2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <trace/events/block.h>
64 #define cpu_to_group(cpu) cpu_to_node(cpu)
65 #define ANY_GROUP NUMA_NO_NODE
67 static struct workqueue_struct *raid5_wq;
72 #define NR_STRIPES 256
73 #define STRIPE_SIZE PAGE_SIZE
74 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
75 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
76 #define IO_THRESHOLD 1
77 #define BYPASS_THRESHOLD 1
78 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
79 #define HASH_MASK (NR_HASH - 1)
80 #define MAX_STRIPE_BATCH 8
82 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
84 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
85 return &conf->stripe_hashtbl[hash];
88 static inline int stripe_hash_locks_hash(sector_t sect)
90 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
93 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
95 spin_lock_irq(conf->hash_locks + hash);
96 spin_lock(&conf->device_lock);
99 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
101 spin_unlock(&conf->device_lock);
102 spin_unlock_irq(conf->hash_locks + hash);
105 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
109 spin_lock(conf->hash_locks);
110 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
111 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
112 spin_lock(&conf->device_lock);
115 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
118 spin_unlock(&conf->device_lock);
119 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
120 spin_unlock(conf->hash_locks + i - 1);
124 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
125 * order without overlap. There may be several bio's per stripe+device, and
126 * a bio could span several devices.
127 * When walking this list for a particular stripe+device, we must never proceed
128 * beyond a bio that extends past this device, as the next bio might no longer
130 * This function is used to determine the 'next' bio in the list, given the sector
131 * of the current stripe+device
133 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
135 int sectors = bio_sectors(bio);
136 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
143 * We maintain a biased count of active stripes in the bottom 16 bits of
144 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
146 static inline int raid5_bi_processed_stripes(struct bio *bio)
148 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
149 return (atomic_read(segments) >> 16) & 0xffff;
152 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
154 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
155 return atomic_sub_return(1, segments) & 0xffff;
158 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
160 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
161 atomic_inc(segments);
164 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
167 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
171 old = atomic_read(segments);
172 new = (old & 0xffff) | (cnt << 16);
173 } while (atomic_cmpxchg(segments, old, new) != old);
176 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
178 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
179 atomic_set(segments, cnt);
182 /* Find first data disk in a raid6 stripe */
183 static inline int raid6_d0(struct stripe_head *sh)
186 /* ddf always start from first device */
188 /* md starts just after Q block */
189 if (sh->qd_idx == sh->disks - 1)
192 return sh->qd_idx + 1;
194 static inline int raid6_next_disk(int disk, int raid_disks)
197 return (disk < raid_disks) ? disk : 0;
200 /* When walking through the disks in a raid5, starting at raid6_d0,
201 * We need to map each disk to a 'slot', where the data disks are slot
202 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
203 * is raid_disks-1. This help does that mapping.
205 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
206 int *count, int syndrome_disks)
212 if (idx == sh->pd_idx)
213 return syndrome_disks;
214 if (idx == sh->qd_idx)
215 return syndrome_disks + 1;
221 static void return_io(struct bio *return_bi)
223 struct bio *bi = return_bi;
226 return_bi = bi->bi_next;
229 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
236 static void print_raid5_conf (struct r5conf *conf);
238 static int stripe_operations_active(struct stripe_head *sh)
240 return sh->check_state || sh->reconstruct_state ||
241 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
242 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
245 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
247 struct r5conf *conf = sh->raid_conf;
248 struct r5worker_group *group;
250 int i, cpu = sh->cpu;
252 if (!cpu_online(cpu)) {
253 cpu = cpumask_any(cpu_online_mask);
257 if (list_empty(&sh->lru)) {
258 struct r5worker_group *group;
259 group = conf->worker_groups + cpu_to_group(cpu);
260 list_add_tail(&sh->lru, &group->handle_list);
261 group->stripes_cnt++;
265 if (conf->worker_cnt_per_group == 0) {
266 md_wakeup_thread(conf->mddev->thread);
270 group = conf->worker_groups + cpu_to_group(sh->cpu);
272 group->workers[0].working = true;
273 /* at least one worker should run to avoid race */
274 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
276 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
277 /* wakeup more workers */
278 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
279 if (group->workers[i].working == false) {
280 group->workers[i].working = true;
281 queue_work_on(sh->cpu, raid5_wq,
282 &group->workers[i].work);
288 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
289 struct list_head *temp_inactive_list)
291 BUG_ON(!list_empty(&sh->lru));
292 BUG_ON(atomic_read(&conf->active_stripes)==0);
293 if (test_bit(STRIPE_HANDLE, &sh->state)) {
294 if (test_bit(STRIPE_DELAYED, &sh->state) &&
295 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
296 list_add_tail(&sh->lru, &conf->delayed_list);
297 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
298 sh->bm_seq - conf->seq_write > 0)
299 list_add_tail(&sh->lru, &conf->bitmap_list);
301 clear_bit(STRIPE_DELAYED, &sh->state);
302 clear_bit(STRIPE_BIT_DELAY, &sh->state);
303 if (conf->worker_cnt_per_group == 0) {
304 list_add_tail(&sh->lru, &conf->handle_list);
306 raid5_wakeup_stripe_thread(sh);
310 md_wakeup_thread(conf->mddev->thread);
312 BUG_ON(stripe_operations_active(sh));
313 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
314 if (atomic_dec_return(&conf->preread_active_stripes)
316 md_wakeup_thread(conf->mddev->thread);
317 atomic_dec(&conf->active_stripes);
318 if (!test_bit(STRIPE_EXPANDING, &sh->state))
319 list_add_tail(&sh->lru, temp_inactive_list);
323 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
324 struct list_head *temp_inactive_list)
326 if (atomic_dec_and_test(&sh->count))
327 do_release_stripe(conf, sh, temp_inactive_list);
331 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
333 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
334 * given time. Adding stripes only takes device lock, while deleting stripes
335 * only takes hash lock.
337 static void release_inactive_stripe_list(struct r5conf *conf,
338 struct list_head *temp_inactive_list,
342 bool do_wakeup = false;
345 if (hash == NR_STRIPE_HASH_LOCKS) {
346 size = NR_STRIPE_HASH_LOCKS;
347 hash = NR_STRIPE_HASH_LOCKS - 1;
351 struct list_head *list = &temp_inactive_list[size - 1];
354 * We don't hold any lock here yet, get_active_stripe() might
355 * remove stripes from the list
357 if (!list_empty_careful(list)) {
358 spin_lock_irqsave(conf->hash_locks + hash, flags);
359 if (list_empty(conf->inactive_list + hash) &&
361 atomic_dec(&conf->empty_inactive_list_nr);
362 list_splice_tail_init(list, conf->inactive_list + hash);
364 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
371 wake_up(&conf->wait_for_stripe);
372 if (conf->retry_read_aligned)
373 md_wakeup_thread(conf->mddev->thread);
377 static struct llist_node *llist_reverse_order(struct llist_node *head)
379 struct llist_node *new_head = NULL;
382 struct llist_node *tmp = head;
384 tmp->next = new_head;
391 /* should hold conf->device_lock already */
392 static int release_stripe_list(struct r5conf *conf,
393 struct list_head *temp_inactive_list)
395 struct stripe_head *sh;
397 struct llist_node *head;
399 head = llist_del_all(&conf->released_stripes);
400 head = llist_reverse_order(head);
404 sh = llist_entry(head, struct stripe_head, release_list);
405 head = llist_next(head);
406 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
408 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
410 * Don't worry the bit is set here, because if the bit is set
411 * again, the count is always > 1. This is true for
412 * STRIPE_ON_UNPLUG_LIST bit too.
414 hash = sh->hash_lock_index;
415 __release_stripe(conf, sh, &temp_inactive_list[hash]);
422 static void release_stripe(struct stripe_head *sh)
424 struct r5conf *conf = sh->raid_conf;
426 struct list_head list;
430 if (unlikely(!conf->mddev->thread) ||
431 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
433 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
435 md_wakeup_thread(conf->mddev->thread);
438 local_irq_save(flags);
439 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
440 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
441 INIT_LIST_HEAD(&list);
442 hash = sh->hash_lock_index;
443 do_release_stripe(conf, sh, &list);
444 spin_unlock(&conf->device_lock);
445 release_inactive_stripe_list(conf, &list, hash);
447 local_irq_restore(flags);
450 static inline void remove_hash(struct stripe_head *sh)
452 pr_debug("remove_hash(), stripe %llu\n",
453 (unsigned long long)sh->sector);
455 hlist_del_init(&sh->hash);
458 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
460 struct hlist_head *hp = stripe_hash(conf, sh->sector);
462 pr_debug("insert_hash(), stripe %llu\n",
463 (unsigned long long)sh->sector);
465 hlist_add_head(&sh->hash, hp);
469 /* find an idle stripe, make sure it is unhashed, and return it. */
470 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
472 struct stripe_head *sh = NULL;
473 struct list_head *first;
475 if (list_empty(conf->inactive_list + hash))
477 first = (conf->inactive_list + hash)->next;
478 sh = list_entry(first, struct stripe_head, lru);
479 list_del_init(first);
481 atomic_inc(&conf->active_stripes);
482 BUG_ON(hash != sh->hash_lock_index);
483 if (list_empty(conf->inactive_list + hash))
484 atomic_inc(&conf->empty_inactive_list_nr);
489 static void shrink_buffers(struct stripe_head *sh)
493 int num = sh->raid_conf->pool_size;
495 for (i = 0; i < num ; i++) {
499 sh->dev[i].page = NULL;
504 static int grow_buffers(struct stripe_head *sh)
507 int num = sh->raid_conf->pool_size;
509 for (i = 0; i < num; i++) {
512 if (!(page = alloc_page(GFP_KERNEL))) {
515 sh->dev[i].page = page;
520 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
521 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
522 struct stripe_head *sh);
524 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
526 struct r5conf *conf = sh->raid_conf;
529 BUG_ON(atomic_read(&sh->count) != 0);
530 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
531 BUG_ON(stripe_operations_active(sh));
533 pr_debug("init_stripe called, stripe %llu\n",
534 (unsigned long long)sh->sector);
538 seq = read_seqcount_begin(&conf->gen_lock);
539 sh->generation = conf->generation - previous;
540 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
542 stripe_set_idx(sector, conf, previous, sh);
546 for (i = sh->disks; i--; ) {
547 struct r5dev *dev = &sh->dev[i];
549 if (dev->toread || dev->read || dev->towrite || dev->written ||
550 test_bit(R5_LOCKED, &dev->flags)) {
551 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
552 (unsigned long long)sh->sector, i, dev->toread,
553 dev->read, dev->towrite, dev->written,
554 test_bit(R5_LOCKED, &dev->flags));
558 raid5_build_block(sh, i, previous);
560 if (read_seqcount_retry(&conf->gen_lock, seq))
562 insert_hash(conf, sh);
563 sh->cpu = smp_processor_id();
566 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
569 struct stripe_head *sh;
571 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
572 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
573 if (sh->sector == sector && sh->generation == generation)
575 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
580 * Need to check if array has failed when deciding whether to:
582 * - remove non-faulty devices
585 * This determination is simple when no reshape is happening.
586 * However if there is a reshape, we need to carefully check
587 * both the before and after sections.
588 * This is because some failed devices may only affect one
589 * of the two sections, and some non-in_sync devices may
590 * be insync in the section most affected by failed devices.
592 static int calc_degraded(struct r5conf *conf)
594 int degraded, degraded2;
599 for (i = 0; i < conf->previous_raid_disks; i++) {
600 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
601 if (rdev && test_bit(Faulty, &rdev->flags))
602 rdev = rcu_dereference(conf->disks[i].replacement);
603 if (!rdev || test_bit(Faulty, &rdev->flags))
605 else if (test_bit(In_sync, &rdev->flags))
608 /* not in-sync or faulty.
609 * If the reshape increases the number of devices,
610 * this is being recovered by the reshape, so
611 * this 'previous' section is not in_sync.
612 * If the number of devices is being reduced however,
613 * the device can only be part of the array if
614 * we are reverting a reshape, so this section will
617 if (conf->raid_disks >= conf->previous_raid_disks)
621 if (conf->raid_disks == conf->previous_raid_disks)
625 for (i = 0; i < conf->raid_disks; i++) {
626 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
627 if (rdev && test_bit(Faulty, &rdev->flags))
628 rdev = rcu_dereference(conf->disks[i].replacement);
629 if (!rdev || test_bit(Faulty, &rdev->flags))
631 else if (test_bit(In_sync, &rdev->flags))
634 /* not in-sync or faulty.
635 * If reshape increases the number of devices, this
636 * section has already been recovered, else it
637 * almost certainly hasn't.
639 if (conf->raid_disks <= conf->previous_raid_disks)
643 if (degraded2 > degraded)
648 static int has_failed(struct r5conf *conf)
652 if (conf->mddev->reshape_position == MaxSector)
653 return conf->mddev->degraded > conf->max_degraded;
655 degraded = calc_degraded(conf);
656 if (degraded > conf->max_degraded)
661 static struct stripe_head *
662 get_active_stripe(struct r5conf *conf, sector_t sector,
663 int previous, int noblock, int noquiesce)
665 struct stripe_head *sh;
666 int hash = stripe_hash_locks_hash(sector);
668 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
670 spin_lock_irq(conf->hash_locks + hash);
673 wait_event_lock_irq(conf->wait_for_stripe,
674 conf->quiesce == 0 || noquiesce,
675 *(conf->hash_locks + hash));
676 sh = __find_stripe(conf, sector, conf->generation - previous);
678 if (!conf->inactive_blocked)
679 sh = get_free_stripe(conf, hash);
680 if (noblock && sh == NULL)
683 conf->inactive_blocked = 1;
685 conf->wait_for_stripe,
686 !list_empty(conf->inactive_list + hash) &&
687 (atomic_read(&conf->active_stripes)
688 < (conf->max_nr_stripes * 3 / 4)
689 || !conf->inactive_blocked),
690 *(conf->hash_locks + hash));
691 conf->inactive_blocked = 0;
693 init_stripe(sh, sector, previous);
695 if (atomic_read(&sh->count)) {
696 BUG_ON(!list_empty(&sh->lru)
697 && !test_bit(STRIPE_EXPANDING, &sh->state)
698 && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state)
699 && !test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
701 spin_lock(&conf->device_lock);
702 if (!test_bit(STRIPE_HANDLE, &sh->state))
703 atomic_inc(&conf->active_stripes);
704 if (list_empty(&sh->lru) &&
705 !test_bit(STRIPE_ON_RELEASE_LIST, &sh->state) &&
706 !test_bit(STRIPE_EXPANDING, &sh->state))
708 list_del_init(&sh->lru);
710 sh->group->stripes_cnt--;
713 spin_unlock(&conf->device_lock);
716 } while (sh == NULL);
719 atomic_inc(&sh->count);
721 spin_unlock_irq(conf->hash_locks + hash);
725 /* Determine if 'data_offset' or 'new_data_offset' should be used
726 * in this stripe_head.
728 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
730 sector_t progress = conf->reshape_progress;
731 /* Need a memory barrier to make sure we see the value
732 * of conf->generation, or ->data_offset that was set before
733 * reshape_progress was updated.
736 if (progress == MaxSector)
738 if (sh->generation == conf->generation - 1)
740 /* We are in a reshape, and this is a new-generation stripe,
741 * so use new_data_offset.
747 raid5_end_read_request(struct bio *bi, int error);
749 raid5_end_write_request(struct bio *bi, int error);
751 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
753 struct r5conf *conf = sh->raid_conf;
754 int i, disks = sh->disks;
758 for (i = disks; i--; ) {
760 int replace_only = 0;
761 struct bio *bi, *rbi;
762 struct md_rdev *rdev, *rrdev = NULL;
763 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
764 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
768 if (test_bit(R5_Discard, &sh->dev[i].flags))
770 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
772 else if (test_and_clear_bit(R5_WantReplace,
773 &sh->dev[i].flags)) {
778 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
781 bi = &sh->dev[i].req;
782 rbi = &sh->dev[i].rreq; /* For writing to replacement */
785 rrdev = rcu_dereference(conf->disks[i].replacement);
786 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
787 rdev = rcu_dereference(conf->disks[i].rdev);
796 /* We raced and saw duplicates */
799 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
804 if (rdev && test_bit(Faulty, &rdev->flags))
807 atomic_inc(&rdev->nr_pending);
808 if (rrdev && test_bit(Faulty, &rrdev->flags))
811 atomic_inc(&rrdev->nr_pending);
814 /* We have already checked bad blocks for reads. Now
815 * need to check for writes. We never accept write errors
816 * on the replacement, so we don't to check rrdev.
818 while ((rw & WRITE) && rdev &&
819 test_bit(WriteErrorSeen, &rdev->flags)) {
822 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
823 &first_bad, &bad_sectors);
828 set_bit(BlockedBadBlocks, &rdev->flags);
829 if (!conf->mddev->external &&
830 conf->mddev->flags) {
831 /* It is very unlikely, but we might
832 * still need to write out the
833 * bad block log - better give it
835 md_check_recovery(conf->mddev);
838 * Because md_wait_for_blocked_rdev
839 * will dec nr_pending, we must
840 * increment it first.
842 atomic_inc(&rdev->nr_pending);
843 md_wait_for_blocked_rdev(rdev, conf->mddev);
845 /* Acknowledged bad block - skip the write */
846 rdev_dec_pending(rdev, conf->mddev);
852 if (s->syncing || s->expanding || s->expanded
854 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
856 set_bit(STRIPE_IO_STARTED, &sh->state);
859 bi->bi_bdev = rdev->bdev;
861 bi->bi_end_io = (rw & WRITE)
862 ? raid5_end_write_request
863 : raid5_end_read_request;
866 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
867 __func__, (unsigned long long)sh->sector,
869 atomic_inc(&sh->count);
870 if (use_new_offset(conf, sh))
871 bi->bi_sector = (sh->sector
872 + rdev->new_data_offset);
874 bi->bi_sector = (sh->sector
875 + rdev->data_offset);
876 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
877 bi->bi_rw |= REQ_NOMERGE;
880 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
881 bi->bi_io_vec[0].bv_offset = 0;
882 bi->bi_size = STRIPE_SIZE;
884 * If this is discard request, set bi_vcnt 0. We don't
885 * want to confuse SCSI because SCSI will replace payload
887 if (rw & REQ_DISCARD)
890 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
892 if (conf->mddev->gendisk)
893 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
894 bi, disk_devt(conf->mddev->gendisk),
896 generic_make_request(bi);
899 if (s->syncing || s->expanding || s->expanded
901 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
903 set_bit(STRIPE_IO_STARTED, &sh->state);
906 rbi->bi_bdev = rrdev->bdev;
908 BUG_ON(!(rw & WRITE));
909 rbi->bi_end_io = raid5_end_write_request;
910 rbi->bi_private = sh;
912 pr_debug("%s: for %llu schedule op %ld on "
913 "replacement disc %d\n",
914 __func__, (unsigned long long)sh->sector,
916 atomic_inc(&sh->count);
917 if (use_new_offset(conf, sh))
918 rbi->bi_sector = (sh->sector
919 + rrdev->new_data_offset);
921 rbi->bi_sector = (sh->sector
922 + rrdev->data_offset);
924 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
925 rbi->bi_io_vec[0].bv_offset = 0;
926 rbi->bi_size = STRIPE_SIZE;
928 * If this is discard request, set bi_vcnt 0. We don't
929 * want to confuse SCSI because SCSI will replace payload
931 if (rw & REQ_DISCARD)
933 if (conf->mddev->gendisk)
934 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
935 rbi, disk_devt(conf->mddev->gendisk),
937 generic_make_request(rbi);
939 if (!rdev && !rrdev) {
941 set_bit(STRIPE_DEGRADED, &sh->state);
942 pr_debug("skip op %ld on disc %d for sector %llu\n",
943 bi->bi_rw, i, (unsigned long long)sh->sector);
944 clear_bit(R5_LOCKED, &sh->dev[i].flags);
945 set_bit(STRIPE_HANDLE, &sh->state);
950 static struct dma_async_tx_descriptor *
951 async_copy_data(int frombio, struct bio *bio, struct page *page,
952 sector_t sector, struct dma_async_tx_descriptor *tx)
955 struct page *bio_page;
958 struct async_submit_ctl submit;
959 enum async_tx_flags flags = 0;
961 if (bio->bi_sector >= sector)
962 page_offset = (signed)(bio->bi_sector - sector) * 512;
964 page_offset = (signed)(sector - bio->bi_sector) * -512;
967 flags |= ASYNC_TX_FENCE;
968 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
970 bio_for_each_segment(bvl, bio, i) {
971 int len = bvl->bv_len;
975 if (page_offset < 0) {
976 b_offset = -page_offset;
977 page_offset += b_offset;
981 if (len > 0 && page_offset + len > STRIPE_SIZE)
982 clen = STRIPE_SIZE - page_offset;
987 b_offset += bvl->bv_offset;
988 bio_page = bvl->bv_page;
990 tx = async_memcpy(page, bio_page, page_offset,
991 b_offset, clen, &submit);
993 tx = async_memcpy(bio_page, page, b_offset,
994 page_offset, clen, &submit);
996 /* chain the operations */
997 submit.depend_tx = tx;
999 if (clen < len) /* hit end of page */
1007 static void ops_complete_biofill(void *stripe_head_ref)
1009 struct stripe_head *sh = stripe_head_ref;
1010 struct bio *return_bi = NULL;
1013 pr_debug("%s: stripe %llu\n", __func__,
1014 (unsigned long long)sh->sector);
1016 /* clear completed biofills */
1017 for (i = sh->disks; i--; ) {
1018 struct r5dev *dev = &sh->dev[i];
1020 /* acknowledge completion of a biofill operation */
1021 /* and check if we need to reply to a read request,
1022 * new R5_Wantfill requests are held off until
1023 * !STRIPE_BIOFILL_RUN
1025 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1026 struct bio *rbi, *rbi2;
1031 while (rbi && rbi->bi_sector <
1032 dev->sector + STRIPE_SECTORS) {
1033 rbi2 = r5_next_bio(rbi, dev->sector);
1034 if (!raid5_dec_bi_active_stripes(rbi)) {
1035 rbi->bi_next = return_bi;
1042 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1044 return_io(return_bi);
1046 set_bit(STRIPE_HANDLE, &sh->state);
1050 static void ops_run_biofill(struct stripe_head *sh)
1052 struct dma_async_tx_descriptor *tx = NULL;
1053 struct async_submit_ctl submit;
1056 pr_debug("%s: stripe %llu\n", __func__,
1057 (unsigned long long)sh->sector);
1059 for (i = sh->disks; i--; ) {
1060 struct r5dev *dev = &sh->dev[i];
1061 if (test_bit(R5_Wantfill, &dev->flags)) {
1063 spin_lock_irq(&sh->stripe_lock);
1064 dev->read = rbi = dev->toread;
1066 spin_unlock_irq(&sh->stripe_lock);
1067 while (rbi && rbi->bi_sector <
1068 dev->sector + STRIPE_SECTORS) {
1069 tx = async_copy_data(0, rbi, dev->page,
1071 rbi = r5_next_bio(rbi, dev->sector);
1076 atomic_inc(&sh->count);
1077 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1078 async_trigger_callback(&submit);
1081 static void mark_target_uptodate(struct stripe_head *sh, int target)
1088 tgt = &sh->dev[target];
1089 set_bit(R5_UPTODATE, &tgt->flags);
1090 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1091 clear_bit(R5_Wantcompute, &tgt->flags);
1094 static void ops_complete_compute(void *stripe_head_ref)
1096 struct stripe_head *sh = stripe_head_ref;
1098 pr_debug("%s: stripe %llu\n", __func__,
1099 (unsigned long long)sh->sector);
1101 /* mark the computed target(s) as uptodate */
1102 mark_target_uptodate(sh, sh->ops.target);
1103 mark_target_uptodate(sh, sh->ops.target2);
1105 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1106 if (sh->check_state == check_state_compute_run)
1107 sh->check_state = check_state_compute_result;
1108 set_bit(STRIPE_HANDLE, &sh->state);
1112 /* return a pointer to the address conversion region of the scribble buffer */
1113 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1114 struct raid5_percpu *percpu)
1116 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
1119 static struct dma_async_tx_descriptor *
1120 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1122 int disks = sh->disks;
1123 struct page **xor_srcs = percpu->scribble;
1124 int target = sh->ops.target;
1125 struct r5dev *tgt = &sh->dev[target];
1126 struct page *xor_dest = tgt->page;
1128 struct dma_async_tx_descriptor *tx;
1129 struct async_submit_ctl submit;
1132 pr_debug("%s: stripe %llu block: %d\n",
1133 __func__, (unsigned long long)sh->sector, target);
1134 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1136 for (i = disks; i--; )
1138 xor_srcs[count++] = sh->dev[i].page;
1140 atomic_inc(&sh->count);
1142 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1143 ops_complete_compute, sh, to_addr_conv(sh, percpu));
1144 if (unlikely(count == 1))
1145 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1147 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1152 /* set_syndrome_sources - populate source buffers for gen_syndrome
1153 * @srcs - (struct page *) array of size sh->disks
1154 * @sh - stripe_head to parse
1156 * Populates srcs in proper layout order for the stripe and returns the
1157 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1158 * destination buffer is recorded in srcs[count] and the Q destination
1159 * is recorded in srcs[count+1]].
1161 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
1163 int disks = sh->disks;
1164 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1165 int d0_idx = raid6_d0(sh);
1169 for (i = 0; i < disks; i++)
1175 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1177 srcs[slot] = sh->dev[i].page;
1178 i = raid6_next_disk(i, disks);
1179 } while (i != d0_idx);
1181 return syndrome_disks;
1184 static struct dma_async_tx_descriptor *
1185 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1187 int disks = sh->disks;
1188 struct page **blocks = percpu->scribble;
1190 int qd_idx = sh->qd_idx;
1191 struct dma_async_tx_descriptor *tx;
1192 struct async_submit_ctl submit;
1198 if (sh->ops.target < 0)
1199 target = sh->ops.target2;
1200 else if (sh->ops.target2 < 0)
1201 target = sh->ops.target;
1203 /* we should only have one valid target */
1206 pr_debug("%s: stripe %llu block: %d\n",
1207 __func__, (unsigned long long)sh->sector, target);
1209 tgt = &sh->dev[target];
1210 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1213 atomic_inc(&sh->count);
1215 if (target == qd_idx) {
1216 count = set_syndrome_sources(blocks, sh);
1217 blocks[count] = NULL; /* regenerating p is not necessary */
1218 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1219 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1220 ops_complete_compute, sh,
1221 to_addr_conv(sh, percpu));
1222 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1224 /* Compute any data- or p-drive using XOR */
1226 for (i = disks; i-- ; ) {
1227 if (i == target || i == qd_idx)
1229 blocks[count++] = sh->dev[i].page;
1232 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1233 NULL, ops_complete_compute, sh,
1234 to_addr_conv(sh, percpu));
1235 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1241 static struct dma_async_tx_descriptor *
1242 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1244 int i, count, disks = sh->disks;
1245 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1246 int d0_idx = raid6_d0(sh);
1247 int faila = -1, failb = -1;
1248 int target = sh->ops.target;
1249 int target2 = sh->ops.target2;
1250 struct r5dev *tgt = &sh->dev[target];
1251 struct r5dev *tgt2 = &sh->dev[target2];
1252 struct dma_async_tx_descriptor *tx;
1253 struct page **blocks = percpu->scribble;
1254 struct async_submit_ctl submit;
1256 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1257 __func__, (unsigned long long)sh->sector, target, target2);
1258 BUG_ON(target < 0 || target2 < 0);
1259 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1260 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1262 /* we need to open-code set_syndrome_sources to handle the
1263 * slot number conversion for 'faila' and 'failb'
1265 for (i = 0; i < disks ; i++)
1270 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1272 blocks[slot] = sh->dev[i].page;
1278 i = raid6_next_disk(i, disks);
1279 } while (i != d0_idx);
1281 BUG_ON(faila == failb);
1284 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1285 __func__, (unsigned long long)sh->sector, faila, failb);
1287 atomic_inc(&sh->count);
1289 if (failb == syndrome_disks+1) {
1290 /* Q disk is one of the missing disks */
1291 if (faila == syndrome_disks) {
1292 /* Missing P+Q, just recompute */
1293 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1294 ops_complete_compute, sh,
1295 to_addr_conv(sh, percpu));
1296 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1297 STRIPE_SIZE, &submit);
1301 int qd_idx = sh->qd_idx;
1303 /* Missing D+Q: recompute D from P, then recompute Q */
1304 if (target == qd_idx)
1305 data_target = target2;
1307 data_target = target;
1310 for (i = disks; i-- ; ) {
1311 if (i == data_target || i == qd_idx)
1313 blocks[count++] = sh->dev[i].page;
1315 dest = sh->dev[data_target].page;
1316 init_async_submit(&submit,
1317 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1319 to_addr_conv(sh, percpu));
1320 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1323 count = set_syndrome_sources(blocks, sh);
1324 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1325 ops_complete_compute, sh,
1326 to_addr_conv(sh, percpu));
1327 return async_gen_syndrome(blocks, 0, count+2,
1328 STRIPE_SIZE, &submit);
1331 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1332 ops_complete_compute, sh,
1333 to_addr_conv(sh, percpu));
1334 if (failb == syndrome_disks) {
1335 /* We're missing D+P. */
1336 return async_raid6_datap_recov(syndrome_disks+2,
1340 /* We're missing D+D. */
1341 return async_raid6_2data_recov(syndrome_disks+2,
1342 STRIPE_SIZE, faila, failb,
1349 static void ops_complete_prexor(void *stripe_head_ref)
1351 struct stripe_head *sh = stripe_head_ref;
1353 pr_debug("%s: stripe %llu\n", __func__,
1354 (unsigned long long)sh->sector);
1357 static struct dma_async_tx_descriptor *
1358 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1359 struct dma_async_tx_descriptor *tx)
1361 int disks = sh->disks;
1362 struct page **xor_srcs = percpu->scribble;
1363 int count = 0, pd_idx = sh->pd_idx, i;
1364 struct async_submit_ctl submit;
1366 /* existing parity data subtracted */
1367 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1369 pr_debug("%s: stripe %llu\n", __func__,
1370 (unsigned long long)sh->sector);
1372 for (i = disks; i--; ) {
1373 struct r5dev *dev = &sh->dev[i];
1374 /* Only process blocks that are known to be uptodate */
1375 if (test_bit(R5_Wantdrain, &dev->flags))
1376 xor_srcs[count++] = dev->page;
1379 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1380 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1381 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1386 static struct dma_async_tx_descriptor *
1387 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1389 int disks = sh->disks;
1392 pr_debug("%s: stripe %llu\n", __func__,
1393 (unsigned long long)sh->sector);
1395 for (i = disks; i--; ) {
1396 struct r5dev *dev = &sh->dev[i];
1399 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1402 spin_lock_irq(&sh->stripe_lock);
1403 chosen = dev->towrite;
1404 dev->towrite = NULL;
1405 BUG_ON(dev->written);
1406 wbi = dev->written = chosen;
1407 spin_unlock_irq(&sh->stripe_lock);
1409 while (wbi && wbi->bi_sector <
1410 dev->sector + STRIPE_SECTORS) {
1411 if (wbi->bi_rw & REQ_FUA)
1412 set_bit(R5_WantFUA, &dev->flags);
1413 if (wbi->bi_rw & REQ_SYNC)
1414 set_bit(R5_SyncIO, &dev->flags);
1415 if (wbi->bi_rw & REQ_DISCARD)
1416 set_bit(R5_Discard, &dev->flags);
1418 tx = async_copy_data(1, wbi, dev->page,
1420 wbi = r5_next_bio(wbi, dev->sector);
1428 static void ops_complete_reconstruct(void *stripe_head_ref)
1430 struct stripe_head *sh = stripe_head_ref;
1431 int disks = sh->disks;
1432 int pd_idx = sh->pd_idx;
1433 int qd_idx = sh->qd_idx;
1435 bool fua = false, sync = false, discard = false;
1437 pr_debug("%s: stripe %llu\n", __func__,
1438 (unsigned long long)sh->sector);
1440 for (i = disks; i--; ) {
1441 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1442 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1443 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1446 for (i = disks; i--; ) {
1447 struct r5dev *dev = &sh->dev[i];
1449 if (dev->written || i == pd_idx || i == qd_idx) {
1451 set_bit(R5_UPTODATE, &dev->flags);
1453 set_bit(R5_WantFUA, &dev->flags);
1455 set_bit(R5_SyncIO, &dev->flags);
1459 if (sh->reconstruct_state == reconstruct_state_drain_run)
1460 sh->reconstruct_state = reconstruct_state_drain_result;
1461 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1462 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1464 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1465 sh->reconstruct_state = reconstruct_state_result;
1468 set_bit(STRIPE_HANDLE, &sh->state);
1473 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1474 struct dma_async_tx_descriptor *tx)
1476 int disks = sh->disks;
1477 struct page **xor_srcs = percpu->scribble;
1478 struct async_submit_ctl submit;
1479 int count = 0, pd_idx = sh->pd_idx, i;
1480 struct page *xor_dest;
1482 unsigned long flags;
1484 pr_debug("%s: stripe %llu\n", __func__,
1485 (unsigned long long)sh->sector);
1487 for (i = 0; i < sh->disks; i++) {
1490 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1493 if (i >= sh->disks) {
1494 atomic_inc(&sh->count);
1495 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1496 ops_complete_reconstruct(sh);
1499 /* check if prexor is active which means only process blocks
1500 * that are part of a read-modify-write (written)
1502 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1504 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1505 for (i = disks; i--; ) {
1506 struct r5dev *dev = &sh->dev[i];
1508 xor_srcs[count++] = dev->page;
1511 xor_dest = sh->dev[pd_idx].page;
1512 for (i = disks; i--; ) {
1513 struct r5dev *dev = &sh->dev[i];
1515 xor_srcs[count++] = dev->page;
1519 /* 1/ if we prexor'd then the dest is reused as a source
1520 * 2/ if we did not prexor then we are redoing the parity
1521 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1522 * for the synchronous xor case
1524 flags = ASYNC_TX_ACK |
1525 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1527 atomic_inc(&sh->count);
1529 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1530 to_addr_conv(sh, percpu));
1531 if (unlikely(count == 1))
1532 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1534 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1538 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1539 struct dma_async_tx_descriptor *tx)
1541 struct async_submit_ctl submit;
1542 struct page **blocks = percpu->scribble;
1545 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1547 for (i = 0; i < sh->disks; i++) {
1548 if (sh->pd_idx == i || sh->qd_idx == i)
1550 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1553 if (i >= sh->disks) {
1554 atomic_inc(&sh->count);
1555 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1556 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1557 ops_complete_reconstruct(sh);
1561 count = set_syndrome_sources(blocks, sh);
1563 atomic_inc(&sh->count);
1565 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1566 sh, to_addr_conv(sh, percpu));
1567 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1570 static void ops_complete_check(void *stripe_head_ref)
1572 struct stripe_head *sh = stripe_head_ref;
1574 pr_debug("%s: stripe %llu\n", __func__,
1575 (unsigned long long)sh->sector);
1577 sh->check_state = check_state_check_result;
1578 set_bit(STRIPE_HANDLE, &sh->state);
1582 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1584 int disks = sh->disks;
1585 int pd_idx = sh->pd_idx;
1586 int qd_idx = sh->qd_idx;
1587 struct page *xor_dest;
1588 struct page **xor_srcs = percpu->scribble;
1589 struct dma_async_tx_descriptor *tx;
1590 struct async_submit_ctl submit;
1594 pr_debug("%s: stripe %llu\n", __func__,
1595 (unsigned long long)sh->sector);
1598 xor_dest = sh->dev[pd_idx].page;
1599 xor_srcs[count++] = xor_dest;
1600 for (i = disks; i--; ) {
1601 if (i == pd_idx || i == qd_idx)
1603 xor_srcs[count++] = sh->dev[i].page;
1606 init_async_submit(&submit, 0, NULL, NULL, NULL,
1607 to_addr_conv(sh, percpu));
1608 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1609 &sh->ops.zero_sum_result, &submit);
1611 atomic_inc(&sh->count);
1612 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1613 tx = async_trigger_callback(&submit);
1616 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1618 struct page **srcs = percpu->scribble;
1619 struct async_submit_ctl submit;
1622 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1623 (unsigned long long)sh->sector, checkp);
1625 count = set_syndrome_sources(srcs, sh);
1629 atomic_inc(&sh->count);
1630 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1631 sh, to_addr_conv(sh, percpu));
1632 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1633 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1636 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1638 int overlap_clear = 0, i, disks = sh->disks;
1639 struct dma_async_tx_descriptor *tx = NULL;
1640 struct r5conf *conf = sh->raid_conf;
1641 int level = conf->level;
1642 struct raid5_percpu *percpu;
1646 percpu = per_cpu_ptr(conf->percpu, cpu);
1647 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1648 ops_run_biofill(sh);
1652 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1654 tx = ops_run_compute5(sh, percpu);
1656 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1657 tx = ops_run_compute6_1(sh, percpu);
1659 tx = ops_run_compute6_2(sh, percpu);
1661 /* terminate the chain if reconstruct is not set to be run */
1662 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1666 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1667 tx = ops_run_prexor(sh, percpu, tx);
1669 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1670 tx = ops_run_biodrain(sh, tx);
1674 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1676 ops_run_reconstruct5(sh, percpu, tx);
1678 ops_run_reconstruct6(sh, percpu, tx);
1681 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1682 if (sh->check_state == check_state_run)
1683 ops_run_check_p(sh, percpu);
1684 else if (sh->check_state == check_state_run_q)
1685 ops_run_check_pq(sh, percpu, 0);
1686 else if (sh->check_state == check_state_run_pq)
1687 ops_run_check_pq(sh, percpu, 1);
1693 for (i = disks; i--; ) {
1694 struct r5dev *dev = &sh->dev[i];
1695 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1696 wake_up(&sh->raid_conf->wait_for_overlap);
1701 static int grow_one_stripe(struct r5conf *conf, int hash)
1703 struct stripe_head *sh;
1704 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1708 sh->raid_conf = conf;
1710 spin_lock_init(&sh->stripe_lock);
1712 if (grow_buffers(sh)) {
1714 kmem_cache_free(conf->slab_cache, sh);
1717 sh->hash_lock_index = hash;
1718 /* we just created an active stripe so... */
1719 atomic_set(&sh->count, 1);
1720 atomic_inc(&conf->active_stripes);
1721 INIT_LIST_HEAD(&sh->lru);
1726 static int grow_stripes(struct r5conf *conf, int num)
1728 struct kmem_cache *sc;
1729 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1732 if (conf->mddev->gendisk)
1733 sprintf(conf->cache_name[0],
1734 "raid%d-%s", conf->level, mdname(conf->mddev));
1736 sprintf(conf->cache_name[0],
1737 "raid%d-%p", conf->level, conf->mddev);
1738 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1740 conf->active_name = 0;
1741 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1742 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1746 conf->slab_cache = sc;
1747 conf->pool_size = devs;
1748 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
1750 if (!grow_one_stripe(conf, hash))
1752 conf->max_nr_stripes++;
1753 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
1759 * scribble_len - return the required size of the scribble region
1760 * @num - total number of disks in the array
1762 * The size must be enough to contain:
1763 * 1/ a struct page pointer for each device in the array +2
1764 * 2/ room to convert each entry in (1) to its corresponding dma
1765 * (dma_map_page()) or page (page_address()) address.
1767 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1768 * calculate over all devices (not just the data blocks), using zeros in place
1769 * of the P and Q blocks.
1771 static size_t scribble_len(int num)
1775 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1780 static int resize_stripes(struct r5conf *conf, int newsize)
1782 /* Make all the stripes able to hold 'newsize' devices.
1783 * New slots in each stripe get 'page' set to a new page.
1785 * This happens in stages:
1786 * 1/ create a new kmem_cache and allocate the required number of
1788 * 2/ gather all the old stripe_heads and transfer the pages across
1789 * to the new stripe_heads. This will have the side effect of
1790 * freezing the array as once all stripe_heads have been collected,
1791 * no IO will be possible. Old stripe heads are freed once their
1792 * pages have been transferred over, and the old kmem_cache is
1793 * freed when all stripes are done.
1794 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1795 * we simple return a failre status - no need to clean anything up.
1796 * 4/ allocate new pages for the new slots in the new stripe_heads.
1797 * If this fails, we don't bother trying the shrink the
1798 * stripe_heads down again, we just leave them as they are.
1799 * As each stripe_head is processed the new one is released into
1802 * Once step2 is started, we cannot afford to wait for a write,
1803 * so we use GFP_NOIO allocations.
1805 struct stripe_head *osh, *nsh;
1806 LIST_HEAD(newstripes);
1807 struct disk_info *ndisks;
1810 struct kmem_cache *sc;
1814 if (newsize <= conf->pool_size)
1815 return 0; /* never bother to shrink */
1817 err = md_allow_write(conf->mddev);
1822 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1823 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1828 for (i = conf->max_nr_stripes; i; i--) {
1829 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1833 nsh->raid_conf = conf;
1834 spin_lock_init(&nsh->stripe_lock);
1836 list_add(&nsh->lru, &newstripes);
1839 /* didn't get enough, give up */
1840 while (!list_empty(&newstripes)) {
1841 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1842 list_del(&nsh->lru);
1843 kmem_cache_free(sc, nsh);
1845 kmem_cache_destroy(sc);
1848 /* Step 2 - Must use GFP_NOIO now.
1849 * OK, we have enough stripes, start collecting inactive
1850 * stripes and copying them over
1854 list_for_each_entry(nsh, &newstripes, lru) {
1855 lock_device_hash_lock(conf, hash);
1856 wait_event_cmd(conf->wait_for_stripe,
1857 !list_empty(conf->inactive_list + hash),
1858 unlock_device_hash_lock(conf, hash),
1859 lock_device_hash_lock(conf, hash));
1860 osh = get_free_stripe(conf, hash);
1861 unlock_device_hash_lock(conf, hash);
1862 atomic_set(&nsh->count, 1);
1863 for(i=0; i<conf->pool_size; i++)
1864 nsh->dev[i].page = osh->dev[i].page;
1865 for( ; i<newsize; i++)
1866 nsh->dev[i].page = NULL;
1867 nsh->hash_lock_index = hash;
1868 kmem_cache_free(conf->slab_cache, osh);
1870 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
1871 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
1876 kmem_cache_destroy(conf->slab_cache);
1879 * At this point, we are holding all the stripes so the array
1880 * is completely stalled, so now is a good time to resize
1881 * conf->disks and the scribble region
1883 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1885 for (i=0; i<conf->raid_disks; i++)
1886 ndisks[i] = conf->disks[i];
1888 conf->disks = ndisks;
1893 conf->scribble_len = scribble_len(newsize);
1894 for_each_present_cpu(cpu) {
1895 struct raid5_percpu *percpu;
1898 percpu = per_cpu_ptr(conf->percpu, cpu);
1899 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1902 kfree(percpu->scribble);
1903 percpu->scribble = scribble;
1911 /* Step 4, return new stripes to service */
1912 while(!list_empty(&newstripes)) {
1913 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1914 list_del_init(&nsh->lru);
1916 for (i=conf->raid_disks; i < newsize; i++)
1917 if (nsh->dev[i].page == NULL) {
1918 struct page *p = alloc_page(GFP_NOIO);
1919 nsh->dev[i].page = p;
1923 release_stripe(nsh);
1925 /* critical section pass, GFP_NOIO no longer needed */
1927 conf->slab_cache = sc;
1928 conf->active_name = 1-conf->active_name;
1929 conf->pool_size = newsize;
1933 static int drop_one_stripe(struct r5conf *conf, int hash)
1935 struct stripe_head *sh;
1937 spin_lock_irq(conf->hash_locks + hash);
1938 sh = get_free_stripe(conf, hash);
1939 spin_unlock_irq(conf->hash_locks + hash);
1942 BUG_ON(atomic_read(&sh->count));
1944 kmem_cache_free(conf->slab_cache, sh);
1945 atomic_dec(&conf->active_stripes);
1949 static void shrink_stripes(struct r5conf *conf)
1952 for (hash = 0; hash < NR_STRIPE_HASH_LOCKS; hash++)
1953 while (drop_one_stripe(conf, hash))
1956 if (conf->slab_cache)
1957 kmem_cache_destroy(conf->slab_cache);
1958 conf->slab_cache = NULL;
1961 static void raid5_end_read_request(struct bio * bi, int error)
1963 struct stripe_head *sh = bi->bi_private;
1964 struct r5conf *conf = sh->raid_conf;
1965 int disks = sh->disks, i;
1966 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1967 char b[BDEVNAME_SIZE];
1968 struct md_rdev *rdev = NULL;
1971 for (i=0 ; i<disks; i++)
1972 if (bi == &sh->dev[i].req)
1975 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1976 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1982 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1983 /* If replacement finished while this request was outstanding,
1984 * 'replacement' might be NULL already.
1985 * In that case it moved down to 'rdev'.
1986 * rdev is not removed until all requests are finished.
1988 rdev = conf->disks[i].replacement;
1990 rdev = conf->disks[i].rdev;
1992 if (use_new_offset(conf, sh))
1993 s = sh->sector + rdev->new_data_offset;
1995 s = sh->sector + rdev->data_offset;
1997 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1998 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1999 /* Note that this cannot happen on a
2000 * replacement device. We just fail those on
2005 "md/raid:%s: read error corrected"
2006 " (%lu sectors at %llu on %s)\n",
2007 mdname(conf->mddev), STRIPE_SECTORS,
2008 (unsigned long long)s,
2009 bdevname(rdev->bdev, b));
2010 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2011 clear_bit(R5_ReadError, &sh->dev[i].flags);
2012 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2013 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2014 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2016 if (atomic_read(&rdev->read_errors))
2017 atomic_set(&rdev->read_errors, 0);
2019 const char *bdn = bdevname(rdev->bdev, b);
2023 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2024 atomic_inc(&rdev->read_errors);
2025 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2028 "md/raid:%s: read error on replacement device "
2029 "(sector %llu on %s).\n",
2030 mdname(conf->mddev),
2031 (unsigned long long)s,
2033 else if (conf->mddev->degraded >= conf->max_degraded) {
2037 "md/raid:%s: read error not correctable "
2038 "(sector %llu on %s).\n",
2039 mdname(conf->mddev),
2040 (unsigned long long)s,
2042 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2047 "md/raid:%s: read error NOT corrected!! "
2048 "(sector %llu on %s).\n",
2049 mdname(conf->mddev),
2050 (unsigned long long)s,
2052 } else if (atomic_read(&rdev->read_errors)
2053 > conf->max_nr_stripes)
2055 "md/raid:%s: Too many read errors, failing device %s.\n",
2056 mdname(conf->mddev), bdn);
2059 if (set_bad && test_bit(In_sync, &rdev->flags)
2060 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2063 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2064 set_bit(R5_ReadError, &sh->dev[i].flags);
2065 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2067 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2069 clear_bit(R5_ReadError, &sh->dev[i].flags);
2070 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2072 && test_bit(In_sync, &rdev->flags)
2073 && rdev_set_badblocks(
2074 rdev, sh->sector, STRIPE_SECTORS, 0)))
2075 md_error(conf->mddev, rdev);
2078 rdev_dec_pending(rdev, conf->mddev);
2079 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2080 set_bit(STRIPE_HANDLE, &sh->state);
2084 static void raid5_end_write_request(struct bio *bi, int error)
2086 struct stripe_head *sh = bi->bi_private;
2087 struct r5conf *conf = sh->raid_conf;
2088 int disks = sh->disks, i;
2089 struct md_rdev *uninitialized_var(rdev);
2090 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2093 int replacement = 0;
2095 for (i = 0 ; i < disks; i++) {
2096 if (bi == &sh->dev[i].req) {
2097 rdev = conf->disks[i].rdev;
2100 if (bi == &sh->dev[i].rreq) {
2101 rdev = conf->disks[i].replacement;
2105 /* rdev was removed and 'replacement'
2106 * replaced it. rdev is not removed
2107 * until all requests are finished.
2109 rdev = conf->disks[i].rdev;
2113 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2114 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2123 md_error(conf->mddev, rdev);
2124 else if (is_badblock(rdev, sh->sector,
2126 &first_bad, &bad_sectors))
2127 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2130 set_bit(WriteErrorSeen, &rdev->flags);
2131 set_bit(R5_WriteError, &sh->dev[i].flags);
2132 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2133 set_bit(MD_RECOVERY_NEEDED,
2134 &rdev->mddev->recovery);
2135 } else if (is_badblock(rdev, sh->sector,
2137 &first_bad, &bad_sectors)) {
2138 set_bit(R5_MadeGood, &sh->dev[i].flags);
2139 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2140 /* That was a successful write so make
2141 * sure it looks like we already did
2144 set_bit(R5_ReWrite, &sh->dev[i].flags);
2147 rdev_dec_pending(rdev, conf->mddev);
2149 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2150 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2151 set_bit(STRIPE_HANDLE, &sh->state);
2155 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2157 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2159 struct r5dev *dev = &sh->dev[i];
2161 bio_init(&dev->req);
2162 dev->req.bi_io_vec = &dev->vec;
2164 dev->req.bi_max_vecs++;
2165 dev->req.bi_private = sh;
2166 dev->vec.bv_page = dev->page;
2168 bio_init(&dev->rreq);
2169 dev->rreq.bi_io_vec = &dev->rvec;
2170 dev->rreq.bi_vcnt++;
2171 dev->rreq.bi_max_vecs++;
2172 dev->rreq.bi_private = sh;
2173 dev->rvec.bv_page = dev->page;
2176 dev->sector = compute_blocknr(sh, i, previous);
2179 static void error(struct mddev *mddev, struct md_rdev *rdev)
2181 char b[BDEVNAME_SIZE];
2182 struct r5conf *conf = mddev->private;
2183 unsigned long flags;
2184 pr_debug("raid456: error called\n");
2186 spin_lock_irqsave(&conf->device_lock, flags);
2187 clear_bit(In_sync, &rdev->flags);
2188 mddev->degraded = calc_degraded(conf);
2189 spin_unlock_irqrestore(&conf->device_lock, flags);
2190 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2192 set_bit(Blocked, &rdev->flags);
2193 set_bit(Faulty, &rdev->flags);
2194 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2196 "md/raid:%s: Disk failure on %s, disabling device.\n"
2197 "md/raid:%s: Operation continuing on %d devices.\n",
2199 bdevname(rdev->bdev, b),
2201 conf->raid_disks - mddev->degraded);
2205 * Input: a 'big' sector number,
2206 * Output: index of the data and parity disk, and the sector # in them.
2208 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2209 int previous, int *dd_idx,
2210 struct stripe_head *sh)
2212 sector_t stripe, stripe2;
2213 sector_t chunk_number;
2214 unsigned int chunk_offset;
2217 sector_t new_sector;
2218 int algorithm = previous ? conf->prev_algo
2220 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2221 : conf->chunk_sectors;
2222 int raid_disks = previous ? conf->previous_raid_disks
2224 int data_disks = raid_disks - conf->max_degraded;
2226 /* First compute the information on this sector */
2229 * Compute the chunk number and the sector offset inside the chunk
2231 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2232 chunk_number = r_sector;
2235 * Compute the stripe number
2237 stripe = chunk_number;
2238 *dd_idx = sector_div(stripe, data_disks);
2241 * Select the parity disk based on the user selected algorithm.
2243 pd_idx = qd_idx = -1;
2244 switch(conf->level) {
2246 pd_idx = data_disks;
2249 switch (algorithm) {
2250 case ALGORITHM_LEFT_ASYMMETRIC:
2251 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2252 if (*dd_idx >= pd_idx)
2255 case ALGORITHM_RIGHT_ASYMMETRIC:
2256 pd_idx = sector_div(stripe2, raid_disks);
2257 if (*dd_idx >= pd_idx)
2260 case ALGORITHM_LEFT_SYMMETRIC:
2261 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2262 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2264 case ALGORITHM_RIGHT_SYMMETRIC:
2265 pd_idx = sector_div(stripe2, raid_disks);
2266 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2268 case ALGORITHM_PARITY_0:
2272 case ALGORITHM_PARITY_N:
2273 pd_idx = data_disks;
2281 switch (algorithm) {
2282 case ALGORITHM_LEFT_ASYMMETRIC:
2283 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2284 qd_idx = pd_idx + 1;
2285 if (pd_idx == raid_disks-1) {
2286 (*dd_idx)++; /* Q D D D P */
2288 } else if (*dd_idx >= pd_idx)
2289 (*dd_idx) += 2; /* D D P Q D */
2291 case ALGORITHM_RIGHT_ASYMMETRIC:
2292 pd_idx = sector_div(stripe2, raid_disks);
2293 qd_idx = pd_idx + 1;
2294 if (pd_idx == raid_disks-1) {
2295 (*dd_idx)++; /* Q D D D P */
2297 } else if (*dd_idx >= pd_idx)
2298 (*dd_idx) += 2; /* D D P Q D */
2300 case ALGORITHM_LEFT_SYMMETRIC:
2301 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2302 qd_idx = (pd_idx + 1) % raid_disks;
2303 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2305 case ALGORITHM_RIGHT_SYMMETRIC:
2306 pd_idx = sector_div(stripe2, raid_disks);
2307 qd_idx = (pd_idx + 1) % raid_disks;
2308 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2311 case ALGORITHM_PARITY_0:
2316 case ALGORITHM_PARITY_N:
2317 pd_idx = data_disks;
2318 qd_idx = data_disks + 1;
2321 case ALGORITHM_ROTATING_ZERO_RESTART:
2322 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2323 * of blocks for computing Q is different.
2325 pd_idx = sector_div(stripe2, raid_disks);
2326 qd_idx = pd_idx + 1;
2327 if (pd_idx == raid_disks-1) {
2328 (*dd_idx)++; /* Q D D D P */
2330 } else if (*dd_idx >= pd_idx)
2331 (*dd_idx) += 2; /* D D P Q D */
2335 case ALGORITHM_ROTATING_N_RESTART:
2336 /* Same a left_asymmetric, by first stripe is
2337 * D D D P Q rather than
2341 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2342 qd_idx = pd_idx + 1;
2343 if (pd_idx == raid_disks-1) {
2344 (*dd_idx)++; /* Q D D D P */
2346 } else if (*dd_idx >= pd_idx)
2347 (*dd_idx) += 2; /* D D P Q D */
2351 case ALGORITHM_ROTATING_N_CONTINUE:
2352 /* Same as left_symmetric but Q is before P */
2353 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2354 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2355 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2359 case ALGORITHM_LEFT_ASYMMETRIC_6:
2360 /* RAID5 left_asymmetric, with Q on last device */
2361 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2362 if (*dd_idx >= pd_idx)
2364 qd_idx = raid_disks - 1;
2367 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2368 pd_idx = sector_div(stripe2, raid_disks-1);
2369 if (*dd_idx >= pd_idx)
2371 qd_idx = raid_disks - 1;
2374 case ALGORITHM_LEFT_SYMMETRIC_6:
2375 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2376 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2377 qd_idx = raid_disks - 1;
2380 case ALGORITHM_RIGHT_SYMMETRIC_6:
2381 pd_idx = sector_div(stripe2, raid_disks-1);
2382 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2383 qd_idx = raid_disks - 1;
2386 case ALGORITHM_PARITY_0_6:
2389 qd_idx = raid_disks - 1;
2399 sh->pd_idx = pd_idx;
2400 sh->qd_idx = qd_idx;
2401 sh->ddf_layout = ddf_layout;
2404 * Finally, compute the new sector number
2406 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2411 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2413 struct r5conf *conf = sh->raid_conf;
2414 int raid_disks = sh->disks;
2415 int data_disks = raid_disks - conf->max_degraded;
2416 sector_t new_sector = sh->sector, check;
2417 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2418 : conf->chunk_sectors;
2419 int algorithm = previous ? conf->prev_algo
2423 sector_t chunk_number;
2424 int dummy1, dd_idx = i;
2426 struct stripe_head sh2;
2429 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2430 stripe = new_sector;
2432 if (i == sh->pd_idx)
2434 switch(conf->level) {
2437 switch (algorithm) {
2438 case ALGORITHM_LEFT_ASYMMETRIC:
2439 case ALGORITHM_RIGHT_ASYMMETRIC:
2443 case ALGORITHM_LEFT_SYMMETRIC:
2444 case ALGORITHM_RIGHT_SYMMETRIC:
2447 i -= (sh->pd_idx + 1);
2449 case ALGORITHM_PARITY_0:
2452 case ALGORITHM_PARITY_N:
2459 if (i == sh->qd_idx)
2460 return 0; /* It is the Q disk */
2461 switch (algorithm) {
2462 case ALGORITHM_LEFT_ASYMMETRIC:
2463 case ALGORITHM_RIGHT_ASYMMETRIC:
2464 case ALGORITHM_ROTATING_ZERO_RESTART:
2465 case ALGORITHM_ROTATING_N_RESTART:
2466 if (sh->pd_idx == raid_disks-1)
2467 i--; /* Q D D D P */
2468 else if (i > sh->pd_idx)
2469 i -= 2; /* D D P Q D */
2471 case ALGORITHM_LEFT_SYMMETRIC:
2472 case ALGORITHM_RIGHT_SYMMETRIC:
2473 if (sh->pd_idx == raid_disks-1)
2474 i--; /* Q D D D P */
2479 i -= (sh->pd_idx + 2);
2482 case ALGORITHM_PARITY_0:
2485 case ALGORITHM_PARITY_N:
2487 case ALGORITHM_ROTATING_N_CONTINUE:
2488 /* Like left_symmetric, but P is before Q */
2489 if (sh->pd_idx == 0)
2490 i--; /* P D D D Q */
2495 i -= (sh->pd_idx + 1);
2498 case ALGORITHM_LEFT_ASYMMETRIC_6:
2499 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2503 case ALGORITHM_LEFT_SYMMETRIC_6:
2504 case ALGORITHM_RIGHT_SYMMETRIC_6:
2506 i += data_disks + 1;
2507 i -= (sh->pd_idx + 1);
2509 case ALGORITHM_PARITY_0_6:
2518 chunk_number = stripe * data_disks + i;
2519 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2521 check = raid5_compute_sector(conf, r_sector,
2522 previous, &dummy1, &sh2);
2523 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2524 || sh2.qd_idx != sh->qd_idx) {
2525 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2526 mdname(conf->mddev));
2534 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2535 int rcw, int expand)
2537 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2538 struct r5conf *conf = sh->raid_conf;
2539 int level = conf->level;
2543 for (i = disks; i--; ) {
2544 struct r5dev *dev = &sh->dev[i];
2547 set_bit(R5_LOCKED, &dev->flags);
2548 set_bit(R5_Wantdrain, &dev->flags);
2550 clear_bit(R5_UPTODATE, &dev->flags);
2554 /* if we are not expanding this is a proper write request, and
2555 * there will be bios with new data to be drained into the
2560 /* False alarm, nothing to do */
2562 sh->reconstruct_state = reconstruct_state_drain_run;
2563 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2565 sh->reconstruct_state = reconstruct_state_run;
2567 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2569 if (s->locked + conf->max_degraded == disks)
2570 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2571 atomic_inc(&conf->pending_full_writes);
2574 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2575 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2577 for (i = disks; i--; ) {
2578 struct r5dev *dev = &sh->dev[i];
2583 (test_bit(R5_UPTODATE, &dev->flags) ||
2584 test_bit(R5_Wantcompute, &dev->flags))) {
2585 set_bit(R5_Wantdrain, &dev->flags);
2586 set_bit(R5_LOCKED, &dev->flags);
2587 clear_bit(R5_UPTODATE, &dev->flags);
2592 /* False alarm - nothing to do */
2594 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2595 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2596 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2597 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2600 /* keep the parity disk(s) locked while asynchronous operations
2603 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2604 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2608 int qd_idx = sh->qd_idx;
2609 struct r5dev *dev = &sh->dev[qd_idx];
2611 set_bit(R5_LOCKED, &dev->flags);
2612 clear_bit(R5_UPTODATE, &dev->flags);
2616 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2617 __func__, (unsigned long long)sh->sector,
2618 s->locked, s->ops_request);
2622 * Each stripe/dev can have one or more bion attached.
2623 * toread/towrite point to the first in a chain.
2624 * The bi_next chain must be in order.
2626 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2629 struct r5conf *conf = sh->raid_conf;
2632 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2633 (unsigned long long)bi->bi_sector,
2634 (unsigned long long)sh->sector);
2637 * If several bio share a stripe. The bio bi_phys_segments acts as a
2638 * reference count to avoid race. The reference count should already be
2639 * increased before this function is called (for example, in
2640 * make_request()), so other bio sharing this stripe will not free the
2641 * stripe. If a stripe is owned by one stripe, the stripe lock will
2644 spin_lock_irq(&sh->stripe_lock);
2646 bip = &sh->dev[dd_idx].towrite;
2650 bip = &sh->dev[dd_idx].toread;
2651 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2652 if (bio_end_sector(*bip) > bi->bi_sector)
2654 bip = & (*bip)->bi_next;
2656 if (*bip && (*bip)->bi_sector < bio_end_sector(bi))
2659 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2663 raid5_inc_bi_active_stripes(bi);
2666 /* check if page is covered */
2667 sector_t sector = sh->dev[dd_idx].sector;
2668 for (bi=sh->dev[dd_idx].towrite;
2669 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2670 bi && bi->bi_sector <= sector;
2671 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2672 if (bio_end_sector(bi) >= sector)
2673 sector = bio_end_sector(bi);
2675 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2676 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2679 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2680 (unsigned long long)(*bip)->bi_sector,
2681 (unsigned long long)sh->sector, dd_idx);
2682 spin_unlock_irq(&sh->stripe_lock);
2684 if (conf->mddev->bitmap && firstwrite) {
2685 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2687 sh->bm_seq = conf->seq_flush+1;
2688 set_bit(STRIPE_BIT_DELAY, &sh->state);
2693 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2694 spin_unlock_irq(&sh->stripe_lock);
2698 static void end_reshape(struct r5conf *conf);
2700 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2701 struct stripe_head *sh)
2703 int sectors_per_chunk =
2704 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2706 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2707 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2709 raid5_compute_sector(conf,
2710 stripe * (disks - conf->max_degraded)
2711 *sectors_per_chunk + chunk_offset,
2717 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2718 struct stripe_head_state *s, int disks,
2719 struct bio **return_bi)
2722 for (i = disks; i--; ) {
2726 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2727 struct md_rdev *rdev;
2729 rdev = rcu_dereference(conf->disks[i].rdev);
2730 if (rdev && test_bit(In_sync, &rdev->flags))
2731 atomic_inc(&rdev->nr_pending);
2736 if (!rdev_set_badblocks(
2740 md_error(conf->mddev, rdev);
2741 rdev_dec_pending(rdev, conf->mddev);
2744 spin_lock_irq(&sh->stripe_lock);
2745 /* fail all writes first */
2746 bi = sh->dev[i].towrite;
2747 sh->dev[i].towrite = NULL;
2748 spin_unlock_irq(&sh->stripe_lock);
2752 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2753 wake_up(&conf->wait_for_overlap);
2755 while (bi && bi->bi_sector <
2756 sh->dev[i].sector + STRIPE_SECTORS) {
2757 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2758 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2759 if (!raid5_dec_bi_active_stripes(bi)) {
2760 md_write_end(conf->mddev);
2761 bi->bi_next = *return_bi;
2767 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2768 STRIPE_SECTORS, 0, 0);
2770 /* and fail all 'written' */
2771 bi = sh->dev[i].written;
2772 sh->dev[i].written = NULL;
2773 if (bi) bitmap_end = 1;
2774 while (bi && bi->bi_sector <
2775 sh->dev[i].sector + STRIPE_SECTORS) {
2776 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2777 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2778 if (!raid5_dec_bi_active_stripes(bi)) {
2779 md_write_end(conf->mddev);
2780 bi->bi_next = *return_bi;
2786 /* fail any reads if this device is non-operational and
2787 * the data has not reached the cache yet.
2789 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2790 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2791 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2792 spin_lock_irq(&sh->stripe_lock);
2793 bi = sh->dev[i].toread;
2794 sh->dev[i].toread = NULL;
2795 spin_unlock_irq(&sh->stripe_lock);
2796 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2797 wake_up(&conf->wait_for_overlap);
2798 while (bi && bi->bi_sector <
2799 sh->dev[i].sector + STRIPE_SECTORS) {
2800 struct bio *nextbi =
2801 r5_next_bio(bi, sh->dev[i].sector);
2802 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2803 if (!raid5_dec_bi_active_stripes(bi)) {
2804 bi->bi_next = *return_bi;
2811 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2812 STRIPE_SECTORS, 0, 0);
2813 /* If we were in the middle of a write the parity block might
2814 * still be locked - so just clear all R5_LOCKED flags
2816 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2819 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2820 if (atomic_dec_and_test(&conf->pending_full_writes))
2821 md_wakeup_thread(conf->mddev->thread);
2825 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2826 struct stripe_head_state *s)
2831 clear_bit(STRIPE_SYNCING, &sh->state);
2832 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2833 wake_up(&conf->wait_for_overlap);
2836 /* There is nothing more to do for sync/check/repair.
2837 * Don't even need to abort as that is handled elsewhere
2838 * if needed, and not always wanted e.g. if there is a known
2840 * For recover/replace we need to record a bad block on all
2841 * non-sync devices, or abort the recovery
2843 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2844 /* During recovery devices cannot be removed, so
2845 * locking and refcounting of rdevs is not needed
2847 for (i = 0; i < conf->raid_disks; i++) {
2848 struct md_rdev *rdev = conf->disks[i].rdev;
2850 && !test_bit(Faulty, &rdev->flags)
2851 && !test_bit(In_sync, &rdev->flags)
2852 && !rdev_set_badblocks(rdev, sh->sector,
2855 rdev = conf->disks[i].replacement;
2857 && !test_bit(Faulty, &rdev->flags)
2858 && !test_bit(In_sync, &rdev->flags)
2859 && !rdev_set_badblocks(rdev, sh->sector,
2864 conf->recovery_disabled =
2865 conf->mddev->recovery_disabled;
2867 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2870 static int want_replace(struct stripe_head *sh, int disk_idx)
2872 struct md_rdev *rdev;
2874 /* Doing recovery so rcu locking not required */
2875 rdev = sh->raid_conf->disks[disk_idx].replacement;
2877 && !test_bit(Faulty, &rdev->flags)
2878 && !test_bit(In_sync, &rdev->flags)
2879 && (rdev->recovery_offset <= sh->sector
2880 || rdev->mddev->recovery_cp <= sh->sector))
2886 /* fetch_block - checks the given member device to see if its data needs
2887 * to be read or computed to satisfy a request.
2889 * Returns 1 when no more member devices need to be checked, otherwise returns
2890 * 0 to tell the loop in handle_stripe_fill to continue
2892 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2893 int disk_idx, int disks)
2895 struct r5dev *dev = &sh->dev[disk_idx];
2896 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2897 &sh->dev[s->failed_num[1]] };
2899 /* is the data in this block needed, and can we get it? */
2900 if (!test_bit(R5_LOCKED, &dev->flags) &&
2901 !test_bit(R5_UPTODATE, &dev->flags) &&
2903 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2904 s->syncing || s->expanding ||
2905 (s->replacing && want_replace(sh, disk_idx)) ||
2906 (s->failed >= 1 && fdev[0]->toread) ||
2907 (s->failed >= 2 && fdev[1]->toread) ||
2908 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2909 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2910 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2911 /* we would like to get this block, possibly by computing it,
2912 * otherwise read it if the backing disk is insync
2914 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2915 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2916 if ((s->uptodate == disks - 1) &&
2917 (s->failed && (disk_idx == s->failed_num[0] ||
2918 disk_idx == s->failed_num[1]))) {
2919 /* have disk failed, and we're requested to fetch it;
2922 pr_debug("Computing stripe %llu block %d\n",
2923 (unsigned long long)sh->sector, disk_idx);
2924 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2925 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2926 set_bit(R5_Wantcompute, &dev->flags);
2927 sh->ops.target = disk_idx;
2928 sh->ops.target2 = -1; /* no 2nd target */
2930 /* Careful: from this point on 'uptodate' is in the eye
2931 * of raid_run_ops which services 'compute' operations
2932 * before writes. R5_Wantcompute flags a block that will
2933 * be R5_UPTODATE by the time it is needed for a
2934 * subsequent operation.
2938 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2939 /* Computing 2-failure is *very* expensive; only
2940 * do it if failed >= 2
2943 for (other = disks; other--; ) {
2944 if (other == disk_idx)
2946 if (!test_bit(R5_UPTODATE,
2947 &sh->dev[other].flags))
2951 pr_debug("Computing stripe %llu blocks %d,%d\n",
2952 (unsigned long long)sh->sector,
2954 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2955 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2956 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2957 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2958 sh->ops.target = disk_idx;
2959 sh->ops.target2 = other;
2963 } else if (test_bit(R5_Insync, &dev->flags)) {
2964 set_bit(R5_LOCKED, &dev->flags);
2965 set_bit(R5_Wantread, &dev->flags);
2967 pr_debug("Reading block %d (sync=%d)\n",
2968 disk_idx, s->syncing);
2976 * handle_stripe_fill - read or compute data to satisfy pending requests.
2978 static void handle_stripe_fill(struct stripe_head *sh,
2979 struct stripe_head_state *s,
2984 /* look for blocks to read/compute, skip this if a compute
2985 * is already in flight, or if the stripe contents are in the
2986 * midst of changing due to a write
2988 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2989 !sh->reconstruct_state)
2990 for (i = disks; i--; )
2991 if (fetch_block(sh, s, i, disks))
2993 set_bit(STRIPE_HANDLE, &sh->state);
2997 /* handle_stripe_clean_event
2998 * any written block on an uptodate or failed drive can be returned.
2999 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3000 * never LOCKED, so we don't need to test 'failed' directly.
3002 static void handle_stripe_clean_event(struct r5conf *conf,
3003 struct stripe_head *sh, int disks, struct bio **return_bi)
3007 int discard_pending = 0;
3009 for (i = disks; i--; )
3010 if (sh->dev[i].written) {
3012 if (!test_bit(R5_LOCKED, &dev->flags) &&
3013 (test_bit(R5_UPTODATE, &dev->flags) ||
3014 test_bit(R5_Discard, &dev->flags))) {
3015 /* We can return any write requests */
3016 struct bio *wbi, *wbi2;
3017 pr_debug("Return write for disc %d\n", i);
3018 if (test_and_clear_bit(R5_Discard, &dev->flags))
3019 clear_bit(R5_UPTODATE, &dev->flags);
3021 dev->written = NULL;
3022 while (wbi && wbi->bi_sector <
3023 dev->sector + STRIPE_SECTORS) {
3024 wbi2 = r5_next_bio(wbi, dev->sector);
3025 if (!raid5_dec_bi_active_stripes(wbi)) {
3026 md_write_end(conf->mddev);
3027 wbi->bi_next = *return_bi;
3032 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3034 !test_bit(STRIPE_DEGRADED, &sh->state),
3036 } else if (test_bit(R5_Discard, &dev->flags))
3037 discard_pending = 1;
3039 if (!discard_pending &&
3040 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3041 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3042 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3043 if (sh->qd_idx >= 0) {
3044 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3045 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3047 /* now that discard is done we can proceed with any sync */
3048 clear_bit(STRIPE_DISCARD, &sh->state);
3050 * SCSI discard will change some bio fields and the stripe has
3051 * no updated data, so remove it from hash list and the stripe
3052 * will be reinitialized
3054 spin_lock_irq(&conf->device_lock);
3056 spin_unlock_irq(&conf->device_lock);
3057 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3058 set_bit(STRIPE_HANDLE, &sh->state);
3062 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3063 if (atomic_dec_and_test(&conf->pending_full_writes))
3064 md_wakeup_thread(conf->mddev->thread);
3067 static void handle_stripe_dirtying(struct r5conf *conf,
3068 struct stripe_head *sh,
3069 struct stripe_head_state *s,
3072 int rmw = 0, rcw = 0, i;
3073 sector_t recovery_cp = conf->mddev->recovery_cp;
3075 /* RAID6 requires 'rcw' in current implementation.
3076 * Otherwise, check whether resync is now happening or should start.
3077 * If yes, then the array is dirty (after unclean shutdown or
3078 * initial creation), so parity in some stripes might be inconsistent.
3079 * In this case, we need to always do reconstruct-write, to ensure
3080 * that in case of drive failure or read-error correction, we
3081 * generate correct data from the parity.
3083 if (conf->max_degraded == 2 ||
3084 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
3085 /* Calculate the real rcw later - for now make it
3086 * look like rcw is cheaper
3089 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
3090 conf->max_degraded, (unsigned long long)recovery_cp,
3091 (unsigned long long)sh->sector);
3092 } else for (i = disks; i--; ) {
3093 /* would I have to read this buffer for read_modify_write */
3094 struct r5dev *dev = &sh->dev[i];
3095 if ((dev->towrite || i == sh->pd_idx) &&
3096 !test_bit(R5_LOCKED, &dev->flags) &&
3097 !(test_bit(R5_UPTODATE, &dev->flags) ||
3098 test_bit(R5_Wantcompute, &dev->flags))) {
3099 if (test_bit(R5_Insync, &dev->flags))
3102 rmw += 2*disks; /* cannot read it */
3104 /* Would I have to read this buffer for reconstruct_write */
3105 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
3106 !test_bit(R5_LOCKED, &dev->flags) &&
3107 !(test_bit(R5_UPTODATE, &dev->flags) ||
3108 test_bit(R5_Wantcompute, &dev->flags))) {
3109 if (test_bit(R5_Insync, &dev->flags)) rcw++;
3114 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3115 (unsigned long long)sh->sector, rmw, rcw);
3116 set_bit(STRIPE_HANDLE, &sh->state);
3117 if (rmw < rcw && rmw > 0) {
3118 /* prefer read-modify-write, but need to get some data */
3119 if (conf->mddev->queue)
3120 blk_add_trace_msg(conf->mddev->queue,
3121 "raid5 rmw %llu %d",
3122 (unsigned long long)sh->sector, rmw);
3123 for (i = disks; i--; ) {
3124 struct r5dev *dev = &sh->dev[i];
3125 if ((dev->towrite || i == sh->pd_idx) &&
3126 !test_bit(R5_LOCKED, &dev->flags) &&
3127 !(test_bit(R5_UPTODATE, &dev->flags) ||
3128 test_bit(R5_Wantcompute, &dev->flags)) &&
3129 test_bit(R5_Insync, &dev->flags)) {
3131 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3132 pr_debug("Read_old block "
3133 "%d for r-m-w\n", i);
3134 set_bit(R5_LOCKED, &dev->flags);
3135 set_bit(R5_Wantread, &dev->flags);
3138 set_bit(STRIPE_DELAYED, &sh->state);
3139 set_bit(STRIPE_HANDLE, &sh->state);
3144 if (rcw <= rmw && rcw > 0) {
3145 /* want reconstruct write, but need to get some data */
3148 for (i = disks; i--; ) {
3149 struct r5dev *dev = &sh->dev[i];
3150 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3151 i != sh->pd_idx && i != sh->qd_idx &&
3152 !test_bit(R5_LOCKED, &dev->flags) &&
3153 !(test_bit(R5_UPTODATE, &dev->flags) ||
3154 test_bit(R5_Wantcompute, &dev->flags))) {
3156 if (!test_bit(R5_Insync, &dev->flags))
3157 continue; /* it's a failed drive */
3159 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3160 pr_debug("Read_old block "
3161 "%d for Reconstruct\n", i);
3162 set_bit(R5_LOCKED, &dev->flags);
3163 set_bit(R5_Wantread, &dev->flags);
3167 set_bit(STRIPE_DELAYED, &sh->state);
3168 set_bit(STRIPE_HANDLE, &sh->state);
3172 if (rcw && conf->mddev->queue)
3173 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3174 (unsigned long long)sh->sector,
3175 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3177 /* now if nothing is locked, and if we have enough data,
3178 * we can start a write request
3180 /* since handle_stripe can be called at any time we need to handle the
3181 * case where a compute block operation has been submitted and then a
3182 * subsequent call wants to start a write request. raid_run_ops only
3183 * handles the case where compute block and reconstruct are requested
3184 * simultaneously. If this is not the case then new writes need to be
3185 * held off until the compute completes.
3187 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3188 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3189 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3190 schedule_reconstruction(sh, s, rcw == 0, 0);
3193 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3194 struct stripe_head_state *s, int disks)
3196 struct r5dev *dev = NULL;
3198 set_bit(STRIPE_HANDLE, &sh->state);
3200 switch (sh->check_state) {
3201 case check_state_idle:
3202 /* start a new check operation if there are no failures */
3203 if (s->failed == 0) {
3204 BUG_ON(s->uptodate != disks);
3205 sh->check_state = check_state_run;
3206 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3207 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3211 dev = &sh->dev[s->failed_num[0]];
3213 case check_state_compute_result:
3214 sh->check_state = check_state_idle;
3216 dev = &sh->dev[sh->pd_idx];
3218 /* check that a write has not made the stripe insync */
3219 if (test_bit(STRIPE_INSYNC, &sh->state))
3222 /* either failed parity check, or recovery is happening */
3223 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3224 BUG_ON(s->uptodate != disks);
3226 set_bit(R5_LOCKED, &dev->flags);
3228 set_bit(R5_Wantwrite, &dev->flags);
3230 clear_bit(STRIPE_DEGRADED, &sh->state);
3231 set_bit(STRIPE_INSYNC, &sh->state);
3233 case check_state_run:
3234 break; /* we will be called again upon completion */
3235 case check_state_check_result:
3236 sh->check_state = check_state_idle;
3238 /* if a failure occurred during the check operation, leave
3239 * STRIPE_INSYNC not set and let the stripe be handled again
3244 /* handle a successful check operation, if parity is correct
3245 * we are done. Otherwise update the mismatch count and repair
3246 * parity if !MD_RECOVERY_CHECK
3248 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3249 /* parity is correct (on disc,
3250 * not in buffer any more)
3252 set_bit(STRIPE_INSYNC, &sh->state);
3254 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3255 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3256 /* don't try to repair!! */
3257 set_bit(STRIPE_INSYNC, &sh->state);
3259 sh->check_state = check_state_compute_run;
3260 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3261 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3262 set_bit(R5_Wantcompute,
3263 &sh->dev[sh->pd_idx].flags);
3264 sh->ops.target = sh->pd_idx;
3265 sh->ops.target2 = -1;
3270 case check_state_compute_run:
3273 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3274 __func__, sh->check_state,
3275 (unsigned long long) sh->sector);
3281 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3282 struct stripe_head_state *s,
3285 int pd_idx = sh->pd_idx;
3286 int qd_idx = sh->qd_idx;
3289 set_bit(STRIPE_HANDLE, &sh->state);
3291 BUG_ON(s->failed > 2);
3293 /* Want to check and possibly repair P and Q.
3294 * However there could be one 'failed' device, in which
3295 * case we can only check one of them, possibly using the
3296 * other to generate missing data
3299 switch (sh->check_state) {
3300 case check_state_idle:
3301 /* start a new check operation if there are < 2 failures */
3302 if (s->failed == s->q_failed) {
3303 /* The only possible failed device holds Q, so it
3304 * makes sense to check P (If anything else were failed,
3305 * we would have used P to recreate it).
3307 sh->check_state = check_state_run;
3309 if (!s->q_failed && s->failed < 2) {
3310 /* Q is not failed, and we didn't use it to generate
3311 * anything, so it makes sense to check it
3313 if (sh->check_state == check_state_run)
3314 sh->check_state = check_state_run_pq;
3316 sh->check_state = check_state_run_q;
3319 /* discard potentially stale zero_sum_result */
3320 sh->ops.zero_sum_result = 0;
3322 if (sh->check_state == check_state_run) {
3323 /* async_xor_zero_sum destroys the contents of P */
3324 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3327 if (sh->check_state >= check_state_run &&
3328 sh->check_state <= check_state_run_pq) {
3329 /* async_syndrome_zero_sum preserves P and Q, so
3330 * no need to mark them !uptodate here
3332 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3336 /* we have 2-disk failure */
3337 BUG_ON(s->failed != 2);
3339 case check_state_compute_result:
3340 sh->check_state = check_state_idle;
3342 /* check that a write has not made the stripe insync */
3343 if (test_bit(STRIPE_INSYNC, &sh->state))
3346 /* now write out any block on a failed drive,
3347 * or P or Q if they were recomputed
3349 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3350 if (s->failed == 2) {
3351 dev = &sh->dev[s->failed_num[1]];
3353 set_bit(R5_LOCKED, &dev->flags);
3354 set_bit(R5_Wantwrite, &dev->flags);
3356 if (s->failed >= 1) {
3357 dev = &sh->dev[s->failed_num[0]];
3359 set_bit(R5_LOCKED, &dev->flags);
3360 set_bit(R5_Wantwrite, &dev->flags);
3362 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3363 dev = &sh->dev[pd_idx];
3365 set_bit(R5_LOCKED, &dev->flags);
3366 set_bit(R5_Wantwrite, &dev->flags);
3368 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3369 dev = &sh->dev[qd_idx];
3371 set_bit(R5_LOCKED, &dev->flags);
3372 set_bit(R5_Wantwrite, &dev->flags);
3374 clear_bit(STRIPE_DEGRADED, &sh->state);
3376 set_bit(STRIPE_INSYNC, &sh->state);
3378 case check_state_run:
3379 case check_state_run_q:
3380 case check_state_run_pq:
3381 break; /* we will be called again upon completion */
3382 case check_state_check_result:
3383 sh->check_state = check_state_idle;
3385 /* handle a successful check operation, if parity is correct
3386 * we are done. Otherwise update the mismatch count and repair
3387 * parity if !MD_RECOVERY_CHECK
3389 if (sh->ops.zero_sum_result == 0) {
3390 /* both parities are correct */
3392 set_bit(STRIPE_INSYNC, &sh->state);
3394 /* in contrast to the raid5 case we can validate
3395 * parity, but still have a failure to write
3398 sh->check_state = check_state_compute_result;
3399 /* Returning at this point means that we may go
3400 * off and bring p and/or q uptodate again so
3401 * we make sure to check zero_sum_result again
3402 * to verify if p or q need writeback
3406 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3407 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3408 /* don't try to repair!! */
3409 set_bit(STRIPE_INSYNC, &sh->state);
3411 int *target = &sh->ops.target;
3413 sh->ops.target = -1;
3414 sh->ops.target2 = -1;
3415 sh->check_state = check_state_compute_run;
3416 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3417 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3418 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3419 set_bit(R5_Wantcompute,
3420 &sh->dev[pd_idx].flags);
3422 target = &sh->ops.target2;
3425 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3426 set_bit(R5_Wantcompute,
3427 &sh->dev[qd_idx].flags);
3434 case check_state_compute_run:
3437 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3438 __func__, sh->check_state,
3439 (unsigned long long) sh->sector);
3444 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3448 /* We have read all the blocks in this stripe and now we need to
3449 * copy some of them into a target stripe for expand.
3451 struct dma_async_tx_descriptor *tx = NULL;
3452 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3453 for (i = 0; i < sh->disks; i++)
3454 if (i != sh->pd_idx && i != sh->qd_idx) {
3456 struct stripe_head *sh2;
3457 struct async_submit_ctl submit;
3459 sector_t bn = compute_blocknr(sh, i, 1);
3460 sector_t s = raid5_compute_sector(conf, bn, 0,
3462 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3464 /* so far only the early blocks of this stripe
3465 * have been requested. When later blocks
3466 * get requested, we will try again
3469 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3470 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3471 /* must have already done this block */
3472 release_stripe(sh2);
3476 /* place all the copies on one channel */
3477 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3478 tx = async_memcpy(sh2->dev[dd_idx].page,
3479 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3482 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3483 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3484 for (j = 0; j < conf->raid_disks; j++)
3485 if (j != sh2->pd_idx &&
3487 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3489 if (j == conf->raid_disks) {
3490 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3491 set_bit(STRIPE_HANDLE, &sh2->state);
3493 release_stripe(sh2);
3496 /* done submitting copies, wait for them to complete */
3497 async_tx_quiesce(&tx);
3501 * handle_stripe - do things to a stripe.
3503 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3504 * state of various bits to see what needs to be done.
3506 * return some read requests which now have data
3507 * return some write requests which are safely on storage
3508 * schedule a read on some buffers
3509 * schedule a write of some buffers
3510 * return confirmation of parity correctness
3514 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3516 struct r5conf *conf = sh->raid_conf;
3517 int disks = sh->disks;
3520 int do_recovery = 0;
3522 memset(s, 0, sizeof(*s));
3524 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3525 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3526 s->failed_num[0] = -1;
3527 s->failed_num[1] = -1;
3529 /* Now to look around and see what can be done */
3531 for (i=disks; i--; ) {
3532 struct md_rdev *rdev;
3539 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3541 dev->toread, dev->towrite, dev->written);
3542 /* maybe we can reply to a read
3544 * new wantfill requests are only permitted while
3545 * ops_complete_biofill is guaranteed to be inactive
3547 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3548 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3549 set_bit(R5_Wantfill, &dev->flags);
3551 /* now count some things */
3552 if (test_bit(R5_LOCKED, &dev->flags))
3554 if (test_bit(R5_UPTODATE, &dev->flags))
3556 if (test_bit(R5_Wantcompute, &dev->flags)) {
3558 BUG_ON(s->compute > 2);
3561 if (test_bit(R5_Wantfill, &dev->flags))
3563 else if (dev->toread)
3567 if (!test_bit(R5_OVERWRITE, &dev->flags))
3572 /* Prefer to use the replacement for reads, but only
3573 * if it is recovered enough and has no bad blocks.
3575 rdev = rcu_dereference(conf->disks[i].replacement);
3576 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3577 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3578 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3579 &first_bad, &bad_sectors))
3580 set_bit(R5_ReadRepl, &dev->flags);
3583 set_bit(R5_NeedReplace, &dev->flags);
3584 rdev = rcu_dereference(conf->disks[i].rdev);
3585 clear_bit(R5_ReadRepl, &dev->flags);
3587 if (rdev && test_bit(Faulty, &rdev->flags))
3590 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3591 &first_bad, &bad_sectors);
3592 if (s->blocked_rdev == NULL
3593 && (test_bit(Blocked, &rdev->flags)
3596 set_bit(BlockedBadBlocks,
3598 s->blocked_rdev = rdev;
3599 atomic_inc(&rdev->nr_pending);
3602 clear_bit(R5_Insync, &dev->flags);
3606 /* also not in-sync */
3607 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3608 test_bit(R5_UPTODATE, &dev->flags)) {
3609 /* treat as in-sync, but with a read error
3610 * which we can now try to correct
3612 set_bit(R5_Insync, &dev->flags);
3613 set_bit(R5_ReadError, &dev->flags);
3615 } else if (test_bit(In_sync, &rdev->flags))
3616 set_bit(R5_Insync, &dev->flags);
3617 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3618 /* in sync if before recovery_offset */
3619 set_bit(R5_Insync, &dev->flags);
3620 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3621 test_bit(R5_Expanded, &dev->flags))
3622 /* If we've reshaped into here, we assume it is Insync.
3623 * We will shortly update recovery_offset to make
3626 set_bit(R5_Insync, &dev->flags);
3628 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3629 /* This flag does not apply to '.replacement'
3630 * only to .rdev, so make sure to check that*/
3631 struct md_rdev *rdev2 = rcu_dereference(
3632 conf->disks[i].rdev);
3634 clear_bit(R5_Insync, &dev->flags);
3635 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3636 s->handle_bad_blocks = 1;
3637 atomic_inc(&rdev2->nr_pending);
3639 clear_bit(R5_WriteError, &dev->flags);
3641 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3642 /* This flag does not apply to '.replacement'
3643 * only to .rdev, so make sure to check that*/
3644 struct md_rdev *rdev2 = rcu_dereference(
3645 conf->disks[i].rdev);
3646 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3647 s->handle_bad_blocks = 1;
3648 atomic_inc(&rdev2->nr_pending);
3650 clear_bit(R5_MadeGood, &dev->flags);
3652 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3653 struct md_rdev *rdev2 = rcu_dereference(
3654 conf->disks[i].replacement);
3655 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3656 s->handle_bad_blocks = 1;
3657 atomic_inc(&rdev2->nr_pending);
3659 clear_bit(R5_MadeGoodRepl, &dev->flags);
3661 if (!test_bit(R5_Insync, &dev->flags)) {
3662 /* The ReadError flag will just be confusing now */
3663 clear_bit(R5_ReadError, &dev->flags);
3664 clear_bit(R5_ReWrite, &dev->flags);
3666 if (test_bit(R5_ReadError, &dev->flags))
3667 clear_bit(R5_Insync, &dev->flags);
3668 if (!test_bit(R5_Insync, &dev->flags)) {
3670 s->failed_num[s->failed] = i;
3672 if (rdev && !test_bit(Faulty, &rdev->flags))
3676 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3677 /* If there is a failed device being replaced,
3678 * we must be recovering.
3679 * else if we are after recovery_cp, we must be syncing
3680 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3681 * else we can only be replacing
3682 * sync and recovery both need to read all devices, and so
3683 * use the same flag.
3686 sh->sector >= conf->mddev->recovery_cp ||
3687 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3695 static void handle_stripe(struct stripe_head *sh)
3697 struct stripe_head_state s;
3698 struct r5conf *conf = sh->raid_conf;
3701 int disks = sh->disks;
3702 struct r5dev *pdev, *qdev;
3704 clear_bit(STRIPE_HANDLE, &sh->state);
3705 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3706 /* already being handled, ensure it gets handled
3707 * again when current action finishes */
3708 set_bit(STRIPE_HANDLE, &sh->state);
3712 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3713 spin_lock(&sh->stripe_lock);
3714 /* Cannot process 'sync' concurrently with 'discard' */
3715 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3716 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3717 set_bit(STRIPE_SYNCING, &sh->state);
3718 clear_bit(STRIPE_INSYNC, &sh->state);
3719 clear_bit(STRIPE_REPLACED, &sh->state);
3721 spin_unlock(&sh->stripe_lock);
3723 clear_bit(STRIPE_DELAYED, &sh->state);
3725 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3726 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3727 (unsigned long long)sh->sector, sh->state,
3728 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3729 sh->check_state, sh->reconstruct_state);
3731 analyse_stripe(sh, &s);
3733 if (s.handle_bad_blocks) {
3734 set_bit(STRIPE_HANDLE, &sh->state);
3738 if (unlikely(s.blocked_rdev)) {
3739 if (s.syncing || s.expanding || s.expanded ||
3740 s.replacing || s.to_write || s.written) {
3741 set_bit(STRIPE_HANDLE, &sh->state);
3744 /* There is nothing for the blocked_rdev to block */
3745 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3746 s.blocked_rdev = NULL;
3749 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3750 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3751 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3754 pr_debug("locked=%d uptodate=%d to_read=%d"
3755 " to_write=%d failed=%d failed_num=%d,%d\n",
3756 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3757 s.failed_num[0], s.failed_num[1]);
3758 /* check if the array has lost more than max_degraded devices and,
3759 * if so, some requests might need to be failed.
3761 if (s.failed > conf->max_degraded) {
3762 sh->check_state = 0;
3763 sh->reconstruct_state = 0;
3764 if (s.to_read+s.to_write+s.written)
3765 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3766 if (s.syncing + s.replacing)
3767 handle_failed_sync(conf, sh, &s);
3770 /* Now we check to see if any write operations have recently
3774 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3776 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3777 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3778 sh->reconstruct_state = reconstruct_state_idle;
3780 /* All the 'written' buffers and the parity block are ready to
3781 * be written back to disk
3783 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3784 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3785 BUG_ON(sh->qd_idx >= 0 &&
3786 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3787 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3788 for (i = disks; i--; ) {
3789 struct r5dev *dev = &sh->dev[i];
3790 if (test_bit(R5_LOCKED, &dev->flags) &&
3791 (i == sh->pd_idx || i == sh->qd_idx ||
3793 pr_debug("Writing block %d\n", i);
3794 set_bit(R5_Wantwrite, &dev->flags);
3797 if (!test_bit(R5_Insync, &dev->flags) ||
3798 ((i == sh->pd_idx || i == sh->qd_idx) &&
3800 set_bit(STRIPE_INSYNC, &sh->state);
3803 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3804 s.dec_preread_active = 1;
3808 * might be able to return some write requests if the parity blocks
3809 * are safe, or on a failed drive
3811 pdev = &sh->dev[sh->pd_idx];
3812 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3813 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3814 qdev = &sh->dev[sh->qd_idx];
3815 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3816 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3820 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3821 && !test_bit(R5_LOCKED, &pdev->flags)
3822 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3823 test_bit(R5_Discard, &pdev->flags))))) &&
3824 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3825 && !test_bit(R5_LOCKED, &qdev->flags)
3826 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3827 test_bit(R5_Discard, &qdev->flags))))))
3828 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3830 /* Now we might consider reading some blocks, either to check/generate
3831 * parity, or to satisfy requests
3832 * or to load a block that is being partially written.
3834 if (s.to_read || s.non_overwrite
3835 || (conf->level == 6 && s.to_write && s.failed)
3836 || (s.syncing && (s.uptodate + s.compute < disks))
3839 handle_stripe_fill(sh, &s, disks);
3841 /* Now to consider new write requests and what else, if anything
3842 * should be read. We do not handle new writes when:
3843 * 1/ A 'write' operation (copy+xor) is already in flight.
3844 * 2/ A 'check' operation is in flight, as it may clobber the parity
3847 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3848 handle_stripe_dirtying(conf, sh, &s, disks);
3850 /* maybe we need to check and possibly fix the parity for this stripe
3851 * Any reads will already have been scheduled, so we just see if enough
3852 * data is available. The parity check is held off while parity
3853 * dependent operations are in flight.
3855 if (sh->check_state ||
3856 (s.syncing && s.locked == 0 &&
3857 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3858 !test_bit(STRIPE_INSYNC, &sh->state))) {
3859 if (conf->level == 6)
3860 handle_parity_checks6(conf, sh, &s, disks);
3862 handle_parity_checks5(conf, sh, &s, disks);
3865 if ((s.replacing || s.syncing) && s.locked == 0
3866 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3867 && !test_bit(STRIPE_REPLACED, &sh->state)) {
3868 /* Write out to replacement devices where possible */
3869 for (i = 0; i < conf->raid_disks; i++)
3870 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3871 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
3872 set_bit(R5_WantReplace, &sh->dev[i].flags);
3873 set_bit(R5_LOCKED, &sh->dev[i].flags);
3877 set_bit(STRIPE_INSYNC, &sh->state);
3878 set_bit(STRIPE_REPLACED, &sh->state);
3880 if ((s.syncing || s.replacing) && s.locked == 0 &&
3881 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3882 test_bit(STRIPE_INSYNC, &sh->state)) {
3883 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3884 clear_bit(STRIPE_SYNCING, &sh->state);
3885 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3886 wake_up(&conf->wait_for_overlap);
3889 /* If the failed drives are just a ReadError, then we might need
3890 * to progress the repair/check process
3892 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3893 for (i = 0; i < s.failed; i++) {
3894 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3895 if (test_bit(R5_ReadError, &dev->flags)
3896 && !test_bit(R5_LOCKED, &dev->flags)
3897 && test_bit(R5_UPTODATE, &dev->flags)
3899 if (!test_bit(R5_ReWrite, &dev->flags)) {
3900 set_bit(R5_Wantwrite, &dev->flags);
3901 set_bit(R5_ReWrite, &dev->flags);
3902 set_bit(R5_LOCKED, &dev->flags);
3905 /* let's read it back */
3906 set_bit(R5_Wantread, &dev->flags);
3907 set_bit(R5_LOCKED, &dev->flags);
3914 /* Finish reconstruct operations initiated by the expansion process */
3915 if (sh->reconstruct_state == reconstruct_state_result) {
3916 struct stripe_head *sh_src
3917 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3918 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3919 /* sh cannot be written until sh_src has been read.
3920 * so arrange for sh to be delayed a little
3922 set_bit(STRIPE_DELAYED, &sh->state);
3923 set_bit(STRIPE_HANDLE, &sh->state);
3924 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3926 atomic_inc(&conf->preread_active_stripes);
3927 release_stripe(sh_src);
3931 release_stripe(sh_src);
3933 sh->reconstruct_state = reconstruct_state_idle;
3934 clear_bit(STRIPE_EXPANDING, &sh->state);
3935 for (i = conf->raid_disks; i--; ) {
3936 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3937 set_bit(R5_LOCKED, &sh->dev[i].flags);
3942 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3943 !sh->reconstruct_state) {
3944 /* Need to write out all blocks after computing parity */
3945 sh->disks = conf->raid_disks;
3946 stripe_set_idx(sh->sector, conf, 0, sh);
3947 schedule_reconstruction(sh, &s, 1, 1);
3948 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3949 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3950 atomic_dec(&conf->reshape_stripes);
3951 wake_up(&conf->wait_for_overlap);
3952 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3955 if (s.expanding && s.locked == 0 &&
3956 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3957 handle_stripe_expansion(conf, sh);
3960 /* wait for this device to become unblocked */
3961 if (unlikely(s.blocked_rdev)) {
3962 if (conf->mddev->external)
3963 md_wait_for_blocked_rdev(s.blocked_rdev,
3966 /* Internal metadata will immediately
3967 * be written by raid5d, so we don't
3968 * need to wait here.
3970 rdev_dec_pending(s.blocked_rdev,
3974 if (s.handle_bad_blocks)
3975 for (i = disks; i--; ) {
3976 struct md_rdev *rdev;
3977 struct r5dev *dev = &sh->dev[i];
3978 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3979 /* We own a safe reference to the rdev */
3980 rdev = conf->disks[i].rdev;
3981 if (!rdev_set_badblocks(rdev, sh->sector,
3983 md_error(conf->mddev, rdev);
3984 rdev_dec_pending(rdev, conf->mddev);
3986 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3987 rdev = conf->disks[i].rdev;
3988 rdev_clear_badblocks(rdev, sh->sector,
3990 rdev_dec_pending(rdev, conf->mddev);
3992 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3993 rdev = conf->disks[i].replacement;
3995 /* rdev have been moved down */
3996 rdev = conf->disks[i].rdev;
3997 rdev_clear_badblocks(rdev, sh->sector,
3999 rdev_dec_pending(rdev, conf->mddev);
4004 raid_run_ops(sh, s.ops_request);
4008 if (s.dec_preread_active) {
4009 /* We delay this until after ops_run_io so that if make_request
4010 * is waiting on a flush, it won't continue until the writes
4011 * have actually been submitted.
4013 atomic_dec(&conf->preread_active_stripes);
4014 if (atomic_read(&conf->preread_active_stripes) <
4016 md_wakeup_thread(conf->mddev->thread);
4019 return_io(s.return_bi);
4021 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4024 static void raid5_activate_delayed(struct r5conf *conf)
4026 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4027 while (!list_empty(&conf->delayed_list)) {
4028 struct list_head *l = conf->delayed_list.next;
4029 struct stripe_head *sh;
4030 sh = list_entry(l, struct stripe_head, lru);
4032 clear_bit(STRIPE_DELAYED, &sh->state);
4033 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4034 atomic_inc(&conf->preread_active_stripes);
4035 list_add_tail(&sh->lru, &conf->hold_list);
4036 raid5_wakeup_stripe_thread(sh);
4041 static void activate_bit_delay(struct r5conf *conf,
4042 struct list_head *temp_inactive_list)
4044 /* device_lock is held */
4045 struct list_head head;
4046 list_add(&head, &conf->bitmap_list);
4047 list_del_init(&conf->bitmap_list);
4048 while (!list_empty(&head)) {
4049 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4051 list_del_init(&sh->lru);
4052 atomic_inc(&sh->count);
4053 hash = sh->hash_lock_index;
4054 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4058 int md_raid5_congested(struct mddev *mddev, int bits)
4060 struct r5conf *conf = mddev->private;
4062 /* No difference between reads and writes. Just check
4063 * how busy the stripe_cache is
4066 if (conf->inactive_blocked)
4070 if (atomic_read(&conf->empty_inactive_list_nr))
4075 EXPORT_SYMBOL_GPL(md_raid5_congested);
4077 static int raid5_congested(void *data, int bits)
4079 struct mddev *mddev = data;
4081 return mddev_congested(mddev, bits) ||
4082 md_raid5_congested(mddev, bits);
4085 /* We want read requests to align with chunks where possible,
4086 * but write requests don't need to.
4088 static int raid5_mergeable_bvec(struct request_queue *q,
4089 struct bvec_merge_data *bvm,
4090 struct bio_vec *biovec)
4092 struct mddev *mddev = q->queuedata;
4093 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4095 unsigned int chunk_sectors = mddev->chunk_sectors;
4096 unsigned int bio_sectors = bvm->bi_size >> 9;
4098 if ((bvm->bi_rw & 1) == WRITE)
4099 return biovec->bv_len; /* always allow writes to be mergeable */
4101 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4102 chunk_sectors = mddev->new_chunk_sectors;
4103 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4104 if (max < 0) max = 0;
4105 if (max <= biovec->bv_len && bio_sectors == 0)
4106 return biovec->bv_len;
4112 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4114 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
4115 unsigned int chunk_sectors = mddev->chunk_sectors;
4116 unsigned int bio_sectors = bio_sectors(bio);
4118 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4119 chunk_sectors = mddev->new_chunk_sectors;
4120 return chunk_sectors >=
4121 ((sector & (chunk_sectors - 1)) + bio_sectors);
4125 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4126 * later sampled by raid5d.
4128 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4130 unsigned long flags;
4132 spin_lock_irqsave(&conf->device_lock, flags);
4134 bi->bi_next = conf->retry_read_aligned_list;
4135 conf->retry_read_aligned_list = bi;
4137 spin_unlock_irqrestore(&conf->device_lock, flags);
4138 md_wakeup_thread(conf->mddev->thread);
4142 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4146 bi = conf->retry_read_aligned;
4148 conf->retry_read_aligned = NULL;
4151 bi = conf->retry_read_aligned_list;
4153 conf->retry_read_aligned_list = bi->bi_next;
4156 * this sets the active strip count to 1 and the processed
4157 * strip count to zero (upper 8 bits)
4159 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4167 * The "raid5_align_endio" should check if the read succeeded and if it
4168 * did, call bio_endio on the original bio (having bio_put the new bio
4170 * If the read failed..
4172 static void raid5_align_endio(struct bio *bi, int error)
4174 struct bio* raid_bi = bi->bi_private;
4175 struct mddev *mddev;
4176 struct r5conf *conf;
4177 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4178 struct md_rdev *rdev;
4182 rdev = (void*)raid_bi->bi_next;
4183 raid_bi->bi_next = NULL;
4184 mddev = rdev->mddev;
4185 conf = mddev->private;
4187 rdev_dec_pending(rdev, conf->mddev);
4189 if (!error && uptodate) {
4190 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4192 bio_endio(raid_bi, 0);
4193 if (atomic_dec_and_test(&conf->active_aligned_reads))
4194 wake_up(&conf->wait_for_stripe);
4199 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4201 add_bio_to_retry(raid_bi, conf);
4204 static int bio_fits_rdev(struct bio *bi)
4206 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4208 if (bio_sectors(bi) > queue_max_sectors(q))
4210 blk_recount_segments(q, bi);
4211 if (bi->bi_phys_segments > queue_max_segments(q))
4214 if (q->merge_bvec_fn)
4215 /* it's too hard to apply the merge_bvec_fn at this stage,
4224 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4226 struct r5conf *conf = mddev->private;
4228 struct bio* align_bi;
4229 struct md_rdev *rdev;
4230 sector_t end_sector;
4232 if (!in_chunk_boundary(mddev, raid_bio)) {
4233 pr_debug("chunk_aligned_read : non aligned\n");
4237 * use bio_clone_mddev to make a copy of the bio
4239 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4243 * set bi_end_io to a new function, and set bi_private to the
4246 align_bi->bi_end_io = raid5_align_endio;
4247 align_bi->bi_private = raid_bio;
4251 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
4255 end_sector = bio_end_sector(align_bi);
4257 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4258 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4259 rdev->recovery_offset < end_sector) {
4260 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4262 (test_bit(Faulty, &rdev->flags) ||
4263 !(test_bit(In_sync, &rdev->flags) ||
4264 rdev->recovery_offset >= end_sector)))
4271 atomic_inc(&rdev->nr_pending);
4273 raid_bio->bi_next = (void*)rdev;
4274 align_bi->bi_bdev = rdev->bdev;
4275 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4277 if (!bio_fits_rdev(align_bi) ||
4278 is_badblock(rdev, align_bi->bi_sector, bio_sectors(align_bi),
4279 &first_bad, &bad_sectors)) {
4280 /* too big in some way, or has a known bad block */
4282 rdev_dec_pending(rdev, mddev);
4286 /* No reshape active, so we can trust rdev->data_offset */
4287 align_bi->bi_sector += rdev->data_offset;
4289 spin_lock_irq(&conf->device_lock);
4290 wait_event_lock_irq(conf->wait_for_stripe,
4293 atomic_inc(&conf->active_aligned_reads);
4294 spin_unlock_irq(&conf->device_lock);
4297 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4298 align_bi, disk_devt(mddev->gendisk),
4299 raid_bio->bi_sector);
4300 generic_make_request(align_bi);
4309 /* __get_priority_stripe - get the next stripe to process
4311 * Full stripe writes are allowed to pass preread active stripes up until
4312 * the bypass_threshold is exceeded. In general the bypass_count
4313 * increments when the handle_list is handled before the hold_list; however, it
4314 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4315 * stripe with in flight i/o. The bypass_count will be reset when the
4316 * head of the hold_list has changed, i.e. the head was promoted to the
4319 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4321 struct stripe_head *sh = NULL, *tmp;
4322 struct list_head *handle_list = NULL;
4323 struct r5worker_group *wg = NULL;
4325 if (conf->worker_cnt_per_group == 0) {
4326 handle_list = &conf->handle_list;
4327 } else if (group != ANY_GROUP) {
4328 handle_list = &conf->worker_groups[group].handle_list;
4329 wg = &conf->worker_groups[group];
4332 for (i = 0; i < conf->group_cnt; i++) {
4333 handle_list = &conf->worker_groups[i].handle_list;
4334 wg = &conf->worker_groups[i];
4335 if (!list_empty(handle_list))
4340 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4342 list_empty(handle_list) ? "empty" : "busy",
4343 list_empty(&conf->hold_list) ? "empty" : "busy",
4344 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4346 if (!list_empty(handle_list)) {
4347 sh = list_entry(handle_list->next, typeof(*sh), lru);
4349 if (list_empty(&conf->hold_list))
4350 conf->bypass_count = 0;
4351 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4352 if (conf->hold_list.next == conf->last_hold)
4353 conf->bypass_count++;
4355 conf->last_hold = conf->hold_list.next;
4356 conf->bypass_count -= conf->bypass_threshold;
4357 if (conf->bypass_count < 0)
4358 conf->bypass_count = 0;
4361 } else if (!list_empty(&conf->hold_list) &&
4362 ((conf->bypass_threshold &&
4363 conf->bypass_count > conf->bypass_threshold) ||
4364 atomic_read(&conf->pending_full_writes) == 0)) {
4366 list_for_each_entry(tmp, &conf->hold_list, lru) {
4367 if (conf->worker_cnt_per_group == 0 ||
4368 group == ANY_GROUP ||
4369 !cpu_online(tmp->cpu) ||
4370 cpu_to_group(tmp->cpu) == group) {
4377 conf->bypass_count -= conf->bypass_threshold;
4378 if (conf->bypass_count < 0)
4379 conf->bypass_count = 0;
4391 list_del_init(&sh->lru);
4392 atomic_inc(&sh->count);
4393 BUG_ON(atomic_read(&sh->count) != 1);
4397 struct raid5_plug_cb {
4398 struct blk_plug_cb cb;
4399 struct list_head list;
4400 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4403 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4405 struct raid5_plug_cb *cb = container_of(
4406 blk_cb, struct raid5_plug_cb, cb);
4407 struct stripe_head *sh;
4408 struct mddev *mddev = cb->cb.data;
4409 struct r5conf *conf = mddev->private;
4413 if (cb->list.next && !list_empty(&cb->list)) {
4414 spin_lock_irq(&conf->device_lock);
4415 while (!list_empty(&cb->list)) {
4416 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4417 list_del_init(&sh->lru);
4419 * avoid race release_stripe_plug() sees
4420 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4421 * is still in our list
4423 smp_mb__before_clear_bit();
4424 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4426 * STRIPE_ON_RELEASE_LIST could be set here. In that
4427 * case, the count is always > 1 here
4429 hash = sh->hash_lock_index;
4430 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
4433 spin_unlock_irq(&conf->device_lock);
4435 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4436 NR_STRIPE_HASH_LOCKS);
4438 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4442 static void release_stripe_plug(struct mddev *mddev,
4443 struct stripe_head *sh)
4445 struct blk_plug_cb *blk_cb = blk_check_plugged(
4446 raid5_unplug, mddev,
4447 sizeof(struct raid5_plug_cb));
4448 struct raid5_plug_cb *cb;
4455 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4457 if (cb->list.next == NULL) {
4459 INIT_LIST_HEAD(&cb->list);
4460 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
4461 INIT_LIST_HEAD(cb->temp_inactive_list + i);
4464 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4465 list_add_tail(&sh->lru, &cb->list);
4470 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4472 struct r5conf *conf = mddev->private;
4473 sector_t logical_sector, last_sector;
4474 struct stripe_head *sh;
4478 if (mddev->reshape_position != MaxSector)
4479 /* Skip discard while reshape is happening */
4482 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4483 last_sector = bi->bi_sector + (bi->bi_size>>9);
4486 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4488 stripe_sectors = conf->chunk_sectors *
4489 (conf->raid_disks - conf->max_degraded);
4490 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4492 sector_div(last_sector, stripe_sectors);
4494 logical_sector *= conf->chunk_sectors;
4495 last_sector *= conf->chunk_sectors;
4497 for (; logical_sector < last_sector;
4498 logical_sector += STRIPE_SECTORS) {
4502 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4503 prepare_to_wait(&conf->wait_for_overlap, &w,
4504 TASK_UNINTERRUPTIBLE);
4505 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4506 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4511 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4512 spin_lock_irq(&sh->stripe_lock);
4513 for (d = 0; d < conf->raid_disks; d++) {
4514 if (d == sh->pd_idx || d == sh->qd_idx)
4516 if (sh->dev[d].towrite || sh->dev[d].toread) {
4517 set_bit(R5_Overlap, &sh->dev[d].flags);
4518 spin_unlock_irq(&sh->stripe_lock);
4524 set_bit(STRIPE_DISCARD, &sh->state);
4525 finish_wait(&conf->wait_for_overlap, &w);
4526 for (d = 0; d < conf->raid_disks; d++) {
4527 if (d == sh->pd_idx || d == sh->qd_idx)
4529 sh->dev[d].towrite = bi;
4530 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4531 raid5_inc_bi_active_stripes(bi);
4533 spin_unlock_irq(&sh->stripe_lock);
4534 if (conf->mddev->bitmap) {
4536 d < conf->raid_disks - conf->max_degraded;
4538 bitmap_startwrite(mddev->bitmap,
4542 sh->bm_seq = conf->seq_flush + 1;
4543 set_bit(STRIPE_BIT_DELAY, &sh->state);
4546 set_bit(STRIPE_HANDLE, &sh->state);
4547 clear_bit(STRIPE_DELAYED, &sh->state);
4548 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4549 atomic_inc(&conf->preread_active_stripes);
4550 release_stripe_plug(mddev, sh);
4553 remaining = raid5_dec_bi_active_stripes(bi);
4554 if (remaining == 0) {
4555 md_write_end(mddev);
4560 static void make_request(struct mddev *mddev, struct bio * bi)
4562 struct r5conf *conf = mddev->private;
4564 sector_t new_sector;
4565 sector_t logical_sector, last_sector;
4566 struct stripe_head *sh;
4567 const int rw = bio_data_dir(bi);
4570 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4571 md_flush_request(mddev, bi);
4575 md_write_start(mddev, bi);
4578 mddev->reshape_position == MaxSector &&
4579 chunk_aligned_read(mddev,bi))
4582 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4583 make_discard_request(mddev, bi);
4587 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4588 last_sector = bio_end_sector(bi);
4590 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4592 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4598 seq = read_seqcount_begin(&conf->gen_lock);
4600 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4601 if (unlikely(conf->reshape_progress != MaxSector)) {
4602 /* spinlock is needed as reshape_progress may be
4603 * 64bit on a 32bit platform, and so it might be
4604 * possible to see a half-updated value
4605 * Of course reshape_progress could change after
4606 * the lock is dropped, so once we get a reference
4607 * to the stripe that we think it is, we will have
4610 spin_lock_irq(&conf->device_lock);
4611 if (mddev->reshape_backwards
4612 ? logical_sector < conf->reshape_progress
4613 : logical_sector >= conf->reshape_progress) {
4616 if (mddev->reshape_backwards
4617 ? logical_sector < conf->reshape_safe
4618 : logical_sector >= conf->reshape_safe) {
4619 spin_unlock_irq(&conf->device_lock);
4624 spin_unlock_irq(&conf->device_lock);
4627 new_sector = raid5_compute_sector(conf, logical_sector,
4630 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4631 (unsigned long long)new_sector,
4632 (unsigned long long)logical_sector);
4634 sh = get_active_stripe(conf, new_sector, previous,
4635 (bi->bi_rw&RWA_MASK), 0);
4637 if (unlikely(previous)) {
4638 /* expansion might have moved on while waiting for a
4639 * stripe, so we must do the range check again.
4640 * Expansion could still move past after this
4641 * test, but as we are holding a reference to
4642 * 'sh', we know that if that happens,
4643 * STRIPE_EXPANDING will get set and the expansion
4644 * won't proceed until we finish with the stripe.
4647 spin_lock_irq(&conf->device_lock);
4648 if (mddev->reshape_backwards
4649 ? logical_sector >= conf->reshape_progress
4650 : logical_sector < conf->reshape_progress)
4651 /* mismatch, need to try again */
4653 spin_unlock_irq(&conf->device_lock);
4660 if (read_seqcount_retry(&conf->gen_lock, seq)) {
4661 /* Might have got the wrong stripe_head
4669 logical_sector >= mddev->suspend_lo &&
4670 logical_sector < mddev->suspend_hi) {
4672 /* As the suspend_* range is controlled by
4673 * userspace, we want an interruptible
4676 flush_signals(current);
4677 prepare_to_wait(&conf->wait_for_overlap,
4678 &w, TASK_INTERRUPTIBLE);
4679 if (logical_sector >= mddev->suspend_lo &&
4680 logical_sector < mddev->suspend_hi)
4685 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4686 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4687 /* Stripe is busy expanding or
4688 * add failed due to overlap. Flush everything
4691 md_wakeup_thread(mddev->thread);
4696 finish_wait(&conf->wait_for_overlap, &w);
4697 set_bit(STRIPE_HANDLE, &sh->state);
4698 clear_bit(STRIPE_DELAYED, &sh->state);
4699 if ((bi->bi_rw & REQ_SYNC) &&
4700 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4701 atomic_inc(&conf->preread_active_stripes);
4702 release_stripe_plug(mddev, sh);
4704 /* cannot get stripe for read-ahead, just give-up */
4705 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4706 finish_wait(&conf->wait_for_overlap, &w);
4711 remaining = raid5_dec_bi_active_stripes(bi);
4712 if (remaining == 0) {
4715 md_write_end(mddev);
4717 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4723 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4725 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4727 /* reshaping is quite different to recovery/resync so it is
4728 * handled quite separately ... here.
4730 * On each call to sync_request, we gather one chunk worth of
4731 * destination stripes and flag them as expanding.
4732 * Then we find all the source stripes and request reads.
4733 * As the reads complete, handle_stripe will copy the data
4734 * into the destination stripe and release that stripe.
4736 struct r5conf *conf = mddev->private;
4737 struct stripe_head *sh;
4738 sector_t first_sector, last_sector;
4739 int raid_disks = conf->previous_raid_disks;
4740 int data_disks = raid_disks - conf->max_degraded;
4741 int new_data_disks = conf->raid_disks - conf->max_degraded;
4744 sector_t writepos, readpos, safepos;
4745 sector_t stripe_addr;
4746 int reshape_sectors;
4747 struct list_head stripes;
4749 if (sector_nr == 0) {
4750 /* If restarting in the middle, skip the initial sectors */
4751 if (mddev->reshape_backwards &&
4752 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4753 sector_nr = raid5_size(mddev, 0, 0)
4754 - conf->reshape_progress;
4755 } else if (!mddev->reshape_backwards &&
4756 conf->reshape_progress > 0)
4757 sector_nr = conf->reshape_progress;
4758 sector_div(sector_nr, new_data_disks);
4760 mddev->curr_resync_completed = sector_nr;
4761 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4767 /* We need to process a full chunk at a time.
4768 * If old and new chunk sizes differ, we need to process the
4771 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4772 reshape_sectors = mddev->new_chunk_sectors;
4774 reshape_sectors = mddev->chunk_sectors;
4776 /* We update the metadata at least every 10 seconds, or when
4777 * the data about to be copied would over-write the source of
4778 * the data at the front of the range. i.e. one new_stripe
4779 * along from reshape_progress new_maps to after where
4780 * reshape_safe old_maps to
4782 writepos = conf->reshape_progress;
4783 sector_div(writepos, new_data_disks);
4784 readpos = conf->reshape_progress;
4785 sector_div(readpos, data_disks);
4786 safepos = conf->reshape_safe;
4787 sector_div(safepos, data_disks);
4788 if (mddev->reshape_backwards) {
4789 writepos -= min_t(sector_t, reshape_sectors, writepos);
4790 readpos += reshape_sectors;
4791 safepos += reshape_sectors;
4793 writepos += reshape_sectors;
4794 readpos -= min_t(sector_t, reshape_sectors, readpos);
4795 safepos -= min_t(sector_t, reshape_sectors, safepos);
4798 /* Having calculated the 'writepos' possibly use it
4799 * to set 'stripe_addr' which is where we will write to.
4801 if (mddev->reshape_backwards) {
4802 BUG_ON(conf->reshape_progress == 0);
4803 stripe_addr = writepos;
4804 BUG_ON((mddev->dev_sectors &
4805 ~((sector_t)reshape_sectors - 1))
4806 - reshape_sectors - stripe_addr
4809 BUG_ON(writepos != sector_nr + reshape_sectors);
4810 stripe_addr = sector_nr;
4813 /* 'writepos' is the most advanced device address we might write.
4814 * 'readpos' is the least advanced device address we might read.
4815 * 'safepos' is the least address recorded in the metadata as having
4817 * If there is a min_offset_diff, these are adjusted either by
4818 * increasing the safepos/readpos if diff is negative, or
4819 * increasing writepos if diff is positive.
4820 * If 'readpos' is then behind 'writepos', there is no way that we can
4821 * ensure safety in the face of a crash - that must be done by userspace
4822 * making a backup of the data. So in that case there is no particular
4823 * rush to update metadata.
4824 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4825 * update the metadata to advance 'safepos' to match 'readpos' so that
4826 * we can be safe in the event of a crash.
4827 * So we insist on updating metadata if safepos is behind writepos and
4828 * readpos is beyond writepos.
4829 * In any case, update the metadata every 10 seconds.
4830 * Maybe that number should be configurable, but I'm not sure it is
4831 * worth it.... maybe it could be a multiple of safemode_delay???
4833 if (conf->min_offset_diff < 0) {
4834 safepos += -conf->min_offset_diff;
4835 readpos += -conf->min_offset_diff;
4837 writepos += conf->min_offset_diff;
4839 if ((mddev->reshape_backwards
4840 ? (safepos > writepos && readpos < writepos)
4841 : (safepos < writepos && readpos > writepos)) ||
4842 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4843 /* Cannot proceed until we've updated the superblock... */
4844 wait_event(conf->wait_for_overlap,
4845 atomic_read(&conf->reshape_stripes)==0
4846 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4847 if (atomic_read(&conf->reshape_stripes) != 0)
4849 mddev->reshape_position = conf->reshape_progress;
4850 mddev->curr_resync_completed = sector_nr;
4851 conf->reshape_checkpoint = jiffies;
4852 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4853 md_wakeup_thread(mddev->thread);
4854 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4855 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4856 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4858 spin_lock_irq(&conf->device_lock);
4859 conf->reshape_safe = mddev->reshape_position;
4860 spin_unlock_irq(&conf->device_lock);
4861 wake_up(&conf->wait_for_overlap);
4862 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4865 INIT_LIST_HEAD(&stripes);
4866 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4868 int skipped_disk = 0;
4869 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4870 set_bit(STRIPE_EXPANDING, &sh->state);
4871 atomic_inc(&conf->reshape_stripes);
4872 /* If any of this stripe is beyond the end of the old
4873 * array, then we need to zero those blocks
4875 for (j=sh->disks; j--;) {
4877 if (j == sh->pd_idx)
4879 if (conf->level == 6 &&
4882 s = compute_blocknr(sh, j, 0);
4883 if (s < raid5_size(mddev, 0, 0)) {
4887 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4888 set_bit(R5_Expanded, &sh->dev[j].flags);
4889 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4891 if (!skipped_disk) {
4892 set_bit(STRIPE_EXPAND_READY, &sh->state);
4893 set_bit(STRIPE_HANDLE, &sh->state);
4895 list_add(&sh->lru, &stripes);
4897 spin_lock_irq(&conf->device_lock);
4898 if (mddev->reshape_backwards)
4899 conf->reshape_progress -= reshape_sectors * new_data_disks;
4901 conf->reshape_progress += reshape_sectors * new_data_disks;
4902 spin_unlock_irq(&conf->device_lock);
4903 /* Ok, those stripe are ready. We can start scheduling
4904 * reads on the source stripes.
4905 * The source stripes are determined by mapping the first and last
4906 * block on the destination stripes.
4909 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4912 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4913 * new_data_disks - 1),
4915 if (last_sector >= mddev->dev_sectors)
4916 last_sector = mddev->dev_sectors - 1;
4917 while (first_sector <= last_sector) {
4918 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4919 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4920 set_bit(STRIPE_HANDLE, &sh->state);
4922 first_sector += STRIPE_SECTORS;
4924 /* Now that the sources are clearly marked, we can release
4925 * the destination stripes
4927 while (!list_empty(&stripes)) {
4928 sh = list_entry(stripes.next, struct stripe_head, lru);
4929 list_del_init(&sh->lru);
4932 /* If this takes us to the resync_max point where we have to pause,
4933 * then we need to write out the superblock.
4935 sector_nr += reshape_sectors;
4936 if ((sector_nr - mddev->curr_resync_completed) * 2
4937 >= mddev->resync_max - mddev->curr_resync_completed) {
4938 /* Cannot proceed until we've updated the superblock... */
4939 wait_event(conf->wait_for_overlap,
4940 atomic_read(&conf->reshape_stripes) == 0
4941 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4942 if (atomic_read(&conf->reshape_stripes) != 0)
4944 mddev->reshape_position = conf->reshape_progress;
4945 mddev->curr_resync_completed = sector_nr;
4946 conf->reshape_checkpoint = jiffies;
4947 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4948 md_wakeup_thread(mddev->thread);
4949 wait_event(mddev->sb_wait,
4950 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4951 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4952 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4954 spin_lock_irq(&conf->device_lock);
4955 conf->reshape_safe = mddev->reshape_position;
4956 spin_unlock_irq(&conf->device_lock);
4957 wake_up(&conf->wait_for_overlap);
4958 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4961 return reshape_sectors;
4964 /* FIXME go_faster isn't used */
4965 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4967 struct r5conf *conf = mddev->private;
4968 struct stripe_head *sh;
4969 sector_t max_sector = mddev->dev_sectors;
4970 sector_t sync_blocks;
4971 int still_degraded = 0;
4974 if (sector_nr >= max_sector) {
4975 /* just being told to finish up .. nothing much to do */
4977 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4982 if (mddev->curr_resync < max_sector) /* aborted */
4983 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4985 else /* completed sync */
4987 bitmap_close_sync(mddev->bitmap);
4992 /* Allow raid5_quiesce to complete */
4993 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4995 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4996 return reshape_request(mddev, sector_nr, skipped);
4998 /* No need to check resync_max as we never do more than one
4999 * stripe, and as resync_max will always be on a chunk boundary,
5000 * if the check in md_do_sync didn't fire, there is no chance
5001 * of overstepping resync_max here
5004 /* if there is too many failed drives and we are trying
5005 * to resync, then assert that we are finished, because there is
5006 * nothing we can do.
5008 if (mddev->degraded >= conf->max_degraded &&
5009 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5010 sector_t rv = mddev->dev_sectors - sector_nr;
5014 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5016 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5017 sync_blocks >= STRIPE_SECTORS) {
5018 /* we can skip this block, and probably more */
5019 sync_blocks /= STRIPE_SECTORS;
5021 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5024 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5026 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5028 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5029 /* make sure we don't swamp the stripe cache if someone else
5030 * is trying to get access
5032 schedule_timeout_uninterruptible(1);
5034 /* Need to check if array will still be degraded after recovery/resync
5035 * We don't need to check the 'failed' flag as when that gets set,
5038 for (i = 0; i < conf->raid_disks; i++)
5039 if (conf->disks[i].rdev == NULL)
5042 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5044 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5049 return STRIPE_SECTORS;
5052 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5054 /* We may not be able to submit a whole bio at once as there
5055 * may not be enough stripe_heads available.
5056 * We cannot pre-allocate enough stripe_heads as we may need
5057 * more than exist in the cache (if we allow ever large chunks).
5058 * So we do one stripe head at a time and record in
5059 * ->bi_hw_segments how many have been done.
5061 * We *know* that this entire raid_bio is in one chunk, so
5062 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5064 struct stripe_head *sh;
5066 sector_t sector, logical_sector, last_sector;
5071 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5072 sector = raid5_compute_sector(conf, logical_sector,
5074 last_sector = bio_end_sector(raid_bio);
5076 for (; logical_sector < last_sector;
5077 logical_sector += STRIPE_SECTORS,
5078 sector += STRIPE_SECTORS,
5081 if (scnt < raid5_bi_processed_stripes(raid_bio))
5082 /* already done this stripe */
5085 sh = get_active_stripe(conf, sector, 0, 1, 0);
5088 /* failed to get a stripe - must wait */
5089 raid5_set_bi_processed_stripes(raid_bio, scnt);
5090 conf->retry_read_aligned = raid_bio;
5094 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
5096 raid5_set_bi_processed_stripes(raid_bio, scnt);
5097 conf->retry_read_aligned = raid_bio;
5101 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5106 remaining = raid5_dec_bi_active_stripes(raid_bio);
5107 if (remaining == 0) {
5108 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5110 bio_endio(raid_bio, 0);
5112 if (atomic_dec_and_test(&conf->active_aligned_reads))
5113 wake_up(&conf->wait_for_stripe);
5117 static int handle_active_stripes(struct r5conf *conf, int group,
5118 struct r5worker *worker,
5119 struct list_head *temp_inactive_list)
5121 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5122 int i, batch_size = 0, hash;
5123 bool release_inactive = false;
5125 while (batch_size < MAX_STRIPE_BATCH &&
5126 (sh = __get_priority_stripe(conf, group)) != NULL)
5127 batch[batch_size++] = sh;
5129 if (batch_size == 0) {
5130 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5131 if (!list_empty(temp_inactive_list + i))
5133 if (i == NR_STRIPE_HASH_LOCKS)
5135 release_inactive = true;
5137 spin_unlock_irq(&conf->device_lock);
5139 release_inactive_stripe_list(conf, temp_inactive_list,
5140 NR_STRIPE_HASH_LOCKS);
5142 if (release_inactive) {
5143 spin_lock_irq(&conf->device_lock);
5147 for (i = 0; i < batch_size; i++)
5148 handle_stripe(batch[i]);
5152 spin_lock_irq(&conf->device_lock);
5153 for (i = 0; i < batch_size; i++) {
5154 hash = batch[i]->hash_lock_index;
5155 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5160 static void raid5_do_work(struct work_struct *work)
5162 struct r5worker *worker = container_of(work, struct r5worker, work);
5163 struct r5worker_group *group = worker->group;
5164 struct r5conf *conf = group->conf;
5165 int group_id = group - conf->worker_groups;
5167 struct blk_plug plug;
5169 pr_debug("+++ raid5worker active\n");
5171 blk_start_plug(&plug);
5173 spin_lock_irq(&conf->device_lock);
5175 int batch_size, released;
5177 released = release_stripe_list(conf, worker->temp_inactive_list);
5179 batch_size = handle_active_stripes(conf, group_id, worker,
5180 worker->temp_inactive_list);
5181 worker->working = false;
5182 if (!batch_size && !released)
5184 handled += batch_size;
5186 pr_debug("%d stripes handled\n", handled);
5188 spin_unlock_irq(&conf->device_lock);
5189 blk_finish_plug(&plug);
5191 pr_debug("--- raid5worker inactive\n");
5195 * This is our raid5 kernel thread.
5197 * We scan the hash table for stripes which can be handled now.
5198 * During the scan, completed stripes are saved for us by the interrupt
5199 * handler, so that they will not have to wait for our next wakeup.
5201 static void raid5d(struct md_thread *thread)
5203 struct mddev *mddev = thread->mddev;
5204 struct r5conf *conf = mddev->private;
5206 struct blk_plug plug;
5208 pr_debug("+++ raid5d active\n");
5210 md_check_recovery(mddev);
5212 blk_start_plug(&plug);
5214 spin_lock_irq(&conf->device_lock);
5217 int batch_size, released;
5219 released = release_stripe_list(conf, conf->temp_inactive_list);
5222 !list_empty(&conf->bitmap_list)) {
5223 /* Now is a good time to flush some bitmap updates */
5225 spin_unlock_irq(&conf->device_lock);
5226 bitmap_unplug(mddev->bitmap);
5227 spin_lock_irq(&conf->device_lock);
5228 conf->seq_write = conf->seq_flush;
5229 activate_bit_delay(conf, conf->temp_inactive_list);
5231 raid5_activate_delayed(conf);
5233 while ((bio = remove_bio_from_retry(conf))) {
5235 spin_unlock_irq(&conf->device_lock);
5236 ok = retry_aligned_read(conf, bio);
5237 spin_lock_irq(&conf->device_lock);
5243 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5244 conf->temp_inactive_list);
5245 if (!batch_size && !released)
5247 handled += batch_size;
5249 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5250 spin_unlock_irq(&conf->device_lock);
5251 md_check_recovery(mddev);
5252 spin_lock_irq(&conf->device_lock);
5255 pr_debug("%d stripes handled\n", handled);
5257 spin_unlock_irq(&conf->device_lock);
5259 async_tx_issue_pending_all();
5260 blk_finish_plug(&plug);
5262 pr_debug("--- raid5d inactive\n");
5266 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5268 struct r5conf *conf = mddev->private;
5270 return sprintf(page, "%d\n", conf->max_nr_stripes);
5276 raid5_set_cache_size(struct mddev *mddev, int size)
5278 struct r5conf *conf = mddev->private;
5282 if (size <= 16 || size > 32768)
5284 hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
5285 while (size < conf->max_nr_stripes) {
5286 if (drop_one_stripe(conf, hash))
5287 conf->max_nr_stripes--;
5292 hash = NR_STRIPE_HASH_LOCKS - 1;
5294 err = md_allow_write(mddev);
5297 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
5298 while (size > conf->max_nr_stripes) {
5299 if (grow_one_stripe(conf, hash))
5300 conf->max_nr_stripes++;
5302 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
5306 EXPORT_SYMBOL(raid5_set_cache_size);
5309 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5311 struct r5conf *conf = mddev->private;
5315 if (len >= PAGE_SIZE)
5320 if (kstrtoul(page, 10, &new))
5322 err = raid5_set_cache_size(mddev, new);
5328 static struct md_sysfs_entry
5329 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5330 raid5_show_stripe_cache_size,
5331 raid5_store_stripe_cache_size);
5334 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5336 struct r5conf *conf = mddev->private;
5338 return sprintf(page, "%d\n", conf->bypass_threshold);
5344 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
5346 struct r5conf *conf = mddev->private;
5348 if (len >= PAGE_SIZE)
5353 if (kstrtoul(page, 10, &new))
5355 if (new > conf->max_nr_stripes)
5357 conf->bypass_threshold = new;
5361 static struct md_sysfs_entry
5362 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5364 raid5_show_preread_threshold,
5365 raid5_store_preread_threshold);
5368 stripe_cache_active_show(struct mddev *mddev, char *page)
5370 struct r5conf *conf = mddev->private;
5372 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5377 static struct md_sysfs_entry
5378 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
5381 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
5383 struct r5conf *conf = mddev->private;
5385 return sprintf(page, "%d\n", conf->worker_cnt_per_group);
5390 static int alloc_thread_groups(struct r5conf *conf, int cnt,
5392 int *worker_cnt_per_group,
5393 struct r5worker_group **worker_groups);
5395 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
5397 struct r5conf *conf = mddev->private;
5400 struct r5worker_group *new_groups, *old_groups;
5401 int group_cnt, worker_cnt_per_group;
5403 if (len >= PAGE_SIZE)
5408 if (kstrtoul(page, 10, &new))
5411 if (new == conf->worker_cnt_per_group)
5414 mddev_suspend(mddev);
5416 old_groups = conf->worker_groups;
5418 flush_workqueue(raid5_wq);
5420 err = alloc_thread_groups(conf, new,
5421 &group_cnt, &worker_cnt_per_group,
5424 spin_lock_irq(&conf->device_lock);
5425 conf->group_cnt = group_cnt;
5426 conf->worker_cnt_per_group = worker_cnt_per_group;
5427 conf->worker_groups = new_groups;
5428 spin_unlock_irq(&conf->device_lock);
5431 kfree(old_groups[0].workers);
5435 mddev_resume(mddev);
5442 static struct md_sysfs_entry
5443 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
5444 raid5_show_group_thread_cnt,
5445 raid5_store_group_thread_cnt);
5447 static struct attribute *raid5_attrs[] = {
5448 &raid5_stripecache_size.attr,
5449 &raid5_stripecache_active.attr,
5450 &raid5_preread_bypass_threshold.attr,
5451 &raid5_group_thread_cnt.attr,
5454 static struct attribute_group raid5_attrs_group = {
5456 .attrs = raid5_attrs,
5459 static int alloc_thread_groups(struct r5conf *conf, int cnt,
5461 int *worker_cnt_per_group,
5462 struct r5worker_group **worker_groups)
5466 struct r5worker *workers;
5468 *worker_cnt_per_group = cnt;
5471 *worker_groups = NULL;
5474 *group_cnt = num_possible_nodes();
5475 size = sizeof(struct r5worker) * cnt;
5476 workers = kzalloc(size * *group_cnt, GFP_NOIO);
5477 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
5478 *group_cnt, GFP_NOIO);
5479 if (!*worker_groups || !workers) {
5481 kfree(*worker_groups);
5485 for (i = 0; i < *group_cnt; i++) {
5486 struct r5worker_group *group;
5488 group = worker_groups[i];
5489 INIT_LIST_HEAD(&group->handle_list);
5491 group->workers = workers + i * cnt;
5493 for (j = 0; j < cnt; j++) {
5494 struct r5worker *worker = group->workers + j;
5495 worker->group = group;
5496 INIT_WORK(&worker->work, raid5_do_work);
5498 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
5499 INIT_LIST_HEAD(worker->temp_inactive_list + k);
5506 static void free_thread_groups(struct r5conf *conf)
5508 if (conf->worker_groups)
5509 kfree(conf->worker_groups[0].workers);
5510 kfree(conf->worker_groups);
5511 conf->worker_groups = NULL;
5515 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
5517 struct r5conf *conf = mddev->private;
5520 sectors = mddev->dev_sectors;
5522 /* size is defined by the smallest of previous and new size */
5523 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
5525 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5526 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
5527 return sectors * (raid_disks - conf->max_degraded);
5530 static void raid5_free_percpu(struct r5conf *conf)
5532 struct raid5_percpu *percpu;
5539 for_each_possible_cpu(cpu) {
5540 percpu = per_cpu_ptr(conf->percpu, cpu);
5541 safe_put_page(percpu->spare_page);
5542 kfree(percpu->scribble);
5544 #ifdef CONFIG_HOTPLUG_CPU
5545 unregister_cpu_notifier(&conf->cpu_notify);
5549 free_percpu(conf->percpu);
5552 static void free_conf(struct r5conf *conf)
5554 free_thread_groups(conf);
5555 shrink_stripes(conf);
5556 raid5_free_percpu(conf);
5558 kfree(conf->stripe_hashtbl);
5562 #ifdef CONFIG_HOTPLUG_CPU
5563 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5566 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5567 long cpu = (long)hcpu;
5568 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5571 case CPU_UP_PREPARE:
5572 case CPU_UP_PREPARE_FROZEN:
5573 if (conf->level == 6 && !percpu->spare_page)
5574 percpu->spare_page = alloc_page(GFP_KERNEL);
5575 if (!percpu->scribble)
5576 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5578 if (!percpu->scribble ||
5579 (conf->level == 6 && !percpu->spare_page)) {
5580 safe_put_page(percpu->spare_page);
5581 kfree(percpu->scribble);
5582 pr_err("%s: failed memory allocation for cpu%ld\n",
5584 return notifier_from_errno(-ENOMEM);
5588 case CPU_DEAD_FROZEN:
5589 safe_put_page(percpu->spare_page);
5590 kfree(percpu->scribble);
5591 percpu->spare_page = NULL;
5592 percpu->scribble = NULL;
5601 static int raid5_alloc_percpu(struct r5conf *conf)
5604 struct page *spare_page;
5605 struct raid5_percpu __percpu *allcpus;
5609 allcpus = alloc_percpu(struct raid5_percpu);
5612 conf->percpu = allcpus;
5616 for_each_present_cpu(cpu) {
5617 if (conf->level == 6) {
5618 spare_page = alloc_page(GFP_KERNEL);
5623 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5625 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5630 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5632 #ifdef CONFIG_HOTPLUG_CPU
5633 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5634 conf->cpu_notify.priority = 0;
5636 err = register_cpu_notifier(&conf->cpu_notify);
5643 static struct r5conf *setup_conf(struct mddev *mddev)
5645 struct r5conf *conf;
5646 int raid_disk, memory, max_disks;
5647 struct md_rdev *rdev;
5648 struct disk_info *disk;
5651 int group_cnt, worker_cnt_per_group;
5652 struct r5worker_group *new_group;
5654 if (mddev->new_level != 5
5655 && mddev->new_level != 4
5656 && mddev->new_level != 6) {
5657 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5658 mdname(mddev), mddev->new_level);
5659 return ERR_PTR(-EIO);
5661 if ((mddev->new_level == 5
5662 && !algorithm_valid_raid5(mddev->new_layout)) ||
5663 (mddev->new_level == 6
5664 && !algorithm_valid_raid6(mddev->new_layout))) {
5665 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5666 mdname(mddev), mddev->new_layout);
5667 return ERR_PTR(-EIO);
5669 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5670 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5671 mdname(mddev), mddev->raid_disks);
5672 return ERR_PTR(-EINVAL);
5675 if (!mddev->new_chunk_sectors ||
5676 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5677 !is_power_of_2(mddev->new_chunk_sectors)) {
5678 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5679 mdname(mddev), mddev->new_chunk_sectors << 9);
5680 return ERR_PTR(-EINVAL);
5683 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5686 /* Don't enable multi-threading by default*/
5687 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
5689 conf->group_cnt = group_cnt;
5690 conf->worker_cnt_per_group = worker_cnt_per_group;
5691 conf->worker_groups = new_group;
5694 spin_lock_init(&conf->device_lock);
5695 seqcount_init(&conf->gen_lock);
5696 init_waitqueue_head(&conf->wait_for_stripe);
5697 init_waitqueue_head(&conf->wait_for_overlap);
5698 INIT_LIST_HEAD(&conf->handle_list);
5699 INIT_LIST_HEAD(&conf->hold_list);
5700 INIT_LIST_HEAD(&conf->delayed_list);
5701 INIT_LIST_HEAD(&conf->bitmap_list);
5702 init_llist_head(&conf->released_stripes);
5703 atomic_set(&conf->active_stripes, 0);
5704 atomic_set(&conf->preread_active_stripes, 0);
5705 atomic_set(&conf->active_aligned_reads, 0);
5706 conf->bypass_threshold = BYPASS_THRESHOLD;
5707 conf->recovery_disabled = mddev->recovery_disabled - 1;
5709 conf->raid_disks = mddev->raid_disks;
5710 if (mddev->reshape_position == MaxSector)
5711 conf->previous_raid_disks = mddev->raid_disks;
5713 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5714 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5715 conf->scribble_len = scribble_len(max_disks);
5717 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5722 conf->mddev = mddev;
5724 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5727 /* We init hash_locks[0] separately to that it can be used
5728 * as the reference lock in the spin_lock_nest_lock() call
5729 * in lock_all_device_hash_locks_irq in order to convince
5730 * lockdep that we know what we are doing.
5732 spin_lock_init(conf->hash_locks);
5733 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
5734 spin_lock_init(conf->hash_locks + i);
5736 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5737 INIT_LIST_HEAD(conf->inactive_list + i);
5739 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5740 INIT_LIST_HEAD(conf->temp_inactive_list + i);
5742 conf->level = mddev->new_level;
5743 if (raid5_alloc_percpu(conf) != 0)
5746 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5748 rdev_for_each(rdev, mddev) {
5749 raid_disk = rdev->raid_disk;
5750 if (raid_disk >= max_disks
5753 disk = conf->disks + raid_disk;
5755 if (test_bit(Replacement, &rdev->flags)) {
5756 if (disk->replacement)
5758 disk->replacement = rdev;
5765 if (test_bit(In_sync, &rdev->flags)) {
5766 char b[BDEVNAME_SIZE];
5767 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5769 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5770 } else if (rdev->saved_raid_disk != raid_disk)
5771 /* Cannot rely on bitmap to complete recovery */
5775 conf->chunk_sectors = mddev->new_chunk_sectors;
5776 conf->level = mddev->new_level;
5777 if (conf->level == 6)
5778 conf->max_degraded = 2;
5780 conf->max_degraded = 1;
5781 conf->algorithm = mddev->new_layout;
5782 conf->reshape_progress = mddev->reshape_position;
5783 if (conf->reshape_progress != MaxSector) {
5784 conf->prev_chunk_sectors = mddev->chunk_sectors;
5785 conf->prev_algo = mddev->layout;
5788 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5789 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5790 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
5791 if (grow_stripes(conf, NR_STRIPES)) {
5793 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5794 mdname(mddev), memory);
5797 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5798 mdname(mddev), memory);
5800 sprintf(pers_name, "raid%d", mddev->new_level);
5801 conf->thread = md_register_thread(raid5d, mddev, pers_name);
5802 if (!conf->thread) {
5804 "md/raid:%s: couldn't allocate thread.\n",
5814 return ERR_PTR(-EIO);
5816 return ERR_PTR(-ENOMEM);
5820 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5823 case ALGORITHM_PARITY_0:
5824 if (raid_disk < max_degraded)
5827 case ALGORITHM_PARITY_N:
5828 if (raid_disk >= raid_disks - max_degraded)
5831 case ALGORITHM_PARITY_0_6:
5832 if (raid_disk == 0 ||
5833 raid_disk == raid_disks - 1)
5836 case ALGORITHM_LEFT_ASYMMETRIC_6:
5837 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5838 case ALGORITHM_LEFT_SYMMETRIC_6:
5839 case ALGORITHM_RIGHT_SYMMETRIC_6:
5840 if (raid_disk == raid_disks - 1)
5846 static int run(struct mddev *mddev)
5848 struct r5conf *conf;
5849 int working_disks = 0;
5850 int dirty_parity_disks = 0;
5851 struct md_rdev *rdev;
5852 sector_t reshape_offset = 0;
5854 long long min_offset_diff = 0;
5857 if (mddev->recovery_cp != MaxSector)
5858 printk(KERN_NOTICE "md/raid:%s: not clean"
5859 " -- starting background reconstruction\n",
5862 rdev_for_each(rdev, mddev) {
5864 if (rdev->raid_disk < 0)
5866 diff = (rdev->new_data_offset - rdev->data_offset);
5868 min_offset_diff = diff;
5870 } else if (mddev->reshape_backwards &&
5871 diff < min_offset_diff)
5872 min_offset_diff = diff;
5873 else if (!mddev->reshape_backwards &&
5874 diff > min_offset_diff)
5875 min_offset_diff = diff;
5878 if (mddev->reshape_position != MaxSector) {
5879 /* Check that we can continue the reshape.
5880 * Difficulties arise if the stripe we would write to
5881 * next is at or after the stripe we would read from next.
5882 * For a reshape that changes the number of devices, this
5883 * is only possible for a very short time, and mdadm makes
5884 * sure that time appears to have past before assembling
5885 * the array. So we fail if that time hasn't passed.
5886 * For a reshape that keeps the number of devices the same
5887 * mdadm must be monitoring the reshape can keeping the
5888 * critical areas read-only and backed up. It will start
5889 * the array in read-only mode, so we check for that.
5891 sector_t here_new, here_old;
5893 int max_degraded = (mddev->level == 6 ? 2 : 1);
5895 if (mddev->new_level != mddev->level) {
5896 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5897 "required - aborting.\n",
5901 old_disks = mddev->raid_disks - mddev->delta_disks;
5902 /* reshape_position must be on a new-stripe boundary, and one
5903 * further up in new geometry must map after here in old
5906 here_new = mddev->reshape_position;
5907 if (sector_div(here_new, mddev->new_chunk_sectors *
5908 (mddev->raid_disks - max_degraded))) {
5909 printk(KERN_ERR "md/raid:%s: reshape_position not "
5910 "on a stripe boundary\n", mdname(mddev));
5913 reshape_offset = here_new * mddev->new_chunk_sectors;
5914 /* here_new is the stripe we will write to */
5915 here_old = mddev->reshape_position;
5916 sector_div(here_old, mddev->chunk_sectors *
5917 (old_disks-max_degraded));
5918 /* here_old is the first stripe that we might need to read
5920 if (mddev->delta_disks == 0) {
5921 if ((here_new * mddev->new_chunk_sectors !=
5922 here_old * mddev->chunk_sectors)) {
5923 printk(KERN_ERR "md/raid:%s: reshape position is"
5924 " confused - aborting\n", mdname(mddev));
5927 /* We cannot be sure it is safe to start an in-place
5928 * reshape. It is only safe if user-space is monitoring
5929 * and taking constant backups.
5930 * mdadm always starts a situation like this in
5931 * readonly mode so it can take control before
5932 * allowing any writes. So just check for that.
5934 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5935 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5936 /* not really in-place - so OK */;
5937 else if (mddev->ro == 0) {
5938 printk(KERN_ERR "md/raid:%s: in-place reshape "
5939 "must be started in read-only mode "
5944 } else if (mddev->reshape_backwards
5945 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5946 here_old * mddev->chunk_sectors)
5947 : (here_new * mddev->new_chunk_sectors >=
5948 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5949 /* Reading from the same stripe as writing to - bad */
5950 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5951 "auto-recovery - aborting.\n",
5955 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5957 /* OK, we should be able to continue; */
5959 BUG_ON(mddev->level != mddev->new_level);
5960 BUG_ON(mddev->layout != mddev->new_layout);
5961 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5962 BUG_ON(mddev->delta_disks != 0);
5965 if (mddev->private == NULL)
5966 conf = setup_conf(mddev);
5968 conf = mddev->private;
5971 return PTR_ERR(conf);
5973 conf->min_offset_diff = min_offset_diff;
5974 mddev->thread = conf->thread;
5975 conf->thread = NULL;
5976 mddev->private = conf;
5978 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5980 rdev = conf->disks[i].rdev;
5981 if (!rdev && conf->disks[i].replacement) {
5982 /* The replacement is all we have yet */
5983 rdev = conf->disks[i].replacement;
5984 conf->disks[i].replacement = NULL;
5985 clear_bit(Replacement, &rdev->flags);
5986 conf->disks[i].rdev = rdev;
5990 if (conf->disks[i].replacement &&
5991 conf->reshape_progress != MaxSector) {
5992 /* replacements and reshape simply do not mix. */
5993 printk(KERN_ERR "md: cannot handle concurrent "
5994 "replacement and reshape.\n");
5997 if (test_bit(In_sync, &rdev->flags)) {
6001 /* This disc is not fully in-sync. However if it
6002 * just stored parity (beyond the recovery_offset),
6003 * when we don't need to be concerned about the
6004 * array being dirty.
6005 * When reshape goes 'backwards', we never have
6006 * partially completed devices, so we only need
6007 * to worry about reshape going forwards.
6009 /* Hack because v0.91 doesn't store recovery_offset properly. */
6010 if (mddev->major_version == 0 &&
6011 mddev->minor_version > 90)
6012 rdev->recovery_offset = reshape_offset;
6014 if (rdev->recovery_offset < reshape_offset) {
6015 /* We need to check old and new layout */
6016 if (!only_parity(rdev->raid_disk,
6019 conf->max_degraded))
6022 if (!only_parity(rdev->raid_disk,
6024 conf->previous_raid_disks,
6025 conf->max_degraded))
6027 dirty_parity_disks++;
6031 * 0 for a fully functional array, 1 or 2 for a degraded array.
6033 mddev->degraded = calc_degraded(conf);
6035 if (has_failed(conf)) {
6036 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6037 " (%d/%d failed)\n",
6038 mdname(mddev), mddev->degraded, conf->raid_disks);
6042 /* device size must be a multiple of chunk size */
6043 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6044 mddev->resync_max_sectors = mddev->dev_sectors;
6046 if (mddev->degraded > dirty_parity_disks &&
6047 mddev->recovery_cp != MaxSector) {
6048 if (mddev->ok_start_degraded)
6050 "md/raid:%s: starting dirty degraded array"
6051 " - data corruption possible.\n",
6055 "md/raid:%s: cannot start dirty degraded array.\n",
6061 if (mddev->degraded == 0)
6062 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6063 " devices, algorithm %d\n", mdname(mddev), conf->level,
6064 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6067 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6068 " out of %d devices, algorithm %d\n",
6069 mdname(mddev), conf->level,
6070 mddev->raid_disks - mddev->degraded,
6071 mddev->raid_disks, mddev->new_layout);
6073 print_raid5_conf(conf);
6075 if (conf->reshape_progress != MaxSector) {
6076 conf->reshape_safe = conf->reshape_progress;
6077 atomic_set(&conf->reshape_stripes, 0);
6078 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6079 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6080 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6081 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6082 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6087 /* Ok, everything is just fine now */
6088 if (mddev->to_remove == &raid5_attrs_group)
6089 mddev->to_remove = NULL;
6090 else if (mddev->kobj.sd &&
6091 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6093 "raid5: failed to create sysfs attributes for %s\n",
6095 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6099 bool discard_supported = true;
6100 /* read-ahead size must cover two whole stripes, which
6101 * is 2 * (datadisks) * chunksize where 'n' is the
6102 * number of raid devices
6104 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6105 int stripe = data_disks *
6106 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6107 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6108 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6110 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
6112 mddev->queue->backing_dev_info.congested_data = mddev;
6113 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
6115 chunk_size = mddev->chunk_sectors << 9;
6116 blk_queue_io_min(mddev->queue, chunk_size);
6117 blk_queue_io_opt(mddev->queue, chunk_size *
6118 (conf->raid_disks - conf->max_degraded));
6120 * We can only discard a whole stripe. It doesn't make sense to
6121 * discard data disk but write parity disk
6123 stripe = stripe * PAGE_SIZE;
6124 /* Round up to power of 2, as discard handling
6125 * currently assumes that */
6126 while ((stripe-1) & stripe)
6127 stripe = (stripe | (stripe-1)) + 1;
6128 mddev->queue->limits.discard_alignment = stripe;
6129 mddev->queue->limits.discard_granularity = stripe;
6131 * unaligned part of discard request will be ignored, so can't
6132 * guarantee discard_zerors_data
6134 mddev->queue->limits.discard_zeroes_data = 0;
6136 blk_queue_max_write_same_sectors(mddev->queue, 0);
6138 rdev_for_each(rdev, mddev) {
6139 disk_stack_limits(mddev->gendisk, rdev->bdev,
6140 rdev->data_offset << 9);
6141 disk_stack_limits(mddev->gendisk, rdev->bdev,
6142 rdev->new_data_offset << 9);
6144 * discard_zeroes_data is required, otherwise data
6145 * could be lost. Consider a scenario: discard a stripe
6146 * (the stripe could be inconsistent if
6147 * discard_zeroes_data is 0); write one disk of the
6148 * stripe (the stripe could be inconsistent again
6149 * depending on which disks are used to calculate
6150 * parity); the disk is broken; The stripe data of this
6153 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6154 !bdev_get_queue(rdev->bdev)->
6155 limits.discard_zeroes_data)
6156 discard_supported = false;
6159 if (discard_supported &&
6160 mddev->queue->limits.max_discard_sectors >= stripe &&
6161 mddev->queue->limits.discard_granularity >= stripe)
6162 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6165 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6171 md_unregister_thread(&mddev->thread);
6172 print_raid5_conf(conf);
6174 mddev->private = NULL;
6175 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6179 static int stop(struct mddev *mddev)
6181 struct r5conf *conf = mddev->private;
6183 md_unregister_thread(&mddev->thread);
6185 mddev->queue->backing_dev_info.congested_fn = NULL;
6187 mddev->private = NULL;
6188 mddev->to_remove = &raid5_attrs_group;
6192 static void status(struct seq_file *seq, struct mddev *mddev)
6194 struct r5conf *conf = mddev->private;
6197 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6198 mddev->chunk_sectors / 2, mddev->layout);
6199 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6200 for (i = 0; i < conf->raid_disks; i++)
6201 seq_printf (seq, "%s",
6202 conf->disks[i].rdev &&
6203 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6204 seq_printf (seq, "]");
6207 static void print_raid5_conf (struct r5conf *conf)
6210 struct disk_info *tmp;
6212 printk(KERN_DEBUG "RAID conf printout:\n");
6214 printk("(conf==NULL)\n");
6217 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6219 conf->raid_disks - conf->mddev->degraded);
6221 for (i = 0; i < conf->raid_disks; i++) {
6222 char b[BDEVNAME_SIZE];
6223 tmp = conf->disks + i;
6225 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6226 i, !test_bit(Faulty, &tmp->rdev->flags),
6227 bdevname(tmp->rdev->bdev, b));
6231 static int raid5_spare_active(struct mddev *mddev)
6234 struct r5conf *conf = mddev->private;
6235 struct disk_info *tmp;
6237 unsigned long flags;
6239 for (i = 0; i < conf->raid_disks; i++) {
6240 tmp = conf->disks + i;
6241 if (tmp->replacement
6242 && tmp->replacement->recovery_offset == MaxSector
6243 && !test_bit(Faulty, &tmp->replacement->flags)
6244 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
6245 /* Replacement has just become active. */
6247 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
6250 /* Replaced device not technically faulty,
6251 * but we need to be sure it gets removed
6252 * and never re-added.
6254 set_bit(Faulty, &tmp->rdev->flags);
6255 sysfs_notify_dirent_safe(
6256 tmp->rdev->sysfs_state);
6258 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
6259 } else if (tmp->rdev
6260 && tmp->rdev->recovery_offset == MaxSector
6261 && !test_bit(Faulty, &tmp->rdev->flags)
6262 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
6264 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
6267 spin_lock_irqsave(&conf->device_lock, flags);
6268 mddev->degraded = calc_degraded(conf);
6269 spin_unlock_irqrestore(&conf->device_lock, flags);
6270 print_raid5_conf(conf);
6274 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
6276 struct r5conf *conf = mddev->private;
6278 int number = rdev->raid_disk;
6279 struct md_rdev **rdevp;
6280 struct disk_info *p = conf->disks + number;
6282 print_raid5_conf(conf);
6283 if (rdev == p->rdev)
6285 else if (rdev == p->replacement)
6286 rdevp = &p->replacement;
6290 if (number >= conf->raid_disks &&
6291 conf->reshape_progress == MaxSector)
6292 clear_bit(In_sync, &rdev->flags);
6294 if (test_bit(In_sync, &rdev->flags) ||
6295 atomic_read(&rdev->nr_pending)) {
6299 /* Only remove non-faulty devices if recovery
6302 if (!test_bit(Faulty, &rdev->flags) &&
6303 mddev->recovery_disabled != conf->recovery_disabled &&
6304 !has_failed(conf) &&
6305 (!p->replacement || p->replacement == rdev) &&
6306 number < conf->raid_disks) {
6312 if (atomic_read(&rdev->nr_pending)) {
6313 /* lost the race, try later */
6316 } else if (p->replacement) {
6317 /* We must have just cleared 'rdev' */
6318 p->rdev = p->replacement;
6319 clear_bit(Replacement, &p->replacement->flags);
6320 smp_mb(); /* Make sure other CPUs may see both as identical
6321 * but will never see neither - if they are careful
6323 p->replacement = NULL;
6324 clear_bit(WantReplacement, &rdev->flags);
6326 /* We might have just removed the Replacement as faulty-
6327 * clear the bit just in case
6329 clear_bit(WantReplacement, &rdev->flags);
6332 print_raid5_conf(conf);
6336 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
6338 struct r5conf *conf = mddev->private;
6341 struct disk_info *p;
6343 int last = conf->raid_disks - 1;
6345 if (mddev->recovery_disabled == conf->recovery_disabled)
6348 if (rdev->saved_raid_disk < 0 && has_failed(conf))
6349 /* no point adding a device */
6352 if (rdev->raid_disk >= 0)
6353 first = last = rdev->raid_disk;
6356 * find the disk ... but prefer rdev->saved_raid_disk
6359 if (rdev->saved_raid_disk >= 0 &&
6360 rdev->saved_raid_disk >= first &&
6361 conf->disks[rdev->saved_raid_disk].rdev == NULL)
6362 first = rdev->saved_raid_disk;
6364 for (disk = first; disk <= last; disk++) {
6365 p = conf->disks + disk;
6366 if (p->rdev == NULL) {
6367 clear_bit(In_sync, &rdev->flags);
6368 rdev->raid_disk = disk;
6370 if (rdev->saved_raid_disk != disk)
6372 rcu_assign_pointer(p->rdev, rdev);
6376 for (disk = first; disk <= last; disk++) {
6377 p = conf->disks + disk;
6378 if (test_bit(WantReplacement, &p->rdev->flags) &&
6379 p->replacement == NULL) {
6380 clear_bit(In_sync, &rdev->flags);
6381 set_bit(Replacement, &rdev->flags);
6382 rdev->raid_disk = disk;
6385 rcu_assign_pointer(p->replacement, rdev);
6390 print_raid5_conf(conf);
6394 static int raid5_resize(struct mddev *mddev, sector_t sectors)
6396 /* no resync is happening, and there is enough space
6397 * on all devices, so we can resize.
6398 * We need to make sure resync covers any new space.
6399 * If the array is shrinking we should possibly wait until
6400 * any io in the removed space completes, but it hardly seems
6404 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6405 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
6406 if (mddev->external_size &&
6407 mddev->array_sectors > newsize)
6409 if (mddev->bitmap) {
6410 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
6414 md_set_array_sectors(mddev, newsize);
6415 set_capacity(mddev->gendisk, mddev->array_sectors);
6416 revalidate_disk(mddev->gendisk);
6417 if (sectors > mddev->dev_sectors &&
6418 mddev->recovery_cp > mddev->dev_sectors) {
6419 mddev->recovery_cp = mddev->dev_sectors;
6420 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
6422 mddev->dev_sectors = sectors;
6423 mddev->resync_max_sectors = sectors;
6427 static int check_stripe_cache(struct mddev *mddev)
6429 /* Can only proceed if there are plenty of stripe_heads.
6430 * We need a minimum of one full stripe,, and for sensible progress
6431 * it is best to have about 4 times that.
6432 * If we require 4 times, then the default 256 4K stripe_heads will
6433 * allow for chunk sizes up to 256K, which is probably OK.
6434 * If the chunk size is greater, user-space should request more
6435 * stripe_heads first.
6437 struct r5conf *conf = mddev->private;
6438 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
6439 > conf->max_nr_stripes ||
6440 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
6441 > conf->max_nr_stripes) {
6442 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
6444 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
6451 static int check_reshape(struct mddev *mddev)
6453 struct r5conf *conf = mddev->private;
6455 if (mddev->delta_disks == 0 &&
6456 mddev->new_layout == mddev->layout &&
6457 mddev->new_chunk_sectors == mddev->chunk_sectors)
6458 return 0; /* nothing to do */
6459 if (has_failed(conf))
6461 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
6462 /* We might be able to shrink, but the devices must
6463 * be made bigger first.
6464 * For raid6, 4 is the minimum size.
6465 * Otherwise 2 is the minimum
6468 if (mddev->level == 6)
6470 if (mddev->raid_disks + mddev->delta_disks < min)
6474 if (!check_stripe_cache(mddev))
6477 return resize_stripes(conf, (conf->previous_raid_disks
6478 + mddev->delta_disks));
6481 static int raid5_start_reshape(struct mddev *mddev)
6483 struct r5conf *conf = mddev->private;
6484 struct md_rdev *rdev;
6486 unsigned long flags;
6488 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
6491 if (!check_stripe_cache(mddev))
6494 if (has_failed(conf))
6497 rdev_for_each(rdev, mddev) {
6498 if (!test_bit(In_sync, &rdev->flags)
6499 && !test_bit(Faulty, &rdev->flags))
6503 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
6504 /* Not enough devices even to make a degraded array
6509 /* Refuse to reduce size of the array. Any reductions in
6510 * array size must be through explicit setting of array_size
6513 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
6514 < mddev->array_sectors) {
6515 printk(KERN_ERR "md/raid:%s: array size must be reduced "
6516 "before number of disks\n", mdname(mddev));
6520 atomic_set(&conf->reshape_stripes, 0);
6521 spin_lock_irq(&conf->device_lock);
6522 write_seqcount_begin(&conf->gen_lock);
6523 conf->previous_raid_disks = conf->raid_disks;
6524 conf->raid_disks += mddev->delta_disks;
6525 conf->prev_chunk_sectors = conf->chunk_sectors;
6526 conf->chunk_sectors = mddev->new_chunk_sectors;
6527 conf->prev_algo = conf->algorithm;
6528 conf->algorithm = mddev->new_layout;
6530 /* Code that selects data_offset needs to see the generation update
6531 * if reshape_progress has been set - so a memory barrier needed.
6534 if (mddev->reshape_backwards)
6535 conf->reshape_progress = raid5_size(mddev, 0, 0);
6537 conf->reshape_progress = 0;
6538 conf->reshape_safe = conf->reshape_progress;
6539 write_seqcount_end(&conf->gen_lock);
6540 spin_unlock_irq(&conf->device_lock);
6542 /* Now make sure any requests that proceeded on the assumption
6543 * the reshape wasn't running - like Discard or Read - have
6546 mddev_suspend(mddev);
6547 mddev_resume(mddev);
6549 /* Add some new drives, as many as will fit.
6550 * We know there are enough to make the newly sized array work.
6551 * Don't add devices if we are reducing the number of
6552 * devices in the array. This is because it is not possible
6553 * to correctly record the "partially reconstructed" state of
6554 * such devices during the reshape and confusion could result.
6556 if (mddev->delta_disks >= 0) {
6557 rdev_for_each(rdev, mddev)
6558 if (rdev->raid_disk < 0 &&
6559 !test_bit(Faulty, &rdev->flags)) {
6560 if (raid5_add_disk(mddev, rdev) == 0) {
6562 >= conf->previous_raid_disks)
6563 set_bit(In_sync, &rdev->flags);
6565 rdev->recovery_offset = 0;
6567 if (sysfs_link_rdev(mddev, rdev))
6568 /* Failure here is OK */;
6570 } else if (rdev->raid_disk >= conf->previous_raid_disks
6571 && !test_bit(Faulty, &rdev->flags)) {
6572 /* This is a spare that was manually added */
6573 set_bit(In_sync, &rdev->flags);
6576 /* When a reshape changes the number of devices,
6577 * ->degraded is measured against the larger of the
6578 * pre and post number of devices.
6580 spin_lock_irqsave(&conf->device_lock, flags);
6581 mddev->degraded = calc_degraded(conf);
6582 spin_unlock_irqrestore(&conf->device_lock, flags);
6584 mddev->raid_disks = conf->raid_disks;
6585 mddev->reshape_position = conf->reshape_progress;
6586 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6588 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6589 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6590 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6591 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6592 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6594 if (!mddev->sync_thread) {
6595 mddev->recovery = 0;
6596 spin_lock_irq(&conf->device_lock);
6597 write_seqcount_begin(&conf->gen_lock);
6598 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6599 mddev->new_chunk_sectors =
6600 conf->chunk_sectors = conf->prev_chunk_sectors;
6601 mddev->new_layout = conf->algorithm = conf->prev_algo;
6602 rdev_for_each(rdev, mddev)
6603 rdev->new_data_offset = rdev->data_offset;
6605 conf->generation --;
6606 conf->reshape_progress = MaxSector;
6607 mddev->reshape_position = MaxSector;
6608 write_seqcount_end(&conf->gen_lock);
6609 spin_unlock_irq(&conf->device_lock);
6612 conf->reshape_checkpoint = jiffies;
6613 md_wakeup_thread(mddev->sync_thread);
6614 md_new_event(mddev);
6618 /* This is called from the reshape thread and should make any
6619 * changes needed in 'conf'
6621 static void end_reshape(struct r5conf *conf)
6624 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6625 struct md_rdev *rdev;
6627 spin_lock_irq(&conf->device_lock);
6628 conf->previous_raid_disks = conf->raid_disks;
6629 rdev_for_each(rdev, conf->mddev)
6630 rdev->data_offset = rdev->new_data_offset;
6632 conf->reshape_progress = MaxSector;
6633 spin_unlock_irq(&conf->device_lock);
6634 wake_up(&conf->wait_for_overlap);
6636 /* read-ahead size must cover two whole stripes, which is
6637 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6639 if (conf->mddev->queue) {
6640 int data_disks = conf->raid_disks - conf->max_degraded;
6641 int stripe = data_disks * ((conf->chunk_sectors << 9)
6643 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6644 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6649 /* This is called from the raid5d thread with mddev_lock held.
6650 * It makes config changes to the device.
6652 static void raid5_finish_reshape(struct mddev *mddev)
6654 struct r5conf *conf = mddev->private;
6656 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6658 if (mddev->delta_disks > 0) {
6659 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6660 set_capacity(mddev->gendisk, mddev->array_sectors);
6661 revalidate_disk(mddev->gendisk);
6664 spin_lock_irq(&conf->device_lock);
6665 mddev->degraded = calc_degraded(conf);
6666 spin_unlock_irq(&conf->device_lock);
6667 for (d = conf->raid_disks ;
6668 d < conf->raid_disks - mddev->delta_disks;
6670 struct md_rdev *rdev = conf->disks[d].rdev;
6672 clear_bit(In_sync, &rdev->flags);
6673 rdev = conf->disks[d].replacement;
6675 clear_bit(In_sync, &rdev->flags);
6678 mddev->layout = conf->algorithm;
6679 mddev->chunk_sectors = conf->chunk_sectors;
6680 mddev->reshape_position = MaxSector;
6681 mddev->delta_disks = 0;
6682 mddev->reshape_backwards = 0;
6686 static void raid5_quiesce(struct mddev *mddev, int state)
6688 struct r5conf *conf = mddev->private;
6691 case 2: /* resume for a suspend */
6692 wake_up(&conf->wait_for_overlap);
6695 case 1: /* stop all writes */
6696 lock_all_device_hash_locks_irq(conf);
6697 /* '2' tells resync/reshape to pause so that all
6698 * active stripes can drain
6701 wait_event_cmd(conf->wait_for_stripe,
6702 atomic_read(&conf->active_stripes) == 0 &&
6703 atomic_read(&conf->active_aligned_reads) == 0,
6704 unlock_all_device_hash_locks_irq(conf),
6705 lock_all_device_hash_locks_irq(conf));
6707 unlock_all_device_hash_locks_irq(conf);
6708 /* allow reshape to continue */
6709 wake_up(&conf->wait_for_overlap);
6712 case 0: /* re-enable writes */
6713 lock_all_device_hash_locks_irq(conf);
6715 wake_up(&conf->wait_for_stripe);
6716 wake_up(&conf->wait_for_overlap);
6717 unlock_all_device_hash_locks_irq(conf);
6723 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6725 struct r0conf *raid0_conf = mddev->private;
6728 /* for raid0 takeover only one zone is supported */
6729 if (raid0_conf->nr_strip_zones > 1) {
6730 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6732 return ERR_PTR(-EINVAL);
6735 sectors = raid0_conf->strip_zone[0].zone_end;
6736 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6737 mddev->dev_sectors = sectors;
6738 mddev->new_level = level;
6739 mddev->new_layout = ALGORITHM_PARITY_N;
6740 mddev->new_chunk_sectors = mddev->chunk_sectors;
6741 mddev->raid_disks += 1;
6742 mddev->delta_disks = 1;
6743 /* make sure it will be not marked as dirty */
6744 mddev->recovery_cp = MaxSector;
6746 return setup_conf(mddev);
6750 static void *raid5_takeover_raid1(struct mddev *mddev)
6754 if (mddev->raid_disks != 2 ||
6755 mddev->degraded > 1)
6756 return ERR_PTR(-EINVAL);
6758 /* Should check if there are write-behind devices? */
6760 chunksect = 64*2; /* 64K by default */
6762 /* The array must be an exact multiple of chunksize */
6763 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6766 if ((chunksect<<9) < STRIPE_SIZE)
6767 /* array size does not allow a suitable chunk size */
6768 return ERR_PTR(-EINVAL);
6770 mddev->new_level = 5;
6771 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6772 mddev->new_chunk_sectors = chunksect;
6774 return setup_conf(mddev);
6777 static void *raid5_takeover_raid6(struct mddev *mddev)
6781 switch (mddev->layout) {
6782 case ALGORITHM_LEFT_ASYMMETRIC_6:
6783 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6785 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6786 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6788 case ALGORITHM_LEFT_SYMMETRIC_6:
6789 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6791 case ALGORITHM_RIGHT_SYMMETRIC_6:
6792 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6794 case ALGORITHM_PARITY_0_6:
6795 new_layout = ALGORITHM_PARITY_0;
6797 case ALGORITHM_PARITY_N:
6798 new_layout = ALGORITHM_PARITY_N;
6801 return ERR_PTR(-EINVAL);
6803 mddev->new_level = 5;
6804 mddev->new_layout = new_layout;
6805 mddev->delta_disks = -1;
6806 mddev->raid_disks -= 1;
6807 return setup_conf(mddev);
6811 static int raid5_check_reshape(struct mddev *mddev)
6813 /* For a 2-drive array, the layout and chunk size can be changed
6814 * immediately as not restriping is needed.
6815 * For larger arrays we record the new value - after validation
6816 * to be used by a reshape pass.
6818 struct r5conf *conf = mddev->private;
6819 int new_chunk = mddev->new_chunk_sectors;
6821 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6823 if (new_chunk > 0) {
6824 if (!is_power_of_2(new_chunk))
6826 if (new_chunk < (PAGE_SIZE>>9))
6828 if (mddev->array_sectors & (new_chunk-1))
6829 /* not factor of array size */
6833 /* They look valid */
6835 if (mddev->raid_disks == 2) {
6836 /* can make the change immediately */
6837 if (mddev->new_layout >= 0) {
6838 conf->algorithm = mddev->new_layout;
6839 mddev->layout = mddev->new_layout;
6841 if (new_chunk > 0) {
6842 conf->chunk_sectors = new_chunk ;
6843 mddev->chunk_sectors = new_chunk;
6845 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6846 md_wakeup_thread(mddev->thread);
6848 return check_reshape(mddev);
6851 static int raid6_check_reshape(struct mddev *mddev)
6853 int new_chunk = mddev->new_chunk_sectors;
6855 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6857 if (new_chunk > 0) {
6858 if (!is_power_of_2(new_chunk))
6860 if (new_chunk < (PAGE_SIZE >> 9))
6862 if (mddev->array_sectors & (new_chunk-1))
6863 /* not factor of array size */
6867 /* They look valid */
6868 return check_reshape(mddev);
6871 static void *raid5_takeover(struct mddev *mddev)
6873 /* raid5 can take over:
6874 * raid0 - if there is only one strip zone - make it a raid4 layout
6875 * raid1 - if there are two drives. We need to know the chunk size
6876 * raid4 - trivial - just use a raid4 layout.
6877 * raid6 - Providing it is a *_6 layout
6879 if (mddev->level == 0)
6880 return raid45_takeover_raid0(mddev, 5);
6881 if (mddev->level == 1)
6882 return raid5_takeover_raid1(mddev);
6883 if (mddev->level == 4) {
6884 mddev->new_layout = ALGORITHM_PARITY_N;
6885 mddev->new_level = 5;
6886 return setup_conf(mddev);
6888 if (mddev->level == 6)
6889 return raid5_takeover_raid6(mddev);
6891 return ERR_PTR(-EINVAL);
6894 static void *raid4_takeover(struct mddev *mddev)
6896 /* raid4 can take over:
6897 * raid0 - if there is only one strip zone
6898 * raid5 - if layout is right
6900 if (mddev->level == 0)
6901 return raid45_takeover_raid0(mddev, 4);
6902 if (mddev->level == 5 &&
6903 mddev->layout == ALGORITHM_PARITY_N) {
6904 mddev->new_layout = 0;
6905 mddev->new_level = 4;
6906 return setup_conf(mddev);
6908 return ERR_PTR(-EINVAL);
6911 static struct md_personality raid5_personality;
6913 static void *raid6_takeover(struct mddev *mddev)
6915 /* Currently can only take over a raid5. We map the
6916 * personality to an equivalent raid6 personality
6917 * with the Q block at the end.
6921 if (mddev->pers != &raid5_personality)
6922 return ERR_PTR(-EINVAL);
6923 if (mddev->degraded > 1)
6924 return ERR_PTR(-EINVAL);
6925 if (mddev->raid_disks > 253)
6926 return ERR_PTR(-EINVAL);
6927 if (mddev->raid_disks < 3)
6928 return ERR_PTR(-EINVAL);
6930 switch (mddev->layout) {
6931 case ALGORITHM_LEFT_ASYMMETRIC:
6932 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6934 case ALGORITHM_RIGHT_ASYMMETRIC:
6935 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6937 case ALGORITHM_LEFT_SYMMETRIC:
6938 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6940 case ALGORITHM_RIGHT_SYMMETRIC:
6941 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6943 case ALGORITHM_PARITY_0:
6944 new_layout = ALGORITHM_PARITY_0_6;
6946 case ALGORITHM_PARITY_N:
6947 new_layout = ALGORITHM_PARITY_N;
6950 return ERR_PTR(-EINVAL);
6952 mddev->new_level = 6;
6953 mddev->new_layout = new_layout;
6954 mddev->delta_disks = 1;
6955 mddev->raid_disks += 1;
6956 return setup_conf(mddev);
6960 static struct md_personality raid6_personality =
6964 .owner = THIS_MODULE,
6965 .make_request = make_request,
6969 .error_handler = error,
6970 .hot_add_disk = raid5_add_disk,
6971 .hot_remove_disk= raid5_remove_disk,
6972 .spare_active = raid5_spare_active,
6973 .sync_request = sync_request,
6974 .resize = raid5_resize,
6976 .check_reshape = raid6_check_reshape,
6977 .start_reshape = raid5_start_reshape,
6978 .finish_reshape = raid5_finish_reshape,
6979 .quiesce = raid5_quiesce,
6980 .takeover = raid6_takeover,
6982 static struct md_personality raid5_personality =
6986 .owner = THIS_MODULE,
6987 .make_request = make_request,
6991 .error_handler = error,
6992 .hot_add_disk = raid5_add_disk,
6993 .hot_remove_disk= raid5_remove_disk,
6994 .spare_active = raid5_spare_active,
6995 .sync_request = sync_request,
6996 .resize = raid5_resize,
6998 .check_reshape = raid5_check_reshape,
6999 .start_reshape = raid5_start_reshape,
7000 .finish_reshape = raid5_finish_reshape,
7001 .quiesce = raid5_quiesce,
7002 .takeover = raid5_takeover,
7005 static struct md_personality raid4_personality =
7009 .owner = THIS_MODULE,
7010 .make_request = make_request,
7014 .error_handler = error,
7015 .hot_add_disk = raid5_add_disk,
7016 .hot_remove_disk= raid5_remove_disk,
7017 .spare_active = raid5_spare_active,
7018 .sync_request = sync_request,
7019 .resize = raid5_resize,
7021 .check_reshape = raid5_check_reshape,
7022 .start_reshape = raid5_start_reshape,
7023 .finish_reshape = raid5_finish_reshape,
7024 .quiesce = raid5_quiesce,
7025 .takeover = raid4_takeover,
7028 static int __init raid5_init(void)
7030 raid5_wq = alloc_workqueue("raid5wq",
7031 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7034 register_md_personality(&raid6_personality);
7035 register_md_personality(&raid5_personality);
7036 register_md_personality(&raid4_personality);
7040 static void raid5_exit(void)
7042 unregister_md_personality(&raid6_personality);
7043 unregister_md_personality(&raid5_personality);
7044 unregister_md_personality(&raid4_personality);
7045 destroy_workqueue(raid5_wq);
7048 module_init(raid5_init);
7049 module_exit(raid5_exit);
7050 MODULE_LICENSE("GPL");
7051 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7052 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7053 MODULE_ALIAS("md-raid5");
7054 MODULE_ALIAS("md-raid4");
7055 MODULE_ALIAS("md-level-5");
7056 MODULE_ALIAS("md-level-4");
7057 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7058 MODULE_ALIAS("md-raid6");
7059 MODULE_ALIAS("md-level-6");
7061 /* This used to be two separate modules, they were: */
7062 MODULE_ALIAS("raid5");
7063 MODULE_ALIAS("raid6");