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 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
302 sh->bm_seq - conf->seq_write > 0)
303 list_add_tail(&sh->lru, &conf->bitmap_list);
305 clear_bit(STRIPE_DELAYED, &sh->state);
306 clear_bit(STRIPE_BIT_DELAY, &sh->state);
307 if (conf->worker_cnt_per_group == 0) {
308 list_add_tail(&sh->lru, &conf->handle_list);
310 raid5_wakeup_stripe_thread(sh);
314 md_wakeup_thread(conf->mddev->thread);
316 BUG_ON(stripe_operations_active(sh));
317 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
318 if (atomic_dec_return(&conf->preread_active_stripes)
320 md_wakeup_thread(conf->mddev->thread);
321 atomic_dec(&conf->active_stripes);
322 if (!test_bit(STRIPE_EXPANDING, &sh->state))
323 list_add_tail(&sh->lru, temp_inactive_list);
327 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
328 struct list_head *temp_inactive_list)
330 if (atomic_dec_and_test(&sh->count))
331 do_release_stripe(conf, sh, temp_inactive_list);
335 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
337 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
338 * given time. Adding stripes only takes device lock, while deleting stripes
339 * only takes hash lock.
341 static void release_inactive_stripe_list(struct r5conf *conf,
342 struct list_head *temp_inactive_list,
346 bool do_wakeup = false;
349 if (hash == NR_STRIPE_HASH_LOCKS) {
350 size = NR_STRIPE_HASH_LOCKS;
351 hash = NR_STRIPE_HASH_LOCKS - 1;
355 struct list_head *list = &temp_inactive_list[size - 1];
358 * We don't hold any lock here yet, get_active_stripe() might
359 * remove stripes from the list
361 if (!list_empty_careful(list)) {
362 spin_lock_irqsave(conf->hash_locks + hash, flags);
363 if (list_empty(conf->inactive_list + hash) &&
365 atomic_dec(&conf->empty_inactive_list_nr);
366 list_splice_tail_init(list, conf->inactive_list + hash);
368 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
375 wake_up(&conf->wait_for_stripe);
376 if (conf->retry_read_aligned)
377 md_wakeup_thread(conf->mddev->thread);
381 /* should hold conf->device_lock already */
382 static int release_stripe_list(struct r5conf *conf,
383 struct list_head *temp_inactive_list)
385 struct stripe_head *sh;
387 struct llist_node *head;
389 head = llist_del_all(&conf->released_stripes);
390 head = llist_reverse_order(head);
394 sh = llist_entry(head, struct stripe_head, release_list);
395 head = llist_next(head);
396 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
398 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
400 * Don't worry the bit is set here, because if the bit is set
401 * again, the count is always > 1. This is true for
402 * STRIPE_ON_UNPLUG_LIST bit too.
404 hash = sh->hash_lock_index;
405 __release_stripe(conf, sh, &temp_inactive_list[hash]);
412 static void release_stripe(struct stripe_head *sh)
414 struct r5conf *conf = sh->raid_conf;
416 struct list_head list;
420 /* Avoid release_list until the last reference.
422 if (atomic_add_unless(&sh->count, -1, 1))
425 if (unlikely(!conf->mddev->thread) ||
426 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
428 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
430 md_wakeup_thread(conf->mddev->thread);
433 local_irq_save(flags);
434 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
435 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
436 INIT_LIST_HEAD(&list);
437 hash = sh->hash_lock_index;
438 do_release_stripe(conf, sh, &list);
439 spin_unlock(&conf->device_lock);
440 release_inactive_stripe_list(conf, &list, hash);
442 local_irq_restore(flags);
445 static inline void remove_hash(struct stripe_head *sh)
447 pr_debug("remove_hash(), stripe %llu\n",
448 (unsigned long long)sh->sector);
450 hlist_del_init(&sh->hash);
453 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
455 struct hlist_head *hp = stripe_hash(conf, sh->sector);
457 pr_debug("insert_hash(), stripe %llu\n",
458 (unsigned long long)sh->sector);
460 hlist_add_head(&sh->hash, hp);
463 /* find an idle stripe, make sure it is unhashed, and return it. */
464 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
466 struct stripe_head *sh = NULL;
467 struct list_head *first;
469 if (list_empty(conf->inactive_list + hash))
471 first = (conf->inactive_list + hash)->next;
472 sh = list_entry(first, struct stripe_head, lru);
473 list_del_init(first);
475 atomic_inc(&conf->active_stripes);
476 BUG_ON(hash != sh->hash_lock_index);
477 if (list_empty(conf->inactive_list + hash))
478 atomic_inc(&conf->empty_inactive_list_nr);
483 static void shrink_buffers(struct stripe_head *sh)
487 int num = sh->raid_conf->pool_size;
489 for (i = 0; i < num ; i++) {
490 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
494 sh->dev[i].page = NULL;
499 static int grow_buffers(struct stripe_head *sh)
502 int num = sh->raid_conf->pool_size;
504 for (i = 0; i < num; i++) {
507 if (!(page = alloc_page(GFP_KERNEL))) {
510 sh->dev[i].page = page;
511 sh->dev[i].orig_page = page;
516 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
517 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
518 struct stripe_head *sh);
520 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
522 struct r5conf *conf = sh->raid_conf;
525 BUG_ON(atomic_read(&sh->count) != 0);
526 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
527 BUG_ON(stripe_operations_active(sh));
529 pr_debug("init_stripe called, stripe %llu\n",
530 (unsigned long long)sector);
532 seq = read_seqcount_begin(&conf->gen_lock);
533 sh->generation = conf->generation - previous;
534 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
536 stripe_set_idx(sector, conf, previous, sh);
539 for (i = sh->disks; i--; ) {
540 struct r5dev *dev = &sh->dev[i];
542 if (dev->toread || dev->read || dev->towrite || dev->written ||
543 test_bit(R5_LOCKED, &dev->flags)) {
544 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
545 (unsigned long long)sh->sector, i, dev->toread,
546 dev->read, dev->towrite, dev->written,
547 test_bit(R5_LOCKED, &dev->flags));
551 raid5_build_block(sh, i, previous);
553 if (read_seqcount_retry(&conf->gen_lock, seq))
555 insert_hash(conf, sh);
556 sh->cpu = smp_processor_id();
559 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
562 struct stripe_head *sh;
564 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
565 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
566 if (sh->sector == sector && sh->generation == generation)
568 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
573 * Need to check if array has failed when deciding whether to:
575 * - remove non-faulty devices
578 * This determination is simple when no reshape is happening.
579 * However if there is a reshape, we need to carefully check
580 * both the before and after sections.
581 * This is because some failed devices may only affect one
582 * of the two sections, and some non-in_sync devices may
583 * be insync in the section most affected by failed devices.
585 static int calc_degraded(struct r5conf *conf)
587 int degraded, degraded2;
592 for (i = 0; i < conf->previous_raid_disks; i++) {
593 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
594 if (rdev && test_bit(Faulty, &rdev->flags))
595 rdev = rcu_dereference(conf->disks[i].replacement);
596 if (!rdev || test_bit(Faulty, &rdev->flags))
598 else if (test_bit(In_sync, &rdev->flags))
601 /* not in-sync or faulty.
602 * If the reshape increases the number of devices,
603 * this is being recovered by the reshape, so
604 * this 'previous' section is not in_sync.
605 * If the number of devices is being reduced however,
606 * the device can only be part of the array if
607 * we are reverting a reshape, so this section will
610 if (conf->raid_disks >= conf->previous_raid_disks)
614 if (conf->raid_disks == conf->previous_raid_disks)
618 for (i = 0; i < conf->raid_disks; i++) {
619 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
620 if (rdev && test_bit(Faulty, &rdev->flags))
621 rdev = rcu_dereference(conf->disks[i].replacement);
622 if (!rdev || test_bit(Faulty, &rdev->flags))
624 else if (test_bit(In_sync, &rdev->flags))
627 /* not in-sync or faulty.
628 * If reshape increases the number of devices, this
629 * section has already been recovered, else it
630 * almost certainly hasn't.
632 if (conf->raid_disks <= conf->previous_raid_disks)
636 if (degraded2 > degraded)
641 static int has_failed(struct r5conf *conf)
645 if (conf->mddev->reshape_position == MaxSector)
646 return conf->mddev->degraded > conf->max_degraded;
648 degraded = calc_degraded(conf);
649 if (degraded > conf->max_degraded)
654 static struct stripe_head *
655 get_active_stripe(struct r5conf *conf, sector_t sector,
656 int previous, int noblock, int noquiesce)
658 struct stripe_head *sh;
659 int hash = stripe_hash_locks_hash(sector);
661 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
663 spin_lock_irq(conf->hash_locks + hash);
666 wait_event_lock_irq(conf->wait_for_stripe,
667 conf->quiesce == 0 || noquiesce,
668 *(conf->hash_locks + hash));
669 sh = __find_stripe(conf, sector, conf->generation - previous);
671 if (!conf->inactive_blocked)
672 sh = get_free_stripe(conf, hash);
673 if (noblock && sh == NULL)
676 conf->inactive_blocked = 1;
678 conf->wait_for_stripe,
679 !list_empty(conf->inactive_list + hash) &&
680 (atomic_read(&conf->active_stripes)
681 < (conf->max_nr_stripes * 3 / 4)
682 || !conf->inactive_blocked),
683 *(conf->hash_locks + hash));
684 conf->inactive_blocked = 0;
686 init_stripe(sh, sector, previous);
687 atomic_inc(&sh->count);
689 } else if (!atomic_inc_not_zero(&sh->count)) {
690 spin_lock(&conf->device_lock);
691 if (!atomic_read(&sh->count)) {
692 if (!test_bit(STRIPE_HANDLE, &sh->state))
693 atomic_inc(&conf->active_stripes);
694 BUG_ON(list_empty(&sh->lru) &&
695 !test_bit(STRIPE_EXPANDING, &sh->state));
696 list_del_init(&sh->lru);
698 sh->group->stripes_cnt--;
702 atomic_inc(&sh->count);
703 spin_unlock(&conf->device_lock);
705 } while (sh == NULL);
707 spin_unlock_irq(conf->hash_locks + hash);
711 /* Determine if 'data_offset' or 'new_data_offset' should be used
712 * in this stripe_head.
714 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
716 sector_t progress = conf->reshape_progress;
717 /* Need a memory barrier to make sure we see the value
718 * of conf->generation, or ->data_offset that was set before
719 * reshape_progress was updated.
722 if (progress == MaxSector)
724 if (sh->generation == conf->generation - 1)
726 /* We are in a reshape, and this is a new-generation stripe,
727 * so use new_data_offset.
733 raid5_end_read_request(struct bio *bi, int error);
735 raid5_end_write_request(struct bio *bi, int error);
737 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
739 struct r5conf *conf = sh->raid_conf;
740 int i, disks = sh->disks;
744 for (i = disks; i--; ) {
746 int replace_only = 0;
747 struct bio *bi, *rbi;
748 struct md_rdev *rdev, *rrdev = NULL;
749 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
750 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
754 if (test_bit(R5_Discard, &sh->dev[i].flags))
756 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
758 else if (test_and_clear_bit(R5_WantReplace,
759 &sh->dev[i].flags)) {
764 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
767 bi = &sh->dev[i].req;
768 rbi = &sh->dev[i].rreq; /* For writing to replacement */
771 rrdev = rcu_dereference(conf->disks[i].replacement);
772 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
773 rdev = rcu_dereference(conf->disks[i].rdev);
782 /* We raced and saw duplicates */
785 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
790 if (rdev && test_bit(Faulty, &rdev->flags))
793 atomic_inc(&rdev->nr_pending);
794 if (rrdev && test_bit(Faulty, &rrdev->flags))
797 atomic_inc(&rrdev->nr_pending);
800 /* We have already checked bad blocks for reads. Now
801 * need to check for writes. We never accept write errors
802 * on the replacement, so we don't to check rrdev.
804 while ((rw & WRITE) && rdev &&
805 test_bit(WriteErrorSeen, &rdev->flags)) {
808 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
809 &first_bad, &bad_sectors);
814 set_bit(BlockedBadBlocks, &rdev->flags);
815 if (!conf->mddev->external &&
816 conf->mddev->flags) {
817 /* It is very unlikely, but we might
818 * still need to write out the
819 * bad block log - better give it
821 md_check_recovery(conf->mddev);
824 * Because md_wait_for_blocked_rdev
825 * will dec nr_pending, we must
826 * increment it first.
828 atomic_inc(&rdev->nr_pending);
829 md_wait_for_blocked_rdev(rdev, conf->mddev);
831 /* Acknowledged bad block - skip the write */
832 rdev_dec_pending(rdev, conf->mddev);
838 if (s->syncing || s->expanding || s->expanded
840 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
842 set_bit(STRIPE_IO_STARTED, &sh->state);
845 bi->bi_bdev = rdev->bdev;
847 bi->bi_end_io = (rw & WRITE)
848 ? raid5_end_write_request
849 : raid5_end_read_request;
852 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
853 __func__, (unsigned long long)sh->sector,
855 atomic_inc(&sh->count);
856 if (use_new_offset(conf, sh))
857 bi->bi_iter.bi_sector = (sh->sector
858 + rdev->new_data_offset);
860 bi->bi_iter.bi_sector = (sh->sector
861 + rdev->data_offset);
862 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
863 bi->bi_rw |= REQ_NOMERGE;
865 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
866 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
867 sh->dev[i].vec.bv_page = sh->dev[i].page;
869 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
870 bi->bi_io_vec[0].bv_offset = 0;
871 bi->bi_iter.bi_size = STRIPE_SIZE;
873 * If this is discard request, set bi_vcnt 0. We don't
874 * want to confuse SCSI because SCSI will replace payload
876 if (rw & REQ_DISCARD)
879 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
881 if (conf->mddev->gendisk)
882 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
883 bi, disk_devt(conf->mddev->gendisk),
885 generic_make_request(bi);
888 if (s->syncing || s->expanding || s->expanded
890 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
892 set_bit(STRIPE_IO_STARTED, &sh->state);
895 rbi->bi_bdev = rrdev->bdev;
897 BUG_ON(!(rw & WRITE));
898 rbi->bi_end_io = raid5_end_write_request;
899 rbi->bi_private = sh;
901 pr_debug("%s: for %llu schedule op %ld on "
902 "replacement disc %d\n",
903 __func__, (unsigned long long)sh->sector,
905 atomic_inc(&sh->count);
906 if (use_new_offset(conf, sh))
907 rbi->bi_iter.bi_sector = (sh->sector
908 + rrdev->new_data_offset);
910 rbi->bi_iter.bi_sector = (sh->sector
911 + rrdev->data_offset);
912 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
913 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
914 sh->dev[i].rvec.bv_page = sh->dev[i].page;
916 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
917 rbi->bi_io_vec[0].bv_offset = 0;
918 rbi->bi_iter.bi_size = STRIPE_SIZE;
920 * If this is discard request, set bi_vcnt 0. We don't
921 * want to confuse SCSI because SCSI will replace payload
923 if (rw & REQ_DISCARD)
925 if (conf->mddev->gendisk)
926 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
927 rbi, disk_devt(conf->mddev->gendisk),
929 generic_make_request(rbi);
931 if (!rdev && !rrdev) {
933 set_bit(STRIPE_DEGRADED, &sh->state);
934 pr_debug("skip op %ld on disc %d for sector %llu\n",
935 bi->bi_rw, i, (unsigned long long)sh->sector);
936 clear_bit(R5_LOCKED, &sh->dev[i].flags);
937 set_bit(STRIPE_HANDLE, &sh->state);
942 static struct dma_async_tx_descriptor *
943 async_copy_data(int frombio, struct bio *bio, struct page **page,
944 sector_t sector, struct dma_async_tx_descriptor *tx,
945 struct stripe_head *sh)
948 struct bvec_iter iter;
949 struct page *bio_page;
951 struct async_submit_ctl submit;
952 enum async_tx_flags flags = 0;
954 if (bio->bi_iter.bi_sector >= sector)
955 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
957 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
960 flags |= ASYNC_TX_FENCE;
961 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
963 bio_for_each_segment(bvl, bio, iter) {
964 int len = bvl.bv_len;
968 if (page_offset < 0) {
969 b_offset = -page_offset;
970 page_offset += b_offset;
974 if (len > 0 && page_offset + len > STRIPE_SIZE)
975 clen = STRIPE_SIZE - page_offset;
980 b_offset += bvl.bv_offset;
981 bio_page = bvl.bv_page;
983 if (sh->raid_conf->skip_copy &&
984 b_offset == 0 && page_offset == 0 &&
988 tx = async_memcpy(*page, bio_page, page_offset,
989 b_offset, clen, &submit);
991 tx = async_memcpy(bio_page, *page, b_offset,
992 page_offset, clen, &submit);
994 /* chain the operations */
995 submit.depend_tx = tx;
997 if (clen < len) /* hit end of page */
1005 static void ops_complete_biofill(void *stripe_head_ref)
1007 struct stripe_head *sh = stripe_head_ref;
1008 struct bio *return_bi = NULL;
1011 pr_debug("%s: stripe %llu\n", __func__,
1012 (unsigned long long)sh->sector);
1014 /* clear completed biofills */
1015 for (i = sh->disks; i--; ) {
1016 struct r5dev *dev = &sh->dev[i];
1018 /* acknowledge completion of a biofill operation */
1019 /* and check if we need to reply to a read request,
1020 * new R5_Wantfill requests are held off until
1021 * !STRIPE_BIOFILL_RUN
1023 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1024 struct bio *rbi, *rbi2;
1029 while (rbi && rbi->bi_iter.bi_sector <
1030 dev->sector + STRIPE_SECTORS) {
1031 rbi2 = r5_next_bio(rbi, dev->sector);
1032 if (!raid5_dec_bi_active_stripes(rbi)) {
1033 rbi->bi_next = return_bi;
1040 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1042 return_io(return_bi);
1044 set_bit(STRIPE_HANDLE, &sh->state);
1048 static void ops_run_biofill(struct stripe_head *sh)
1050 struct dma_async_tx_descriptor *tx = NULL;
1051 struct async_submit_ctl submit;
1054 pr_debug("%s: stripe %llu\n", __func__,
1055 (unsigned long long)sh->sector);
1057 for (i = sh->disks; i--; ) {
1058 struct r5dev *dev = &sh->dev[i];
1059 if (test_bit(R5_Wantfill, &dev->flags)) {
1061 spin_lock_irq(&sh->stripe_lock);
1062 dev->read = rbi = dev->toread;
1064 spin_unlock_irq(&sh->stripe_lock);
1065 while (rbi && rbi->bi_iter.bi_sector <
1066 dev->sector + STRIPE_SECTORS) {
1067 tx = async_copy_data(0, rbi, &dev->page,
1068 dev->sector, tx, sh);
1069 rbi = r5_next_bio(rbi, dev->sector);
1074 atomic_inc(&sh->count);
1075 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1076 async_trigger_callback(&submit);
1079 static void mark_target_uptodate(struct stripe_head *sh, int target)
1086 tgt = &sh->dev[target];
1087 set_bit(R5_UPTODATE, &tgt->flags);
1088 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1089 clear_bit(R5_Wantcompute, &tgt->flags);
1092 static void ops_complete_compute(void *stripe_head_ref)
1094 struct stripe_head *sh = stripe_head_ref;
1096 pr_debug("%s: stripe %llu\n", __func__,
1097 (unsigned long long)sh->sector);
1099 /* mark the computed target(s) as uptodate */
1100 mark_target_uptodate(sh, sh->ops.target);
1101 mark_target_uptodate(sh, sh->ops.target2);
1103 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1104 if (sh->check_state == check_state_compute_run)
1105 sh->check_state = check_state_compute_result;
1106 set_bit(STRIPE_HANDLE, &sh->state);
1110 /* return a pointer to the address conversion region of the scribble buffer */
1111 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1112 struct raid5_percpu *percpu)
1114 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
1117 static struct dma_async_tx_descriptor *
1118 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1120 int disks = sh->disks;
1121 struct page **xor_srcs = percpu->scribble;
1122 int target = sh->ops.target;
1123 struct r5dev *tgt = &sh->dev[target];
1124 struct page *xor_dest = tgt->page;
1126 struct dma_async_tx_descriptor *tx;
1127 struct async_submit_ctl submit;
1130 pr_debug("%s: stripe %llu block: %d\n",
1131 __func__, (unsigned long long)sh->sector, target);
1132 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1134 for (i = disks; i--; )
1136 xor_srcs[count++] = sh->dev[i].page;
1138 atomic_inc(&sh->count);
1140 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1141 ops_complete_compute, sh, to_addr_conv(sh, percpu));
1142 if (unlikely(count == 1))
1143 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1145 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1150 /* set_syndrome_sources - populate source buffers for gen_syndrome
1151 * @srcs - (struct page *) array of size sh->disks
1152 * @sh - stripe_head to parse
1154 * Populates srcs in proper layout order for the stripe and returns the
1155 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1156 * destination buffer is recorded in srcs[count] and the Q destination
1157 * is recorded in srcs[count+1]].
1159 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
1161 int disks = sh->disks;
1162 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1163 int d0_idx = raid6_d0(sh);
1167 for (i = 0; i < disks; i++)
1173 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1175 srcs[slot] = sh->dev[i].page;
1176 i = raid6_next_disk(i, disks);
1177 } while (i != d0_idx);
1179 return syndrome_disks;
1182 static struct dma_async_tx_descriptor *
1183 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1185 int disks = sh->disks;
1186 struct page **blocks = percpu->scribble;
1188 int qd_idx = sh->qd_idx;
1189 struct dma_async_tx_descriptor *tx;
1190 struct async_submit_ctl submit;
1196 if (sh->ops.target < 0)
1197 target = sh->ops.target2;
1198 else if (sh->ops.target2 < 0)
1199 target = sh->ops.target;
1201 /* we should only have one valid target */
1204 pr_debug("%s: stripe %llu block: %d\n",
1205 __func__, (unsigned long long)sh->sector, target);
1207 tgt = &sh->dev[target];
1208 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1211 atomic_inc(&sh->count);
1213 if (target == qd_idx) {
1214 count = set_syndrome_sources(blocks, sh);
1215 blocks[count] = NULL; /* regenerating p is not necessary */
1216 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1217 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1218 ops_complete_compute, sh,
1219 to_addr_conv(sh, percpu));
1220 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1222 /* Compute any data- or p-drive using XOR */
1224 for (i = disks; i-- ; ) {
1225 if (i == target || i == qd_idx)
1227 blocks[count++] = sh->dev[i].page;
1230 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1231 NULL, ops_complete_compute, sh,
1232 to_addr_conv(sh, percpu));
1233 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1239 static struct dma_async_tx_descriptor *
1240 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1242 int i, count, disks = sh->disks;
1243 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1244 int d0_idx = raid6_d0(sh);
1245 int faila = -1, failb = -1;
1246 int target = sh->ops.target;
1247 int target2 = sh->ops.target2;
1248 struct r5dev *tgt = &sh->dev[target];
1249 struct r5dev *tgt2 = &sh->dev[target2];
1250 struct dma_async_tx_descriptor *tx;
1251 struct page **blocks = percpu->scribble;
1252 struct async_submit_ctl submit;
1254 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1255 __func__, (unsigned long long)sh->sector, target, target2);
1256 BUG_ON(target < 0 || target2 < 0);
1257 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1258 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1260 /* we need to open-code set_syndrome_sources to handle the
1261 * slot number conversion for 'faila' and 'failb'
1263 for (i = 0; i < disks ; i++)
1268 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1270 blocks[slot] = sh->dev[i].page;
1276 i = raid6_next_disk(i, disks);
1277 } while (i != d0_idx);
1279 BUG_ON(faila == failb);
1282 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1283 __func__, (unsigned long long)sh->sector, faila, failb);
1285 atomic_inc(&sh->count);
1287 if (failb == syndrome_disks+1) {
1288 /* Q disk is one of the missing disks */
1289 if (faila == syndrome_disks) {
1290 /* Missing P+Q, just recompute */
1291 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1292 ops_complete_compute, sh,
1293 to_addr_conv(sh, percpu));
1294 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1295 STRIPE_SIZE, &submit);
1299 int qd_idx = sh->qd_idx;
1301 /* Missing D+Q: recompute D from P, then recompute Q */
1302 if (target == qd_idx)
1303 data_target = target2;
1305 data_target = target;
1308 for (i = disks; i-- ; ) {
1309 if (i == data_target || i == qd_idx)
1311 blocks[count++] = sh->dev[i].page;
1313 dest = sh->dev[data_target].page;
1314 init_async_submit(&submit,
1315 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1317 to_addr_conv(sh, percpu));
1318 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1321 count = set_syndrome_sources(blocks, sh);
1322 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1323 ops_complete_compute, sh,
1324 to_addr_conv(sh, percpu));
1325 return async_gen_syndrome(blocks, 0, count+2,
1326 STRIPE_SIZE, &submit);
1329 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1330 ops_complete_compute, sh,
1331 to_addr_conv(sh, percpu));
1332 if (failb == syndrome_disks) {
1333 /* We're missing D+P. */
1334 return async_raid6_datap_recov(syndrome_disks+2,
1338 /* We're missing D+D. */
1339 return async_raid6_2data_recov(syndrome_disks+2,
1340 STRIPE_SIZE, faila, failb,
1346 static void ops_complete_prexor(void *stripe_head_ref)
1348 struct stripe_head *sh = stripe_head_ref;
1350 pr_debug("%s: stripe %llu\n", __func__,
1351 (unsigned long long)sh->sector);
1354 static struct dma_async_tx_descriptor *
1355 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1356 struct dma_async_tx_descriptor *tx)
1358 int disks = sh->disks;
1359 struct page **xor_srcs = percpu->scribble;
1360 int count = 0, pd_idx = sh->pd_idx, i;
1361 struct async_submit_ctl submit;
1363 /* existing parity data subtracted */
1364 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1366 pr_debug("%s: stripe %llu\n", __func__,
1367 (unsigned long long)sh->sector);
1369 for (i = disks; i--; ) {
1370 struct r5dev *dev = &sh->dev[i];
1371 /* Only process blocks that are known to be uptodate */
1372 if (test_bit(R5_Wantdrain, &dev->flags))
1373 xor_srcs[count++] = dev->page;
1376 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1377 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1378 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1383 static struct dma_async_tx_descriptor *
1384 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1386 int disks = sh->disks;
1389 pr_debug("%s: stripe %llu\n", __func__,
1390 (unsigned long long)sh->sector);
1392 for (i = disks; i--; ) {
1393 struct r5dev *dev = &sh->dev[i];
1396 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1399 spin_lock_irq(&sh->stripe_lock);
1400 chosen = dev->towrite;
1401 dev->towrite = NULL;
1402 BUG_ON(dev->written);
1403 wbi = dev->written = chosen;
1404 spin_unlock_irq(&sh->stripe_lock);
1405 WARN_ON(dev->page != dev->orig_page);
1407 while (wbi && wbi->bi_iter.bi_sector <
1408 dev->sector + STRIPE_SECTORS) {
1409 if (wbi->bi_rw & REQ_FUA)
1410 set_bit(R5_WantFUA, &dev->flags);
1411 if (wbi->bi_rw & REQ_SYNC)
1412 set_bit(R5_SyncIO, &dev->flags);
1413 if (wbi->bi_rw & REQ_DISCARD)
1414 set_bit(R5_Discard, &dev->flags);
1416 tx = async_copy_data(1, wbi, &dev->page,
1417 dev->sector, tx, sh);
1418 if (dev->page != dev->orig_page) {
1419 set_bit(R5_SkipCopy, &dev->flags);
1420 clear_bit(R5_UPTODATE, &dev->flags);
1421 clear_bit(R5_OVERWRITE, &dev->flags);
1424 wbi = r5_next_bio(wbi, dev->sector);
1432 static void ops_complete_reconstruct(void *stripe_head_ref)
1434 struct stripe_head *sh = stripe_head_ref;
1435 int disks = sh->disks;
1436 int pd_idx = sh->pd_idx;
1437 int qd_idx = sh->qd_idx;
1439 bool fua = false, sync = false, discard = false;
1441 pr_debug("%s: stripe %llu\n", __func__,
1442 (unsigned long long)sh->sector);
1444 for (i = disks; i--; ) {
1445 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1446 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1447 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1450 for (i = disks; i--; ) {
1451 struct r5dev *dev = &sh->dev[i];
1453 if (dev->written || i == pd_idx || i == qd_idx) {
1454 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1455 set_bit(R5_UPTODATE, &dev->flags);
1457 set_bit(R5_WantFUA, &dev->flags);
1459 set_bit(R5_SyncIO, &dev->flags);
1463 if (sh->reconstruct_state == reconstruct_state_drain_run)
1464 sh->reconstruct_state = reconstruct_state_drain_result;
1465 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1466 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1468 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1469 sh->reconstruct_state = reconstruct_state_result;
1472 set_bit(STRIPE_HANDLE, &sh->state);
1477 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1478 struct dma_async_tx_descriptor *tx)
1480 int disks = sh->disks;
1481 struct page **xor_srcs = percpu->scribble;
1482 struct async_submit_ctl submit;
1483 int count = 0, pd_idx = sh->pd_idx, i;
1484 struct page *xor_dest;
1486 unsigned long flags;
1488 pr_debug("%s: stripe %llu\n", __func__,
1489 (unsigned long long)sh->sector);
1491 for (i = 0; i < sh->disks; i++) {
1494 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1497 if (i >= sh->disks) {
1498 atomic_inc(&sh->count);
1499 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1500 ops_complete_reconstruct(sh);
1503 /* check if prexor is active which means only process blocks
1504 * that are part of a read-modify-write (written)
1506 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1508 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1509 for (i = disks; i--; ) {
1510 struct r5dev *dev = &sh->dev[i];
1512 xor_srcs[count++] = dev->page;
1515 xor_dest = sh->dev[pd_idx].page;
1516 for (i = disks; i--; ) {
1517 struct r5dev *dev = &sh->dev[i];
1519 xor_srcs[count++] = dev->page;
1523 /* 1/ if we prexor'd then the dest is reused as a source
1524 * 2/ if we did not prexor then we are redoing the parity
1525 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1526 * for the synchronous xor case
1528 flags = ASYNC_TX_ACK |
1529 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1531 atomic_inc(&sh->count);
1533 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1534 to_addr_conv(sh, percpu));
1535 if (unlikely(count == 1))
1536 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1538 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1542 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1543 struct dma_async_tx_descriptor *tx)
1545 struct async_submit_ctl submit;
1546 struct page **blocks = percpu->scribble;
1549 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1551 for (i = 0; i < sh->disks; i++) {
1552 if (sh->pd_idx == i || sh->qd_idx == i)
1554 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1557 if (i >= sh->disks) {
1558 atomic_inc(&sh->count);
1559 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1560 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1561 ops_complete_reconstruct(sh);
1565 count = set_syndrome_sources(blocks, sh);
1567 atomic_inc(&sh->count);
1569 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1570 sh, to_addr_conv(sh, percpu));
1571 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1574 static void ops_complete_check(void *stripe_head_ref)
1576 struct stripe_head *sh = stripe_head_ref;
1578 pr_debug("%s: stripe %llu\n", __func__,
1579 (unsigned long long)sh->sector);
1581 sh->check_state = check_state_check_result;
1582 set_bit(STRIPE_HANDLE, &sh->state);
1586 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1588 int disks = sh->disks;
1589 int pd_idx = sh->pd_idx;
1590 int qd_idx = sh->qd_idx;
1591 struct page *xor_dest;
1592 struct page **xor_srcs = percpu->scribble;
1593 struct dma_async_tx_descriptor *tx;
1594 struct async_submit_ctl submit;
1598 pr_debug("%s: stripe %llu\n", __func__,
1599 (unsigned long long)sh->sector);
1602 xor_dest = sh->dev[pd_idx].page;
1603 xor_srcs[count++] = xor_dest;
1604 for (i = disks; i--; ) {
1605 if (i == pd_idx || i == qd_idx)
1607 xor_srcs[count++] = sh->dev[i].page;
1610 init_async_submit(&submit, 0, NULL, NULL, NULL,
1611 to_addr_conv(sh, percpu));
1612 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1613 &sh->ops.zero_sum_result, &submit);
1615 atomic_inc(&sh->count);
1616 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1617 tx = async_trigger_callback(&submit);
1620 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1622 struct page **srcs = percpu->scribble;
1623 struct async_submit_ctl submit;
1626 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1627 (unsigned long long)sh->sector, checkp);
1629 count = set_syndrome_sources(srcs, sh);
1633 atomic_inc(&sh->count);
1634 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1635 sh, to_addr_conv(sh, percpu));
1636 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1637 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1640 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1642 int overlap_clear = 0, i, disks = sh->disks;
1643 struct dma_async_tx_descriptor *tx = NULL;
1644 struct r5conf *conf = sh->raid_conf;
1645 int level = conf->level;
1646 struct raid5_percpu *percpu;
1650 percpu = per_cpu_ptr(conf->percpu, cpu);
1651 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1652 ops_run_biofill(sh);
1656 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1658 tx = ops_run_compute5(sh, percpu);
1660 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1661 tx = ops_run_compute6_1(sh, percpu);
1663 tx = ops_run_compute6_2(sh, percpu);
1665 /* terminate the chain if reconstruct is not set to be run */
1666 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1670 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1671 tx = ops_run_prexor(sh, percpu, tx);
1673 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1674 tx = ops_run_biodrain(sh, tx);
1678 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1680 ops_run_reconstruct5(sh, percpu, tx);
1682 ops_run_reconstruct6(sh, percpu, tx);
1685 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1686 if (sh->check_state == check_state_run)
1687 ops_run_check_p(sh, percpu);
1688 else if (sh->check_state == check_state_run_q)
1689 ops_run_check_pq(sh, percpu, 0);
1690 else if (sh->check_state == check_state_run_pq)
1691 ops_run_check_pq(sh, percpu, 1);
1697 for (i = disks; i--; ) {
1698 struct r5dev *dev = &sh->dev[i];
1699 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1700 wake_up(&sh->raid_conf->wait_for_overlap);
1705 static int grow_one_stripe(struct r5conf *conf, int hash)
1707 struct stripe_head *sh;
1708 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1712 sh->raid_conf = conf;
1714 spin_lock_init(&sh->stripe_lock);
1716 if (grow_buffers(sh)) {
1718 kmem_cache_free(conf->slab_cache, sh);
1721 sh->hash_lock_index = hash;
1722 /* we just created an active stripe so... */
1723 atomic_set(&sh->count, 1);
1724 atomic_inc(&conf->active_stripes);
1725 INIT_LIST_HEAD(&sh->lru);
1730 static int grow_stripes(struct r5conf *conf, int num)
1732 struct kmem_cache *sc;
1733 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1736 if (conf->mddev->gendisk)
1737 sprintf(conf->cache_name[0],
1738 "raid%d-%s", conf->level, mdname(conf->mddev));
1740 sprintf(conf->cache_name[0],
1741 "raid%d-%p", conf->level, conf->mddev);
1742 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1744 conf->active_name = 0;
1745 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1746 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1750 conf->slab_cache = sc;
1751 conf->pool_size = devs;
1752 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
1754 if (!grow_one_stripe(conf, hash))
1756 conf->max_nr_stripes++;
1757 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
1763 * scribble_len - return the required size of the scribble region
1764 * @num - total number of disks in the array
1766 * The size must be enough to contain:
1767 * 1/ a struct page pointer for each device in the array +2
1768 * 2/ room to convert each entry in (1) to its corresponding dma
1769 * (dma_map_page()) or page (page_address()) address.
1771 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1772 * calculate over all devices (not just the data blocks), using zeros in place
1773 * of the P and Q blocks.
1775 static size_t scribble_len(int num)
1779 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1784 static int resize_stripes(struct r5conf *conf, int newsize)
1786 /* Make all the stripes able to hold 'newsize' devices.
1787 * New slots in each stripe get 'page' set to a new page.
1789 * This happens in stages:
1790 * 1/ create a new kmem_cache and allocate the required number of
1792 * 2/ gather all the old stripe_heads and transfer the pages across
1793 * to the new stripe_heads. This will have the side effect of
1794 * freezing the array as once all stripe_heads have been collected,
1795 * no IO will be possible. Old stripe heads are freed once their
1796 * pages have been transferred over, and the old kmem_cache is
1797 * freed when all stripes are done.
1798 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1799 * we simple return a failre status - no need to clean anything up.
1800 * 4/ allocate new pages for the new slots in the new stripe_heads.
1801 * If this fails, we don't bother trying the shrink the
1802 * stripe_heads down again, we just leave them as they are.
1803 * As each stripe_head is processed the new one is released into
1806 * Once step2 is started, we cannot afford to wait for a write,
1807 * so we use GFP_NOIO allocations.
1809 struct stripe_head *osh, *nsh;
1810 LIST_HEAD(newstripes);
1811 struct disk_info *ndisks;
1814 struct kmem_cache *sc;
1818 if (newsize <= conf->pool_size)
1819 return 0; /* never bother to shrink */
1821 err = md_allow_write(conf->mddev);
1826 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1827 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1832 for (i = conf->max_nr_stripes; i; i--) {
1833 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1837 nsh->raid_conf = conf;
1838 spin_lock_init(&nsh->stripe_lock);
1840 list_add(&nsh->lru, &newstripes);
1843 /* didn't get enough, give up */
1844 while (!list_empty(&newstripes)) {
1845 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1846 list_del(&nsh->lru);
1847 kmem_cache_free(sc, nsh);
1849 kmem_cache_destroy(sc);
1852 /* Step 2 - Must use GFP_NOIO now.
1853 * OK, we have enough stripes, start collecting inactive
1854 * stripes and copying them over
1858 list_for_each_entry(nsh, &newstripes, lru) {
1859 lock_device_hash_lock(conf, hash);
1860 wait_event_cmd(conf->wait_for_stripe,
1861 !list_empty(conf->inactive_list + hash),
1862 unlock_device_hash_lock(conf, hash),
1863 lock_device_hash_lock(conf, hash));
1864 osh = get_free_stripe(conf, hash);
1865 unlock_device_hash_lock(conf, hash);
1866 atomic_set(&nsh->count, 1);
1867 for(i=0; i<conf->pool_size; i++) {
1868 nsh->dev[i].page = osh->dev[i].page;
1869 nsh->dev[i].orig_page = osh->dev[i].page;
1871 for( ; i<newsize; i++)
1872 nsh->dev[i].page = NULL;
1873 nsh->hash_lock_index = hash;
1874 kmem_cache_free(conf->slab_cache, osh);
1876 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
1877 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
1882 kmem_cache_destroy(conf->slab_cache);
1885 * At this point, we are holding all the stripes so the array
1886 * is completely stalled, so now is a good time to resize
1887 * conf->disks and the scribble region
1889 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1891 for (i=0; i<conf->raid_disks; i++)
1892 ndisks[i] = conf->disks[i];
1894 conf->disks = ndisks;
1899 conf->scribble_len = scribble_len(newsize);
1900 for_each_present_cpu(cpu) {
1901 struct raid5_percpu *percpu;
1904 percpu = per_cpu_ptr(conf->percpu, cpu);
1905 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1908 kfree(percpu->scribble);
1909 percpu->scribble = scribble;
1917 /* Step 4, return new stripes to service */
1918 while(!list_empty(&newstripes)) {
1919 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1920 list_del_init(&nsh->lru);
1922 for (i=conf->raid_disks; i < newsize; i++)
1923 if (nsh->dev[i].page == NULL) {
1924 struct page *p = alloc_page(GFP_NOIO);
1925 nsh->dev[i].page = p;
1926 nsh->dev[i].orig_page = p;
1930 release_stripe(nsh);
1932 /* critical section pass, GFP_NOIO no longer needed */
1934 conf->slab_cache = sc;
1935 conf->active_name = 1-conf->active_name;
1936 conf->pool_size = newsize;
1940 static int drop_one_stripe(struct r5conf *conf, int hash)
1942 struct stripe_head *sh;
1944 spin_lock_irq(conf->hash_locks + hash);
1945 sh = get_free_stripe(conf, hash);
1946 spin_unlock_irq(conf->hash_locks + hash);
1949 BUG_ON(atomic_read(&sh->count));
1951 kmem_cache_free(conf->slab_cache, sh);
1952 atomic_dec(&conf->active_stripes);
1956 static void shrink_stripes(struct r5conf *conf)
1959 for (hash = 0; hash < NR_STRIPE_HASH_LOCKS; hash++)
1960 while (drop_one_stripe(conf, hash))
1963 if (conf->slab_cache)
1964 kmem_cache_destroy(conf->slab_cache);
1965 conf->slab_cache = NULL;
1968 static void raid5_end_read_request(struct bio * bi, int error)
1970 struct stripe_head *sh = bi->bi_private;
1971 struct r5conf *conf = sh->raid_conf;
1972 int disks = sh->disks, i;
1973 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1974 char b[BDEVNAME_SIZE];
1975 struct md_rdev *rdev = NULL;
1978 for (i=0 ; i<disks; i++)
1979 if (bi == &sh->dev[i].req)
1982 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1983 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1989 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1990 /* If replacement finished while this request was outstanding,
1991 * 'replacement' might be NULL already.
1992 * In that case it moved down to 'rdev'.
1993 * rdev is not removed until all requests are finished.
1995 rdev = conf->disks[i].replacement;
1997 rdev = conf->disks[i].rdev;
1999 if (use_new_offset(conf, sh))
2000 s = sh->sector + rdev->new_data_offset;
2002 s = sh->sector + rdev->data_offset;
2004 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2005 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2006 /* Note that this cannot happen on a
2007 * replacement device. We just fail those on
2012 "md/raid:%s: read error corrected"
2013 " (%lu sectors at %llu on %s)\n",
2014 mdname(conf->mddev), STRIPE_SECTORS,
2015 (unsigned long long)s,
2016 bdevname(rdev->bdev, b));
2017 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2018 clear_bit(R5_ReadError, &sh->dev[i].flags);
2019 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2020 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2021 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2023 if (atomic_read(&rdev->read_errors))
2024 atomic_set(&rdev->read_errors, 0);
2026 const char *bdn = bdevname(rdev->bdev, b);
2030 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2031 atomic_inc(&rdev->read_errors);
2032 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2035 "md/raid:%s: read error on replacement device "
2036 "(sector %llu on %s).\n",
2037 mdname(conf->mddev),
2038 (unsigned long long)s,
2040 else if (conf->mddev->degraded >= conf->max_degraded) {
2044 "md/raid:%s: read error not correctable "
2045 "(sector %llu on %s).\n",
2046 mdname(conf->mddev),
2047 (unsigned long long)s,
2049 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2054 "md/raid:%s: read error NOT corrected!! "
2055 "(sector %llu on %s).\n",
2056 mdname(conf->mddev),
2057 (unsigned long long)s,
2059 } else if (atomic_read(&rdev->read_errors)
2060 > conf->max_nr_stripes)
2062 "md/raid:%s: Too many read errors, failing device %s.\n",
2063 mdname(conf->mddev), bdn);
2066 if (set_bad && test_bit(In_sync, &rdev->flags)
2067 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2070 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2071 set_bit(R5_ReadError, &sh->dev[i].flags);
2072 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2074 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2076 clear_bit(R5_ReadError, &sh->dev[i].flags);
2077 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2079 && test_bit(In_sync, &rdev->flags)
2080 && rdev_set_badblocks(
2081 rdev, sh->sector, STRIPE_SECTORS, 0)))
2082 md_error(conf->mddev, rdev);
2085 rdev_dec_pending(rdev, conf->mddev);
2086 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2087 set_bit(STRIPE_HANDLE, &sh->state);
2091 static void raid5_end_write_request(struct bio *bi, int error)
2093 struct stripe_head *sh = bi->bi_private;
2094 struct r5conf *conf = sh->raid_conf;
2095 int disks = sh->disks, i;
2096 struct md_rdev *uninitialized_var(rdev);
2097 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2100 int replacement = 0;
2102 for (i = 0 ; i < disks; i++) {
2103 if (bi == &sh->dev[i].req) {
2104 rdev = conf->disks[i].rdev;
2107 if (bi == &sh->dev[i].rreq) {
2108 rdev = conf->disks[i].replacement;
2112 /* rdev was removed and 'replacement'
2113 * replaced it. rdev is not removed
2114 * until all requests are finished.
2116 rdev = conf->disks[i].rdev;
2120 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2121 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2130 md_error(conf->mddev, rdev);
2131 else if (is_badblock(rdev, sh->sector,
2133 &first_bad, &bad_sectors))
2134 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2137 set_bit(STRIPE_DEGRADED, &sh->state);
2138 set_bit(WriteErrorSeen, &rdev->flags);
2139 set_bit(R5_WriteError, &sh->dev[i].flags);
2140 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2141 set_bit(MD_RECOVERY_NEEDED,
2142 &rdev->mddev->recovery);
2143 } else if (is_badblock(rdev, sh->sector,
2145 &first_bad, &bad_sectors)) {
2146 set_bit(R5_MadeGood, &sh->dev[i].flags);
2147 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2148 /* That was a successful write so make
2149 * sure it looks like we already did
2152 set_bit(R5_ReWrite, &sh->dev[i].flags);
2155 rdev_dec_pending(rdev, conf->mddev);
2157 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2158 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2159 set_bit(STRIPE_HANDLE, &sh->state);
2163 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2165 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2167 struct r5dev *dev = &sh->dev[i];
2169 bio_init(&dev->req);
2170 dev->req.bi_io_vec = &dev->vec;
2171 dev->req.bi_max_vecs = 1;
2172 dev->req.bi_private = sh;
2174 bio_init(&dev->rreq);
2175 dev->rreq.bi_io_vec = &dev->rvec;
2176 dev->rreq.bi_max_vecs = 1;
2177 dev->rreq.bi_private = sh;
2180 dev->sector = compute_blocknr(sh, i, previous);
2183 static void error(struct mddev *mddev, struct md_rdev *rdev)
2185 char b[BDEVNAME_SIZE];
2186 struct r5conf *conf = mddev->private;
2187 unsigned long flags;
2188 pr_debug("raid456: error called\n");
2190 spin_lock_irqsave(&conf->device_lock, flags);
2191 clear_bit(In_sync, &rdev->flags);
2192 mddev->degraded = calc_degraded(conf);
2193 spin_unlock_irqrestore(&conf->device_lock, flags);
2194 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2196 set_bit(Blocked, &rdev->flags);
2197 set_bit(Faulty, &rdev->flags);
2198 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2200 "md/raid:%s: Disk failure on %s, disabling device.\n"
2201 "md/raid:%s: Operation continuing on %d devices.\n",
2203 bdevname(rdev->bdev, b),
2205 conf->raid_disks - mddev->degraded);
2209 * Input: a 'big' sector number,
2210 * Output: index of the data and parity disk, and the sector # in them.
2212 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2213 int previous, int *dd_idx,
2214 struct stripe_head *sh)
2216 sector_t stripe, stripe2;
2217 sector_t chunk_number;
2218 unsigned int chunk_offset;
2221 sector_t new_sector;
2222 int algorithm = previous ? conf->prev_algo
2224 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2225 : conf->chunk_sectors;
2226 int raid_disks = previous ? conf->previous_raid_disks
2228 int data_disks = raid_disks - conf->max_degraded;
2230 /* First compute the information on this sector */
2233 * Compute the chunk number and the sector offset inside the chunk
2235 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2236 chunk_number = r_sector;
2239 * Compute the stripe number
2241 stripe = chunk_number;
2242 *dd_idx = sector_div(stripe, data_disks);
2245 * Select the parity disk based on the user selected algorithm.
2247 pd_idx = qd_idx = -1;
2248 switch(conf->level) {
2250 pd_idx = data_disks;
2253 switch (algorithm) {
2254 case ALGORITHM_LEFT_ASYMMETRIC:
2255 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2256 if (*dd_idx >= pd_idx)
2259 case ALGORITHM_RIGHT_ASYMMETRIC:
2260 pd_idx = sector_div(stripe2, raid_disks);
2261 if (*dd_idx >= pd_idx)
2264 case ALGORITHM_LEFT_SYMMETRIC:
2265 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2266 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2268 case ALGORITHM_RIGHT_SYMMETRIC:
2269 pd_idx = sector_div(stripe2, raid_disks);
2270 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2272 case ALGORITHM_PARITY_0:
2276 case ALGORITHM_PARITY_N:
2277 pd_idx = data_disks;
2285 switch (algorithm) {
2286 case ALGORITHM_LEFT_ASYMMETRIC:
2287 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2288 qd_idx = pd_idx + 1;
2289 if (pd_idx == raid_disks-1) {
2290 (*dd_idx)++; /* Q D D D P */
2292 } else if (*dd_idx >= pd_idx)
2293 (*dd_idx) += 2; /* D D P Q D */
2295 case ALGORITHM_RIGHT_ASYMMETRIC:
2296 pd_idx = sector_div(stripe2, raid_disks);
2297 qd_idx = pd_idx + 1;
2298 if (pd_idx == raid_disks-1) {
2299 (*dd_idx)++; /* Q D D D P */
2301 } else if (*dd_idx >= pd_idx)
2302 (*dd_idx) += 2; /* D D P Q D */
2304 case ALGORITHM_LEFT_SYMMETRIC:
2305 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2306 qd_idx = (pd_idx + 1) % raid_disks;
2307 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2309 case ALGORITHM_RIGHT_SYMMETRIC:
2310 pd_idx = sector_div(stripe2, raid_disks);
2311 qd_idx = (pd_idx + 1) % raid_disks;
2312 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2315 case ALGORITHM_PARITY_0:
2320 case ALGORITHM_PARITY_N:
2321 pd_idx = data_disks;
2322 qd_idx = data_disks + 1;
2325 case ALGORITHM_ROTATING_ZERO_RESTART:
2326 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2327 * of blocks for computing Q is different.
2329 pd_idx = sector_div(stripe2, raid_disks);
2330 qd_idx = pd_idx + 1;
2331 if (pd_idx == raid_disks-1) {
2332 (*dd_idx)++; /* Q D D D P */
2334 } else if (*dd_idx >= pd_idx)
2335 (*dd_idx) += 2; /* D D P Q D */
2339 case ALGORITHM_ROTATING_N_RESTART:
2340 /* Same a left_asymmetric, by first stripe is
2341 * D D D P Q rather than
2345 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2346 qd_idx = pd_idx + 1;
2347 if (pd_idx == raid_disks-1) {
2348 (*dd_idx)++; /* Q D D D P */
2350 } else if (*dd_idx >= pd_idx)
2351 (*dd_idx) += 2; /* D D P Q D */
2355 case ALGORITHM_ROTATING_N_CONTINUE:
2356 /* Same as left_symmetric but Q is before P */
2357 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2358 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2359 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2363 case ALGORITHM_LEFT_ASYMMETRIC_6:
2364 /* RAID5 left_asymmetric, with Q on last device */
2365 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2366 if (*dd_idx >= pd_idx)
2368 qd_idx = raid_disks - 1;
2371 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2372 pd_idx = sector_div(stripe2, raid_disks-1);
2373 if (*dd_idx >= pd_idx)
2375 qd_idx = raid_disks - 1;
2378 case ALGORITHM_LEFT_SYMMETRIC_6:
2379 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2380 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2381 qd_idx = raid_disks - 1;
2384 case ALGORITHM_RIGHT_SYMMETRIC_6:
2385 pd_idx = sector_div(stripe2, raid_disks-1);
2386 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2387 qd_idx = raid_disks - 1;
2390 case ALGORITHM_PARITY_0_6:
2393 qd_idx = raid_disks - 1;
2403 sh->pd_idx = pd_idx;
2404 sh->qd_idx = qd_idx;
2405 sh->ddf_layout = ddf_layout;
2408 * Finally, compute the new sector number
2410 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2414 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2416 struct r5conf *conf = sh->raid_conf;
2417 int raid_disks = sh->disks;
2418 int data_disks = raid_disks - conf->max_degraded;
2419 sector_t new_sector = sh->sector, check;
2420 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2421 : conf->chunk_sectors;
2422 int algorithm = previous ? conf->prev_algo
2426 sector_t chunk_number;
2427 int dummy1, dd_idx = i;
2429 struct stripe_head sh2;
2431 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2432 stripe = new_sector;
2434 if (i == sh->pd_idx)
2436 switch(conf->level) {
2439 switch (algorithm) {
2440 case ALGORITHM_LEFT_ASYMMETRIC:
2441 case ALGORITHM_RIGHT_ASYMMETRIC:
2445 case ALGORITHM_LEFT_SYMMETRIC:
2446 case ALGORITHM_RIGHT_SYMMETRIC:
2449 i -= (sh->pd_idx + 1);
2451 case ALGORITHM_PARITY_0:
2454 case ALGORITHM_PARITY_N:
2461 if (i == sh->qd_idx)
2462 return 0; /* It is the Q disk */
2463 switch (algorithm) {
2464 case ALGORITHM_LEFT_ASYMMETRIC:
2465 case ALGORITHM_RIGHT_ASYMMETRIC:
2466 case ALGORITHM_ROTATING_ZERO_RESTART:
2467 case ALGORITHM_ROTATING_N_RESTART:
2468 if (sh->pd_idx == raid_disks-1)
2469 i--; /* Q D D D P */
2470 else if (i > sh->pd_idx)
2471 i -= 2; /* D D P Q D */
2473 case ALGORITHM_LEFT_SYMMETRIC:
2474 case ALGORITHM_RIGHT_SYMMETRIC:
2475 if (sh->pd_idx == raid_disks-1)
2476 i--; /* Q D D D P */
2481 i -= (sh->pd_idx + 2);
2484 case ALGORITHM_PARITY_0:
2487 case ALGORITHM_PARITY_N:
2489 case ALGORITHM_ROTATING_N_CONTINUE:
2490 /* Like left_symmetric, but P is before Q */
2491 if (sh->pd_idx == 0)
2492 i--; /* P D D D Q */
2497 i -= (sh->pd_idx + 1);
2500 case ALGORITHM_LEFT_ASYMMETRIC_6:
2501 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2505 case ALGORITHM_LEFT_SYMMETRIC_6:
2506 case ALGORITHM_RIGHT_SYMMETRIC_6:
2508 i += data_disks + 1;
2509 i -= (sh->pd_idx + 1);
2511 case ALGORITHM_PARITY_0_6:
2520 chunk_number = stripe * data_disks + i;
2521 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2523 check = raid5_compute_sector(conf, r_sector,
2524 previous, &dummy1, &sh2);
2525 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2526 || sh2.qd_idx != sh->qd_idx) {
2527 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2528 mdname(conf->mddev));
2535 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2536 int rcw, int expand)
2538 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2539 struct r5conf *conf = sh->raid_conf;
2540 int level = conf->level;
2544 for (i = disks; i--; ) {
2545 struct r5dev *dev = &sh->dev[i];
2548 set_bit(R5_LOCKED, &dev->flags);
2549 set_bit(R5_Wantdrain, &dev->flags);
2551 clear_bit(R5_UPTODATE, &dev->flags);
2555 /* if we are not expanding this is a proper write request, and
2556 * there will be bios with new data to be drained into the
2561 /* False alarm, nothing to do */
2563 sh->reconstruct_state = reconstruct_state_drain_run;
2564 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2566 sh->reconstruct_state = reconstruct_state_run;
2568 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2570 if (s->locked + conf->max_degraded == disks)
2571 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2572 atomic_inc(&conf->pending_full_writes);
2575 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2576 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2578 for (i = disks; i--; ) {
2579 struct r5dev *dev = &sh->dev[i];
2584 (test_bit(R5_UPTODATE, &dev->flags) ||
2585 test_bit(R5_Wantcompute, &dev->flags))) {
2586 set_bit(R5_Wantdrain, &dev->flags);
2587 set_bit(R5_LOCKED, &dev->flags);
2588 clear_bit(R5_UPTODATE, &dev->flags);
2593 /* False alarm - nothing to do */
2595 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2596 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2597 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2598 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2601 /* keep the parity disk(s) locked while asynchronous operations
2604 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2605 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2609 int qd_idx = sh->qd_idx;
2610 struct r5dev *dev = &sh->dev[qd_idx];
2612 set_bit(R5_LOCKED, &dev->flags);
2613 clear_bit(R5_UPTODATE, &dev->flags);
2617 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2618 __func__, (unsigned long long)sh->sector,
2619 s->locked, s->ops_request);
2623 * Each stripe/dev can have one or more bion attached.
2624 * toread/towrite point to the first in a chain.
2625 * The bi_next chain must be in order.
2627 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2630 struct r5conf *conf = sh->raid_conf;
2633 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2634 (unsigned long long)bi->bi_iter.bi_sector,
2635 (unsigned long long)sh->sector);
2638 * If several bio share a stripe. The bio bi_phys_segments acts as a
2639 * reference count to avoid race. The reference count should already be
2640 * increased before this function is called (for example, in
2641 * make_request()), so other bio sharing this stripe will not free the
2642 * stripe. If a stripe is owned by one stripe, the stripe lock will
2645 spin_lock_irq(&sh->stripe_lock);
2647 bip = &sh->dev[dd_idx].towrite;
2651 bip = &sh->dev[dd_idx].toread;
2652 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2653 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2655 bip = & (*bip)->bi_next;
2657 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2660 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2664 raid5_inc_bi_active_stripes(bi);
2667 /* check if page is covered */
2668 sector_t sector = sh->dev[dd_idx].sector;
2669 for (bi=sh->dev[dd_idx].towrite;
2670 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2671 bi && bi->bi_iter.bi_sector <= sector;
2672 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2673 if (bio_end_sector(bi) >= sector)
2674 sector = bio_end_sector(bi);
2676 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2677 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2680 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2681 (unsigned long long)(*bip)->bi_iter.bi_sector,
2682 (unsigned long long)sh->sector, dd_idx);
2683 spin_unlock_irq(&sh->stripe_lock);
2685 if (conf->mddev->bitmap && firstwrite) {
2686 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2688 sh->bm_seq = conf->seq_flush+1;
2689 set_bit(STRIPE_BIT_DELAY, &sh->state);
2694 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2695 spin_unlock_irq(&sh->stripe_lock);
2699 static void end_reshape(struct r5conf *conf);
2701 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2702 struct stripe_head *sh)
2704 int sectors_per_chunk =
2705 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2707 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2708 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2710 raid5_compute_sector(conf,
2711 stripe * (disks - conf->max_degraded)
2712 *sectors_per_chunk + chunk_offset,
2718 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2719 struct stripe_head_state *s, int disks,
2720 struct bio **return_bi)
2723 for (i = disks; i--; ) {
2727 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2728 struct md_rdev *rdev;
2730 rdev = rcu_dereference(conf->disks[i].rdev);
2731 if (rdev && test_bit(In_sync, &rdev->flags))
2732 atomic_inc(&rdev->nr_pending);
2737 if (!rdev_set_badblocks(
2741 md_error(conf->mddev, rdev);
2742 rdev_dec_pending(rdev, conf->mddev);
2745 spin_lock_irq(&sh->stripe_lock);
2746 /* fail all writes first */
2747 bi = sh->dev[i].towrite;
2748 sh->dev[i].towrite = NULL;
2749 spin_unlock_irq(&sh->stripe_lock);
2753 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2754 wake_up(&conf->wait_for_overlap);
2756 while (bi && bi->bi_iter.bi_sector <
2757 sh->dev[i].sector + STRIPE_SECTORS) {
2758 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2759 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2760 if (!raid5_dec_bi_active_stripes(bi)) {
2761 md_write_end(conf->mddev);
2762 bi->bi_next = *return_bi;
2768 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2769 STRIPE_SECTORS, 0, 0);
2771 /* and fail all 'written' */
2772 bi = sh->dev[i].written;
2773 sh->dev[i].written = NULL;
2774 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
2775 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
2776 sh->dev[i].page = sh->dev[i].orig_page;
2779 if (bi) bitmap_end = 1;
2780 while (bi && bi->bi_iter.bi_sector <
2781 sh->dev[i].sector + STRIPE_SECTORS) {
2782 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2783 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2784 if (!raid5_dec_bi_active_stripes(bi)) {
2785 md_write_end(conf->mddev);
2786 bi->bi_next = *return_bi;
2792 /* fail any reads if this device is non-operational and
2793 * the data has not reached the cache yet.
2795 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2796 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2797 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2798 spin_lock_irq(&sh->stripe_lock);
2799 bi = sh->dev[i].toread;
2800 sh->dev[i].toread = NULL;
2801 spin_unlock_irq(&sh->stripe_lock);
2802 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2803 wake_up(&conf->wait_for_overlap);
2804 while (bi && bi->bi_iter.bi_sector <
2805 sh->dev[i].sector + STRIPE_SECTORS) {
2806 struct bio *nextbi =
2807 r5_next_bio(bi, sh->dev[i].sector);
2808 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2809 if (!raid5_dec_bi_active_stripes(bi)) {
2810 bi->bi_next = *return_bi;
2817 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2818 STRIPE_SECTORS, 0, 0);
2819 /* If we were in the middle of a write the parity block might
2820 * still be locked - so just clear all R5_LOCKED flags
2822 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2825 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2826 if (atomic_dec_and_test(&conf->pending_full_writes))
2827 md_wakeup_thread(conf->mddev->thread);
2831 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2832 struct stripe_head_state *s)
2837 clear_bit(STRIPE_SYNCING, &sh->state);
2838 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2839 wake_up(&conf->wait_for_overlap);
2842 /* There is nothing more to do for sync/check/repair.
2843 * Don't even need to abort as that is handled elsewhere
2844 * if needed, and not always wanted e.g. if there is a known
2846 * For recover/replace we need to record a bad block on all
2847 * non-sync devices, or abort the recovery
2849 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2850 /* During recovery devices cannot be removed, so
2851 * locking and refcounting of rdevs is not needed
2853 for (i = 0; i < conf->raid_disks; i++) {
2854 struct md_rdev *rdev = conf->disks[i].rdev;
2856 && !test_bit(Faulty, &rdev->flags)
2857 && !test_bit(In_sync, &rdev->flags)
2858 && !rdev_set_badblocks(rdev, sh->sector,
2861 rdev = conf->disks[i].replacement;
2863 && !test_bit(Faulty, &rdev->flags)
2864 && !test_bit(In_sync, &rdev->flags)
2865 && !rdev_set_badblocks(rdev, sh->sector,
2870 conf->recovery_disabled =
2871 conf->mddev->recovery_disabled;
2873 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2876 static int want_replace(struct stripe_head *sh, int disk_idx)
2878 struct md_rdev *rdev;
2880 /* Doing recovery so rcu locking not required */
2881 rdev = sh->raid_conf->disks[disk_idx].replacement;
2883 && !test_bit(Faulty, &rdev->flags)
2884 && !test_bit(In_sync, &rdev->flags)
2885 && (rdev->recovery_offset <= sh->sector
2886 || rdev->mddev->recovery_cp <= sh->sector))
2892 /* fetch_block - checks the given member device to see if its data needs
2893 * to be read or computed to satisfy a request.
2895 * Returns 1 when no more member devices need to be checked, otherwise returns
2896 * 0 to tell the loop in handle_stripe_fill to continue
2898 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2899 int disk_idx, int disks)
2901 struct r5dev *dev = &sh->dev[disk_idx];
2902 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2903 &sh->dev[s->failed_num[1]] };
2905 /* is the data in this block needed, and can we get it? */
2906 if (!test_bit(R5_LOCKED, &dev->flags) &&
2907 !test_bit(R5_UPTODATE, &dev->flags) &&
2909 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2910 s->syncing || s->expanding ||
2911 (s->replacing && want_replace(sh, disk_idx)) ||
2912 (s->failed >= 1 && fdev[0]->toread) ||
2913 (s->failed >= 2 && fdev[1]->toread) ||
2914 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2915 (!test_bit(R5_Insync, &dev->flags) || test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) &&
2916 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2917 ((sh->raid_conf->level == 6 ||
2918 sh->sector >= sh->raid_conf->mddev->recovery_cp)
2919 && s->failed && s->to_write &&
2920 (s->to_write - s->non_overwrite <
2921 sh->raid_conf->raid_disks - sh->raid_conf->max_degraded) &&
2922 (!test_bit(R5_Insync, &dev->flags) || test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))))) {
2923 /* we would like to get this block, possibly by computing it,
2924 * otherwise read it if the backing disk is insync
2926 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2927 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2928 if ((s->uptodate == disks - 1) &&
2929 (s->failed && (disk_idx == s->failed_num[0] ||
2930 disk_idx == s->failed_num[1]))) {
2931 /* have disk failed, and we're requested to fetch it;
2934 pr_debug("Computing stripe %llu block %d\n",
2935 (unsigned long long)sh->sector, disk_idx);
2936 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2937 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2938 set_bit(R5_Wantcompute, &dev->flags);
2939 sh->ops.target = disk_idx;
2940 sh->ops.target2 = -1; /* no 2nd target */
2942 /* Careful: from this point on 'uptodate' is in the eye
2943 * of raid_run_ops which services 'compute' operations
2944 * before writes. R5_Wantcompute flags a block that will
2945 * be R5_UPTODATE by the time it is needed for a
2946 * subsequent operation.
2950 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2951 /* Computing 2-failure is *very* expensive; only
2952 * do it if failed >= 2
2955 for (other = disks; other--; ) {
2956 if (other == disk_idx)
2958 if (!test_bit(R5_UPTODATE,
2959 &sh->dev[other].flags))
2963 pr_debug("Computing stripe %llu blocks %d,%d\n",
2964 (unsigned long long)sh->sector,
2966 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2967 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2968 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2969 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2970 sh->ops.target = disk_idx;
2971 sh->ops.target2 = other;
2975 } else if (test_bit(R5_Insync, &dev->flags)) {
2976 set_bit(R5_LOCKED, &dev->flags);
2977 set_bit(R5_Wantread, &dev->flags);
2979 pr_debug("Reading block %d (sync=%d)\n",
2980 disk_idx, s->syncing);
2988 * handle_stripe_fill - read or compute data to satisfy pending requests.
2990 static void handle_stripe_fill(struct stripe_head *sh,
2991 struct stripe_head_state *s,
2996 /* look for blocks to read/compute, skip this if a compute
2997 * is already in flight, or if the stripe contents are in the
2998 * midst of changing due to a write
3000 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3001 !sh->reconstruct_state)
3002 for (i = disks; i--; )
3003 if (fetch_block(sh, s, i, disks))
3005 set_bit(STRIPE_HANDLE, &sh->state);
3008 /* handle_stripe_clean_event
3009 * any written block on an uptodate or failed drive can be returned.
3010 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3011 * never LOCKED, so we don't need to test 'failed' directly.
3013 static void handle_stripe_clean_event(struct r5conf *conf,
3014 struct stripe_head *sh, int disks, struct bio **return_bi)
3018 int discard_pending = 0;
3020 for (i = disks; i--; )
3021 if (sh->dev[i].written) {
3023 if (!test_bit(R5_LOCKED, &dev->flags) &&
3024 (test_bit(R5_UPTODATE, &dev->flags) ||
3025 test_bit(R5_Discard, &dev->flags) ||
3026 test_bit(R5_SkipCopy, &dev->flags))) {
3027 /* We can return any write requests */
3028 struct bio *wbi, *wbi2;
3029 pr_debug("Return write for disc %d\n", i);
3030 if (test_and_clear_bit(R5_Discard, &dev->flags))
3031 clear_bit(R5_UPTODATE, &dev->flags);
3032 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3033 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3034 dev->page = dev->orig_page;
3037 dev->written = NULL;
3038 while (wbi && wbi->bi_iter.bi_sector <
3039 dev->sector + STRIPE_SECTORS) {
3040 wbi2 = r5_next_bio(wbi, dev->sector);
3041 if (!raid5_dec_bi_active_stripes(wbi)) {
3042 md_write_end(conf->mddev);
3043 wbi->bi_next = *return_bi;
3048 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3050 !test_bit(STRIPE_DEGRADED, &sh->state),
3052 } else if (test_bit(R5_Discard, &dev->flags))
3053 discard_pending = 1;
3054 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3055 WARN_ON(dev->page != dev->orig_page);
3057 if (!discard_pending &&
3058 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3059 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3060 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3061 if (sh->qd_idx >= 0) {
3062 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3063 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3065 /* now that discard is done we can proceed with any sync */
3066 clear_bit(STRIPE_DISCARD, &sh->state);
3068 * SCSI discard will change some bio fields and the stripe has
3069 * no updated data, so remove it from hash list and the stripe
3070 * will be reinitialized
3072 spin_lock_irq(&conf->device_lock);
3074 spin_unlock_irq(&conf->device_lock);
3075 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3076 set_bit(STRIPE_HANDLE, &sh->state);
3080 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3081 if (atomic_dec_and_test(&conf->pending_full_writes))
3082 md_wakeup_thread(conf->mddev->thread);
3085 static void handle_stripe_dirtying(struct r5conf *conf,
3086 struct stripe_head *sh,
3087 struct stripe_head_state *s,
3090 int rmw = 0, rcw = 0, i;
3091 sector_t recovery_cp = conf->mddev->recovery_cp;
3093 /* RAID6 requires 'rcw' in current implementation.
3094 * Otherwise, check whether resync is now happening or should start.
3095 * If yes, then the array is dirty (after unclean shutdown or
3096 * initial creation), so parity in some stripes might be inconsistent.
3097 * In this case, we need to always do reconstruct-write, to ensure
3098 * that in case of drive failure or read-error correction, we
3099 * generate correct data from the parity.
3101 if (conf->max_degraded == 2 ||
3102 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
3103 /* Calculate the real rcw later - for now make it
3104 * look like rcw is cheaper
3107 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
3108 conf->max_degraded, (unsigned long long)recovery_cp,
3109 (unsigned long long)sh->sector);
3110 } else for (i = disks; i--; ) {
3111 /* would I have to read this buffer for read_modify_write */
3112 struct r5dev *dev = &sh->dev[i];
3113 if ((dev->towrite || i == sh->pd_idx) &&
3114 !test_bit(R5_LOCKED, &dev->flags) &&
3115 !(test_bit(R5_UPTODATE, &dev->flags) ||
3116 test_bit(R5_Wantcompute, &dev->flags))) {
3117 if (test_bit(R5_Insync, &dev->flags))
3120 rmw += 2*disks; /* cannot read it */
3122 /* Would I have to read this buffer for reconstruct_write */
3123 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
3124 !test_bit(R5_LOCKED, &dev->flags) &&
3125 !(test_bit(R5_UPTODATE, &dev->flags) ||
3126 test_bit(R5_Wantcompute, &dev->flags))) {
3127 if (test_bit(R5_Insync, &dev->flags))
3133 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3134 (unsigned long long)sh->sector, rmw, rcw);
3135 set_bit(STRIPE_HANDLE, &sh->state);
3136 if (rmw < rcw && rmw > 0) {
3137 /* prefer read-modify-write, but need to get some data */
3138 if (conf->mddev->queue)
3139 blk_add_trace_msg(conf->mddev->queue,
3140 "raid5 rmw %llu %d",
3141 (unsigned long long)sh->sector, rmw);
3142 for (i = disks; i--; ) {
3143 struct r5dev *dev = &sh->dev[i];
3144 if ((dev->towrite || i == sh->pd_idx) &&
3145 !test_bit(R5_LOCKED, &dev->flags) &&
3146 !(test_bit(R5_UPTODATE, &dev->flags) ||
3147 test_bit(R5_Wantcompute, &dev->flags)) &&
3148 test_bit(R5_Insync, &dev->flags)) {
3149 if (test_bit(STRIPE_PREREAD_ACTIVE,
3151 pr_debug("Read_old block %d for r-m-w\n",
3153 set_bit(R5_LOCKED, &dev->flags);
3154 set_bit(R5_Wantread, &dev->flags);
3157 set_bit(STRIPE_DELAYED, &sh->state);
3158 set_bit(STRIPE_HANDLE, &sh->state);
3163 if (rcw <= rmw && rcw > 0) {
3164 /* want reconstruct write, but need to get some data */
3167 for (i = disks; i--; ) {
3168 struct r5dev *dev = &sh->dev[i];
3169 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3170 i != sh->pd_idx && i != sh->qd_idx &&
3171 !test_bit(R5_LOCKED, &dev->flags) &&
3172 !(test_bit(R5_UPTODATE, &dev->flags) ||
3173 test_bit(R5_Wantcompute, &dev->flags))) {
3175 if (test_bit(R5_Insync, &dev->flags) &&
3176 test_bit(STRIPE_PREREAD_ACTIVE,
3178 pr_debug("Read_old block "
3179 "%d for Reconstruct\n", i);
3180 set_bit(R5_LOCKED, &dev->flags);
3181 set_bit(R5_Wantread, &dev->flags);
3185 set_bit(STRIPE_DELAYED, &sh->state);
3186 set_bit(STRIPE_HANDLE, &sh->state);
3190 if (rcw && conf->mddev->queue)
3191 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3192 (unsigned long long)sh->sector,
3193 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3196 if (rcw > disks && rmw > disks &&
3197 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3198 set_bit(STRIPE_DELAYED, &sh->state);
3200 /* now if nothing is locked, and if we have enough data,
3201 * we can start a write request
3203 /* since handle_stripe can be called at any time we need to handle the
3204 * case where a compute block operation has been submitted and then a
3205 * subsequent call wants to start a write request. raid_run_ops only
3206 * handles the case where compute block and reconstruct are requested
3207 * simultaneously. If this is not the case then new writes need to be
3208 * held off until the compute completes.
3210 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3211 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3212 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3213 schedule_reconstruction(sh, s, rcw == 0, 0);
3216 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3217 struct stripe_head_state *s, int disks)
3219 struct r5dev *dev = NULL;
3221 set_bit(STRIPE_HANDLE, &sh->state);
3223 switch (sh->check_state) {
3224 case check_state_idle:
3225 /* start a new check operation if there are no failures */
3226 if (s->failed == 0) {
3227 BUG_ON(s->uptodate != disks);
3228 sh->check_state = check_state_run;
3229 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3230 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3234 dev = &sh->dev[s->failed_num[0]];
3236 case check_state_compute_result:
3237 sh->check_state = check_state_idle;
3239 dev = &sh->dev[sh->pd_idx];
3241 /* check that a write has not made the stripe insync */
3242 if (test_bit(STRIPE_INSYNC, &sh->state))
3245 /* either failed parity check, or recovery is happening */
3246 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3247 BUG_ON(s->uptodate != disks);
3249 set_bit(R5_LOCKED, &dev->flags);
3251 set_bit(R5_Wantwrite, &dev->flags);
3253 clear_bit(STRIPE_DEGRADED, &sh->state);
3254 set_bit(STRIPE_INSYNC, &sh->state);
3256 case check_state_run:
3257 break; /* we will be called again upon completion */
3258 case check_state_check_result:
3259 sh->check_state = check_state_idle;
3261 /* if a failure occurred during the check operation, leave
3262 * STRIPE_INSYNC not set and let the stripe be handled again
3267 /* handle a successful check operation, if parity is correct
3268 * we are done. Otherwise update the mismatch count and repair
3269 * parity if !MD_RECOVERY_CHECK
3271 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3272 /* parity is correct (on disc,
3273 * not in buffer any more)
3275 set_bit(STRIPE_INSYNC, &sh->state);
3277 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3278 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3279 /* don't try to repair!! */
3280 set_bit(STRIPE_INSYNC, &sh->state);
3282 sh->check_state = check_state_compute_run;
3283 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3284 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3285 set_bit(R5_Wantcompute,
3286 &sh->dev[sh->pd_idx].flags);
3287 sh->ops.target = sh->pd_idx;
3288 sh->ops.target2 = -1;
3293 case check_state_compute_run:
3296 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3297 __func__, sh->check_state,
3298 (unsigned long long) sh->sector);
3303 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3304 struct stripe_head_state *s,
3307 int pd_idx = sh->pd_idx;
3308 int qd_idx = sh->qd_idx;
3311 set_bit(STRIPE_HANDLE, &sh->state);
3313 BUG_ON(s->failed > 2);
3315 /* Want to check and possibly repair P and Q.
3316 * However there could be one 'failed' device, in which
3317 * case we can only check one of them, possibly using the
3318 * other to generate missing data
3321 switch (sh->check_state) {
3322 case check_state_idle:
3323 /* start a new check operation if there are < 2 failures */
3324 if (s->failed == s->q_failed) {
3325 /* The only possible failed device holds Q, so it
3326 * makes sense to check P (If anything else were failed,
3327 * we would have used P to recreate it).
3329 sh->check_state = check_state_run;
3331 if (!s->q_failed && s->failed < 2) {
3332 /* Q is not failed, and we didn't use it to generate
3333 * anything, so it makes sense to check it
3335 if (sh->check_state == check_state_run)
3336 sh->check_state = check_state_run_pq;
3338 sh->check_state = check_state_run_q;
3341 /* discard potentially stale zero_sum_result */
3342 sh->ops.zero_sum_result = 0;
3344 if (sh->check_state == check_state_run) {
3345 /* async_xor_zero_sum destroys the contents of P */
3346 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3349 if (sh->check_state >= check_state_run &&
3350 sh->check_state <= check_state_run_pq) {
3351 /* async_syndrome_zero_sum preserves P and Q, so
3352 * no need to mark them !uptodate here
3354 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3358 /* we have 2-disk failure */
3359 BUG_ON(s->failed != 2);
3361 case check_state_compute_result:
3362 sh->check_state = check_state_idle;
3364 /* check that a write has not made the stripe insync */
3365 if (test_bit(STRIPE_INSYNC, &sh->state))
3368 /* now write out any block on a failed drive,
3369 * or P or Q if they were recomputed
3371 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3372 if (s->failed == 2) {
3373 dev = &sh->dev[s->failed_num[1]];
3375 set_bit(R5_LOCKED, &dev->flags);
3376 set_bit(R5_Wantwrite, &dev->flags);
3378 if (s->failed >= 1) {
3379 dev = &sh->dev[s->failed_num[0]];
3381 set_bit(R5_LOCKED, &dev->flags);
3382 set_bit(R5_Wantwrite, &dev->flags);
3384 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3385 dev = &sh->dev[pd_idx];
3387 set_bit(R5_LOCKED, &dev->flags);
3388 set_bit(R5_Wantwrite, &dev->flags);
3390 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3391 dev = &sh->dev[qd_idx];
3393 set_bit(R5_LOCKED, &dev->flags);
3394 set_bit(R5_Wantwrite, &dev->flags);
3396 clear_bit(STRIPE_DEGRADED, &sh->state);
3398 set_bit(STRIPE_INSYNC, &sh->state);
3400 case check_state_run:
3401 case check_state_run_q:
3402 case check_state_run_pq:
3403 break; /* we will be called again upon completion */
3404 case check_state_check_result:
3405 sh->check_state = check_state_idle;
3407 /* handle a successful check operation, if parity is correct
3408 * we are done. Otherwise update the mismatch count and repair
3409 * parity if !MD_RECOVERY_CHECK
3411 if (sh->ops.zero_sum_result == 0) {
3412 /* both parities are correct */
3414 set_bit(STRIPE_INSYNC, &sh->state);
3416 /* in contrast to the raid5 case we can validate
3417 * parity, but still have a failure to write
3420 sh->check_state = check_state_compute_result;
3421 /* Returning at this point means that we may go
3422 * off and bring p and/or q uptodate again so
3423 * we make sure to check zero_sum_result again
3424 * to verify if p or q need writeback
3428 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3429 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3430 /* don't try to repair!! */
3431 set_bit(STRIPE_INSYNC, &sh->state);
3433 int *target = &sh->ops.target;
3435 sh->ops.target = -1;
3436 sh->ops.target2 = -1;
3437 sh->check_state = check_state_compute_run;
3438 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3439 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3440 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3441 set_bit(R5_Wantcompute,
3442 &sh->dev[pd_idx].flags);
3444 target = &sh->ops.target2;
3447 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3448 set_bit(R5_Wantcompute,
3449 &sh->dev[qd_idx].flags);
3456 case check_state_compute_run:
3459 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3460 __func__, sh->check_state,
3461 (unsigned long long) sh->sector);
3466 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3470 /* We have read all the blocks in this stripe and now we need to
3471 * copy some of them into a target stripe for expand.
3473 struct dma_async_tx_descriptor *tx = NULL;
3474 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3475 for (i = 0; i < sh->disks; i++)
3476 if (i != sh->pd_idx && i != sh->qd_idx) {
3478 struct stripe_head *sh2;
3479 struct async_submit_ctl submit;
3481 sector_t bn = compute_blocknr(sh, i, 1);
3482 sector_t s = raid5_compute_sector(conf, bn, 0,
3484 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3486 /* so far only the early blocks of this stripe
3487 * have been requested. When later blocks
3488 * get requested, we will try again
3491 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3492 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3493 /* must have already done this block */
3494 release_stripe(sh2);
3498 /* place all the copies on one channel */
3499 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3500 tx = async_memcpy(sh2->dev[dd_idx].page,
3501 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3504 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3505 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3506 for (j = 0; j < conf->raid_disks; j++)
3507 if (j != sh2->pd_idx &&
3509 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3511 if (j == conf->raid_disks) {
3512 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3513 set_bit(STRIPE_HANDLE, &sh2->state);
3515 release_stripe(sh2);
3518 /* done submitting copies, wait for them to complete */
3519 async_tx_quiesce(&tx);
3523 * handle_stripe - do things to a stripe.
3525 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3526 * state of various bits to see what needs to be done.
3528 * return some read requests which now have data
3529 * return some write requests which are safely on storage
3530 * schedule a read on some buffers
3531 * schedule a write of some buffers
3532 * return confirmation of parity correctness
3536 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3538 struct r5conf *conf = sh->raid_conf;
3539 int disks = sh->disks;
3542 int do_recovery = 0;
3544 memset(s, 0, sizeof(*s));
3546 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3547 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3548 s->failed_num[0] = -1;
3549 s->failed_num[1] = -1;
3551 /* Now to look around and see what can be done */
3553 for (i=disks; i--; ) {
3554 struct md_rdev *rdev;
3561 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3563 dev->toread, dev->towrite, dev->written);
3564 /* maybe we can reply to a read
3566 * new wantfill requests are only permitted while
3567 * ops_complete_biofill is guaranteed to be inactive
3569 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3570 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3571 set_bit(R5_Wantfill, &dev->flags);
3573 /* now count some things */
3574 if (test_bit(R5_LOCKED, &dev->flags))
3576 if (test_bit(R5_UPTODATE, &dev->flags))
3578 if (test_bit(R5_Wantcompute, &dev->flags)) {
3580 BUG_ON(s->compute > 2);
3583 if (test_bit(R5_Wantfill, &dev->flags))
3585 else if (dev->toread)
3589 if (!test_bit(R5_OVERWRITE, &dev->flags))
3594 /* Prefer to use the replacement for reads, but only
3595 * if it is recovered enough and has no bad blocks.
3597 rdev = rcu_dereference(conf->disks[i].replacement);
3598 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3599 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3600 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3601 &first_bad, &bad_sectors))
3602 set_bit(R5_ReadRepl, &dev->flags);
3605 set_bit(R5_NeedReplace, &dev->flags);
3606 rdev = rcu_dereference(conf->disks[i].rdev);
3607 clear_bit(R5_ReadRepl, &dev->flags);
3609 if (rdev && test_bit(Faulty, &rdev->flags))
3612 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3613 &first_bad, &bad_sectors);
3614 if (s->blocked_rdev == NULL
3615 && (test_bit(Blocked, &rdev->flags)
3618 set_bit(BlockedBadBlocks,
3620 s->blocked_rdev = rdev;
3621 atomic_inc(&rdev->nr_pending);
3624 clear_bit(R5_Insync, &dev->flags);
3628 /* also not in-sync */
3629 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3630 test_bit(R5_UPTODATE, &dev->flags)) {
3631 /* treat as in-sync, but with a read error
3632 * which we can now try to correct
3634 set_bit(R5_Insync, &dev->flags);
3635 set_bit(R5_ReadError, &dev->flags);
3637 } else if (test_bit(In_sync, &rdev->flags))
3638 set_bit(R5_Insync, &dev->flags);
3639 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3640 /* in sync if before recovery_offset */
3641 set_bit(R5_Insync, &dev->flags);
3642 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3643 test_bit(R5_Expanded, &dev->flags))
3644 /* If we've reshaped into here, we assume it is Insync.
3645 * We will shortly update recovery_offset to make
3648 set_bit(R5_Insync, &dev->flags);
3650 if (test_bit(R5_WriteError, &dev->flags)) {
3651 /* This flag does not apply to '.replacement'
3652 * only to .rdev, so make sure to check that*/
3653 struct md_rdev *rdev2 = rcu_dereference(
3654 conf->disks[i].rdev);
3656 clear_bit(R5_Insync, &dev->flags);
3657 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3658 s->handle_bad_blocks = 1;
3659 atomic_inc(&rdev2->nr_pending);
3661 clear_bit(R5_WriteError, &dev->flags);
3663 if (test_bit(R5_MadeGood, &dev->flags)) {
3664 /* This flag does not apply to '.replacement'
3665 * only to .rdev, so make sure to check that*/
3666 struct md_rdev *rdev2 = rcu_dereference(
3667 conf->disks[i].rdev);
3668 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3669 s->handle_bad_blocks = 1;
3670 atomic_inc(&rdev2->nr_pending);
3672 clear_bit(R5_MadeGood, &dev->flags);
3674 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3675 struct md_rdev *rdev2 = rcu_dereference(
3676 conf->disks[i].replacement);
3677 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3678 s->handle_bad_blocks = 1;
3679 atomic_inc(&rdev2->nr_pending);
3681 clear_bit(R5_MadeGoodRepl, &dev->flags);
3683 if (!test_bit(R5_Insync, &dev->flags)) {
3684 /* The ReadError flag will just be confusing now */
3685 clear_bit(R5_ReadError, &dev->flags);
3686 clear_bit(R5_ReWrite, &dev->flags);
3688 if (test_bit(R5_ReadError, &dev->flags))
3689 clear_bit(R5_Insync, &dev->flags);
3690 if (!test_bit(R5_Insync, &dev->flags)) {
3692 s->failed_num[s->failed] = i;
3694 if (rdev && !test_bit(Faulty, &rdev->flags))
3698 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3699 /* If there is a failed device being replaced,
3700 * we must be recovering.
3701 * else if we are after recovery_cp, we must be syncing
3702 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3703 * else we can only be replacing
3704 * sync and recovery both need to read all devices, and so
3705 * use the same flag.
3708 sh->sector >= conf->mddev->recovery_cp ||
3709 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3717 static void handle_stripe(struct stripe_head *sh)
3719 struct stripe_head_state s;
3720 struct r5conf *conf = sh->raid_conf;
3723 int disks = sh->disks;
3724 struct r5dev *pdev, *qdev;
3726 clear_bit(STRIPE_HANDLE, &sh->state);
3727 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3728 /* already being handled, ensure it gets handled
3729 * again when current action finishes */
3730 set_bit(STRIPE_HANDLE, &sh->state);
3734 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3735 spin_lock(&sh->stripe_lock);
3736 /* Cannot process 'sync' concurrently with 'discard' */
3737 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3738 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3739 set_bit(STRIPE_SYNCING, &sh->state);
3740 clear_bit(STRIPE_INSYNC, &sh->state);
3741 clear_bit(STRIPE_REPLACED, &sh->state);
3743 spin_unlock(&sh->stripe_lock);
3745 clear_bit(STRIPE_DELAYED, &sh->state);
3747 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3748 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3749 (unsigned long long)sh->sector, sh->state,
3750 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3751 sh->check_state, sh->reconstruct_state);
3753 analyse_stripe(sh, &s);
3755 if (s.handle_bad_blocks) {
3756 set_bit(STRIPE_HANDLE, &sh->state);
3760 if (unlikely(s.blocked_rdev)) {
3761 if (s.syncing || s.expanding || s.expanded ||
3762 s.replacing || s.to_write || s.written) {
3763 set_bit(STRIPE_HANDLE, &sh->state);
3766 /* There is nothing for the blocked_rdev to block */
3767 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3768 s.blocked_rdev = NULL;
3771 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3772 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3773 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3776 pr_debug("locked=%d uptodate=%d to_read=%d"
3777 " to_write=%d failed=%d failed_num=%d,%d\n",
3778 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3779 s.failed_num[0], s.failed_num[1]);
3780 /* check if the array has lost more than max_degraded devices and,
3781 * if so, some requests might need to be failed.
3783 if (s.failed > conf->max_degraded) {
3784 sh->check_state = 0;
3785 sh->reconstruct_state = 0;
3786 if (s.to_read+s.to_write+s.written)
3787 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3788 if (s.syncing + s.replacing)
3789 handle_failed_sync(conf, sh, &s);
3792 /* Now we check to see if any write operations have recently
3796 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3798 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3799 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3800 sh->reconstruct_state = reconstruct_state_idle;
3802 /* All the 'written' buffers and the parity block are ready to
3803 * be written back to disk
3805 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3806 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3807 BUG_ON(sh->qd_idx >= 0 &&
3808 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3809 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3810 for (i = disks; i--; ) {
3811 struct r5dev *dev = &sh->dev[i];
3812 if (test_bit(R5_LOCKED, &dev->flags) &&
3813 (i == sh->pd_idx || i == sh->qd_idx ||
3815 pr_debug("Writing block %d\n", i);
3816 set_bit(R5_Wantwrite, &dev->flags);
3821 if (!test_bit(R5_Insync, &dev->flags) ||
3822 ((i == sh->pd_idx || i == sh->qd_idx) &&
3824 set_bit(STRIPE_INSYNC, &sh->state);
3827 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3828 s.dec_preread_active = 1;
3832 * might be able to return some write requests if the parity blocks
3833 * are safe, or on a failed drive
3835 pdev = &sh->dev[sh->pd_idx];
3836 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3837 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3838 qdev = &sh->dev[sh->qd_idx];
3839 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3840 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3844 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3845 && !test_bit(R5_LOCKED, &pdev->flags)
3846 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3847 test_bit(R5_Discard, &pdev->flags))))) &&
3848 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3849 && !test_bit(R5_LOCKED, &qdev->flags)
3850 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3851 test_bit(R5_Discard, &qdev->flags))))))
3852 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3854 /* Now we might consider reading some blocks, either to check/generate
3855 * parity, or to satisfy requests
3856 * or to load a block that is being partially written.
3858 if (s.to_read || s.non_overwrite
3859 || (conf->level == 6 && s.to_write && s.failed)
3860 || (s.syncing && (s.uptodate + s.compute < disks))
3863 handle_stripe_fill(sh, &s, disks);
3865 /* Now to consider new write requests and what else, if anything
3866 * should be read. We do not handle new writes when:
3867 * 1/ A 'write' operation (copy+xor) is already in flight.
3868 * 2/ A 'check' operation is in flight, as it may clobber the parity
3871 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3872 handle_stripe_dirtying(conf, sh, &s, disks);
3874 /* maybe we need to check and possibly fix the parity for this stripe
3875 * Any reads will already have been scheduled, so we just see if enough
3876 * data is available. The parity check is held off while parity
3877 * dependent operations are in flight.
3879 if (sh->check_state ||
3880 (s.syncing && s.locked == 0 &&
3881 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3882 !test_bit(STRIPE_INSYNC, &sh->state))) {
3883 if (conf->level == 6)
3884 handle_parity_checks6(conf, sh, &s, disks);
3886 handle_parity_checks5(conf, sh, &s, disks);
3889 if ((s.replacing || s.syncing) && s.locked == 0
3890 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3891 && !test_bit(STRIPE_REPLACED, &sh->state)) {
3892 /* Write out to replacement devices where possible */
3893 for (i = 0; i < conf->raid_disks; i++)
3894 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3895 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
3896 set_bit(R5_WantReplace, &sh->dev[i].flags);
3897 set_bit(R5_LOCKED, &sh->dev[i].flags);
3901 set_bit(STRIPE_INSYNC, &sh->state);
3902 set_bit(STRIPE_REPLACED, &sh->state);
3904 if ((s.syncing || s.replacing) && s.locked == 0 &&
3905 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3906 test_bit(STRIPE_INSYNC, &sh->state)) {
3907 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3908 clear_bit(STRIPE_SYNCING, &sh->state);
3909 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3910 wake_up(&conf->wait_for_overlap);
3913 /* If the failed drives are just a ReadError, then we might need
3914 * to progress the repair/check process
3916 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3917 for (i = 0; i < s.failed; i++) {
3918 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3919 if (test_bit(R5_ReadError, &dev->flags)
3920 && !test_bit(R5_LOCKED, &dev->flags)
3921 && test_bit(R5_UPTODATE, &dev->flags)
3923 if (!test_bit(R5_ReWrite, &dev->flags)) {
3924 set_bit(R5_Wantwrite, &dev->flags);
3925 set_bit(R5_ReWrite, &dev->flags);
3926 set_bit(R5_LOCKED, &dev->flags);
3929 /* let's read it back */
3930 set_bit(R5_Wantread, &dev->flags);
3931 set_bit(R5_LOCKED, &dev->flags);
3937 /* Finish reconstruct operations initiated by the expansion process */
3938 if (sh->reconstruct_state == reconstruct_state_result) {
3939 struct stripe_head *sh_src
3940 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3941 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3942 /* sh cannot be written until sh_src has been read.
3943 * so arrange for sh to be delayed a little
3945 set_bit(STRIPE_DELAYED, &sh->state);
3946 set_bit(STRIPE_HANDLE, &sh->state);
3947 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3949 atomic_inc(&conf->preread_active_stripes);
3950 release_stripe(sh_src);
3954 release_stripe(sh_src);
3956 sh->reconstruct_state = reconstruct_state_idle;
3957 clear_bit(STRIPE_EXPANDING, &sh->state);
3958 for (i = conf->raid_disks; i--; ) {
3959 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3960 set_bit(R5_LOCKED, &sh->dev[i].flags);
3965 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3966 !sh->reconstruct_state) {
3967 /* Need to write out all blocks after computing parity */
3968 sh->disks = conf->raid_disks;
3969 stripe_set_idx(sh->sector, conf, 0, sh);
3970 schedule_reconstruction(sh, &s, 1, 1);
3971 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3972 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3973 atomic_dec(&conf->reshape_stripes);
3974 wake_up(&conf->wait_for_overlap);
3975 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3978 if (s.expanding && s.locked == 0 &&
3979 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3980 handle_stripe_expansion(conf, sh);
3983 /* wait for this device to become unblocked */
3984 if (unlikely(s.blocked_rdev)) {
3985 if (conf->mddev->external)
3986 md_wait_for_blocked_rdev(s.blocked_rdev,
3989 /* Internal metadata will immediately
3990 * be written by raid5d, so we don't
3991 * need to wait here.
3993 rdev_dec_pending(s.blocked_rdev,
3997 if (s.handle_bad_blocks)
3998 for (i = disks; i--; ) {
3999 struct md_rdev *rdev;
4000 struct r5dev *dev = &sh->dev[i];
4001 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4002 /* We own a safe reference to the rdev */
4003 rdev = conf->disks[i].rdev;
4004 if (!rdev_set_badblocks(rdev, sh->sector,
4006 md_error(conf->mddev, rdev);
4007 rdev_dec_pending(rdev, conf->mddev);
4009 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4010 rdev = conf->disks[i].rdev;
4011 rdev_clear_badblocks(rdev, sh->sector,
4013 rdev_dec_pending(rdev, conf->mddev);
4015 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4016 rdev = conf->disks[i].replacement;
4018 /* rdev have been moved down */
4019 rdev = conf->disks[i].rdev;
4020 rdev_clear_badblocks(rdev, sh->sector,
4022 rdev_dec_pending(rdev, conf->mddev);
4027 raid_run_ops(sh, s.ops_request);
4031 if (s.dec_preread_active) {
4032 /* We delay this until after ops_run_io so that if make_request
4033 * is waiting on a flush, it won't continue until the writes
4034 * have actually been submitted.
4036 atomic_dec(&conf->preread_active_stripes);
4037 if (atomic_read(&conf->preread_active_stripes) <
4039 md_wakeup_thread(conf->mddev->thread);
4042 return_io(s.return_bi);
4044 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4047 static void raid5_activate_delayed(struct r5conf *conf)
4049 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4050 while (!list_empty(&conf->delayed_list)) {
4051 struct list_head *l = conf->delayed_list.next;
4052 struct stripe_head *sh;
4053 sh = list_entry(l, struct stripe_head, lru);
4055 clear_bit(STRIPE_DELAYED, &sh->state);
4056 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4057 atomic_inc(&conf->preread_active_stripes);
4058 list_add_tail(&sh->lru, &conf->hold_list);
4059 raid5_wakeup_stripe_thread(sh);
4064 static void activate_bit_delay(struct r5conf *conf,
4065 struct list_head *temp_inactive_list)
4067 /* device_lock is held */
4068 struct list_head head;
4069 list_add(&head, &conf->bitmap_list);
4070 list_del_init(&conf->bitmap_list);
4071 while (!list_empty(&head)) {
4072 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4074 list_del_init(&sh->lru);
4075 atomic_inc(&sh->count);
4076 hash = sh->hash_lock_index;
4077 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4081 int md_raid5_congested(struct mddev *mddev, int bits)
4083 struct r5conf *conf = mddev->private;
4085 /* No difference between reads and writes. Just check
4086 * how busy the stripe_cache is
4089 if (conf->inactive_blocked)
4093 if (atomic_read(&conf->empty_inactive_list_nr))
4098 EXPORT_SYMBOL_GPL(md_raid5_congested);
4100 static int raid5_congested(void *data, int bits)
4102 struct mddev *mddev = data;
4104 return mddev_congested(mddev, bits) ||
4105 md_raid5_congested(mddev, bits);
4108 /* We want read requests to align with chunks where possible,
4109 * but write requests don't need to.
4111 static int raid5_mergeable_bvec(struct request_queue *q,
4112 struct bvec_merge_data *bvm,
4113 struct bio_vec *biovec)
4115 struct mddev *mddev = q->queuedata;
4116 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4118 unsigned int chunk_sectors = mddev->chunk_sectors;
4119 unsigned int bio_sectors = bvm->bi_size >> 9;
4121 if ((bvm->bi_rw & 1) == WRITE)
4122 return biovec->bv_len; /* always allow writes to be mergeable */
4124 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4125 chunk_sectors = mddev->new_chunk_sectors;
4126 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4127 if (max < 0) max = 0;
4128 if (max <= biovec->bv_len && bio_sectors == 0)
4129 return biovec->bv_len;
4134 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4136 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4137 unsigned int chunk_sectors = mddev->chunk_sectors;
4138 unsigned int bio_sectors = bio_sectors(bio);
4140 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4141 chunk_sectors = mddev->new_chunk_sectors;
4142 return chunk_sectors >=
4143 ((sector & (chunk_sectors - 1)) + bio_sectors);
4147 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4148 * later sampled by raid5d.
4150 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4152 unsigned long flags;
4154 spin_lock_irqsave(&conf->device_lock, flags);
4156 bi->bi_next = conf->retry_read_aligned_list;
4157 conf->retry_read_aligned_list = bi;
4159 spin_unlock_irqrestore(&conf->device_lock, flags);
4160 md_wakeup_thread(conf->mddev->thread);
4163 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4167 bi = conf->retry_read_aligned;
4169 conf->retry_read_aligned = NULL;
4172 bi = conf->retry_read_aligned_list;
4174 conf->retry_read_aligned_list = bi->bi_next;
4177 * this sets the active strip count to 1 and the processed
4178 * strip count to zero (upper 8 bits)
4180 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4187 * The "raid5_align_endio" should check if the read succeeded and if it
4188 * did, call bio_endio on the original bio (having bio_put the new bio
4190 * If the read failed..
4192 static void raid5_align_endio(struct bio *bi, int error)
4194 struct bio* raid_bi = bi->bi_private;
4195 struct mddev *mddev;
4196 struct r5conf *conf;
4197 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4198 struct md_rdev *rdev;
4202 rdev = (void*)raid_bi->bi_next;
4203 raid_bi->bi_next = NULL;
4204 mddev = rdev->mddev;
4205 conf = mddev->private;
4207 rdev_dec_pending(rdev, conf->mddev);
4209 if (!error && uptodate) {
4210 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4212 bio_endio(raid_bi, 0);
4213 if (atomic_dec_and_test(&conf->active_aligned_reads))
4214 wake_up(&conf->wait_for_stripe);
4218 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4220 add_bio_to_retry(raid_bi, conf);
4223 static int bio_fits_rdev(struct bio *bi)
4225 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4227 if (bio_sectors(bi) > queue_max_sectors(q))
4229 blk_recount_segments(q, bi);
4230 if (bi->bi_phys_segments > queue_max_segments(q))
4233 if (q->merge_bvec_fn)
4234 /* it's too hard to apply the merge_bvec_fn at this stage,
4242 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4244 struct r5conf *conf = mddev->private;
4246 struct bio* align_bi;
4247 struct md_rdev *rdev;
4248 sector_t end_sector;
4250 if (!in_chunk_boundary(mddev, raid_bio)) {
4251 pr_debug("chunk_aligned_read : non aligned\n");
4255 * use bio_clone_mddev to make a copy of the bio
4257 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4261 * set bi_end_io to a new function, and set bi_private to the
4264 align_bi->bi_end_io = raid5_align_endio;
4265 align_bi->bi_private = raid_bio;
4269 align_bi->bi_iter.bi_sector =
4270 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4273 end_sector = bio_end_sector(align_bi);
4275 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4276 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4277 rdev->recovery_offset < end_sector) {
4278 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4280 (test_bit(Faulty, &rdev->flags) ||
4281 !(test_bit(In_sync, &rdev->flags) ||
4282 rdev->recovery_offset >= end_sector)))
4289 atomic_inc(&rdev->nr_pending);
4291 raid_bio->bi_next = (void*)rdev;
4292 align_bi->bi_bdev = rdev->bdev;
4293 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
4295 if (!bio_fits_rdev(align_bi) ||
4296 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4297 bio_sectors(align_bi),
4298 &first_bad, &bad_sectors)) {
4299 /* too big in some way, or has a known bad block */
4301 rdev_dec_pending(rdev, mddev);
4305 /* No reshape active, so we can trust rdev->data_offset */
4306 align_bi->bi_iter.bi_sector += rdev->data_offset;
4308 spin_lock_irq(&conf->device_lock);
4309 wait_event_lock_irq(conf->wait_for_stripe,
4312 atomic_inc(&conf->active_aligned_reads);
4313 spin_unlock_irq(&conf->device_lock);
4316 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4317 align_bi, disk_devt(mddev->gendisk),
4318 raid_bio->bi_iter.bi_sector);
4319 generic_make_request(align_bi);
4328 /* __get_priority_stripe - get the next stripe to process
4330 * Full stripe writes are allowed to pass preread active stripes up until
4331 * the bypass_threshold is exceeded. In general the bypass_count
4332 * increments when the handle_list is handled before the hold_list; however, it
4333 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4334 * stripe with in flight i/o. The bypass_count will be reset when the
4335 * head of the hold_list has changed, i.e. the head was promoted to the
4338 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4340 struct stripe_head *sh = NULL, *tmp;
4341 struct list_head *handle_list = NULL;
4342 struct r5worker_group *wg = NULL;
4344 if (conf->worker_cnt_per_group == 0) {
4345 handle_list = &conf->handle_list;
4346 } else if (group != ANY_GROUP) {
4347 handle_list = &conf->worker_groups[group].handle_list;
4348 wg = &conf->worker_groups[group];
4351 for (i = 0; i < conf->group_cnt; i++) {
4352 handle_list = &conf->worker_groups[i].handle_list;
4353 wg = &conf->worker_groups[i];
4354 if (!list_empty(handle_list))
4359 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4361 list_empty(handle_list) ? "empty" : "busy",
4362 list_empty(&conf->hold_list) ? "empty" : "busy",
4363 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4365 if (!list_empty(handle_list)) {
4366 sh = list_entry(handle_list->next, typeof(*sh), lru);
4368 if (list_empty(&conf->hold_list))
4369 conf->bypass_count = 0;
4370 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4371 if (conf->hold_list.next == conf->last_hold)
4372 conf->bypass_count++;
4374 conf->last_hold = conf->hold_list.next;
4375 conf->bypass_count -= conf->bypass_threshold;
4376 if (conf->bypass_count < 0)
4377 conf->bypass_count = 0;
4380 } else if (!list_empty(&conf->hold_list) &&
4381 ((conf->bypass_threshold &&
4382 conf->bypass_count > conf->bypass_threshold) ||
4383 atomic_read(&conf->pending_full_writes) == 0)) {
4385 list_for_each_entry(tmp, &conf->hold_list, lru) {
4386 if (conf->worker_cnt_per_group == 0 ||
4387 group == ANY_GROUP ||
4388 !cpu_online(tmp->cpu) ||
4389 cpu_to_group(tmp->cpu) == group) {
4396 conf->bypass_count -= conf->bypass_threshold;
4397 if (conf->bypass_count < 0)
4398 conf->bypass_count = 0;
4410 list_del_init(&sh->lru);
4411 BUG_ON(atomic_inc_return(&sh->count) != 1);
4415 struct raid5_plug_cb {
4416 struct blk_plug_cb cb;
4417 struct list_head list;
4418 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4421 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4423 struct raid5_plug_cb *cb = container_of(
4424 blk_cb, struct raid5_plug_cb, cb);
4425 struct stripe_head *sh;
4426 struct mddev *mddev = cb->cb.data;
4427 struct r5conf *conf = mddev->private;
4431 if (cb->list.next && !list_empty(&cb->list)) {
4432 spin_lock_irq(&conf->device_lock);
4433 while (!list_empty(&cb->list)) {
4434 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4435 list_del_init(&sh->lru);
4437 * avoid race release_stripe_plug() sees
4438 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4439 * is still in our list
4441 smp_mb__before_atomic();
4442 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4444 * STRIPE_ON_RELEASE_LIST could be set here. In that
4445 * case, the count is always > 1 here
4447 hash = sh->hash_lock_index;
4448 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
4451 spin_unlock_irq(&conf->device_lock);
4453 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4454 NR_STRIPE_HASH_LOCKS);
4456 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4460 static void release_stripe_plug(struct mddev *mddev,
4461 struct stripe_head *sh)
4463 struct blk_plug_cb *blk_cb = blk_check_plugged(
4464 raid5_unplug, mddev,
4465 sizeof(struct raid5_plug_cb));
4466 struct raid5_plug_cb *cb;
4473 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4475 if (cb->list.next == NULL) {
4477 INIT_LIST_HEAD(&cb->list);
4478 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
4479 INIT_LIST_HEAD(cb->temp_inactive_list + i);
4482 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4483 list_add_tail(&sh->lru, &cb->list);
4488 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4490 struct r5conf *conf = mddev->private;
4491 sector_t logical_sector, last_sector;
4492 struct stripe_head *sh;
4496 if (mddev->reshape_position != MaxSector)
4497 /* Skip discard while reshape is happening */
4500 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4501 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
4504 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4506 stripe_sectors = conf->chunk_sectors *
4507 (conf->raid_disks - conf->max_degraded);
4508 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4510 sector_div(last_sector, stripe_sectors);
4512 logical_sector *= conf->chunk_sectors;
4513 last_sector *= conf->chunk_sectors;
4515 for (; logical_sector < last_sector;
4516 logical_sector += STRIPE_SECTORS) {
4520 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4521 prepare_to_wait(&conf->wait_for_overlap, &w,
4522 TASK_UNINTERRUPTIBLE);
4523 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4524 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4529 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4530 spin_lock_irq(&sh->stripe_lock);
4531 for (d = 0; d < conf->raid_disks; d++) {
4532 if (d == sh->pd_idx || d == sh->qd_idx)
4534 if (sh->dev[d].towrite || sh->dev[d].toread) {
4535 set_bit(R5_Overlap, &sh->dev[d].flags);
4536 spin_unlock_irq(&sh->stripe_lock);
4542 set_bit(STRIPE_DISCARD, &sh->state);
4543 finish_wait(&conf->wait_for_overlap, &w);
4544 for (d = 0; d < conf->raid_disks; d++) {
4545 if (d == sh->pd_idx || d == sh->qd_idx)
4547 sh->dev[d].towrite = bi;
4548 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4549 raid5_inc_bi_active_stripes(bi);
4551 spin_unlock_irq(&sh->stripe_lock);
4552 if (conf->mddev->bitmap) {
4554 d < conf->raid_disks - conf->max_degraded;
4556 bitmap_startwrite(mddev->bitmap,
4560 sh->bm_seq = conf->seq_flush + 1;
4561 set_bit(STRIPE_BIT_DELAY, &sh->state);
4564 set_bit(STRIPE_HANDLE, &sh->state);
4565 clear_bit(STRIPE_DELAYED, &sh->state);
4566 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4567 atomic_inc(&conf->preread_active_stripes);
4568 release_stripe_plug(mddev, sh);
4571 remaining = raid5_dec_bi_active_stripes(bi);
4572 if (remaining == 0) {
4573 md_write_end(mddev);
4578 static void make_request(struct mddev *mddev, struct bio * bi)
4580 struct r5conf *conf = mddev->private;
4582 sector_t new_sector;
4583 sector_t logical_sector, last_sector;
4584 struct stripe_head *sh;
4585 const int rw = bio_data_dir(bi);
4590 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4591 md_flush_request(mddev, bi);
4595 md_write_start(mddev, bi);
4598 mddev->reshape_position == MaxSector &&
4599 chunk_aligned_read(mddev,bi))
4602 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4603 make_discard_request(mddev, bi);
4607 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4608 last_sector = bio_end_sector(bi);
4610 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4612 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4613 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4619 seq = read_seqcount_begin(&conf->gen_lock);
4622 prepare_to_wait(&conf->wait_for_overlap, &w,
4623 TASK_UNINTERRUPTIBLE);
4624 if (unlikely(conf->reshape_progress != MaxSector)) {
4625 /* spinlock is needed as reshape_progress may be
4626 * 64bit on a 32bit platform, and so it might be
4627 * possible to see a half-updated value
4628 * Of course reshape_progress could change after
4629 * the lock is dropped, so once we get a reference
4630 * to the stripe that we think it is, we will have
4633 spin_lock_irq(&conf->device_lock);
4634 if (mddev->reshape_backwards
4635 ? logical_sector < conf->reshape_progress
4636 : logical_sector >= conf->reshape_progress) {
4639 if (mddev->reshape_backwards
4640 ? logical_sector < conf->reshape_safe
4641 : logical_sector >= conf->reshape_safe) {
4642 spin_unlock_irq(&conf->device_lock);
4648 spin_unlock_irq(&conf->device_lock);
4651 new_sector = raid5_compute_sector(conf, logical_sector,
4654 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4655 (unsigned long long)new_sector,
4656 (unsigned long long)logical_sector);
4658 sh = get_active_stripe(conf, new_sector, previous,
4659 (bi->bi_rw&RWA_MASK), 0);
4661 if (unlikely(previous)) {
4662 /* expansion might have moved on while waiting for a
4663 * stripe, so we must do the range check again.
4664 * Expansion could still move past after this
4665 * test, but as we are holding a reference to
4666 * 'sh', we know that if that happens,
4667 * STRIPE_EXPANDING will get set and the expansion
4668 * won't proceed until we finish with the stripe.
4671 spin_lock_irq(&conf->device_lock);
4672 if (mddev->reshape_backwards
4673 ? logical_sector >= conf->reshape_progress
4674 : logical_sector < conf->reshape_progress)
4675 /* mismatch, need to try again */
4677 spin_unlock_irq(&conf->device_lock);
4685 if (read_seqcount_retry(&conf->gen_lock, seq)) {
4686 /* Might have got the wrong stripe_head
4694 logical_sector >= mddev->suspend_lo &&
4695 logical_sector < mddev->suspend_hi) {
4697 /* As the suspend_* range is controlled by
4698 * userspace, we want an interruptible
4701 flush_signals(current);
4702 prepare_to_wait(&conf->wait_for_overlap,
4703 &w, TASK_INTERRUPTIBLE);
4704 if (logical_sector >= mddev->suspend_lo &&
4705 logical_sector < mddev->suspend_hi) {
4712 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4713 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4714 /* Stripe is busy expanding or
4715 * add failed due to overlap. Flush everything
4718 md_wakeup_thread(mddev->thread);
4724 set_bit(STRIPE_HANDLE, &sh->state);
4725 clear_bit(STRIPE_DELAYED, &sh->state);
4726 if ((bi->bi_rw & REQ_SYNC) &&
4727 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4728 atomic_inc(&conf->preread_active_stripes);
4729 release_stripe_plug(mddev, sh);
4731 /* cannot get stripe for read-ahead, just give-up */
4732 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4736 finish_wait(&conf->wait_for_overlap, &w);
4738 remaining = raid5_dec_bi_active_stripes(bi);
4739 if (remaining == 0) {
4742 md_write_end(mddev);
4744 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4750 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4752 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4754 /* reshaping is quite different to recovery/resync so it is
4755 * handled quite separately ... here.
4757 * On each call to sync_request, we gather one chunk worth of
4758 * destination stripes and flag them as expanding.
4759 * Then we find all the source stripes and request reads.
4760 * As the reads complete, handle_stripe will copy the data
4761 * into the destination stripe and release that stripe.
4763 struct r5conf *conf = mddev->private;
4764 struct stripe_head *sh;
4765 sector_t first_sector, last_sector;
4766 int raid_disks = conf->previous_raid_disks;
4767 int data_disks = raid_disks - conf->max_degraded;
4768 int new_data_disks = conf->raid_disks - conf->max_degraded;
4771 sector_t writepos, readpos, safepos;
4772 sector_t stripe_addr;
4773 int reshape_sectors;
4774 struct list_head stripes;
4776 if (sector_nr == 0) {
4777 /* If restarting in the middle, skip the initial sectors */
4778 if (mddev->reshape_backwards &&
4779 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4780 sector_nr = raid5_size(mddev, 0, 0)
4781 - conf->reshape_progress;
4782 } else if (!mddev->reshape_backwards &&
4783 conf->reshape_progress > 0)
4784 sector_nr = conf->reshape_progress;
4785 sector_div(sector_nr, new_data_disks);
4787 mddev->curr_resync_completed = sector_nr;
4788 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4794 /* We need to process a full chunk at a time.
4795 * If old and new chunk sizes differ, we need to process the
4798 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4799 reshape_sectors = mddev->new_chunk_sectors;
4801 reshape_sectors = mddev->chunk_sectors;
4803 /* We update the metadata at least every 10 seconds, or when
4804 * the data about to be copied would over-write the source of
4805 * the data at the front of the range. i.e. one new_stripe
4806 * along from reshape_progress new_maps to after where
4807 * reshape_safe old_maps to
4809 writepos = conf->reshape_progress;
4810 sector_div(writepos, new_data_disks);
4811 readpos = conf->reshape_progress;
4812 sector_div(readpos, data_disks);
4813 safepos = conf->reshape_safe;
4814 sector_div(safepos, data_disks);
4815 if (mddev->reshape_backwards) {
4816 writepos -= min_t(sector_t, reshape_sectors, writepos);
4817 readpos += reshape_sectors;
4818 safepos += reshape_sectors;
4820 writepos += reshape_sectors;
4821 readpos -= min_t(sector_t, reshape_sectors, readpos);
4822 safepos -= min_t(sector_t, reshape_sectors, safepos);
4825 /* Having calculated the 'writepos' possibly use it
4826 * to set 'stripe_addr' which is where we will write to.
4828 if (mddev->reshape_backwards) {
4829 BUG_ON(conf->reshape_progress == 0);
4830 stripe_addr = writepos;
4831 BUG_ON((mddev->dev_sectors &
4832 ~((sector_t)reshape_sectors - 1))
4833 - reshape_sectors - stripe_addr
4836 BUG_ON(writepos != sector_nr + reshape_sectors);
4837 stripe_addr = sector_nr;
4840 /* 'writepos' is the most advanced device address we might write.
4841 * 'readpos' is the least advanced device address we might read.
4842 * 'safepos' is the least address recorded in the metadata as having
4844 * If there is a min_offset_diff, these are adjusted either by
4845 * increasing the safepos/readpos if diff is negative, or
4846 * increasing writepos if diff is positive.
4847 * If 'readpos' is then behind 'writepos', there is no way that we can
4848 * ensure safety in the face of a crash - that must be done by userspace
4849 * making a backup of the data. So in that case there is no particular
4850 * rush to update metadata.
4851 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4852 * update the metadata to advance 'safepos' to match 'readpos' so that
4853 * we can be safe in the event of a crash.
4854 * So we insist on updating metadata if safepos is behind writepos and
4855 * readpos is beyond writepos.
4856 * In any case, update the metadata every 10 seconds.
4857 * Maybe that number should be configurable, but I'm not sure it is
4858 * worth it.... maybe it could be a multiple of safemode_delay???
4860 if (conf->min_offset_diff < 0) {
4861 safepos += -conf->min_offset_diff;
4862 readpos += -conf->min_offset_diff;
4864 writepos += conf->min_offset_diff;
4866 if ((mddev->reshape_backwards
4867 ? (safepos > writepos && readpos < writepos)
4868 : (safepos < writepos && readpos > writepos)) ||
4869 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4870 /* Cannot proceed until we've updated the superblock... */
4871 wait_event(conf->wait_for_overlap,
4872 atomic_read(&conf->reshape_stripes)==0
4873 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4874 if (atomic_read(&conf->reshape_stripes) != 0)
4876 mddev->reshape_position = conf->reshape_progress;
4877 mddev->curr_resync_completed = sector_nr;
4878 conf->reshape_checkpoint = jiffies;
4879 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4880 md_wakeup_thread(mddev->thread);
4881 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4882 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4883 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4885 spin_lock_irq(&conf->device_lock);
4886 conf->reshape_safe = mddev->reshape_position;
4887 spin_unlock_irq(&conf->device_lock);
4888 wake_up(&conf->wait_for_overlap);
4889 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4892 INIT_LIST_HEAD(&stripes);
4893 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4895 int skipped_disk = 0;
4896 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4897 set_bit(STRIPE_EXPANDING, &sh->state);
4898 atomic_inc(&conf->reshape_stripes);
4899 /* If any of this stripe is beyond the end of the old
4900 * array, then we need to zero those blocks
4902 for (j=sh->disks; j--;) {
4904 if (j == sh->pd_idx)
4906 if (conf->level == 6 &&
4909 s = compute_blocknr(sh, j, 0);
4910 if (s < raid5_size(mddev, 0, 0)) {
4914 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4915 set_bit(R5_Expanded, &sh->dev[j].flags);
4916 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4918 if (!skipped_disk) {
4919 set_bit(STRIPE_EXPAND_READY, &sh->state);
4920 set_bit(STRIPE_HANDLE, &sh->state);
4922 list_add(&sh->lru, &stripes);
4924 spin_lock_irq(&conf->device_lock);
4925 if (mddev->reshape_backwards)
4926 conf->reshape_progress -= reshape_sectors * new_data_disks;
4928 conf->reshape_progress += reshape_sectors * new_data_disks;
4929 spin_unlock_irq(&conf->device_lock);
4930 /* Ok, those stripe are ready. We can start scheduling
4931 * reads on the source stripes.
4932 * The source stripes are determined by mapping the first and last
4933 * block on the destination stripes.
4936 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4939 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4940 * new_data_disks - 1),
4942 if (last_sector >= mddev->dev_sectors)
4943 last_sector = mddev->dev_sectors - 1;
4944 while (first_sector <= last_sector) {
4945 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4946 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4947 set_bit(STRIPE_HANDLE, &sh->state);
4949 first_sector += STRIPE_SECTORS;
4951 /* Now that the sources are clearly marked, we can release
4952 * the destination stripes
4954 while (!list_empty(&stripes)) {
4955 sh = list_entry(stripes.next, struct stripe_head, lru);
4956 list_del_init(&sh->lru);
4959 /* If this takes us to the resync_max point where we have to pause,
4960 * then we need to write out the superblock.
4962 sector_nr += reshape_sectors;
4963 if ((sector_nr - mddev->curr_resync_completed) * 2
4964 >= mddev->resync_max - mddev->curr_resync_completed) {
4965 /* Cannot proceed until we've updated the superblock... */
4966 wait_event(conf->wait_for_overlap,
4967 atomic_read(&conf->reshape_stripes) == 0
4968 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4969 if (atomic_read(&conf->reshape_stripes) != 0)
4971 mddev->reshape_position = conf->reshape_progress;
4972 mddev->curr_resync_completed = sector_nr;
4973 conf->reshape_checkpoint = jiffies;
4974 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4975 md_wakeup_thread(mddev->thread);
4976 wait_event(mddev->sb_wait,
4977 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4978 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4979 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4981 spin_lock_irq(&conf->device_lock);
4982 conf->reshape_safe = mddev->reshape_position;
4983 spin_unlock_irq(&conf->device_lock);
4984 wake_up(&conf->wait_for_overlap);
4985 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4988 return reshape_sectors;
4991 /* FIXME go_faster isn't used */
4992 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4994 struct r5conf *conf = mddev->private;
4995 struct stripe_head *sh;
4996 sector_t max_sector = mddev->dev_sectors;
4997 sector_t sync_blocks;
4998 int still_degraded = 0;
5001 if (sector_nr >= max_sector) {
5002 /* just being told to finish up .. nothing much to do */
5004 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5009 if (mddev->curr_resync < max_sector) /* aborted */
5010 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5012 else /* completed sync */
5014 bitmap_close_sync(mddev->bitmap);
5019 /* Allow raid5_quiesce to complete */
5020 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5022 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5023 return reshape_request(mddev, sector_nr, skipped);
5025 /* No need to check resync_max as we never do more than one
5026 * stripe, and as resync_max will always be on a chunk boundary,
5027 * if the check in md_do_sync didn't fire, there is no chance
5028 * of overstepping resync_max here
5031 /* if there is too many failed drives and we are trying
5032 * to resync, then assert that we are finished, because there is
5033 * nothing we can do.
5035 if (mddev->degraded >= conf->max_degraded &&
5036 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5037 sector_t rv = mddev->dev_sectors - sector_nr;
5041 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5043 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5044 sync_blocks >= STRIPE_SECTORS) {
5045 /* we can skip this block, and probably more */
5046 sync_blocks /= STRIPE_SECTORS;
5048 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5051 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5053 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5055 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5056 /* make sure we don't swamp the stripe cache if someone else
5057 * is trying to get access
5059 schedule_timeout_uninterruptible(1);
5061 /* Need to check if array will still be degraded after recovery/resync
5062 * We don't need to check the 'failed' flag as when that gets set,
5065 for (i = 0; i < conf->raid_disks; i++)
5066 if (conf->disks[i].rdev == NULL)
5069 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5071 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5072 set_bit(STRIPE_HANDLE, &sh->state);
5076 return STRIPE_SECTORS;
5079 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5081 /* We may not be able to submit a whole bio at once as there
5082 * may not be enough stripe_heads available.
5083 * We cannot pre-allocate enough stripe_heads as we may need
5084 * more than exist in the cache (if we allow ever large chunks).
5085 * So we do one stripe head at a time and record in
5086 * ->bi_hw_segments how many have been done.
5088 * We *know* that this entire raid_bio is in one chunk, so
5089 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5091 struct stripe_head *sh;
5093 sector_t sector, logical_sector, last_sector;
5098 logical_sector = raid_bio->bi_iter.bi_sector &
5099 ~((sector_t)STRIPE_SECTORS-1);
5100 sector = raid5_compute_sector(conf, logical_sector,
5102 last_sector = bio_end_sector(raid_bio);
5104 for (; logical_sector < last_sector;
5105 logical_sector += STRIPE_SECTORS,
5106 sector += STRIPE_SECTORS,
5109 if (scnt < raid5_bi_processed_stripes(raid_bio))
5110 /* already done this stripe */
5113 sh = get_active_stripe(conf, sector, 0, 1, 1);
5116 /* failed to get a stripe - must wait */
5117 raid5_set_bi_processed_stripes(raid_bio, scnt);
5118 conf->retry_read_aligned = raid_bio;
5122 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
5124 raid5_set_bi_processed_stripes(raid_bio, scnt);
5125 conf->retry_read_aligned = raid_bio;
5129 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5134 remaining = raid5_dec_bi_active_stripes(raid_bio);
5135 if (remaining == 0) {
5136 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5138 bio_endio(raid_bio, 0);
5140 if (atomic_dec_and_test(&conf->active_aligned_reads))
5141 wake_up(&conf->wait_for_stripe);
5145 static int handle_active_stripes(struct r5conf *conf, int group,
5146 struct r5worker *worker,
5147 struct list_head *temp_inactive_list)
5149 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5150 int i, batch_size = 0, hash;
5151 bool release_inactive = false;
5153 while (batch_size < MAX_STRIPE_BATCH &&
5154 (sh = __get_priority_stripe(conf, group)) != NULL)
5155 batch[batch_size++] = sh;
5157 if (batch_size == 0) {
5158 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5159 if (!list_empty(temp_inactive_list + i))
5161 if (i == NR_STRIPE_HASH_LOCKS)
5163 release_inactive = true;
5165 spin_unlock_irq(&conf->device_lock);
5167 release_inactive_stripe_list(conf, temp_inactive_list,
5168 NR_STRIPE_HASH_LOCKS);
5170 if (release_inactive) {
5171 spin_lock_irq(&conf->device_lock);
5175 for (i = 0; i < batch_size; i++)
5176 handle_stripe(batch[i]);
5180 spin_lock_irq(&conf->device_lock);
5181 for (i = 0; i < batch_size; i++) {
5182 hash = batch[i]->hash_lock_index;
5183 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5188 static void raid5_do_work(struct work_struct *work)
5190 struct r5worker *worker = container_of(work, struct r5worker, work);
5191 struct r5worker_group *group = worker->group;
5192 struct r5conf *conf = group->conf;
5193 int group_id = group - conf->worker_groups;
5195 struct blk_plug plug;
5197 pr_debug("+++ raid5worker active\n");
5199 blk_start_plug(&plug);
5201 spin_lock_irq(&conf->device_lock);
5203 int batch_size, released;
5205 released = release_stripe_list(conf, worker->temp_inactive_list);
5207 batch_size = handle_active_stripes(conf, group_id, worker,
5208 worker->temp_inactive_list);
5209 worker->working = false;
5210 if (!batch_size && !released)
5212 handled += batch_size;
5214 pr_debug("%d stripes handled\n", handled);
5216 spin_unlock_irq(&conf->device_lock);
5217 blk_finish_plug(&plug);
5219 pr_debug("--- raid5worker inactive\n");
5223 * This is our raid5 kernel thread.
5225 * We scan the hash table for stripes which can be handled now.
5226 * During the scan, completed stripes are saved for us by the interrupt
5227 * handler, so that they will not have to wait for our next wakeup.
5229 static void raid5d(struct md_thread *thread)
5231 struct mddev *mddev = thread->mddev;
5232 struct r5conf *conf = mddev->private;
5234 struct blk_plug plug;
5236 pr_debug("+++ raid5d active\n");
5238 md_check_recovery(mddev);
5240 blk_start_plug(&plug);
5242 spin_lock_irq(&conf->device_lock);
5245 int batch_size, released;
5247 released = release_stripe_list(conf, conf->temp_inactive_list);
5250 !list_empty(&conf->bitmap_list)) {
5251 /* Now is a good time to flush some bitmap updates */
5253 spin_unlock_irq(&conf->device_lock);
5254 bitmap_unplug(mddev->bitmap);
5255 spin_lock_irq(&conf->device_lock);
5256 conf->seq_write = conf->seq_flush;
5257 activate_bit_delay(conf, conf->temp_inactive_list);
5259 raid5_activate_delayed(conf);
5261 while ((bio = remove_bio_from_retry(conf))) {
5263 spin_unlock_irq(&conf->device_lock);
5264 ok = retry_aligned_read(conf, bio);
5265 spin_lock_irq(&conf->device_lock);
5271 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5272 conf->temp_inactive_list);
5273 if (!batch_size && !released)
5275 handled += batch_size;
5277 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5278 spin_unlock_irq(&conf->device_lock);
5279 md_check_recovery(mddev);
5280 spin_lock_irq(&conf->device_lock);
5283 pr_debug("%d stripes handled\n", handled);
5285 spin_unlock_irq(&conf->device_lock);
5287 async_tx_issue_pending_all();
5288 blk_finish_plug(&plug);
5290 pr_debug("--- raid5d inactive\n");
5294 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5296 struct r5conf *conf = mddev->private;
5298 return sprintf(page, "%d\n", conf->max_nr_stripes);
5304 raid5_set_cache_size(struct mddev *mddev, int size)
5306 struct r5conf *conf = mddev->private;
5310 if (size <= 16 || size > 32768)
5312 hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
5313 while (size < conf->max_nr_stripes) {
5314 if (drop_one_stripe(conf, hash))
5315 conf->max_nr_stripes--;
5320 hash = NR_STRIPE_HASH_LOCKS - 1;
5322 err = md_allow_write(mddev);
5325 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
5326 while (size > conf->max_nr_stripes) {
5327 if (grow_one_stripe(conf, hash))
5328 conf->max_nr_stripes++;
5330 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
5334 EXPORT_SYMBOL(raid5_set_cache_size);
5337 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5339 struct r5conf *conf = mddev->private;
5343 if (len >= PAGE_SIZE)
5348 if (kstrtoul(page, 10, &new))
5350 err = raid5_set_cache_size(mddev, new);
5356 static struct md_sysfs_entry
5357 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5358 raid5_show_stripe_cache_size,
5359 raid5_store_stripe_cache_size);
5362 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5364 struct r5conf *conf = mddev->private;
5366 return sprintf(page, "%d\n", conf->bypass_threshold);
5372 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
5374 struct r5conf *conf = mddev->private;
5376 if (len >= PAGE_SIZE)
5381 if (kstrtoul(page, 10, &new))
5383 if (new > conf->max_nr_stripes)
5385 conf->bypass_threshold = new;
5389 static struct md_sysfs_entry
5390 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5392 raid5_show_preread_threshold,
5393 raid5_store_preread_threshold);
5396 raid5_show_skip_copy(struct mddev *mddev, char *page)
5398 struct r5conf *conf = mddev->private;
5400 return sprintf(page, "%d\n", conf->skip_copy);
5406 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
5408 struct r5conf *conf = mddev->private;
5410 if (len >= PAGE_SIZE)
5415 if (kstrtoul(page, 10, &new))
5418 if (new == conf->skip_copy)
5421 mddev_suspend(mddev);
5422 conf->skip_copy = new;
5424 mddev->queue->backing_dev_info.capabilities |=
5425 BDI_CAP_STABLE_WRITES;
5427 mddev->queue->backing_dev_info.capabilities &=
5428 ~BDI_CAP_STABLE_WRITES;
5429 mddev_resume(mddev);
5433 static struct md_sysfs_entry
5434 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
5435 raid5_show_skip_copy,
5436 raid5_store_skip_copy);
5439 stripe_cache_active_show(struct mddev *mddev, char *page)
5441 struct r5conf *conf = mddev->private;
5443 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5448 static struct md_sysfs_entry
5449 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
5452 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
5454 struct r5conf *conf = mddev->private;
5456 return sprintf(page, "%d\n", conf->worker_cnt_per_group);
5461 static int alloc_thread_groups(struct r5conf *conf, int cnt,
5463 int *worker_cnt_per_group,
5464 struct r5worker_group **worker_groups);
5466 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
5468 struct r5conf *conf = mddev->private;
5471 struct r5worker_group *new_groups, *old_groups;
5472 int group_cnt, worker_cnt_per_group;
5474 if (len >= PAGE_SIZE)
5479 if (kstrtoul(page, 10, &new))
5482 if (new == conf->worker_cnt_per_group)
5485 mddev_suspend(mddev);
5487 old_groups = conf->worker_groups;
5489 flush_workqueue(raid5_wq);
5491 err = alloc_thread_groups(conf, new,
5492 &group_cnt, &worker_cnt_per_group,
5495 spin_lock_irq(&conf->device_lock);
5496 conf->group_cnt = group_cnt;
5497 conf->worker_cnt_per_group = worker_cnt_per_group;
5498 conf->worker_groups = new_groups;
5499 spin_unlock_irq(&conf->device_lock);
5502 kfree(old_groups[0].workers);
5506 mddev_resume(mddev);
5513 static struct md_sysfs_entry
5514 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
5515 raid5_show_group_thread_cnt,
5516 raid5_store_group_thread_cnt);
5518 static struct attribute *raid5_attrs[] = {
5519 &raid5_stripecache_size.attr,
5520 &raid5_stripecache_active.attr,
5521 &raid5_preread_bypass_threshold.attr,
5522 &raid5_group_thread_cnt.attr,
5523 &raid5_skip_copy.attr,
5526 static struct attribute_group raid5_attrs_group = {
5528 .attrs = raid5_attrs,
5531 static int alloc_thread_groups(struct r5conf *conf, int cnt,
5533 int *worker_cnt_per_group,
5534 struct r5worker_group **worker_groups)
5538 struct r5worker *workers;
5540 *worker_cnt_per_group = cnt;
5543 *worker_groups = NULL;
5546 *group_cnt = num_possible_nodes();
5547 size = sizeof(struct r5worker) * cnt;
5548 workers = kzalloc(size * *group_cnt, GFP_NOIO);
5549 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
5550 *group_cnt, GFP_NOIO);
5551 if (!*worker_groups || !workers) {
5553 kfree(*worker_groups);
5557 for (i = 0; i < *group_cnt; i++) {
5558 struct r5worker_group *group;
5560 group = &(*worker_groups)[i];
5561 INIT_LIST_HEAD(&group->handle_list);
5563 group->workers = workers + i * cnt;
5565 for (j = 0; j < cnt; j++) {
5566 struct r5worker *worker = group->workers + j;
5567 worker->group = group;
5568 INIT_WORK(&worker->work, raid5_do_work);
5570 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
5571 INIT_LIST_HEAD(worker->temp_inactive_list + k);
5578 static void free_thread_groups(struct r5conf *conf)
5580 if (conf->worker_groups)
5581 kfree(conf->worker_groups[0].workers);
5582 kfree(conf->worker_groups);
5583 conf->worker_groups = NULL;
5587 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
5589 struct r5conf *conf = mddev->private;
5592 sectors = mddev->dev_sectors;
5594 /* size is defined by the smallest of previous and new size */
5595 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
5597 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5598 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
5599 return sectors * (raid_disks - conf->max_degraded);
5602 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5604 safe_put_page(percpu->spare_page);
5605 kfree(percpu->scribble);
5606 percpu->spare_page = NULL;
5607 percpu->scribble = NULL;
5610 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5612 if (conf->level == 6 && !percpu->spare_page)
5613 percpu->spare_page = alloc_page(GFP_KERNEL);
5614 if (!percpu->scribble)
5615 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5617 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
5618 free_scratch_buffer(conf, percpu);
5625 static void raid5_free_percpu(struct r5conf *conf)
5632 #ifdef CONFIG_HOTPLUG_CPU
5633 unregister_cpu_notifier(&conf->cpu_notify);
5637 for_each_possible_cpu(cpu)
5638 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5641 free_percpu(conf->percpu);
5644 static void free_conf(struct r5conf *conf)
5646 free_thread_groups(conf);
5647 shrink_stripes(conf);
5648 raid5_free_percpu(conf);
5650 kfree(conf->stripe_hashtbl);
5654 #ifdef CONFIG_HOTPLUG_CPU
5655 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5658 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5659 long cpu = (long)hcpu;
5660 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5663 case CPU_UP_PREPARE:
5664 case CPU_UP_PREPARE_FROZEN:
5665 if (alloc_scratch_buffer(conf, percpu)) {
5666 pr_err("%s: failed memory allocation for cpu%ld\n",
5668 return notifier_from_errno(-ENOMEM);
5672 case CPU_DEAD_FROZEN:
5673 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5682 static int raid5_alloc_percpu(struct r5conf *conf)
5687 conf->percpu = alloc_percpu(struct raid5_percpu);
5691 #ifdef CONFIG_HOTPLUG_CPU
5692 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5693 conf->cpu_notify.priority = 0;
5694 err = register_cpu_notifier(&conf->cpu_notify);
5700 for_each_present_cpu(cpu) {
5701 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5703 pr_err("%s: failed memory allocation for cpu%ld\n",
5713 static struct r5conf *setup_conf(struct mddev *mddev)
5715 struct r5conf *conf;
5716 int raid_disk, memory, max_disks;
5717 struct md_rdev *rdev;
5718 struct disk_info *disk;
5721 int group_cnt, worker_cnt_per_group;
5722 struct r5worker_group *new_group;
5724 if (mddev->new_level != 5
5725 && mddev->new_level != 4
5726 && mddev->new_level != 6) {
5727 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5728 mdname(mddev), mddev->new_level);
5729 return ERR_PTR(-EIO);
5731 if ((mddev->new_level == 5
5732 && !algorithm_valid_raid5(mddev->new_layout)) ||
5733 (mddev->new_level == 6
5734 && !algorithm_valid_raid6(mddev->new_layout))) {
5735 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5736 mdname(mddev), mddev->new_layout);
5737 return ERR_PTR(-EIO);
5739 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5740 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5741 mdname(mddev), mddev->raid_disks);
5742 return ERR_PTR(-EINVAL);
5745 if (!mddev->new_chunk_sectors ||
5746 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5747 !is_power_of_2(mddev->new_chunk_sectors)) {
5748 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5749 mdname(mddev), mddev->new_chunk_sectors << 9);
5750 return ERR_PTR(-EINVAL);
5753 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5756 /* Don't enable multi-threading by default*/
5757 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
5759 conf->group_cnt = group_cnt;
5760 conf->worker_cnt_per_group = worker_cnt_per_group;
5761 conf->worker_groups = new_group;
5764 spin_lock_init(&conf->device_lock);
5765 seqcount_init(&conf->gen_lock);
5766 init_waitqueue_head(&conf->wait_for_stripe);
5767 init_waitqueue_head(&conf->wait_for_overlap);
5768 INIT_LIST_HEAD(&conf->handle_list);
5769 INIT_LIST_HEAD(&conf->hold_list);
5770 INIT_LIST_HEAD(&conf->delayed_list);
5771 INIT_LIST_HEAD(&conf->bitmap_list);
5772 init_llist_head(&conf->released_stripes);
5773 atomic_set(&conf->active_stripes, 0);
5774 atomic_set(&conf->preread_active_stripes, 0);
5775 atomic_set(&conf->active_aligned_reads, 0);
5776 conf->bypass_threshold = BYPASS_THRESHOLD;
5777 conf->recovery_disabled = mddev->recovery_disabled - 1;
5779 conf->raid_disks = mddev->raid_disks;
5780 if (mddev->reshape_position == MaxSector)
5781 conf->previous_raid_disks = mddev->raid_disks;
5783 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5784 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5785 conf->scribble_len = scribble_len(max_disks);
5787 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5792 conf->mddev = mddev;
5794 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5797 /* We init hash_locks[0] separately to that it can be used
5798 * as the reference lock in the spin_lock_nest_lock() call
5799 * in lock_all_device_hash_locks_irq in order to convince
5800 * lockdep that we know what we are doing.
5802 spin_lock_init(conf->hash_locks);
5803 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
5804 spin_lock_init(conf->hash_locks + i);
5806 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5807 INIT_LIST_HEAD(conf->inactive_list + i);
5809 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5810 INIT_LIST_HEAD(conf->temp_inactive_list + i);
5812 conf->level = mddev->new_level;
5813 if (raid5_alloc_percpu(conf) != 0)
5816 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5818 rdev_for_each(rdev, mddev) {
5819 raid_disk = rdev->raid_disk;
5820 if (raid_disk >= max_disks
5823 disk = conf->disks + raid_disk;
5825 if (test_bit(Replacement, &rdev->flags)) {
5826 if (disk->replacement)
5828 disk->replacement = rdev;
5835 if (test_bit(In_sync, &rdev->flags)) {
5836 char b[BDEVNAME_SIZE];
5837 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5839 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5840 } else if (rdev->saved_raid_disk != raid_disk)
5841 /* Cannot rely on bitmap to complete recovery */
5845 conf->chunk_sectors = mddev->new_chunk_sectors;
5846 conf->level = mddev->new_level;
5847 if (conf->level == 6)
5848 conf->max_degraded = 2;
5850 conf->max_degraded = 1;
5851 conf->algorithm = mddev->new_layout;
5852 conf->reshape_progress = mddev->reshape_position;
5853 if (conf->reshape_progress != MaxSector) {
5854 conf->prev_chunk_sectors = mddev->chunk_sectors;
5855 conf->prev_algo = mddev->layout;
5858 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5859 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5860 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
5861 if (grow_stripes(conf, NR_STRIPES)) {
5863 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5864 mdname(mddev), memory);
5867 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5868 mdname(mddev), memory);
5870 sprintf(pers_name, "raid%d", mddev->new_level);
5871 conf->thread = md_register_thread(raid5d, mddev, pers_name);
5872 if (!conf->thread) {
5874 "md/raid:%s: couldn't allocate thread.\n",
5884 return ERR_PTR(-EIO);
5886 return ERR_PTR(-ENOMEM);
5889 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5892 case ALGORITHM_PARITY_0:
5893 if (raid_disk < max_degraded)
5896 case ALGORITHM_PARITY_N:
5897 if (raid_disk >= raid_disks - max_degraded)
5900 case ALGORITHM_PARITY_0_6:
5901 if (raid_disk == 0 ||
5902 raid_disk == raid_disks - 1)
5905 case ALGORITHM_LEFT_ASYMMETRIC_6:
5906 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5907 case ALGORITHM_LEFT_SYMMETRIC_6:
5908 case ALGORITHM_RIGHT_SYMMETRIC_6:
5909 if (raid_disk == raid_disks - 1)
5915 static int run(struct mddev *mddev)
5917 struct r5conf *conf;
5918 int working_disks = 0;
5919 int dirty_parity_disks = 0;
5920 struct md_rdev *rdev;
5921 sector_t reshape_offset = 0;
5923 long long min_offset_diff = 0;
5926 if (mddev->recovery_cp != MaxSector)
5927 printk(KERN_NOTICE "md/raid:%s: not clean"
5928 " -- starting background reconstruction\n",
5931 rdev_for_each(rdev, mddev) {
5933 if (rdev->raid_disk < 0)
5935 diff = (rdev->new_data_offset - rdev->data_offset);
5937 min_offset_diff = diff;
5939 } else if (mddev->reshape_backwards &&
5940 diff < min_offset_diff)
5941 min_offset_diff = diff;
5942 else if (!mddev->reshape_backwards &&
5943 diff > min_offset_diff)
5944 min_offset_diff = diff;
5947 if (mddev->reshape_position != MaxSector) {
5948 /* Check that we can continue the reshape.
5949 * Difficulties arise if the stripe we would write to
5950 * next is at or after the stripe we would read from next.
5951 * For a reshape that changes the number of devices, this
5952 * is only possible for a very short time, and mdadm makes
5953 * sure that time appears to have past before assembling
5954 * the array. So we fail if that time hasn't passed.
5955 * For a reshape that keeps the number of devices the same
5956 * mdadm must be monitoring the reshape can keeping the
5957 * critical areas read-only and backed up. It will start
5958 * the array in read-only mode, so we check for that.
5960 sector_t here_new, here_old;
5962 int max_degraded = (mddev->level == 6 ? 2 : 1);
5964 if (mddev->new_level != mddev->level) {
5965 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5966 "required - aborting.\n",
5970 old_disks = mddev->raid_disks - mddev->delta_disks;
5971 /* reshape_position must be on a new-stripe boundary, and one
5972 * further up in new geometry must map after here in old
5975 here_new = mddev->reshape_position;
5976 if (sector_div(here_new, mddev->new_chunk_sectors *
5977 (mddev->raid_disks - max_degraded))) {
5978 printk(KERN_ERR "md/raid:%s: reshape_position not "
5979 "on a stripe boundary\n", mdname(mddev));
5982 reshape_offset = here_new * mddev->new_chunk_sectors;
5983 /* here_new is the stripe we will write to */
5984 here_old = mddev->reshape_position;
5985 sector_div(here_old, mddev->chunk_sectors *
5986 (old_disks-max_degraded));
5987 /* here_old is the first stripe that we might need to read
5989 if (mddev->delta_disks == 0) {
5990 if ((here_new * mddev->new_chunk_sectors !=
5991 here_old * mddev->chunk_sectors)) {
5992 printk(KERN_ERR "md/raid:%s: reshape position is"
5993 " confused - aborting\n", mdname(mddev));
5996 /* We cannot be sure it is safe to start an in-place
5997 * reshape. It is only safe if user-space is monitoring
5998 * and taking constant backups.
5999 * mdadm always starts a situation like this in
6000 * readonly mode so it can take control before
6001 * allowing any writes. So just check for that.
6003 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6004 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6005 /* not really in-place - so OK */;
6006 else if (mddev->ro == 0) {
6007 printk(KERN_ERR "md/raid:%s: in-place reshape "
6008 "must be started in read-only mode "
6013 } else if (mddev->reshape_backwards
6014 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6015 here_old * mddev->chunk_sectors)
6016 : (here_new * mddev->new_chunk_sectors >=
6017 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6018 /* Reading from the same stripe as writing to - bad */
6019 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6020 "auto-recovery - aborting.\n",
6024 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6026 /* OK, we should be able to continue; */
6028 BUG_ON(mddev->level != mddev->new_level);
6029 BUG_ON(mddev->layout != mddev->new_layout);
6030 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6031 BUG_ON(mddev->delta_disks != 0);
6034 if (mddev->private == NULL)
6035 conf = setup_conf(mddev);
6037 conf = mddev->private;
6040 return PTR_ERR(conf);
6042 conf->min_offset_diff = min_offset_diff;
6043 mddev->thread = conf->thread;
6044 conf->thread = NULL;
6045 mddev->private = conf;
6047 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6049 rdev = conf->disks[i].rdev;
6050 if (!rdev && conf->disks[i].replacement) {
6051 /* The replacement is all we have yet */
6052 rdev = conf->disks[i].replacement;
6053 conf->disks[i].replacement = NULL;
6054 clear_bit(Replacement, &rdev->flags);
6055 conf->disks[i].rdev = rdev;
6059 if (conf->disks[i].replacement &&
6060 conf->reshape_progress != MaxSector) {
6061 /* replacements and reshape simply do not mix. */
6062 printk(KERN_ERR "md: cannot handle concurrent "
6063 "replacement and reshape.\n");
6066 if (test_bit(In_sync, &rdev->flags)) {
6070 /* This disc is not fully in-sync. However if it
6071 * just stored parity (beyond the recovery_offset),
6072 * when we don't need to be concerned about the
6073 * array being dirty.
6074 * When reshape goes 'backwards', we never have
6075 * partially completed devices, so we only need
6076 * to worry about reshape going forwards.
6078 /* Hack because v0.91 doesn't store recovery_offset properly. */
6079 if (mddev->major_version == 0 &&
6080 mddev->minor_version > 90)
6081 rdev->recovery_offset = reshape_offset;
6083 if (rdev->recovery_offset < reshape_offset) {
6084 /* We need to check old and new layout */
6085 if (!only_parity(rdev->raid_disk,
6088 conf->max_degraded))
6091 if (!only_parity(rdev->raid_disk,
6093 conf->previous_raid_disks,
6094 conf->max_degraded))
6096 dirty_parity_disks++;
6100 * 0 for a fully functional array, 1 or 2 for a degraded array.
6102 mddev->degraded = calc_degraded(conf);
6104 if (has_failed(conf)) {
6105 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6106 " (%d/%d failed)\n",
6107 mdname(mddev), mddev->degraded, conf->raid_disks);
6111 /* device size must be a multiple of chunk size */
6112 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6113 mddev->resync_max_sectors = mddev->dev_sectors;
6115 if (mddev->degraded > dirty_parity_disks &&
6116 mddev->recovery_cp != MaxSector) {
6117 if (mddev->ok_start_degraded)
6119 "md/raid:%s: starting dirty degraded array"
6120 " - data corruption possible.\n",
6124 "md/raid:%s: cannot start dirty degraded array.\n",
6130 if (mddev->degraded == 0)
6131 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6132 " devices, algorithm %d\n", mdname(mddev), conf->level,
6133 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6136 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6137 " out of %d devices, algorithm %d\n",
6138 mdname(mddev), conf->level,
6139 mddev->raid_disks - mddev->degraded,
6140 mddev->raid_disks, mddev->new_layout);
6142 print_raid5_conf(conf);
6144 if (conf->reshape_progress != MaxSector) {
6145 conf->reshape_safe = conf->reshape_progress;
6146 atomic_set(&conf->reshape_stripes, 0);
6147 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6148 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6149 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6150 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6151 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6155 /* Ok, everything is just fine now */
6156 if (mddev->to_remove == &raid5_attrs_group)
6157 mddev->to_remove = NULL;
6158 else if (mddev->kobj.sd &&
6159 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6161 "raid5: failed to create sysfs attributes for %s\n",
6163 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6167 bool discard_supported = true;
6168 /* read-ahead size must cover two whole stripes, which
6169 * is 2 * (datadisks) * chunksize where 'n' is the
6170 * number of raid devices
6172 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6173 int stripe = data_disks *
6174 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6175 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6176 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6178 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
6180 mddev->queue->backing_dev_info.congested_data = mddev;
6181 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
6183 chunk_size = mddev->chunk_sectors << 9;
6184 blk_queue_io_min(mddev->queue, chunk_size);
6185 blk_queue_io_opt(mddev->queue, chunk_size *
6186 (conf->raid_disks - conf->max_degraded));
6187 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6189 * We can only discard a whole stripe. It doesn't make sense to
6190 * discard data disk but write parity disk
6192 stripe = stripe * PAGE_SIZE;
6193 /* Round up to power of 2, as discard handling
6194 * currently assumes that */
6195 while ((stripe-1) & stripe)
6196 stripe = (stripe | (stripe-1)) + 1;
6197 mddev->queue->limits.discard_alignment = stripe;
6198 mddev->queue->limits.discard_granularity = stripe;
6200 * unaligned part of discard request will be ignored, so can't
6201 * guarantee discard_zeroes_data
6203 mddev->queue->limits.discard_zeroes_data = 0;
6205 blk_queue_max_write_same_sectors(mddev->queue, 0);
6207 rdev_for_each(rdev, mddev) {
6208 disk_stack_limits(mddev->gendisk, rdev->bdev,
6209 rdev->data_offset << 9);
6210 disk_stack_limits(mddev->gendisk, rdev->bdev,
6211 rdev->new_data_offset << 9);
6213 * discard_zeroes_data is required, otherwise data
6214 * could be lost. Consider a scenario: discard a stripe
6215 * (the stripe could be inconsistent if
6216 * discard_zeroes_data is 0); write one disk of the
6217 * stripe (the stripe could be inconsistent again
6218 * depending on which disks are used to calculate
6219 * parity); the disk is broken; The stripe data of this
6222 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6223 !bdev_get_queue(rdev->bdev)->
6224 limits.discard_zeroes_data)
6225 discard_supported = false;
6226 /* Unfortunately, discard_zeroes_data is not currently
6227 * a guarantee - just a hint. So we only allow DISCARD
6228 * if the sysadmin has confirmed that only safe devices
6229 * are in use by setting a module parameter.
6231 if (!devices_handle_discard_safely) {
6232 if (discard_supported) {
6233 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6234 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6236 discard_supported = false;
6240 if (discard_supported &&
6241 mddev->queue->limits.max_discard_sectors >= stripe &&
6242 mddev->queue->limits.discard_granularity >= stripe)
6243 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6246 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6252 md_unregister_thread(&mddev->thread);
6253 print_raid5_conf(conf);
6255 mddev->private = NULL;
6256 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6260 static int stop(struct mddev *mddev)
6262 struct r5conf *conf = mddev->private;
6264 md_unregister_thread(&mddev->thread);
6266 mddev->queue->backing_dev_info.congested_fn = NULL;
6268 mddev->private = NULL;
6269 mddev->to_remove = &raid5_attrs_group;
6273 static void status(struct seq_file *seq, struct mddev *mddev)
6275 struct r5conf *conf = mddev->private;
6278 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6279 mddev->chunk_sectors / 2, mddev->layout);
6280 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6281 for (i = 0; i < conf->raid_disks; i++)
6282 seq_printf (seq, "%s",
6283 conf->disks[i].rdev &&
6284 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6285 seq_printf (seq, "]");
6288 static void print_raid5_conf (struct r5conf *conf)
6291 struct disk_info *tmp;
6293 printk(KERN_DEBUG "RAID conf printout:\n");
6295 printk("(conf==NULL)\n");
6298 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6300 conf->raid_disks - conf->mddev->degraded);
6302 for (i = 0; i < conf->raid_disks; i++) {
6303 char b[BDEVNAME_SIZE];
6304 tmp = conf->disks + i;
6306 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6307 i, !test_bit(Faulty, &tmp->rdev->flags),
6308 bdevname(tmp->rdev->bdev, b));
6312 static int raid5_spare_active(struct mddev *mddev)
6315 struct r5conf *conf = mddev->private;
6316 struct disk_info *tmp;
6318 unsigned long flags;
6320 for (i = 0; i < conf->raid_disks; i++) {
6321 tmp = conf->disks + i;
6322 if (tmp->replacement
6323 && tmp->replacement->recovery_offset == MaxSector
6324 && !test_bit(Faulty, &tmp->replacement->flags)
6325 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
6326 /* Replacement has just become active. */
6328 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
6331 /* Replaced device not technically faulty,
6332 * but we need to be sure it gets removed
6333 * and never re-added.
6335 set_bit(Faulty, &tmp->rdev->flags);
6336 sysfs_notify_dirent_safe(
6337 tmp->rdev->sysfs_state);
6339 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
6340 } else if (tmp->rdev
6341 && tmp->rdev->recovery_offset == MaxSector
6342 && !test_bit(Faulty, &tmp->rdev->flags)
6343 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
6345 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
6348 spin_lock_irqsave(&conf->device_lock, flags);
6349 mddev->degraded = calc_degraded(conf);
6350 spin_unlock_irqrestore(&conf->device_lock, flags);
6351 print_raid5_conf(conf);
6355 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
6357 struct r5conf *conf = mddev->private;
6359 int number = rdev->raid_disk;
6360 struct md_rdev **rdevp;
6361 struct disk_info *p = conf->disks + number;
6363 print_raid5_conf(conf);
6364 if (rdev == p->rdev)
6366 else if (rdev == p->replacement)
6367 rdevp = &p->replacement;
6371 if (number >= conf->raid_disks &&
6372 conf->reshape_progress == MaxSector)
6373 clear_bit(In_sync, &rdev->flags);
6375 if (test_bit(In_sync, &rdev->flags) ||
6376 atomic_read(&rdev->nr_pending)) {
6380 /* Only remove non-faulty devices if recovery
6383 if (!test_bit(Faulty, &rdev->flags) &&
6384 mddev->recovery_disabled != conf->recovery_disabled &&
6385 !has_failed(conf) &&
6386 (!p->replacement || p->replacement == rdev) &&
6387 number < conf->raid_disks) {
6393 if (atomic_read(&rdev->nr_pending)) {
6394 /* lost the race, try later */
6397 } else if (p->replacement) {
6398 /* We must have just cleared 'rdev' */
6399 p->rdev = p->replacement;
6400 clear_bit(Replacement, &p->replacement->flags);
6401 smp_mb(); /* Make sure other CPUs may see both as identical
6402 * but will never see neither - if they are careful
6404 p->replacement = NULL;
6405 clear_bit(WantReplacement, &rdev->flags);
6407 /* We might have just removed the Replacement as faulty-
6408 * clear the bit just in case
6410 clear_bit(WantReplacement, &rdev->flags);
6413 print_raid5_conf(conf);
6417 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
6419 struct r5conf *conf = mddev->private;
6422 struct disk_info *p;
6424 int last = conf->raid_disks - 1;
6426 if (mddev->recovery_disabled == conf->recovery_disabled)
6429 if (rdev->saved_raid_disk < 0 && has_failed(conf))
6430 /* no point adding a device */
6433 if (rdev->raid_disk >= 0)
6434 first = last = rdev->raid_disk;
6437 * find the disk ... but prefer rdev->saved_raid_disk
6440 if (rdev->saved_raid_disk >= 0 &&
6441 rdev->saved_raid_disk >= first &&
6442 conf->disks[rdev->saved_raid_disk].rdev == NULL)
6443 first = rdev->saved_raid_disk;
6445 for (disk = first; disk <= last; disk++) {
6446 p = conf->disks + disk;
6447 if (p->rdev == NULL) {
6448 clear_bit(In_sync, &rdev->flags);
6449 rdev->raid_disk = disk;
6451 if (rdev->saved_raid_disk != disk)
6453 rcu_assign_pointer(p->rdev, rdev);
6457 for (disk = first; disk <= last; disk++) {
6458 p = conf->disks + disk;
6459 if (test_bit(WantReplacement, &p->rdev->flags) &&
6460 p->replacement == NULL) {
6461 clear_bit(In_sync, &rdev->flags);
6462 set_bit(Replacement, &rdev->flags);
6463 rdev->raid_disk = disk;
6466 rcu_assign_pointer(p->replacement, rdev);
6471 print_raid5_conf(conf);
6475 static int raid5_resize(struct mddev *mddev, sector_t sectors)
6477 /* no resync is happening, and there is enough space
6478 * on all devices, so we can resize.
6479 * We need to make sure resync covers any new space.
6480 * If the array is shrinking we should possibly wait until
6481 * any io in the removed space completes, but it hardly seems
6485 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6486 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
6487 if (mddev->external_size &&
6488 mddev->array_sectors > newsize)
6490 if (mddev->bitmap) {
6491 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
6495 md_set_array_sectors(mddev, newsize);
6496 set_capacity(mddev->gendisk, mddev->array_sectors);
6497 revalidate_disk(mddev->gendisk);
6498 if (sectors > mddev->dev_sectors &&
6499 mddev->recovery_cp > mddev->dev_sectors) {
6500 mddev->recovery_cp = mddev->dev_sectors;
6501 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
6503 mddev->dev_sectors = sectors;
6504 mddev->resync_max_sectors = sectors;
6508 static int check_stripe_cache(struct mddev *mddev)
6510 /* Can only proceed if there are plenty of stripe_heads.
6511 * We need a minimum of one full stripe,, and for sensible progress
6512 * it is best to have about 4 times that.
6513 * If we require 4 times, then the default 256 4K stripe_heads will
6514 * allow for chunk sizes up to 256K, which is probably OK.
6515 * If the chunk size is greater, user-space should request more
6516 * stripe_heads first.
6518 struct r5conf *conf = mddev->private;
6519 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
6520 > conf->max_nr_stripes ||
6521 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
6522 > conf->max_nr_stripes) {
6523 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
6525 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
6532 static int check_reshape(struct mddev *mddev)
6534 struct r5conf *conf = mddev->private;
6536 if (mddev->delta_disks == 0 &&
6537 mddev->new_layout == mddev->layout &&
6538 mddev->new_chunk_sectors == mddev->chunk_sectors)
6539 return 0; /* nothing to do */
6540 if (has_failed(conf))
6542 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
6543 /* We might be able to shrink, but the devices must
6544 * be made bigger first.
6545 * For raid6, 4 is the minimum size.
6546 * Otherwise 2 is the minimum
6549 if (mddev->level == 6)
6551 if (mddev->raid_disks + mddev->delta_disks < min)
6555 if (!check_stripe_cache(mddev))
6558 return resize_stripes(conf, (conf->previous_raid_disks
6559 + mddev->delta_disks));
6562 static int raid5_start_reshape(struct mddev *mddev)
6564 struct r5conf *conf = mddev->private;
6565 struct md_rdev *rdev;
6567 unsigned long flags;
6569 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
6572 if (!check_stripe_cache(mddev))
6575 if (has_failed(conf))
6578 rdev_for_each(rdev, mddev) {
6579 if (!test_bit(In_sync, &rdev->flags)
6580 && !test_bit(Faulty, &rdev->flags))
6584 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
6585 /* Not enough devices even to make a degraded array
6590 /* Refuse to reduce size of the array. Any reductions in
6591 * array size must be through explicit setting of array_size
6594 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
6595 < mddev->array_sectors) {
6596 printk(KERN_ERR "md/raid:%s: array size must be reduced "
6597 "before number of disks\n", mdname(mddev));
6601 atomic_set(&conf->reshape_stripes, 0);
6602 spin_lock_irq(&conf->device_lock);
6603 write_seqcount_begin(&conf->gen_lock);
6604 conf->previous_raid_disks = conf->raid_disks;
6605 conf->raid_disks += mddev->delta_disks;
6606 conf->prev_chunk_sectors = conf->chunk_sectors;
6607 conf->chunk_sectors = mddev->new_chunk_sectors;
6608 conf->prev_algo = conf->algorithm;
6609 conf->algorithm = mddev->new_layout;
6611 /* Code that selects data_offset needs to see the generation update
6612 * if reshape_progress has been set - so a memory barrier needed.
6615 if (mddev->reshape_backwards)
6616 conf->reshape_progress = raid5_size(mddev, 0, 0);
6618 conf->reshape_progress = 0;
6619 conf->reshape_safe = conf->reshape_progress;
6620 write_seqcount_end(&conf->gen_lock);
6621 spin_unlock_irq(&conf->device_lock);
6623 /* Now make sure any requests that proceeded on the assumption
6624 * the reshape wasn't running - like Discard or Read - have
6627 mddev_suspend(mddev);
6628 mddev_resume(mddev);
6630 /* Add some new drives, as many as will fit.
6631 * We know there are enough to make the newly sized array work.
6632 * Don't add devices if we are reducing the number of
6633 * devices in the array. This is because it is not possible
6634 * to correctly record the "partially reconstructed" state of
6635 * such devices during the reshape and confusion could result.
6637 if (mddev->delta_disks >= 0) {
6638 rdev_for_each(rdev, mddev)
6639 if (rdev->raid_disk < 0 &&
6640 !test_bit(Faulty, &rdev->flags)) {
6641 if (raid5_add_disk(mddev, rdev) == 0) {
6643 >= conf->previous_raid_disks)
6644 set_bit(In_sync, &rdev->flags);
6646 rdev->recovery_offset = 0;
6648 if (sysfs_link_rdev(mddev, rdev))
6649 /* Failure here is OK */;
6651 } else if (rdev->raid_disk >= conf->previous_raid_disks
6652 && !test_bit(Faulty, &rdev->flags)) {
6653 /* This is a spare that was manually added */
6654 set_bit(In_sync, &rdev->flags);
6657 /* When a reshape changes the number of devices,
6658 * ->degraded is measured against the larger of the
6659 * pre and post number of devices.
6661 spin_lock_irqsave(&conf->device_lock, flags);
6662 mddev->degraded = calc_degraded(conf);
6663 spin_unlock_irqrestore(&conf->device_lock, flags);
6665 mddev->raid_disks = conf->raid_disks;
6666 mddev->reshape_position = conf->reshape_progress;
6667 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6669 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6670 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6671 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6672 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6673 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6675 if (!mddev->sync_thread) {
6676 mddev->recovery = 0;
6677 spin_lock_irq(&conf->device_lock);
6678 write_seqcount_begin(&conf->gen_lock);
6679 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6680 mddev->new_chunk_sectors =
6681 conf->chunk_sectors = conf->prev_chunk_sectors;
6682 mddev->new_layout = conf->algorithm = conf->prev_algo;
6683 rdev_for_each(rdev, mddev)
6684 rdev->new_data_offset = rdev->data_offset;
6686 conf->generation --;
6687 conf->reshape_progress = MaxSector;
6688 mddev->reshape_position = MaxSector;
6689 write_seqcount_end(&conf->gen_lock);
6690 spin_unlock_irq(&conf->device_lock);
6693 conf->reshape_checkpoint = jiffies;
6694 md_wakeup_thread(mddev->sync_thread);
6695 md_new_event(mddev);
6699 /* This is called from the reshape thread and should make any
6700 * changes needed in 'conf'
6702 static void end_reshape(struct r5conf *conf)
6705 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6706 struct md_rdev *rdev;
6708 spin_lock_irq(&conf->device_lock);
6709 conf->previous_raid_disks = conf->raid_disks;
6710 rdev_for_each(rdev, conf->mddev)
6711 rdev->data_offset = rdev->new_data_offset;
6713 conf->reshape_progress = MaxSector;
6714 spin_unlock_irq(&conf->device_lock);
6715 wake_up(&conf->wait_for_overlap);
6717 /* read-ahead size must cover two whole stripes, which is
6718 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6720 if (conf->mddev->queue) {
6721 int data_disks = conf->raid_disks - conf->max_degraded;
6722 int stripe = data_disks * ((conf->chunk_sectors << 9)
6724 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6725 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6730 /* This is called from the raid5d thread with mddev_lock held.
6731 * It makes config changes to the device.
6733 static void raid5_finish_reshape(struct mddev *mddev)
6735 struct r5conf *conf = mddev->private;
6737 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6739 if (mddev->delta_disks > 0) {
6740 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6741 set_capacity(mddev->gendisk, mddev->array_sectors);
6742 revalidate_disk(mddev->gendisk);
6745 spin_lock_irq(&conf->device_lock);
6746 mddev->degraded = calc_degraded(conf);
6747 spin_unlock_irq(&conf->device_lock);
6748 for (d = conf->raid_disks ;
6749 d < conf->raid_disks - mddev->delta_disks;
6751 struct md_rdev *rdev = conf->disks[d].rdev;
6753 clear_bit(In_sync, &rdev->flags);
6754 rdev = conf->disks[d].replacement;
6756 clear_bit(In_sync, &rdev->flags);
6759 mddev->layout = conf->algorithm;
6760 mddev->chunk_sectors = conf->chunk_sectors;
6761 mddev->reshape_position = MaxSector;
6762 mddev->delta_disks = 0;
6763 mddev->reshape_backwards = 0;
6767 static void raid5_quiesce(struct mddev *mddev, int state)
6769 struct r5conf *conf = mddev->private;
6772 case 2: /* resume for a suspend */
6773 wake_up(&conf->wait_for_overlap);
6776 case 1: /* stop all writes */
6777 lock_all_device_hash_locks_irq(conf);
6778 /* '2' tells resync/reshape to pause so that all
6779 * active stripes can drain
6782 wait_event_cmd(conf->wait_for_stripe,
6783 atomic_read(&conf->active_stripes) == 0 &&
6784 atomic_read(&conf->active_aligned_reads) == 0,
6785 unlock_all_device_hash_locks_irq(conf),
6786 lock_all_device_hash_locks_irq(conf));
6788 unlock_all_device_hash_locks_irq(conf);
6789 /* allow reshape to continue */
6790 wake_up(&conf->wait_for_overlap);
6793 case 0: /* re-enable writes */
6794 lock_all_device_hash_locks_irq(conf);
6796 wake_up(&conf->wait_for_stripe);
6797 wake_up(&conf->wait_for_overlap);
6798 unlock_all_device_hash_locks_irq(conf);
6803 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6805 struct r0conf *raid0_conf = mddev->private;
6808 /* for raid0 takeover only one zone is supported */
6809 if (raid0_conf->nr_strip_zones > 1) {
6810 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6812 return ERR_PTR(-EINVAL);
6815 sectors = raid0_conf->strip_zone[0].zone_end;
6816 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6817 mddev->dev_sectors = sectors;
6818 mddev->new_level = level;
6819 mddev->new_layout = ALGORITHM_PARITY_N;
6820 mddev->new_chunk_sectors = mddev->chunk_sectors;
6821 mddev->raid_disks += 1;
6822 mddev->delta_disks = 1;
6823 /* make sure it will be not marked as dirty */
6824 mddev->recovery_cp = MaxSector;
6826 return setup_conf(mddev);
6829 static void *raid5_takeover_raid1(struct mddev *mddev)
6833 if (mddev->raid_disks != 2 ||
6834 mddev->degraded > 1)
6835 return ERR_PTR(-EINVAL);
6837 /* Should check if there are write-behind devices? */
6839 chunksect = 64*2; /* 64K by default */
6841 /* The array must be an exact multiple of chunksize */
6842 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6845 if ((chunksect<<9) < STRIPE_SIZE)
6846 /* array size does not allow a suitable chunk size */
6847 return ERR_PTR(-EINVAL);
6849 mddev->new_level = 5;
6850 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6851 mddev->new_chunk_sectors = chunksect;
6853 return setup_conf(mddev);
6856 static void *raid5_takeover_raid6(struct mddev *mddev)
6860 switch (mddev->layout) {
6861 case ALGORITHM_LEFT_ASYMMETRIC_6:
6862 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6864 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6865 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6867 case ALGORITHM_LEFT_SYMMETRIC_6:
6868 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6870 case ALGORITHM_RIGHT_SYMMETRIC_6:
6871 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6873 case ALGORITHM_PARITY_0_6:
6874 new_layout = ALGORITHM_PARITY_0;
6876 case ALGORITHM_PARITY_N:
6877 new_layout = ALGORITHM_PARITY_N;
6880 return ERR_PTR(-EINVAL);
6882 mddev->new_level = 5;
6883 mddev->new_layout = new_layout;
6884 mddev->delta_disks = -1;
6885 mddev->raid_disks -= 1;
6886 return setup_conf(mddev);
6889 static int raid5_check_reshape(struct mddev *mddev)
6891 /* For a 2-drive array, the layout and chunk size can be changed
6892 * immediately as not restriping is needed.
6893 * For larger arrays we record the new value - after validation
6894 * to be used by a reshape pass.
6896 struct r5conf *conf = mddev->private;
6897 int new_chunk = mddev->new_chunk_sectors;
6899 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6901 if (new_chunk > 0) {
6902 if (!is_power_of_2(new_chunk))
6904 if (new_chunk < (PAGE_SIZE>>9))
6906 if (mddev->array_sectors & (new_chunk-1))
6907 /* not factor of array size */
6911 /* They look valid */
6913 if (mddev->raid_disks == 2) {
6914 /* can make the change immediately */
6915 if (mddev->new_layout >= 0) {
6916 conf->algorithm = mddev->new_layout;
6917 mddev->layout = mddev->new_layout;
6919 if (new_chunk > 0) {
6920 conf->chunk_sectors = new_chunk ;
6921 mddev->chunk_sectors = new_chunk;
6923 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6924 md_wakeup_thread(mddev->thread);
6926 return check_reshape(mddev);
6929 static int raid6_check_reshape(struct mddev *mddev)
6931 int new_chunk = mddev->new_chunk_sectors;
6933 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6935 if (new_chunk > 0) {
6936 if (!is_power_of_2(new_chunk))
6938 if (new_chunk < (PAGE_SIZE >> 9))
6940 if (mddev->array_sectors & (new_chunk-1))
6941 /* not factor of array size */
6945 /* They look valid */
6946 return check_reshape(mddev);
6949 static void *raid5_takeover(struct mddev *mddev)
6951 /* raid5 can take over:
6952 * raid0 - if there is only one strip zone - make it a raid4 layout
6953 * raid1 - if there are two drives. We need to know the chunk size
6954 * raid4 - trivial - just use a raid4 layout.
6955 * raid6 - Providing it is a *_6 layout
6957 if (mddev->level == 0)
6958 return raid45_takeover_raid0(mddev, 5);
6959 if (mddev->level == 1)
6960 return raid5_takeover_raid1(mddev);
6961 if (mddev->level == 4) {
6962 mddev->new_layout = ALGORITHM_PARITY_N;
6963 mddev->new_level = 5;
6964 return setup_conf(mddev);
6966 if (mddev->level == 6)
6967 return raid5_takeover_raid6(mddev);
6969 return ERR_PTR(-EINVAL);
6972 static void *raid4_takeover(struct mddev *mddev)
6974 /* raid4 can take over:
6975 * raid0 - if there is only one strip zone
6976 * raid5 - if layout is right
6978 if (mddev->level == 0)
6979 return raid45_takeover_raid0(mddev, 4);
6980 if (mddev->level == 5 &&
6981 mddev->layout == ALGORITHM_PARITY_N) {
6982 mddev->new_layout = 0;
6983 mddev->new_level = 4;
6984 return setup_conf(mddev);
6986 return ERR_PTR(-EINVAL);
6989 static struct md_personality raid5_personality;
6991 static void *raid6_takeover(struct mddev *mddev)
6993 /* Currently can only take over a raid5. We map the
6994 * personality to an equivalent raid6 personality
6995 * with the Q block at the end.
6999 if (mddev->pers != &raid5_personality)
7000 return ERR_PTR(-EINVAL);
7001 if (mddev->degraded > 1)
7002 return ERR_PTR(-EINVAL);
7003 if (mddev->raid_disks > 253)
7004 return ERR_PTR(-EINVAL);
7005 if (mddev->raid_disks < 3)
7006 return ERR_PTR(-EINVAL);
7008 switch (mddev->layout) {
7009 case ALGORITHM_LEFT_ASYMMETRIC:
7010 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7012 case ALGORITHM_RIGHT_ASYMMETRIC:
7013 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7015 case ALGORITHM_LEFT_SYMMETRIC:
7016 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7018 case ALGORITHM_RIGHT_SYMMETRIC:
7019 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7021 case ALGORITHM_PARITY_0:
7022 new_layout = ALGORITHM_PARITY_0_6;
7024 case ALGORITHM_PARITY_N:
7025 new_layout = ALGORITHM_PARITY_N;
7028 return ERR_PTR(-EINVAL);
7030 mddev->new_level = 6;
7031 mddev->new_layout = new_layout;
7032 mddev->delta_disks = 1;
7033 mddev->raid_disks += 1;
7034 return setup_conf(mddev);
7037 static struct md_personality raid6_personality =
7041 .owner = THIS_MODULE,
7042 .make_request = make_request,
7046 .error_handler = error,
7047 .hot_add_disk = raid5_add_disk,
7048 .hot_remove_disk= raid5_remove_disk,
7049 .spare_active = raid5_spare_active,
7050 .sync_request = sync_request,
7051 .resize = raid5_resize,
7053 .check_reshape = raid6_check_reshape,
7054 .start_reshape = raid5_start_reshape,
7055 .finish_reshape = raid5_finish_reshape,
7056 .quiesce = raid5_quiesce,
7057 .takeover = raid6_takeover,
7059 static struct md_personality raid5_personality =
7063 .owner = THIS_MODULE,
7064 .make_request = make_request,
7068 .error_handler = error,
7069 .hot_add_disk = raid5_add_disk,
7070 .hot_remove_disk= raid5_remove_disk,
7071 .spare_active = raid5_spare_active,
7072 .sync_request = sync_request,
7073 .resize = raid5_resize,
7075 .check_reshape = raid5_check_reshape,
7076 .start_reshape = raid5_start_reshape,
7077 .finish_reshape = raid5_finish_reshape,
7078 .quiesce = raid5_quiesce,
7079 .takeover = raid5_takeover,
7082 static struct md_personality raid4_personality =
7086 .owner = THIS_MODULE,
7087 .make_request = make_request,
7091 .error_handler = error,
7092 .hot_add_disk = raid5_add_disk,
7093 .hot_remove_disk= raid5_remove_disk,
7094 .spare_active = raid5_spare_active,
7095 .sync_request = sync_request,
7096 .resize = raid5_resize,
7098 .check_reshape = raid5_check_reshape,
7099 .start_reshape = raid5_start_reshape,
7100 .finish_reshape = raid5_finish_reshape,
7101 .quiesce = raid5_quiesce,
7102 .takeover = raid4_takeover,
7105 static int __init raid5_init(void)
7107 raid5_wq = alloc_workqueue("raid5wq",
7108 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7111 register_md_personality(&raid6_personality);
7112 register_md_personality(&raid5_personality);
7113 register_md_personality(&raid4_personality);
7117 static void raid5_exit(void)
7119 unregister_md_personality(&raid6_personality);
7120 unregister_md_personality(&raid5_personality);
7121 unregister_md_personality(&raid4_personality);
7122 destroy_workqueue(raid5_wq);
7125 module_init(raid5_init);
7126 module_exit(raid5_exit);
7127 MODULE_LICENSE("GPL");
7128 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7129 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7130 MODULE_ALIAS("md-raid5");
7131 MODULE_ALIAS("md-raid4");
7132 MODULE_ALIAS("md-level-5");
7133 MODULE_ALIAS("md-level-4");
7134 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7135 MODULE_ALIAS("md-raid6");
7136 MODULE_ALIAS("md-level-6");
7138 /* This used to be two separate modules, they were: */
7139 MODULE_ALIAS("raid5");
7140 MODULE_ALIAS("raid6");