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 bool devices_handle_discard_safely = false;
68 module_param(devices_handle_discard_safely, bool, 0644);
69 MODULE_PARM_DESC(devices_handle_discard_safely,
70 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
71 static struct workqueue_struct *raid5_wq;
76 #define NR_STRIPES 256
77 #define STRIPE_SIZE PAGE_SIZE
78 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
79 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
80 #define IO_THRESHOLD 1
81 #define BYPASS_THRESHOLD 1
82 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
83 #define HASH_MASK (NR_HASH - 1)
84 #define MAX_STRIPE_BATCH 8
86 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
88 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
89 return &conf->stripe_hashtbl[hash];
92 static inline int stripe_hash_locks_hash(sector_t sect)
94 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
97 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
99 spin_lock_irq(conf->hash_locks + hash);
100 spin_lock(&conf->device_lock);
103 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
105 spin_unlock(&conf->device_lock);
106 spin_unlock_irq(conf->hash_locks + hash);
109 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
113 spin_lock(conf->hash_locks);
114 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
115 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
116 spin_lock(&conf->device_lock);
119 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
122 spin_unlock(&conf->device_lock);
123 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
124 spin_unlock(conf->hash_locks + i - 1);
128 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
129 * order without overlap. There may be several bio's per stripe+device, and
130 * a bio could span several devices.
131 * When walking this list for a particular stripe+device, we must never proceed
132 * beyond a bio that extends past this device, as the next bio might no longer
134 * This function is used to determine the 'next' bio in the list, given the sector
135 * of the current stripe+device
137 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
139 int sectors = bio_sectors(bio);
140 if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
147 * We maintain a biased count of active stripes in the bottom 16 bits of
148 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
150 static inline int raid5_bi_processed_stripes(struct bio *bio)
152 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
153 return (atomic_read(segments) >> 16) & 0xffff;
156 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
158 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
159 return atomic_sub_return(1, segments) & 0xffff;
162 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
164 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
165 atomic_inc(segments);
168 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
171 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
175 old = atomic_read(segments);
176 new = (old & 0xffff) | (cnt << 16);
177 } while (atomic_cmpxchg(segments, old, new) != old);
180 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
182 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
183 atomic_set(segments, cnt);
186 /* Find first data disk in a raid6 stripe */
187 static inline int raid6_d0(struct stripe_head *sh)
190 /* ddf always start from first device */
192 /* md starts just after Q block */
193 if (sh->qd_idx == sh->disks - 1)
196 return sh->qd_idx + 1;
198 static inline int raid6_next_disk(int disk, int raid_disks)
201 return (disk < raid_disks) ? disk : 0;
204 /* When walking through the disks in a raid5, starting at raid6_d0,
205 * We need to map each disk to a 'slot', where the data disks are slot
206 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
207 * is raid_disks-1. This help does that mapping.
209 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
210 int *count, int syndrome_disks)
216 if (idx == sh->pd_idx)
217 return syndrome_disks;
218 if (idx == sh->qd_idx)
219 return syndrome_disks + 1;
225 static void return_io(struct bio *return_bi)
227 struct bio *bi = return_bi;
230 return_bi = bi->bi_next;
232 bi->bi_iter.bi_size = 0;
233 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
240 static void print_raid5_conf (struct r5conf *conf);
242 static int stripe_operations_active(struct stripe_head *sh)
244 return sh->check_state || sh->reconstruct_state ||
245 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
246 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
249 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
251 struct r5conf *conf = sh->raid_conf;
252 struct r5worker_group *group;
254 int i, cpu = sh->cpu;
256 if (!cpu_online(cpu)) {
257 cpu = cpumask_any(cpu_online_mask);
261 if (list_empty(&sh->lru)) {
262 struct r5worker_group *group;
263 group = conf->worker_groups + cpu_to_group(cpu);
264 list_add_tail(&sh->lru, &group->handle_list);
265 group->stripes_cnt++;
269 if (conf->worker_cnt_per_group == 0) {
270 md_wakeup_thread(conf->mddev->thread);
274 group = conf->worker_groups + cpu_to_group(sh->cpu);
276 group->workers[0].working = true;
277 /* at least one worker should run to avoid race */
278 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
280 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
281 /* wakeup more workers */
282 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
283 if (group->workers[i].working == false) {
284 group->workers[i].working = true;
285 queue_work_on(sh->cpu, raid5_wq,
286 &group->workers[i].work);
292 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
293 struct list_head *temp_inactive_list)
295 BUG_ON(!list_empty(&sh->lru));
296 BUG_ON(atomic_read(&conf->active_stripes)==0);
297 if (test_bit(STRIPE_HANDLE, &sh->state)) {
298 if (test_bit(STRIPE_DELAYED, &sh->state) &&
299 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
300 list_add_tail(&sh->lru, &conf->delayed_list);
301 if (atomic_read(&conf->preread_active_stripes)
303 md_wakeup_thread(conf->mddev->thread);
304 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
305 sh->bm_seq - conf->seq_write > 0)
306 list_add_tail(&sh->lru, &conf->bitmap_list);
308 clear_bit(STRIPE_DELAYED, &sh->state);
309 clear_bit(STRIPE_BIT_DELAY, &sh->state);
310 if (conf->worker_cnt_per_group == 0) {
311 list_add_tail(&sh->lru, &conf->handle_list);
313 raid5_wakeup_stripe_thread(sh);
317 md_wakeup_thread(conf->mddev->thread);
319 BUG_ON(stripe_operations_active(sh));
320 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
321 if (atomic_dec_return(&conf->preread_active_stripes)
323 md_wakeup_thread(conf->mddev->thread);
324 atomic_dec(&conf->active_stripes);
325 if (!test_bit(STRIPE_EXPANDING, &sh->state))
326 list_add_tail(&sh->lru, temp_inactive_list);
330 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
331 struct list_head *temp_inactive_list)
333 if (atomic_dec_and_test(&sh->count))
334 do_release_stripe(conf, sh, temp_inactive_list);
338 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
340 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
341 * given time. Adding stripes only takes device lock, while deleting stripes
342 * only takes hash lock.
344 static void release_inactive_stripe_list(struct r5conf *conf,
345 struct list_head *temp_inactive_list,
349 bool do_wakeup = false;
352 if (hash == NR_STRIPE_HASH_LOCKS) {
353 size = NR_STRIPE_HASH_LOCKS;
354 hash = NR_STRIPE_HASH_LOCKS - 1;
358 struct list_head *list = &temp_inactive_list[size - 1];
361 * We don't hold any lock here yet, get_active_stripe() might
362 * remove stripes from the list
364 if (!list_empty_careful(list)) {
365 spin_lock_irqsave(conf->hash_locks + hash, flags);
366 if (list_empty(conf->inactive_list + hash) &&
368 atomic_dec(&conf->empty_inactive_list_nr);
369 list_splice_tail_init(list, conf->inactive_list + hash);
371 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
378 wake_up(&conf->wait_for_stripe);
379 if (conf->retry_read_aligned)
380 md_wakeup_thread(conf->mddev->thread);
384 /* should hold conf->device_lock already */
385 static int release_stripe_list(struct r5conf *conf,
386 struct list_head *temp_inactive_list)
388 struct stripe_head *sh;
390 struct llist_node *head;
392 head = llist_del_all(&conf->released_stripes);
393 head = llist_reverse_order(head);
397 sh = llist_entry(head, struct stripe_head, release_list);
398 head = llist_next(head);
399 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
401 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
403 * Don't worry the bit is set here, because if the bit is set
404 * again, the count is always > 1. This is true for
405 * STRIPE_ON_UNPLUG_LIST bit too.
407 hash = sh->hash_lock_index;
408 __release_stripe(conf, sh, &temp_inactive_list[hash]);
415 static void release_stripe(struct stripe_head *sh)
417 struct r5conf *conf = sh->raid_conf;
419 struct list_head list;
423 /* Avoid release_list until the last reference.
425 if (atomic_add_unless(&sh->count, -1, 1))
428 if (unlikely(!conf->mddev->thread) ||
429 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
431 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
433 md_wakeup_thread(conf->mddev->thread);
436 local_irq_save(flags);
437 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
438 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
439 INIT_LIST_HEAD(&list);
440 hash = sh->hash_lock_index;
441 do_release_stripe(conf, sh, &list);
442 spin_unlock(&conf->device_lock);
443 release_inactive_stripe_list(conf, &list, hash);
445 local_irq_restore(flags);
448 static inline void remove_hash(struct stripe_head *sh)
450 pr_debug("remove_hash(), stripe %llu\n",
451 (unsigned long long)sh->sector);
453 hlist_del_init(&sh->hash);
456 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
458 struct hlist_head *hp = stripe_hash(conf, sh->sector);
460 pr_debug("insert_hash(), stripe %llu\n",
461 (unsigned long long)sh->sector);
463 hlist_add_head(&sh->hash, hp);
466 /* find an idle stripe, make sure it is unhashed, and return it. */
467 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
469 struct stripe_head *sh = NULL;
470 struct list_head *first;
472 if (list_empty(conf->inactive_list + hash))
474 first = (conf->inactive_list + hash)->next;
475 sh = list_entry(first, struct stripe_head, lru);
476 list_del_init(first);
478 atomic_inc(&conf->active_stripes);
479 BUG_ON(hash != sh->hash_lock_index);
480 if (list_empty(conf->inactive_list + hash))
481 atomic_inc(&conf->empty_inactive_list_nr);
486 static void shrink_buffers(struct stripe_head *sh)
490 int num = sh->raid_conf->pool_size;
492 for (i = 0; i < num ; i++) {
493 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
497 sh->dev[i].page = NULL;
502 static int grow_buffers(struct stripe_head *sh)
505 int num = sh->raid_conf->pool_size;
507 for (i = 0; i < num; i++) {
510 if (!(page = alloc_page(GFP_KERNEL))) {
513 sh->dev[i].page = page;
514 sh->dev[i].orig_page = page;
519 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
520 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
521 struct stripe_head *sh);
523 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
525 struct r5conf *conf = sh->raid_conf;
528 BUG_ON(atomic_read(&sh->count) != 0);
529 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
530 BUG_ON(stripe_operations_active(sh));
532 pr_debug("init_stripe called, stripe %llu\n",
533 (unsigned long long)sector);
535 seq = read_seqcount_begin(&conf->gen_lock);
536 sh->generation = conf->generation - previous;
537 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
539 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,
1349 static void ops_complete_prexor(void *stripe_head_ref)
1351 struct stripe_head *sh = stripe_head_ref;
1353 pr_debug("%s: stripe %llu\n", __func__,
1354 (unsigned long long)sh->sector);
1357 static struct dma_async_tx_descriptor *
1358 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1359 struct dma_async_tx_descriptor *tx)
1361 int disks = sh->disks;
1362 struct page **xor_srcs = percpu->scribble;
1363 int count = 0, pd_idx = sh->pd_idx, i;
1364 struct async_submit_ctl submit;
1366 /* existing parity data subtracted */
1367 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1369 pr_debug("%s: stripe %llu\n", __func__,
1370 (unsigned long long)sh->sector);
1372 for (i = disks; i--; ) {
1373 struct r5dev *dev = &sh->dev[i];
1374 /* Only process blocks that are known to be uptodate */
1375 if (test_bit(R5_Wantdrain, &dev->flags))
1376 xor_srcs[count++] = dev->page;
1379 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1380 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1381 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1386 static struct dma_async_tx_descriptor *
1387 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1389 int disks = sh->disks;
1392 pr_debug("%s: stripe %llu\n", __func__,
1393 (unsigned long long)sh->sector);
1395 for (i = disks; i--; ) {
1396 struct r5dev *dev = &sh->dev[i];
1399 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1402 spin_lock_irq(&sh->stripe_lock);
1403 chosen = dev->towrite;
1404 dev->towrite = NULL;
1405 BUG_ON(dev->written);
1406 wbi = dev->written = chosen;
1407 spin_unlock_irq(&sh->stripe_lock);
1408 WARN_ON(dev->page != dev->orig_page);
1410 while (wbi && wbi->bi_iter.bi_sector <
1411 dev->sector + STRIPE_SECTORS) {
1412 if (wbi->bi_rw & REQ_FUA)
1413 set_bit(R5_WantFUA, &dev->flags);
1414 if (wbi->bi_rw & REQ_SYNC)
1415 set_bit(R5_SyncIO, &dev->flags);
1416 if (wbi->bi_rw & REQ_DISCARD)
1417 set_bit(R5_Discard, &dev->flags);
1419 tx = async_copy_data(1, wbi, &dev->page,
1420 dev->sector, tx, sh);
1421 if (dev->page != dev->orig_page) {
1422 set_bit(R5_SkipCopy, &dev->flags);
1423 clear_bit(R5_UPTODATE, &dev->flags);
1424 clear_bit(R5_OVERWRITE, &dev->flags);
1427 wbi = r5_next_bio(wbi, dev->sector);
1435 static void ops_complete_reconstruct(void *stripe_head_ref)
1437 struct stripe_head *sh = stripe_head_ref;
1438 int disks = sh->disks;
1439 int pd_idx = sh->pd_idx;
1440 int qd_idx = sh->qd_idx;
1442 bool fua = false, sync = false, discard = false;
1444 pr_debug("%s: stripe %llu\n", __func__,
1445 (unsigned long long)sh->sector);
1447 for (i = disks; i--; ) {
1448 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1449 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1450 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1453 for (i = disks; i--; ) {
1454 struct r5dev *dev = &sh->dev[i];
1456 if (dev->written || i == pd_idx || i == qd_idx) {
1457 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1458 set_bit(R5_UPTODATE, &dev->flags);
1460 set_bit(R5_WantFUA, &dev->flags);
1462 set_bit(R5_SyncIO, &dev->flags);
1466 if (sh->reconstruct_state == reconstruct_state_drain_run)
1467 sh->reconstruct_state = reconstruct_state_drain_result;
1468 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1469 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1471 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1472 sh->reconstruct_state = reconstruct_state_result;
1475 set_bit(STRIPE_HANDLE, &sh->state);
1480 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1481 struct dma_async_tx_descriptor *tx)
1483 int disks = sh->disks;
1484 struct page **xor_srcs = percpu->scribble;
1485 struct async_submit_ctl submit;
1486 int count = 0, pd_idx = sh->pd_idx, i;
1487 struct page *xor_dest;
1489 unsigned long flags;
1491 pr_debug("%s: stripe %llu\n", __func__,
1492 (unsigned long long)sh->sector);
1494 for (i = 0; i < sh->disks; i++) {
1497 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1500 if (i >= sh->disks) {
1501 atomic_inc(&sh->count);
1502 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1503 ops_complete_reconstruct(sh);
1506 /* check if prexor is active which means only process blocks
1507 * that are part of a read-modify-write (written)
1509 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1511 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1512 for (i = disks; i--; ) {
1513 struct r5dev *dev = &sh->dev[i];
1515 xor_srcs[count++] = dev->page;
1518 xor_dest = sh->dev[pd_idx].page;
1519 for (i = disks; i--; ) {
1520 struct r5dev *dev = &sh->dev[i];
1522 xor_srcs[count++] = dev->page;
1526 /* 1/ if we prexor'd then the dest is reused as a source
1527 * 2/ if we did not prexor then we are redoing the parity
1528 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1529 * for the synchronous xor case
1531 flags = ASYNC_TX_ACK |
1532 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1534 atomic_inc(&sh->count);
1536 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1537 to_addr_conv(sh, percpu));
1538 if (unlikely(count == 1))
1539 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1541 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1545 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1546 struct dma_async_tx_descriptor *tx)
1548 struct async_submit_ctl submit;
1549 struct page **blocks = percpu->scribble;
1552 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1554 for (i = 0; i < sh->disks; i++) {
1555 if (sh->pd_idx == i || sh->qd_idx == i)
1557 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1560 if (i >= sh->disks) {
1561 atomic_inc(&sh->count);
1562 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1563 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1564 ops_complete_reconstruct(sh);
1568 count = set_syndrome_sources(blocks, sh);
1570 atomic_inc(&sh->count);
1572 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1573 sh, to_addr_conv(sh, percpu));
1574 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1577 static void ops_complete_check(void *stripe_head_ref)
1579 struct stripe_head *sh = stripe_head_ref;
1581 pr_debug("%s: stripe %llu\n", __func__,
1582 (unsigned long long)sh->sector);
1584 sh->check_state = check_state_check_result;
1585 set_bit(STRIPE_HANDLE, &sh->state);
1589 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1591 int disks = sh->disks;
1592 int pd_idx = sh->pd_idx;
1593 int qd_idx = sh->qd_idx;
1594 struct page *xor_dest;
1595 struct page **xor_srcs = percpu->scribble;
1596 struct dma_async_tx_descriptor *tx;
1597 struct async_submit_ctl submit;
1601 pr_debug("%s: stripe %llu\n", __func__,
1602 (unsigned long long)sh->sector);
1605 xor_dest = sh->dev[pd_idx].page;
1606 xor_srcs[count++] = xor_dest;
1607 for (i = disks; i--; ) {
1608 if (i == pd_idx || i == qd_idx)
1610 xor_srcs[count++] = sh->dev[i].page;
1613 init_async_submit(&submit, 0, NULL, NULL, NULL,
1614 to_addr_conv(sh, percpu));
1615 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1616 &sh->ops.zero_sum_result, &submit);
1618 atomic_inc(&sh->count);
1619 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1620 tx = async_trigger_callback(&submit);
1623 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1625 struct page **srcs = percpu->scribble;
1626 struct async_submit_ctl submit;
1629 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1630 (unsigned long long)sh->sector, checkp);
1632 count = set_syndrome_sources(srcs, sh);
1636 atomic_inc(&sh->count);
1637 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1638 sh, to_addr_conv(sh, percpu));
1639 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1640 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1643 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1645 int overlap_clear = 0, i, disks = sh->disks;
1646 struct dma_async_tx_descriptor *tx = NULL;
1647 struct r5conf *conf = sh->raid_conf;
1648 int level = conf->level;
1649 struct raid5_percpu *percpu;
1653 percpu = per_cpu_ptr(conf->percpu, cpu);
1654 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1655 ops_run_biofill(sh);
1659 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1661 tx = ops_run_compute5(sh, percpu);
1663 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1664 tx = ops_run_compute6_1(sh, percpu);
1666 tx = ops_run_compute6_2(sh, percpu);
1668 /* terminate the chain if reconstruct is not set to be run */
1669 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1673 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1674 tx = ops_run_prexor(sh, percpu, tx);
1676 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1677 tx = ops_run_biodrain(sh, tx);
1681 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1683 ops_run_reconstruct5(sh, percpu, tx);
1685 ops_run_reconstruct6(sh, percpu, tx);
1688 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1689 if (sh->check_state == check_state_run)
1690 ops_run_check_p(sh, percpu);
1691 else if (sh->check_state == check_state_run_q)
1692 ops_run_check_pq(sh, percpu, 0);
1693 else if (sh->check_state == check_state_run_pq)
1694 ops_run_check_pq(sh, percpu, 1);
1700 for (i = disks; i--; ) {
1701 struct r5dev *dev = &sh->dev[i];
1702 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1703 wake_up(&sh->raid_conf->wait_for_overlap);
1708 static int grow_one_stripe(struct r5conf *conf, int hash)
1710 struct stripe_head *sh;
1711 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1715 sh->raid_conf = conf;
1717 spin_lock_init(&sh->stripe_lock);
1719 if (grow_buffers(sh)) {
1721 kmem_cache_free(conf->slab_cache, sh);
1724 sh->hash_lock_index = hash;
1725 /* we just created an active stripe so... */
1726 atomic_set(&sh->count, 1);
1727 atomic_inc(&conf->active_stripes);
1728 INIT_LIST_HEAD(&sh->lru);
1733 static int grow_stripes(struct r5conf *conf, int num)
1735 struct kmem_cache *sc;
1736 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1739 if (conf->mddev->gendisk)
1740 sprintf(conf->cache_name[0],
1741 "raid%d-%s", conf->level, mdname(conf->mddev));
1743 sprintf(conf->cache_name[0],
1744 "raid%d-%p", conf->level, conf->mddev);
1745 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1747 conf->active_name = 0;
1748 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1749 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1753 conf->slab_cache = sc;
1754 conf->pool_size = devs;
1755 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
1757 if (!grow_one_stripe(conf, hash))
1759 conf->max_nr_stripes++;
1760 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
1766 * scribble_len - return the required size of the scribble region
1767 * @num - total number of disks in the array
1769 * The size must be enough to contain:
1770 * 1/ a struct page pointer for each device in the array +2
1771 * 2/ room to convert each entry in (1) to its corresponding dma
1772 * (dma_map_page()) or page (page_address()) address.
1774 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1775 * calculate over all devices (not just the data blocks), using zeros in place
1776 * of the P and Q blocks.
1778 static size_t scribble_len(int num)
1782 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1787 static int resize_stripes(struct r5conf *conf, int newsize)
1789 /* Make all the stripes able to hold 'newsize' devices.
1790 * New slots in each stripe get 'page' set to a new page.
1792 * This happens in stages:
1793 * 1/ create a new kmem_cache and allocate the required number of
1795 * 2/ gather all the old stripe_heads and transfer the pages across
1796 * to the new stripe_heads. This will have the side effect of
1797 * freezing the array as once all stripe_heads have been collected,
1798 * no IO will be possible. Old stripe heads are freed once their
1799 * pages have been transferred over, and the old kmem_cache is
1800 * freed when all stripes are done.
1801 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1802 * we simple return a failre status - no need to clean anything up.
1803 * 4/ allocate new pages for the new slots in the new stripe_heads.
1804 * If this fails, we don't bother trying the shrink the
1805 * stripe_heads down again, we just leave them as they are.
1806 * As each stripe_head is processed the new one is released into
1809 * Once step2 is started, we cannot afford to wait for a write,
1810 * so we use GFP_NOIO allocations.
1812 struct stripe_head *osh, *nsh;
1813 LIST_HEAD(newstripes);
1814 struct disk_info *ndisks;
1817 struct kmem_cache *sc;
1821 if (newsize <= conf->pool_size)
1822 return 0; /* never bother to shrink */
1824 err = md_allow_write(conf->mddev);
1829 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1830 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1835 for (i = conf->max_nr_stripes; i; i--) {
1836 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1840 nsh->raid_conf = conf;
1841 spin_lock_init(&nsh->stripe_lock);
1843 list_add(&nsh->lru, &newstripes);
1846 /* didn't get enough, give up */
1847 while (!list_empty(&newstripes)) {
1848 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1849 list_del(&nsh->lru);
1850 kmem_cache_free(sc, nsh);
1852 kmem_cache_destroy(sc);
1855 /* Step 2 - Must use GFP_NOIO now.
1856 * OK, we have enough stripes, start collecting inactive
1857 * stripes and copying them over
1861 list_for_each_entry(nsh, &newstripes, lru) {
1862 lock_device_hash_lock(conf, hash);
1863 wait_event_cmd(conf->wait_for_stripe,
1864 !list_empty(conf->inactive_list + hash),
1865 unlock_device_hash_lock(conf, hash),
1866 lock_device_hash_lock(conf, hash));
1867 osh = get_free_stripe(conf, hash);
1868 unlock_device_hash_lock(conf, hash);
1869 atomic_set(&nsh->count, 1);
1870 for(i=0; i<conf->pool_size; i++) {
1871 nsh->dev[i].page = osh->dev[i].page;
1872 nsh->dev[i].orig_page = osh->dev[i].page;
1874 for( ; i<newsize; i++)
1875 nsh->dev[i].page = NULL;
1876 nsh->hash_lock_index = hash;
1877 kmem_cache_free(conf->slab_cache, osh);
1879 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
1880 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
1885 kmem_cache_destroy(conf->slab_cache);
1888 * At this point, we are holding all the stripes so the array
1889 * is completely stalled, so now is a good time to resize
1890 * conf->disks and the scribble region
1892 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1894 for (i=0; i<conf->raid_disks; i++)
1895 ndisks[i] = conf->disks[i];
1897 conf->disks = ndisks;
1902 conf->scribble_len = scribble_len(newsize);
1903 for_each_present_cpu(cpu) {
1904 struct raid5_percpu *percpu;
1907 percpu = per_cpu_ptr(conf->percpu, cpu);
1908 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1911 kfree(percpu->scribble);
1912 percpu->scribble = scribble;
1920 /* Step 4, return new stripes to service */
1921 while(!list_empty(&newstripes)) {
1922 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1923 list_del_init(&nsh->lru);
1925 for (i=conf->raid_disks; i < newsize; i++)
1926 if (nsh->dev[i].page == NULL) {
1927 struct page *p = alloc_page(GFP_NOIO);
1928 nsh->dev[i].page = p;
1929 nsh->dev[i].orig_page = p;
1933 release_stripe(nsh);
1935 /* critical section pass, GFP_NOIO no longer needed */
1937 conf->slab_cache = sc;
1938 conf->active_name = 1-conf->active_name;
1939 conf->pool_size = newsize;
1943 static int drop_one_stripe(struct r5conf *conf, int hash)
1945 struct stripe_head *sh;
1947 spin_lock_irq(conf->hash_locks + hash);
1948 sh = get_free_stripe(conf, hash);
1949 spin_unlock_irq(conf->hash_locks + hash);
1952 BUG_ON(atomic_read(&sh->count));
1954 kmem_cache_free(conf->slab_cache, sh);
1955 atomic_dec(&conf->active_stripes);
1959 static void shrink_stripes(struct r5conf *conf)
1962 for (hash = 0; hash < NR_STRIPE_HASH_LOCKS; hash++)
1963 while (drop_one_stripe(conf, hash))
1966 if (conf->slab_cache)
1967 kmem_cache_destroy(conf->slab_cache);
1968 conf->slab_cache = NULL;
1971 static void raid5_end_read_request(struct bio * bi, int error)
1973 struct stripe_head *sh = bi->bi_private;
1974 struct r5conf *conf = sh->raid_conf;
1975 int disks = sh->disks, i;
1976 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1977 char b[BDEVNAME_SIZE];
1978 struct md_rdev *rdev = NULL;
1981 for (i=0 ; i<disks; i++)
1982 if (bi == &sh->dev[i].req)
1985 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1986 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1992 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1993 /* If replacement finished while this request was outstanding,
1994 * 'replacement' might be NULL already.
1995 * In that case it moved down to 'rdev'.
1996 * rdev is not removed until all requests are finished.
1998 rdev = conf->disks[i].replacement;
2000 rdev = conf->disks[i].rdev;
2002 if (use_new_offset(conf, sh))
2003 s = sh->sector + rdev->new_data_offset;
2005 s = sh->sector + rdev->data_offset;
2007 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2008 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2009 /* Note that this cannot happen on a
2010 * replacement device. We just fail those on
2015 "md/raid:%s: read error corrected"
2016 " (%lu sectors at %llu on %s)\n",
2017 mdname(conf->mddev), STRIPE_SECTORS,
2018 (unsigned long long)s,
2019 bdevname(rdev->bdev, b));
2020 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2021 clear_bit(R5_ReadError, &sh->dev[i].flags);
2022 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2023 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2024 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2026 if (atomic_read(&rdev->read_errors))
2027 atomic_set(&rdev->read_errors, 0);
2029 const char *bdn = bdevname(rdev->bdev, b);
2033 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2034 atomic_inc(&rdev->read_errors);
2035 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2038 "md/raid:%s: read error on replacement device "
2039 "(sector %llu on %s).\n",
2040 mdname(conf->mddev),
2041 (unsigned long long)s,
2043 else if (conf->mddev->degraded >= conf->max_degraded) {
2047 "md/raid:%s: read error not correctable "
2048 "(sector %llu on %s).\n",
2049 mdname(conf->mddev),
2050 (unsigned long long)s,
2052 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2057 "md/raid:%s: read error NOT corrected!! "
2058 "(sector %llu on %s).\n",
2059 mdname(conf->mddev),
2060 (unsigned long long)s,
2062 } else if (atomic_read(&rdev->read_errors)
2063 > conf->max_nr_stripes)
2065 "md/raid:%s: Too many read errors, failing device %s.\n",
2066 mdname(conf->mddev), bdn);
2069 if (set_bad && test_bit(In_sync, &rdev->flags)
2070 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2073 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2074 set_bit(R5_ReadError, &sh->dev[i].flags);
2075 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2077 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2079 clear_bit(R5_ReadError, &sh->dev[i].flags);
2080 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2082 && test_bit(In_sync, &rdev->flags)
2083 && rdev_set_badblocks(
2084 rdev, sh->sector, STRIPE_SECTORS, 0)))
2085 md_error(conf->mddev, rdev);
2088 rdev_dec_pending(rdev, conf->mddev);
2089 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2090 set_bit(STRIPE_HANDLE, &sh->state);
2094 static void raid5_end_write_request(struct bio *bi, int error)
2096 struct stripe_head *sh = bi->bi_private;
2097 struct r5conf *conf = sh->raid_conf;
2098 int disks = sh->disks, i;
2099 struct md_rdev *uninitialized_var(rdev);
2100 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2103 int replacement = 0;
2105 for (i = 0 ; i < disks; i++) {
2106 if (bi == &sh->dev[i].req) {
2107 rdev = conf->disks[i].rdev;
2110 if (bi == &sh->dev[i].rreq) {
2111 rdev = conf->disks[i].replacement;
2115 /* rdev was removed and 'replacement'
2116 * replaced it. rdev is not removed
2117 * until all requests are finished.
2119 rdev = conf->disks[i].rdev;
2123 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2124 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2133 md_error(conf->mddev, rdev);
2134 else if (is_badblock(rdev, sh->sector,
2136 &first_bad, &bad_sectors))
2137 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2140 set_bit(STRIPE_DEGRADED, &sh->state);
2141 set_bit(WriteErrorSeen, &rdev->flags);
2142 set_bit(R5_WriteError, &sh->dev[i].flags);
2143 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2144 set_bit(MD_RECOVERY_NEEDED,
2145 &rdev->mddev->recovery);
2146 } else if (is_badblock(rdev, sh->sector,
2148 &first_bad, &bad_sectors)) {
2149 set_bit(R5_MadeGood, &sh->dev[i].flags);
2150 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2151 /* That was a successful write so make
2152 * sure it looks like we already did
2155 set_bit(R5_ReWrite, &sh->dev[i].flags);
2158 rdev_dec_pending(rdev, conf->mddev);
2160 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2161 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2162 set_bit(STRIPE_HANDLE, &sh->state);
2166 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2168 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2170 struct r5dev *dev = &sh->dev[i];
2172 bio_init(&dev->req);
2173 dev->req.bi_io_vec = &dev->vec;
2174 dev->req.bi_max_vecs = 1;
2175 dev->req.bi_private = sh;
2177 bio_init(&dev->rreq);
2178 dev->rreq.bi_io_vec = &dev->rvec;
2179 dev->rreq.bi_max_vecs = 1;
2180 dev->rreq.bi_private = sh;
2183 dev->sector = compute_blocknr(sh, i, previous);
2186 static void error(struct mddev *mddev, struct md_rdev *rdev)
2188 char b[BDEVNAME_SIZE];
2189 struct r5conf *conf = mddev->private;
2190 unsigned long flags;
2191 pr_debug("raid456: error called\n");
2193 spin_lock_irqsave(&conf->device_lock, flags);
2194 clear_bit(In_sync, &rdev->flags);
2195 mddev->degraded = calc_degraded(conf);
2196 spin_unlock_irqrestore(&conf->device_lock, flags);
2197 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2199 set_bit(Blocked, &rdev->flags);
2200 set_bit(Faulty, &rdev->flags);
2201 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2203 "md/raid:%s: Disk failure on %s, disabling device.\n"
2204 "md/raid:%s: Operation continuing on %d devices.\n",
2206 bdevname(rdev->bdev, b),
2208 conf->raid_disks - mddev->degraded);
2212 * Input: a 'big' sector number,
2213 * Output: index of the data and parity disk, and the sector # in them.
2215 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2216 int previous, int *dd_idx,
2217 struct stripe_head *sh)
2219 sector_t stripe, stripe2;
2220 sector_t chunk_number;
2221 unsigned int chunk_offset;
2224 sector_t new_sector;
2225 int algorithm = previous ? conf->prev_algo
2227 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2228 : conf->chunk_sectors;
2229 int raid_disks = previous ? conf->previous_raid_disks
2231 int data_disks = raid_disks - conf->max_degraded;
2233 /* First compute the information on this sector */
2236 * Compute the chunk number and the sector offset inside the chunk
2238 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2239 chunk_number = r_sector;
2242 * Compute the stripe number
2244 stripe = chunk_number;
2245 *dd_idx = sector_div(stripe, data_disks);
2248 * Select the parity disk based on the user selected algorithm.
2250 pd_idx = qd_idx = -1;
2251 switch(conf->level) {
2253 pd_idx = data_disks;
2256 switch (algorithm) {
2257 case ALGORITHM_LEFT_ASYMMETRIC:
2258 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2259 if (*dd_idx >= pd_idx)
2262 case ALGORITHM_RIGHT_ASYMMETRIC:
2263 pd_idx = sector_div(stripe2, raid_disks);
2264 if (*dd_idx >= pd_idx)
2267 case ALGORITHM_LEFT_SYMMETRIC:
2268 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2269 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2271 case ALGORITHM_RIGHT_SYMMETRIC:
2272 pd_idx = sector_div(stripe2, raid_disks);
2273 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2275 case ALGORITHM_PARITY_0:
2279 case ALGORITHM_PARITY_N:
2280 pd_idx = data_disks;
2288 switch (algorithm) {
2289 case ALGORITHM_LEFT_ASYMMETRIC:
2290 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2291 qd_idx = pd_idx + 1;
2292 if (pd_idx == raid_disks-1) {
2293 (*dd_idx)++; /* Q D D D P */
2295 } else if (*dd_idx >= pd_idx)
2296 (*dd_idx) += 2; /* D D P Q D */
2298 case ALGORITHM_RIGHT_ASYMMETRIC:
2299 pd_idx = sector_div(stripe2, raid_disks);
2300 qd_idx = pd_idx + 1;
2301 if (pd_idx == raid_disks-1) {
2302 (*dd_idx)++; /* Q D D D P */
2304 } else if (*dd_idx >= pd_idx)
2305 (*dd_idx) += 2; /* D D P Q D */
2307 case ALGORITHM_LEFT_SYMMETRIC:
2308 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2309 qd_idx = (pd_idx + 1) % raid_disks;
2310 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2312 case ALGORITHM_RIGHT_SYMMETRIC:
2313 pd_idx = sector_div(stripe2, raid_disks);
2314 qd_idx = (pd_idx + 1) % raid_disks;
2315 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2318 case ALGORITHM_PARITY_0:
2323 case ALGORITHM_PARITY_N:
2324 pd_idx = data_disks;
2325 qd_idx = data_disks + 1;
2328 case ALGORITHM_ROTATING_ZERO_RESTART:
2329 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2330 * of blocks for computing Q is different.
2332 pd_idx = sector_div(stripe2, raid_disks);
2333 qd_idx = pd_idx + 1;
2334 if (pd_idx == raid_disks-1) {
2335 (*dd_idx)++; /* Q D D D P */
2337 } else if (*dd_idx >= pd_idx)
2338 (*dd_idx) += 2; /* D D P Q D */
2342 case ALGORITHM_ROTATING_N_RESTART:
2343 /* Same a left_asymmetric, by first stripe is
2344 * D D D P Q rather than
2348 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2349 qd_idx = pd_idx + 1;
2350 if (pd_idx == raid_disks-1) {
2351 (*dd_idx)++; /* Q D D D P */
2353 } else if (*dd_idx >= pd_idx)
2354 (*dd_idx) += 2; /* D D P Q D */
2358 case ALGORITHM_ROTATING_N_CONTINUE:
2359 /* Same as left_symmetric but Q is before P */
2360 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2361 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2362 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2366 case ALGORITHM_LEFT_ASYMMETRIC_6:
2367 /* RAID5 left_asymmetric, with Q on last device */
2368 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2369 if (*dd_idx >= pd_idx)
2371 qd_idx = raid_disks - 1;
2374 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2375 pd_idx = sector_div(stripe2, raid_disks-1);
2376 if (*dd_idx >= pd_idx)
2378 qd_idx = raid_disks - 1;
2381 case ALGORITHM_LEFT_SYMMETRIC_6:
2382 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2383 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2384 qd_idx = raid_disks - 1;
2387 case ALGORITHM_RIGHT_SYMMETRIC_6:
2388 pd_idx = sector_div(stripe2, raid_disks-1);
2389 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2390 qd_idx = raid_disks - 1;
2393 case ALGORITHM_PARITY_0_6:
2396 qd_idx = raid_disks - 1;
2406 sh->pd_idx = pd_idx;
2407 sh->qd_idx = qd_idx;
2408 sh->ddf_layout = ddf_layout;
2411 * Finally, compute the new sector number
2413 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2417 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2419 struct r5conf *conf = sh->raid_conf;
2420 int raid_disks = sh->disks;
2421 int data_disks = raid_disks - conf->max_degraded;
2422 sector_t new_sector = sh->sector, check;
2423 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2424 : conf->chunk_sectors;
2425 int algorithm = previous ? conf->prev_algo
2429 sector_t chunk_number;
2430 int dummy1, dd_idx = i;
2432 struct stripe_head sh2;
2434 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2435 stripe = new_sector;
2437 if (i == sh->pd_idx)
2439 switch(conf->level) {
2442 switch (algorithm) {
2443 case ALGORITHM_LEFT_ASYMMETRIC:
2444 case ALGORITHM_RIGHT_ASYMMETRIC:
2448 case ALGORITHM_LEFT_SYMMETRIC:
2449 case ALGORITHM_RIGHT_SYMMETRIC:
2452 i -= (sh->pd_idx + 1);
2454 case ALGORITHM_PARITY_0:
2457 case ALGORITHM_PARITY_N:
2464 if (i == sh->qd_idx)
2465 return 0; /* It is the Q disk */
2466 switch (algorithm) {
2467 case ALGORITHM_LEFT_ASYMMETRIC:
2468 case ALGORITHM_RIGHT_ASYMMETRIC:
2469 case ALGORITHM_ROTATING_ZERO_RESTART:
2470 case ALGORITHM_ROTATING_N_RESTART:
2471 if (sh->pd_idx == raid_disks-1)
2472 i--; /* Q D D D P */
2473 else if (i > sh->pd_idx)
2474 i -= 2; /* D D P Q D */
2476 case ALGORITHM_LEFT_SYMMETRIC:
2477 case ALGORITHM_RIGHT_SYMMETRIC:
2478 if (sh->pd_idx == raid_disks-1)
2479 i--; /* Q D D D P */
2484 i -= (sh->pd_idx + 2);
2487 case ALGORITHM_PARITY_0:
2490 case ALGORITHM_PARITY_N:
2492 case ALGORITHM_ROTATING_N_CONTINUE:
2493 /* Like left_symmetric, but P is before Q */
2494 if (sh->pd_idx == 0)
2495 i--; /* P D D D Q */
2500 i -= (sh->pd_idx + 1);
2503 case ALGORITHM_LEFT_ASYMMETRIC_6:
2504 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2508 case ALGORITHM_LEFT_SYMMETRIC_6:
2509 case ALGORITHM_RIGHT_SYMMETRIC_6:
2511 i += data_disks + 1;
2512 i -= (sh->pd_idx + 1);
2514 case ALGORITHM_PARITY_0_6:
2523 chunk_number = stripe * data_disks + i;
2524 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2526 check = raid5_compute_sector(conf, r_sector,
2527 previous, &dummy1, &sh2);
2528 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2529 || sh2.qd_idx != sh->qd_idx) {
2530 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2531 mdname(conf->mddev));
2538 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2539 int rcw, int expand)
2541 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2542 struct r5conf *conf = sh->raid_conf;
2543 int level = conf->level;
2547 for (i = disks; i--; ) {
2548 struct r5dev *dev = &sh->dev[i];
2551 set_bit(R5_LOCKED, &dev->flags);
2552 set_bit(R5_Wantdrain, &dev->flags);
2554 clear_bit(R5_UPTODATE, &dev->flags);
2558 /* if we are not expanding this is a proper write request, and
2559 * there will be bios with new data to be drained into the
2564 /* False alarm, nothing to do */
2566 sh->reconstruct_state = reconstruct_state_drain_run;
2567 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2569 sh->reconstruct_state = reconstruct_state_run;
2571 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2573 if (s->locked + conf->max_degraded == disks)
2574 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2575 atomic_inc(&conf->pending_full_writes);
2578 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2579 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2581 for (i = disks; i--; ) {
2582 struct r5dev *dev = &sh->dev[i];
2587 (test_bit(R5_UPTODATE, &dev->flags) ||
2588 test_bit(R5_Wantcompute, &dev->flags))) {
2589 set_bit(R5_Wantdrain, &dev->flags);
2590 set_bit(R5_LOCKED, &dev->flags);
2591 clear_bit(R5_UPTODATE, &dev->flags);
2596 /* False alarm - nothing to do */
2598 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2599 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2600 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2601 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2604 /* keep the parity disk(s) locked while asynchronous operations
2607 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2608 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2612 int qd_idx = sh->qd_idx;
2613 struct r5dev *dev = &sh->dev[qd_idx];
2615 set_bit(R5_LOCKED, &dev->flags);
2616 clear_bit(R5_UPTODATE, &dev->flags);
2620 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2621 __func__, (unsigned long long)sh->sector,
2622 s->locked, s->ops_request);
2626 * Each stripe/dev can have one or more bion attached.
2627 * toread/towrite point to the first in a chain.
2628 * The bi_next chain must be in order.
2630 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2633 struct r5conf *conf = sh->raid_conf;
2636 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2637 (unsigned long long)bi->bi_iter.bi_sector,
2638 (unsigned long long)sh->sector);
2641 * If several bio share a stripe. The bio bi_phys_segments acts as a
2642 * reference count to avoid race. The reference count should already be
2643 * increased before this function is called (for example, in
2644 * make_request()), so other bio sharing this stripe will not free the
2645 * stripe. If a stripe is owned by one stripe, the stripe lock will
2648 spin_lock_irq(&sh->stripe_lock);
2650 bip = &sh->dev[dd_idx].towrite;
2654 bip = &sh->dev[dd_idx].toread;
2655 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2656 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2658 bip = & (*bip)->bi_next;
2660 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2663 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2667 raid5_inc_bi_active_stripes(bi);
2670 /* check if page is covered */
2671 sector_t sector = sh->dev[dd_idx].sector;
2672 for (bi=sh->dev[dd_idx].towrite;
2673 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2674 bi && bi->bi_iter.bi_sector <= sector;
2675 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2676 if (bio_end_sector(bi) >= sector)
2677 sector = bio_end_sector(bi);
2679 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2680 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2683 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2684 (unsigned long long)(*bip)->bi_iter.bi_sector,
2685 (unsigned long long)sh->sector, dd_idx);
2686 spin_unlock_irq(&sh->stripe_lock);
2688 if (conf->mddev->bitmap && firstwrite) {
2689 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2691 sh->bm_seq = conf->seq_flush+1;
2692 set_bit(STRIPE_BIT_DELAY, &sh->state);
2697 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2698 spin_unlock_irq(&sh->stripe_lock);
2702 static void end_reshape(struct r5conf *conf);
2704 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2705 struct stripe_head *sh)
2707 int sectors_per_chunk =
2708 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2710 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2711 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2713 raid5_compute_sector(conf,
2714 stripe * (disks - conf->max_degraded)
2715 *sectors_per_chunk + chunk_offset,
2721 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2722 struct stripe_head_state *s, int disks,
2723 struct bio **return_bi)
2726 for (i = disks; i--; ) {
2730 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2731 struct md_rdev *rdev;
2733 rdev = rcu_dereference(conf->disks[i].rdev);
2734 if (rdev && test_bit(In_sync, &rdev->flags))
2735 atomic_inc(&rdev->nr_pending);
2740 if (!rdev_set_badblocks(
2744 md_error(conf->mddev, rdev);
2745 rdev_dec_pending(rdev, conf->mddev);
2748 spin_lock_irq(&sh->stripe_lock);
2749 /* fail all writes first */
2750 bi = sh->dev[i].towrite;
2751 sh->dev[i].towrite = NULL;
2752 spin_unlock_irq(&sh->stripe_lock);
2756 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2757 wake_up(&conf->wait_for_overlap);
2759 while (bi && bi->bi_iter.bi_sector <
2760 sh->dev[i].sector + STRIPE_SECTORS) {
2761 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2762 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2763 if (!raid5_dec_bi_active_stripes(bi)) {
2764 md_write_end(conf->mddev);
2765 bi->bi_next = *return_bi;
2771 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2772 STRIPE_SECTORS, 0, 0);
2774 /* and fail all 'written' */
2775 bi = sh->dev[i].written;
2776 sh->dev[i].written = NULL;
2777 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
2778 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
2779 sh->dev[i].page = sh->dev[i].orig_page;
2782 if (bi) bitmap_end = 1;
2783 while (bi && bi->bi_iter.bi_sector <
2784 sh->dev[i].sector + STRIPE_SECTORS) {
2785 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2786 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2787 if (!raid5_dec_bi_active_stripes(bi)) {
2788 md_write_end(conf->mddev);
2789 bi->bi_next = *return_bi;
2795 /* fail any reads if this device is non-operational and
2796 * the data has not reached the cache yet.
2798 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2799 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2800 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2801 spin_lock_irq(&sh->stripe_lock);
2802 bi = sh->dev[i].toread;
2803 sh->dev[i].toread = NULL;
2804 spin_unlock_irq(&sh->stripe_lock);
2805 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2806 wake_up(&conf->wait_for_overlap);
2807 while (bi && bi->bi_iter.bi_sector <
2808 sh->dev[i].sector + STRIPE_SECTORS) {
2809 struct bio *nextbi =
2810 r5_next_bio(bi, sh->dev[i].sector);
2811 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2812 if (!raid5_dec_bi_active_stripes(bi)) {
2813 bi->bi_next = *return_bi;
2820 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2821 STRIPE_SECTORS, 0, 0);
2822 /* If we were in the middle of a write the parity block might
2823 * still be locked - so just clear all R5_LOCKED flags
2825 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2828 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2829 if (atomic_dec_and_test(&conf->pending_full_writes))
2830 md_wakeup_thread(conf->mddev->thread);
2834 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2835 struct stripe_head_state *s)
2840 clear_bit(STRIPE_SYNCING, &sh->state);
2841 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2842 wake_up(&conf->wait_for_overlap);
2845 /* There is nothing more to do for sync/check/repair.
2846 * Don't even need to abort as that is handled elsewhere
2847 * if needed, and not always wanted e.g. if there is a known
2849 * For recover/replace we need to record a bad block on all
2850 * non-sync devices, or abort the recovery
2852 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2853 /* During recovery devices cannot be removed, so
2854 * locking and refcounting of rdevs is not needed
2856 for (i = 0; i < conf->raid_disks; i++) {
2857 struct md_rdev *rdev = conf->disks[i].rdev;
2859 && !test_bit(Faulty, &rdev->flags)
2860 && !test_bit(In_sync, &rdev->flags)
2861 && !rdev_set_badblocks(rdev, sh->sector,
2864 rdev = conf->disks[i].replacement;
2866 && !test_bit(Faulty, &rdev->flags)
2867 && !test_bit(In_sync, &rdev->flags)
2868 && !rdev_set_badblocks(rdev, sh->sector,
2873 conf->recovery_disabled =
2874 conf->mddev->recovery_disabled;
2876 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2879 static int want_replace(struct stripe_head *sh, int disk_idx)
2881 struct md_rdev *rdev;
2883 /* Doing recovery so rcu locking not required */
2884 rdev = sh->raid_conf->disks[disk_idx].replacement;
2886 && !test_bit(Faulty, &rdev->flags)
2887 && !test_bit(In_sync, &rdev->flags)
2888 && (rdev->recovery_offset <= sh->sector
2889 || rdev->mddev->recovery_cp <= sh->sector))
2895 /* fetch_block - checks the given member device to see if its data needs
2896 * to be read or computed to satisfy a request.
2898 * Returns 1 when no more member devices need to be checked, otherwise returns
2899 * 0 to tell the loop in handle_stripe_fill to continue
2901 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2902 int disk_idx, int disks)
2904 struct r5dev *dev = &sh->dev[disk_idx];
2905 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2906 &sh->dev[s->failed_num[1]] };
2908 /* is the data in this block needed, and can we get it? */
2909 if (!test_bit(R5_LOCKED, &dev->flags) &&
2910 !test_bit(R5_UPTODATE, &dev->flags) &&
2912 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2913 s->syncing || s->expanding ||
2914 (s->replacing && want_replace(sh, disk_idx)) ||
2915 (s->failed >= 1 && fdev[0]->toread) ||
2916 (s->failed >= 2 && fdev[1]->toread) ||
2917 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2918 (!test_bit(R5_Insync, &dev->flags) || test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) &&
2919 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2920 ((sh->raid_conf->level == 6 ||
2921 sh->sector >= sh->raid_conf->mddev->recovery_cp)
2922 && s->failed && s->to_write &&
2923 (s->to_write - s->non_overwrite <
2924 sh->raid_conf->raid_disks - sh->raid_conf->max_degraded) &&
2925 (!test_bit(R5_Insync, &dev->flags) || test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))))) {
2926 /* we would like to get this block, possibly by computing it,
2927 * otherwise read it if the backing disk is insync
2929 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2930 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2931 if ((s->uptodate == disks - 1) &&
2932 (s->failed && (disk_idx == s->failed_num[0] ||
2933 disk_idx == s->failed_num[1]))) {
2934 /* have disk failed, and we're requested to fetch it;
2937 pr_debug("Computing stripe %llu block %d\n",
2938 (unsigned long long)sh->sector, disk_idx);
2939 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2940 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2941 set_bit(R5_Wantcompute, &dev->flags);
2942 sh->ops.target = disk_idx;
2943 sh->ops.target2 = -1; /* no 2nd target */
2945 /* Careful: from this point on 'uptodate' is in the eye
2946 * of raid_run_ops which services 'compute' operations
2947 * before writes. R5_Wantcompute flags a block that will
2948 * be R5_UPTODATE by the time it is needed for a
2949 * subsequent operation.
2953 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2954 /* Computing 2-failure is *very* expensive; only
2955 * do it if failed >= 2
2958 for (other = disks; other--; ) {
2959 if (other == disk_idx)
2961 if (!test_bit(R5_UPTODATE,
2962 &sh->dev[other].flags))
2966 pr_debug("Computing stripe %llu blocks %d,%d\n",
2967 (unsigned long long)sh->sector,
2969 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2970 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2971 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2972 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2973 sh->ops.target = disk_idx;
2974 sh->ops.target2 = other;
2978 } else if (test_bit(R5_Insync, &dev->flags)) {
2979 set_bit(R5_LOCKED, &dev->flags);
2980 set_bit(R5_Wantread, &dev->flags);
2982 pr_debug("Reading block %d (sync=%d)\n",
2983 disk_idx, s->syncing);
2991 * handle_stripe_fill - read or compute data to satisfy pending requests.
2993 static void handle_stripe_fill(struct stripe_head *sh,
2994 struct stripe_head_state *s,
2999 /* look for blocks to read/compute, skip this if a compute
3000 * is already in flight, or if the stripe contents are in the
3001 * midst of changing due to a write
3003 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3004 !sh->reconstruct_state)
3005 for (i = disks; i--; )
3006 if (fetch_block(sh, s, i, disks))
3008 set_bit(STRIPE_HANDLE, &sh->state);
3011 /* handle_stripe_clean_event
3012 * any written block on an uptodate or failed drive can be returned.
3013 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3014 * never LOCKED, so we don't need to test 'failed' directly.
3016 static void handle_stripe_clean_event(struct r5conf *conf,
3017 struct stripe_head *sh, int disks, struct bio **return_bi)
3021 int discard_pending = 0;
3023 for (i = disks; i--; )
3024 if (sh->dev[i].written) {
3026 if (!test_bit(R5_LOCKED, &dev->flags) &&
3027 (test_bit(R5_UPTODATE, &dev->flags) ||
3028 test_bit(R5_Discard, &dev->flags) ||
3029 test_bit(R5_SkipCopy, &dev->flags))) {
3030 /* We can return any write requests */
3031 struct bio *wbi, *wbi2;
3032 pr_debug("Return write for disc %d\n", i);
3033 if (test_and_clear_bit(R5_Discard, &dev->flags))
3034 clear_bit(R5_UPTODATE, &dev->flags);
3035 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3036 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3037 dev->page = dev->orig_page;
3040 dev->written = NULL;
3041 while (wbi && wbi->bi_iter.bi_sector <
3042 dev->sector + STRIPE_SECTORS) {
3043 wbi2 = r5_next_bio(wbi, dev->sector);
3044 if (!raid5_dec_bi_active_stripes(wbi)) {
3045 md_write_end(conf->mddev);
3046 wbi->bi_next = *return_bi;
3051 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3053 !test_bit(STRIPE_DEGRADED, &sh->state),
3055 } else if (test_bit(R5_Discard, &dev->flags))
3056 discard_pending = 1;
3057 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3058 WARN_ON(dev->page != dev->orig_page);
3060 if (!discard_pending &&
3061 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3062 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3063 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3064 if (sh->qd_idx >= 0) {
3065 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3066 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3068 /* now that discard is done we can proceed with any sync */
3069 clear_bit(STRIPE_DISCARD, &sh->state);
3071 * SCSI discard will change some bio fields and the stripe has
3072 * no updated data, so remove it from hash list and the stripe
3073 * will be reinitialized
3075 spin_lock_irq(&conf->device_lock);
3077 spin_unlock_irq(&conf->device_lock);
3078 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3079 set_bit(STRIPE_HANDLE, &sh->state);
3083 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3084 if (atomic_dec_and_test(&conf->pending_full_writes))
3085 md_wakeup_thread(conf->mddev->thread);
3088 static void handle_stripe_dirtying(struct r5conf *conf,
3089 struct stripe_head *sh,
3090 struct stripe_head_state *s,
3093 int rmw = 0, rcw = 0, i;
3094 sector_t recovery_cp = conf->mddev->recovery_cp;
3096 /* RAID6 requires 'rcw' in current implementation.
3097 * Otherwise, check whether resync is now happening or should start.
3098 * If yes, then the array is dirty (after unclean shutdown or
3099 * initial creation), so parity in some stripes might be inconsistent.
3100 * In this case, we need to always do reconstruct-write, to ensure
3101 * that in case of drive failure or read-error correction, we
3102 * generate correct data from the parity.
3104 if (conf->max_degraded == 2 ||
3105 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
3106 /* Calculate the real rcw later - for now make it
3107 * look like rcw is cheaper
3110 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
3111 conf->max_degraded, (unsigned long long)recovery_cp,
3112 (unsigned long long)sh->sector);
3113 } else for (i = disks; i--; ) {
3114 /* would I have to read this buffer for read_modify_write */
3115 struct r5dev *dev = &sh->dev[i];
3116 if ((dev->towrite || i == sh->pd_idx) &&
3117 !test_bit(R5_LOCKED, &dev->flags) &&
3118 !(test_bit(R5_UPTODATE, &dev->flags) ||
3119 test_bit(R5_Wantcompute, &dev->flags))) {
3120 if (test_bit(R5_Insync, &dev->flags))
3123 rmw += 2*disks; /* cannot read it */
3125 /* Would I have to read this buffer for reconstruct_write */
3126 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
3127 !test_bit(R5_LOCKED, &dev->flags) &&
3128 !(test_bit(R5_UPTODATE, &dev->flags) ||
3129 test_bit(R5_Wantcompute, &dev->flags))) {
3130 if (test_bit(R5_Insync, &dev->flags))
3136 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3137 (unsigned long long)sh->sector, rmw, rcw);
3138 set_bit(STRIPE_HANDLE, &sh->state);
3139 if (rmw < rcw && rmw > 0) {
3140 /* prefer read-modify-write, but need to get some data */
3141 if (conf->mddev->queue)
3142 blk_add_trace_msg(conf->mddev->queue,
3143 "raid5 rmw %llu %d",
3144 (unsigned long long)sh->sector, rmw);
3145 for (i = disks; i--; ) {
3146 struct r5dev *dev = &sh->dev[i];
3147 if ((dev->towrite || i == sh->pd_idx) &&
3148 !test_bit(R5_LOCKED, &dev->flags) &&
3149 !(test_bit(R5_UPTODATE, &dev->flags) ||
3150 test_bit(R5_Wantcompute, &dev->flags)) &&
3151 test_bit(R5_Insync, &dev->flags)) {
3152 if (test_bit(STRIPE_PREREAD_ACTIVE,
3154 pr_debug("Read_old block %d for r-m-w\n",
3156 set_bit(R5_LOCKED, &dev->flags);
3157 set_bit(R5_Wantread, &dev->flags);
3160 set_bit(STRIPE_DELAYED, &sh->state);
3161 set_bit(STRIPE_HANDLE, &sh->state);
3166 if (rcw <= rmw && rcw > 0) {
3167 /* want reconstruct write, but need to get some data */
3170 for (i = disks; i--; ) {
3171 struct r5dev *dev = &sh->dev[i];
3172 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3173 i != sh->pd_idx && i != sh->qd_idx &&
3174 !test_bit(R5_LOCKED, &dev->flags) &&
3175 !(test_bit(R5_UPTODATE, &dev->flags) ||
3176 test_bit(R5_Wantcompute, &dev->flags))) {
3178 if (test_bit(R5_Insync, &dev->flags) &&
3179 test_bit(STRIPE_PREREAD_ACTIVE,
3181 pr_debug("Read_old block "
3182 "%d for Reconstruct\n", i);
3183 set_bit(R5_LOCKED, &dev->flags);
3184 set_bit(R5_Wantread, &dev->flags);
3188 set_bit(STRIPE_DELAYED, &sh->state);
3189 set_bit(STRIPE_HANDLE, &sh->state);
3193 if (rcw && conf->mddev->queue)
3194 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3195 (unsigned long long)sh->sector,
3196 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3198 /* now if nothing is locked, and if we have enough data,
3199 * we can start a write request
3201 /* since handle_stripe can be called at any time we need to handle the
3202 * case where a compute block operation has been submitted and then a
3203 * subsequent call wants to start a write request. raid_run_ops only
3204 * handles the case where compute block and reconstruct are requested
3205 * simultaneously. If this is not the case then new writes need to be
3206 * held off until the compute completes.
3208 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3209 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3210 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3211 schedule_reconstruction(sh, s, rcw == 0, 0);
3214 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3215 struct stripe_head_state *s, int disks)
3217 struct r5dev *dev = NULL;
3219 set_bit(STRIPE_HANDLE, &sh->state);
3221 switch (sh->check_state) {
3222 case check_state_idle:
3223 /* start a new check operation if there are no failures */
3224 if (s->failed == 0) {
3225 BUG_ON(s->uptodate != disks);
3226 sh->check_state = check_state_run;
3227 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3228 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3232 dev = &sh->dev[s->failed_num[0]];
3234 case check_state_compute_result:
3235 sh->check_state = check_state_idle;
3237 dev = &sh->dev[sh->pd_idx];
3239 /* check that a write has not made the stripe insync */
3240 if (test_bit(STRIPE_INSYNC, &sh->state))
3243 /* either failed parity check, or recovery is happening */
3244 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3245 BUG_ON(s->uptodate != disks);
3247 set_bit(R5_LOCKED, &dev->flags);
3249 set_bit(R5_Wantwrite, &dev->flags);
3251 clear_bit(STRIPE_DEGRADED, &sh->state);
3252 set_bit(STRIPE_INSYNC, &sh->state);
3254 case check_state_run:
3255 break; /* we will be called again upon completion */
3256 case check_state_check_result:
3257 sh->check_state = check_state_idle;
3259 /* if a failure occurred during the check operation, leave
3260 * STRIPE_INSYNC not set and let the stripe be handled again
3265 /* handle a successful check operation, if parity is correct
3266 * we are done. Otherwise update the mismatch count and repair
3267 * parity if !MD_RECOVERY_CHECK
3269 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3270 /* parity is correct (on disc,
3271 * not in buffer any more)
3273 set_bit(STRIPE_INSYNC, &sh->state);
3275 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3276 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3277 /* don't try to repair!! */
3278 set_bit(STRIPE_INSYNC, &sh->state);
3280 sh->check_state = check_state_compute_run;
3281 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3282 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3283 set_bit(R5_Wantcompute,
3284 &sh->dev[sh->pd_idx].flags);
3285 sh->ops.target = sh->pd_idx;
3286 sh->ops.target2 = -1;
3291 case check_state_compute_run:
3294 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3295 __func__, sh->check_state,
3296 (unsigned long long) sh->sector);
3301 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3302 struct stripe_head_state *s,
3305 int pd_idx = sh->pd_idx;
3306 int qd_idx = sh->qd_idx;
3309 set_bit(STRIPE_HANDLE, &sh->state);
3311 BUG_ON(s->failed > 2);
3313 /* Want to check and possibly repair P and Q.
3314 * However there could be one 'failed' device, in which
3315 * case we can only check one of them, possibly using the
3316 * other to generate missing data
3319 switch (sh->check_state) {
3320 case check_state_idle:
3321 /* start a new check operation if there are < 2 failures */
3322 if (s->failed == s->q_failed) {
3323 /* The only possible failed device holds Q, so it
3324 * makes sense to check P (If anything else were failed,
3325 * we would have used P to recreate it).
3327 sh->check_state = check_state_run;
3329 if (!s->q_failed && s->failed < 2) {
3330 /* Q is not failed, and we didn't use it to generate
3331 * anything, so it makes sense to check it
3333 if (sh->check_state == check_state_run)
3334 sh->check_state = check_state_run_pq;
3336 sh->check_state = check_state_run_q;
3339 /* discard potentially stale zero_sum_result */
3340 sh->ops.zero_sum_result = 0;
3342 if (sh->check_state == check_state_run) {
3343 /* async_xor_zero_sum destroys the contents of P */
3344 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3347 if (sh->check_state >= check_state_run &&
3348 sh->check_state <= check_state_run_pq) {
3349 /* async_syndrome_zero_sum preserves P and Q, so
3350 * no need to mark them !uptodate here
3352 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3356 /* we have 2-disk failure */
3357 BUG_ON(s->failed != 2);
3359 case check_state_compute_result:
3360 sh->check_state = check_state_idle;
3362 /* check that a write has not made the stripe insync */
3363 if (test_bit(STRIPE_INSYNC, &sh->state))
3366 /* now write out any block on a failed drive,
3367 * or P or Q if they were recomputed
3369 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3370 if (s->failed == 2) {
3371 dev = &sh->dev[s->failed_num[1]];
3373 set_bit(R5_LOCKED, &dev->flags);
3374 set_bit(R5_Wantwrite, &dev->flags);
3376 if (s->failed >= 1) {
3377 dev = &sh->dev[s->failed_num[0]];
3379 set_bit(R5_LOCKED, &dev->flags);
3380 set_bit(R5_Wantwrite, &dev->flags);
3382 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3383 dev = &sh->dev[pd_idx];
3385 set_bit(R5_LOCKED, &dev->flags);
3386 set_bit(R5_Wantwrite, &dev->flags);
3388 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3389 dev = &sh->dev[qd_idx];
3391 set_bit(R5_LOCKED, &dev->flags);
3392 set_bit(R5_Wantwrite, &dev->flags);
3394 clear_bit(STRIPE_DEGRADED, &sh->state);
3396 set_bit(STRIPE_INSYNC, &sh->state);
3398 case check_state_run:
3399 case check_state_run_q:
3400 case check_state_run_pq:
3401 break; /* we will be called again upon completion */
3402 case check_state_check_result:
3403 sh->check_state = check_state_idle;
3405 /* handle a successful check operation, if parity is correct
3406 * we are done. Otherwise update the mismatch count and repair
3407 * parity if !MD_RECOVERY_CHECK
3409 if (sh->ops.zero_sum_result == 0) {
3410 /* both parities are correct */
3412 set_bit(STRIPE_INSYNC, &sh->state);
3414 /* in contrast to the raid5 case we can validate
3415 * parity, but still have a failure to write
3418 sh->check_state = check_state_compute_result;
3419 /* Returning at this point means that we may go
3420 * off and bring p and/or q uptodate again so
3421 * we make sure to check zero_sum_result again
3422 * to verify if p or q need writeback
3426 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3427 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3428 /* don't try to repair!! */
3429 set_bit(STRIPE_INSYNC, &sh->state);
3431 int *target = &sh->ops.target;
3433 sh->ops.target = -1;
3434 sh->ops.target2 = -1;
3435 sh->check_state = check_state_compute_run;
3436 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3437 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3438 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3439 set_bit(R5_Wantcompute,
3440 &sh->dev[pd_idx].flags);
3442 target = &sh->ops.target2;
3445 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3446 set_bit(R5_Wantcompute,
3447 &sh->dev[qd_idx].flags);
3454 case check_state_compute_run:
3457 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3458 __func__, sh->check_state,
3459 (unsigned long long) sh->sector);
3464 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3468 /* We have read all the blocks in this stripe and now we need to
3469 * copy some of them into a target stripe for expand.
3471 struct dma_async_tx_descriptor *tx = NULL;
3472 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3473 for (i = 0; i < sh->disks; i++)
3474 if (i != sh->pd_idx && i != sh->qd_idx) {
3476 struct stripe_head *sh2;
3477 struct async_submit_ctl submit;
3479 sector_t bn = compute_blocknr(sh, i, 1);
3480 sector_t s = raid5_compute_sector(conf, bn, 0,
3482 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3484 /* so far only the early blocks of this stripe
3485 * have been requested. When later blocks
3486 * get requested, we will try again
3489 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3490 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3491 /* must have already done this block */
3492 release_stripe(sh2);
3496 /* place all the copies on one channel */
3497 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3498 tx = async_memcpy(sh2->dev[dd_idx].page,
3499 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3502 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3503 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3504 for (j = 0; j < conf->raid_disks; j++)
3505 if (j != sh2->pd_idx &&
3507 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3509 if (j == conf->raid_disks) {
3510 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3511 set_bit(STRIPE_HANDLE, &sh2->state);
3513 release_stripe(sh2);
3516 /* done submitting copies, wait for them to complete */
3517 async_tx_quiesce(&tx);
3521 * handle_stripe - do things to a stripe.
3523 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3524 * state of various bits to see what needs to be done.
3526 * return some read requests which now have data
3527 * return some write requests which are safely on storage
3528 * schedule a read on some buffers
3529 * schedule a write of some buffers
3530 * return confirmation of parity correctness
3534 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3536 struct r5conf *conf = sh->raid_conf;
3537 int disks = sh->disks;
3540 int do_recovery = 0;
3542 memset(s, 0, sizeof(*s));
3544 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3545 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3546 s->failed_num[0] = -1;
3547 s->failed_num[1] = -1;
3549 /* Now to look around and see what can be done */
3551 for (i=disks; i--; ) {
3552 struct md_rdev *rdev;
3559 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3561 dev->toread, dev->towrite, dev->written);
3562 /* maybe we can reply to a read
3564 * new wantfill requests are only permitted while
3565 * ops_complete_biofill is guaranteed to be inactive
3567 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3568 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3569 set_bit(R5_Wantfill, &dev->flags);
3571 /* now count some things */
3572 if (test_bit(R5_LOCKED, &dev->flags))
3574 if (test_bit(R5_UPTODATE, &dev->flags))
3576 if (test_bit(R5_Wantcompute, &dev->flags)) {
3578 BUG_ON(s->compute > 2);
3581 if (test_bit(R5_Wantfill, &dev->flags))
3583 else if (dev->toread)
3587 if (!test_bit(R5_OVERWRITE, &dev->flags))
3592 /* Prefer to use the replacement for reads, but only
3593 * if it is recovered enough and has no bad blocks.
3595 rdev = rcu_dereference(conf->disks[i].replacement);
3596 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3597 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3598 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3599 &first_bad, &bad_sectors))
3600 set_bit(R5_ReadRepl, &dev->flags);
3603 set_bit(R5_NeedReplace, &dev->flags);
3604 rdev = rcu_dereference(conf->disks[i].rdev);
3605 clear_bit(R5_ReadRepl, &dev->flags);
3607 if (rdev && test_bit(Faulty, &rdev->flags))
3610 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3611 &first_bad, &bad_sectors);
3612 if (s->blocked_rdev == NULL
3613 && (test_bit(Blocked, &rdev->flags)
3616 set_bit(BlockedBadBlocks,
3618 s->blocked_rdev = rdev;
3619 atomic_inc(&rdev->nr_pending);
3622 clear_bit(R5_Insync, &dev->flags);
3626 /* also not in-sync */
3627 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3628 test_bit(R5_UPTODATE, &dev->flags)) {
3629 /* treat as in-sync, but with a read error
3630 * which we can now try to correct
3632 set_bit(R5_Insync, &dev->flags);
3633 set_bit(R5_ReadError, &dev->flags);
3635 } else if (test_bit(In_sync, &rdev->flags))
3636 set_bit(R5_Insync, &dev->flags);
3637 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3638 /* in sync if before recovery_offset */
3639 set_bit(R5_Insync, &dev->flags);
3640 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3641 test_bit(R5_Expanded, &dev->flags))
3642 /* If we've reshaped into here, we assume it is Insync.
3643 * We will shortly update recovery_offset to make
3646 set_bit(R5_Insync, &dev->flags);
3648 if (test_bit(R5_WriteError, &dev->flags)) {
3649 /* This flag does not apply to '.replacement'
3650 * only to .rdev, so make sure to check that*/
3651 struct md_rdev *rdev2 = rcu_dereference(
3652 conf->disks[i].rdev);
3654 clear_bit(R5_Insync, &dev->flags);
3655 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3656 s->handle_bad_blocks = 1;
3657 atomic_inc(&rdev2->nr_pending);
3659 clear_bit(R5_WriteError, &dev->flags);
3661 if (test_bit(R5_MadeGood, &dev->flags)) {
3662 /* This flag does not apply to '.replacement'
3663 * only to .rdev, so make sure to check that*/
3664 struct md_rdev *rdev2 = rcu_dereference(
3665 conf->disks[i].rdev);
3666 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3667 s->handle_bad_blocks = 1;
3668 atomic_inc(&rdev2->nr_pending);
3670 clear_bit(R5_MadeGood, &dev->flags);
3672 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3673 struct md_rdev *rdev2 = rcu_dereference(
3674 conf->disks[i].replacement);
3675 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3676 s->handle_bad_blocks = 1;
3677 atomic_inc(&rdev2->nr_pending);
3679 clear_bit(R5_MadeGoodRepl, &dev->flags);
3681 if (!test_bit(R5_Insync, &dev->flags)) {
3682 /* The ReadError flag will just be confusing now */
3683 clear_bit(R5_ReadError, &dev->flags);
3684 clear_bit(R5_ReWrite, &dev->flags);
3686 if (test_bit(R5_ReadError, &dev->flags))
3687 clear_bit(R5_Insync, &dev->flags);
3688 if (!test_bit(R5_Insync, &dev->flags)) {
3690 s->failed_num[s->failed] = i;
3692 if (rdev && !test_bit(Faulty, &rdev->flags))
3696 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3697 /* If there is a failed device being replaced,
3698 * we must be recovering.
3699 * else if we are after recovery_cp, we must be syncing
3700 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3701 * else we can only be replacing
3702 * sync and recovery both need to read all devices, and so
3703 * use the same flag.
3706 sh->sector >= conf->mddev->recovery_cp ||
3707 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3715 static void handle_stripe(struct stripe_head *sh)
3717 struct stripe_head_state s;
3718 struct r5conf *conf = sh->raid_conf;
3721 int disks = sh->disks;
3722 struct r5dev *pdev, *qdev;
3724 clear_bit(STRIPE_HANDLE, &sh->state);
3725 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3726 /* already being handled, ensure it gets handled
3727 * again when current action finishes */
3728 set_bit(STRIPE_HANDLE, &sh->state);
3732 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3733 spin_lock(&sh->stripe_lock);
3734 /* Cannot process 'sync' concurrently with 'discard' */
3735 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3736 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3737 set_bit(STRIPE_SYNCING, &sh->state);
3738 clear_bit(STRIPE_INSYNC, &sh->state);
3739 clear_bit(STRIPE_REPLACED, &sh->state);
3741 spin_unlock(&sh->stripe_lock);
3743 clear_bit(STRIPE_DELAYED, &sh->state);
3745 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3746 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3747 (unsigned long long)sh->sector, sh->state,
3748 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3749 sh->check_state, sh->reconstruct_state);
3751 analyse_stripe(sh, &s);
3753 if (s.handle_bad_blocks) {
3754 set_bit(STRIPE_HANDLE, &sh->state);
3758 if (unlikely(s.blocked_rdev)) {
3759 if (s.syncing || s.expanding || s.expanded ||
3760 s.replacing || s.to_write || s.written) {
3761 set_bit(STRIPE_HANDLE, &sh->state);
3764 /* There is nothing for the blocked_rdev to block */
3765 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3766 s.blocked_rdev = NULL;
3769 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3770 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3771 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3774 pr_debug("locked=%d uptodate=%d to_read=%d"
3775 " to_write=%d failed=%d failed_num=%d,%d\n",
3776 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3777 s.failed_num[0], s.failed_num[1]);
3778 /* check if the array has lost more than max_degraded devices and,
3779 * if so, some requests might need to be failed.
3781 if (s.failed > conf->max_degraded) {
3782 sh->check_state = 0;
3783 sh->reconstruct_state = 0;
3784 if (s.to_read+s.to_write+s.written)
3785 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3786 if (s.syncing + s.replacing)
3787 handle_failed_sync(conf, sh, &s);
3790 /* Now we check to see if any write operations have recently
3794 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3796 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3797 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3798 sh->reconstruct_state = reconstruct_state_idle;
3800 /* All the 'written' buffers and the parity block are ready to
3801 * be written back to disk
3803 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3804 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3805 BUG_ON(sh->qd_idx >= 0 &&
3806 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3807 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3808 for (i = disks; i--; ) {
3809 struct r5dev *dev = &sh->dev[i];
3810 if (test_bit(R5_LOCKED, &dev->flags) &&
3811 (i == sh->pd_idx || i == sh->qd_idx ||
3813 pr_debug("Writing block %d\n", i);
3814 set_bit(R5_Wantwrite, &dev->flags);
3819 if (!test_bit(R5_Insync, &dev->flags) ||
3820 ((i == sh->pd_idx || i == sh->qd_idx) &&
3822 set_bit(STRIPE_INSYNC, &sh->state);
3825 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3826 s.dec_preread_active = 1;
3830 * might be able to return some write requests if the parity blocks
3831 * are safe, or on a failed drive
3833 pdev = &sh->dev[sh->pd_idx];
3834 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3835 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3836 qdev = &sh->dev[sh->qd_idx];
3837 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3838 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3842 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3843 && !test_bit(R5_LOCKED, &pdev->flags)
3844 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3845 test_bit(R5_Discard, &pdev->flags))))) &&
3846 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3847 && !test_bit(R5_LOCKED, &qdev->flags)
3848 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3849 test_bit(R5_Discard, &qdev->flags))))))
3850 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3852 /* Now we might consider reading some blocks, either to check/generate
3853 * parity, or to satisfy requests
3854 * or to load a block that is being partially written.
3856 if (s.to_read || s.non_overwrite
3857 || (conf->level == 6 && s.to_write && s.failed)
3858 || (s.syncing && (s.uptodate + s.compute < disks))
3861 handle_stripe_fill(sh, &s, disks);
3863 /* Now to consider new write requests and what else, if anything
3864 * should be read. We do not handle new writes when:
3865 * 1/ A 'write' operation (copy+xor) is already in flight.
3866 * 2/ A 'check' operation is in flight, as it may clobber the parity
3869 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3870 handle_stripe_dirtying(conf, sh, &s, disks);
3872 /* maybe we need to check and possibly fix the parity for this stripe
3873 * Any reads will already have been scheduled, so we just see if enough
3874 * data is available. The parity check is held off while parity
3875 * dependent operations are in flight.
3877 if (sh->check_state ||
3878 (s.syncing && s.locked == 0 &&
3879 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3880 !test_bit(STRIPE_INSYNC, &sh->state))) {
3881 if (conf->level == 6)
3882 handle_parity_checks6(conf, sh, &s, disks);
3884 handle_parity_checks5(conf, sh, &s, disks);
3887 if ((s.replacing || s.syncing) && s.locked == 0
3888 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3889 && !test_bit(STRIPE_REPLACED, &sh->state)) {
3890 /* Write out to replacement devices where possible */
3891 for (i = 0; i < conf->raid_disks; i++)
3892 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3893 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
3894 set_bit(R5_WantReplace, &sh->dev[i].flags);
3895 set_bit(R5_LOCKED, &sh->dev[i].flags);
3899 set_bit(STRIPE_INSYNC, &sh->state);
3900 set_bit(STRIPE_REPLACED, &sh->state);
3902 if ((s.syncing || s.replacing) && s.locked == 0 &&
3903 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3904 test_bit(STRIPE_INSYNC, &sh->state)) {
3905 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3906 clear_bit(STRIPE_SYNCING, &sh->state);
3907 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3908 wake_up(&conf->wait_for_overlap);
3911 /* If the failed drives are just a ReadError, then we might need
3912 * to progress the repair/check process
3914 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3915 for (i = 0; i < s.failed; i++) {
3916 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3917 if (test_bit(R5_ReadError, &dev->flags)
3918 && !test_bit(R5_LOCKED, &dev->flags)
3919 && test_bit(R5_UPTODATE, &dev->flags)
3921 if (!test_bit(R5_ReWrite, &dev->flags)) {
3922 set_bit(R5_Wantwrite, &dev->flags);
3923 set_bit(R5_ReWrite, &dev->flags);
3924 set_bit(R5_LOCKED, &dev->flags);
3927 /* let's read it back */
3928 set_bit(R5_Wantread, &dev->flags);
3929 set_bit(R5_LOCKED, &dev->flags);
3935 /* Finish reconstruct operations initiated by the expansion process */
3936 if (sh->reconstruct_state == reconstruct_state_result) {
3937 struct stripe_head *sh_src
3938 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3939 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3940 /* sh cannot be written until sh_src has been read.
3941 * so arrange for sh to be delayed a little
3943 set_bit(STRIPE_DELAYED, &sh->state);
3944 set_bit(STRIPE_HANDLE, &sh->state);
3945 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3947 atomic_inc(&conf->preread_active_stripes);
3948 release_stripe(sh_src);
3952 release_stripe(sh_src);
3954 sh->reconstruct_state = reconstruct_state_idle;
3955 clear_bit(STRIPE_EXPANDING, &sh->state);
3956 for (i = conf->raid_disks; i--; ) {
3957 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3958 set_bit(R5_LOCKED, &sh->dev[i].flags);
3963 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3964 !sh->reconstruct_state) {
3965 /* Need to write out all blocks after computing parity */
3966 sh->disks = conf->raid_disks;
3967 stripe_set_idx(sh->sector, conf, 0, sh);
3968 schedule_reconstruction(sh, &s, 1, 1);
3969 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3970 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3971 atomic_dec(&conf->reshape_stripes);
3972 wake_up(&conf->wait_for_overlap);
3973 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3976 if (s.expanding && s.locked == 0 &&
3977 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3978 handle_stripe_expansion(conf, sh);
3981 /* wait for this device to become unblocked */
3982 if (unlikely(s.blocked_rdev)) {
3983 if (conf->mddev->external)
3984 md_wait_for_blocked_rdev(s.blocked_rdev,
3987 /* Internal metadata will immediately
3988 * be written by raid5d, so we don't
3989 * need to wait here.
3991 rdev_dec_pending(s.blocked_rdev,
3995 if (s.handle_bad_blocks)
3996 for (i = disks; i--; ) {
3997 struct md_rdev *rdev;
3998 struct r5dev *dev = &sh->dev[i];
3999 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4000 /* We own a safe reference to the rdev */
4001 rdev = conf->disks[i].rdev;
4002 if (!rdev_set_badblocks(rdev, sh->sector,
4004 md_error(conf->mddev, rdev);
4005 rdev_dec_pending(rdev, conf->mddev);
4007 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4008 rdev = conf->disks[i].rdev;
4009 rdev_clear_badblocks(rdev, sh->sector,
4011 rdev_dec_pending(rdev, conf->mddev);
4013 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4014 rdev = conf->disks[i].replacement;
4016 /* rdev have been moved down */
4017 rdev = conf->disks[i].rdev;
4018 rdev_clear_badblocks(rdev, sh->sector,
4020 rdev_dec_pending(rdev, conf->mddev);
4025 raid_run_ops(sh, s.ops_request);
4029 if (s.dec_preread_active) {
4030 /* We delay this until after ops_run_io so that if make_request
4031 * is waiting on a flush, it won't continue until the writes
4032 * have actually been submitted.
4034 atomic_dec(&conf->preread_active_stripes);
4035 if (atomic_read(&conf->preread_active_stripes) <
4037 md_wakeup_thread(conf->mddev->thread);
4040 return_io(s.return_bi);
4042 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4045 static void raid5_activate_delayed(struct r5conf *conf)
4047 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4048 while (!list_empty(&conf->delayed_list)) {
4049 struct list_head *l = conf->delayed_list.next;
4050 struct stripe_head *sh;
4051 sh = list_entry(l, struct stripe_head, lru);
4053 clear_bit(STRIPE_DELAYED, &sh->state);
4054 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4055 atomic_inc(&conf->preread_active_stripes);
4056 list_add_tail(&sh->lru, &conf->hold_list);
4057 raid5_wakeup_stripe_thread(sh);
4062 static void activate_bit_delay(struct r5conf *conf,
4063 struct list_head *temp_inactive_list)
4065 /* device_lock is held */
4066 struct list_head head;
4067 list_add(&head, &conf->bitmap_list);
4068 list_del_init(&conf->bitmap_list);
4069 while (!list_empty(&head)) {
4070 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4072 list_del_init(&sh->lru);
4073 atomic_inc(&sh->count);
4074 hash = sh->hash_lock_index;
4075 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4079 int md_raid5_congested(struct mddev *mddev, int bits)
4081 struct r5conf *conf = mddev->private;
4083 /* No difference between reads and writes. Just check
4084 * how busy the stripe_cache is
4087 if (conf->inactive_blocked)
4091 if (atomic_read(&conf->empty_inactive_list_nr))
4096 EXPORT_SYMBOL_GPL(md_raid5_congested);
4098 static int raid5_congested(void *data, int bits)
4100 struct mddev *mddev = data;
4102 return mddev_congested(mddev, bits) ||
4103 md_raid5_congested(mddev, bits);
4106 /* We want read requests to align with chunks where possible,
4107 * but write requests don't need to.
4109 static int raid5_mergeable_bvec(struct request_queue *q,
4110 struct bvec_merge_data *bvm,
4111 struct bio_vec *biovec)
4113 struct mddev *mddev = q->queuedata;
4114 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4116 unsigned int chunk_sectors = mddev->chunk_sectors;
4117 unsigned int bio_sectors = bvm->bi_size >> 9;
4119 if ((bvm->bi_rw & 1) == WRITE)
4120 return biovec->bv_len; /* always allow writes to be mergeable */
4122 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4123 chunk_sectors = mddev->new_chunk_sectors;
4124 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4125 if (max < 0) max = 0;
4126 if (max <= biovec->bv_len && bio_sectors == 0)
4127 return biovec->bv_len;
4132 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4134 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4135 unsigned int chunk_sectors = mddev->chunk_sectors;
4136 unsigned int bio_sectors = bio_sectors(bio);
4138 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4139 chunk_sectors = mddev->new_chunk_sectors;
4140 return chunk_sectors >=
4141 ((sector & (chunk_sectors - 1)) + bio_sectors);
4145 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4146 * later sampled by raid5d.
4148 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4150 unsigned long flags;
4152 spin_lock_irqsave(&conf->device_lock, flags);
4154 bi->bi_next = conf->retry_read_aligned_list;
4155 conf->retry_read_aligned_list = bi;
4157 spin_unlock_irqrestore(&conf->device_lock, flags);
4158 md_wakeup_thread(conf->mddev->thread);
4161 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4165 bi = conf->retry_read_aligned;
4167 conf->retry_read_aligned = NULL;
4170 bi = conf->retry_read_aligned_list;
4172 conf->retry_read_aligned_list = bi->bi_next;
4175 * this sets the active strip count to 1 and the processed
4176 * strip count to zero (upper 8 bits)
4178 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4185 * The "raid5_align_endio" should check if the read succeeded and if it
4186 * did, call bio_endio on the original bio (having bio_put the new bio
4188 * If the read failed..
4190 static void raid5_align_endio(struct bio *bi, int error)
4192 struct bio* raid_bi = bi->bi_private;
4193 struct mddev *mddev;
4194 struct r5conf *conf;
4195 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4196 struct md_rdev *rdev;
4200 rdev = (void*)raid_bi->bi_next;
4201 raid_bi->bi_next = NULL;
4202 mddev = rdev->mddev;
4203 conf = mddev->private;
4205 rdev_dec_pending(rdev, conf->mddev);
4207 if (!error && uptodate) {
4208 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4210 bio_endio(raid_bi, 0);
4211 if (atomic_dec_and_test(&conf->active_aligned_reads))
4212 wake_up(&conf->wait_for_stripe);
4216 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4218 add_bio_to_retry(raid_bi, conf);
4221 static int bio_fits_rdev(struct bio *bi)
4223 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4225 if (bio_sectors(bi) > queue_max_sectors(q))
4227 blk_recount_segments(q, bi);
4228 if (bi->bi_phys_segments > queue_max_segments(q))
4231 if (q->merge_bvec_fn)
4232 /* it's too hard to apply the merge_bvec_fn at this stage,
4240 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4242 struct r5conf *conf = mddev->private;
4244 struct bio* align_bi;
4245 struct md_rdev *rdev;
4246 sector_t end_sector;
4248 if (!in_chunk_boundary(mddev, raid_bio)) {
4249 pr_debug("chunk_aligned_read : non aligned\n");
4253 * use bio_clone_mddev to make a copy of the bio
4255 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4259 * set bi_end_io to a new function, and set bi_private to the
4262 align_bi->bi_end_io = raid5_align_endio;
4263 align_bi->bi_private = raid_bio;
4267 align_bi->bi_iter.bi_sector =
4268 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4271 end_sector = bio_end_sector(align_bi);
4273 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4274 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4275 rdev->recovery_offset < end_sector) {
4276 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4278 (test_bit(Faulty, &rdev->flags) ||
4279 !(test_bit(In_sync, &rdev->flags) ||
4280 rdev->recovery_offset >= end_sector)))
4287 atomic_inc(&rdev->nr_pending);
4289 raid_bio->bi_next = (void*)rdev;
4290 align_bi->bi_bdev = rdev->bdev;
4291 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
4293 if (!bio_fits_rdev(align_bi) ||
4294 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4295 bio_sectors(align_bi),
4296 &first_bad, &bad_sectors)) {
4297 /* too big in some way, or has a known bad block */
4299 rdev_dec_pending(rdev, mddev);
4303 /* No reshape active, so we can trust rdev->data_offset */
4304 align_bi->bi_iter.bi_sector += rdev->data_offset;
4306 spin_lock_irq(&conf->device_lock);
4307 wait_event_lock_irq(conf->wait_for_stripe,
4310 atomic_inc(&conf->active_aligned_reads);
4311 spin_unlock_irq(&conf->device_lock);
4314 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4315 align_bi, disk_devt(mddev->gendisk),
4316 raid_bio->bi_iter.bi_sector);
4317 generic_make_request(align_bi);
4326 /* __get_priority_stripe - get the next stripe to process
4328 * Full stripe writes are allowed to pass preread active stripes up until
4329 * the bypass_threshold is exceeded. In general the bypass_count
4330 * increments when the handle_list is handled before the hold_list; however, it
4331 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4332 * stripe with in flight i/o. The bypass_count will be reset when the
4333 * head of the hold_list has changed, i.e. the head was promoted to the
4336 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4338 struct stripe_head *sh = NULL, *tmp;
4339 struct list_head *handle_list = NULL;
4340 struct r5worker_group *wg = NULL;
4342 if (conf->worker_cnt_per_group == 0) {
4343 handle_list = &conf->handle_list;
4344 } else if (group != ANY_GROUP) {
4345 handle_list = &conf->worker_groups[group].handle_list;
4346 wg = &conf->worker_groups[group];
4349 for (i = 0; i < conf->group_cnt; i++) {
4350 handle_list = &conf->worker_groups[i].handle_list;
4351 wg = &conf->worker_groups[i];
4352 if (!list_empty(handle_list))
4357 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4359 list_empty(handle_list) ? "empty" : "busy",
4360 list_empty(&conf->hold_list) ? "empty" : "busy",
4361 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4363 if (!list_empty(handle_list)) {
4364 sh = list_entry(handle_list->next, typeof(*sh), lru);
4366 if (list_empty(&conf->hold_list))
4367 conf->bypass_count = 0;
4368 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4369 if (conf->hold_list.next == conf->last_hold)
4370 conf->bypass_count++;
4372 conf->last_hold = conf->hold_list.next;
4373 conf->bypass_count -= conf->bypass_threshold;
4374 if (conf->bypass_count < 0)
4375 conf->bypass_count = 0;
4378 } else if (!list_empty(&conf->hold_list) &&
4379 ((conf->bypass_threshold &&
4380 conf->bypass_count > conf->bypass_threshold) ||
4381 atomic_read(&conf->pending_full_writes) == 0)) {
4383 list_for_each_entry(tmp, &conf->hold_list, lru) {
4384 if (conf->worker_cnt_per_group == 0 ||
4385 group == ANY_GROUP ||
4386 !cpu_online(tmp->cpu) ||
4387 cpu_to_group(tmp->cpu) == group) {
4394 conf->bypass_count -= conf->bypass_threshold;
4395 if (conf->bypass_count < 0)
4396 conf->bypass_count = 0;
4408 list_del_init(&sh->lru);
4409 BUG_ON(atomic_inc_return(&sh->count) != 1);
4413 struct raid5_plug_cb {
4414 struct blk_plug_cb cb;
4415 struct list_head list;
4416 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4419 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4421 struct raid5_plug_cb *cb = container_of(
4422 blk_cb, struct raid5_plug_cb, cb);
4423 struct stripe_head *sh;
4424 struct mddev *mddev = cb->cb.data;
4425 struct r5conf *conf = mddev->private;
4429 if (cb->list.next && !list_empty(&cb->list)) {
4430 spin_lock_irq(&conf->device_lock);
4431 while (!list_empty(&cb->list)) {
4432 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4433 list_del_init(&sh->lru);
4435 * avoid race release_stripe_plug() sees
4436 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4437 * is still in our list
4439 smp_mb__before_atomic();
4440 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4442 * STRIPE_ON_RELEASE_LIST could be set here. In that
4443 * case, the count is always > 1 here
4445 hash = sh->hash_lock_index;
4446 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
4449 spin_unlock_irq(&conf->device_lock);
4451 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4452 NR_STRIPE_HASH_LOCKS);
4454 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4458 static void release_stripe_plug(struct mddev *mddev,
4459 struct stripe_head *sh)
4461 struct blk_plug_cb *blk_cb = blk_check_plugged(
4462 raid5_unplug, mddev,
4463 sizeof(struct raid5_plug_cb));
4464 struct raid5_plug_cb *cb;
4471 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4473 if (cb->list.next == NULL) {
4475 INIT_LIST_HEAD(&cb->list);
4476 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
4477 INIT_LIST_HEAD(cb->temp_inactive_list + i);
4480 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4481 list_add_tail(&sh->lru, &cb->list);
4486 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4488 struct r5conf *conf = mddev->private;
4489 sector_t logical_sector, last_sector;
4490 struct stripe_head *sh;
4494 if (mddev->reshape_position != MaxSector)
4495 /* Skip discard while reshape is happening */
4498 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4499 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
4502 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4504 stripe_sectors = conf->chunk_sectors *
4505 (conf->raid_disks - conf->max_degraded);
4506 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4508 sector_div(last_sector, stripe_sectors);
4510 logical_sector *= conf->chunk_sectors;
4511 last_sector *= conf->chunk_sectors;
4513 for (; logical_sector < last_sector;
4514 logical_sector += STRIPE_SECTORS) {
4518 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4519 prepare_to_wait(&conf->wait_for_overlap, &w,
4520 TASK_UNINTERRUPTIBLE);
4521 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4522 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4527 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4528 spin_lock_irq(&sh->stripe_lock);
4529 for (d = 0; d < conf->raid_disks; d++) {
4530 if (d == sh->pd_idx || d == sh->qd_idx)
4532 if (sh->dev[d].towrite || sh->dev[d].toread) {
4533 set_bit(R5_Overlap, &sh->dev[d].flags);
4534 spin_unlock_irq(&sh->stripe_lock);
4540 set_bit(STRIPE_DISCARD, &sh->state);
4541 finish_wait(&conf->wait_for_overlap, &w);
4542 for (d = 0; d < conf->raid_disks; d++) {
4543 if (d == sh->pd_idx || d == sh->qd_idx)
4545 sh->dev[d].towrite = bi;
4546 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4547 raid5_inc_bi_active_stripes(bi);
4549 spin_unlock_irq(&sh->stripe_lock);
4550 if (conf->mddev->bitmap) {
4552 d < conf->raid_disks - conf->max_degraded;
4554 bitmap_startwrite(mddev->bitmap,
4558 sh->bm_seq = conf->seq_flush + 1;
4559 set_bit(STRIPE_BIT_DELAY, &sh->state);
4562 set_bit(STRIPE_HANDLE, &sh->state);
4563 clear_bit(STRIPE_DELAYED, &sh->state);
4564 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4565 atomic_inc(&conf->preread_active_stripes);
4566 release_stripe_plug(mddev, sh);
4569 remaining = raid5_dec_bi_active_stripes(bi);
4570 if (remaining == 0) {
4571 md_write_end(mddev);
4576 static void make_request(struct mddev *mddev, struct bio * bi)
4578 struct r5conf *conf = mddev->private;
4580 sector_t new_sector;
4581 sector_t logical_sector, last_sector;
4582 struct stripe_head *sh;
4583 const int rw = bio_data_dir(bi);
4588 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4589 md_flush_request(mddev, bi);
4593 md_write_start(mddev, bi);
4596 mddev->reshape_position == MaxSector &&
4597 chunk_aligned_read(mddev,bi))
4600 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4601 make_discard_request(mddev, bi);
4605 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4606 last_sector = bio_end_sector(bi);
4608 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4610 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4611 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4617 seq = read_seqcount_begin(&conf->gen_lock);
4620 prepare_to_wait(&conf->wait_for_overlap, &w,
4621 TASK_UNINTERRUPTIBLE);
4622 if (unlikely(conf->reshape_progress != MaxSector)) {
4623 /* spinlock is needed as reshape_progress may be
4624 * 64bit on a 32bit platform, and so it might be
4625 * possible to see a half-updated value
4626 * Of course reshape_progress could change after
4627 * the lock is dropped, so once we get a reference
4628 * to the stripe that we think it is, we will have
4631 spin_lock_irq(&conf->device_lock);
4632 if (mddev->reshape_backwards
4633 ? logical_sector < conf->reshape_progress
4634 : logical_sector >= conf->reshape_progress) {
4637 if (mddev->reshape_backwards
4638 ? logical_sector < conf->reshape_safe
4639 : logical_sector >= conf->reshape_safe) {
4640 spin_unlock_irq(&conf->device_lock);
4646 spin_unlock_irq(&conf->device_lock);
4649 new_sector = raid5_compute_sector(conf, logical_sector,
4652 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4653 (unsigned long long)new_sector,
4654 (unsigned long long)logical_sector);
4656 sh = get_active_stripe(conf, new_sector, previous,
4657 (bi->bi_rw&RWA_MASK), 0);
4659 if (unlikely(previous)) {
4660 /* expansion might have moved on while waiting for a
4661 * stripe, so we must do the range check again.
4662 * Expansion could still move past after this
4663 * test, but as we are holding a reference to
4664 * 'sh', we know that if that happens,
4665 * STRIPE_EXPANDING will get set and the expansion
4666 * won't proceed until we finish with the stripe.
4669 spin_lock_irq(&conf->device_lock);
4670 if (mddev->reshape_backwards
4671 ? logical_sector >= conf->reshape_progress
4672 : logical_sector < conf->reshape_progress)
4673 /* mismatch, need to try again */
4675 spin_unlock_irq(&conf->device_lock);
4683 if (read_seqcount_retry(&conf->gen_lock, seq)) {
4684 /* Might have got the wrong stripe_head
4692 logical_sector >= mddev->suspend_lo &&
4693 logical_sector < mddev->suspend_hi) {
4695 /* As the suspend_* range is controlled by
4696 * userspace, we want an interruptible
4699 flush_signals(current);
4700 prepare_to_wait(&conf->wait_for_overlap,
4701 &w, TASK_INTERRUPTIBLE);
4702 if (logical_sector >= mddev->suspend_lo &&
4703 logical_sector < mddev->suspend_hi) {
4710 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4711 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4712 /* Stripe is busy expanding or
4713 * add failed due to overlap. Flush everything
4716 md_wakeup_thread(mddev->thread);
4722 set_bit(STRIPE_HANDLE, &sh->state);
4723 clear_bit(STRIPE_DELAYED, &sh->state);
4724 if ((bi->bi_rw & REQ_SYNC) &&
4725 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4726 atomic_inc(&conf->preread_active_stripes);
4727 release_stripe_plug(mddev, sh);
4729 /* cannot get stripe for read-ahead, just give-up */
4730 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4734 finish_wait(&conf->wait_for_overlap, &w);
4736 remaining = raid5_dec_bi_active_stripes(bi);
4737 if (remaining == 0) {
4740 md_write_end(mddev);
4742 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4748 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4750 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4752 /* reshaping is quite different to recovery/resync so it is
4753 * handled quite separately ... here.
4755 * On each call to sync_request, we gather one chunk worth of
4756 * destination stripes and flag them as expanding.
4757 * Then we find all the source stripes and request reads.
4758 * As the reads complete, handle_stripe will copy the data
4759 * into the destination stripe and release that stripe.
4761 struct r5conf *conf = mddev->private;
4762 struct stripe_head *sh;
4763 sector_t first_sector, last_sector;
4764 int raid_disks = conf->previous_raid_disks;
4765 int data_disks = raid_disks - conf->max_degraded;
4766 int new_data_disks = conf->raid_disks - conf->max_degraded;
4769 sector_t writepos, readpos, safepos;
4770 sector_t stripe_addr;
4771 int reshape_sectors;
4772 struct list_head stripes;
4774 if (sector_nr == 0) {
4775 /* If restarting in the middle, skip the initial sectors */
4776 if (mddev->reshape_backwards &&
4777 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4778 sector_nr = raid5_size(mddev, 0, 0)
4779 - conf->reshape_progress;
4780 } else if (!mddev->reshape_backwards &&
4781 conf->reshape_progress > 0)
4782 sector_nr = conf->reshape_progress;
4783 sector_div(sector_nr, new_data_disks);
4785 mddev->curr_resync_completed = sector_nr;
4786 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4792 /* We need to process a full chunk at a time.
4793 * If old and new chunk sizes differ, we need to process the
4796 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4797 reshape_sectors = mddev->new_chunk_sectors;
4799 reshape_sectors = mddev->chunk_sectors;
4801 /* We update the metadata at least every 10 seconds, or when
4802 * the data about to be copied would over-write the source of
4803 * the data at the front of the range. i.e. one new_stripe
4804 * along from reshape_progress new_maps to after where
4805 * reshape_safe old_maps to
4807 writepos = conf->reshape_progress;
4808 sector_div(writepos, new_data_disks);
4809 readpos = conf->reshape_progress;
4810 sector_div(readpos, data_disks);
4811 safepos = conf->reshape_safe;
4812 sector_div(safepos, data_disks);
4813 if (mddev->reshape_backwards) {
4814 writepos -= min_t(sector_t, reshape_sectors, writepos);
4815 readpos += reshape_sectors;
4816 safepos += reshape_sectors;
4818 writepos += reshape_sectors;
4819 readpos -= min_t(sector_t, reshape_sectors, readpos);
4820 safepos -= min_t(sector_t, reshape_sectors, safepos);
4823 /* Having calculated the 'writepos' possibly use it
4824 * to set 'stripe_addr' which is where we will write to.
4826 if (mddev->reshape_backwards) {
4827 BUG_ON(conf->reshape_progress == 0);
4828 stripe_addr = writepos;
4829 BUG_ON((mddev->dev_sectors &
4830 ~((sector_t)reshape_sectors - 1))
4831 - reshape_sectors - stripe_addr
4834 BUG_ON(writepos != sector_nr + reshape_sectors);
4835 stripe_addr = sector_nr;
4838 /* 'writepos' is the most advanced device address we might write.
4839 * 'readpos' is the least advanced device address we might read.
4840 * 'safepos' is the least address recorded in the metadata as having
4842 * If there is a min_offset_diff, these are adjusted either by
4843 * increasing the safepos/readpos if diff is negative, or
4844 * increasing writepos if diff is positive.
4845 * If 'readpos' is then behind 'writepos', there is no way that we can
4846 * ensure safety in the face of a crash - that must be done by userspace
4847 * making a backup of the data. So in that case there is no particular
4848 * rush to update metadata.
4849 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4850 * update the metadata to advance 'safepos' to match 'readpos' so that
4851 * we can be safe in the event of a crash.
4852 * So we insist on updating metadata if safepos is behind writepos and
4853 * readpos is beyond writepos.
4854 * In any case, update the metadata every 10 seconds.
4855 * Maybe that number should be configurable, but I'm not sure it is
4856 * worth it.... maybe it could be a multiple of safemode_delay???
4858 if (conf->min_offset_diff < 0) {
4859 safepos += -conf->min_offset_diff;
4860 readpos += -conf->min_offset_diff;
4862 writepos += conf->min_offset_diff;
4864 if ((mddev->reshape_backwards
4865 ? (safepos > writepos && readpos < writepos)
4866 : (safepos < writepos && readpos > writepos)) ||
4867 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4868 /* Cannot proceed until we've updated the superblock... */
4869 wait_event(conf->wait_for_overlap,
4870 atomic_read(&conf->reshape_stripes)==0
4871 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4872 if (atomic_read(&conf->reshape_stripes) != 0)
4874 mddev->reshape_position = conf->reshape_progress;
4875 mddev->curr_resync_completed = sector_nr;
4876 conf->reshape_checkpoint = jiffies;
4877 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4878 md_wakeup_thread(mddev->thread);
4879 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4880 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4881 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4883 spin_lock_irq(&conf->device_lock);
4884 conf->reshape_safe = mddev->reshape_position;
4885 spin_unlock_irq(&conf->device_lock);
4886 wake_up(&conf->wait_for_overlap);
4887 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4890 INIT_LIST_HEAD(&stripes);
4891 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4893 int skipped_disk = 0;
4894 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4895 set_bit(STRIPE_EXPANDING, &sh->state);
4896 atomic_inc(&conf->reshape_stripes);
4897 /* If any of this stripe is beyond the end of the old
4898 * array, then we need to zero those blocks
4900 for (j=sh->disks; j--;) {
4902 if (j == sh->pd_idx)
4904 if (conf->level == 6 &&
4907 s = compute_blocknr(sh, j, 0);
4908 if (s < raid5_size(mddev, 0, 0)) {
4912 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4913 set_bit(R5_Expanded, &sh->dev[j].flags);
4914 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4916 if (!skipped_disk) {
4917 set_bit(STRIPE_EXPAND_READY, &sh->state);
4918 set_bit(STRIPE_HANDLE, &sh->state);
4920 list_add(&sh->lru, &stripes);
4922 spin_lock_irq(&conf->device_lock);
4923 if (mddev->reshape_backwards)
4924 conf->reshape_progress -= reshape_sectors * new_data_disks;
4926 conf->reshape_progress += reshape_sectors * new_data_disks;
4927 spin_unlock_irq(&conf->device_lock);
4928 /* Ok, those stripe are ready. We can start scheduling
4929 * reads on the source stripes.
4930 * The source stripes are determined by mapping the first and last
4931 * block on the destination stripes.
4934 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4937 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4938 * new_data_disks - 1),
4940 if (last_sector >= mddev->dev_sectors)
4941 last_sector = mddev->dev_sectors - 1;
4942 while (first_sector <= last_sector) {
4943 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4944 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4945 set_bit(STRIPE_HANDLE, &sh->state);
4947 first_sector += STRIPE_SECTORS;
4949 /* Now that the sources are clearly marked, we can release
4950 * the destination stripes
4952 while (!list_empty(&stripes)) {
4953 sh = list_entry(stripes.next, struct stripe_head, lru);
4954 list_del_init(&sh->lru);
4957 /* If this takes us to the resync_max point where we have to pause,
4958 * then we need to write out the superblock.
4960 sector_nr += reshape_sectors;
4961 if ((sector_nr - mddev->curr_resync_completed) * 2
4962 >= mddev->resync_max - mddev->curr_resync_completed) {
4963 /* Cannot proceed until we've updated the superblock... */
4964 wait_event(conf->wait_for_overlap,
4965 atomic_read(&conf->reshape_stripes) == 0
4966 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4967 if (atomic_read(&conf->reshape_stripes) != 0)
4969 mddev->reshape_position = conf->reshape_progress;
4970 mddev->curr_resync_completed = sector_nr;
4971 conf->reshape_checkpoint = jiffies;
4972 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4973 md_wakeup_thread(mddev->thread);
4974 wait_event(mddev->sb_wait,
4975 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4976 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4977 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4979 spin_lock_irq(&conf->device_lock);
4980 conf->reshape_safe = mddev->reshape_position;
4981 spin_unlock_irq(&conf->device_lock);
4982 wake_up(&conf->wait_for_overlap);
4983 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4986 return reshape_sectors;
4989 /* FIXME go_faster isn't used */
4990 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4992 struct r5conf *conf = mddev->private;
4993 struct stripe_head *sh;
4994 sector_t max_sector = mddev->dev_sectors;
4995 sector_t sync_blocks;
4996 int still_degraded = 0;
4999 if (sector_nr >= max_sector) {
5000 /* just being told to finish up .. nothing much to do */
5002 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5007 if (mddev->curr_resync < max_sector) /* aborted */
5008 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5010 else /* completed sync */
5012 bitmap_close_sync(mddev->bitmap);
5017 /* Allow raid5_quiesce to complete */
5018 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5020 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5021 return reshape_request(mddev, sector_nr, skipped);
5023 /* No need to check resync_max as we never do more than one
5024 * stripe, and as resync_max will always be on a chunk boundary,
5025 * if the check in md_do_sync didn't fire, there is no chance
5026 * of overstepping resync_max here
5029 /* if there is too many failed drives and we are trying
5030 * to resync, then assert that we are finished, because there is
5031 * nothing we can do.
5033 if (mddev->degraded >= conf->max_degraded &&
5034 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5035 sector_t rv = mddev->dev_sectors - sector_nr;
5039 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5041 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5042 sync_blocks >= STRIPE_SECTORS) {
5043 /* we can skip this block, and probably more */
5044 sync_blocks /= STRIPE_SECTORS;
5046 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5049 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5051 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5053 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5054 /* make sure we don't swamp the stripe cache if someone else
5055 * is trying to get access
5057 schedule_timeout_uninterruptible(1);
5059 /* Need to check if array will still be degraded after recovery/resync
5060 * We don't need to check the 'failed' flag as when that gets set,
5063 for (i = 0; i < conf->raid_disks; i++)
5064 if (conf->disks[i].rdev == NULL)
5067 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5069 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5070 set_bit(STRIPE_HANDLE, &sh->state);
5074 return STRIPE_SECTORS;
5077 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5079 /* We may not be able to submit a whole bio at once as there
5080 * may not be enough stripe_heads available.
5081 * We cannot pre-allocate enough stripe_heads as we may need
5082 * more than exist in the cache (if we allow ever large chunks).
5083 * So we do one stripe head at a time and record in
5084 * ->bi_hw_segments how many have been done.
5086 * We *know* that this entire raid_bio is in one chunk, so
5087 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5089 struct stripe_head *sh;
5091 sector_t sector, logical_sector, last_sector;
5096 logical_sector = raid_bio->bi_iter.bi_sector &
5097 ~((sector_t)STRIPE_SECTORS-1);
5098 sector = raid5_compute_sector(conf, logical_sector,
5100 last_sector = bio_end_sector(raid_bio);
5102 for (; logical_sector < last_sector;
5103 logical_sector += STRIPE_SECTORS,
5104 sector += STRIPE_SECTORS,
5107 if (scnt < raid5_bi_processed_stripes(raid_bio))
5108 /* already done this stripe */
5111 sh = get_active_stripe(conf, sector, 0, 1, 1);
5114 /* failed to get a stripe - must wait */
5115 raid5_set_bi_processed_stripes(raid_bio, scnt);
5116 conf->retry_read_aligned = raid_bio;
5120 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
5122 raid5_set_bi_processed_stripes(raid_bio, scnt);
5123 conf->retry_read_aligned = raid_bio;
5127 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5132 remaining = raid5_dec_bi_active_stripes(raid_bio);
5133 if (remaining == 0) {
5134 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5136 bio_endio(raid_bio, 0);
5138 if (atomic_dec_and_test(&conf->active_aligned_reads))
5139 wake_up(&conf->wait_for_stripe);
5143 static int handle_active_stripes(struct r5conf *conf, int group,
5144 struct r5worker *worker,
5145 struct list_head *temp_inactive_list)
5147 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5148 int i, batch_size = 0, hash;
5149 bool release_inactive = false;
5151 while (batch_size < MAX_STRIPE_BATCH &&
5152 (sh = __get_priority_stripe(conf, group)) != NULL)
5153 batch[batch_size++] = sh;
5155 if (batch_size == 0) {
5156 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5157 if (!list_empty(temp_inactive_list + i))
5159 if (i == NR_STRIPE_HASH_LOCKS)
5161 release_inactive = true;
5163 spin_unlock_irq(&conf->device_lock);
5165 release_inactive_stripe_list(conf, temp_inactive_list,
5166 NR_STRIPE_HASH_LOCKS);
5168 if (release_inactive) {
5169 spin_lock_irq(&conf->device_lock);
5173 for (i = 0; i < batch_size; i++)
5174 handle_stripe(batch[i]);
5178 spin_lock_irq(&conf->device_lock);
5179 for (i = 0; i < batch_size; i++) {
5180 hash = batch[i]->hash_lock_index;
5181 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5186 static void raid5_do_work(struct work_struct *work)
5188 struct r5worker *worker = container_of(work, struct r5worker, work);
5189 struct r5worker_group *group = worker->group;
5190 struct r5conf *conf = group->conf;
5191 int group_id = group - conf->worker_groups;
5193 struct blk_plug plug;
5195 pr_debug("+++ raid5worker active\n");
5197 blk_start_plug(&plug);
5199 spin_lock_irq(&conf->device_lock);
5201 int batch_size, released;
5203 released = release_stripe_list(conf, worker->temp_inactive_list);
5205 batch_size = handle_active_stripes(conf, group_id, worker,
5206 worker->temp_inactive_list);
5207 worker->working = false;
5208 if (!batch_size && !released)
5210 handled += batch_size;
5212 pr_debug("%d stripes handled\n", handled);
5214 spin_unlock_irq(&conf->device_lock);
5215 blk_finish_plug(&plug);
5217 pr_debug("--- raid5worker inactive\n");
5221 * This is our raid5 kernel thread.
5223 * We scan the hash table for stripes which can be handled now.
5224 * During the scan, completed stripes are saved for us by the interrupt
5225 * handler, so that they will not have to wait for our next wakeup.
5227 static void raid5d(struct md_thread *thread)
5229 struct mddev *mddev = thread->mddev;
5230 struct r5conf *conf = mddev->private;
5232 struct blk_plug plug;
5234 pr_debug("+++ raid5d active\n");
5236 md_check_recovery(mddev);
5238 blk_start_plug(&plug);
5240 spin_lock_irq(&conf->device_lock);
5243 int batch_size, released;
5245 released = release_stripe_list(conf, conf->temp_inactive_list);
5248 !list_empty(&conf->bitmap_list)) {
5249 /* Now is a good time to flush some bitmap updates */
5251 spin_unlock_irq(&conf->device_lock);
5252 bitmap_unplug(mddev->bitmap);
5253 spin_lock_irq(&conf->device_lock);
5254 conf->seq_write = conf->seq_flush;
5255 activate_bit_delay(conf, conf->temp_inactive_list);
5257 raid5_activate_delayed(conf);
5259 while ((bio = remove_bio_from_retry(conf))) {
5261 spin_unlock_irq(&conf->device_lock);
5262 ok = retry_aligned_read(conf, bio);
5263 spin_lock_irq(&conf->device_lock);
5269 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5270 conf->temp_inactive_list);
5271 if (!batch_size && !released)
5273 handled += batch_size;
5275 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5276 spin_unlock_irq(&conf->device_lock);
5277 md_check_recovery(mddev);
5278 spin_lock_irq(&conf->device_lock);
5281 pr_debug("%d stripes handled\n", handled);
5283 spin_unlock_irq(&conf->device_lock);
5285 async_tx_issue_pending_all();
5286 blk_finish_plug(&plug);
5288 pr_debug("--- raid5d inactive\n");
5292 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5294 struct r5conf *conf = mddev->private;
5296 return sprintf(page, "%d\n", conf->max_nr_stripes);
5302 raid5_set_cache_size(struct mddev *mddev, int size)
5304 struct r5conf *conf = mddev->private;
5308 if (size <= 16 || size > 32768)
5310 hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
5311 while (size < conf->max_nr_stripes) {
5312 if (drop_one_stripe(conf, hash))
5313 conf->max_nr_stripes--;
5318 hash = NR_STRIPE_HASH_LOCKS - 1;
5320 err = md_allow_write(mddev);
5323 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
5324 while (size > conf->max_nr_stripes) {
5325 if (grow_one_stripe(conf, hash))
5326 conf->max_nr_stripes++;
5328 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
5332 EXPORT_SYMBOL(raid5_set_cache_size);
5335 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5337 struct r5conf *conf = mddev->private;
5341 if (len >= PAGE_SIZE)
5346 if (kstrtoul(page, 10, &new))
5348 err = raid5_set_cache_size(mddev, new);
5354 static struct md_sysfs_entry
5355 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5356 raid5_show_stripe_cache_size,
5357 raid5_store_stripe_cache_size);
5360 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5362 struct r5conf *conf = mddev->private;
5364 return sprintf(page, "%d\n", conf->bypass_threshold);
5370 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
5372 struct r5conf *conf = mddev->private;
5374 if (len >= PAGE_SIZE)
5379 if (kstrtoul(page, 10, &new))
5381 if (new > conf->max_nr_stripes)
5383 conf->bypass_threshold = new;
5387 static struct md_sysfs_entry
5388 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5390 raid5_show_preread_threshold,
5391 raid5_store_preread_threshold);
5394 raid5_show_skip_copy(struct mddev *mddev, char *page)
5396 struct r5conf *conf = mddev->private;
5398 return sprintf(page, "%d\n", conf->skip_copy);
5404 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
5406 struct r5conf *conf = mddev->private;
5408 if (len >= PAGE_SIZE)
5413 if (kstrtoul(page, 10, &new))
5416 if (new == conf->skip_copy)
5419 mddev_suspend(mddev);
5420 conf->skip_copy = new;
5422 mddev->queue->backing_dev_info.capabilities |=
5423 BDI_CAP_STABLE_WRITES;
5425 mddev->queue->backing_dev_info.capabilities &=
5426 ~BDI_CAP_STABLE_WRITES;
5427 mddev_resume(mddev);
5431 static struct md_sysfs_entry
5432 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
5433 raid5_show_skip_copy,
5434 raid5_store_skip_copy);
5437 stripe_cache_active_show(struct mddev *mddev, char *page)
5439 struct r5conf *conf = mddev->private;
5441 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5446 static struct md_sysfs_entry
5447 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
5450 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
5452 struct r5conf *conf = mddev->private;
5454 return sprintf(page, "%d\n", conf->worker_cnt_per_group);
5459 static int alloc_thread_groups(struct r5conf *conf, int cnt,
5461 int *worker_cnt_per_group,
5462 struct r5worker_group **worker_groups);
5464 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
5466 struct r5conf *conf = mddev->private;
5469 struct r5worker_group *new_groups, *old_groups;
5470 int group_cnt, worker_cnt_per_group;
5472 if (len >= PAGE_SIZE)
5477 if (kstrtoul(page, 10, &new))
5480 if (new == conf->worker_cnt_per_group)
5483 mddev_suspend(mddev);
5485 old_groups = conf->worker_groups;
5487 flush_workqueue(raid5_wq);
5489 err = alloc_thread_groups(conf, new,
5490 &group_cnt, &worker_cnt_per_group,
5493 spin_lock_irq(&conf->device_lock);
5494 conf->group_cnt = group_cnt;
5495 conf->worker_cnt_per_group = worker_cnt_per_group;
5496 conf->worker_groups = new_groups;
5497 spin_unlock_irq(&conf->device_lock);
5500 kfree(old_groups[0].workers);
5504 mddev_resume(mddev);
5511 static struct md_sysfs_entry
5512 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
5513 raid5_show_group_thread_cnt,
5514 raid5_store_group_thread_cnt);
5516 static struct attribute *raid5_attrs[] = {
5517 &raid5_stripecache_size.attr,
5518 &raid5_stripecache_active.attr,
5519 &raid5_preread_bypass_threshold.attr,
5520 &raid5_group_thread_cnt.attr,
5521 &raid5_skip_copy.attr,
5524 static struct attribute_group raid5_attrs_group = {
5526 .attrs = raid5_attrs,
5529 static int alloc_thread_groups(struct r5conf *conf, int cnt,
5531 int *worker_cnt_per_group,
5532 struct r5worker_group **worker_groups)
5536 struct r5worker *workers;
5538 *worker_cnt_per_group = cnt;
5541 *worker_groups = NULL;
5544 *group_cnt = num_possible_nodes();
5545 size = sizeof(struct r5worker) * cnt;
5546 workers = kzalloc(size * *group_cnt, GFP_NOIO);
5547 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
5548 *group_cnt, GFP_NOIO);
5549 if (!*worker_groups || !workers) {
5551 kfree(*worker_groups);
5555 for (i = 0; i < *group_cnt; i++) {
5556 struct r5worker_group *group;
5558 group = &(*worker_groups)[i];
5559 INIT_LIST_HEAD(&group->handle_list);
5561 group->workers = workers + i * cnt;
5563 for (j = 0; j < cnt; j++) {
5564 struct r5worker *worker = group->workers + j;
5565 worker->group = group;
5566 INIT_WORK(&worker->work, raid5_do_work);
5568 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
5569 INIT_LIST_HEAD(worker->temp_inactive_list + k);
5576 static void free_thread_groups(struct r5conf *conf)
5578 if (conf->worker_groups)
5579 kfree(conf->worker_groups[0].workers);
5580 kfree(conf->worker_groups);
5581 conf->worker_groups = NULL;
5585 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
5587 struct r5conf *conf = mddev->private;
5590 sectors = mddev->dev_sectors;
5592 /* size is defined by the smallest of previous and new size */
5593 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
5595 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5596 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
5597 return sectors * (raid_disks - conf->max_degraded);
5600 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5602 safe_put_page(percpu->spare_page);
5603 kfree(percpu->scribble);
5604 percpu->spare_page = NULL;
5605 percpu->scribble = NULL;
5608 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5610 if (conf->level == 6 && !percpu->spare_page)
5611 percpu->spare_page = alloc_page(GFP_KERNEL);
5612 if (!percpu->scribble)
5613 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5615 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
5616 free_scratch_buffer(conf, percpu);
5623 static void raid5_free_percpu(struct r5conf *conf)
5630 #ifdef CONFIG_HOTPLUG_CPU
5631 unregister_cpu_notifier(&conf->cpu_notify);
5635 for_each_possible_cpu(cpu)
5636 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5639 free_percpu(conf->percpu);
5642 static void free_conf(struct r5conf *conf)
5644 free_thread_groups(conf);
5645 shrink_stripes(conf);
5646 raid5_free_percpu(conf);
5648 kfree(conf->stripe_hashtbl);
5652 #ifdef CONFIG_HOTPLUG_CPU
5653 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5656 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5657 long cpu = (long)hcpu;
5658 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5661 case CPU_UP_PREPARE:
5662 case CPU_UP_PREPARE_FROZEN:
5663 if (alloc_scratch_buffer(conf, percpu)) {
5664 pr_err("%s: failed memory allocation for cpu%ld\n",
5666 return notifier_from_errno(-ENOMEM);
5670 case CPU_DEAD_FROZEN:
5671 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5680 static int raid5_alloc_percpu(struct r5conf *conf)
5685 conf->percpu = alloc_percpu(struct raid5_percpu);
5689 #ifdef CONFIG_HOTPLUG_CPU
5690 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5691 conf->cpu_notify.priority = 0;
5692 err = register_cpu_notifier(&conf->cpu_notify);
5698 for_each_present_cpu(cpu) {
5699 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5701 pr_err("%s: failed memory allocation for cpu%ld\n",
5711 static struct r5conf *setup_conf(struct mddev *mddev)
5713 struct r5conf *conf;
5714 int raid_disk, memory, max_disks;
5715 struct md_rdev *rdev;
5716 struct disk_info *disk;
5719 int group_cnt, worker_cnt_per_group;
5720 struct r5worker_group *new_group;
5722 if (mddev->new_level != 5
5723 && mddev->new_level != 4
5724 && mddev->new_level != 6) {
5725 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5726 mdname(mddev), mddev->new_level);
5727 return ERR_PTR(-EIO);
5729 if ((mddev->new_level == 5
5730 && !algorithm_valid_raid5(mddev->new_layout)) ||
5731 (mddev->new_level == 6
5732 && !algorithm_valid_raid6(mddev->new_layout))) {
5733 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5734 mdname(mddev), mddev->new_layout);
5735 return ERR_PTR(-EIO);
5737 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5738 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5739 mdname(mddev), mddev->raid_disks);
5740 return ERR_PTR(-EINVAL);
5743 if (!mddev->new_chunk_sectors ||
5744 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5745 !is_power_of_2(mddev->new_chunk_sectors)) {
5746 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5747 mdname(mddev), mddev->new_chunk_sectors << 9);
5748 return ERR_PTR(-EINVAL);
5751 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5754 /* Don't enable multi-threading by default*/
5755 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
5757 conf->group_cnt = group_cnt;
5758 conf->worker_cnt_per_group = worker_cnt_per_group;
5759 conf->worker_groups = new_group;
5762 spin_lock_init(&conf->device_lock);
5763 seqcount_init(&conf->gen_lock);
5764 init_waitqueue_head(&conf->wait_for_stripe);
5765 init_waitqueue_head(&conf->wait_for_overlap);
5766 INIT_LIST_HEAD(&conf->handle_list);
5767 INIT_LIST_HEAD(&conf->hold_list);
5768 INIT_LIST_HEAD(&conf->delayed_list);
5769 INIT_LIST_HEAD(&conf->bitmap_list);
5770 init_llist_head(&conf->released_stripes);
5771 atomic_set(&conf->active_stripes, 0);
5772 atomic_set(&conf->preread_active_stripes, 0);
5773 atomic_set(&conf->active_aligned_reads, 0);
5774 conf->bypass_threshold = BYPASS_THRESHOLD;
5775 conf->recovery_disabled = mddev->recovery_disabled - 1;
5777 conf->raid_disks = mddev->raid_disks;
5778 if (mddev->reshape_position == MaxSector)
5779 conf->previous_raid_disks = mddev->raid_disks;
5781 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5782 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5783 conf->scribble_len = scribble_len(max_disks);
5785 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5790 conf->mddev = mddev;
5792 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5795 /* We init hash_locks[0] separately to that it can be used
5796 * as the reference lock in the spin_lock_nest_lock() call
5797 * in lock_all_device_hash_locks_irq in order to convince
5798 * lockdep that we know what we are doing.
5800 spin_lock_init(conf->hash_locks);
5801 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
5802 spin_lock_init(conf->hash_locks + i);
5804 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5805 INIT_LIST_HEAD(conf->inactive_list + i);
5807 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5808 INIT_LIST_HEAD(conf->temp_inactive_list + i);
5810 conf->level = mddev->new_level;
5811 if (raid5_alloc_percpu(conf) != 0)
5814 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5816 rdev_for_each(rdev, mddev) {
5817 raid_disk = rdev->raid_disk;
5818 if (raid_disk >= max_disks
5821 disk = conf->disks + raid_disk;
5823 if (test_bit(Replacement, &rdev->flags)) {
5824 if (disk->replacement)
5826 disk->replacement = rdev;
5833 if (test_bit(In_sync, &rdev->flags)) {
5834 char b[BDEVNAME_SIZE];
5835 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5837 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5838 } else if (rdev->saved_raid_disk != raid_disk)
5839 /* Cannot rely on bitmap to complete recovery */
5843 conf->chunk_sectors = mddev->new_chunk_sectors;
5844 conf->level = mddev->new_level;
5845 if (conf->level == 6)
5846 conf->max_degraded = 2;
5848 conf->max_degraded = 1;
5849 conf->algorithm = mddev->new_layout;
5850 conf->reshape_progress = mddev->reshape_position;
5851 if (conf->reshape_progress != MaxSector) {
5852 conf->prev_chunk_sectors = mddev->chunk_sectors;
5853 conf->prev_algo = mddev->layout;
5856 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5857 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5858 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
5859 if (grow_stripes(conf, NR_STRIPES)) {
5861 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5862 mdname(mddev), memory);
5865 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5866 mdname(mddev), memory);
5868 sprintf(pers_name, "raid%d", mddev->new_level);
5869 conf->thread = md_register_thread(raid5d, mddev, pers_name);
5870 if (!conf->thread) {
5872 "md/raid:%s: couldn't allocate thread.\n",
5882 return ERR_PTR(-EIO);
5884 return ERR_PTR(-ENOMEM);
5887 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5890 case ALGORITHM_PARITY_0:
5891 if (raid_disk < max_degraded)
5894 case ALGORITHM_PARITY_N:
5895 if (raid_disk >= raid_disks - max_degraded)
5898 case ALGORITHM_PARITY_0_6:
5899 if (raid_disk == 0 ||
5900 raid_disk == raid_disks - 1)
5903 case ALGORITHM_LEFT_ASYMMETRIC_6:
5904 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5905 case ALGORITHM_LEFT_SYMMETRIC_6:
5906 case ALGORITHM_RIGHT_SYMMETRIC_6:
5907 if (raid_disk == raid_disks - 1)
5913 static int run(struct mddev *mddev)
5915 struct r5conf *conf;
5916 int working_disks = 0;
5917 int dirty_parity_disks = 0;
5918 struct md_rdev *rdev;
5919 sector_t reshape_offset = 0;
5921 long long min_offset_diff = 0;
5924 if (mddev->recovery_cp != MaxSector)
5925 printk(KERN_NOTICE "md/raid:%s: not clean"
5926 " -- starting background reconstruction\n",
5929 rdev_for_each(rdev, mddev) {
5931 if (rdev->raid_disk < 0)
5933 diff = (rdev->new_data_offset - rdev->data_offset);
5935 min_offset_diff = diff;
5937 } else if (mddev->reshape_backwards &&
5938 diff < min_offset_diff)
5939 min_offset_diff = diff;
5940 else if (!mddev->reshape_backwards &&
5941 diff > min_offset_diff)
5942 min_offset_diff = diff;
5945 if (mddev->reshape_position != MaxSector) {
5946 /* Check that we can continue the reshape.
5947 * Difficulties arise if the stripe we would write to
5948 * next is at or after the stripe we would read from next.
5949 * For a reshape that changes the number of devices, this
5950 * is only possible for a very short time, and mdadm makes
5951 * sure that time appears to have past before assembling
5952 * the array. So we fail if that time hasn't passed.
5953 * For a reshape that keeps the number of devices the same
5954 * mdadm must be monitoring the reshape can keeping the
5955 * critical areas read-only and backed up. It will start
5956 * the array in read-only mode, so we check for that.
5958 sector_t here_new, here_old;
5960 int max_degraded = (mddev->level == 6 ? 2 : 1);
5962 if (mddev->new_level != mddev->level) {
5963 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5964 "required - aborting.\n",
5968 old_disks = mddev->raid_disks - mddev->delta_disks;
5969 /* reshape_position must be on a new-stripe boundary, and one
5970 * further up in new geometry must map after here in old
5973 here_new = mddev->reshape_position;
5974 if (sector_div(here_new, mddev->new_chunk_sectors *
5975 (mddev->raid_disks - max_degraded))) {
5976 printk(KERN_ERR "md/raid:%s: reshape_position not "
5977 "on a stripe boundary\n", mdname(mddev));
5980 reshape_offset = here_new * mddev->new_chunk_sectors;
5981 /* here_new is the stripe we will write to */
5982 here_old = mddev->reshape_position;
5983 sector_div(here_old, mddev->chunk_sectors *
5984 (old_disks-max_degraded));
5985 /* here_old is the first stripe that we might need to read
5987 if (mddev->delta_disks == 0) {
5988 if ((here_new * mddev->new_chunk_sectors !=
5989 here_old * mddev->chunk_sectors)) {
5990 printk(KERN_ERR "md/raid:%s: reshape position is"
5991 " confused - aborting\n", mdname(mddev));
5994 /* We cannot be sure it is safe to start an in-place
5995 * reshape. It is only safe if user-space is monitoring
5996 * and taking constant backups.
5997 * mdadm always starts a situation like this in
5998 * readonly mode so it can take control before
5999 * allowing any writes. So just check for that.
6001 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6002 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6003 /* not really in-place - so OK */;
6004 else if (mddev->ro == 0) {
6005 printk(KERN_ERR "md/raid:%s: in-place reshape "
6006 "must be started in read-only mode "
6011 } else if (mddev->reshape_backwards
6012 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6013 here_old * mddev->chunk_sectors)
6014 : (here_new * mddev->new_chunk_sectors >=
6015 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6016 /* Reading from the same stripe as writing to - bad */
6017 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6018 "auto-recovery - aborting.\n",
6022 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6024 /* OK, we should be able to continue; */
6026 BUG_ON(mddev->level != mddev->new_level);
6027 BUG_ON(mddev->layout != mddev->new_layout);
6028 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6029 BUG_ON(mddev->delta_disks != 0);
6032 if (mddev->private == NULL)
6033 conf = setup_conf(mddev);
6035 conf = mddev->private;
6038 return PTR_ERR(conf);
6040 conf->min_offset_diff = min_offset_diff;
6041 mddev->thread = conf->thread;
6042 conf->thread = NULL;
6043 mddev->private = conf;
6045 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6047 rdev = conf->disks[i].rdev;
6048 if (!rdev && conf->disks[i].replacement) {
6049 /* The replacement is all we have yet */
6050 rdev = conf->disks[i].replacement;
6051 conf->disks[i].replacement = NULL;
6052 clear_bit(Replacement, &rdev->flags);
6053 conf->disks[i].rdev = rdev;
6057 if (conf->disks[i].replacement &&
6058 conf->reshape_progress != MaxSector) {
6059 /* replacements and reshape simply do not mix. */
6060 printk(KERN_ERR "md: cannot handle concurrent "
6061 "replacement and reshape.\n");
6064 if (test_bit(In_sync, &rdev->flags)) {
6068 /* This disc is not fully in-sync. However if it
6069 * just stored parity (beyond the recovery_offset),
6070 * when we don't need to be concerned about the
6071 * array being dirty.
6072 * When reshape goes 'backwards', we never have
6073 * partially completed devices, so we only need
6074 * to worry about reshape going forwards.
6076 /* Hack because v0.91 doesn't store recovery_offset properly. */
6077 if (mddev->major_version == 0 &&
6078 mddev->minor_version > 90)
6079 rdev->recovery_offset = reshape_offset;
6081 if (rdev->recovery_offset < reshape_offset) {
6082 /* We need to check old and new layout */
6083 if (!only_parity(rdev->raid_disk,
6086 conf->max_degraded))
6089 if (!only_parity(rdev->raid_disk,
6091 conf->previous_raid_disks,
6092 conf->max_degraded))
6094 dirty_parity_disks++;
6098 * 0 for a fully functional array, 1 or 2 for a degraded array.
6100 mddev->degraded = calc_degraded(conf);
6102 if (has_failed(conf)) {
6103 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6104 " (%d/%d failed)\n",
6105 mdname(mddev), mddev->degraded, conf->raid_disks);
6109 /* device size must be a multiple of chunk size */
6110 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6111 mddev->resync_max_sectors = mddev->dev_sectors;
6113 if (mddev->degraded > dirty_parity_disks &&
6114 mddev->recovery_cp != MaxSector) {
6115 if (mddev->ok_start_degraded)
6117 "md/raid:%s: starting dirty degraded array"
6118 " - data corruption possible.\n",
6122 "md/raid:%s: cannot start dirty degraded array.\n",
6128 if (mddev->degraded == 0)
6129 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6130 " devices, algorithm %d\n", mdname(mddev), conf->level,
6131 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6134 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6135 " out of %d devices, algorithm %d\n",
6136 mdname(mddev), conf->level,
6137 mddev->raid_disks - mddev->degraded,
6138 mddev->raid_disks, mddev->new_layout);
6140 print_raid5_conf(conf);
6142 if (conf->reshape_progress != MaxSector) {
6143 conf->reshape_safe = conf->reshape_progress;
6144 atomic_set(&conf->reshape_stripes, 0);
6145 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6146 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6147 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6148 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6149 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6153 /* Ok, everything is just fine now */
6154 if (mddev->to_remove == &raid5_attrs_group)
6155 mddev->to_remove = NULL;
6156 else if (mddev->kobj.sd &&
6157 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6159 "raid5: failed to create sysfs attributes for %s\n",
6161 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6165 bool discard_supported = true;
6166 /* read-ahead size must cover two whole stripes, which
6167 * is 2 * (datadisks) * chunksize where 'n' is the
6168 * number of raid devices
6170 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6171 int stripe = data_disks *
6172 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6173 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6174 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6176 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
6178 mddev->queue->backing_dev_info.congested_data = mddev;
6179 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
6181 chunk_size = mddev->chunk_sectors << 9;
6182 blk_queue_io_min(mddev->queue, chunk_size);
6183 blk_queue_io_opt(mddev->queue, chunk_size *
6184 (conf->raid_disks - conf->max_degraded));
6185 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6187 * We can only discard a whole stripe. It doesn't make sense to
6188 * discard data disk but write parity disk
6190 stripe = stripe * PAGE_SIZE;
6191 /* Round up to power of 2, as discard handling
6192 * currently assumes that */
6193 while ((stripe-1) & stripe)
6194 stripe = (stripe | (stripe-1)) + 1;
6195 mddev->queue->limits.discard_alignment = stripe;
6196 mddev->queue->limits.discard_granularity = stripe;
6198 * unaligned part of discard request will be ignored, so can't
6199 * guarantee discard_zeroes_data
6201 mddev->queue->limits.discard_zeroes_data = 0;
6203 blk_queue_max_write_same_sectors(mddev->queue, 0);
6205 rdev_for_each(rdev, mddev) {
6206 disk_stack_limits(mddev->gendisk, rdev->bdev,
6207 rdev->data_offset << 9);
6208 disk_stack_limits(mddev->gendisk, rdev->bdev,
6209 rdev->new_data_offset << 9);
6211 * discard_zeroes_data is required, otherwise data
6212 * could be lost. Consider a scenario: discard a stripe
6213 * (the stripe could be inconsistent if
6214 * discard_zeroes_data is 0); write one disk of the
6215 * stripe (the stripe could be inconsistent again
6216 * depending on which disks are used to calculate
6217 * parity); the disk is broken; The stripe data of this
6220 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6221 !bdev_get_queue(rdev->bdev)->
6222 limits.discard_zeroes_data)
6223 discard_supported = false;
6224 /* Unfortunately, discard_zeroes_data is not currently
6225 * a guarantee - just a hint. So we only allow DISCARD
6226 * if the sysadmin has confirmed that only safe devices
6227 * are in use by setting a module parameter.
6229 if (!devices_handle_discard_safely) {
6230 if (discard_supported) {
6231 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6232 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6234 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);
6801 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6803 struct r0conf *raid0_conf = mddev->private;
6806 /* for raid0 takeover only one zone is supported */
6807 if (raid0_conf->nr_strip_zones > 1) {
6808 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6810 return ERR_PTR(-EINVAL);
6813 sectors = raid0_conf->strip_zone[0].zone_end;
6814 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6815 mddev->dev_sectors = sectors;
6816 mddev->new_level = level;
6817 mddev->new_layout = ALGORITHM_PARITY_N;
6818 mddev->new_chunk_sectors = mddev->chunk_sectors;
6819 mddev->raid_disks += 1;
6820 mddev->delta_disks = 1;
6821 /* make sure it will be not marked as dirty */
6822 mddev->recovery_cp = MaxSector;
6824 return setup_conf(mddev);
6827 static void *raid5_takeover_raid1(struct mddev *mddev)
6831 if (mddev->raid_disks != 2 ||
6832 mddev->degraded > 1)
6833 return ERR_PTR(-EINVAL);
6835 /* Should check if there are write-behind devices? */
6837 chunksect = 64*2; /* 64K by default */
6839 /* The array must be an exact multiple of chunksize */
6840 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6843 if ((chunksect<<9) < STRIPE_SIZE)
6844 /* array size does not allow a suitable chunk size */
6845 return ERR_PTR(-EINVAL);
6847 mddev->new_level = 5;
6848 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6849 mddev->new_chunk_sectors = chunksect;
6851 return setup_conf(mddev);
6854 static void *raid5_takeover_raid6(struct mddev *mddev)
6858 switch (mddev->layout) {
6859 case ALGORITHM_LEFT_ASYMMETRIC_6:
6860 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6862 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6863 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6865 case ALGORITHM_LEFT_SYMMETRIC_6:
6866 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6868 case ALGORITHM_RIGHT_SYMMETRIC_6:
6869 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6871 case ALGORITHM_PARITY_0_6:
6872 new_layout = ALGORITHM_PARITY_0;
6874 case ALGORITHM_PARITY_N:
6875 new_layout = ALGORITHM_PARITY_N;
6878 return ERR_PTR(-EINVAL);
6880 mddev->new_level = 5;
6881 mddev->new_layout = new_layout;
6882 mddev->delta_disks = -1;
6883 mddev->raid_disks -= 1;
6884 return setup_conf(mddev);
6887 static int raid5_check_reshape(struct mddev *mddev)
6889 /* For a 2-drive array, the layout and chunk size can be changed
6890 * immediately as not restriping is needed.
6891 * For larger arrays we record the new value - after validation
6892 * to be used by a reshape pass.
6894 struct r5conf *conf = mddev->private;
6895 int new_chunk = mddev->new_chunk_sectors;
6897 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6899 if (new_chunk > 0) {
6900 if (!is_power_of_2(new_chunk))
6902 if (new_chunk < (PAGE_SIZE>>9))
6904 if (mddev->array_sectors & (new_chunk-1))
6905 /* not factor of array size */
6909 /* They look valid */
6911 if (mddev->raid_disks == 2) {
6912 /* can make the change immediately */
6913 if (mddev->new_layout >= 0) {
6914 conf->algorithm = mddev->new_layout;
6915 mddev->layout = mddev->new_layout;
6917 if (new_chunk > 0) {
6918 conf->chunk_sectors = new_chunk ;
6919 mddev->chunk_sectors = new_chunk;
6921 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6922 md_wakeup_thread(mddev->thread);
6924 return check_reshape(mddev);
6927 static int raid6_check_reshape(struct mddev *mddev)
6929 int new_chunk = mddev->new_chunk_sectors;
6931 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6933 if (new_chunk > 0) {
6934 if (!is_power_of_2(new_chunk))
6936 if (new_chunk < (PAGE_SIZE >> 9))
6938 if (mddev->array_sectors & (new_chunk-1))
6939 /* not factor of array size */
6943 /* They look valid */
6944 return check_reshape(mddev);
6947 static void *raid5_takeover(struct mddev *mddev)
6949 /* raid5 can take over:
6950 * raid0 - if there is only one strip zone - make it a raid4 layout
6951 * raid1 - if there are two drives. We need to know the chunk size
6952 * raid4 - trivial - just use a raid4 layout.
6953 * raid6 - Providing it is a *_6 layout
6955 if (mddev->level == 0)
6956 return raid45_takeover_raid0(mddev, 5);
6957 if (mddev->level == 1)
6958 return raid5_takeover_raid1(mddev);
6959 if (mddev->level == 4) {
6960 mddev->new_layout = ALGORITHM_PARITY_N;
6961 mddev->new_level = 5;
6962 return setup_conf(mddev);
6964 if (mddev->level == 6)
6965 return raid5_takeover_raid6(mddev);
6967 return ERR_PTR(-EINVAL);
6970 static void *raid4_takeover(struct mddev *mddev)
6972 /* raid4 can take over:
6973 * raid0 - if there is only one strip zone
6974 * raid5 - if layout is right
6976 if (mddev->level == 0)
6977 return raid45_takeover_raid0(mddev, 4);
6978 if (mddev->level == 5 &&
6979 mddev->layout == ALGORITHM_PARITY_N) {
6980 mddev->new_layout = 0;
6981 mddev->new_level = 4;
6982 return setup_conf(mddev);
6984 return ERR_PTR(-EINVAL);
6987 static struct md_personality raid5_personality;
6989 static void *raid6_takeover(struct mddev *mddev)
6991 /* Currently can only take over a raid5. We map the
6992 * personality to an equivalent raid6 personality
6993 * with the Q block at the end.
6997 if (mddev->pers != &raid5_personality)
6998 return ERR_PTR(-EINVAL);
6999 if (mddev->degraded > 1)
7000 return ERR_PTR(-EINVAL);
7001 if (mddev->raid_disks > 253)
7002 return ERR_PTR(-EINVAL);
7003 if (mddev->raid_disks < 3)
7004 return ERR_PTR(-EINVAL);
7006 switch (mddev->layout) {
7007 case ALGORITHM_LEFT_ASYMMETRIC:
7008 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7010 case ALGORITHM_RIGHT_ASYMMETRIC:
7011 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7013 case ALGORITHM_LEFT_SYMMETRIC:
7014 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7016 case ALGORITHM_RIGHT_SYMMETRIC:
7017 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7019 case ALGORITHM_PARITY_0:
7020 new_layout = ALGORITHM_PARITY_0_6;
7022 case ALGORITHM_PARITY_N:
7023 new_layout = ALGORITHM_PARITY_N;
7026 return ERR_PTR(-EINVAL);
7028 mddev->new_level = 6;
7029 mddev->new_layout = new_layout;
7030 mddev->delta_disks = 1;
7031 mddev->raid_disks += 1;
7032 return setup_conf(mddev);
7035 static struct md_personality raid6_personality =
7039 .owner = THIS_MODULE,
7040 .make_request = make_request,
7044 .error_handler = error,
7045 .hot_add_disk = raid5_add_disk,
7046 .hot_remove_disk= raid5_remove_disk,
7047 .spare_active = raid5_spare_active,
7048 .sync_request = sync_request,
7049 .resize = raid5_resize,
7051 .check_reshape = raid6_check_reshape,
7052 .start_reshape = raid5_start_reshape,
7053 .finish_reshape = raid5_finish_reshape,
7054 .quiesce = raid5_quiesce,
7055 .takeover = raid6_takeover,
7057 static struct md_personality raid5_personality =
7061 .owner = THIS_MODULE,
7062 .make_request = make_request,
7066 .error_handler = error,
7067 .hot_add_disk = raid5_add_disk,
7068 .hot_remove_disk= raid5_remove_disk,
7069 .spare_active = raid5_spare_active,
7070 .sync_request = sync_request,
7071 .resize = raid5_resize,
7073 .check_reshape = raid5_check_reshape,
7074 .start_reshape = raid5_start_reshape,
7075 .finish_reshape = raid5_finish_reshape,
7076 .quiesce = raid5_quiesce,
7077 .takeover = raid5_takeover,
7080 static struct md_personality raid4_personality =
7084 .owner = THIS_MODULE,
7085 .make_request = make_request,
7089 .error_handler = error,
7090 .hot_add_disk = raid5_add_disk,
7091 .hot_remove_disk= raid5_remove_disk,
7092 .spare_active = raid5_spare_active,
7093 .sync_request = sync_request,
7094 .resize = raid5_resize,
7096 .check_reshape = raid5_check_reshape,
7097 .start_reshape = raid5_start_reshape,
7098 .finish_reshape = raid5_finish_reshape,
7099 .quiesce = raid5_quiesce,
7100 .takeover = raid4_takeover,
7103 static int __init raid5_init(void)
7105 raid5_wq = alloc_workqueue("raid5wq",
7106 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7109 register_md_personality(&raid6_personality);
7110 register_md_personality(&raid5_personality);
7111 register_md_personality(&raid4_personality);
7115 static void raid5_exit(void)
7117 unregister_md_personality(&raid6_personality);
7118 unregister_md_personality(&raid5_personality);
7119 unregister_md_personality(&raid4_personality);
7120 destroy_workqueue(raid5_wq);
7123 module_init(raid5_init);
7124 module_exit(raid5_exit);
7125 MODULE_LICENSE("GPL");
7126 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7127 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7128 MODULE_ALIAS("md-raid5");
7129 MODULE_ALIAS("md-raid4");
7130 MODULE_ALIAS("md-level-5");
7131 MODULE_ALIAS("md-level-4");
7132 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7133 MODULE_ALIAS("md-raid6");
7134 MODULE_ALIAS("md-level-6");
7136 /* This used to be two separate modules, they were: */
7137 MODULE_ALIAS("raid5");
7138 MODULE_ALIAS("raid6");