_bit(R5_Wantwrite, &dev->flags);
3397 clear_bit(STRIPE_DEGRADED, &sh->state);
3399 set_bit(STRIPE_INSYNC, &sh->state);
3401 case check_state_run:
3402 case check_state_run_q:
3403 case check_state_run_pq:
3404 break; /* we will be called again upon completion */
3405 case check_state_check_result:
3406 sh->check_state = check_state_idle;
3408 /* handle a successful check operation, if parity is correct
3409 * we are done. Otherwise update the mismatch count and repair
3410 * parity if !MD_RECOVERY_CHECK
3412 if (sh->ops.zero_sum_result == 0) {
3413 /* both parities are correct */
3415 set_bit(STRIPE_INSYNC, &sh->state);
3417 /* in contrast to the raid5 case we can validate
3418 * parity, but still have a failure to write
3421 sh->check_state = check_state_compute_result;
3422 /* Returning at this point means that we may go
3423 * off and bring p and/or q uptodate again so
3424 * we make sure to check zero_sum_result again
3425 * to verify if p or q need writeback
3429 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3430 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3431 /* don't try to repair!! */
3432 set_bit(STRIPE_INSYNC, &sh->state);
3434 int *target = &sh->ops.target;
3436 sh->ops.target = -1;
3437 sh->ops.target2 = -1;
3438 sh->check_state = check_state_compute_run;
3439 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3440 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3441 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3442 set_bit(R5_Wantcompute,
3443 &sh->dev[pd_idx].flags);
3445 target = &sh->ops.target2;
3448 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3449 set_bit(R5_Wantcompute,
3450 &sh->dev[qd_idx].flags);
3457 case check_state_compute_run:
3460 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3461 __func__, sh->check_state,
3462 (unsigned long long) sh->sector);
3467 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3471 /* We have read all the blocks in this stripe and now we need to
3472 * copy some of them into a target stripe for expand.
3474 struct dma_async_tx_descriptor *tx = NULL;
3475 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3476 for (i = 0; i < sh->disks; i++)
3477 if (i != sh->pd_idx && i != sh->qd_idx) {
3479 struct stripe_head *sh2;
3480 struct async_submit_ctl submit;
3482 sector_t bn = compute_blocknr(sh, i, 1);
3483 sector_t s = raid5_compute_sector(conf, bn, 0,
3485 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3487 /* so far only the early blocks of this stripe
3488 * have been requested. When later blocks
3489 * get requested, we will try again
3492 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3493 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3494 /* must have already done this block */
3495 release_stripe(sh2);
3499 /* place all the copies on one channel */
3500 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3501 tx = async_memcpy(sh2->dev[dd_idx].page,
3502 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3505 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3506 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3507 for (j = 0; j < conf->raid_disks; j++)
3508 if (j != sh2->pd_idx &&
3510 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3512 if (j == conf->raid_disks) {
3513 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3514 set_bit(STRIPE_HANDLE, &sh2->state);
3516 release_stripe(sh2);
3519 /* done submitting copies, wait for them to complete */
3520 async_tx_quiesce(&tx);
3524 * handle_stripe - do things to a stripe.
3526 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3527 * state of various bits to see what needs to be done.
3529 * return some read requests which now have data
3530 * return some write requests which are safely on storage
3531 * schedule a read on some buffers
3532 * schedule a write of some buffers
3533 * return confirmation of parity correctness
3537 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3539 struct r5conf *conf = sh->raid_conf;
3540 int disks = sh->disks;
3543 int do_recovery = 0;
3545 memset(s, 0, sizeof(*s));
3547 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3548 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3549 s->failed_num[0] = -1;
3550 s->failed_num[1] = -1;
3552 /* Now to look around and see what can be done */
3554 for (i=disks; i--; ) {
3555 struct md_rdev *rdev;
3562 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3564 dev->toread, dev->towrite, dev->written);
3565 /* maybe we can reply to a read
3567 * new wantfill requests are only permitted while
3568 * ops_complete_biofill is guaranteed to be inactive
3570 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3571 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3572 set_bit(R5_Wantfill, &dev->flags);
3574 /* now count some things */
3575 if (test_bit(R5_LOCKED, &dev->flags))
3577 if (test_bit(R5_UPTODATE, &dev->flags))
3579 if (test_bit(R5_Wantcompute, &dev->flags)) {
3581 BUG_ON(s->compute > 2);
3584 if (test_bit(R5_Wantfill, &dev->flags))
3586 else if (dev->toread)
3590 if (!test_bit(R5_OVERWRITE, &dev->flags))
3595 /* Prefer to use the replacement for reads, but only
3596 * if it is recovered enough and has no bad blocks.
3598 rdev = rcu_dereference(conf->disks[i].replacement);
3599 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3600 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3601 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3602 &first_bad, &bad_sectors))
3603 set_bit(R5_ReadRepl, &dev->flags);
3606 set_bit(R5_NeedReplace, &dev->flags);
3607 rdev = rcu_dereference(conf->disks[i].rdev);
3608 clear_bit(R5_ReadRepl, &dev->flags);
3610 if (rdev && test_bit(Faulty, &rdev->flags))
3613 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3614 &first_bad, &bad_sectors);
3615 if (s->blocked_rdev == NULL
3616 && (test_bit(Blocked, &rdev->flags)
3619 set_bit(BlockedBadBlocks,
3621 s->blocked_rdev = rdev;
3622 atomic_inc(&rdev->nr_pending);
3625 clear_bit(R5_Insync, &dev->flags);
3629 /* also not in-sync */
3630 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3631 test_bit(R5_UPTODATE, &dev->flags)) {
3632 /* treat as in-sync, but with a read error
3633 * which we can now try to correct
3635 set_bit(R5_Insync, &dev->flags);
3636 set_bit(R5_ReadError, &dev->flags);
3638 } else if (test_bit(In_sync, &rdev->flags))
3639 set_bit(R5_Insync, &dev->flags);
3640 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3641 /* in sync if before recovery_offset */
3642 set_bit(R5_Insync, &dev->flags);
3643 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3644 test_bit(R5_Expanded, &dev->flags))
3645 /* If we've reshaped into here, we assume it is Insync.
3646 * We will shortly update recovery_offset to make
3649 set_bit(R5_Insync, &dev->flags);
3651 if (test_bit(R5_WriteError, &dev->flags)) {
3652 /* This flag does not apply to '.replacement'
3653 * only to .rdev, so make sure to check that*/
3654 struct md_rdev *rdev2 = rcu_dereference(
3655 conf->disks[i].rdev);
3657 clear_bit(R5_Insync, &dev->flags);
3658 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3659 s->handle_bad_blocks = 1;
3660 atomic_inc(&rdev2->nr_pending);
3662 clear_bit(R5_WriteError, &dev->flags);
3664 if (test_bit(R5_MadeGood, &dev->flags)) {
3665 /* This flag does not apply to '.replacement'
3666 * only to .rdev, so make sure to check that*/
3667 struct md_rdev *rdev2 = rcu_dereference(
3668 conf->disks[i].rdev);
3669 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3670 s->handle_bad_blocks = 1;
3671 atomic_inc(&rdev2->nr_pending);
3673 clear_bit(R5_MadeGood, &dev->flags);
3675 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3676 struct md_rdev *rdev2 = rcu_dereference(
3677 conf->disks[i].replacement);
3678 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3679 s->handle_bad_blocks = 1;
3680 atomic_inc(&rdev2->nr_pending);
3682 clear_bit(R5_MadeGoodRepl, &dev->flags);
3684 if (!test_bit(R5_Insync, &dev->flags)) {
3685 /* The ReadError flag will just be confusing now */
3686 clear_bit(R5_ReadError, &dev->flags);
3687 clear_bit(R5_ReWrite, &dev->flags);
3689 if (test_bit(R5_ReadError, &dev->flags))
3690 clear_bit(R5_Insync, &dev->flags);
3691 if (!test_bit(R5_Insync, &dev->flags)) {
3693 s->failed_num[s->failed] = i;
3695 if (rdev && !test_bit(Faulty, &rdev->flags))
3699 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3700 /* If there is a failed device being replaced,
3701 * we must be recovering.
3702 * else if we are after recovery_cp, we must be syncing
3703 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3704 * else we can only be replacing
3705 * sync and recovery both need to read all devices, and so
3706 * use the same flag.
3709 sh->sector >= conf->mddev->recovery_cp ||
3710 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3718 static void handle_stripe(struct stripe_head *sh)
3720 struct stripe_head_state s;
3721 struct r5conf *conf = sh->raid_conf;
3724 int disks = sh->disks;
3725 struct r5dev *pdev, *qdev;
3727 clear_bit(STRIPE_HANDLE, &sh->state);
3728 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3729 /* already being handled, ensure it gets handled
3730 * again when current action finishes */
3731 set_bit(STRIPE_HANDLE, &sh->state);
3735 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3736 spin_lock(&sh->stripe_lock);
3737 /* Cannot process 'sync' concurrently with 'discard' */
3738 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3739 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3740 set_bit(STRIPE_SYNCING, &sh->state);
3741 clear_bit(STRIPE_INSYNC, &sh->state);
3742 clear_bit(STRIPE_REPLACED, &sh->state);
3744 spin_unlock(&sh->stripe_lock);
3746 clear_bit(STRIPE_DELAYED, &sh->state);
3748 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3749 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3750 (unsigned long long)sh->sector, sh->state,
3751 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3752 sh->check_state, sh->reconstruct_state);
3754 analyse_stripe(sh, &s);
3756 if (s.handle_bad_blocks) {
3757 set_bit(STRIPE_HANDLE, &sh->state);
3761 if (unlikely(s.blocked_rdev)) {
3762 if (s.syncing || s.expanding || s.expanded ||
3763 s.replacing || s.to_write || s.written) {
3764 set_bit(STRIPE_HANDLE, &sh->state);
3767 /* There is nothing for the blocked_rdev to block */
3768 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3769 s.blocked_rdev = NULL;
3772 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3773 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3774 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3777 pr_debug("locked=%d uptodate=%d to_read=%d"
3778 " to_write=%d failed=%d failed_num=%d,%d\n",
3779 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3780 s.failed_num[0], s.failed_num[1]);
3781 /* check if the array has lost more than max_degraded devices and,
3782 * if so, some requests might need to be failed.
3784 if (s.failed > conf->max_degraded) {
3785 sh->check_state = 0;
3786 sh->reconstruct_state = 0;
3787 if (s.to_read+s.to_write+s.written)
3788 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3789 if (s.syncing + s.replacing)
3790 handle_failed_sync(conf, sh, &s);
3793 /* Now we check to see if any write operations have recently
3797 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3799 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3800 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3801 sh->reconstruct_state = reconstruct_state_idle;
3803 /* All the 'written' buffers and the parity block are ready to
3804 * be written back to disk
3806 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3807 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3808 BUG_ON(sh->qd_idx >= 0 &&
3809 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3810 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3811 for (i = disks; i--; ) {
3812 struct r5dev *dev = &sh->dev[i];
3813 if (test_bit(R5_LOCKED, &dev->flags) &&
3814 (i == sh->pd_idx || i == sh->qd_idx ||
3816 pr_debug("Writing block %d\n", i);
3817 set_bit(R5_Wantwrite, &dev->flags);
3822 if (!test_bit(R5_Insync, &dev->flags) ||
3823 ((i == sh->pd_idx || i == sh->qd_idx) &&
3825 set_bit(STRIPE_INSYNC, &sh->state);
3828 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3829 s.dec_preread_active = 1;
3833 * might be able to return some write requests if the parity blocks
3834 * are safe, or on a failed drive
3836 pdev = &sh->dev[sh->pd_idx];
3837 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3838 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3839 qdev = &sh->dev[sh->qd_idx];
3840 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3841 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3845 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3846 && !test_bit(R5_LOCKED, &pdev->flags)
3847 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3848 test_bit(R5_Discard, &pdev->flags))))) &&
3849 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3850 && !test_bit(R5_LOCKED, &qdev->flags)
3851 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3852 test_bit(R5_Discard, &qdev->flags))))))
3853 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3855 /* Now we might consider reading some blocks, either to check/generate
3856 * parity, or to satisfy requests
3857 * or to load a block that is being partially written.
3859 if (s.to_read || s.non_overwrite
3860 || (conf->level == 6 && s.to_write && s.failed)
3861 || (s.syncing && (s.uptodate + s.compute < disks))
3864 handle_stripe_fill(sh, &s, disks);
3866 /* Now to consider new write requests and what else, if anything
3867 * should be read. We do not handle new writes when:
3868 * 1/ A 'write' operation (copy+xor) is already in flight.
3869 * 2/ A 'check' operation is in flight, as it may clobber the parity
3872 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3873 handle_stripe_dirtying(conf, sh, &s, disks);
3875 /* maybe we need to check and possibly fix the parity for this stripe
3876 * Any reads will already have been scheduled, so we just see if enough
3877 * data is available. The parity check is held off while parity
3878 * dependent operations are in flight.
3880 if (sh->check_state ||
3881 (s.syncing && s.locked == 0 &&
3882 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3883 !test_bit(STRIPE_INSYNC, &sh->state))) {
3884 if (conf->level == 6)
3885 handle_parity_checks6(conf, sh, &s, disks);
3887 handle_parity_checks5(conf, sh, &s, disks);
3890 if ((s.replacing || s.syncing) && s.locked == 0
3891 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3892 && !test_bit(STRIPE_REPLACED, &sh->state)) {
3893 /* Write out to replacement devices where possible */
3894 for (i = 0; i < conf->raid_disks; i++)
3895 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3896 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
3897 set_bit(R5_WantReplace, &sh->dev[i].flags);
3898 set_bit(R5_LOCKED, &sh->dev[i].flags);
3902 set_bit(STRIPE_INSYNC, &sh->state);
3903 set_bit(STRIPE_REPLACED, &sh->state);
3905 if ((s.syncing || s.replacing) && s.locked == 0 &&
3906 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3907 test_bit(STRIPE_INSYNC, &sh->state)) {
3908 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3909 clear_bit(STRIPE_SYNCING, &sh->state);
3910 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3911 wake_up(&conf->wait_for_overlap);
3914 /* If the failed drives are just a ReadError, then we might need
3915 * to progress the repair/check process
3917 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3918 for (i = 0; i < s.failed; i++) {
3919 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3920 if (test_bit(R5_ReadError, &dev->flags)
3921 && !test_bit(R5_LOCKED, &dev->flags)
3922 && test_bit(R5_UPTODATE, &dev->flags)
3924 if (!test_bit(R5_ReWrite, &dev->flags)) {
3925 set_bit(R5_Wantwrite, &dev->flags);
3926 set_bit(R5_ReWrite, &dev->flags);
3927 set_bit(R5_LOCKED, &dev->flags);
3930 /* let's read it back */
3931 set_bit(R5_Wantread, &dev->flags);
3932 set_bit(R5_LOCKED, &dev->flags);
3939 /* Finish reconstruct operations initiated by the expansion process */
3940 if (sh->reconstruct_state == reconstruct_state_result) {
3941 struct stripe_head *sh_src
3942 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3943 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3944 /* sh cannot be written until sh_src has been read.
3945 * so arrange for sh to be delayed a little
3947 set_bit(STRIPE_DELAYED, &sh->state);
3948 set_bit(STRIPE_HANDLE, &sh->state);
3949 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3951 atomic_inc(&conf->preread_active_stripes);
3952 release_stripe(sh_src);
3956 release_stripe(sh_src);
3958 sh->reconstruct_state = reconstruct_state_idle;
3959 clear_bit(STRIPE_EXPANDING, &sh->state);
3960 for (i = conf->raid_disks; i--; ) {
3961 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3962 set_bit(R5_LOCKED, &sh->dev[i].flags);
3967 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3968 !sh->reconstruct_state) {
3969 /* Need to write out all blocks after computing parity */
3970 sh->disks = conf->raid_disks;
3971 stripe_set_idx(sh->sector, conf, 0, sh);
3972 schedule_reconstruction(sh, &s, 1, 1);
3973 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3974 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3975 atomic_dec(&conf->reshape_stripes);
3976 wake_up(&conf->wait_for_overlap);
3977 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3980 if (s.expanding && s.locked == 0 &&
3981 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3982 handle_stripe_expansion(conf, sh);
3985 /* wait for this device to become unblocked */
3986 if (unlikely(s.blocked_rdev)) {
3987 if (conf->mddev->external)
3988 md_wait_for_blocked_rdev(s.blocked_rdev,
3991 /* Internal metadata will immediately
3992 * be written by raid5d, so we don't
3993 * need to wait here.
3995 rdev_dec_pending(s.blocked_rdev,
3999 if (s.handle_bad_blocks)
4000 for (i = disks; i--; ) {
4001 struct md_rdev *rdev;
4002 struct r5dev *dev = &sh->dev[i];
4003 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4004 /* We own a safe reference to the rdev */
4005 rdev = conf->disks[i].rdev;
4006 if (!rdev_set_badblocks(rdev, sh->sector,
4008 md_error(conf->mddev, rdev);
4009 rdev_dec_pending(rdev, conf->mddev);
4011 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4012 rdev = conf->disks[i].rdev;
4013 rdev_clear_badblocks(rdev, sh->sector,
4015 rdev_dec_pending(rdev, conf->mddev);
4017 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4018 rdev = conf->disks[i].replacement;
4020 /* rdev have been moved down */
4021 rdev = conf->disks[i].rdev;
4022 rdev_clear_badblocks(rdev, sh->sector,
4024 rdev_dec_pending(rdev, conf->mddev);
4029 raid_run_ops(sh, s.ops_request);
4033 if (s.dec_preread_active) {
4034 /* We delay this until after ops_run_io so that if make_request
4035 * is waiting on a flush, it won't continue until the writes
4036 * have actually been submitted.
4038 atomic_dec(&conf->preread_active_stripes);
4039 if (atomic_read(&conf->preread_active_stripes) <
4041 md_wakeup_thread(conf->mddev->thread);
4044 return_io(s.return_bi);
4046 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4049 static void raid5_activate_delayed(struct r5conf *conf)
4051 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4052 while (!list_empty(&conf->delayed_list)) {
4053 struct list_head *l = conf->delayed_list.next;
4054 struct stripe_head *sh;
4055 sh = list_entry(l, struct stripe_head, lru);
4057 clear_bit(STRIPE_DELAYED, &sh->state);
4058 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4059 atomic_inc(&conf->preread_active_stripes);
4060 list_add_tail(&sh->lru, &conf->hold_list);
4061 raid5_wakeup_stripe_thread(sh);
4066 static void activate_bit_delay(struct r5conf *conf,
4067 struct list_head *temp_inactive_list)
4069 /* device_lock is held */
4070 struct list_head head;
4071 list_add(&head, &conf->bitmap_list);
4072 list_del_init(&conf->bitmap_list);
4073 while (!list_empty(&head)) {
4074 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4076 list_del_init(&sh->lru);
4077 atomic_inc(&sh->count);
4078 hash = sh->hash_lock_index;
4079 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4083 int md_raid5_congested(struct mddev *mddev, int bits)
4085 struct r5conf *conf = mddev->private;
4087 /* No difference between reads and writes. Just check
4088 * how busy the stripe_cache is
4091 if (conf->inactive_blocked)
4095 if (atomic_read(&conf->empty_inactive_list_nr))
4100 EXPORT_SYMBOL_GPL(md_raid5_congested);
4102 static int raid5_congested(void *data, int bits)
4104 struct mddev *mddev = data;
4106 return mddev_congested(mddev, bits) ||
4107 md_raid5_congested(mddev, bits);
4110 /* We want read requests to align with chunks where possible,
4111 * but write requests don't need to.
4113 static int raid5_mergeable_bvec(struct request_queue *q,
4114 struct bvec_merge_data *bvm,
4115 struct bio_vec *biovec)
4117 struct mddev *mddev = q->queuedata;
4118 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4120 unsigned int chunk_sectors = mddev->chunk_sectors;
4121 unsigned int bio_sectors = bvm->bi_size >> 9;
4123 if ((bvm->bi_rw & 1) == WRITE)
4124 return biovec->bv_len; /* always allow writes to be mergeable */
4126 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4127 chunk_sectors = mddev->new_chunk_sectors;
4128 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4129 if (max < 0) max = 0;
4130 if (max <= biovec->bv_len && bio_sectors == 0)
4131 return biovec->bv_len;
4137 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4139 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4140 unsigned int chunk_sectors = mddev->chunk_sectors;
4141 unsigned int bio_sectors = bio_sectors(bio);
4143 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4144 chunk_sectors = mddev->new_chunk_sectors;
4145 return chunk_sectors >=
4146 ((sector & (chunk_sectors - 1)) + bio_sectors);
4150 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4151 * later sampled by raid5d.
4153 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4155 unsigned long flags;
4157 spin_lock_irqsave(&conf->device_lock, flags);
4159 bi->bi_next = conf->retry_read_aligned_list;
4160 conf->retry_read_aligned_list = bi;
4162 spin_unlock_irqrestore(&conf->device_lock, flags);
4163 md_wakeup_thread(conf->mddev->thread);
4167 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4171 bi = conf->retry_read_aligned;
4173 conf->retry_read_aligned = NULL;
4176 bi = conf->retry_read_aligned_list;
4178 conf->retry_read_aligned_list = bi->bi_next;
4181 * this sets the active strip count to 1 and the processed
4182 * strip count to zero (upper 8 bits)
4184 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4192 * The "raid5_align_endio" should check if the read succeeded and if it
4193 * did, call bio_endio on the original bio (having bio_put the new bio
4195 * If the read failed..
4197 static void raid5_align_endio(struct bio *bi, int error)
4199 struct bio* raid_bi = bi->bi_private;
4200 struct mddev *mddev;
4201 struct r5conf *conf;
4202 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4203 struct md_rdev *rdev;
4207 rdev = (void*)raid_bi->bi_next;
4208 raid_bi->bi_next = NULL;
4209 mddev = rdev->mddev;
4210 conf = mddev->private;
4212 rdev_dec_pending(rdev, conf->mddev);
4214 if (!error && uptodate) {
4215 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4217 bio_endio(raid_bi, 0);
4218 if (atomic_dec_and_test(&conf->active_aligned_reads))
4219 wake_up(&conf->wait_for_stripe);
4224 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4226 add_bio_to_retry(raid_bi, conf);
4229 static int bio_fits_rdev(struct bio *bi)
4231 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4233 if (bio_sectors(bi) > queue_max_sectors(q))
4235 blk_recount_segments(q, bi);
4236 if (bi->bi_phys_segments > queue_max_segments(q))
4239 if (q->merge_bvec_fn)
4240 /* it's too hard to apply the merge_bvec_fn at this stage,
4249 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4251 struct r5conf *conf = mddev->private;
4253 struct bio* align_bi;
4254 struct md_rdev *rdev;
4255 sector_t end_sector;
4257 if (!in_chunk_boundary(mddev, raid_bio)) {
4258 pr_debug("chunk_aligned_read : non aligned\n");
4262 * use bio_clone_mddev to make a copy of the bio
4264 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4268 * set bi_end_io to a new function, and set bi_private to the
4271 align_bi->bi_end_io = raid5_align_endio;
4272 align_bi->bi_private = raid_bio;
4276 align_bi->bi_iter.bi_sector =
4277 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4280 end_sector = bio_end_sector(align_bi);
4282 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4283 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4284 rdev->recovery_offset < end_sector) {
4285 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4287 (test_bit(Faulty, &rdev->flags) ||
4288 !(test_bit(In_sync, &rdev->flags) ||
4289 rdev->recovery_offset >= end_sector)))
4296 atomic_inc(&rdev->nr_pending);
4298 raid_bio->bi_next = (void*)rdev;
4299 align_bi->bi_bdev = rdev->bdev;
4300 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4302 if (!bio_fits_rdev(align_bi) ||
4303 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4304 bio_sectors(align_bi),
4305 &first_bad, &bad_sectors)) {
4306 /* too big in some way, or has a known bad block */
4308 rdev_dec_pending(rdev, mddev);
4312 /* No reshape active, so we can trust rdev->data_offset */
4313 align_bi->bi_iter.bi_sector += rdev->data_offset;
4315 spin_lock_irq(&conf->device_lock);
4316 wait_event_lock_irq(conf->wait_for_stripe,
4319 atomic_inc(&conf->active_aligned_reads);
4320 spin_unlock_irq(&conf->device_lock);
4323 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4324 align_bi, disk_devt(mddev->gendisk),
4325 raid_bio->bi_iter.bi_sector);
4326 generic_make_request(align_bi);
4335 /* __get_priority_stripe - get the next stripe to process
4337 * Full stripe writes are allowed to pass preread active stripes up until
4338 * the bypass_threshold is exceeded. In general the bypass_count
4339 * increments when the handle_list is handled before the hold_list; however, it
4340 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4341 * stripe with in flight i/o. The bypass_count will be reset when the
4342 * head of the hold_list has changed, i.e. the head was promoted to the
4345 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4347 struct stripe_head *sh = NULL, *tmp;
4348 struct list_head *handle_list = NULL;
4349 struct r5worker_group *wg = NULL;
4351 if (conf->worker_cnt_per_group == 0) {
4352 handle_list = &conf->handle_list;
4353 } else if (group != ANY_GROUP) {
4354 handle_list = &conf->worker_groups[group].handle_list;
4355 wg = &conf->worker_groups[group];
4358 for (i = 0; i < conf->group_cnt; i++) {
4359 handle_list = &conf->worker_groups[i].handle_list;
4360 wg = &conf->worker_groups[i];
4361 if (!list_empty(handle_list))
4366 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4368 list_empty(handle_list) ? "empty" : "busy",
4369 list_empty(&conf->hold_list) ? "empty" : "busy",
4370 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4372 if (!list_empty(handle_list)) {
4373 sh = list_entry(handle_list->next, typeof(*sh), lru);
4375 if (list_empty(&conf->hold_list))
4376 conf->bypass_count = 0;
4377 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4378 if (conf->hold_list.next == conf->last_hold)
4379 conf->bypass_count++;
4381 conf->last_hold = conf->hold_list.next;
4382 conf->bypass_count -= conf->bypass_threshold;
4383 if (conf->bypass_count < 0)
4384 conf->bypass_count = 0;
4387 } else if (!list_empty(&conf->hold_list) &&
4388 ((conf->bypass_threshold &&
4389 conf->bypass_count > conf->bypass_threshold) ||
4390 atomic_read(&conf->pending_full_writes) == 0)) {
4392 list_for_each_entry(tmp, &conf->hold_list, lru) {
4393 if (conf->worker_cnt_per_group == 0 ||
4394 group == ANY_GROUP ||
4395 !cpu_online(tmp->cpu) ||
4396 cpu_to_group(tmp->cpu) == group) {
4403 conf->bypass_count -= conf->bypass_threshold;
4404 if (conf->bypass_count < 0)
4405 conf->bypass_count = 0;
4417 list_del_init(&sh->lru);
4418 BUG_ON(atomic_inc_return(&sh->count) != 1);
4422 struct raid5_plug_cb {
4423 struct blk_plug_cb cb;
4424 struct list_head list;
4425 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4428 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4430 struct raid5_plug_cb *cb = container_of(
4431 blk_cb, struct raid5_plug_cb, cb);
4432 struct stripe_head *sh;
4433 struct mddev *mddev = cb->cb.data;
4434 struct r5conf *conf = mddev->private;
4438 if (cb->list.next && !list_empty(&cb->list)) {
4439 spin_lock_irq(&conf->device_lock);
4440 while (!list_empty(&cb->list)) {
4441 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4442 list_del_init(&sh->lru);
4444 * avoid race release_stripe_plug() sees
4445 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4446 * is still in our list
4448 smp_mb__before_atomic();
4449 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4451 * STRIPE_ON_RELEASE_LIST could be set here. In that
4452 * case, the count is always > 1 here
4454 hash = sh->hash_lock_index;
4455 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
4458 spin_unlock_irq(&conf->device_lock);
4460 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4461 NR_STRIPE_HASH_LOCKS);
4463 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4467 static void release_stripe_plug(struct mddev *mddev,
4468 struct stripe_head *sh)
4470 struct blk_plug_cb *blk_cb = blk_check_plugged(
4471 raid5_unplug, mddev,
4472 sizeof(struct raid5_plug_cb));
4473 struct raid5_plug_cb *cb;
4480 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4482 if (cb->list.next == NULL) {
4484 INIT_LIST_HEAD(&cb->list);
4485 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
4486 INIT_LIST_HEAD(cb->temp_inactive_list + i);
4489 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4490 list_add_tail(&sh->lru, &cb->list);
4495 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4497 struct r5conf *conf = mddev->private;
4498 sector_t logical_sector, last_sector;
4499 struct stripe_head *sh;
4503 if (mddev->reshape_position != MaxSector)
4504 /* Skip discard while reshape is happening */
4507 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4508 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
4511 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4513 stripe_sectors = conf->chunk_sectors *
4514 (conf->raid_disks - conf->max_degraded);
4515 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4517 sector_div(last_sector, stripe_sectors);
4519 logical_sector *= conf->chunk_sectors;
4520 last_sector *= conf->chunk_sectors;
4522 for (; logical_sector < last_sector;
4523 logical_sector += STRIPE_SECTORS) {
4527 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4528 prepare_to_wait(&conf->wait_for_overlap, &w,
4529 TASK_UNINTERRUPTIBLE);
4530 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4531 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4536 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4537 spin_lock_irq(&sh->stripe_lock);
4538 for (d = 0; d < conf->raid_disks; d++) {
4539 if (d == sh->pd_idx || d == sh->qd_idx)
4541 if (sh->dev[d].towrite || sh->dev[d].toread) {
4542 set_bit(R5_Overlap, &sh->dev[d].flags);
4543 spin_unlock_irq(&sh->stripe_lock);
4549 set_bit(STRIPE_DISCARD, &sh->state);
4550 finish_wait(&conf->wait_for_overlap, &w);
4551 for (d = 0; d < conf->raid_disks; d++) {
4552 if (d == sh->pd_idx || d == sh->qd_idx)
4554 sh->dev[d].towrite = bi;
4555 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4556 raid5_inc_bi_active_stripes(bi);
4558 spin_unlock_irq(&sh->stripe_lock);
4559 if (conf->mddev->bitmap) {
4561 d < conf->raid_disks - conf->max_degraded;
4563 bitmap_startwrite(mddev->bitmap,
4567 sh->bm_seq = conf->seq_flush + 1;
4568 set_bit(STRIPE_BIT_DELAY, &sh->state);
4571 set_bit(STRIPE_HANDLE, &sh->state);
4572 clear_bit(STRIPE_DELAYED, &sh->state);
4573 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4574 atomic_inc(&conf->preread_active_stripes);
4575 release_stripe_plug(mddev, sh);
4578 remaining = raid5_dec_bi_active_stripes(bi);
4579 if (remaining == 0) {
4580 md_write_end(mddev);
4585 static void make_request(struct mddev *mddev, struct bio * bi)
4587 struct r5conf *conf = mddev->private;
4589 sector_t new_sector;
4590 sector_t logical_sector, last_sector;
4591 struct stripe_head *sh;
4592 const int rw = bio_data_dir(bi);
4597 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4598 md_flush_request(mddev, bi);
4602 md_write_start(mddev, bi);
4605 mddev->reshape_position == MaxSector &&
4606 chunk_aligned_read(mddev,bi))
4609 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4610 make_discard_request(mddev, bi);
4614 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4615 last_sector = bio_end_sector(bi);
4617 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4619 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4620 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4626 seq = read_seqcount_begin(&conf->gen_lock);
4629 prepare_to_wait(&conf->wait_for_overlap, &w,
4630 TASK_UNINTERRUPTIBLE);
4631 if (unlikely(conf->reshape_progress != MaxSector)) {
4632 /* spinlock is needed as reshape_progress may be
4633 * 64bit on a 32bit platform, and so it might be
4634 * possible to see a half-updated value
4635 * Of course reshape_progress could change after
4636 * the lock is dropped, so once we get a reference
4637 * to the stripe that we think it is, we will have
4640 spin_lock_irq(&conf->device_lock);
4641 if (mddev->reshape_backwards
4642 ? logical_sector < conf->reshape_progress
4643 : logical_sector >= conf->reshape_progress) {
4646 if (mddev->reshape_backwards
4647 ? logical_sector < conf->reshape_safe
4648 : logical_sector >= conf->reshape_safe) {
4649 spin_unlock_irq(&conf->device_lock);
4655 spin_unlock_irq(&conf->device_lock);
4658 new_sector = raid5_compute_sector(conf, logical_sector,
4661 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4662 (unsigned long long)new_sector,
4663 (unsigned long long)logical_sector);
4665 sh = get_active_stripe(conf, new_sector, previous,
4666 (bi->bi_rw&RWA_MASK), 0);
4668 if (unlikely(previous)) {
4669 /* expansion might have moved on while waiting for a
4670 * stripe, so we must do the range check again.
4671 * Expansion could still move past after this
4672 * test, but as we are holding a reference to
4673 * 'sh', we know that if that happens,
4674 * STRIPE_EXPANDING will get set and the expansion
4675 * won't proceed until we finish with the stripe.
4678 spin_lock_irq(&conf->device_lock);
4679 if (mddev->reshape_backwards
4680 ? logical_sector >= conf->reshape_progress
4681 : logical_sector < conf->reshape_progress)
4682 /* mismatch, need to try again */
4684 spin_unlock_irq(&conf->device_lock);
4692 if (read_seqcount_retry(&conf->gen_lock, seq)) {
4693 /* Might have got the wrong stripe_head
4701 logical_sector >= mddev->suspend_lo &&
4702 logical_sector < mddev->suspend_hi) {
4704 /* As the suspend_* range is controlled by
4705 * userspace, we want an interruptible
4708 flush_signals(current);
4709 prepare_to_wait(&conf->wait_for_overlap,
4710 &w, TASK_INTERRUPTIBLE);
4711 if (logical_sector >= mddev->suspend_lo &&
4712 logical_sector < mddev->suspend_hi) {
4719 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4720 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4721 /* Stripe is busy expanding or
4722 * add failed due to overlap. Flush everything
4725 md_wakeup_thread(mddev->thread);
4731 set_bit(STRIPE_HANDLE, &sh->state);
4732 clear_bit(STRIPE_DELAYED, &sh->state);
4733 if ((bi->bi_rw & REQ_SYNC) &&
4734 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4735 atomic_inc(&conf->preread_active_stripes);
4736 release_stripe_plug(mddev, sh);
4738 /* cannot get stripe for read-ahead, just give-up */
4739 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4743 finish_wait(&conf->wait_for_overlap, &w);
4745 remaining = raid5_dec_bi_active_stripes(bi);
4746 if (remaining == 0) {
4749 md_write_end(mddev);
4751 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4757 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4759 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4761 /* reshaping is quite different to recovery/resync so it is
4762 * handled quite separately ... here.
4764 * On each call to sync_request, we gather one chunk worth of
4765 * destination stripes and flag them as expanding.
4766 * Then we find all the source stripes and request reads.
4767 * As the reads complete, handle_stripe will copy the data
4768 * into the destination stripe and release that stripe.
4770 struct r5conf *conf = mddev->private;
4771 struct stripe_head *sh;
4772 sector_t first_sector, last_sector;
4773 int raid_disks = conf->previous_raid_disks;
4774 int data_disks = raid_disks - conf->max_degraded;
4775 int new_data_disks = conf->raid_disks - conf->max_degraded;
4778 sector_t writepos, readpos, safepos;
4779 sector_t stripe_addr;
4780 int reshape_sectors;
4781 struct list_head stripes;
4783 if (sector_nr == 0) {
4784 /* If restarting in the middle, skip the initial sectors */
4785 if (mddev->reshape_backwards &&
4786 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4787 sector_nr = raid5_size(mddev, 0, 0)
4788 - conf->reshape_progress;
4789 } else if (!mddev->reshape_backwards &&
4790 conf->reshape_progress > 0)
4791 sector_nr = conf->reshape_progress;
4792 sector_div(sector_nr, new_data_disks);
4794 mddev->curr_resync_completed = sector_nr;
4795 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4801 /* We need to process a full chunk at a time.
4802 * If old and new chunk sizes differ, we need to process the
4805 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4806 reshape_sectors = mddev->new_chunk_sectors;
4808 reshape_sectors = mddev->chunk_sectors;
4810 /* We update the metadata at least every 10 seconds, or when
4811 * the data about to be copied would over-write the source of
4812 * the data at the front of the range. i.e. one new_stripe
4813 * along from reshape_progress new_maps to after where
4814 * reshape_safe old_maps to
4816 writepos = conf->reshape_progress;
4817 sector_div(writepos, new_data_disks);
4818 readpos = conf->reshape_progress;
4819 sector_div(readpos, data_disks);
4820 safepos = conf->reshape_safe;
4821 sector_div(safepos, data_disks);
4822 if (mddev->reshape_backwards) {
4823 writepos -= min_t(sector_t, reshape_sectors, writepos);
4824 readpos += reshape_sectors;
4825 safepos += reshape_sectors;
4827 writepos += reshape_sectors;
4828 readpos -= min_t(sector_t, reshape_sectors, readpos);
4829 safepos -= min_t(sector_t, reshape_sectors, safepos);
4832 /* Having calculated the 'writepos' possibly use it
4833 * to set 'stripe_addr' which is where we will write to.
4835 if (mddev->reshape_backwards) {
4836 BUG_ON(conf->reshape_progress == 0);
4837 stripe_addr = writepos;
4838 BUG_ON((mddev->dev_sectors &
4839 ~((sector_t)reshape_sectors - 1))
4840 - reshape_sectors - stripe_addr
4843 BUG_ON(writepos != sector_nr + reshape_sectors);
4844 stripe_addr = sector_nr;
4847 /* 'writepos' is the most advanced device address we might write.
4848 * 'readpos' is the least advanced device address we might read.
4849 * 'safepos' is the least address recorded in the metadata as having
4851 * If there is a min_offset_diff, these are adjusted either by
4852 * increasing the safepos/readpos if diff is negative, or
4853 * increasing writepos if diff is positive.
4854 * If 'readpos' is then behind 'writepos', there is no way that we can
4855 * ensure safety in the face of a crash - that must be done by userspace
4856 * making a backup of the data. So in that case there is no particular
4857 * rush to update metadata.
4858 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4859 * update the metadata to advance 'safepos' to match 'readpos' so that
4860 * we can be safe in the event of a crash.
4861 * So we insist on updating metadata if safepos is behind writepos and
4862 * readpos is beyond writepos.
4863 * In any case, update the metadata every 10 seconds.
4864 * Maybe that number should be configurable, but I'm not sure it is
4865 * worth it.... maybe it could be a multiple of safemode_delay???
4867 if (conf->min_offset_diff < 0) {
4868 safepos += -conf->min_offset_diff;
4869 readpos += -conf->min_offset_diff;
4871 writepos += conf->min_offset_diff;
4873 if ((mddev->reshape_backwards
4874 ? (safepos > writepos && readpos < writepos)
4875 : (safepos < writepos && readpos > writepos)) ||
4876 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4877 /* Cannot proceed until we've updated the superblock... */
4878 wait_event(conf->wait_for_overlap,
4879 atomic_read(&conf->reshape_stripes)==0
4880 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4881 if (atomic_read(&conf->reshape_stripes) != 0)
4883 mddev->reshape_position = conf->reshape_progress;
4884 mddev->curr_resync_completed = sector_nr;
4885 conf->reshape_checkpoint = jiffies;
4886 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4887 md_wakeup_thread(mddev->thread);
4888 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4889 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4890 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4892 spin_lock_irq(&conf->device_lock);
4893 conf->reshape_safe = mddev->reshape_position;
4894 spin_unlock_irq(&conf->device_lock);
4895 wake_up(&conf->wait_for_overlap);
4896 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4899 INIT_LIST_HEAD(&stripes);
4900 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4902 int skipped_disk = 0;
4903 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4904 set_bit(STRIPE_EXPANDING, &sh->state);
4905 atomic_inc(&conf->reshape_stripes);
4906 /* If any of this stripe is beyond the end of the old
4907 * array, then we need to zero those blocks
4909 for (j=sh->disks; j--;) {
4911 if (j == sh->pd_idx)
4913 if (conf->level == 6 &&
4916 s = compute_blocknr(sh, j, 0);
4917 if (s < raid5_size(mddev, 0, 0)) {
4921 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4922 set_bit(R5_Expanded, &sh->dev[j].flags);
4923 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4925 if (!skipped_disk) {
4926 set_bit(STRIPE_EXPAND_READY, &sh->state);
4927 set_bit(STRIPE_HANDLE, &sh->state);
4929 list_add(&sh->lru, &stripes);
4931 spin_lock_irq(&conf->device_lock);
4932 if (mddev->reshape_backwards)
4933 conf->reshape_progress -= reshape_sectors * new_data_disks;
4935 conf->reshape_progress += reshape_sectors * new_data_disks;
4936 spin_unlock_irq(&conf->device_lock);
4937 /* Ok, those stripe are ready. We can start scheduling
4938 * reads on the source stripes.
4939 * The source stripes are determined by mapping the first and last
4940 * block on the destination stripes.
4943 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4946 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4947 * new_data_disks - 1),
4949 if (last_sector >= mddev->dev_sectors)
4950 last_sector = mddev->dev_sectors - 1;
4951 while (first_sector <= last_sector) {
4952 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4953 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4954 set_bit(STRIPE_HANDLE, &sh->state);
4956 first_sector += STRIPE_SECTORS;
4958 /* Now that the sources are clearly marked, we can release
4959 * the destination stripes
4961 while (!list_empty(&stripes)) {
4962 sh = list_entry(stripes.next, struct stripe_head, lru);
4963 list_del_init(&sh->lru);
4966 /* If this takes us to the resync_max point where we have to pause,
4967 * then we need to write out the superblock.
4969 sector_nr += reshape_sectors;
4970 if ((sector_nr - mddev->curr_resync_completed) * 2
4971 >= mddev->resync_max - mddev->curr_resync_completed) {
4972 /* Cannot proceed until we've updated the superblock... */
4973 wait_event(conf->wait_for_overlap,
4974 atomic_read(&conf->reshape_stripes) == 0
4975 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4976 if (atomic_read(&conf->reshape_stripes) != 0)
4978 mddev->reshape_position = conf->reshape_progress;
4979 mddev->curr_resync_completed = sector_nr;
4980 conf->reshape_checkpoint = jiffies;
4981 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4982 md_wakeup_thread(mddev->thread);
4983 wait_event(mddev->sb_wait,
4984 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4985 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4986 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4988 spin_lock_irq(&conf->device_lock);
4989 conf->reshape_safe = mddev->reshape_position;
4990 spin_unlock_irq(&conf->device_lock);
4991 wake_up(&conf->wait_for_overlap);
4992 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4995 return reshape_sectors;
4998 /* FIXME go_faster isn't used */
4999 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
5001 struct r5conf *conf = mddev->private;
5002 struct stripe_head *sh;
5003 sector_t max_sector = mddev->dev_sectors;
5004 sector_t sync_blocks;
5005 int still_degraded = 0;
5008 if (sector_nr >= max_sector) {
5009 /* just being told to finish up .. nothing much to do */
5011 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5016 if (mddev->curr_resync < max_sector) /* aborted */
5017 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5019 else /* completed sync */
5021 bitmap_close_sync(mddev->bitmap);
5026 /* Allow raid5_quiesce to complete */
5027 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5029 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5030 return reshape_request(mddev, sector_nr, skipped);
5032 /* No need to check resync_max as we never do more than one
5033 * stripe, and as resync_max will always be on a chunk boundary,
5034 * if the check in md_do_sync didn't fire, there is no chance
5035 * of overstepping resync_max here
5038 /* if there is too many failed drives and we are trying
5039 * to resync, then assert that we are finished, because there is
5040 * nothing we can do.
5042 if (mddev->degraded >= conf->max_degraded &&
5043 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5044 sector_t rv = mddev->dev_sectors - sector_nr;
5048 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5050 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5051 sync_blocks >= STRIPE_SECTORS) {
5052 /* we can skip this block, and probably more */
5053 sync_blocks /= STRIPE_SECTORS;
5055 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5058 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5060 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5062 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5063 /* make sure we don't swamp the stripe cache if someone else
5064 * is trying to get access
5066 schedule_timeout_uninterruptible(1);
5068 /* Need to check if array will still be degraded after recovery/resync
5069 * We don't need to check the 'failed' flag as when that gets set,
5072 for (i = 0; i < conf->raid_disks; i++)
5073 if (conf->disks[i].rdev == NULL)
5076 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5078 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5079 set_bit(STRIPE_HANDLE, &sh->state);
5083 return STRIPE_SECTORS;
5086 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5088 /* We may not be able to submit a whole bio at once as there
5089 * may not be enough stripe_heads available.
5090 * We cannot pre-allocate enough stripe_heads as we may need
5091 * more than exist in the cache (if we allow ever large chunks).
5092 * So we do one stripe head at a time and record in
5093 * ->bi_hw_segments how many have been done.
5095 * We *know* that this entire raid_bio is in one chunk, so
5096 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5098 struct stripe_head *sh;
5100 sector_t sector, logical_sector, last_sector;
5105 logical_sector = raid_bio->bi_iter.bi_sector &
5106 ~((sector_t)STRIPE_SECTORS-1);
5107 sector = raid5_compute_sector(conf, logical_sector,
5109 last_sector = bio_end_sector(raid_bio);
5111 for (; logical_sector < last_sector;
5112 logical_sector += STRIPE_SECTORS,
5113 sector += STRIPE_SECTORS,
5116 if (scnt < raid5_bi_processed_stripes(raid_bio))
5117 /* already done this stripe */
5120 sh = get_active_stripe(conf, sector, 0, 1, 1);
5123 /* failed to get a stripe - must wait */
5124 raid5_set_bi_processed_stripes(raid_bio, scnt);
5125 conf->retry_read_aligned = raid_bio;
5129 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
5131 raid5_set_bi_processed_stripes(raid_bio, scnt);
5132 conf->retry_read_aligned = raid_bio;
5136 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5141 remaining = raid5_dec_bi_active_stripes(raid_bio);
5142 if (remaining == 0) {
5143 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5145 bio_endio(raid_bio, 0);
5147 if (atomic_dec_and_test(&conf->active_aligned_reads))
5148 wake_up(&conf->wait_for_stripe);
5152 static int handle_active_stripes(struct r5conf *conf, int group,
5153 struct r5worker *worker,
5154 struct list_head *temp_inactive_list)
5156 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5157 int i, batch_size = 0, hash;
5158 bool release_inactive = false;
5160 while (batch_size < MAX_STRIPE_BATCH &&
5161 (sh = __get_priority_stripe(conf, group)) != NULL)
5162 batch[batch_size++] = sh;
5164 if (batch_size == 0) {
5165 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5166 if (!list_empty(temp_inactive_list + i))
5168 if (i == NR_STRIPE_HASH_LOCKS)
5170 release_inactive = true;
5172 spin_unlock_irq(&conf->device_lock);
5174 release_inactive_stripe_list(conf, temp_inactive_list,
5175 NR_STRIPE_HASH_LOCKS);
5177 if (release_inactive) {
5178 spin_lock_irq(&conf->device_lock);
5182 for (i = 0; i < batch_size; i++)
5183 handle_stripe(batch[i]);
5187 spin_lock_irq(&conf->device_lock);
5188 for (i = 0; i < batch_size; i++) {
5189 hash = batch[i]->hash_lock_index;
5190 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5195 static void raid5_do_work(struct work_struct *work)
5197 struct r5worker *worker = container_of(work, struct r5worker, work);
5198 struct r5worker_group *group = worker->group;
5199 struct r5conf *conf = group->conf;
5200 int group_id = group - conf->worker_groups;
5202 struct blk_plug plug;
5204 pr_debug("+++ raid5worker active\n");
5206 blk_start_plug(&plug);
5208 spin_lock_irq(&conf->device_lock);
5210 int batch_size, released;
5212 released = release_stripe_list(conf, worker->temp_inactive_list);
5214 batch_size = handle_active_stripes(conf, group_id, worker,
5215 worker->temp_inactive_list);
5216 worker->working = false;
5217 if (!batch_size && !released)
5219 handled += batch_size;
5221 pr_debug("%d stripes handled\n", handled);
5223 spin_unlock_irq(&conf->device_lock);
5224 blk_finish_plug(&plug);
5226 pr_debug("--- raid5worker inactive\n");
5230 * This is our raid5 kernel thread.
5232 * We scan the hash table for stripes which can be handled now.
5233 * During the scan, completed stripes are saved for us by the interrupt
5234 * handler, so that they will not have to wait for our next wakeup.
5236 static void raid5d(struct md_thread *thread)
5238 struct mddev *mddev = thread->mddev;
5239 struct r5conf *conf = mddev->private;
5241 struct blk_plug plug;
5243 pr_debug("+++ raid5d active\n");
5245 md_check_recovery(mddev);
5247 blk_start_plug(&plug);
5249 spin_lock_irq(&conf->device_lock);
5252 int batch_size, released;
5254 released = release_stripe_list(conf, conf->temp_inactive_list);
5257 !list_empty(&conf->bitmap_list)) {
5258 /* Now is a good time to flush some bitmap updates */
5260 spin_unlock_irq(&conf->device_lock);
5261 bitmap_unplug(mddev->bitmap);
5262 spin_lock_irq(&conf->device_lock);
5263 conf->seq_write = conf->seq_flush;
5264 activate_bit_delay(conf, conf->temp_inactive_list);
5266 raid5_activate_delayed(conf);
5268 while ((bio = remove_bio_from_retry(conf))) {
5270 spin_unlock_irq(&conf->device_lock);
5271 ok = retry_aligned_read(conf, bio);
5272 spin_lock_irq(&conf->device_lock);
5278 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5279 conf->temp_inactive_list);
5280 if (!batch_size && !released)
5282 handled += batch_size;
5284 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5285 spin_unlock_irq(&conf->device_lock);
5286 md_check_recovery(mddev);
5287 spin_lock_irq(&conf->device_lock);
5290 pr_debug("%d stripes handled\n", handled);
5292 spin_unlock_irq(&conf->device_lock);
5294 async_tx_issue_pending_all();
5295 blk_finish_plug(&plug);
5297 pr_debug("--- raid5d inactive\n");
5301 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5303 struct r5conf *conf = mddev->private;
5305 return sprintf(page, "%d\n", conf->max_nr_stripes);
5311 raid5_set_cache_size(struct mddev *mddev, int size)
5313 struct r5conf *conf = mddev->private;
5317 if (size <= 16 || size > 32768)
5319 hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
5320 while (size < conf->max_nr_stripes) {
5321 if (drop_one_stripe(conf, hash))
5322 conf->max_nr_stripes--;
5327 hash = NR_STRIPE_HASH_LOCKS - 1;
5329 err = md_allow_write(mddev);
5332 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
5333 while (size > conf->max_nr_stripes) {
5334 if (grow_one_stripe(conf, hash))
5335 conf->max_nr_stripes++;
5337 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
5341 EXPORT_SYMBOL(raid5_set_cache_size);
5344 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5346 struct r5conf *conf = mddev->private;
5350 if (len >= PAGE_SIZE)
5355 if (kstrtoul(page, 10, &new))
5357 err = raid5_set_cache_size(mddev, new);
5363 static struct md_sysfs_entry
5364 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5365 raid5_show_stripe_cache_size,
5366 raid5_store_stripe_cache_size);
5369 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5371 struct r5conf *conf = mddev->private;
5373 return sprintf(page, "%d\n", conf->bypass_threshold);
5379 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
5381 struct r5conf *conf = mddev->private;
5383 if (len >= PAGE_SIZE)
5388 if (kstrtoul(page, 10, &new))
5390 if (new > conf->max_nr_stripes)
5392 conf->bypass_threshold = new;
5396 static struct md_sysfs_entry
5397 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5399 raid5_show_preread_threshold,
5400 raid5_store_preread_threshold);
5403 raid5_show_skip_copy(struct mddev *mddev, char *page)
5405 struct r5conf *conf = mddev->private;
5407 return sprintf(page, "%d\n", conf->skip_copy);
5413 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
5415 struct r5conf *conf = mddev->private;
5417 if (len >= PAGE_SIZE)
5422 if (kstrtoul(page, 10, &new))
5425 if (new == conf->skip_copy)
5428 mddev_suspend(mddev);
5429 conf->skip_copy = new;
5431 mddev->queue->backing_dev_info.capabilities |=
5432 BDI_CAP_STABLE_WRITES;
5434 mddev->queue->backing_dev_info.capabilities &=
5435 ~BDI_CAP_STABLE_WRITES;
5436 mddev_resume(mddev);
5440 static struct md_sysfs_entry
5441 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
5442 raid5_show_skip_copy,
5443 raid5_store_skip_copy);
5447 stripe_cache_active_show(struct mddev *mddev, char *page)
5449 struct r5conf *conf = mddev->private;
5451 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5456 static struct md_sysfs_entry
5457 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
5460 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
5462 struct r5conf *conf = mddev->private;
5464 return sprintf(page, "%d\n", conf->worker_cnt_per_group);
5469 static int alloc_thread_groups(struct r5conf *conf, int cnt,
5471 int *worker_cnt_per_group,
5472 struct r5worker_group **worker_groups);
5474 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
5476 struct r5conf *conf = mddev->private;
5479 struct r5worker_group *new_groups, *old_groups;
5480 int group_cnt, worker_cnt_per_group;
5482 if (len >= PAGE_SIZE)
5487 if (kstrtoul(page, 10, &new))
5490 if (new == conf->worker_cnt_per_group)
5493 mddev_suspend(mddev);
5495 old_groups = conf->worker_groups;
5497 flush_workqueue(raid5_wq);
5499 err = alloc_thread_groups(conf, new,
5500 &group_cnt, &worker_cnt_per_group,
5503 spin_lock_irq(&conf->device_lock);
5504 conf->group_cnt = group_cnt;
5505 conf->worker_cnt_per_group = worker_cnt_per_group;
5506 conf->worker_groups = new_groups;
5507 spin_unlock_irq(&conf->device_lock);
5510 kfree(old_groups[0].workers);
5514 mddev_resume(mddev);
5521 static struct md_sysfs_entry
5522 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
5523 raid5_show_group_thread_cnt,
5524 raid5_store_group_thread_cnt);
5526 static struct attribute *raid5_attrs[] = {
5527 &raid5_stripecache_size.attr,
5528 &raid5_stripecache_active.attr,
5529 &raid5_preread_bypass_threshold.attr,
5530 &raid5_group_thread_cnt.attr,
5531 &raid5_skip_copy.attr,
5534 static struct attribute_group raid5_attrs_group = {
5536 .attrs = raid5_attrs,
5539 static int alloc_thread_groups(struct r5conf *conf, int cnt,
5541 int *worker_cnt_per_group,
5542 struct r5worker_group **worker_groups)
5546 struct r5worker *workers;
5548 *worker_cnt_per_group = cnt;
5551 *worker_groups = NULL;
5554 *group_cnt = num_possible_nodes();
5555 size = sizeof(struct r5worker) * cnt;
5556 workers = kzalloc(size * *group_cnt, GFP_NOIO);
5557 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
5558 *group_cnt, GFP_NOIO);
5559 if (!*worker_groups || !workers) {
5561 kfree(*worker_groups);
5565 for (i = 0; i < *group_cnt; i++) {
5566 struct r5worker_group *group;
5568 group = &(*worker_groups)[i];
5569 INIT_LIST_HEAD(&group->handle_list);
5571 group->workers = workers + i * cnt;
5573 for (j = 0; j < cnt; j++) {
5574 struct r5worker *worker = group->workers + j;
5575 worker->group = group;
5576 INIT_WORK(&worker->work, raid5_do_work);
5578 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
5579 INIT_LIST_HEAD(worker->temp_inactive_list + k);
5586 static void free_thread_groups(struct r5conf *conf)
5588 if (conf->worker_groups)
5589 kfree(conf->worker_groups[0].workers);
5590 kfree(conf->worker_groups);
5591 conf->worker_groups = NULL;
5595 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
5597 struct r5conf *conf = mddev->private;
5600 sectors = mddev->dev_sectors;
5602 /* size is defined by the smallest of previous and new size */
5603 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
5605 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5606 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
5607 return sectors * (raid_disks - conf->max_degraded);
5610 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5612 safe_put_page(percpu->spare_page);
5613 kfree(percpu->scribble);
5614 percpu->spare_page = NULL;
5615 percpu->scribble = NULL;
5618 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5620 if (conf->level == 6 && !percpu->spare_page)
5621 percpu->spare_page = alloc_page(GFP_KERNEL);
5622 if (!percpu->scribble)
5623 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5625 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
5626 free_scratch_buffer(conf, percpu);
5633 static void raid5_free_percpu(struct r5conf *conf)
5640 #ifdef CONFIG_HOTPLUG_CPU
5641 unregister_cpu_notifier(&conf->cpu_notify);
5645 for_each_possible_cpu(cpu)
5646 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5649 free_percpu(conf->percpu);
5652 static void free_conf(struct r5conf *conf)
5654 free_thread_groups(conf);
5655 shrink_stripes(conf);
5656 raid5_free_percpu(conf);
5658 kfree(conf->stripe_hashtbl);
5662 #ifdef CONFIG_HOTPLUG_CPU
5663 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5666 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5667 long cpu = (long)hcpu;
5668 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5671 case CPU_UP_PREPARE:
5672 case CPU_UP_PREPARE_FROZEN:
5673 if (alloc_scratch_buffer(conf, percpu)) {
5674 pr_err("%s: failed memory allocation for cpu%ld\n",
5676 return notifier_from_errno(-ENOMEM);
5680 case CPU_DEAD_FROZEN:
5681 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5690 static int raid5_alloc_percpu(struct r5conf *conf)
5695 conf->percpu = alloc_percpu(struct raid5_percpu);
5699 #ifdef CONFIG_HOTPLUG_CPU
5700 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5701 conf->cpu_notify.priority = 0;
5702 err = register_cpu_notifier(&conf->cpu_notify);
5708 for_each_present_cpu(cpu) {
5709 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5711 pr_err("%s: failed memory allocation for cpu%ld\n",
5721 static struct r5conf *setup_conf(struct mddev *mddev)
5723 struct r5conf *conf;
5724 int raid_disk, memory, max_disks;
5725 struct md_rdev *rdev;
5726 struct disk_info *disk;
5729 int group_cnt, worker_cnt_per_group;
5730 struct r5worker_group *new_group;
5732 if (mddev->new_level != 5
5733 && mddev->new_level != 4
5734 && mddev->new_level != 6) {
5735 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5736 mdname(mddev), mddev->new_level);
5737 return ERR_PTR(-EIO);
5739 if ((mddev->new_level == 5
5740 && !algorithm_valid_raid5(mddev->new_layout)) ||
5741 (mddev->new_level == 6
5742 && !algorithm_valid_raid6(mddev->new_layout))) {
5743 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5744 mdname(mddev), mddev->new_layout);
5745 return ERR_PTR(-EIO);
5747 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5748 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5749 mdname(mddev), mddev->raid_disks);
5750 return ERR_PTR(-EINVAL);
5753 if (!mddev->new_chunk_sectors ||
5754 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5755 !is_power_of_2(mddev->new_chunk_sectors)) {
5756 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5757 mdname(mddev), mddev->new_chunk_sectors << 9);
5758 return ERR_PTR(-EINVAL);
5761 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5764 /* Don't enable multi-threading by default*/
5765 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
5767 conf->group_cnt = group_cnt;
5768 conf->worker_cnt_per_group = worker_cnt_per_group;
5769 conf->worker_groups = new_group;
5772 spin_lock_init(&conf->device_lock);
5773 seqcount_init(&conf->gen_lock);
5774 init_waitqueue_head(&conf->wait_for_stripe);
5775 init_waitqueue_head(&conf->wait_for_overlap);
5776 INIT_LIST_HEAD(&conf->handle_list);
5777 INIT_LIST_HEAD(&conf->hold_list);
5778 INIT_LIST_HEAD(&conf->delayed_list);
5779 INIT_LIST_HEAD(&conf->bitmap_list);
5780 init_llist_head(&conf->released_stripes);
5781 atomic_set(&conf->active_stripes, 0);
5782 atomic_set(&conf->preread_active_stripes, 0);
5783 atomic_set(&conf->active_aligned_reads, 0);
5784 conf->bypass_threshold = BYPASS_THRESHOLD;
5785 conf->recovery_disabled = mddev->recovery_disabled - 1;
5787 conf->raid_disks = mddev->raid_disks;
5788 if (mddev->reshape_position == MaxSector)
5789 conf->previous_raid_disks = mddev->raid_disks;
5791 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5792 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5793 conf->scribble_len = scribble_len(max_disks);
5795 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5800 conf->mddev = mddev;
5802 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5805 /* We init hash_locks[0] separately to that it can be used
5806 * as the reference lock in the spin_lock_nest_lock() call
5807 * in lock_all_device_hash_locks_irq in order to convince
5808 * lockdep that we know what we are doing.
5810 spin_lock_init(conf->hash_locks);
5811 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
5812 spin_lock_init(conf->hash_locks + i);
5814 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5815 INIT_LIST_HEAD(conf->inactive_list + i);
5817 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5818 INIT_LIST_HEAD(conf->temp_inactive_list + i);
5820 conf->level = mddev->new_level;
5821 if (raid5_alloc_percpu(conf) != 0)
5824 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5826 rdev_for_each(rdev, mddev) {
5827 raid_disk = rdev->raid_disk;
5828 if (raid_disk >= max_disks
5831 disk = conf->disks + raid_disk;
5833 if (test_bit(Replacement, &rdev->flags)) {
5834 if (disk->replacement)
5836 disk->replacement = rdev;
5843 if (test_bit(In_sync, &rdev->flags)) {
5844 char b[BDEVNAME_SIZE];
5845 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5847 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5848 } else if (rdev->saved_raid_disk != raid_disk)
5849 /* Cannot rely on bitmap to complete recovery */
5853 conf->chunk_sectors = mddev->new_chunk_sectors;
5854 conf->level = mddev->new_level;
5855 if (conf->level == 6)
5856 conf->max_degraded = 2;
5858 conf->max_degraded = 1;
5859 conf->algorithm = mddev->new_layout;
5860 conf->reshape_progress = mddev->reshape_position;
5861 if (conf->reshape_progress != MaxSector) {
5862 conf->prev_chunk_sectors = mddev->chunk_sectors;
5863 conf->prev_algo = mddev->layout;
5866 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5867 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5868 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
5869 if (grow_stripes(conf, NR_STRIPES)) {
5871 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5872 mdname(mddev), memory);
5875 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5876 mdname(mddev), memory);
5878 sprintf(pers_name, "raid%d", mddev->new_level);
5879 conf->thread = md_register_thread(raid5d, mddev, pers_name);
5880 if (!conf->thread) {
5882 "md/raid:%s: couldn't allocate thread.\n",
5892 return ERR_PTR(-EIO);
5894 return ERR_PTR(-ENOMEM);
5898 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5901 case ALGORITHM_PARITY_0:
5902 if (raid_disk < max_degraded)
5905 case ALGORITHM_PARITY_N:
5906 if (raid_disk >= raid_disks - max_degraded)
5909 case ALGORITHM_PARITY_0_6:
5910 if (raid_disk == 0 ||
5911 raid_disk == raid_disks - 1)
5914 case ALGORITHM_LEFT_ASYMMETRIC_6:
5915 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5916 case ALGORITHM_LEFT_SYMMETRIC_6:
5917 case ALGORITHM_RIGHT_SYMMETRIC_6:
5918 if (raid_disk == raid_disks - 1)
5924 static int run(struct mddev *mddev)
5926 struct r5conf *conf;
5927 int working_disks = 0;
5928 int dirty_parity_disks = 0;
5929 struct md_rdev *rdev;
5930 sector_t reshape_offset = 0;
5932 long long min_offset_diff = 0;
5935 if (mddev->recovery_cp != MaxSector)
5936 printk(KERN_NOTICE "md/raid:%s: not clean"
5937 " -- starting background reconstruction\n",
5940 rdev_for_each(rdev, mddev) {
5942 if (rdev->raid_disk < 0)
5944 diff = (rdev->new_data_offset - rdev->data_offset);
5946 min_offset_diff = diff;
5948 } else if (mddev->reshape_backwards &&
5949 diff < min_offset_diff)
5950 min_offset_diff = diff;
5951 else if (!mddev->reshape_backwards &&
5952 diff > min_offset_diff)
5953 min_offset_diff = diff;
5956 if (mddev->reshape_position != MaxSector) {
5957 /* Check that we can continue the reshape.
5958 * Difficulties arise if the stripe we would write to
5959 * next is at or after the stripe we would read from next.
5960 * For a reshape that changes the number of devices, this
5961 * is only possible for a very short time, and mdadm makes
5962 * sure that time appears to have past before assembling
5963 * the array. So we fail if that time hasn't passed.
5964 * For a reshape that keeps the number of devices the same
5965 * mdadm must be monitoring the reshape can keeping the
5966 * critical areas read-only and backed up. It will start
5967 * the array in read-only mode, so we check for that.
5969 sector_t here_new, here_old;
5971 int max_degraded = (mddev->level == 6 ? 2 : 1);
5973 if (mddev->new_level != mddev->level) {
5974 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5975 "required - aborting.\n",
5979 old_disks = mddev->raid_disks - mddev->delta_disks;
5980 /* reshape_position must be on a new-stripe boundary, and one
5981 * further up in new geometry must map after here in old
5984 here_new = mddev->reshape_position;
5985 if (sector_div(here_new, mddev->new_chunk_sectors *
5986 (mddev->raid_disks - max_degraded))) {
5987 printk(KERN_ERR "md/raid:%s: reshape_position not "
5988 "on a stripe boundary\n", mdname(mddev));
5991 reshape_offset = here_new * mddev->new_chunk_sectors;
5992 /* here_new is the stripe we will write to */
5993 here_old = mddev->reshape_position;
5994 sector_div(here_old, mddev->chunk_sectors *
5995 (old_disks-max_degraded));
5996 /* here_old is the first stripe that we might need to read
5998 if (mddev->delta_disks == 0) {
5999 if ((here_new * mddev->new_chunk_sectors !=
6000 here_old * mddev->chunk_sectors)) {
6001 printk(KERN_ERR "md/raid:%s: reshape position is"
6002 " confused - aborting\n", mdname(mddev));
6005 /* We cannot be sure it is safe to start an in-place
6006 * reshape. It is only safe if user-space is monitoring
6007 * and taking constant backups.
6008 * mdadm always starts a situation like this in
6009 * readonly mode so it can take control before
6010 * allowing any writes. So just check for that.
6012 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6013 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6014 /* not really in-place - so OK */;
6015 else if (mddev->ro == 0) {
6016 printk(KERN_ERR "md/raid:%s: in-place reshape "
6017 "must be started in read-only mode "
6022 } else if (mddev->reshape_backwards
6023 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6024 here_old * mddev->chunk_sectors)
6025 : (here_new * mddev->new_chunk_sectors >=
6026 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6027 /* Reading from the same stripe as writing to - bad */
6028 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6029 "auto-recovery - aborting.\n",
6033 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6035 /* OK, we should be able to continue; */
6037 BUG_ON(mddev->level != mddev->new_level);
6038 BUG_ON(mddev->layout != mddev->new_layout);
6039 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6040 BUG_ON(mddev->delta_disks != 0);
6043 if (mddev->private == NULL)
6044 conf = setup_conf(mddev);
6046 conf = mddev->private;
6049 return PTR_ERR(conf);
6051 conf->min_offset_diff = min_offset_diff;
6052 mddev->thread = conf->thread;
6053 conf->thread = NULL;
6054 mddev->private = conf;
6056 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6058 rdev = conf->disks[i].rdev;
6059 if (!rdev && conf->disks[i].replacement) {
6060 /* The replacement is all we have yet */
6061 rdev = conf->disks[i].replacement;
6062 conf->disks[i].replacement = NULL;
6063 clear_bit(Replacement, &rdev->flags);
6064 conf->disks[i].rdev = rdev;
6068 if (conf->disks[i].replacement &&
6069 conf->reshape_progress != MaxSector) {
6070 /* replacements and reshape simply do not mix. */
6071 printk(KERN_ERR "md: cannot handle concurrent "
6072 "replacement and reshape.\n");
6075 if (test_bit(In_sync, &rdev->flags)) {
6079 /* This disc is not fully in-sync. However if it
6080 * just stored parity (beyond the recovery_offset),
6081 * when we don't need to be concerned about the
6082 * array being dirty.
6083 * When reshape goes 'backwards', we never have
6084 * partially completed devices, so we only need
6085 * to worry about reshape going forwards.
6087 /* Hack because v0.91 doesn't store recovery_offset properly. */
6088 if (mddev->major_version == 0 &&
6089 mddev->minor_version > 90)
6090 rdev->recovery_offset = reshape_offset;
6092 if (rdev->recovery_offset < reshape_offset) {
6093 /* We need to check old and new layout */
6094 if (!only_parity(rdev->raid_disk,
6097 conf->max_degraded))
6100 if (!only_parity(rdev->raid_disk,
6102 conf->previous_raid_disks,
6103 conf->max_degraded))
6105 dirty_parity_disks++;
6109 * 0 for a fully functional array, 1 or 2 for a degraded array.
6111 mddev->degraded = calc_degraded(conf);
6113 if (has_failed(conf)) {
6114 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6115 " (%d/%d failed)\n",
6116 mdname(mddev), mddev->degraded, conf->raid_disks);
6120 /* device size must be a multiple of chunk size */
6121 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6122 mddev->resync_max_sectors = mddev->dev_sectors;
6124 if (mddev->degraded > dirty_parity_disks &&
6125 mddev->recovery_cp != MaxSector) {
6126 if (mddev->ok_start_degraded)
6128 "md/raid:%s: starting dirty degraded array"
6129 " - data corruption possible.\n",
6133 "md/raid:%s: cannot start dirty degraded array.\n",
6139 if (mddev->degraded == 0)
6140 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6141 " devices, algorithm %d\n", mdname(mddev), conf->level,
6142 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6145 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6146 " out of %d devices, algorithm %d\n",
6147 mdname(mddev), conf->level,
6148 mddev->raid_disks - mddev->degraded,
6149 mddev->raid_disks, mddev->new_layout);
6151 print_raid5_conf(conf);
6153 if (conf->reshape_progress != MaxSector) {
6154 conf->reshape_safe = conf->reshape_progress;
6155 atomic_set(&conf->reshape_stripes, 0);
6156 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6157 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6158 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6159 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6160 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6165 /* Ok, everything is just fine now */
6166 if (mddev->to_remove == &raid5_attrs_group)
6167 mddev->to_remove = NULL;
6168 else if (mddev->kobj.sd &&
6169 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6171 "raid5: failed to create sysfs attributes for %s\n",
6173 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6177 bool discard_supported = true;
6178 /* read-ahead size must cover two whole stripes, which
6179 * is 2 * (datadisks) * chunksize where 'n' is the
6180 * number of raid devices
6182 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6183 int stripe = data_disks *
6184 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6185 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6186 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6188 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
6190 mddev->queue->backing_dev_info.congested_data = mddev;
6191 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
6193 chunk_size = mddev->chunk_sectors << 9;
6194 blk_queue_io_min(mddev->queue, chunk_size);
6195 blk_queue_io_opt(mddev->queue, chunk_size *
6196 (conf->raid_disks - conf->max_degraded));
6197 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6199 * We can only discard a whole stripe. It doesn't make sense to
6200 * discard data disk but write parity disk
6202 stripe = stripe * PAGE_SIZE;
6203 /* Round up to power of 2, as discard handling
6204 * currently assumes that */
6205 while ((stripe-1) & stripe)
6206 stripe = (stripe | (stripe-1)) + 1;
6207 mddev->queue->limits.discard_alignment = stripe;
6208 mddev->queue->limits.discard_granularity = stripe;
6210 * unaligned part of discard request will be ignored, so can't
6211 * guarantee discard_zerors_data
6213 mddev->queue->limits.discard_zeroes_data = 0;
6215 blk_queue_max_write_same_sectors(mddev->queue, 0);
6217 rdev_for_each(rdev, mddev) {
6218 disk_stack_limits(mddev->gendisk, rdev->bdev,
6219 rdev->data_offset << 9);
6220 disk_stack_limits(mddev->gendisk, rdev->bdev,
6221 rdev->new_data_offset << 9);
6223 * discard_zeroes_data is required, otherwise data
6224 * could be lost. Consider a scenario: discard a stripe
6225 * (the stripe could be inconsistent if
6226 * discard_zeroes_data is 0); write one disk of the
6227 * stripe (the stripe could be inconsistent again
6228 * depending on which disks are used to calculate
6229 * parity); the disk is broken; The stripe data of this
6232 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6233 !bdev_get_queue(rdev->bdev)->
6234 limits.discard_zeroes_data)
6235 discard_supported = false;
6238 if (discard_supported &&
6239 mddev->queue->limits.max_discard_sectors >= stripe &&
6240 mddev->queue->limits.discard_granularity >= stripe)
6241 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6244 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6250 md_unregister_thread(&mddev->thread);
6251 print_raid5_conf(conf);
6253 mddev->private = NULL;
6254 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6258 static int stop(struct mddev *mddev)
6260 struct r5conf *conf = mddev->private;
6262 md_unregister_thread(&mddev->thread);
6264 mddev->queue->backing_dev_info.congested_fn = NULL;
6266 mddev->private = NULL;
6267 mddev->to_remove = &raid5_attrs_group;
6271 static void status(struct seq_file *seq, struct mddev *mddev)
6273 struct r5conf *conf = mddev->private;
6276 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6277 mddev->chunk_sectors / 2, mddev->layout);
6278 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6279 for (i = 0; i < conf->raid_disks; i++)
6280 seq_printf (seq, "%s",
6281 conf->disks[i].rdev &&
6282 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6283 seq_printf (seq, "]");
6286 static void print_raid5_conf (struct r5conf *conf)
6289 struct disk_info *tmp;
6291 printk(KERN_DEBUG "RAID conf printout:\n");
6293 printk("(conf==NULL)\n");
6296 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6298 conf->raid_disks - conf->mddev->degraded);
6300 for (i = 0; i < conf->raid_disks; i++) {
6301 char b[BDEVNAME_SIZE];
6302 tmp = conf->disks + i;
6304 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6305 i, !test_bit(Faulty, &tmp->rdev->flags),
6306 bdevname(tmp->rdev->bdev, b));
6310 static int raid5_spare_active(struct mddev *mddev)
6313 struct r5conf *conf = mddev->private;
6314 struct disk_info *tmp;
6316 unsigned long flags;
6318 for (i = 0; i < conf->raid_disks; i++) {
6319 tmp = conf->disks + i;
6320 if (tmp->replacement
6321 && tmp->replacement->recovery_offset == MaxSector
6322 && !test_bit(Faulty, &tmp->replacement->flags)
6323 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
6324 /* Replacement has just become active. */
6326 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
6329 /* Replaced device not technically faulty,
6330 * but we need to be sure it gets removed
6331 * and never re-added.
6333 set_bit(Faulty, &tmp->rdev->flags);
6334 sysfs_notify_dirent_safe(
6335 tmp->rdev->sysfs_state);
6337 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
6338 } else if (tmp->rdev
6339 && tmp->rdev->recovery_offset == MaxSector
6340 && !test_bit(Faulty, &tmp->rdev->flags)
6341 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
6343 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
6346 spin_lock_irqsave(&conf->device_lock, flags);
6347 mddev->degraded = calc_degraded(conf);
6348 spin_unlock_irqrestore(&conf->device_lock, flags);
6349 print_raid5_conf(conf);
6353 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
6355 struct r5conf *conf = mddev->private;
6357 int number = rdev->raid_disk;
6358 struct md_rdev **rdevp;
6359 struct disk_info *p = conf->disks + number;
6361 print_raid5_conf(conf);
6362 if (rdev == p->rdev)
6364 else if (rdev == p->replacement)
6365 rdevp = &p->replacement;
6369 if (number >= conf->raid_disks &&
6370 conf->reshape_progress == MaxSector)
6371 clear_bit(In_sync, &rdev->flags);
6373 if (test_bit(In_sync, &rdev->flags) ||
6374 atomic_read(&rdev->nr_pending)) {
6378 /* Only remove non-faulty devices if recovery
6381 if (!test_bit(Faulty, &rdev->flags) &&
6382 mddev->recovery_disabled != conf->recovery_disabled &&
6383 !has_failed(conf) &&
6384 (!p->replacement || p->replacement == rdev) &&
6385 number < conf->raid_disks) {
6391 if (atomic_read(&rdev->nr_pending)) {
6392 /* lost the race, try later */
6395 } else if (p->replacement) {
6396 /* We must have just cleared 'rdev' */
6397 p->rdev = p->replacement;
6398 clear_bit(Replacement, &p->replacement->flags);
6399 smp_mb(); /* Make sure other CPUs may see both as identical
6400 * but will never see neither - if they are careful
6402 p->replacement = NULL;
6403 clear_bit(WantReplacement, &rdev->flags);
6405 /* We might have just removed the Replacement as faulty-
6406 * clear the bit just in case
6408 clear_bit(WantReplacement, &rdev->flags);
6411 print_raid5_conf(conf);
6415 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
6417 struct r5conf *conf = mddev->private;
6420 struct disk_info *p;
6422 int last = conf->raid_disks - 1;
6424 if (mddev->recovery_disabled == conf->recovery_disabled)
6427 if (rdev->saved_raid_disk < 0 && has_failed(conf))
6428 /* no point adding a device */
6431 if (rdev->raid_disk >= 0)
6432 first = last = rdev->raid_disk;
6435 * find the disk ... but prefer rdev->saved_raid_disk
6438 if (rdev->saved_raid_disk >= 0 &&
6439 rdev->saved_raid_disk >= first &&
6440 conf->disks[rdev->saved_raid_disk].rdev == NULL)
6441 first = rdev->saved_raid_disk;
6443 for (disk = first; disk <= last; disk++) {
6444 p = conf->disks + disk;
6445 if (p->rdev == NULL) {
6446 clear_bit(In_sync, &rdev->flags);
6447 rdev->raid_disk = disk;
6449 if (rdev->saved_raid_disk != disk)
6451 rcu_assign_pointer(p->rdev, rdev);
6455 for (disk = first; disk <= last; disk++) {
6456 p = conf->disks + disk;
6457 if (test_bit(WantReplacement, &p->rdev->flags) &&
6458 p->replacement == NULL) {
6459 clear_bit(In_sync, &rdev->flags);
6460 set_bit(Replacement, &rdev->flags);
6461 rdev->raid_disk = disk;
6464 rcu_assign_pointer(p->replacement, rdev);
6469 print_raid5_conf(conf);
6473 static int raid5_resize(struct mddev *mddev, sector_t sectors)
6475 /* no resync is happening, and there is enough space
6476 * on all devices, so we can resize.
6477 * We need to make sure resync covers any new space.
6478 * If the array is shrinking we should possibly wait until
6479 * any io in the removed space completes, but it hardly seems
6483 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6484 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
6485 if (mddev->external_size &&
6486 mddev->array_sectors > newsize)
6488 if (mddev->bitmap) {
6489 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
6493 md_set_array_sectors(mddev, newsize);
6494 set_capacity(mddev->gendisk, mddev->array_sectors);
6495 revalidate_disk(mddev->gendisk);
6496 if (sectors > mddev->dev_sectors &&
6497 mddev->recovery_cp > mddev->dev_sectors) {
6498 mddev->recovery_cp = mddev->dev_sectors;
6499 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
6501 mddev->dev_sectors = sectors;
6502 mddev->resync_max_sectors = sectors;
6506 static int check_stripe_cache(struct mddev *mddev)
6508 /* Can only proceed if there are plenty of stripe_heads.
6509 * We need a minimum of one full stripe,, and for sensible progress
6510 * it is best to have about 4 times that.
6511 * If we require 4 times, then the default 256 4K stripe_heads will
6512 * allow for chunk sizes up to 256K, which is probably OK.
6513 * If the chunk size is greater, user-space should request more
6514 * stripe_heads first.
6516 struct r5conf *conf = mddev->private;
6517 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
6518 > conf->max_nr_stripes ||
6519 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
6520 > conf->max_nr_stripes) {
6521 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
6523 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
6530 static int check_reshape(struct mddev *mddev)
6532 struct r5conf *conf = mddev->private;
6534 if (mddev->delta_disks == 0 &&
6535 mddev->new_layout == mddev->layout &&
6536 mddev->new_chunk_sectors == mddev->chunk_sectors)
6537 return 0; /* nothing to do */
6538 if (has_failed(conf))
6540 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
6541 /* We might be able to shrink, but the devices must
6542 * be made bigger first.
6543 * For raid6, 4 is the minimum size.
6544 * Otherwise 2 is the minimum
6547 if (mddev->level == 6)
6549 if (mddev->raid_disks + mddev->delta_disks < min)
6553 if (!check_stripe_cache(mddev))
6556 return resize_stripes(conf, (conf->previous_raid_disks
6557 + mddev->delta_disks));
6560 static int raid5_start_reshape(struct mddev *mddev)
6562 struct r5conf *conf = mddev->private;
6563 struct md_rdev *rdev;
6565 unsigned long flags;
6567 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
6570 if (!check_stripe_cache(mddev))
6573 if (has_failed(conf))
6576 rdev_for_each(rdev, mddev) {
6577 if (!test_bit(In_sync, &rdev->flags)
6578 && !test_bit(Faulty, &rdev->flags))
6582 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
6583 /* Not enough devices even to make a degraded array
6588 /* Refuse to reduce size of the array. Any reductions in
6589 * array size must be through explicit setting of array_size
6592 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
6593 < mddev->array_sectors) {
6594 printk(KERN_ERR "md/raid:%s: array size must be reduced "
6595 "before number of disks\n", mdname(mddev));
6599 atomic_set(&conf->reshape_stripes, 0);
6600 spin_lock_irq(&conf->device_lock);
6601 write_seqcount_begin(&conf->gen_lock);
6602 conf->previous_raid_disks = conf->raid_disks;
6603 conf->raid_disks += mddev->delta_disks;
6604 conf->prev_chunk_sectors = conf->chunk_sectors;
6605 conf->chunk_sectors = mddev->new_chunk_sectors;
6606 conf->prev_algo = conf->algorithm;
6607 conf->algorithm = mddev->new_layout;
6609 /* Code that selects data_offset needs to see the generation update
6610 * if reshape_progress has been set - so a memory barrier needed.
6613 if (mddev->reshape_backwards)
6614 conf->reshape_progress = raid5_size(mddev, 0, 0);
6616 conf->reshape_progress = 0;
6617 conf->reshape_safe = conf->reshape_progress;
6618 write_seqcount_end(&conf->gen_lock);
6619 spin_unlock_irq(&conf->device_lock);
6621 /* Now make sure any requests that proceeded on the assumption
6622 * the reshape wasn't running - like Discard or Read - have
6625 mddev_suspend(mddev);
6626 mddev_resume(mddev);
6628 /* Add some new drives, as many as will fit.
6629 * We know there are enough to make the newly sized array work.
6630 * Don't add devices if we are reducing the number of
6631 * devices in the array. This is because it is not possible
6632 * to correctly record the "partially reconstructed" state of
6633 * such devices during the reshape and confusion could result.
6635 if (mddev->delta_disks >= 0) {
6636 rdev_for_each(rdev, mddev)
6637 if (rdev->raid_disk < 0 &&
6638 !test_bit(Faulty, &rdev->flags)) {
6639 if (raid5_add_disk(mddev, rdev) == 0) {
6641 >= conf->previous_raid_disks)
6642 set_bit(In_sync, &rdev->flags);
6644 rdev->recovery_offset = 0;
6646 if (sysfs_link_rdev(mddev, rdev))
6647 /* Failure here is OK */;
6649 } else if (rdev->raid_disk >= conf->previous_raid_disks
6650 && !test_bit(Faulty, &rdev->flags)) {
6651 /* This is a spare that was manually added */
6652 set_bit(In_sync, &rdev->flags);
6655 /* When a reshape changes the number of devices,
6656 * ->degraded is measured against the larger of the
6657 * pre and post number of devices.
6659 spin_lock_irqsave(&conf->device_lock, flags);
6660 mddev->degraded = calc_degraded(conf);
6661 spin_unlock_irqrestore(&conf->device_lock, flags);
6663 mddev->raid_disks = conf->raid_disks;
6664 mddev->reshape_position = conf->reshape_progress;
6665 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6667 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6668 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6669 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6670 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6671 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6673 if (!mddev->sync_thread) {
6674 mddev->recovery = 0;
6675 spin_lock_irq(&conf->device_lock);
6676 write_seqcount_begin(&conf->gen_lock);
6677 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6678 mddev->new_chunk_sectors =
6679 conf->chunk_sectors = conf->prev_chunk_sectors;
6680 mddev->new_layout = conf->algorithm = conf->prev_algo;
6681 rdev_for_each(rdev, mddev)
6682 rdev->new_data_offset = rdev->data_offset;
6684 conf->generation --;
6685 conf->reshape_progress = MaxSector;
6686 mddev->reshape_position = MaxSector;
6687 write_seqcount_end(&conf->gen_lock);
6688 spin_unlock_irq(&conf->device_lock);
6691 conf->reshape_checkpoint = jiffies;
6692 md_wakeup_thread(mddev->sync_thread);
6693 md_new_event(mddev);
6697 /* This is called from the reshape thread and should make any
6698 * changes needed in 'conf'
6700 static void end_reshape(struct r5conf *conf)
6703 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6704 struct md_rdev *rdev;
6706 spin_lock_irq(&conf->device_lock);
6707 conf->previous_raid_disks = conf->raid_disks;
6708 rdev_for_each(rdev, conf->mddev)
6709 rdev->data_offset = rdev->new_data_offset;
6711 conf->reshape_progress = MaxSector;
6712 spin_unlock_irq(&conf->device_lock);
6713 wake_up(&conf->wait_for_overlap);
6715 /* read-ahead size must cover two whole stripes, which is
6716 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6718 if (conf->mddev->queue) {
6719 int data_disks = conf->raid_disks - conf->max_degraded;
6720 int stripe = data_disks * ((conf->chunk_sectors << 9)
6722 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6723 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6728 /* This is called from the raid5d thread with mddev_lock held.
6729 * It makes config changes to the device.
6731 static void raid5_finish_reshape(struct mddev *mddev)
6733 struct r5conf *conf = mddev->private;
6735 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6737 if (mddev->delta_disks > 0) {
6738 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6739 set_capacity(mddev->gendisk, mddev->array_sectors);
6740 revalidate_disk(mddev->gendisk);
6743 spin_lock_irq(&conf->device_lock);
6744 mddev->degraded = calc_degraded(conf);
6745 spin_unlock_irq(&conf->device_lock);
6746 for (d = conf->raid_disks ;
6747 d < conf->raid_disks - mddev->delta_disks;
6749 struct md_rdev *rdev = conf->disks[d].rdev;
6751 clear_bit(In_sync, &rdev->flags);
6752 rdev = conf->disks[d].replacement;
6754 clear_bit(In_sync, &rdev->flags);
6757 mddev->layout = conf->algorithm;
6758 mddev->chunk_sectors = conf->chunk_sectors;
6759 mddev->reshape_position = MaxSector;
6760 mddev->delta_disks = 0;
6761 mddev->reshape_backwards = 0;
6765 static void raid5_quiesce(struct mddev *mddev, int state)
6767 struct r5conf *conf = mddev->private;
6770 case 2: /* resume for a suspend */
6771 wake_up(&conf->wait_for_overlap);
6774 case 1: /* stop all writes */
6775 lock_all_device_hash_locks_irq(conf);
6776 /* '2' tells resync/reshape to pause so that all
6777 * active stripes can drain
6780 wait_event_cmd(conf->wait_for_stripe,
6781 atomic_read(&conf->active_stripes) == 0 &&
6782 atomic_read(&conf->active_aligned_reads) == 0,
6783 unlock_all_device_hash_locks_irq(conf),
6784 lock_all_device_hash_locks_irq(conf));
6786 unlock_all_device_hash_locks_irq(conf);
6787 /* allow reshape to continue */
6788 wake_up(&conf->wait_for_overlap);
6791 case 0: /* re-enable writes */
6792 lock_all_device_hash_locks_irq(conf);
6794 wake_up(&conf->wait_for_stripe);
6795 wake_up(&conf->wait_for_overlap);
6796 unlock_all_device_hash_locks_irq(conf);
6802 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6804 struct r0conf *raid0_conf = mddev->private;
6807 /* for raid0 takeover only one zone is supported */
6808 if (raid0_conf->nr_strip_zones > 1) {
6809 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6811 return ERR_PTR(-EINVAL);
6814 sectors = raid0_conf->strip_zone[0].zone_end;
6815 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6816 mddev->dev_sectors = sectors;
6817 mddev->new_level = level;
6818 mddev->new_layout = ALGORITHM_PARITY_N;
6819 mddev->new_chunk_sectors = mddev->chunk_sectors;
6820 mddev->raid_disks += 1;
6821 mddev->delta_disks = 1;
6822 /* make sure it will be not marked as dirty */
6823 mddev->recovery_cp = MaxSector;
6825 return setup_conf(mddev);
6829 static void *raid5_takeover_raid1(struct mddev *mddev)
6833 if (mddev->raid_disks != 2 ||
6834 mddev->degraded > 1)
6835 return ERR_PTR(-EINVAL);
6837 /* Should check if there are write-behind devices? */
6839 chunksect = 64*2; /* 64K by default */
6841 /* The array must be an exact multiple of chunksize */
6842 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6845 if ((chunksect<<9) < STRIPE_SIZE)
6846 /* array size does not allow a suitable chunk size */
6847 return ERR_PTR(-EINVAL);
6849 mddev->new_level = 5;
6850 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6851 mddev->new_chunk_sectors = chunksect;
6853 return setup_conf(mddev);
6856 static void *raid5_takeover_raid6(struct mddev *mddev)
6860 switch (mddev->layout) {
6861 case ALGORITHM_LEFT_ASYMMETRIC_6:
6862 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6864 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6865 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6867 case ALGORITHM_LEFT_SYMMETRIC_6:
6868 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6870 case ALGORITHM_RIGHT_SYMMETRIC_6:
6871 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6873 case ALGORITHM_PARITY_0_6:
6874 new_layout = ALGORITHM_PARITY_0;
6876 case ALGORITHM_PARITY_N:
6877 new_layout = ALGORITHM_PARITY_N;
6880 return ERR_PTR(-EINVAL);
6882 mddev->new_level = 5;
6883 mddev->new_layout = new_layout;
6884 mddev->delta_disks = -1;
6885 mddev->raid_disks -= 1;
6886 return setup_conf(mddev);
6890 static int raid5_check_reshape(struct mddev *mddev)
6892 /* For a 2-drive array, the layout and chunk size can be changed
6893 * immediately as not restriping is needed.
6894 * For larger arrays we record the new value - after validation
6895 * to be used by a reshape pass.
6897 struct r5conf *conf = mddev->private;
6898 int new_chunk = mddev->new_chunk_sectors;
6900 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6902 if (new_chunk > 0) {
6903 if (!is_power_of_2(new_chunk))
6905 if (new_chunk < (PAGE_SIZE>>9))
6907 if (mddev->array_sectors & (new_chunk-1))
6908 /* not factor of array size */
6912 /* They look valid */
6914 if (mddev->raid_disks == 2) {
6915 /* can make the change immediately */
6916 if (mddev->new_layout >= 0) {
6917 conf->algorithm = mddev->new_layout;
6918 mddev->layout = mddev->new_layout;
6920 if (new_chunk > 0) {
6921 conf->chunk_sectors = new_chunk ;
6922 mddev->chunk_sectors = new_chunk;
6924 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6925 md_wakeup_thread(mddev->thread);
6927 return check_reshape(mddev);
6930 static int raid6_check_reshape(struct mddev *mddev)
6932 int new_chunk = mddev->new_chunk_sectors;
6934 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6936 if (new_chunk > 0) {
6937 if (!is_power_of_2(new_chunk))
6939 if (new_chunk < (PAGE_SIZE >> 9))
6941 if (mddev->array_sectors & (new_chunk-1))
6942 /* not factor of array size */
6946 /* They look valid */
6947 return check_reshape(mddev);
6950 static void *raid5_takeover(struct mddev *mddev)
6952 /* raid5 can take over:
6953 * raid0 - if there is only one strip zone - make it a raid4 layout
6954 * raid1 - if there are two drives. We need to know the chunk size
6955 * raid4 - trivial - just use a raid4 layout.
6956 * raid6 - Providing it is a *_6 layout
6958 if (mddev->level == 0)
6959 return raid45_takeover_raid0(mddev, 5);
6960 if (mddev->level == 1)
6961 return raid5_takeover_raid1(mddev);
6962 if (mddev->level == 4) {
6963 mddev->new_layout = ALGORITHM_PARITY_N;
6964 mddev->new_level = 5;
6965 return setup_conf(mddev);
6967 if (mddev->level == 6)
6968 return raid5_takeover_raid6(mddev);
6970 return ERR_PTR(-EINVAL);
6973 static void *raid4_takeover(struct mddev *mddev)
6975 /* raid4 can take over:
6976 * raid0 - if there is only one strip zone
6977 * raid5 - if layout is right
6979 if (mddev->level == 0)
6980 return raid45_takeover_raid0(mddev, 4);
6981 if (mddev->level == 5 &&
6982 mddev->layout == ALGORITHM_PARITY_N) {
6983 mddev->new_layout = 0;
6984 mddev->new_level = 4;
6985 return setup_conf(mddev);
6987 return ERR_PTR(-EINVAL);
6990 static struct md_personality raid5_personality;
6992 static void *raid6_takeover(struct mddev *mddev)
6994 /* Currently can only take over a raid5. We map the
6995 * personality to an equivalent raid6 personality
6996 * with the Q block at the end.
7000 if (mddev->pers != &raid5_personality)
7001 return ERR_PTR(-EINVAL);
7002 if (mddev->degraded > 1)
7003 return ERR_PTR(-EINVAL);
7004 if (mddev->raid_disks > 253)
7005 return ERR_PTR(-EINVAL);
7006 if (mddev->raid_disks < 3)
7007 return ERR_PTR(-EINVAL);
7009 switch (mddev->layout) {
7010 case ALGORITHM_LEFT_ASYMMETRIC:
7011 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7013 case ALGORITHM_RIGHT_ASYMMETRIC:
7014 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7016 case ALGORITHM_LEFT_SYMMETRIC:
7017 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7019 case ALGORITHM_RIGHT_SYMMETRIC:
7020 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7022 case ALGORITHM_PARITY_0:
7023 new_layout = ALGORITHM_PARITY_0_6;
7025 case ALGORITHM_PARITY_N:
7026 new_layout = ALGORITHM_PARITY_N;
7029 return ERR_PTR(-EINVAL);
7031 mddev->new_level = 6;
7032 mddev->new_layout = new_layout;
7033 mddev->delta_disks = 1;
7034 mddev->raid_disks += 1;
7035 return setup_conf(mddev);
7039 static struct md_personality raid6_personality =
7043 .owner = THIS_MODULE,
7044 .make_request = make_request,
7048 .error_handler = error,
7049 .hot_add_disk = raid5_add_disk,
7050 .hot_remove_disk= raid5_remove_disk,
7051 .spare_active = raid5_spare_active,
7052 .sync_request = sync_request,
7053 .resize = raid5_resize,
7055 .check_reshape = raid6_check_reshape,
7056 .start_reshape = raid5_start_reshape,
7057 .finish_reshape = raid5_finish_reshape,
7058 .quiesce = raid5_quiesce,
7059 .takeover = raid6_takeover,
7061 static struct md_personality raid5_personality =
7065 .owner = THIS_MODULE,
7066 .make_request = make_request,
7070 .error_handler = error,
7071 .hot_add_disk = raid5_add_disk,
7072 .hot_remove_disk= raid5_remove_disk,
7073 .spare_active = raid5_spare_active,
7074 .sync_request = sync_request,
7075 .resize = raid5_resize,
7077 .check_reshape = raid5_check_reshape,
7078 .start_reshape = raid5_start_reshape,
7079 .finish_reshape = raid5_finish_reshape,
7080 .quiesce = raid5_quiesce,
7081 .takeover = raid5_takeover,
7084 static struct md_personality raid4_personality =
7088 .owner = THIS_MODULE,
7089 .make_request = make_request,
7093 .error_handler = error,
7094 .hot_add_disk = raid5_add_disk,
7095 .hot_remove_disk= raid5_remove_disk,
7096 .spare_active = raid5_spare_active,
7097 .sync_request = sync_request,
7098 .resize = raid5_resize,
7100 .check_reshape = raid5_check_reshape,
7101 .start_reshape = raid5_start_reshape,
7102 .finish_reshape = raid5_finish_reshape,
7103 .quiesce = raid5_quiesce,
7104 .takeover = raid4_takeover,
7107 static int __init raid5_init(void)
7109 raid5_wq = alloc_workqueue("raid5wq",
7110 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7113 register_md_personality(&raid6_personality);
7114 register_md_personality(&raid5_personality);
7115 register_md_personality(&raid4_personality);
7119 static void raid5_exit(void)
7121 unregister_md_personality(&raid6_personality);
7122 unregister_md_personality(&raid5_personality);
7123 unregister_md_personality(&raid4_personality);
7124 destroy_workqueue(raid5_wq);
7127 module_init(raid5_init);
7128 module_exit(raid5_exit);
7129 MODULE_LICENSE("GPL");
7130 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7131 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7132 MODULE_ALIAS("md-raid5");
7133 MODULE_ALIAS("md-raid4");
7134 MODULE_ALIAS("md-level-5");
7135 MODULE_ALIAS("md-level-4");
7136 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7137 MODULE_ALIAS("md-raid6");
7138 MODULE_ALIAS("md-level-6");
7140 /* This used to be two separate modules, they were: */
7141 MODULE_ALIAS("raid5");
7142 MODULE_ALIAS("raid6");