2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
58 #include <trace/events/block.h>
65 #define cpu_to_group(cpu) cpu_to_node(cpu)
66 #define ANY_GROUP NUMA_NO_NODE
68 static bool devices_handle_discard_safely = false;
69 module_param(devices_handle_discard_safely, bool, 0644);
70 MODULE_PARM_DESC(devices_handle_discard_safely,
71 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
72 static struct workqueue_struct *raid5_wq;
77 #define NR_STRIPES 256
78 #define STRIPE_SIZE PAGE_SIZE
79 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
80 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
81 #define IO_THRESHOLD 1
82 #define BYPASS_THRESHOLD 1
83 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
84 #define HASH_MASK (NR_HASH - 1)
85 #define MAX_STRIPE_BATCH 8
87 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
89 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
90 return &conf->stripe_hashtbl[hash];
93 static inline int stripe_hash_locks_hash(sector_t sect)
95 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
98 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
100 spin_lock_irq(conf->hash_locks + hash);
101 spin_lock(&conf->device_lock);
104 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
106 spin_unlock(&conf->device_lock);
107 spin_unlock_irq(conf->hash_locks + hash);
110 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
114 spin_lock(conf->hash_locks);
115 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
116 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
117 spin_lock(&conf->device_lock);
120 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
123 spin_unlock(&conf->device_lock);
124 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
125 spin_unlock(conf->hash_locks + i - 1);
129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
130 * order without overlap. There may be several bio's per stripe+device, and
131 * a bio could span several devices.
132 * When walking this list for a particular stripe+device, we must never proceed
133 * beyond a bio that extends past this device, as the next bio might no longer
135 * This function is used to determine the 'next' bio in the list, given the sector
136 * of the current stripe+device
138 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
140 int sectors = bio_sectors(bio);
141 if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
148 * We maintain a biased count of active stripes in the bottom 16 bits of
149 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
151 static inline int raid5_bi_processed_stripes(struct bio *bio)
153 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
154 return (atomic_read(segments) >> 16) & 0xffff;
157 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
159 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
160 return atomic_sub_return(1, segments) & 0xffff;
163 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
165 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
166 atomic_inc(segments);
169 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
172 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
176 old = atomic_read(segments);
177 new = (old & 0xffff) | (cnt << 16);
178 } while (atomic_cmpxchg(segments, old, new) != old);
181 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
183 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
184 atomic_set(segments, cnt);
187 /* Find first data disk in a raid6 stripe */
188 static inline int raid6_d0(struct stripe_head *sh)
191 /* ddf always start from first device */
193 /* md starts just after Q block */
194 if (sh->qd_idx == sh->disks - 1)
197 return sh->qd_idx + 1;
199 static inline int raid6_next_disk(int disk, int raid_disks)
202 return (disk < raid_disks) ? disk : 0;
205 /* When walking through the disks in a raid5, starting at raid6_d0,
206 * We need to map each disk to a 'slot', where the data disks are slot
207 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
208 * is raid_disks-1. This help does that mapping.
210 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
211 int *count, int syndrome_disks)
217 if (idx == sh->pd_idx)
218 return syndrome_disks;
219 if (idx == sh->qd_idx)
220 return syndrome_disks + 1;
226 static void return_io(struct bio *return_bi)
228 struct bio *bi = return_bi;
231 return_bi = bi->bi_next;
233 bi->bi_iter.bi_size = 0;
234 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
241 static void print_raid5_conf (struct r5conf *conf);
243 static int stripe_operations_active(struct stripe_head *sh)
245 return sh->check_state || sh->reconstruct_state ||
246 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
247 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
250 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
252 struct r5conf *conf = sh->raid_conf;
253 struct r5worker_group *group;
255 int i, cpu = sh->cpu;
257 if (!cpu_online(cpu)) {
258 cpu = cpumask_any(cpu_online_mask);
262 if (list_empty(&sh->lru)) {
263 struct r5worker_group *group;
264 group = conf->worker_groups + cpu_to_group(cpu);
265 list_add_tail(&sh->lru, &group->handle_list);
266 group->stripes_cnt++;
270 if (conf->worker_cnt_per_group == 0) {
271 md_wakeup_thread(conf->mddev->thread);
275 group = conf->worker_groups + cpu_to_group(sh->cpu);
277 group->workers[0].working = true;
278 /* at least one worker should run to avoid race */
279 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
281 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
282 /* wakeup more workers */
283 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
284 if (group->workers[i].working == false) {
285 group->workers[i].working = true;
286 queue_work_on(sh->cpu, raid5_wq,
287 &group->workers[i].work);
293 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
294 struct list_head *temp_inactive_list)
296 BUG_ON(!list_empty(&sh->lru));
297 BUG_ON(atomic_read(&conf->active_stripes)==0);
298 if (test_bit(STRIPE_HANDLE, &sh->state)) {
299 if (test_bit(STRIPE_DELAYED, &sh->state) &&
300 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
301 list_add_tail(&sh->lru, &conf->delayed_list);
302 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
303 sh->bm_seq - conf->seq_write > 0)
304 list_add_tail(&sh->lru, &conf->bitmap_list);
306 clear_bit(STRIPE_DELAYED, &sh->state);
307 clear_bit(STRIPE_BIT_DELAY, &sh->state);
308 if (conf->worker_cnt_per_group == 0) {
309 list_add_tail(&sh->lru, &conf->handle_list);
311 raid5_wakeup_stripe_thread(sh);
315 md_wakeup_thread(conf->mddev->thread);
317 BUG_ON(stripe_operations_active(sh));
318 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
319 if (atomic_dec_return(&conf->preread_active_stripes)
321 md_wakeup_thread(conf->mddev->thread);
322 atomic_dec(&conf->active_stripes);
323 if (!test_bit(STRIPE_EXPANDING, &sh->state))
324 list_add_tail(&sh->lru, temp_inactive_list);
328 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
329 struct list_head *temp_inactive_list)
331 if (atomic_dec_and_test(&sh->count))
332 do_release_stripe(conf, sh, temp_inactive_list);
336 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
338 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
339 * given time. Adding stripes only takes device lock, while deleting stripes
340 * only takes hash lock.
342 static void release_inactive_stripe_list(struct r5conf *conf,
343 struct list_head *temp_inactive_list,
347 bool do_wakeup = false;
350 if (hash == NR_STRIPE_HASH_LOCKS) {
351 size = NR_STRIPE_HASH_LOCKS;
352 hash = NR_STRIPE_HASH_LOCKS - 1;
356 struct list_head *list = &temp_inactive_list[size - 1];
359 * We don't hold any lock here yet, get_active_stripe() might
360 * remove stripes from the list
362 if (!list_empty_careful(list)) {
363 spin_lock_irqsave(conf->hash_locks + hash, flags);
364 if (list_empty(conf->inactive_list + hash) &&
366 atomic_dec(&conf->empty_inactive_list_nr);
367 list_splice_tail_init(list, conf->inactive_list + hash);
369 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
376 wake_up(&conf->wait_for_stripe);
377 if (conf->retry_read_aligned)
378 md_wakeup_thread(conf->mddev->thread);
382 /* should hold conf->device_lock already */
383 static int release_stripe_list(struct r5conf *conf,
384 struct list_head *temp_inactive_list)
386 struct stripe_head *sh;
388 struct llist_node *head;
390 head = llist_del_all(&conf->released_stripes);
391 head = llist_reverse_order(head);
395 sh = llist_entry(head, struct stripe_head, release_list);
396 head = llist_next(head);
397 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
399 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
401 * Don't worry the bit is set here, because if the bit is set
402 * again, the count is always > 1. This is true for
403 * STRIPE_ON_UNPLUG_LIST bit too.
405 hash = sh->hash_lock_index;
406 __release_stripe(conf, sh, &temp_inactive_list[hash]);
413 static void release_stripe(struct stripe_head *sh)
415 struct r5conf *conf = sh->raid_conf;
417 struct list_head list;
421 /* Avoid release_list until the last reference.
423 if (atomic_add_unless(&sh->count, -1, 1))
426 if (unlikely(!conf->mddev->thread) ||
427 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
429 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
431 md_wakeup_thread(conf->mddev->thread);
434 local_irq_save(flags);
435 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
436 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
437 INIT_LIST_HEAD(&list);
438 hash = sh->hash_lock_index;
439 do_release_stripe(conf, sh, &list);
440 spin_unlock(&conf->device_lock);
441 release_inactive_stripe_list(conf, &list, hash);
443 local_irq_restore(flags);
446 static inline void remove_hash(struct stripe_head *sh)
448 pr_debug("remove_hash(), stripe %llu\n",
449 (unsigned long long)sh->sector);
451 hlist_del_init(&sh->hash);
454 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
456 struct hlist_head *hp = stripe_hash(conf, sh->sector);
458 pr_debug("insert_hash(), stripe %llu\n",
459 (unsigned long long)sh->sector);
461 hlist_add_head(&sh->hash, hp);
464 /* find an idle stripe, make sure it is unhashed, and return it. */
465 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
467 struct stripe_head *sh = NULL;
468 struct list_head *first;
470 if (list_empty(conf->inactive_list + hash))
472 first = (conf->inactive_list + hash)->next;
473 sh = list_entry(first, struct stripe_head, lru);
474 list_del_init(first);
476 atomic_inc(&conf->active_stripes);
477 BUG_ON(hash != sh->hash_lock_index);
478 if (list_empty(conf->inactive_list + hash))
479 atomic_inc(&conf->empty_inactive_list_nr);
484 static void shrink_buffers(struct stripe_head *sh)
488 int num = sh->raid_conf->pool_size;
490 for (i = 0; i < num ; i++) {
491 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
495 sh->dev[i].page = NULL;
500 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
503 int num = sh->raid_conf->pool_size;
505 for (i = 0; i < num; i++) {
508 if (!(page = alloc_page(gfp))) {
511 sh->dev[i].page = page;
512 sh->dev[i].orig_page = page;
517 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
518 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
519 struct stripe_head *sh);
521 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
523 struct r5conf *conf = sh->raid_conf;
526 BUG_ON(atomic_read(&sh->count) != 0);
527 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
528 BUG_ON(stripe_operations_active(sh));
529 BUG_ON(sh->batch_head);
531 pr_debug("init_stripe called, stripe %llu\n",
532 (unsigned long long)sector);
534 seq = read_seqcount_begin(&conf->gen_lock);
535 sh->generation = conf->generation - previous;
536 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
538 stripe_set_idx(sector, conf, previous, sh);
541 for (i = sh->disks; i--; ) {
542 struct r5dev *dev = &sh->dev[i];
544 if (dev->toread || dev->read || dev->towrite || dev->written ||
545 test_bit(R5_LOCKED, &dev->flags)) {
546 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
547 (unsigned long long)sh->sector, i, dev->toread,
548 dev->read, dev->towrite, dev->written,
549 test_bit(R5_LOCKED, &dev->flags));
553 raid5_build_block(sh, i, previous);
555 if (read_seqcount_retry(&conf->gen_lock, seq))
557 sh->overwrite_disks = 0;
558 insert_hash(conf, sh);
559 sh->cpu = smp_processor_id();
560 set_bit(STRIPE_BATCH_READY, &sh->state);
563 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
566 struct stripe_head *sh;
568 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
569 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
570 if (sh->sector == sector && sh->generation == generation)
572 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
577 * Need to check if array has failed when deciding whether to:
579 * - remove non-faulty devices
582 * This determination is simple when no reshape is happening.
583 * However if there is a reshape, we need to carefully check
584 * both the before and after sections.
585 * This is because some failed devices may only affect one
586 * of the two sections, and some non-in_sync devices may
587 * be insync in the section most affected by failed devices.
589 static int calc_degraded(struct r5conf *conf)
591 int degraded, degraded2;
596 for (i = 0; i < conf->previous_raid_disks; i++) {
597 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
598 if (rdev && test_bit(Faulty, &rdev->flags))
599 rdev = rcu_dereference(conf->disks[i].replacement);
600 if (!rdev || test_bit(Faulty, &rdev->flags))
602 else if (test_bit(In_sync, &rdev->flags))
605 /* not in-sync or faulty.
606 * If the reshape increases the number of devices,
607 * this is being recovered by the reshape, so
608 * this 'previous' section is not in_sync.
609 * If the number of devices is being reduced however,
610 * the device can only be part of the array if
611 * we are reverting a reshape, so this section will
614 if (conf->raid_disks >= conf->previous_raid_disks)
618 if (conf->raid_disks == conf->previous_raid_disks)
622 for (i = 0; i < conf->raid_disks; i++) {
623 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
624 if (rdev && test_bit(Faulty, &rdev->flags))
625 rdev = rcu_dereference(conf->disks[i].replacement);
626 if (!rdev || test_bit(Faulty, &rdev->flags))
628 else if (test_bit(In_sync, &rdev->flags))
631 /* not in-sync or faulty.
632 * If reshape increases the number of devices, this
633 * section has already been recovered, else it
634 * almost certainly hasn't.
636 if (conf->raid_disks <= conf->previous_raid_disks)
640 if (degraded2 > degraded)
645 static int has_failed(struct r5conf *conf)
649 if (conf->mddev->reshape_position == MaxSector)
650 return conf->mddev->degraded > conf->max_degraded;
652 degraded = calc_degraded(conf);
653 if (degraded > conf->max_degraded)
658 static struct stripe_head *
659 get_active_stripe(struct r5conf *conf, sector_t sector,
660 int previous, int noblock, int noquiesce)
662 struct stripe_head *sh;
663 int hash = stripe_hash_locks_hash(sector);
665 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
667 spin_lock_irq(conf->hash_locks + hash);
670 wait_event_lock_irq(conf->wait_for_stripe,
671 conf->quiesce == 0 || noquiesce,
672 *(conf->hash_locks + hash));
673 sh = __find_stripe(conf, sector, conf->generation - previous);
675 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
676 sh = get_free_stripe(conf, hash);
677 if (!sh && llist_empty(&conf->released_stripes) &&
678 !test_bit(R5_DID_ALLOC, &conf->cache_state))
679 set_bit(R5_ALLOC_MORE,
682 if (noblock && sh == NULL)
685 set_bit(R5_INACTIVE_BLOCKED,
688 conf->wait_for_stripe,
689 !list_empty(conf->inactive_list + hash) &&
690 (atomic_read(&conf->active_stripes)
691 < (conf->max_nr_stripes * 3 / 4)
692 || !test_bit(R5_INACTIVE_BLOCKED,
693 &conf->cache_state)),
694 *(conf->hash_locks + hash));
695 clear_bit(R5_INACTIVE_BLOCKED,
698 init_stripe(sh, sector, previous);
699 atomic_inc(&sh->count);
701 } else if (!atomic_inc_not_zero(&sh->count)) {
702 spin_lock(&conf->device_lock);
703 if (!atomic_read(&sh->count)) {
704 if (!test_bit(STRIPE_HANDLE, &sh->state))
705 atomic_inc(&conf->active_stripes);
706 BUG_ON(list_empty(&sh->lru) &&
707 !test_bit(STRIPE_EXPANDING, &sh->state));
708 list_del_init(&sh->lru);
710 sh->group->stripes_cnt--;
714 atomic_inc(&sh->count);
715 spin_unlock(&conf->device_lock);
717 } while (sh == NULL);
719 spin_unlock_irq(conf->hash_locks + hash);
723 static bool is_full_stripe_write(struct stripe_head *sh)
725 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
726 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
729 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
733 spin_lock(&sh2->stripe_lock);
734 spin_lock_nested(&sh1->stripe_lock, 1);
736 spin_lock(&sh1->stripe_lock);
737 spin_lock_nested(&sh2->stripe_lock, 1);
741 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
743 spin_unlock(&sh1->stripe_lock);
744 spin_unlock(&sh2->stripe_lock);
748 /* Only freshly new full stripe normal write stripe can be added to a batch list */
749 static bool stripe_can_batch(struct stripe_head *sh)
751 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
752 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
753 is_full_stripe_write(sh);
756 /* we only do back search */
757 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
759 struct stripe_head *head;
760 sector_t head_sector, tmp_sec;
764 if (!stripe_can_batch(sh))
766 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
767 tmp_sec = sh->sector;
768 if (!sector_div(tmp_sec, conf->chunk_sectors))
770 head_sector = sh->sector - STRIPE_SECTORS;
772 hash = stripe_hash_locks_hash(head_sector);
773 spin_lock_irq(conf->hash_locks + hash);
774 head = __find_stripe(conf, head_sector, conf->generation);
775 if (head && !atomic_inc_not_zero(&head->count)) {
776 spin_lock(&conf->device_lock);
777 if (!atomic_read(&head->count)) {
778 if (!test_bit(STRIPE_HANDLE, &head->state))
779 atomic_inc(&conf->active_stripes);
780 BUG_ON(list_empty(&head->lru) &&
781 !test_bit(STRIPE_EXPANDING, &head->state));
782 list_del_init(&head->lru);
784 head->group->stripes_cnt--;
788 atomic_inc(&head->count);
789 spin_unlock(&conf->device_lock);
791 spin_unlock_irq(conf->hash_locks + hash);
795 if (!stripe_can_batch(head))
798 lock_two_stripes(head, sh);
799 /* clear_batch_ready clear the flag */
800 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
807 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
809 if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
812 if (head->batch_head) {
813 spin_lock(&head->batch_head->batch_lock);
814 /* This batch list is already running */
815 if (!stripe_can_batch(head)) {
816 spin_unlock(&head->batch_head->batch_lock);
821 * at this point, head's BATCH_READY could be cleared, but we
822 * can still add the stripe to batch list
824 list_add(&sh->batch_list, &head->batch_list);
825 spin_unlock(&head->batch_head->batch_lock);
827 sh->batch_head = head->batch_head;
829 head->batch_head = head;
830 sh->batch_head = head->batch_head;
831 spin_lock(&head->batch_lock);
832 list_add_tail(&sh->batch_list, &head->batch_list);
833 spin_unlock(&head->batch_lock);
836 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
837 if (atomic_dec_return(&conf->preread_active_stripes)
839 md_wakeup_thread(conf->mddev->thread);
841 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
842 int seq = sh->bm_seq;
843 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
844 sh->batch_head->bm_seq > seq)
845 seq = sh->batch_head->bm_seq;
846 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
847 sh->batch_head->bm_seq = seq;
850 atomic_inc(&sh->count);
852 unlock_two_stripes(head, sh);
854 release_stripe(head);
857 /* Determine if 'data_offset' or 'new_data_offset' should be used
858 * in this stripe_head.
860 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
862 sector_t progress = conf->reshape_progress;
863 /* Need a memory barrier to make sure we see the value
864 * of conf->generation, or ->data_offset that was set before
865 * reshape_progress was updated.
868 if (progress == MaxSector)
870 if (sh->generation == conf->generation - 1)
872 /* We are in a reshape, and this is a new-generation stripe,
873 * so use new_data_offset.
879 raid5_end_read_request(struct bio *bi, int error);
881 raid5_end_write_request(struct bio *bi, int error);
883 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
885 struct r5conf *conf = sh->raid_conf;
886 int i, disks = sh->disks;
887 struct stripe_head *head_sh = sh;
891 for (i = disks; i--; ) {
893 int replace_only = 0;
894 struct bio *bi, *rbi;
895 struct md_rdev *rdev, *rrdev = NULL;
898 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
899 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
903 if (test_bit(R5_Discard, &sh->dev[i].flags))
905 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
907 else if (test_and_clear_bit(R5_WantReplace,
908 &sh->dev[i].flags)) {
913 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
917 bi = &sh->dev[i].req;
918 rbi = &sh->dev[i].rreq; /* For writing to replacement */
921 rrdev = rcu_dereference(conf->disks[i].replacement);
922 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
923 rdev = rcu_dereference(conf->disks[i].rdev);
932 /* We raced and saw duplicates */
935 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
940 if (rdev && test_bit(Faulty, &rdev->flags))
943 atomic_inc(&rdev->nr_pending);
944 if (rrdev && test_bit(Faulty, &rrdev->flags))
947 atomic_inc(&rrdev->nr_pending);
950 /* We have already checked bad blocks for reads. Now
951 * need to check for writes. We never accept write errors
952 * on the replacement, so we don't to check rrdev.
954 while ((rw & WRITE) && rdev &&
955 test_bit(WriteErrorSeen, &rdev->flags)) {
958 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
959 &first_bad, &bad_sectors);
964 set_bit(BlockedBadBlocks, &rdev->flags);
965 if (!conf->mddev->external &&
966 conf->mddev->flags) {
967 /* It is very unlikely, but we might
968 * still need to write out the
969 * bad block log - better give it
971 md_check_recovery(conf->mddev);
974 * Because md_wait_for_blocked_rdev
975 * will dec nr_pending, we must
976 * increment it first.
978 atomic_inc(&rdev->nr_pending);
979 md_wait_for_blocked_rdev(rdev, conf->mddev);
981 /* Acknowledged bad block - skip the write */
982 rdev_dec_pending(rdev, conf->mddev);
988 if (s->syncing || s->expanding || s->expanded
990 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
992 set_bit(STRIPE_IO_STARTED, &sh->state);
995 bi->bi_bdev = rdev->bdev;
997 bi->bi_end_io = (rw & WRITE)
998 ? raid5_end_write_request
999 : raid5_end_read_request;
1000 bi->bi_private = sh;
1002 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1003 __func__, (unsigned long long)sh->sector,
1005 atomic_inc(&sh->count);
1007 atomic_inc(&head_sh->count);
1008 if (use_new_offset(conf, sh))
1009 bi->bi_iter.bi_sector = (sh->sector
1010 + rdev->new_data_offset);
1012 bi->bi_iter.bi_sector = (sh->sector
1013 + rdev->data_offset);
1014 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1015 bi->bi_rw |= REQ_NOMERGE;
1017 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1018 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1019 sh->dev[i].vec.bv_page = sh->dev[i].page;
1021 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1022 bi->bi_io_vec[0].bv_offset = 0;
1023 bi->bi_iter.bi_size = STRIPE_SIZE;
1025 * If this is discard request, set bi_vcnt 0. We don't
1026 * want to confuse SCSI because SCSI will replace payload
1028 if (rw & REQ_DISCARD)
1031 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1033 if (conf->mddev->gendisk)
1034 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1035 bi, disk_devt(conf->mddev->gendisk),
1037 generic_make_request(bi);
1040 if (s->syncing || s->expanding || s->expanded
1042 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1044 set_bit(STRIPE_IO_STARTED, &sh->state);
1047 rbi->bi_bdev = rrdev->bdev;
1049 BUG_ON(!(rw & WRITE));
1050 rbi->bi_end_io = raid5_end_write_request;
1051 rbi->bi_private = sh;
1053 pr_debug("%s: for %llu schedule op %ld on "
1054 "replacement disc %d\n",
1055 __func__, (unsigned long long)sh->sector,
1057 atomic_inc(&sh->count);
1059 atomic_inc(&head_sh->count);
1060 if (use_new_offset(conf, sh))
1061 rbi->bi_iter.bi_sector = (sh->sector
1062 + rrdev->new_data_offset);
1064 rbi->bi_iter.bi_sector = (sh->sector
1065 + rrdev->data_offset);
1066 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1067 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1068 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1070 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1071 rbi->bi_io_vec[0].bv_offset = 0;
1072 rbi->bi_iter.bi_size = STRIPE_SIZE;
1074 * If this is discard request, set bi_vcnt 0. We don't
1075 * want to confuse SCSI because SCSI will replace payload
1077 if (rw & REQ_DISCARD)
1079 if (conf->mddev->gendisk)
1080 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1081 rbi, disk_devt(conf->mddev->gendisk),
1083 generic_make_request(rbi);
1085 if (!rdev && !rrdev) {
1087 set_bit(STRIPE_DEGRADED, &sh->state);
1088 pr_debug("skip op %ld on disc %d for sector %llu\n",
1089 bi->bi_rw, i, (unsigned long long)sh->sector);
1090 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1091 set_bit(STRIPE_HANDLE, &sh->state);
1094 if (!head_sh->batch_head)
1096 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1103 static struct dma_async_tx_descriptor *
1104 async_copy_data(int frombio, struct bio *bio, struct page **page,
1105 sector_t sector, struct dma_async_tx_descriptor *tx,
1106 struct stripe_head *sh)
1109 struct bvec_iter iter;
1110 struct page *bio_page;
1112 struct async_submit_ctl submit;
1113 enum async_tx_flags flags = 0;
1115 if (bio->bi_iter.bi_sector >= sector)
1116 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1118 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1121 flags |= ASYNC_TX_FENCE;
1122 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1124 bio_for_each_segment(bvl, bio, iter) {
1125 int len = bvl.bv_len;
1129 if (page_offset < 0) {
1130 b_offset = -page_offset;
1131 page_offset += b_offset;
1135 if (len > 0 && page_offset + len > STRIPE_SIZE)
1136 clen = STRIPE_SIZE - page_offset;
1141 b_offset += bvl.bv_offset;
1142 bio_page = bvl.bv_page;
1144 if (sh->raid_conf->skip_copy &&
1145 b_offset == 0 && page_offset == 0 &&
1146 clen == STRIPE_SIZE)
1149 tx = async_memcpy(*page, bio_page, page_offset,
1150 b_offset, clen, &submit);
1152 tx = async_memcpy(bio_page, *page, b_offset,
1153 page_offset, clen, &submit);
1155 /* chain the operations */
1156 submit.depend_tx = tx;
1158 if (clen < len) /* hit end of page */
1166 static void ops_complete_biofill(void *stripe_head_ref)
1168 struct stripe_head *sh = stripe_head_ref;
1169 struct bio *return_bi = NULL;
1172 pr_debug("%s: stripe %llu\n", __func__,
1173 (unsigned long long)sh->sector);
1175 /* clear completed biofills */
1176 for (i = sh->disks; i--; ) {
1177 struct r5dev *dev = &sh->dev[i];
1179 /* acknowledge completion of a biofill operation */
1180 /* and check if we need to reply to a read request,
1181 * new R5_Wantfill requests are held off until
1182 * !STRIPE_BIOFILL_RUN
1184 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1185 struct bio *rbi, *rbi2;
1190 while (rbi && rbi->bi_iter.bi_sector <
1191 dev->sector + STRIPE_SECTORS) {
1192 rbi2 = r5_next_bio(rbi, dev->sector);
1193 if (!raid5_dec_bi_active_stripes(rbi)) {
1194 rbi->bi_next = return_bi;
1201 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1203 return_io(return_bi);
1205 set_bit(STRIPE_HANDLE, &sh->state);
1209 static void ops_run_biofill(struct stripe_head *sh)
1211 struct dma_async_tx_descriptor *tx = NULL;
1212 struct async_submit_ctl submit;
1215 BUG_ON(sh->batch_head);
1216 pr_debug("%s: stripe %llu\n", __func__,
1217 (unsigned long long)sh->sector);
1219 for (i = sh->disks; i--; ) {
1220 struct r5dev *dev = &sh->dev[i];
1221 if (test_bit(R5_Wantfill, &dev->flags)) {
1223 spin_lock_irq(&sh->stripe_lock);
1224 dev->read = rbi = dev->toread;
1226 spin_unlock_irq(&sh->stripe_lock);
1227 while (rbi && rbi->bi_iter.bi_sector <
1228 dev->sector + STRIPE_SECTORS) {
1229 tx = async_copy_data(0, rbi, &dev->page,
1230 dev->sector, tx, sh);
1231 rbi = r5_next_bio(rbi, dev->sector);
1236 atomic_inc(&sh->count);
1237 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1238 async_trigger_callback(&submit);
1241 static void mark_target_uptodate(struct stripe_head *sh, int target)
1248 tgt = &sh->dev[target];
1249 set_bit(R5_UPTODATE, &tgt->flags);
1250 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1251 clear_bit(R5_Wantcompute, &tgt->flags);
1254 static void ops_complete_compute(void *stripe_head_ref)
1256 struct stripe_head *sh = stripe_head_ref;
1258 pr_debug("%s: stripe %llu\n", __func__,
1259 (unsigned long long)sh->sector);
1261 /* mark the computed target(s) as uptodate */
1262 mark_target_uptodate(sh, sh->ops.target);
1263 mark_target_uptodate(sh, sh->ops.target2);
1265 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1266 if (sh->check_state == check_state_compute_run)
1267 sh->check_state = check_state_compute_result;
1268 set_bit(STRIPE_HANDLE, &sh->state);
1272 /* return a pointer to the address conversion region of the scribble buffer */
1273 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1274 struct raid5_percpu *percpu, int i)
1278 addr = flex_array_get(percpu->scribble, i);
1279 return addr + sizeof(struct page *) * (sh->disks + 2);
1282 /* return a pointer to the address conversion region of the scribble buffer */
1283 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1287 addr = flex_array_get(percpu->scribble, i);
1291 static struct dma_async_tx_descriptor *
1292 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1294 int disks = sh->disks;
1295 struct page **xor_srcs = to_addr_page(percpu, 0);
1296 int target = sh->ops.target;
1297 struct r5dev *tgt = &sh->dev[target];
1298 struct page *xor_dest = tgt->page;
1300 struct dma_async_tx_descriptor *tx;
1301 struct async_submit_ctl submit;
1304 BUG_ON(sh->batch_head);
1306 pr_debug("%s: stripe %llu block: %d\n",
1307 __func__, (unsigned long long)sh->sector, target);
1308 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1310 for (i = disks; i--; )
1312 xor_srcs[count++] = sh->dev[i].page;
1314 atomic_inc(&sh->count);
1316 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1317 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1318 if (unlikely(count == 1))
1319 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1321 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1326 /* set_syndrome_sources - populate source buffers for gen_syndrome
1327 * @srcs - (struct page *) array of size sh->disks
1328 * @sh - stripe_head to parse
1330 * Populates srcs in proper layout order for the stripe and returns the
1331 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1332 * destination buffer is recorded in srcs[count] and the Q destination
1333 * is recorded in srcs[count+1]].
1335 static int set_syndrome_sources(struct page **srcs,
1336 struct stripe_head *sh,
1339 int disks = sh->disks;
1340 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1341 int d0_idx = raid6_d0(sh);
1345 for (i = 0; i < disks; i++)
1351 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1352 struct r5dev *dev = &sh->dev[i];
1354 if (i == sh->qd_idx || i == sh->pd_idx ||
1355 (srctype == SYNDROME_SRC_ALL) ||
1356 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1357 test_bit(R5_Wantdrain, &dev->flags)) ||
1358 (srctype == SYNDROME_SRC_WRITTEN &&
1360 srcs[slot] = sh->dev[i].page;
1361 i = raid6_next_disk(i, disks);
1362 } while (i != d0_idx);
1364 return syndrome_disks;
1367 static struct dma_async_tx_descriptor *
1368 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1370 int disks = sh->disks;
1371 struct page **blocks = to_addr_page(percpu, 0);
1373 int qd_idx = sh->qd_idx;
1374 struct dma_async_tx_descriptor *tx;
1375 struct async_submit_ctl submit;
1381 BUG_ON(sh->batch_head);
1382 if (sh->ops.target < 0)
1383 target = sh->ops.target2;
1384 else if (sh->ops.target2 < 0)
1385 target = sh->ops.target;
1387 /* we should only have one valid target */
1390 pr_debug("%s: stripe %llu block: %d\n",
1391 __func__, (unsigned long long)sh->sector, target);
1393 tgt = &sh->dev[target];
1394 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1397 atomic_inc(&sh->count);
1399 if (target == qd_idx) {
1400 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1401 blocks[count] = NULL; /* regenerating p is not necessary */
1402 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1403 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1404 ops_complete_compute, sh,
1405 to_addr_conv(sh, percpu, 0));
1406 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1408 /* Compute any data- or p-drive using XOR */
1410 for (i = disks; i-- ; ) {
1411 if (i == target || i == qd_idx)
1413 blocks[count++] = sh->dev[i].page;
1416 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1417 NULL, ops_complete_compute, sh,
1418 to_addr_conv(sh, percpu, 0));
1419 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1425 static struct dma_async_tx_descriptor *
1426 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1428 int i, count, disks = sh->disks;
1429 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1430 int d0_idx = raid6_d0(sh);
1431 int faila = -1, failb = -1;
1432 int target = sh->ops.target;
1433 int target2 = sh->ops.target2;
1434 struct r5dev *tgt = &sh->dev[target];
1435 struct r5dev *tgt2 = &sh->dev[target2];
1436 struct dma_async_tx_descriptor *tx;
1437 struct page **blocks = to_addr_page(percpu, 0);
1438 struct async_submit_ctl submit;
1440 BUG_ON(sh->batch_head);
1441 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1442 __func__, (unsigned long long)sh->sector, target, target2);
1443 BUG_ON(target < 0 || target2 < 0);
1444 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1445 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1447 /* we need to open-code set_syndrome_sources to handle the
1448 * slot number conversion for 'faila' and 'failb'
1450 for (i = 0; i < disks ; i++)
1455 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1457 blocks[slot] = sh->dev[i].page;
1463 i = raid6_next_disk(i, disks);
1464 } while (i != d0_idx);
1466 BUG_ON(faila == failb);
1469 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1470 __func__, (unsigned long long)sh->sector, faila, failb);
1472 atomic_inc(&sh->count);
1474 if (failb == syndrome_disks+1) {
1475 /* Q disk is one of the missing disks */
1476 if (faila == syndrome_disks) {
1477 /* Missing P+Q, just recompute */
1478 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1479 ops_complete_compute, sh,
1480 to_addr_conv(sh, percpu, 0));
1481 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1482 STRIPE_SIZE, &submit);
1486 int qd_idx = sh->qd_idx;
1488 /* Missing D+Q: recompute D from P, then recompute Q */
1489 if (target == qd_idx)
1490 data_target = target2;
1492 data_target = target;
1495 for (i = disks; i-- ; ) {
1496 if (i == data_target || i == qd_idx)
1498 blocks[count++] = sh->dev[i].page;
1500 dest = sh->dev[data_target].page;
1501 init_async_submit(&submit,
1502 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1504 to_addr_conv(sh, percpu, 0));
1505 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1508 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1509 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1510 ops_complete_compute, sh,
1511 to_addr_conv(sh, percpu, 0));
1512 return async_gen_syndrome(blocks, 0, count+2,
1513 STRIPE_SIZE, &submit);
1516 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1517 ops_complete_compute, sh,
1518 to_addr_conv(sh, percpu, 0));
1519 if (failb == syndrome_disks) {
1520 /* We're missing D+P. */
1521 return async_raid6_datap_recov(syndrome_disks+2,
1525 /* We're missing D+D. */
1526 return async_raid6_2data_recov(syndrome_disks+2,
1527 STRIPE_SIZE, faila, failb,
1533 static void ops_complete_prexor(void *stripe_head_ref)
1535 struct stripe_head *sh = stripe_head_ref;
1537 pr_debug("%s: stripe %llu\n", __func__,
1538 (unsigned long long)sh->sector);
1541 static struct dma_async_tx_descriptor *
1542 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1543 struct dma_async_tx_descriptor *tx)
1545 int disks = sh->disks;
1546 struct page **xor_srcs = to_addr_page(percpu, 0);
1547 int count = 0, pd_idx = sh->pd_idx, i;
1548 struct async_submit_ctl submit;
1550 /* existing parity data subtracted */
1551 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1553 BUG_ON(sh->batch_head);
1554 pr_debug("%s: stripe %llu\n", __func__,
1555 (unsigned long long)sh->sector);
1557 for (i = disks; i--; ) {
1558 struct r5dev *dev = &sh->dev[i];
1559 /* Only process blocks that are known to be uptodate */
1560 if (test_bit(R5_Wantdrain, &dev->flags))
1561 xor_srcs[count++] = dev->page;
1564 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1565 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1566 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1571 static struct dma_async_tx_descriptor *
1572 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1573 struct dma_async_tx_descriptor *tx)
1575 struct page **blocks = to_addr_page(percpu, 0);
1577 struct async_submit_ctl submit;
1579 pr_debug("%s: stripe %llu\n", __func__,
1580 (unsigned long long)sh->sector);
1582 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1584 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1585 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1586 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1591 static struct dma_async_tx_descriptor *
1592 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1594 int disks = sh->disks;
1596 struct stripe_head *head_sh = sh;
1598 pr_debug("%s: stripe %llu\n", __func__,
1599 (unsigned long long)sh->sector);
1601 for (i = disks; i--; ) {
1606 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1611 spin_lock_irq(&sh->stripe_lock);
1612 chosen = dev->towrite;
1613 dev->towrite = NULL;
1614 sh->overwrite_disks = 0;
1615 BUG_ON(dev->written);
1616 wbi = dev->written = chosen;
1617 spin_unlock_irq(&sh->stripe_lock);
1618 WARN_ON(dev->page != dev->orig_page);
1620 while (wbi && wbi->bi_iter.bi_sector <
1621 dev->sector + STRIPE_SECTORS) {
1622 if (wbi->bi_rw & REQ_FUA)
1623 set_bit(R5_WantFUA, &dev->flags);
1624 if (wbi->bi_rw & REQ_SYNC)
1625 set_bit(R5_SyncIO, &dev->flags);
1626 if (wbi->bi_rw & REQ_DISCARD)
1627 set_bit(R5_Discard, &dev->flags);
1629 tx = async_copy_data(1, wbi, &dev->page,
1630 dev->sector, tx, sh);
1631 if (dev->page != dev->orig_page) {
1632 set_bit(R5_SkipCopy, &dev->flags);
1633 clear_bit(R5_UPTODATE, &dev->flags);
1634 clear_bit(R5_OVERWRITE, &dev->flags);
1637 wbi = r5_next_bio(wbi, dev->sector);
1640 if (head_sh->batch_head) {
1641 sh = list_first_entry(&sh->batch_list,
1654 static void ops_complete_reconstruct(void *stripe_head_ref)
1656 struct stripe_head *sh = stripe_head_ref;
1657 int disks = sh->disks;
1658 int pd_idx = sh->pd_idx;
1659 int qd_idx = sh->qd_idx;
1661 bool fua = false, sync = false, discard = false;
1663 pr_debug("%s: stripe %llu\n", __func__,
1664 (unsigned long long)sh->sector);
1666 for (i = disks; i--; ) {
1667 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1668 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1669 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1672 for (i = disks; i--; ) {
1673 struct r5dev *dev = &sh->dev[i];
1675 if (dev->written || i == pd_idx || i == qd_idx) {
1676 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1677 set_bit(R5_UPTODATE, &dev->flags);
1679 set_bit(R5_WantFUA, &dev->flags);
1681 set_bit(R5_SyncIO, &dev->flags);
1685 if (sh->reconstruct_state == reconstruct_state_drain_run)
1686 sh->reconstruct_state = reconstruct_state_drain_result;
1687 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1688 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1690 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1691 sh->reconstruct_state = reconstruct_state_result;
1694 set_bit(STRIPE_HANDLE, &sh->state);
1699 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1700 struct dma_async_tx_descriptor *tx)
1702 int disks = sh->disks;
1703 struct page **xor_srcs;
1704 struct async_submit_ctl submit;
1705 int count, pd_idx = sh->pd_idx, i;
1706 struct page *xor_dest;
1708 unsigned long flags;
1710 struct stripe_head *head_sh = sh;
1713 pr_debug("%s: stripe %llu\n", __func__,
1714 (unsigned long long)sh->sector);
1716 for (i = 0; i < sh->disks; i++) {
1719 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1722 if (i >= sh->disks) {
1723 atomic_inc(&sh->count);
1724 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1725 ops_complete_reconstruct(sh);
1730 xor_srcs = to_addr_page(percpu, j);
1731 /* check if prexor is active which means only process blocks
1732 * that are part of a read-modify-write (written)
1734 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1736 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1737 for (i = disks; i--; ) {
1738 struct r5dev *dev = &sh->dev[i];
1739 if (head_sh->dev[i].written)
1740 xor_srcs[count++] = dev->page;
1743 xor_dest = sh->dev[pd_idx].page;
1744 for (i = disks; i--; ) {
1745 struct r5dev *dev = &sh->dev[i];
1747 xor_srcs[count++] = dev->page;
1751 /* 1/ if we prexor'd then the dest is reused as a source
1752 * 2/ if we did not prexor then we are redoing the parity
1753 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1754 * for the synchronous xor case
1756 last_stripe = !head_sh->batch_head ||
1757 list_first_entry(&sh->batch_list,
1758 struct stripe_head, batch_list) == head_sh;
1760 flags = ASYNC_TX_ACK |
1761 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1763 atomic_inc(&head_sh->count);
1764 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1765 to_addr_conv(sh, percpu, j));
1767 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1768 init_async_submit(&submit, flags, tx, NULL, NULL,
1769 to_addr_conv(sh, percpu, j));
1772 if (unlikely(count == 1))
1773 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1775 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1778 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1785 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1786 struct dma_async_tx_descriptor *tx)
1788 struct async_submit_ctl submit;
1789 struct page **blocks;
1790 int count, i, j = 0;
1791 struct stripe_head *head_sh = sh;
1794 unsigned long txflags;
1796 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1798 for (i = 0; i < sh->disks; i++) {
1799 if (sh->pd_idx == i || sh->qd_idx == i)
1801 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1804 if (i >= sh->disks) {
1805 atomic_inc(&sh->count);
1806 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1807 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1808 ops_complete_reconstruct(sh);
1813 blocks = to_addr_page(percpu, j);
1815 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1816 synflags = SYNDROME_SRC_WRITTEN;
1817 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1819 synflags = SYNDROME_SRC_ALL;
1820 txflags = ASYNC_TX_ACK;
1823 count = set_syndrome_sources(blocks, sh, synflags);
1824 last_stripe = !head_sh->batch_head ||
1825 list_first_entry(&sh->batch_list,
1826 struct stripe_head, batch_list) == head_sh;
1829 atomic_inc(&head_sh->count);
1830 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1831 head_sh, to_addr_conv(sh, percpu, j));
1833 init_async_submit(&submit, 0, tx, NULL, NULL,
1834 to_addr_conv(sh, percpu, j));
1835 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1838 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1844 static void ops_complete_check(void *stripe_head_ref)
1846 struct stripe_head *sh = stripe_head_ref;
1848 pr_debug("%s: stripe %llu\n", __func__,
1849 (unsigned long long)sh->sector);
1851 sh->check_state = check_state_check_result;
1852 set_bit(STRIPE_HANDLE, &sh->state);
1856 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1858 int disks = sh->disks;
1859 int pd_idx = sh->pd_idx;
1860 int qd_idx = sh->qd_idx;
1861 struct page *xor_dest;
1862 struct page **xor_srcs = to_addr_page(percpu, 0);
1863 struct dma_async_tx_descriptor *tx;
1864 struct async_submit_ctl submit;
1868 pr_debug("%s: stripe %llu\n", __func__,
1869 (unsigned long long)sh->sector);
1871 BUG_ON(sh->batch_head);
1873 xor_dest = sh->dev[pd_idx].page;
1874 xor_srcs[count++] = xor_dest;
1875 for (i = disks; i--; ) {
1876 if (i == pd_idx || i == qd_idx)
1878 xor_srcs[count++] = sh->dev[i].page;
1881 init_async_submit(&submit, 0, NULL, NULL, NULL,
1882 to_addr_conv(sh, percpu, 0));
1883 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1884 &sh->ops.zero_sum_result, &submit);
1886 atomic_inc(&sh->count);
1887 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1888 tx = async_trigger_callback(&submit);
1891 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1893 struct page **srcs = to_addr_page(percpu, 0);
1894 struct async_submit_ctl submit;
1897 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1898 (unsigned long long)sh->sector, checkp);
1900 BUG_ON(sh->batch_head);
1901 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1905 atomic_inc(&sh->count);
1906 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1907 sh, to_addr_conv(sh, percpu, 0));
1908 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1909 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1912 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1914 int overlap_clear = 0, i, disks = sh->disks;
1915 struct dma_async_tx_descriptor *tx = NULL;
1916 struct r5conf *conf = sh->raid_conf;
1917 int level = conf->level;
1918 struct raid5_percpu *percpu;
1921 cpu = get_cpu_light();
1922 percpu = per_cpu_ptr(conf->percpu, cpu);
1923 spin_lock(&percpu->lock);
1924 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1925 ops_run_biofill(sh);
1929 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1931 tx = ops_run_compute5(sh, percpu);
1933 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1934 tx = ops_run_compute6_1(sh, percpu);
1936 tx = ops_run_compute6_2(sh, percpu);
1938 /* terminate the chain if reconstruct is not set to be run */
1939 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1943 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1945 tx = ops_run_prexor5(sh, percpu, tx);
1947 tx = ops_run_prexor6(sh, percpu, tx);
1950 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1951 tx = ops_run_biodrain(sh, tx);
1955 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1957 ops_run_reconstruct5(sh, percpu, tx);
1959 ops_run_reconstruct6(sh, percpu, tx);
1962 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1963 if (sh->check_state == check_state_run)
1964 ops_run_check_p(sh, percpu);
1965 else if (sh->check_state == check_state_run_q)
1966 ops_run_check_pq(sh, percpu, 0);
1967 else if (sh->check_state == check_state_run_pq)
1968 ops_run_check_pq(sh, percpu, 1);
1973 if (overlap_clear && !sh->batch_head)
1974 for (i = disks; i--; ) {
1975 struct r5dev *dev = &sh->dev[i];
1976 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1977 wake_up(&sh->raid_conf->wait_for_overlap);
1979 spin_unlock(&percpu->lock);
1983 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1985 struct stripe_head *sh;
1987 sh = kmem_cache_zalloc(sc, gfp);
1989 spin_lock_init(&sh->stripe_lock);
1990 spin_lock_init(&sh->batch_lock);
1991 INIT_LIST_HEAD(&sh->batch_list);
1992 INIT_LIST_HEAD(&sh->lru);
1993 atomic_set(&sh->count, 1);
1997 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
1999 struct stripe_head *sh;
2001 sh = alloc_stripe(conf->slab_cache, gfp);
2005 sh->raid_conf = conf;
2007 if (grow_buffers(sh, gfp)) {
2009 kmem_cache_free(conf->slab_cache, sh);
2012 sh->hash_lock_index =
2013 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2014 /* we just created an active stripe so... */
2015 atomic_inc(&conf->active_stripes);
2018 conf->max_nr_stripes++;
2022 static int grow_stripes(struct r5conf *conf, int num)
2024 struct kmem_cache *sc;
2025 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2027 if (conf->mddev->gendisk)
2028 sprintf(conf->cache_name[0],
2029 "raid%d-%s", conf->level, mdname(conf->mddev));
2031 sprintf(conf->cache_name[0],
2032 "raid%d-%p", conf->level, conf->mddev);
2033 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2035 conf->active_name = 0;
2036 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2037 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2041 conf->slab_cache = sc;
2042 conf->pool_size = devs;
2044 if (!grow_one_stripe(conf, GFP_KERNEL))
2051 * scribble_len - return the required size of the scribble region
2052 * @num - total number of disks in the array
2054 * The size must be enough to contain:
2055 * 1/ a struct page pointer for each device in the array +2
2056 * 2/ room to convert each entry in (1) to its corresponding dma
2057 * (dma_map_page()) or page (page_address()) address.
2059 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2060 * calculate over all devices (not just the data blocks), using zeros in place
2061 * of the P and Q blocks.
2063 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2065 struct flex_array *ret;
2068 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2069 ret = flex_array_alloc(len, cnt, flags);
2072 /* always prealloc all elements, so no locking is required */
2073 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2074 flex_array_free(ret);
2080 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2085 mddev_suspend(conf->mddev);
2087 for_each_present_cpu(cpu) {
2088 struct raid5_percpu *percpu;
2089 struct flex_array *scribble;
2091 percpu = per_cpu_ptr(conf->percpu, cpu);
2092 scribble = scribble_alloc(new_disks,
2093 new_sectors / STRIPE_SECTORS,
2097 flex_array_free(percpu->scribble);
2098 percpu->scribble = scribble;
2105 mddev_resume(conf->mddev);
2109 static int resize_stripes(struct r5conf *conf, int newsize)
2111 /* Make all the stripes able to hold 'newsize' devices.
2112 * New slots in each stripe get 'page' set to a new page.
2114 * This happens in stages:
2115 * 1/ create a new kmem_cache and allocate the required number of
2117 * 2/ gather all the old stripe_heads and transfer the pages across
2118 * to the new stripe_heads. This will have the side effect of
2119 * freezing the array as once all stripe_heads have been collected,
2120 * no IO will be possible. Old stripe heads are freed once their
2121 * pages have been transferred over, and the old kmem_cache is
2122 * freed when all stripes are done.
2123 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2124 * we simple return a failre status - no need to clean anything up.
2125 * 4/ allocate new pages for the new slots in the new stripe_heads.
2126 * If this fails, we don't bother trying the shrink the
2127 * stripe_heads down again, we just leave them as they are.
2128 * As each stripe_head is processed the new one is released into
2131 * Once step2 is started, we cannot afford to wait for a write,
2132 * so we use GFP_NOIO allocations.
2134 struct stripe_head *osh, *nsh;
2135 LIST_HEAD(newstripes);
2136 struct disk_info *ndisks;
2138 struct kmem_cache *sc;
2142 if (newsize <= conf->pool_size)
2143 return 0; /* never bother to shrink */
2145 err = md_allow_write(conf->mddev);
2150 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2151 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2156 /* Need to ensure auto-resizing doesn't interfere */
2157 mutex_lock(&conf->cache_size_mutex);
2159 for (i = conf->max_nr_stripes; i; i--) {
2160 nsh = alloc_stripe(sc, GFP_KERNEL);
2164 nsh->raid_conf = conf;
2165 list_add(&nsh->lru, &newstripes);
2168 /* didn't get enough, give up */
2169 while (!list_empty(&newstripes)) {
2170 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2171 list_del(&nsh->lru);
2172 kmem_cache_free(sc, nsh);
2174 kmem_cache_destroy(sc);
2175 mutex_unlock(&conf->cache_size_mutex);
2178 /* Step 2 - Must use GFP_NOIO now.
2179 * OK, we have enough stripes, start collecting inactive
2180 * stripes and copying them over
2184 list_for_each_entry(nsh, &newstripes, lru) {
2185 lock_device_hash_lock(conf, hash);
2186 wait_event_cmd(conf->wait_for_stripe,
2187 !list_empty(conf->inactive_list + hash),
2188 unlock_device_hash_lock(conf, hash),
2189 lock_device_hash_lock(conf, hash));
2190 osh = get_free_stripe(conf, hash);
2191 unlock_device_hash_lock(conf, hash);
2193 for(i=0; i<conf->pool_size; i++) {
2194 nsh->dev[i].page = osh->dev[i].page;
2195 nsh->dev[i].orig_page = osh->dev[i].page;
2197 nsh->hash_lock_index = hash;
2198 kmem_cache_free(conf->slab_cache, osh);
2200 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2201 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2206 kmem_cache_destroy(conf->slab_cache);
2209 * At this point, we are holding all the stripes so the array
2210 * is completely stalled, so now is a good time to resize
2211 * conf->disks and the scribble region
2213 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2215 for (i=0; i<conf->raid_disks; i++)
2216 ndisks[i] = conf->disks[i];
2218 conf->disks = ndisks;
2222 mutex_unlock(&conf->cache_size_mutex);
2223 /* Step 4, return new stripes to service */
2224 while(!list_empty(&newstripes)) {
2225 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2226 list_del_init(&nsh->lru);
2228 for (i=conf->raid_disks; i < newsize; i++)
2229 if (nsh->dev[i].page == NULL) {
2230 struct page *p = alloc_page(GFP_NOIO);
2231 nsh->dev[i].page = p;
2232 nsh->dev[i].orig_page = p;
2236 release_stripe(nsh);
2238 /* critical section pass, GFP_NOIO no longer needed */
2240 conf->slab_cache = sc;
2241 conf->active_name = 1-conf->active_name;
2243 conf->pool_size = newsize;
2247 static int drop_one_stripe(struct r5conf *conf)
2249 struct stripe_head *sh;
2250 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2252 spin_lock_irq(conf->hash_locks + hash);
2253 sh = get_free_stripe(conf, hash);
2254 spin_unlock_irq(conf->hash_locks + hash);
2257 BUG_ON(atomic_read(&sh->count));
2259 kmem_cache_free(conf->slab_cache, sh);
2260 atomic_dec(&conf->active_stripes);
2261 conf->max_nr_stripes--;
2265 static void shrink_stripes(struct r5conf *conf)
2267 while (conf->max_nr_stripes &&
2268 drop_one_stripe(conf))
2271 if (conf->slab_cache)
2272 kmem_cache_destroy(conf->slab_cache);
2273 conf->slab_cache = NULL;
2276 static void raid5_end_read_request(struct bio * bi, int error)
2278 struct stripe_head *sh = bi->bi_private;
2279 struct r5conf *conf = sh->raid_conf;
2280 int disks = sh->disks, i;
2281 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2282 char b[BDEVNAME_SIZE];
2283 struct md_rdev *rdev = NULL;
2286 for (i=0 ; i<disks; i++)
2287 if (bi == &sh->dev[i].req)
2290 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2291 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2297 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2298 /* If replacement finished while this request was outstanding,
2299 * 'replacement' might be NULL already.
2300 * In that case it moved down to 'rdev'.
2301 * rdev is not removed until all requests are finished.
2303 rdev = conf->disks[i].replacement;
2305 rdev = conf->disks[i].rdev;
2307 if (use_new_offset(conf, sh))
2308 s = sh->sector + rdev->new_data_offset;
2310 s = sh->sector + rdev->data_offset;
2312 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2313 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2314 /* Note that this cannot happen on a
2315 * replacement device. We just fail those on
2320 "md/raid:%s: read error corrected"
2321 " (%lu sectors at %llu on %s)\n",
2322 mdname(conf->mddev), STRIPE_SECTORS,
2323 (unsigned long long)s,
2324 bdevname(rdev->bdev, b));
2325 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2326 clear_bit(R5_ReadError, &sh->dev[i].flags);
2327 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2328 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2329 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2331 if (atomic_read(&rdev->read_errors))
2332 atomic_set(&rdev->read_errors, 0);
2334 const char *bdn = bdevname(rdev->bdev, b);
2338 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2339 atomic_inc(&rdev->read_errors);
2340 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2343 "md/raid:%s: read error on replacement device "
2344 "(sector %llu on %s).\n",
2345 mdname(conf->mddev),
2346 (unsigned long long)s,
2348 else if (conf->mddev->degraded >= conf->max_degraded) {
2352 "md/raid:%s: read error not correctable "
2353 "(sector %llu on %s).\n",
2354 mdname(conf->mddev),
2355 (unsigned long long)s,
2357 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2362 "md/raid:%s: read error NOT corrected!! "
2363 "(sector %llu on %s).\n",
2364 mdname(conf->mddev),
2365 (unsigned long long)s,
2367 } else if (atomic_read(&rdev->read_errors)
2368 > conf->max_nr_stripes)
2370 "md/raid:%s: Too many read errors, failing device %s.\n",
2371 mdname(conf->mddev), bdn);
2374 if (set_bad && test_bit(In_sync, &rdev->flags)
2375 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2378 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2379 set_bit(R5_ReadError, &sh->dev[i].flags);
2380 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2382 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2384 clear_bit(R5_ReadError, &sh->dev[i].flags);
2385 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2387 && test_bit(In_sync, &rdev->flags)
2388 && rdev_set_badblocks(
2389 rdev, sh->sector, STRIPE_SECTORS, 0)))
2390 md_error(conf->mddev, rdev);
2393 rdev_dec_pending(rdev, conf->mddev);
2394 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2395 set_bit(STRIPE_HANDLE, &sh->state);
2399 static void raid5_end_write_request(struct bio *bi, int error)
2401 struct stripe_head *sh = bi->bi_private;
2402 struct r5conf *conf = sh->raid_conf;
2403 int disks = sh->disks, i;
2404 struct md_rdev *uninitialized_var(rdev);
2405 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2408 int replacement = 0;
2410 for (i = 0 ; i < disks; i++) {
2411 if (bi == &sh->dev[i].req) {
2412 rdev = conf->disks[i].rdev;
2415 if (bi == &sh->dev[i].rreq) {
2416 rdev = conf->disks[i].replacement;
2420 /* rdev was removed and 'replacement'
2421 * replaced it. rdev is not removed
2422 * until all requests are finished.
2424 rdev = conf->disks[i].rdev;
2428 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2429 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2438 md_error(conf->mddev, rdev);
2439 else if (is_badblock(rdev, sh->sector,
2441 &first_bad, &bad_sectors))
2442 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2445 set_bit(STRIPE_DEGRADED, &sh->state);
2446 set_bit(WriteErrorSeen, &rdev->flags);
2447 set_bit(R5_WriteError, &sh->dev[i].flags);
2448 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2449 set_bit(MD_RECOVERY_NEEDED,
2450 &rdev->mddev->recovery);
2451 } else if (is_badblock(rdev, sh->sector,
2453 &first_bad, &bad_sectors)) {
2454 set_bit(R5_MadeGood, &sh->dev[i].flags);
2455 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2456 /* That was a successful write so make
2457 * sure it looks like we already did
2460 set_bit(R5_ReWrite, &sh->dev[i].flags);
2463 rdev_dec_pending(rdev, conf->mddev);
2465 if (sh->batch_head && !uptodate && !replacement)
2466 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2468 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2469 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2470 set_bit(STRIPE_HANDLE, &sh->state);
2473 if (sh->batch_head && sh != sh->batch_head)
2474 release_stripe(sh->batch_head);
2477 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2479 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2481 struct r5dev *dev = &sh->dev[i];
2483 bio_init(&dev->req);
2484 dev->req.bi_io_vec = &dev->vec;
2485 dev->req.bi_max_vecs = 1;
2486 dev->req.bi_private = sh;
2488 bio_init(&dev->rreq);
2489 dev->rreq.bi_io_vec = &dev->rvec;
2490 dev->rreq.bi_max_vecs = 1;
2491 dev->rreq.bi_private = sh;
2494 dev->sector = compute_blocknr(sh, i, previous);
2497 static void error(struct mddev *mddev, struct md_rdev *rdev)
2499 char b[BDEVNAME_SIZE];
2500 struct r5conf *conf = mddev->private;
2501 unsigned long flags;
2502 pr_debug("raid456: error called\n");
2504 spin_lock_irqsave(&conf->device_lock, flags);
2505 clear_bit(In_sync, &rdev->flags);
2506 mddev->degraded = calc_degraded(conf);
2507 spin_unlock_irqrestore(&conf->device_lock, flags);
2508 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2510 set_bit(Blocked, &rdev->flags);
2511 set_bit(Faulty, &rdev->flags);
2512 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2514 "md/raid:%s: Disk failure on %s, disabling device.\n"
2515 "md/raid:%s: Operation continuing on %d devices.\n",
2517 bdevname(rdev->bdev, b),
2519 conf->raid_disks - mddev->degraded);
2523 * Input: a 'big' sector number,
2524 * Output: index of the data and parity disk, and the sector # in them.
2526 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2527 int previous, int *dd_idx,
2528 struct stripe_head *sh)
2530 sector_t stripe, stripe2;
2531 sector_t chunk_number;
2532 unsigned int chunk_offset;
2535 sector_t new_sector;
2536 int algorithm = previous ? conf->prev_algo
2538 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2539 : conf->chunk_sectors;
2540 int raid_disks = previous ? conf->previous_raid_disks
2542 int data_disks = raid_disks - conf->max_degraded;
2544 /* First compute the information on this sector */
2547 * Compute the chunk number and the sector offset inside the chunk
2549 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2550 chunk_number = r_sector;
2553 * Compute the stripe number
2555 stripe = chunk_number;
2556 *dd_idx = sector_div(stripe, data_disks);
2559 * Select the parity disk based on the user selected algorithm.
2561 pd_idx = qd_idx = -1;
2562 switch(conf->level) {
2564 pd_idx = data_disks;
2567 switch (algorithm) {
2568 case ALGORITHM_LEFT_ASYMMETRIC:
2569 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2570 if (*dd_idx >= pd_idx)
2573 case ALGORITHM_RIGHT_ASYMMETRIC:
2574 pd_idx = sector_div(stripe2, raid_disks);
2575 if (*dd_idx >= pd_idx)
2578 case ALGORITHM_LEFT_SYMMETRIC:
2579 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2580 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2582 case ALGORITHM_RIGHT_SYMMETRIC:
2583 pd_idx = sector_div(stripe2, raid_disks);
2584 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2586 case ALGORITHM_PARITY_0:
2590 case ALGORITHM_PARITY_N:
2591 pd_idx = data_disks;
2599 switch (algorithm) {
2600 case ALGORITHM_LEFT_ASYMMETRIC:
2601 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2602 qd_idx = pd_idx + 1;
2603 if (pd_idx == raid_disks-1) {
2604 (*dd_idx)++; /* Q D D D P */
2606 } else if (*dd_idx >= pd_idx)
2607 (*dd_idx) += 2; /* D D P Q D */
2609 case ALGORITHM_RIGHT_ASYMMETRIC:
2610 pd_idx = sector_div(stripe2, raid_disks);
2611 qd_idx = pd_idx + 1;
2612 if (pd_idx == raid_disks-1) {
2613 (*dd_idx)++; /* Q D D D P */
2615 } else if (*dd_idx >= pd_idx)
2616 (*dd_idx) += 2; /* D D P Q D */
2618 case ALGORITHM_LEFT_SYMMETRIC:
2619 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2620 qd_idx = (pd_idx + 1) % raid_disks;
2621 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2623 case ALGORITHM_RIGHT_SYMMETRIC:
2624 pd_idx = sector_div(stripe2, raid_disks);
2625 qd_idx = (pd_idx + 1) % raid_disks;
2626 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2629 case ALGORITHM_PARITY_0:
2634 case ALGORITHM_PARITY_N:
2635 pd_idx = data_disks;
2636 qd_idx = data_disks + 1;
2639 case ALGORITHM_ROTATING_ZERO_RESTART:
2640 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2641 * of blocks for computing Q is different.
2643 pd_idx = sector_div(stripe2, raid_disks);
2644 qd_idx = pd_idx + 1;
2645 if (pd_idx == raid_disks-1) {
2646 (*dd_idx)++; /* Q D D D P */
2648 } else if (*dd_idx >= pd_idx)
2649 (*dd_idx) += 2; /* D D P Q D */
2653 case ALGORITHM_ROTATING_N_RESTART:
2654 /* Same a left_asymmetric, by first stripe is
2655 * D D D P Q rather than
2659 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2660 qd_idx = pd_idx + 1;
2661 if (pd_idx == raid_disks-1) {
2662 (*dd_idx)++; /* Q D D D P */
2664 } else if (*dd_idx >= pd_idx)
2665 (*dd_idx) += 2; /* D D P Q D */
2669 case ALGORITHM_ROTATING_N_CONTINUE:
2670 /* Same as left_symmetric but Q is before P */
2671 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2672 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2673 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2677 case ALGORITHM_LEFT_ASYMMETRIC_6:
2678 /* RAID5 left_asymmetric, with Q on last device */
2679 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2680 if (*dd_idx >= pd_idx)
2682 qd_idx = raid_disks - 1;
2685 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2686 pd_idx = sector_div(stripe2, raid_disks-1);
2687 if (*dd_idx >= pd_idx)
2689 qd_idx = raid_disks - 1;
2692 case ALGORITHM_LEFT_SYMMETRIC_6:
2693 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2694 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2695 qd_idx = raid_disks - 1;
2698 case ALGORITHM_RIGHT_SYMMETRIC_6:
2699 pd_idx = sector_div(stripe2, raid_disks-1);
2700 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2701 qd_idx = raid_disks - 1;
2704 case ALGORITHM_PARITY_0_6:
2707 qd_idx = raid_disks - 1;
2717 sh->pd_idx = pd_idx;
2718 sh->qd_idx = qd_idx;
2719 sh->ddf_layout = ddf_layout;
2722 * Finally, compute the new sector number
2724 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2728 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2730 struct r5conf *conf = sh->raid_conf;
2731 int raid_disks = sh->disks;
2732 int data_disks = raid_disks - conf->max_degraded;
2733 sector_t new_sector = sh->sector, check;
2734 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2735 : conf->chunk_sectors;
2736 int algorithm = previous ? conf->prev_algo
2740 sector_t chunk_number;
2741 int dummy1, dd_idx = i;
2743 struct stripe_head sh2;
2745 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2746 stripe = new_sector;
2748 if (i == sh->pd_idx)
2750 switch(conf->level) {
2753 switch (algorithm) {
2754 case ALGORITHM_LEFT_ASYMMETRIC:
2755 case ALGORITHM_RIGHT_ASYMMETRIC:
2759 case ALGORITHM_LEFT_SYMMETRIC:
2760 case ALGORITHM_RIGHT_SYMMETRIC:
2763 i -= (sh->pd_idx + 1);
2765 case ALGORITHM_PARITY_0:
2768 case ALGORITHM_PARITY_N:
2775 if (i == sh->qd_idx)
2776 return 0; /* It is the Q disk */
2777 switch (algorithm) {
2778 case ALGORITHM_LEFT_ASYMMETRIC:
2779 case ALGORITHM_RIGHT_ASYMMETRIC:
2780 case ALGORITHM_ROTATING_ZERO_RESTART:
2781 case ALGORITHM_ROTATING_N_RESTART:
2782 if (sh->pd_idx == raid_disks-1)
2783 i--; /* Q D D D P */
2784 else if (i > sh->pd_idx)
2785 i -= 2; /* D D P Q D */
2787 case ALGORITHM_LEFT_SYMMETRIC:
2788 case ALGORITHM_RIGHT_SYMMETRIC:
2789 if (sh->pd_idx == raid_disks-1)
2790 i--; /* Q D D D P */
2795 i -= (sh->pd_idx + 2);
2798 case ALGORITHM_PARITY_0:
2801 case ALGORITHM_PARITY_N:
2803 case ALGORITHM_ROTATING_N_CONTINUE:
2804 /* Like left_symmetric, but P is before Q */
2805 if (sh->pd_idx == 0)
2806 i--; /* P D D D Q */
2811 i -= (sh->pd_idx + 1);
2814 case ALGORITHM_LEFT_ASYMMETRIC_6:
2815 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2819 case ALGORITHM_LEFT_SYMMETRIC_6:
2820 case ALGORITHM_RIGHT_SYMMETRIC_6:
2822 i += data_disks + 1;
2823 i -= (sh->pd_idx + 1);
2825 case ALGORITHM_PARITY_0_6:
2834 chunk_number = stripe * data_disks + i;
2835 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2837 check = raid5_compute_sector(conf, r_sector,
2838 previous, &dummy1, &sh2);
2839 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2840 || sh2.qd_idx != sh->qd_idx) {
2841 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2842 mdname(conf->mddev));
2849 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2850 int rcw, int expand)
2852 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2853 struct r5conf *conf = sh->raid_conf;
2854 int level = conf->level;
2858 for (i = disks; i--; ) {
2859 struct r5dev *dev = &sh->dev[i];
2862 set_bit(R5_LOCKED, &dev->flags);
2863 set_bit(R5_Wantdrain, &dev->flags);
2865 clear_bit(R5_UPTODATE, &dev->flags);
2869 /* if we are not expanding this is a proper write request, and
2870 * there will be bios with new data to be drained into the
2875 /* False alarm, nothing to do */
2877 sh->reconstruct_state = reconstruct_state_drain_run;
2878 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2880 sh->reconstruct_state = reconstruct_state_run;
2882 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2884 if (s->locked + conf->max_degraded == disks)
2885 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2886 atomic_inc(&conf->pending_full_writes);
2888 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2889 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2890 BUG_ON(level == 6 &&
2891 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2892 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2894 for (i = disks; i--; ) {
2895 struct r5dev *dev = &sh->dev[i];
2896 if (i == pd_idx || i == qd_idx)
2900 (test_bit(R5_UPTODATE, &dev->flags) ||
2901 test_bit(R5_Wantcompute, &dev->flags))) {
2902 set_bit(R5_Wantdrain, &dev->flags);
2903 set_bit(R5_LOCKED, &dev->flags);
2904 clear_bit(R5_UPTODATE, &dev->flags);
2909 /* False alarm - nothing to do */
2911 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2912 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2913 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2914 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2917 /* keep the parity disk(s) locked while asynchronous operations
2920 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2921 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2925 int qd_idx = sh->qd_idx;
2926 struct r5dev *dev = &sh->dev[qd_idx];
2928 set_bit(R5_LOCKED, &dev->flags);
2929 clear_bit(R5_UPTODATE, &dev->flags);
2933 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2934 __func__, (unsigned long long)sh->sector,
2935 s->locked, s->ops_request);
2939 * Each stripe/dev can have one or more bion attached.
2940 * toread/towrite point to the first in a chain.
2941 * The bi_next chain must be in order.
2943 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2944 int forwrite, int previous)
2947 struct r5conf *conf = sh->raid_conf;
2950 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2951 (unsigned long long)bi->bi_iter.bi_sector,
2952 (unsigned long long)sh->sector);
2955 * If several bio share a stripe. The bio bi_phys_segments acts as a
2956 * reference count to avoid race. The reference count should already be
2957 * increased before this function is called (for example, in
2958 * make_request()), so other bio sharing this stripe will not free the
2959 * stripe. If a stripe is owned by one stripe, the stripe lock will
2962 spin_lock_irq(&sh->stripe_lock);
2963 /* Don't allow new IO added to stripes in batch list */
2967 bip = &sh->dev[dd_idx].towrite;
2971 bip = &sh->dev[dd_idx].toread;
2972 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2973 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2975 bip = & (*bip)->bi_next;
2977 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2980 if (!forwrite || previous)
2981 clear_bit(STRIPE_BATCH_READY, &sh->state);
2983 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2987 raid5_inc_bi_active_stripes(bi);
2990 /* check if page is covered */
2991 sector_t sector = sh->dev[dd_idx].sector;
2992 for (bi=sh->dev[dd_idx].towrite;
2993 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2994 bi && bi->bi_iter.bi_sector <= sector;
2995 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2996 if (bio_end_sector(bi) >= sector)
2997 sector = bio_end_sector(bi);
2999 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3000 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3001 sh->overwrite_disks++;
3004 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3005 (unsigned long long)(*bip)->bi_iter.bi_sector,
3006 (unsigned long long)sh->sector, dd_idx);
3008 if (conf->mddev->bitmap && firstwrite) {
3009 /* Cannot hold spinlock over bitmap_startwrite,
3010 * but must ensure this isn't added to a batch until
3011 * we have added to the bitmap and set bm_seq.
3012 * So set STRIPE_BITMAP_PENDING to prevent
3014 * If multiple add_stripe_bio() calls race here they
3015 * much all set STRIPE_BITMAP_PENDING. So only the first one
3016 * to complete "bitmap_startwrite" gets to set
3017 * STRIPE_BIT_DELAY. This is important as once a stripe
3018 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3021 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3022 spin_unlock_irq(&sh->stripe_lock);
3023 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3025 spin_lock_irq(&sh->stripe_lock);
3026 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3027 if (!sh->batch_head) {
3028 sh->bm_seq = conf->seq_flush+1;
3029 set_bit(STRIPE_BIT_DELAY, &sh->state);
3032 spin_unlock_irq(&sh->stripe_lock);
3034 if (stripe_can_batch(sh))
3035 stripe_add_to_batch_list(conf, sh);
3039 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3040 spin_unlock_irq(&sh->stripe_lock);
3044 static void end_reshape(struct r5conf *conf);
3046 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3047 struct stripe_head *sh)
3049 int sectors_per_chunk =
3050 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3052 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3053 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3055 raid5_compute_sector(conf,
3056 stripe * (disks - conf->max_degraded)
3057 *sectors_per_chunk + chunk_offset,
3063 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3064 struct stripe_head_state *s, int disks,
3065 struct bio **return_bi)
3068 BUG_ON(sh->batch_head);
3069 for (i = disks; i--; ) {
3073 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3074 struct md_rdev *rdev;
3076 rdev = rcu_dereference(conf->disks[i].rdev);
3077 if (rdev && test_bit(In_sync, &rdev->flags))
3078 atomic_inc(&rdev->nr_pending);
3083 if (!rdev_set_badblocks(
3087 md_error(conf->mddev, rdev);
3088 rdev_dec_pending(rdev, conf->mddev);
3091 spin_lock_irq(&sh->stripe_lock);
3092 /* fail all writes first */
3093 bi = sh->dev[i].towrite;
3094 sh->dev[i].towrite = NULL;
3095 sh->overwrite_disks = 0;
3096 spin_unlock_irq(&sh->stripe_lock);
3100 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3101 wake_up(&conf->wait_for_overlap);
3103 while (bi && bi->bi_iter.bi_sector <
3104 sh->dev[i].sector + STRIPE_SECTORS) {
3105 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3106 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3107 if (!raid5_dec_bi_active_stripes(bi)) {
3108 md_write_end(conf->mddev);
3109 bi->bi_next = *return_bi;
3115 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3116 STRIPE_SECTORS, 0, 0);
3118 /* and fail all 'written' */
3119 bi = sh->dev[i].written;
3120 sh->dev[i].written = NULL;
3121 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3122 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3123 sh->dev[i].page = sh->dev[i].orig_page;
3126 if (bi) bitmap_end = 1;
3127 while (bi && bi->bi_iter.bi_sector <
3128 sh->dev[i].sector + STRIPE_SECTORS) {
3129 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3130 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3131 if (!raid5_dec_bi_active_stripes(bi)) {
3132 md_write_end(conf->mddev);
3133 bi->bi_next = *return_bi;
3139 /* fail any reads if this device is non-operational and
3140 * the data has not reached the cache yet.
3142 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3143 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3144 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3145 spin_lock_irq(&sh->stripe_lock);
3146 bi = sh->dev[i].toread;
3147 sh->dev[i].toread = NULL;
3148 spin_unlock_irq(&sh->stripe_lock);
3149 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3150 wake_up(&conf->wait_for_overlap);
3151 while (bi && bi->bi_iter.bi_sector <
3152 sh->dev[i].sector + STRIPE_SECTORS) {
3153 struct bio *nextbi =
3154 r5_next_bio(bi, sh->dev[i].sector);
3155 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3156 if (!raid5_dec_bi_active_stripes(bi)) {
3157 bi->bi_next = *return_bi;
3164 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3165 STRIPE_SECTORS, 0, 0);
3166 /* If we were in the middle of a write the parity block might
3167 * still be locked - so just clear all R5_LOCKED flags
3169 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3172 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3173 if (atomic_dec_and_test(&conf->pending_full_writes))
3174 md_wakeup_thread(conf->mddev->thread);
3178 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3179 struct stripe_head_state *s)
3184 BUG_ON(sh->batch_head);
3185 clear_bit(STRIPE_SYNCING, &sh->state);
3186 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3187 wake_up(&conf->wait_for_overlap);
3190 /* There is nothing more to do for sync/check/repair.
3191 * Don't even need to abort as that is handled elsewhere
3192 * if needed, and not always wanted e.g. if there is a known
3194 * For recover/replace we need to record a bad block on all
3195 * non-sync devices, or abort the recovery
3197 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3198 /* During recovery devices cannot be removed, so
3199 * locking and refcounting of rdevs is not needed
3201 for (i = 0; i < conf->raid_disks; i++) {
3202 struct md_rdev *rdev = conf->disks[i].rdev;
3204 && !test_bit(Faulty, &rdev->flags)
3205 && !test_bit(In_sync, &rdev->flags)
3206 && !rdev_set_badblocks(rdev, sh->sector,
3209 rdev = conf->disks[i].replacement;
3211 && !test_bit(Faulty, &rdev->flags)
3212 && !test_bit(In_sync, &rdev->flags)
3213 && !rdev_set_badblocks(rdev, sh->sector,
3218 conf->recovery_disabled =
3219 conf->mddev->recovery_disabled;
3221 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3224 static int want_replace(struct stripe_head *sh, int disk_idx)
3226 struct md_rdev *rdev;
3228 /* Doing recovery so rcu locking not required */
3229 rdev = sh->raid_conf->disks[disk_idx].replacement;
3231 && !test_bit(Faulty, &rdev->flags)
3232 && !test_bit(In_sync, &rdev->flags)
3233 && (rdev->recovery_offset <= sh->sector
3234 || rdev->mddev->recovery_cp <= sh->sector))
3240 /* fetch_block - checks the given member device to see if its data needs
3241 * to be read or computed to satisfy a request.
3243 * Returns 1 when no more member devices need to be checked, otherwise returns
3244 * 0 to tell the loop in handle_stripe_fill to continue
3247 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3248 int disk_idx, int disks)
3250 struct r5dev *dev = &sh->dev[disk_idx];
3251 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3252 &sh->dev[s->failed_num[1]] };
3256 if (test_bit(R5_LOCKED, &dev->flags) ||
3257 test_bit(R5_UPTODATE, &dev->flags))
3258 /* No point reading this as we already have it or have
3259 * decided to get it.
3264 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3265 /* We need this block to directly satisfy a request */
3268 if (s->syncing || s->expanding ||
3269 (s->replacing && want_replace(sh, disk_idx)))
3270 /* When syncing, or expanding we read everything.
3271 * When replacing, we need the replaced block.
3275 if ((s->failed >= 1 && fdev[0]->toread) ||
3276 (s->failed >= 2 && fdev[1]->toread))
3277 /* If we want to read from a failed device, then
3278 * we need to actually read every other device.
3282 /* Sometimes neither read-modify-write nor reconstruct-write
3283 * cycles can work. In those cases we read every block we
3284 * can. Then the parity-update is certain to have enough to
3286 * This can only be a problem when we need to write something,
3287 * and some device has failed. If either of those tests
3288 * fail we need look no further.
3290 if (!s->failed || !s->to_write)
3293 if (test_bit(R5_Insync, &dev->flags) &&
3294 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3295 /* Pre-reads at not permitted until after short delay
3296 * to gather multiple requests. However if this
3297 * device is no Insync, the block could only be be computed
3298 * and there is no need to delay that.
3302 for (i = 0; i < s->failed; i++) {
3303 if (fdev[i]->towrite &&
3304 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3305 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3306 /* If we have a partial write to a failed
3307 * device, then we will need to reconstruct
3308 * the content of that device, so all other
3309 * devices must be read.
3314 /* If we are forced to do a reconstruct-write, either because
3315 * the current RAID6 implementation only supports that, or
3316 * or because parity cannot be trusted and we are currently
3317 * recovering it, there is extra need to be careful.
3318 * If one of the devices that we would need to read, because
3319 * it is not being overwritten (and maybe not written at all)
3320 * is missing/faulty, then we need to read everything we can.
3322 if (sh->raid_conf->level != 6 &&
3323 sh->sector < sh->raid_conf->mddev->recovery_cp)
3324 /* reconstruct-write isn't being forced */
3326 for (i = 0; i < s->failed; i++) {
3327 if (s->failed_num[i] != sh->pd_idx &&
3328 s->failed_num[i] != sh->qd_idx &&
3329 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3330 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3337 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3338 int disk_idx, int disks)
3340 struct r5dev *dev = &sh->dev[disk_idx];
3342 /* is the data in this block needed, and can we get it? */
3343 if (need_this_block(sh, s, disk_idx, disks)) {
3344 /* we would like to get this block, possibly by computing it,
3345 * otherwise read it if the backing disk is insync
3347 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3348 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3349 BUG_ON(sh->batch_head);
3350 if ((s->uptodate == disks - 1) &&
3351 (s->failed && (disk_idx == s->failed_num[0] ||
3352 disk_idx == s->failed_num[1]))) {
3353 /* have disk failed, and we're requested to fetch it;
3356 pr_debug("Computing stripe %llu block %d\n",
3357 (unsigned long long)sh->sector, disk_idx);
3358 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3359 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3360 set_bit(R5_Wantcompute, &dev->flags);
3361 sh->ops.target = disk_idx;
3362 sh->ops.target2 = -1; /* no 2nd target */
3364 /* Careful: from this point on 'uptodate' is in the eye
3365 * of raid_run_ops which services 'compute' operations
3366 * before writes. R5_Wantcompute flags a block that will
3367 * be R5_UPTODATE by the time it is needed for a
3368 * subsequent operation.
3372 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3373 /* Computing 2-failure is *very* expensive; only
3374 * do it if failed >= 2
3377 for (other = disks; other--; ) {
3378 if (other == disk_idx)
3380 if (!test_bit(R5_UPTODATE,
3381 &sh->dev[other].flags))
3385 pr_debug("Computing stripe %llu blocks %d,%d\n",
3386 (unsigned long long)sh->sector,
3388 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3389 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3390 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3391 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3392 sh->ops.target = disk_idx;
3393 sh->ops.target2 = other;
3397 } else if (test_bit(R5_Insync, &dev->flags)) {
3398 set_bit(R5_LOCKED, &dev->flags);
3399 set_bit(R5_Wantread, &dev->flags);
3401 pr_debug("Reading block %d (sync=%d)\n",
3402 disk_idx, s->syncing);
3410 * handle_stripe_fill - read or compute data to satisfy pending requests.
3412 static void handle_stripe_fill(struct stripe_head *sh,
3413 struct stripe_head_state *s,
3418 /* look for blocks to read/compute, skip this if a compute
3419 * is already in flight, or if the stripe contents are in the
3420 * midst of changing due to a write
3422 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3423 !sh->reconstruct_state)
3424 for (i = disks; i--; )
3425 if (fetch_block(sh, s, i, disks))
3427 set_bit(STRIPE_HANDLE, &sh->state);
3430 static void break_stripe_batch_list(struct stripe_head *head_sh,
3431 unsigned long handle_flags);
3432 /* handle_stripe_clean_event
3433 * any written block on an uptodate or failed drive can be returned.
3434 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3435 * never LOCKED, so we don't need to test 'failed' directly.
3437 static void handle_stripe_clean_event(struct r5conf *conf,
3438 struct stripe_head *sh, int disks, struct bio **return_bi)
3442 int discard_pending = 0;
3443 struct stripe_head *head_sh = sh;
3444 bool do_endio = false;
3446 for (i = disks; i--; )
3447 if (sh->dev[i].written) {
3449 if (!test_bit(R5_LOCKED, &dev->flags) &&
3450 (test_bit(R5_UPTODATE, &dev->flags) ||
3451 test_bit(R5_Discard, &dev->flags) ||
3452 test_bit(R5_SkipCopy, &dev->flags))) {
3453 /* We can return any write requests */
3454 struct bio *wbi, *wbi2;
3455 pr_debug("Return write for disc %d\n", i);
3456 if (test_and_clear_bit(R5_Discard, &dev->flags))
3457 clear_bit(R5_UPTODATE, &dev->flags);
3458 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3459 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3464 dev->page = dev->orig_page;
3466 dev->written = NULL;
3467 while (wbi && wbi->bi_iter.bi_sector <
3468 dev->sector + STRIPE_SECTORS) {
3469 wbi2 = r5_next_bio(wbi, dev->sector);
3470 if (!raid5_dec_bi_active_stripes(wbi)) {
3471 md_write_end(conf->mddev);
3472 wbi->bi_next = *return_bi;
3477 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3479 !test_bit(STRIPE_DEGRADED, &sh->state),
3481 if (head_sh->batch_head) {
3482 sh = list_first_entry(&sh->batch_list,
3485 if (sh != head_sh) {
3492 } else if (test_bit(R5_Discard, &dev->flags))
3493 discard_pending = 1;
3494 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3495 WARN_ON(dev->page != dev->orig_page);
3497 if (!discard_pending &&
3498 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3499 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3500 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3501 if (sh->qd_idx >= 0) {
3502 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3503 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3505 /* now that discard is done we can proceed with any sync */
3506 clear_bit(STRIPE_DISCARD, &sh->state);
3508 * SCSI discard will change some bio fields and the stripe has
3509 * no updated data, so remove it from hash list and the stripe
3510 * will be reinitialized
3512 spin_lock_irq(&conf->device_lock);
3515 if (head_sh->batch_head) {
3516 sh = list_first_entry(&sh->batch_list,
3517 struct stripe_head, batch_list);
3521 spin_unlock_irq(&conf->device_lock);
3524 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3525 set_bit(STRIPE_HANDLE, &sh->state);
3529 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3530 if (atomic_dec_and_test(&conf->pending_full_writes))
3531 md_wakeup_thread(conf->mddev->thread);
3533 if (head_sh->batch_head && do_endio)
3534 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3537 static void handle_stripe_dirtying(struct r5conf *conf,
3538 struct stripe_head *sh,
3539 struct stripe_head_state *s,
3542 int rmw = 0, rcw = 0, i;
3543 sector_t recovery_cp = conf->mddev->recovery_cp;
3545 /* Check whether resync is now happening or should start.
3546 * If yes, then the array is dirty (after unclean shutdown or
3547 * initial creation), so parity in some stripes might be inconsistent.
3548 * In this case, we need to always do reconstruct-write, to ensure
3549 * that in case of drive failure or read-error correction, we
3550 * generate correct data from the parity.
3552 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3553 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3555 /* Calculate the real rcw later - for now make it
3556 * look like rcw is cheaper
3559 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3560 conf->rmw_level, (unsigned long long)recovery_cp,
3561 (unsigned long long)sh->sector);
3562 } else for (i = disks; i--; ) {
3563 /* would I have to read this buffer for read_modify_write */
3564 struct r5dev *dev = &sh->dev[i];
3565 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3566 !test_bit(R5_LOCKED, &dev->flags) &&
3567 !(test_bit(R5_UPTODATE, &dev->flags) ||
3568 test_bit(R5_Wantcompute, &dev->flags))) {
3569 if (test_bit(R5_Insync, &dev->flags))
3572 rmw += 2*disks; /* cannot read it */
3574 /* Would I have to read this buffer for reconstruct_write */
3575 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3576 i != sh->pd_idx && i != sh->qd_idx &&
3577 !test_bit(R5_LOCKED, &dev->flags) &&
3578 !(test_bit(R5_UPTODATE, &dev->flags) ||
3579 test_bit(R5_Wantcompute, &dev->flags))) {
3580 if (test_bit(R5_Insync, &dev->flags))
3586 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3587 (unsigned long long)sh->sector, rmw, rcw);
3588 set_bit(STRIPE_HANDLE, &sh->state);
3589 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3590 /* prefer read-modify-write, but need to get some data */
3591 if (conf->mddev->queue)
3592 blk_add_trace_msg(conf->mddev->queue,
3593 "raid5 rmw %llu %d",
3594 (unsigned long long)sh->sector, rmw);
3595 for (i = disks; i--; ) {
3596 struct r5dev *dev = &sh->dev[i];
3597 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3598 !test_bit(R5_LOCKED, &dev->flags) &&
3599 !(test_bit(R5_UPTODATE, &dev->flags) ||
3600 test_bit(R5_Wantcompute, &dev->flags)) &&
3601 test_bit(R5_Insync, &dev->flags)) {
3602 if (test_bit(STRIPE_PREREAD_ACTIVE,
3604 pr_debug("Read_old block %d for r-m-w\n",
3606 set_bit(R5_LOCKED, &dev->flags);
3607 set_bit(R5_Wantread, &dev->flags);
3610 set_bit(STRIPE_DELAYED, &sh->state);
3611 set_bit(STRIPE_HANDLE, &sh->state);
3616 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3617 /* want reconstruct write, but need to get some data */
3620 for (i = disks; i--; ) {
3621 struct r5dev *dev = &sh->dev[i];
3622 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3623 i != sh->pd_idx && i != sh->qd_idx &&
3624 !test_bit(R5_LOCKED, &dev->flags) &&
3625 !(test_bit(R5_UPTODATE, &dev->flags) ||
3626 test_bit(R5_Wantcompute, &dev->flags))) {
3628 if (test_bit(R5_Insync, &dev->flags) &&
3629 test_bit(STRIPE_PREREAD_ACTIVE,
3631 pr_debug("Read_old block "
3632 "%d for Reconstruct\n", i);
3633 set_bit(R5_LOCKED, &dev->flags);
3634 set_bit(R5_Wantread, &dev->flags);
3638 set_bit(STRIPE_DELAYED, &sh->state);
3639 set_bit(STRIPE_HANDLE, &sh->state);
3643 if (rcw && conf->mddev->queue)
3644 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3645 (unsigned long long)sh->sector,
3646 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3649 if (rcw > disks && rmw > disks &&
3650 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3651 set_bit(STRIPE_DELAYED, &sh->state);
3653 /* now if nothing is locked, and if we have enough data,
3654 * we can start a write request
3656 /* since handle_stripe can be called at any time we need to handle the
3657 * case where a compute block operation has been submitted and then a
3658 * subsequent call wants to start a write request. raid_run_ops only
3659 * handles the case where compute block and reconstruct are requested
3660 * simultaneously. If this is not the case then new writes need to be
3661 * held off until the compute completes.
3663 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3664 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3665 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3666 schedule_reconstruction(sh, s, rcw == 0, 0);
3669 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3670 struct stripe_head_state *s, int disks)
3672 struct r5dev *dev = NULL;
3674 BUG_ON(sh->batch_head);
3675 set_bit(STRIPE_HANDLE, &sh->state);
3677 switch (sh->check_state) {
3678 case check_state_idle:
3679 /* start a new check operation if there are no failures */
3680 if (s->failed == 0) {
3681 BUG_ON(s->uptodate != disks);
3682 sh->check_state = check_state_run;
3683 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3684 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3688 dev = &sh->dev[s->failed_num[0]];
3690 case check_state_compute_result:
3691 sh->check_state = check_state_idle;
3693 dev = &sh->dev[sh->pd_idx];
3695 /* check that a write has not made the stripe insync */
3696 if (test_bit(STRIPE_INSYNC, &sh->state))
3699 /* either failed parity check, or recovery is happening */
3700 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3701 BUG_ON(s->uptodate != disks);
3703 set_bit(R5_LOCKED, &dev->flags);
3705 set_bit(R5_Wantwrite, &dev->flags);
3707 clear_bit(STRIPE_DEGRADED, &sh->state);
3708 set_bit(STRIPE_INSYNC, &sh->state);
3710 case check_state_run:
3711 break; /* we will be called again upon completion */
3712 case check_state_check_result:
3713 sh->check_state = check_state_idle;
3715 /* if a failure occurred during the check operation, leave
3716 * STRIPE_INSYNC not set and let the stripe be handled again
3721 /* handle a successful check operation, if parity is correct
3722 * we are done. Otherwise update the mismatch count and repair
3723 * parity if !MD_RECOVERY_CHECK
3725 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3726 /* parity is correct (on disc,
3727 * not in buffer any more)
3729 set_bit(STRIPE_INSYNC, &sh->state);
3731 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3732 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3733 /* don't try to repair!! */
3734 set_bit(STRIPE_INSYNC, &sh->state);
3736 sh->check_state = check_state_compute_run;
3737 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3738 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3739 set_bit(R5_Wantcompute,
3740 &sh->dev[sh->pd_idx].flags);
3741 sh->ops.target = sh->pd_idx;
3742 sh->ops.target2 = -1;
3747 case check_state_compute_run:
3750 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3751 __func__, sh->check_state,
3752 (unsigned long long) sh->sector);
3757 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3758 struct stripe_head_state *s,
3761 int pd_idx = sh->pd_idx;
3762 int qd_idx = sh->qd_idx;
3765 BUG_ON(sh->batch_head);
3766 set_bit(STRIPE_HANDLE, &sh->state);
3768 BUG_ON(s->failed > 2);
3770 /* Want to check and possibly repair P and Q.
3771 * However there could be one 'failed' device, in which
3772 * case we can only check one of them, possibly using the
3773 * other to generate missing data
3776 switch (sh->check_state) {
3777 case check_state_idle:
3778 /* start a new check operation if there are < 2 failures */
3779 if (s->failed == s->q_failed) {
3780 /* The only possible failed device holds Q, so it
3781 * makes sense to check P (If anything else were failed,
3782 * we would have used P to recreate it).
3784 sh->check_state = check_state_run;
3786 if (!s->q_failed && s->failed < 2) {
3787 /* Q is not failed, and we didn't use it to generate
3788 * anything, so it makes sense to check it
3790 if (sh->check_state == check_state_run)
3791 sh->check_state = check_state_run_pq;
3793 sh->check_state = check_state_run_q;
3796 /* discard potentially stale zero_sum_result */
3797 sh->ops.zero_sum_result = 0;
3799 if (sh->check_state == check_state_run) {
3800 /* async_xor_zero_sum destroys the contents of P */
3801 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3804 if (sh->check_state >= check_state_run &&
3805 sh->check_state <= check_state_run_pq) {
3806 /* async_syndrome_zero_sum preserves P and Q, so
3807 * no need to mark them !uptodate here
3809 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3813 /* we have 2-disk failure */
3814 BUG_ON(s->failed != 2);
3816 case check_state_compute_result:
3817 sh->check_state = check_state_idle;
3819 /* check that a write has not made the stripe insync */
3820 if (test_bit(STRIPE_INSYNC, &sh->state))
3823 /* now write out any block on a failed drive,
3824 * or P or Q if they were recomputed
3826 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3827 if (s->failed == 2) {
3828 dev = &sh->dev[s->failed_num[1]];
3830 set_bit(R5_LOCKED, &dev->flags);
3831 set_bit(R5_Wantwrite, &dev->flags);
3833 if (s->failed >= 1) {
3834 dev = &sh->dev[s->failed_num[0]];
3836 set_bit(R5_LOCKED, &dev->flags);
3837 set_bit(R5_Wantwrite, &dev->flags);
3839 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3840 dev = &sh->dev[pd_idx];
3842 set_bit(R5_LOCKED, &dev->flags);
3843 set_bit(R5_Wantwrite, &dev->flags);
3845 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3846 dev = &sh->dev[qd_idx];
3848 set_bit(R5_LOCKED, &dev->flags);
3849 set_bit(R5_Wantwrite, &dev->flags);
3851 clear_bit(STRIPE_DEGRADED, &sh->state);
3853 set_bit(STRIPE_INSYNC, &sh->state);
3855 case check_state_run:
3856 case check_state_run_q:
3857 case check_state_run_pq:
3858 break; /* we will be called again upon completion */
3859 case check_state_check_result:
3860 sh->check_state = check_state_idle;
3862 /* handle a successful check operation, if parity is correct
3863 * we are done. Otherwise update the mismatch count and repair
3864 * parity if !MD_RECOVERY_CHECK
3866 if (sh->ops.zero_sum_result == 0) {
3867 /* both parities are correct */
3869 set_bit(STRIPE_INSYNC, &sh->state);
3871 /* in contrast to the raid5 case we can validate
3872 * parity, but still have a failure to write
3875 sh->check_state = check_state_compute_result;
3876 /* Returning at this point means that we may go
3877 * off and bring p and/or q uptodate again so
3878 * we make sure to check zero_sum_result again
3879 * to verify if p or q need writeback
3883 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3884 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3885 /* don't try to repair!! */
3886 set_bit(STRIPE_INSYNC, &sh->state);
3888 int *target = &sh->ops.target;
3890 sh->ops.target = -1;
3891 sh->ops.target2 = -1;
3892 sh->check_state = check_state_compute_run;
3893 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3894 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3895 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3896 set_bit(R5_Wantcompute,
3897 &sh->dev[pd_idx].flags);
3899 target = &sh->ops.target2;
3902 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3903 set_bit(R5_Wantcompute,
3904 &sh->dev[qd_idx].flags);
3911 case check_state_compute_run:
3914 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3915 __func__, sh->check_state,
3916 (unsigned long long) sh->sector);
3921 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3925 /* We have read all the blocks in this stripe and now we need to
3926 * copy some of them into a target stripe for expand.
3928 struct dma_async_tx_descriptor *tx = NULL;
3929 BUG_ON(sh->batch_head);
3930 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3931 for (i = 0; i < sh->disks; i++)
3932 if (i != sh->pd_idx && i != sh->qd_idx) {
3934 struct stripe_head *sh2;
3935 struct async_submit_ctl submit;
3937 sector_t bn = compute_blocknr(sh, i, 1);
3938 sector_t s = raid5_compute_sector(conf, bn, 0,
3940 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3942 /* so far only the early blocks of this stripe
3943 * have been requested. When later blocks
3944 * get requested, we will try again
3947 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3948 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3949 /* must have already done this block */
3950 release_stripe(sh2);
3954 /* place all the copies on one channel */
3955 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3956 tx = async_memcpy(sh2->dev[dd_idx].page,
3957 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3960 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3961 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3962 for (j = 0; j < conf->raid_disks; j++)
3963 if (j != sh2->pd_idx &&
3965 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3967 if (j == conf->raid_disks) {
3968 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3969 set_bit(STRIPE_HANDLE, &sh2->state);
3971 release_stripe(sh2);
3974 /* done submitting copies, wait for them to complete */
3975 async_tx_quiesce(&tx);
3979 * handle_stripe - do things to a stripe.
3981 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3982 * state of various bits to see what needs to be done.
3984 * return some read requests which now have data
3985 * return some write requests which are safely on storage
3986 * schedule a read on some buffers
3987 * schedule a write of some buffers
3988 * return confirmation of parity correctness
3992 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3994 struct r5conf *conf = sh->raid_conf;
3995 int disks = sh->disks;
3998 int do_recovery = 0;
4000 memset(s, 0, sizeof(*s));
4002 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4003 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4004 s->failed_num[0] = -1;
4005 s->failed_num[1] = -1;
4007 /* Now to look around and see what can be done */
4009 for (i=disks; i--; ) {
4010 struct md_rdev *rdev;
4017 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4019 dev->toread, dev->towrite, dev->written);
4020 /* maybe we can reply to a read
4022 * new wantfill requests are only permitted while
4023 * ops_complete_biofill is guaranteed to be inactive
4025 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4026 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4027 set_bit(R5_Wantfill, &dev->flags);
4029 /* now count some things */
4030 if (test_bit(R5_LOCKED, &dev->flags))
4032 if (test_bit(R5_UPTODATE, &dev->flags))
4034 if (test_bit(R5_Wantcompute, &dev->flags)) {
4036 BUG_ON(s->compute > 2);
4039 if (test_bit(R5_Wantfill, &dev->flags))
4041 else if (dev->toread)
4045 if (!test_bit(R5_OVERWRITE, &dev->flags))
4050 /* Prefer to use the replacement for reads, but only
4051 * if it is recovered enough and has no bad blocks.
4053 rdev = rcu_dereference(conf->disks[i].replacement);
4054 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4055 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4056 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4057 &first_bad, &bad_sectors))
4058 set_bit(R5_ReadRepl, &dev->flags);
4061 set_bit(R5_NeedReplace, &dev->flags);
4062 rdev = rcu_dereference(conf->disks[i].rdev);
4063 clear_bit(R5_ReadRepl, &dev->flags);
4065 if (rdev && test_bit(Faulty, &rdev->flags))
4068 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4069 &first_bad, &bad_sectors);
4070 if (s->blocked_rdev == NULL
4071 && (test_bit(Blocked, &rdev->flags)
4074 set_bit(BlockedBadBlocks,
4076 s->blocked_rdev = rdev;
4077 atomic_inc(&rdev->nr_pending);
4080 clear_bit(R5_Insync, &dev->flags);
4084 /* also not in-sync */
4085 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4086 test_bit(R5_UPTODATE, &dev->flags)) {
4087 /* treat as in-sync, but with a read error
4088 * which we can now try to correct
4090 set_bit(R5_Insync, &dev->flags);
4091 set_bit(R5_ReadError, &dev->flags);
4093 } else if (test_bit(In_sync, &rdev->flags))
4094 set_bit(R5_Insync, &dev->flags);
4095 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4096 /* in sync if before recovery_offset */
4097 set_bit(R5_Insync, &dev->flags);
4098 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4099 test_bit(R5_Expanded, &dev->flags))
4100 /* If we've reshaped into here, we assume it is Insync.
4101 * We will shortly update recovery_offset to make
4104 set_bit(R5_Insync, &dev->flags);
4106 if (test_bit(R5_WriteError, &dev->flags)) {
4107 /* This flag does not apply to '.replacement'
4108 * only to .rdev, so make sure to check that*/
4109 struct md_rdev *rdev2 = rcu_dereference(
4110 conf->disks[i].rdev);
4112 clear_bit(R5_Insync, &dev->flags);
4113 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4114 s->handle_bad_blocks = 1;
4115 atomic_inc(&rdev2->nr_pending);
4117 clear_bit(R5_WriteError, &dev->flags);
4119 if (test_bit(R5_MadeGood, &dev->flags)) {
4120 /* This flag does not apply to '.replacement'
4121 * only to .rdev, so make sure to check that*/
4122 struct md_rdev *rdev2 = rcu_dereference(
4123 conf->disks[i].rdev);
4124 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4125 s->handle_bad_blocks = 1;
4126 atomic_inc(&rdev2->nr_pending);
4128 clear_bit(R5_MadeGood, &dev->flags);
4130 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4131 struct md_rdev *rdev2 = rcu_dereference(
4132 conf->disks[i].replacement);
4133 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4134 s->handle_bad_blocks = 1;
4135 atomic_inc(&rdev2->nr_pending);
4137 clear_bit(R5_MadeGoodRepl, &dev->flags);
4139 if (!test_bit(R5_Insync, &dev->flags)) {
4140 /* The ReadError flag will just be confusing now */
4141 clear_bit(R5_ReadError, &dev->flags);
4142 clear_bit(R5_ReWrite, &dev->flags);
4144 if (test_bit(R5_ReadError, &dev->flags))
4145 clear_bit(R5_Insync, &dev->flags);
4146 if (!test_bit(R5_Insync, &dev->flags)) {
4148 s->failed_num[s->failed] = i;
4150 if (rdev && !test_bit(Faulty, &rdev->flags))
4154 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4155 /* If there is a failed device being replaced,
4156 * we must be recovering.
4157 * else if we are after recovery_cp, we must be syncing
4158 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4159 * else we can only be replacing
4160 * sync and recovery both need to read all devices, and so
4161 * use the same flag.
4164 sh->sector >= conf->mddev->recovery_cp ||
4165 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4173 static int clear_batch_ready(struct stripe_head *sh)
4175 /* Return '1' if this is a member of batch, or
4176 * '0' if it is a lone stripe or a head which can now be
4179 struct stripe_head *tmp;
4180 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4181 return (sh->batch_head && sh->batch_head != sh);
4182 spin_lock(&sh->stripe_lock);
4183 if (!sh->batch_head) {
4184 spin_unlock(&sh->stripe_lock);
4189 * this stripe could be added to a batch list before we check
4190 * BATCH_READY, skips it
4192 if (sh->batch_head != sh) {
4193 spin_unlock(&sh->stripe_lock);
4196 spin_lock(&sh->batch_lock);
4197 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4198 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4199 spin_unlock(&sh->batch_lock);
4200 spin_unlock(&sh->stripe_lock);
4203 * BATCH_READY is cleared, no new stripes can be added.
4204 * batch_list can be accessed without lock
4209 static void break_stripe_batch_list(struct stripe_head *head_sh,
4210 unsigned long handle_flags)
4212 struct stripe_head *sh, *next;
4216 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4218 list_del_init(&sh->batch_list);
4220 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4221 (1 << STRIPE_SYNCING) |
4222 (1 << STRIPE_REPLACED) |
4223 (1 << STRIPE_PREREAD_ACTIVE) |
4224 (1 << STRIPE_DELAYED) |
4225 (1 << STRIPE_BIT_DELAY) |
4226 (1 << STRIPE_FULL_WRITE) |
4227 (1 << STRIPE_BIOFILL_RUN) |
4228 (1 << STRIPE_COMPUTE_RUN) |
4229 (1 << STRIPE_OPS_REQ_PENDING) |
4230 (1 << STRIPE_DISCARD) |
4231 (1 << STRIPE_BATCH_READY) |
4232 (1 << STRIPE_BATCH_ERR) |
4233 (1 << STRIPE_BITMAP_PENDING)));
4234 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4235 (1 << STRIPE_REPLACED)));
4237 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4238 (1 << STRIPE_DEGRADED)),
4239 head_sh->state & (1 << STRIPE_INSYNC));
4241 sh->check_state = head_sh->check_state;
4242 sh->reconstruct_state = head_sh->reconstruct_state;
4243 for (i = 0; i < sh->disks; i++) {
4244 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4246 sh->dev[i].flags = head_sh->dev[i].flags &
4247 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4249 spin_lock_irq(&sh->stripe_lock);
4250 sh->batch_head = NULL;
4251 spin_unlock_irq(&sh->stripe_lock);
4252 if (handle_flags == 0 ||
4253 sh->state & handle_flags)
4254 set_bit(STRIPE_HANDLE, &sh->state);
4257 spin_lock_irq(&head_sh->stripe_lock);
4258 head_sh->batch_head = NULL;
4259 spin_unlock_irq(&head_sh->stripe_lock);
4260 for (i = 0; i < head_sh->disks; i++)
4261 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4263 if (head_sh->state & handle_flags)
4264 set_bit(STRIPE_HANDLE, &head_sh->state);
4267 wake_up(&head_sh->raid_conf->wait_for_overlap);
4270 static void handle_stripe(struct stripe_head *sh)
4272 struct stripe_head_state s;
4273 struct r5conf *conf = sh->raid_conf;
4276 int disks = sh->disks;
4277 struct r5dev *pdev, *qdev;
4279 clear_bit(STRIPE_HANDLE, &sh->state);
4280 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4281 /* already being handled, ensure it gets handled
4282 * again when current action finishes */
4283 set_bit(STRIPE_HANDLE, &sh->state);
4287 if (clear_batch_ready(sh) ) {
4288 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4292 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4293 break_stripe_batch_list(sh, 0);
4295 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4296 spin_lock(&sh->stripe_lock);
4297 /* Cannot process 'sync' concurrently with 'discard' */
4298 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4299 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4300 set_bit(STRIPE_SYNCING, &sh->state);
4301 clear_bit(STRIPE_INSYNC, &sh->state);
4302 clear_bit(STRIPE_REPLACED, &sh->state);
4304 spin_unlock(&sh->stripe_lock);
4306 clear_bit(STRIPE_DELAYED, &sh->state);
4308 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4309 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4310 (unsigned long long)sh->sector, sh->state,
4311 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4312 sh->check_state, sh->reconstruct_state);
4314 analyse_stripe(sh, &s);
4316 if (s.handle_bad_blocks) {
4317 set_bit(STRIPE_HANDLE, &sh->state);
4321 if (unlikely(s.blocked_rdev)) {
4322 if (s.syncing || s.expanding || s.expanded ||
4323 s.replacing || s.to_write || s.written) {
4324 set_bit(STRIPE_HANDLE, &sh->state);
4327 /* There is nothing for the blocked_rdev to block */
4328 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4329 s.blocked_rdev = NULL;
4332 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4333 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4334 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4337 pr_debug("locked=%d uptodate=%d to_read=%d"
4338 " to_write=%d failed=%d failed_num=%d,%d\n",
4339 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4340 s.failed_num[0], s.failed_num[1]);
4341 /* check if the array has lost more than max_degraded devices and,
4342 * if so, some requests might need to be failed.
4344 if (s.failed > conf->max_degraded) {
4345 sh->check_state = 0;
4346 sh->reconstruct_state = 0;
4347 break_stripe_batch_list(sh, 0);
4348 if (s.to_read+s.to_write+s.written)
4349 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4350 if (s.syncing + s.replacing)
4351 handle_failed_sync(conf, sh, &s);
4354 /* Now we check to see if any write operations have recently
4358 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4360 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4361 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4362 sh->reconstruct_state = reconstruct_state_idle;
4364 /* All the 'written' buffers and the parity block are ready to
4365 * be written back to disk
4367 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4368 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4369 BUG_ON(sh->qd_idx >= 0 &&
4370 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4371 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4372 for (i = disks; i--; ) {
4373 struct r5dev *dev = &sh->dev[i];
4374 if (test_bit(R5_LOCKED, &dev->flags) &&
4375 (i == sh->pd_idx || i == sh->qd_idx ||
4377 pr_debug("Writing block %d\n", i);
4378 set_bit(R5_Wantwrite, &dev->flags);
4383 if (!test_bit(R5_Insync, &dev->flags) ||
4384 ((i == sh->pd_idx || i == sh->qd_idx) &&
4386 set_bit(STRIPE_INSYNC, &sh->state);
4389 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4390 s.dec_preread_active = 1;
4394 * might be able to return some write requests if the parity blocks
4395 * are safe, or on a failed drive
4397 pdev = &sh->dev[sh->pd_idx];
4398 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4399 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4400 qdev = &sh->dev[sh->qd_idx];
4401 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4402 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4406 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4407 && !test_bit(R5_LOCKED, &pdev->flags)
4408 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4409 test_bit(R5_Discard, &pdev->flags))))) &&
4410 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4411 && !test_bit(R5_LOCKED, &qdev->flags)
4412 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4413 test_bit(R5_Discard, &qdev->flags))))))
4414 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4416 /* Now we might consider reading some blocks, either to check/generate
4417 * parity, or to satisfy requests
4418 * or to load a block that is being partially written.
4420 if (s.to_read || s.non_overwrite
4421 || (conf->level == 6 && s.to_write && s.failed)
4422 || (s.syncing && (s.uptodate + s.compute < disks))
4425 handle_stripe_fill(sh, &s, disks);
4427 /* Now to consider new write requests and what else, if anything
4428 * should be read. We do not handle new writes when:
4429 * 1/ A 'write' operation (copy+xor) is already in flight.
4430 * 2/ A 'check' operation is in flight, as it may clobber the parity
4433 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4434 handle_stripe_dirtying(conf, sh, &s, disks);
4436 /* maybe we need to check and possibly fix the parity for this stripe
4437 * Any reads will already have been scheduled, so we just see if enough
4438 * data is available. The parity check is held off while parity
4439 * dependent operations are in flight.
4441 if (sh->check_state ||
4442 (s.syncing && s.locked == 0 &&
4443 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4444 !test_bit(STRIPE_INSYNC, &sh->state))) {
4445 if (conf->level == 6)
4446 handle_parity_checks6(conf, sh, &s, disks);
4448 handle_parity_checks5(conf, sh, &s, disks);
4451 if ((s.replacing || s.syncing) && s.locked == 0
4452 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4453 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4454 /* Write out to replacement devices where possible */
4455 for (i = 0; i < conf->raid_disks; i++)
4456 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4457 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4458 set_bit(R5_WantReplace, &sh->dev[i].flags);
4459 set_bit(R5_LOCKED, &sh->dev[i].flags);
4463 set_bit(STRIPE_INSYNC, &sh->state);
4464 set_bit(STRIPE_REPLACED, &sh->state);
4466 if ((s.syncing || s.replacing) && s.locked == 0 &&
4467 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4468 test_bit(STRIPE_INSYNC, &sh->state)) {
4469 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4470 clear_bit(STRIPE_SYNCING, &sh->state);
4471 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4472 wake_up(&conf->wait_for_overlap);
4475 /* If the failed drives are just a ReadError, then we might need
4476 * to progress the repair/check process
4478 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4479 for (i = 0; i < s.failed; i++) {
4480 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4481 if (test_bit(R5_ReadError, &dev->flags)
4482 && !test_bit(R5_LOCKED, &dev->flags)
4483 && test_bit(R5_UPTODATE, &dev->flags)
4485 if (!test_bit(R5_ReWrite, &dev->flags)) {
4486 set_bit(R5_Wantwrite, &dev->flags);
4487 set_bit(R5_ReWrite, &dev->flags);
4488 set_bit(R5_LOCKED, &dev->flags);
4491 /* let's read it back */
4492 set_bit(R5_Wantread, &dev->flags);
4493 set_bit(R5_LOCKED, &dev->flags);
4499 /* Finish reconstruct operations initiated by the expansion process */
4500 if (sh->reconstruct_state == reconstruct_state_result) {
4501 struct stripe_head *sh_src
4502 = get_active_stripe(conf, sh->sector, 1, 1, 1);
4503 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4504 /* sh cannot be written until sh_src has been read.
4505 * so arrange for sh to be delayed a little
4507 set_bit(STRIPE_DELAYED, &sh->state);
4508 set_bit(STRIPE_HANDLE, &sh->state);
4509 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4511 atomic_inc(&conf->preread_active_stripes);
4512 release_stripe(sh_src);
4516 release_stripe(sh_src);
4518 sh->reconstruct_state = reconstruct_state_idle;
4519 clear_bit(STRIPE_EXPANDING, &sh->state);
4520 for (i = conf->raid_disks; i--; ) {
4521 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4522 set_bit(R5_LOCKED, &sh->dev[i].flags);
4527 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4528 !sh->reconstruct_state) {
4529 /* Need to write out all blocks after computing parity */
4530 sh->disks = conf->raid_disks;
4531 stripe_set_idx(sh->sector, conf, 0, sh);
4532 schedule_reconstruction(sh, &s, 1, 1);
4533 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4534 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4535 atomic_dec(&conf->reshape_stripes);
4536 wake_up(&conf->wait_for_overlap);
4537 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4540 if (s.expanding && s.locked == 0 &&
4541 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4542 handle_stripe_expansion(conf, sh);
4545 /* wait for this device to become unblocked */
4546 if (unlikely(s.blocked_rdev)) {
4547 if (conf->mddev->external)
4548 md_wait_for_blocked_rdev(s.blocked_rdev,
4551 /* Internal metadata will immediately
4552 * be written by raid5d, so we don't
4553 * need to wait here.
4555 rdev_dec_pending(s.blocked_rdev,
4559 if (s.handle_bad_blocks)
4560 for (i = disks; i--; ) {
4561 struct md_rdev *rdev;
4562 struct r5dev *dev = &sh->dev[i];
4563 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4564 /* We own a safe reference to the rdev */
4565 rdev = conf->disks[i].rdev;
4566 if (!rdev_set_badblocks(rdev, sh->sector,
4568 md_error(conf->mddev, rdev);
4569 rdev_dec_pending(rdev, conf->mddev);
4571 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4572 rdev = conf->disks[i].rdev;
4573 rdev_clear_badblocks(rdev, sh->sector,
4575 rdev_dec_pending(rdev, conf->mddev);
4577 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4578 rdev = conf->disks[i].replacement;
4580 /* rdev have been moved down */
4581 rdev = conf->disks[i].rdev;
4582 rdev_clear_badblocks(rdev, sh->sector,
4584 rdev_dec_pending(rdev, conf->mddev);
4589 raid_run_ops(sh, s.ops_request);
4593 if (s.dec_preread_active) {
4594 /* We delay this until after ops_run_io so that if make_request
4595 * is waiting on a flush, it won't continue until the writes
4596 * have actually been submitted.
4598 atomic_dec(&conf->preread_active_stripes);
4599 if (atomic_read(&conf->preread_active_stripes) <
4601 md_wakeup_thread(conf->mddev->thread);
4604 return_io(s.return_bi);
4606 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4609 static void raid5_activate_delayed(struct r5conf *conf)
4611 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4612 while (!list_empty(&conf->delayed_list)) {
4613 struct list_head *l = conf->delayed_list.next;
4614 struct stripe_head *sh;
4615 sh = list_entry(l, struct stripe_head, lru);
4617 clear_bit(STRIPE_DELAYED, &sh->state);
4618 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4619 atomic_inc(&conf->preread_active_stripes);
4620 list_add_tail(&sh->lru, &conf->hold_list);
4621 raid5_wakeup_stripe_thread(sh);
4626 static void activate_bit_delay(struct r5conf *conf,
4627 struct list_head *temp_inactive_list)
4629 /* device_lock is held */
4630 struct list_head head;
4631 list_add(&head, &conf->bitmap_list);
4632 list_del_init(&conf->bitmap_list);
4633 while (!list_empty(&head)) {
4634 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4636 list_del_init(&sh->lru);
4637 atomic_inc(&sh->count);
4638 hash = sh->hash_lock_index;
4639 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4643 static int raid5_congested(struct mddev *mddev, int bits)
4645 struct r5conf *conf = mddev->private;
4647 /* No difference between reads and writes. Just check
4648 * how busy the stripe_cache is
4651 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4655 if (atomic_read(&conf->empty_inactive_list_nr))
4661 /* We want read requests to align with chunks where possible,
4662 * but write requests don't need to.
4664 static int raid5_mergeable_bvec(struct mddev *mddev,
4665 struct bvec_merge_data *bvm,
4666 struct bio_vec *biovec)
4668 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4670 unsigned int chunk_sectors = mddev->chunk_sectors;
4671 unsigned int bio_sectors = bvm->bi_size >> 9;
4674 * always allow writes to be mergeable, read as well if array
4675 * is degraded as we'll go through stripe cache anyway.
4677 if ((bvm->bi_rw & 1) == WRITE || mddev->degraded)
4678 return biovec->bv_len;
4680 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4681 chunk_sectors = mddev->new_chunk_sectors;
4682 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4683 if (max < 0) max = 0;
4684 if (max <= biovec->bv_len && bio_sectors == 0)
4685 return biovec->bv_len;
4690 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4692 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4693 unsigned int chunk_sectors = mddev->chunk_sectors;
4694 unsigned int bio_sectors = bio_sectors(bio);
4696 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4697 chunk_sectors = mddev->new_chunk_sectors;
4698 return chunk_sectors >=
4699 ((sector & (chunk_sectors - 1)) + bio_sectors);
4703 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4704 * later sampled by raid5d.
4706 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4708 unsigned long flags;
4710 spin_lock_irqsave(&conf->device_lock, flags);
4712 bi->bi_next = conf->retry_read_aligned_list;
4713 conf->retry_read_aligned_list = bi;
4715 spin_unlock_irqrestore(&conf->device_lock, flags);
4716 md_wakeup_thread(conf->mddev->thread);
4719 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4723 bi = conf->retry_read_aligned;
4725 conf->retry_read_aligned = NULL;
4728 bi = conf->retry_read_aligned_list;
4730 conf->retry_read_aligned_list = bi->bi_next;
4733 * this sets the active strip count to 1 and the processed
4734 * strip count to zero (upper 8 bits)
4736 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4743 * The "raid5_align_endio" should check if the read succeeded and if it
4744 * did, call bio_endio on the original bio (having bio_put the new bio
4746 * If the read failed..
4748 static void raid5_align_endio(struct bio *bi, int error)
4750 struct bio* raid_bi = bi->bi_private;
4751 struct mddev *mddev;
4752 struct r5conf *conf;
4753 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4754 struct md_rdev *rdev;
4758 rdev = (void*)raid_bi->bi_next;
4759 raid_bi->bi_next = NULL;
4760 mddev = rdev->mddev;
4761 conf = mddev->private;
4763 rdev_dec_pending(rdev, conf->mddev);
4765 if (!error && uptodate) {
4766 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4768 bio_endio(raid_bi, 0);
4769 if (atomic_dec_and_test(&conf->active_aligned_reads))
4770 wake_up(&conf->wait_for_stripe);
4774 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4776 add_bio_to_retry(raid_bi, conf);
4779 static int bio_fits_rdev(struct bio *bi)
4781 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4783 if (bio_sectors(bi) > queue_max_sectors(q))
4785 blk_recount_segments(q, bi);
4786 if (bi->bi_phys_segments > queue_max_segments(q))
4789 if (q->merge_bvec_fn)
4790 /* it's too hard to apply the merge_bvec_fn at this stage,
4798 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4800 struct r5conf *conf = mddev->private;
4802 struct bio* align_bi;
4803 struct md_rdev *rdev;
4804 sector_t end_sector;
4806 if (!in_chunk_boundary(mddev, raid_bio)) {
4807 pr_debug("chunk_aligned_read : non aligned\n");
4811 * use bio_clone_mddev to make a copy of the bio
4813 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4817 * set bi_end_io to a new function, and set bi_private to the
4820 align_bi->bi_end_io = raid5_align_endio;
4821 align_bi->bi_private = raid_bio;
4825 align_bi->bi_iter.bi_sector =
4826 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4829 end_sector = bio_end_sector(align_bi);
4831 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4832 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4833 rdev->recovery_offset < end_sector) {
4834 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4836 (test_bit(Faulty, &rdev->flags) ||
4837 !(test_bit(In_sync, &rdev->flags) ||
4838 rdev->recovery_offset >= end_sector)))
4845 atomic_inc(&rdev->nr_pending);
4847 raid_bio->bi_next = (void*)rdev;
4848 align_bi->bi_bdev = rdev->bdev;
4849 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
4851 if (!bio_fits_rdev(align_bi) ||
4852 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4853 bio_sectors(align_bi),
4854 &first_bad, &bad_sectors)) {
4855 /* too big in some way, or has a known bad block */
4857 rdev_dec_pending(rdev, mddev);
4861 /* No reshape active, so we can trust rdev->data_offset */
4862 align_bi->bi_iter.bi_sector += rdev->data_offset;
4864 spin_lock_irq(&conf->device_lock);
4865 wait_event_lock_irq(conf->wait_for_stripe,
4868 atomic_inc(&conf->active_aligned_reads);
4869 spin_unlock_irq(&conf->device_lock);
4872 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4873 align_bi, disk_devt(mddev->gendisk),
4874 raid_bio->bi_iter.bi_sector);
4875 generic_make_request(align_bi);
4884 /* __get_priority_stripe - get the next stripe to process
4886 * Full stripe writes are allowed to pass preread active stripes up until
4887 * the bypass_threshold is exceeded. In general the bypass_count
4888 * increments when the handle_list is handled before the hold_list; however, it
4889 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4890 * stripe with in flight i/o. The bypass_count will be reset when the
4891 * head of the hold_list has changed, i.e. the head was promoted to the
4894 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4896 struct stripe_head *sh = NULL, *tmp;
4897 struct list_head *handle_list = NULL;
4898 struct r5worker_group *wg = NULL;
4900 if (conf->worker_cnt_per_group == 0) {
4901 handle_list = &conf->handle_list;
4902 } else if (group != ANY_GROUP) {
4903 handle_list = &conf->worker_groups[group].handle_list;
4904 wg = &conf->worker_groups[group];
4907 for (i = 0; i < conf->group_cnt; i++) {
4908 handle_list = &conf->worker_groups[i].handle_list;
4909 wg = &conf->worker_groups[i];
4910 if (!list_empty(handle_list))
4915 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4917 list_empty(handle_list) ? "empty" : "busy",
4918 list_empty(&conf->hold_list) ? "empty" : "busy",
4919 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4921 if (!list_empty(handle_list)) {
4922 sh = list_entry(handle_list->next, typeof(*sh), lru);
4924 if (list_empty(&conf->hold_list))
4925 conf->bypass_count = 0;
4926 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4927 if (conf->hold_list.next == conf->last_hold)
4928 conf->bypass_count++;
4930 conf->last_hold = conf->hold_list.next;
4931 conf->bypass_count -= conf->bypass_threshold;
4932 if (conf->bypass_count < 0)
4933 conf->bypass_count = 0;
4936 } else if (!list_empty(&conf->hold_list) &&
4937 ((conf->bypass_threshold &&
4938 conf->bypass_count > conf->bypass_threshold) ||
4939 atomic_read(&conf->pending_full_writes) == 0)) {
4941 list_for_each_entry(tmp, &conf->hold_list, lru) {
4942 if (conf->worker_cnt_per_group == 0 ||
4943 group == ANY_GROUP ||
4944 !cpu_online(tmp->cpu) ||
4945 cpu_to_group(tmp->cpu) == group) {
4952 conf->bypass_count -= conf->bypass_threshold;
4953 if (conf->bypass_count < 0)
4954 conf->bypass_count = 0;
4966 list_del_init(&sh->lru);
4967 BUG_ON(atomic_inc_return(&sh->count) != 1);
4971 struct raid5_plug_cb {
4972 struct blk_plug_cb cb;
4973 struct list_head list;
4974 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4977 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4979 struct raid5_plug_cb *cb = container_of(
4980 blk_cb, struct raid5_plug_cb, cb);
4981 struct stripe_head *sh;
4982 struct mddev *mddev = cb->cb.data;
4983 struct r5conf *conf = mddev->private;
4987 if (cb->list.next && !list_empty(&cb->list)) {
4988 spin_lock_irq(&conf->device_lock);
4989 while (!list_empty(&cb->list)) {
4990 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4991 list_del_init(&sh->lru);
4993 * avoid race release_stripe_plug() sees
4994 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4995 * is still in our list
4997 smp_mb__before_atomic();
4998 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5000 * STRIPE_ON_RELEASE_LIST could be set here. In that
5001 * case, the count is always > 1 here
5003 hash = sh->hash_lock_index;
5004 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5007 spin_unlock_irq(&conf->device_lock);
5009 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5010 NR_STRIPE_HASH_LOCKS);
5012 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5016 static void release_stripe_plug(struct mddev *mddev,
5017 struct stripe_head *sh)
5019 struct blk_plug_cb *blk_cb = blk_check_plugged(
5020 raid5_unplug, mddev,
5021 sizeof(struct raid5_plug_cb));
5022 struct raid5_plug_cb *cb;
5029 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5031 if (cb->list.next == NULL) {
5033 INIT_LIST_HEAD(&cb->list);
5034 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5035 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5038 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5039 list_add_tail(&sh->lru, &cb->list);
5044 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5046 struct r5conf *conf = mddev->private;
5047 sector_t logical_sector, last_sector;
5048 struct stripe_head *sh;
5052 if (mddev->reshape_position != MaxSector)
5053 /* Skip discard while reshape is happening */
5056 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5057 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5060 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5062 stripe_sectors = conf->chunk_sectors *
5063 (conf->raid_disks - conf->max_degraded);
5064 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5066 sector_div(last_sector, stripe_sectors);
5068 logical_sector *= conf->chunk_sectors;
5069 last_sector *= conf->chunk_sectors;
5071 for (; logical_sector < last_sector;
5072 logical_sector += STRIPE_SECTORS) {
5076 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
5077 prepare_to_wait(&conf->wait_for_overlap, &w,
5078 TASK_UNINTERRUPTIBLE);
5079 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5080 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5085 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5086 spin_lock_irq(&sh->stripe_lock);
5087 for (d = 0; d < conf->raid_disks; d++) {
5088 if (d == sh->pd_idx || d == sh->qd_idx)
5090 if (sh->dev[d].towrite || sh->dev[d].toread) {
5091 set_bit(R5_Overlap, &sh->dev[d].flags);
5092 spin_unlock_irq(&sh->stripe_lock);
5098 set_bit(STRIPE_DISCARD, &sh->state);
5099 finish_wait(&conf->wait_for_overlap, &w);
5100 sh->overwrite_disks = 0;
5101 for (d = 0; d < conf->raid_disks; d++) {
5102 if (d == sh->pd_idx || d == sh->qd_idx)
5104 sh->dev[d].towrite = bi;
5105 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5106 raid5_inc_bi_active_stripes(bi);
5107 sh->overwrite_disks++;
5109 spin_unlock_irq(&sh->stripe_lock);
5110 if (conf->mddev->bitmap) {
5112 d < conf->raid_disks - conf->max_degraded;
5114 bitmap_startwrite(mddev->bitmap,
5118 sh->bm_seq = conf->seq_flush + 1;
5119 set_bit(STRIPE_BIT_DELAY, &sh->state);
5122 set_bit(STRIPE_HANDLE, &sh->state);
5123 clear_bit(STRIPE_DELAYED, &sh->state);
5124 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5125 atomic_inc(&conf->preread_active_stripes);
5126 release_stripe_plug(mddev, sh);
5129 remaining = raid5_dec_bi_active_stripes(bi);
5130 if (remaining == 0) {
5131 md_write_end(mddev);
5136 static void make_request(struct mddev *mddev, struct bio * bi)
5138 struct r5conf *conf = mddev->private;
5140 sector_t new_sector;
5141 sector_t logical_sector, last_sector;
5142 struct stripe_head *sh;
5143 const int rw = bio_data_dir(bi);
5148 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5149 md_flush_request(mddev, bi);
5153 md_write_start(mddev, bi);
5156 * If array is degraded, better not do chunk aligned read because
5157 * later we might have to read it again in order to reconstruct
5158 * data on failed drives.
5160 if (rw == READ && mddev->degraded == 0 &&
5161 mddev->reshape_position == MaxSector &&
5162 chunk_aligned_read(mddev,bi))
5165 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5166 make_discard_request(mddev, bi);
5170 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5171 last_sector = bio_end_sector(bi);
5173 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5175 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5176 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5182 seq = read_seqcount_begin(&conf->gen_lock);
5185 prepare_to_wait(&conf->wait_for_overlap, &w,
5186 TASK_UNINTERRUPTIBLE);
5187 if (unlikely(conf->reshape_progress != MaxSector)) {
5188 /* spinlock is needed as reshape_progress may be
5189 * 64bit on a 32bit platform, and so it might be
5190 * possible to see a half-updated value
5191 * Of course reshape_progress could change after
5192 * the lock is dropped, so once we get a reference
5193 * to the stripe that we think it is, we will have
5196 spin_lock_irq(&conf->device_lock);
5197 if (mddev->reshape_backwards
5198 ? logical_sector < conf->reshape_progress
5199 : logical_sector >= conf->reshape_progress) {
5202 if (mddev->reshape_backwards
5203 ? logical_sector < conf->reshape_safe
5204 : logical_sector >= conf->reshape_safe) {
5205 spin_unlock_irq(&conf->device_lock);
5211 spin_unlock_irq(&conf->device_lock);
5214 new_sector = raid5_compute_sector(conf, logical_sector,
5217 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5218 (unsigned long long)new_sector,
5219 (unsigned long long)logical_sector);
5221 sh = get_active_stripe(conf, new_sector, previous,
5222 (bi->bi_rw&RWA_MASK), 0);
5224 if (unlikely(previous)) {
5225 /* expansion might have moved on while waiting for a
5226 * stripe, so we must do the range check again.
5227 * Expansion could still move past after this
5228 * test, but as we are holding a reference to
5229 * 'sh', we know that if that happens,
5230 * STRIPE_EXPANDING will get set and the expansion
5231 * won't proceed until we finish with the stripe.
5234 spin_lock_irq(&conf->device_lock);
5235 if (mddev->reshape_backwards
5236 ? logical_sector >= conf->reshape_progress
5237 : logical_sector < conf->reshape_progress)
5238 /* mismatch, need to try again */
5240 spin_unlock_irq(&conf->device_lock);
5248 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5249 /* Might have got the wrong stripe_head
5257 logical_sector >= mddev->suspend_lo &&
5258 logical_sector < mddev->suspend_hi) {
5260 /* As the suspend_* range is controlled by
5261 * userspace, we want an interruptible
5264 flush_signals(current);
5265 prepare_to_wait(&conf->wait_for_overlap,
5266 &w, TASK_INTERRUPTIBLE);
5267 if (logical_sector >= mddev->suspend_lo &&
5268 logical_sector < mddev->suspend_hi) {
5275 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5276 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5277 /* Stripe is busy expanding or
5278 * add failed due to overlap. Flush everything
5281 md_wakeup_thread(mddev->thread);
5287 set_bit(STRIPE_HANDLE, &sh->state);
5288 clear_bit(STRIPE_DELAYED, &sh->state);
5289 if ((!sh->batch_head || sh == sh->batch_head) &&
5290 (bi->bi_rw & REQ_SYNC) &&
5291 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5292 atomic_inc(&conf->preread_active_stripes);
5293 release_stripe_plug(mddev, sh);
5295 /* cannot get stripe for read-ahead, just give-up */
5296 clear_bit(BIO_UPTODATE, &bi->bi_flags);
5300 finish_wait(&conf->wait_for_overlap, &w);
5302 remaining = raid5_dec_bi_active_stripes(bi);
5303 if (remaining == 0) {
5306 md_write_end(mddev);
5308 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5314 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5316 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5318 /* reshaping is quite different to recovery/resync so it is
5319 * handled quite separately ... here.
5321 * On each call to sync_request, we gather one chunk worth of
5322 * destination stripes and flag them as expanding.
5323 * Then we find all the source stripes and request reads.
5324 * As the reads complete, handle_stripe will copy the data
5325 * into the destination stripe and release that stripe.
5327 struct r5conf *conf = mddev->private;
5328 struct stripe_head *sh;
5329 sector_t first_sector, last_sector;
5330 int raid_disks = conf->previous_raid_disks;
5331 int data_disks = raid_disks - conf->max_degraded;
5332 int new_data_disks = conf->raid_disks - conf->max_degraded;
5335 sector_t writepos, readpos, safepos;
5336 sector_t stripe_addr;
5337 int reshape_sectors;
5338 struct list_head stripes;
5340 if (sector_nr == 0) {
5341 /* If restarting in the middle, skip the initial sectors */
5342 if (mddev->reshape_backwards &&
5343 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5344 sector_nr = raid5_size(mddev, 0, 0)
5345 - conf->reshape_progress;
5346 } else if (!mddev->reshape_backwards &&
5347 conf->reshape_progress > 0)
5348 sector_nr = conf->reshape_progress;
5349 sector_div(sector_nr, new_data_disks);
5351 mddev->curr_resync_completed = sector_nr;
5352 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5358 /* We need to process a full chunk at a time.
5359 * If old and new chunk sizes differ, we need to process the
5362 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
5363 reshape_sectors = mddev->new_chunk_sectors;
5365 reshape_sectors = mddev->chunk_sectors;
5367 /* We update the metadata at least every 10 seconds, or when
5368 * the data about to be copied would over-write the source of
5369 * the data at the front of the range. i.e. one new_stripe
5370 * along from reshape_progress new_maps to after where
5371 * reshape_safe old_maps to
5373 writepos = conf->reshape_progress;
5374 sector_div(writepos, new_data_disks);
5375 readpos = conf->reshape_progress;
5376 sector_div(readpos, data_disks);
5377 safepos = conf->reshape_safe;
5378 sector_div(safepos, data_disks);
5379 if (mddev->reshape_backwards) {
5380 writepos -= min_t(sector_t, reshape_sectors, writepos);
5381 readpos += reshape_sectors;
5382 safepos += reshape_sectors;
5384 writepos += reshape_sectors;
5385 readpos -= min_t(sector_t, reshape_sectors, readpos);
5386 safepos -= min_t(sector_t, reshape_sectors, safepos);
5389 /* Having calculated the 'writepos' possibly use it
5390 * to set 'stripe_addr' which is where we will write to.
5392 if (mddev->reshape_backwards) {
5393 BUG_ON(conf->reshape_progress == 0);
5394 stripe_addr = writepos;
5395 BUG_ON((mddev->dev_sectors &
5396 ~((sector_t)reshape_sectors - 1))
5397 - reshape_sectors - stripe_addr
5400 BUG_ON(writepos != sector_nr + reshape_sectors);
5401 stripe_addr = sector_nr;
5404 /* 'writepos' is the most advanced device address we might write.
5405 * 'readpos' is the least advanced device address we might read.
5406 * 'safepos' is the least address recorded in the metadata as having
5408 * If there is a min_offset_diff, these are adjusted either by
5409 * increasing the safepos/readpos if diff is negative, or
5410 * increasing writepos if diff is positive.
5411 * If 'readpos' is then behind 'writepos', there is no way that we can
5412 * ensure safety in the face of a crash - that must be done by userspace
5413 * making a backup of the data. So in that case there is no particular
5414 * rush to update metadata.
5415 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5416 * update the metadata to advance 'safepos' to match 'readpos' so that
5417 * we can be safe in the event of a crash.
5418 * So we insist on updating metadata if safepos is behind writepos and
5419 * readpos is beyond writepos.
5420 * In any case, update the metadata every 10 seconds.
5421 * Maybe that number should be configurable, but I'm not sure it is
5422 * worth it.... maybe it could be a multiple of safemode_delay???
5424 if (conf->min_offset_diff < 0) {
5425 safepos += -conf->min_offset_diff;
5426 readpos += -conf->min_offset_diff;
5428 writepos += conf->min_offset_diff;
5430 if ((mddev->reshape_backwards
5431 ? (safepos > writepos && readpos < writepos)
5432 : (safepos < writepos && readpos > writepos)) ||
5433 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5434 /* Cannot proceed until we've updated the superblock... */
5435 wait_event(conf->wait_for_overlap,
5436 atomic_read(&conf->reshape_stripes)==0
5437 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5438 if (atomic_read(&conf->reshape_stripes) != 0)
5440 mddev->reshape_position = conf->reshape_progress;
5441 mddev->curr_resync_completed = sector_nr;
5442 conf->reshape_checkpoint = jiffies;
5443 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5444 md_wakeup_thread(mddev->thread);
5445 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5446 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5447 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5449 spin_lock_irq(&conf->device_lock);
5450 conf->reshape_safe = mddev->reshape_position;
5451 spin_unlock_irq(&conf->device_lock);
5452 wake_up(&conf->wait_for_overlap);
5453 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5456 INIT_LIST_HEAD(&stripes);
5457 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5459 int skipped_disk = 0;
5460 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5461 set_bit(STRIPE_EXPANDING, &sh->state);
5462 atomic_inc(&conf->reshape_stripes);
5463 /* If any of this stripe is beyond the end of the old
5464 * array, then we need to zero those blocks
5466 for (j=sh->disks; j--;) {
5468 if (j == sh->pd_idx)
5470 if (conf->level == 6 &&
5473 s = compute_blocknr(sh, j, 0);
5474 if (s < raid5_size(mddev, 0, 0)) {
5478 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5479 set_bit(R5_Expanded, &sh->dev[j].flags);
5480 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5482 if (!skipped_disk) {
5483 set_bit(STRIPE_EXPAND_READY, &sh->state);
5484 set_bit(STRIPE_HANDLE, &sh->state);
5486 list_add(&sh->lru, &stripes);
5488 spin_lock_irq(&conf->device_lock);
5489 if (mddev->reshape_backwards)
5490 conf->reshape_progress -= reshape_sectors * new_data_disks;
5492 conf->reshape_progress += reshape_sectors * new_data_disks;
5493 spin_unlock_irq(&conf->device_lock);
5494 /* Ok, those stripe are ready. We can start scheduling
5495 * reads on the source stripes.
5496 * The source stripes are determined by mapping the first and last
5497 * block on the destination stripes.
5500 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5503 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5504 * new_data_disks - 1),
5506 if (last_sector >= mddev->dev_sectors)
5507 last_sector = mddev->dev_sectors - 1;
5508 while (first_sector <= last_sector) {
5509 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
5510 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5511 set_bit(STRIPE_HANDLE, &sh->state);
5513 first_sector += STRIPE_SECTORS;
5515 /* Now that the sources are clearly marked, we can release
5516 * the destination stripes
5518 while (!list_empty(&stripes)) {
5519 sh = list_entry(stripes.next, struct stripe_head, lru);
5520 list_del_init(&sh->lru);
5523 /* If this takes us to the resync_max point where we have to pause,
5524 * then we need to write out the superblock.
5526 sector_nr += reshape_sectors;
5527 if ((sector_nr - mddev->curr_resync_completed) * 2
5528 >= mddev->resync_max - mddev->curr_resync_completed) {
5529 /* Cannot proceed until we've updated the superblock... */
5530 wait_event(conf->wait_for_overlap,
5531 atomic_read(&conf->reshape_stripes) == 0
5532 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5533 if (atomic_read(&conf->reshape_stripes) != 0)
5535 mddev->reshape_position = conf->reshape_progress;
5536 mddev->curr_resync_completed = sector_nr;
5537 conf->reshape_checkpoint = jiffies;
5538 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5539 md_wakeup_thread(mddev->thread);
5540 wait_event(mddev->sb_wait,
5541 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5542 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5543 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5545 spin_lock_irq(&conf->device_lock);
5546 conf->reshape_safe = mddev->reshape_position;
5547 spin_unlock_irq(&conf->device_lock);
5548 wake_up(&conf->wait_for_overlap);
5549 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5552 return reshape_sectors;
5555 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5557 struct r5conf *conf = mddev->private;
5558 struct stripe_head *sh;
5559 sector_t max_sector = mddev->dev_sectors;
5560 sector_t sync_blocks;
5561 int still_degraded = 0;
5564 if (sector_nr >= max_sector) {
5565 /* just being told to finish up .. nothing much to do */
5567 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5572 if (mddev->curr_resync < max_sector) /* aborted */
5573 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5575 else /* completed sync */
5577 bitmap_close_sync(mddev->bitmap);
5582 /* Allow raid5_quiesce to complete */
5583 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5585 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5586 return reshape_request(mddev, sector_nr, skipped);
5588 /* No need to check resync_max as we never do more than one
5589 * stripe, and as resync_max will always be on a chunk boundary,
5590 * if the check in md_do_sync didn't fire, there is no chance
5591 * of overstepping resync_max here
5594 /* if there is too many failed drives and we are trying
5595 * to resync, then assert that we are finished, because there is
5596 * nothing we can do.
5598 if (mddev->degraded >= conf->max_degraded &&
5599 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5600 sector_t rv = mddev->dev_sectors - sector_nr;
5604 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5606 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5607 sync_blocks >= STRIPE_SECTORS) {
5608 /* we can skip this block, and probably more */
5609 sync_blocks /= STRIPE_SECTORS;
5611 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5614 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5616 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5618 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5619 /* make sure we don't swamp the stripe cache if someone else
5620 * is trying to get access
5622 schedule_timeout_uninterruptible(1);
5624 /* Need to check if array will still be degraded after recovery/resync
5625 * Note in case of > 1 drive failures it's possible we're rebuilding
5626 * one drive while leaving another faulty drive in array.
5629 for (i = 0; i < conf->raid_disks; i++) {
5630 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5632 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5637 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5639 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5640 set_bit(STRIPE_HANDLE, &sh->state);
5644 return STRIPE_SECTORS;
5647 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5649 /* We may not be able to submit a whole bio at once as there
5650 * may not be enough stripe_heads available.
5651 * We cannot pre-allocate enough stripe_heads as we may need
5652 * more than exist in the cache (if we allow ever large chunks).
5653 * So we do one stripe head at a time and record in
5654 * ->bi_hw_segments how many have been done.
5656 * We *know* that this entire raid_bio is in one chunk, so
5657 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5659 struct stripe_head *sh;
5661 sector_t sector, logical_sector, last_sector;
5666 logical_sector = raid_bio->bi_iter.bi_sector &
5667 ~((sector_t)STRIPE_SECTORS-1);
5668 sector = raid5_compute_sector(conf, logical_sector,
5670 last_sector = bio_end_sector(raid_bio);
5672 for (; logical_sector < last_sector;
5673 logical_sector += STRIPE_SECTORS,
5674 sector += STRIPE_SECTORS,
5677 if (scnt < raid5_bi_processed_stripes(raid_bio))
5678 /* already done this stripe */
5681 sh = get_active_stripe(conf, sector, 0, 1, 1);
5684 /* failed to get a stripe - must wait */
5685 raid5_set_bi_processed_stripes(raid_bio, scnt);
5686 conf->retry_read_aligned = raid_bio;
5690 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5692 raid5_set_bi_processed_stripes(raid_bio, scnt);
5693 conf->retry_read_aligned = raid_bio;
5697 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5702 remaining = raid5_dec_bi_active_stripes(raid_bio);
5703 if (remaining == 0) {
5704 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5706 bio_endio(raid_bio, 0);
5708 if (atomic_dec_and_test(&conf->active_aligned_reads))
5709 wake_up(&conf->wait_for_stripe);
5713 static int handle_active_stripes(struct r5conf *conf, int group,
5714 struct r5worker *worker,
5715 struct list_head *temp_inactive_list)
5717 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5718 int i, batch_size = 0, hash;
5719 bool release_inactive = false;
5721 while (batch_size < MAX_STRIPE_BATCH &&
5722 (sh = __get_priority_stripe(conf, group)) != NULL)
5723 batch[batch_size++] = sh;
5725 if (batch_size == 0) {
5726 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5727 if (!list_empty(temp_inactive_list + i))
5729 if (i == NR_STRIPE_HASH_LOCKS)
5731 release_inactive = true;
5733 spin_unlock_irq(&conf->device_lock);
5735 release_inactive_stripe_list(conf, temp_inactive_list,
5736 NR_STRIPE_HASH_LOCKS);
5738 if (release_inactive) {
5739 spin_lock_irq(&conf->device_lock);
5743 for (i = 0; i < batch_size; i++)
5744 handle_stripe(batch[i]);
5748 spin_lock_irq(&conf->device_lock);
5749 for (i = 0; i < batch_size; i++) {
5750 hash = batch[i]->hash_lock_index;
5751 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5756 static void raid5_do_work(struct work_struct *work)
5758 struct r5worker *worker = container_of(work, struct r5worker, work);
5759 struct r5worker_group *group = worker->group;
5760 struct r5conf *conf = group->conf;
5761 int group_id = group - conf->worker_groups;
5763 struct blk_plug plug;
5765 pr_debug("+++ raid5worker active\n");
5767 blk_start_plug(&plug);
5769 spin_lock_irq(&conf->device_lock);
5771 int batch_size, released;
5773 released = release_stripe_list(conf, worker->temp_inactive_list);
5775 batch_size = handle_active_stripes(conf, group_id, worker,
5776 worker->temp_inactive_list);
5777 worker->working = false;
5778 if (!batch_size && !released)
5780 handled += batch_size;
5782 pr_debug("%d stripes handled\n", handled);
5784 spin_unlock_irq(&conf->device_lock);
5785 blk_finish_plug(&plug);
5787 pr_debug("--- raid5worker inactive\n");
5791 * This is our raid5 kernel thread.
5793 * We scan the hash table for stripes which can be handled now.
5794 * During the scan, completed stripes are saved for us by the interrupt
5795 * handler, so that they will not have to wait for our next wakeup.
5797 static void raid5d(struct md_thread *thread)
5799 struct mddev *mddev = thread->mddev;
5800 struct r5conf *conf = mddev->private;
5802 struct blk_plug plug;
5804 pr_debug("+++ raid5d active\n");
5806 md_check_recovery(mddev);
5808 blk_start_plug(&plug);
5810 spin_lock_irq(&conf->device_lock);
5813 int batch_size, released;
5815 released = release_stripe_list(conf, conf->temp_inactive_list);
5817 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5820 !list_empty(&conf->bitmap_list)) {
5821 /* Now is a good time to flush some bitmap updates */
5823 spin_unlock_irq(&conf->device_lock);
5824 bitmap_unplug(mddev->bitmap);
5825 spin_lock_irq(&conf->device_lock);
5826 conf->seq_write = conf->seq_flush;
5827 activate_bit_delay(conf, conf->temp_inactive_list);
5829 raid5_activate_delayed(conf);
5831 while ((bio = remove_bio_from_retry(conf))) {
5833 spin_unlock_irq(&conf->device_lock);
5834 ok = retry_aligned_read(conf, bio);
5835 spin_lock_irq(&conf->device_lock);
5841 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5842 conf->temp_inactive_list);
5843 if (!batch_size && !released)
5845 handled += batch_size;
5847 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5848 spin_unlock_irq(&conf->device_lock);
5849 md_check_recovery(mddev);
5850 spin_lock_irq(&conf->device_lock);
5853 pr_debug("%d stripes handled\n", handled);
5855 spin_unlock_irq(&conf->device_lock);
5856 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5857 mutex_trylock(&conf->cache_size_mutex)) {
5858 grow_one_stripe(conf, __GFP_NOWARN);
5859 /* Set flag even if allocation failed. This helps
5860 * slow down allocation requests when mem is short
5862 set_bit(R5_DID_ALLOC, &conf->cache_state);
5863 mutex_unlock(&conf->cache_size_mutex);
5866 async_tx_issue_pending_all();
5867 blk_finish_plug(&plug);
5869 pr_debug("--- raid5d inactive\n");
5873 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5875 struct r5conf *conf;
5877 spin_lock(&mddev->lock);
5878 conf = mddev->private;
5880 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5881 spin_unlock(&mddev->lock);
5886 raid5_set_cache_size(struct mddev *mddev, int size)
5888 struct r5conf *conf = mddev->private;
5891 if (size <= 16 || size > 32768)
5894 conf->min_nr_stripes = size;
5895 mutex_lock(&conf->cache_size_mutex);
5896 while (size < conf->max_nr_stripes &&
5897 drop_one_stripe(conf))
5899 mutex_unlock(&conf->cache_size_mutex);
5902 err = md_allow_write(mddev);
5906 mutex_lock(&conf->cache_size_mutex);
5907 while (size > conf->max_nr_stripes)
5908 if (!grow_one_stripe(conf, GFP_KERNEL))
5910 mutex_unlock(&conf->cache_size_mutex);
5914 EXPORT_SYMBOL(raid5_set_cache_size);
5917 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5919 struct r5conf *conf;
5923 if (len >= PAGE_SIZE)
5925 if (kstrtoul(page, 10, &new))
5927 err = mddev_lock(mddev);
5930 conf = mddev->private;
5934 err = raid5_set_cache_size(mddev, new);
5935 mddev_unlock(mddev);
5940 static struct md_sysfs_entry
5941 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5942 raid5_show_stripe_cache_size,
5943 raid5_store_stripe_cache_size);
5946 raid5_show_rmw_level(struct mddev *mddev, char *page)
5948 struct r5conf *conf = mddev->private;
5950 return sprintf(page, "%d\n", conf->rmw_level);
5956 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
5958 struct r5conf *conf = mddev->private;
5964 if (len >= PAGE_SIZE)
5967 if (kstrtoul(page, 10, &new))
5970 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
5973 if (new != PARITY_DISABLE_RMW &&
5974 new != PARITY_ENABLE_RMW &&
5975 new != PARITY_PREFER_RMW)
5978 conf->rmw_level = new;
5982 static struct md_sysfs_entry
5983 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
5984 raid5_show_rmw_level,
5985 raid5_store_rmw_level);
5989 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5991 struct r5conf *conf;
5993 spin_lock(&mddev->lock);
5994 conf = mddev->private;
5996 ret = sprintf(page, "%d\n", conf->bypass_threshold);
5997 spin_unlock(&mddev->lock);
6002 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6004 struct r5conf *conf;
6008 if (len >= PAGE_SIZE)
6010 if (kstrtoul(page, 10, &new))
6013 err = mddev_lock(mddev);
6016 conf = mddev->private;
6019 else if (new > conf->min_nr_stripes)
6022 conf->bypass_threshold = new;
6023 mddev_unlock(mddev);
6027 static struct md_sysfs_entry
6028 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6030 raid5_show_preread_threshold,
6031 raid5_store_preread_threshold);
6034 raid5_show_skip_copy(struct mddev *mddev, char *page)
6036 struct r5conf *conf;
6038 spin_lock(&mddev->lock);
6039 conf = mddev->private;
6041 ret = sprintf(page, "%d\n", conf->skip_copy);
6042 spin_unlock(&mddev->lock);
6047 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6049 struct r5conf *conf;
6053 if (len >= PAGE_SIZE)
6055 if (kstrtoul(page, 10, &new))
6059 err = mddev_lock(mddev);
6062 conf = mddev->private;
6065 else if (new != conf->skip_copy) {
6066 mddev_suspend(mddev);
6067 conf->skip_copy = new;
6069 mddev->queue->backing_dev_info.capabilities |=
6070 BDI_CAP_STABLE_WRITES;
6072 mddev->queue->backing_dev_info.capabilities &=
6073 ~BDI_CAP_STABLE_WRITES;
6074 mddev_resume(mddev);
6076 mddev_unlock(mddev);
6080 static struct md_sysfs_entry
6081 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6082 raid5_show_skip_copy,
6083 raid5_store_skip_copy);
6086 stripe_cache_active_show(struct mddev *mddev, char *page)
6088 struct r5conf *conf = mddev->private;
6090 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6095 static struct md_sysfs_entry
6096 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6099 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6101 struct r5conf *conf;
6103 spin_lock(&mddev->lock);
6104 conf = mddev->private;
6106 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6107 spin_unlock(&mddev->lock);
6111 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6113 int *worker_cnt_per_group,
6114 struct r5worker_group **worker_groups);
6116 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6118 struct r5conf *conf;
6121 struct r5worker_group *new_groups, *old_groups;
6122 int group_cnt, worker_cnt_per_group;
6124 if (len >= PAGE_SIZE)
6126 if (kstrtoul(page, 10, &new))
6129 err = mddev_lock(mddev);
6132 conf = mddev->private;
6135 else if (new != conf->worker_cnt_per_group) {
6136 mddev_suspend(mddev);
6138 old_groups = conf->worker_groups;
6140 flush_workqueue(raid5_wq);
6142 err = alloc_thread_groups(conf, new,
6143 &group_cnt, &worker_cnt_per_group,
6146 spin_lock_irq(&conf->device_lock);
6147 conf->group_cnt = group_cnt;
6148 conf->worker_cnt_per_group = worker_cnt_per_group;
6149 conf->worker_groups = new_groups;
6150 spin_unlock_irq(&conf->device_lock);
6153 kfree(old_groups[0].workers);
6156 mddev_resume(mddev);
6158 mddev_unlock(mddev);
6163 static struct md_sysfs_entry
6164 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6165 raid5_show_group_thread_cnt,
6166 raid5_store_group_thread_cnt);
6168 static struct attribute *raid5_attrs[] = {
6169 &raid5_stripecache_size.attr,
6170 &raid5_stripecache_active.attr,
6171 &raid5_preread_bypass_threshold.attr,
6172 &raid5_group_thread_cnt.attr,
6173 &raid5_skip_copy.attr,
6174 &raid5_rmw_level.attr,
6177 static struct attribute_group raid5_attrs_group = {
6179 .attrs = raid5_attrs,
6182 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6184 int *worker_cnt_per_group,
6185 struct r5worker_group **worker_groups)
6189 struct r5worker *workers;
6191 *worker_cnt_per_group = cnt;
6194 *worker_groups = NULL;
6197 *group_cnt = num_possible_nodes();
6198 size = sizeof(struct r5worker) * cnt;
6199 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6200 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6201 *group_cnt, GFP_NOIO);
6202 if (!*worker_groups || !workers) {
6204 kfree(*worker_groups);
6208 for (i = 0; i < *group_cnt; i++) {
6209 struct r5worker_group *group;
6211 group = &(*worker_groups)[i];
6212 INIT_LIST_HEAD(&group->handle_list);
6214 group->workers = workers + i * cnt;
6216 for (j = 0; j < cnt; j++) {
6217 struct r5worker *worker = group->workers + j;
6218 worker->group = group;
6219 INIT_WORK(&worker->work, raid5_do_work);
6221 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6222 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6229 static void free_thread_groups(struct r5conf *conf)
6231 if (conf->worker_groups)
6232 kfree(conf->worker_groups[0].workers);
6233 kfree(conf->worker_groups);
6234 conf->worker_groups = NULL;
6238 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6240 struct r5conf *conf = mddev->private;
6243 sectors = mddev->dev_sectors;
6245 /* size is defined by the smallest of previous and new size */
6246 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6248 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6249 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
6250 return sectors * (raid_disks - conf->max_degraded);
6253 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6255 safe_put_page(percpu->spare_page);
6256 if (percpu->scribble)
6257 flex_array_free(percpu->scribble);
6258 percpu->spare_page = NULL;
6259 percpu->scribble = NULL;
6262 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6264 if (conf->level == 6 && !percpu->spare_page)
6265 percpu->spare_page = alloc_page(GFP_KERNEL);
6266 if (!percpu->scribble)
6267 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6268 conf->previous_raid_disks),
6269 max(conf->chunk_sectors,
6270 conf->prev_chunk_sectors)
6274 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6275 free_scratch_buffer(conf, percpu);
6282 static void raid5_free_percpu(struct r5conf *conf)
6289 #ifdef CONFIG_HOTPLUG_CPU
6290 unregister_cpu_notifier(&conf->cpu_notify);
6294 for_each_possible_cpu(cpu)
6295 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6298 free_percpu(conf->percpu);
6301 static void free_conf(struct r5conf *conf)
6303 if (conf->shrinker.seeks)
6304 unregister_shrinker(&conf->shrinker);
6305 free_thread_groups(conf);
6306 shrink_stripes(conf);
6307 raid5_free_percpu(conf);
6309 kfree(conf->stripe_hashtbl);
6313 #ifdef CONFIG_HOTPLUG_CPU
6314 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6317 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6318 long cpu = (long)hcpu;
6319 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6322 case CPU_UP_PREPARE:
6323 case CPU_UP_PREPARE_FROZEN:
6324 if (alloc_scratch_buffer(conf, percpu)) {
6325 pr_err("%s: failed memory allocation for cpu%ld\n",
6327 return notifier_from_errno(-ENOMEM);
6331 case CPU_DEAD_FROZEN:
6332 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6341 static int raid5_alloc_percpu(struct r5conf *conf)
6346 conf->percpu = alloc_percpu(struct raid5_percpu);
6350 #ifdef CONFIG_HOTPLUG_CPU
6351 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6352 conf->cpu_notify.priority = 0;
6353 err = register_cpu_notifier(&conf->cpu_notify);
6359 for_each_present_cpu(cpu) {
6360 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6362 pr_err("%s: failed memory allocation for cpu%ld\n",
6366 spin_lock_init(&per_cpu_ptr(conf->percpu, cpu)->lock);
6373 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6374 struct shrink_control *sc)
6376 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6377 unsigned long ret = SHRINK_STOP;
6379 if (mutex_trylock(&conf->cache_size_mutex)) {
6381 while (ret < sc->nr_to_scan &&
6382 conf->max_nr_stripes > conf->min_nr_stripes) {
6383 if (drop_one_stripe(conf) == 0) {
6389 mutex_unlock(&conf->cache_size_mutex);
6394 static unsigned long raid5_cache_count(struct shrinker *shrink,
6395 struct shrink_control *sc)
6397 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6399 if (conf->max_nr_stripes < conf->min_nr_stripes)
6400 /* unlikely, but not impossible */
6402 return conf->max_nr_stripes - conf->min_nr_stripes;
6405 static struct r5conf *setup_conf(struct mddev *mddev)
6407 struct r5conf *conf;
6408 int raid_disk, memory, max_disks;
6409 struct md_rdev *rdev;
6410 struct disk_info *disk;
6413 int group_cnt, worker_cnt_per_group;
6414 struct r5worker_group *new_group;
6416 if (mddev->new_level != 5
6417 && mddev->new_level != 4
6418 && mddev->new_level != 6) {
6419 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6420 mdname(mddev), mddev->new_level);
6421 return ERR_PTR(-EIO);
6423 if ((mddev->new_level == 5
6424 && !algorithm_valid_raid5(mddev->new_layout)) ||
6425 (mddev->new_level == 6
6426 && !algorithm_valid_raid6(mddev->new_layout))) {
6427 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6428 mdname(mddev), mddev->new_layout);
6429 return ERR_PTR(-EIO);
6431 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6432 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6433 mdname(mddev), mddev->raid_disks);
6434 return ERR_PTR(-EINVAL);
6437 if (!mddev->new_chunk_sectors ||
6438 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6439 !is_power_of_2(mddev->new_chunk_sectors)) {
6440 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6441 mdname(mddev), mddev->new_chunk_sectors << 9);
6442 return ERR_PTR(-EINVAL);
6445 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6448 /* Don't enable multi-threading by default*/
6449 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6451 conf->group_cnt = group_cnt;
6452 conf->worker_cnt_per_group = worker_cnt_per_group;
6453 conf->worker_groups = new_group;
6456 spin_lock_init(&conf->device_lock);
6457 seqcount_init(&conf->gen_lock);
6458 mutex_init(&conf->cache_size_mutex);
6459 init_waitqueue_head(&conf->wait_for_stripe);
6460 init_waitqueue_head(&conf->wait_for_overlap);
6461 INIT_LIST_HEAD(&conf->handle_list);
6462 INIT_LIST_HEAD(&conf->hold_list);
6463 INIT_LIST_HEAD(&conf->delayed_list);
6464 INIT_LIST_HEAD(&conf->bitmap_list);
6465 init_llist_head(&conf->released_stripes);
6466 atomic_set(&conf->active_stripes, 0);
6467 atomic_set(&conf->preread_active_stripes, 0);
6468 atomic_set(&conf->active_aligned_reads, 0);
6469 conf->bypass_threshold = BYPASS_THRESHOLD;
6470 conf->recovery_disabled = mddev->recovery_disabled - 1;
6472 conf->raid_disks = mddev->raid_disks;
6473 if (mddev->reshape_position == MaxSector)
6474 conf->previous_raid_disks = mddev->raid_disks;
6476 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6477 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6479 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6484 conf->mddev = mddev;
6486 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6489 /* We init hash_locks[0] separately to that it can be used
6490 * as the reference lock in the spin_lock_nest_lock() call
6491 * in lock_all_device_hash_locks_irq in order to convince
6492 * lockdep that we know what we are doing.
6494 spin_lock_init(conf->hash_locks);
6495 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6496 spin_lock_init(conf->hash_locks + i);
6498 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6499 INIT_LIST_HEAD(conf->inactive_list + i);
6501 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6502 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6504 conf->level = mddev->new_level;
6505 conf->chunk_sectors = mddev->new_chunk_sectors;
6506 if (raid5_alloc_percpu(conf) != 0)
6509 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6511 rdev_for_each(rdev, mddev) {
6512 raid_disk = rdev->raid_disk;
6513 if (raid_disk >= max_disks
6516 disk = conf->disks + raid_disk;
6518 if (test_bit(Replacement, &rdev->flags)) {
6519 if (disk->replacement)
6521 disk->replacement = rdev;
6528 if (test_bit(In_sync, &rdev->flags)) {
6529 char b[BDEVNAME_SIZE];
6530 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6532 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6533 } else if (rdev->saved_raid_disk != raid_disk)
6534 /* Cannot rely on bitmap to complete recovery */
6538 conf->level = mddev->new_level;
6539 if (conf->level == 6) {
6540 conf->max_degraded = 2;
6541 if (raid6_call.xor_syndrome)
6542 conf->rmw_level = PARITY_ENABLE_RMW;
6544 conf->rmw_level = PARITY_DISABLE_RMW;
6546 conf->max_degraded = 1;
6547 conf->rmw_level = PARITY_ENABLE_RMW;
6549 conf->algorithm = mddev->new_layout;
6550 conf->reshape_progress = mddev->reshape_position;
6551 if (conf->reshape_progress != MaxSector) {
6552 conf->prev_chunk_sectors = mddev->chunk_sectors;
6553 conf->prev_algo = mddev->layout;
6556 conf->min_nr_stripes = NR_STRIPES;
6557 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6558 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6559 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6560 if (grow_stripes(conf, conf->min_nr_stripes)) {
6562 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6563 mdname(mddev), memory);
6566 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6567 mdname(mddev), memory);
6569 * Losing a stripe head costs more than the time to refill it,
6570 * it reduces the queue depth and so can hurt throughput.
6571 * So set it rather large, scaled by number of devices.
6573 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6574 conf->shrinker.scan_objects = raid5_cache_scan;
6575 conf->shrinker.count_objects = raid5_cache_count;
6576 conf->shrinker.batch = 128;
6577 conf->shrinker.flags = 0;
6578 register_shrinker(&conf->shrinker);
6580 sprintf(pers_name, "raid%d", mddev->new_level);
6581 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6582 if (!conf->thread) {
6584 "md/raid:%s: couldn't allocate thread.\n",
6594 return ERR_PTR(-EIO);
6596 return ERR_PTR(-ENOMEM);
6599 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6602 case ALGORITHM_PARITY_0:
6603 if (raid_disk < max_degraded)
6606 case ALGORITHM_PARITY_N:
6607 if (raid_disk >= raid_disks - max_degraded)
6610 case ALGORITHM_PARITY_0_6:
6611 if (raid_disk == 0 ||
6612 raid_disk == raid_disks - 1)
6615 case ALGORITHM_LEFT_ASYMMETRIC_6:
6616 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6617 case ALGORITHM_LEFT_SYMMETRIC_6:
6618 case ALGORITHM_RIGHT_SYMMETRIC_6:
6619 if (raid_disk == raid_disks - 1)
6625 static int run(struct mddev *mddev)
6627 struct r5conf *conf;
6628 int working_disks = 0;
6629 int dirty_parity_disks = 0;
6630 struct md_rdev *rdev;
6631 sector_t reshape_offset = 0;
6633 long long min_offset_diff = 0;
6636 if (mddev->recovery_cp != MaxSector)
6637 printk(KERN_NOTICE "md/raid:%s: not clean"
6638 " -- starting background reconstruction\n",
6641 rdev_for_each(rdev, mddev) {
6643 if (rdev->raid_disk < 0)
6645 diff = (rdev->new_data_offset - rdev->data_offset);
6647 min_offset_diff = diff;
6649 } else if (mddev->reshape_backwards &&
6650 diff < min_offset_diff)
6651 min_offset_diff = diff;
6652 else if (!mddev->reshape_backwards &&
6653 diff > min_offset_diff)
6654 min_offset_diff = diff;
6657 if (mddev->reshape_position != MaxSector) {
6658 /* Check that we can continue the reshape.
6659 * Difficulties arise if the stripe we would write to
6660 * next is at or after the stripe we would read from next.
6661 * For a reshape that changes the number of devices, this
6662 * is only possible for a very short time, and mdadm makes
6663 * sure that time appears to have past before assembling
6664 * the array. So we fail if that time hasn't passed.
6665 * For a reshape that keeps the number of devices the same
6666 * mdadm must be monitoring the reshape can keeping the
6667 * critical areas read-only and backed up. It will start
6668 * the array in read-only mode, so we check for that.
6670 sector_t here_new, here_old;
6672 int max_degraded = (mddev->level == 6 ? 2 : 1);
6674 if (mddev->new_level != mddev->level) {
6675 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6676 "required - aborting.\n",
6680 old_disks = mddev->raid_disks - mddev->delta_disks;
6681 /* reshape_position must be on a new-stripe boundary, and one
6682 * further up in new geometry must map after here in old
6685 here_new = mddev->reshape_position;
6686 if (sector_div(here_new, mddev->new_chunk_sectors *
6687 (mddev->raid_disks - max_degraded))) {
6688 printk(KERN_ERR "md/raid:%s: reshape_position not "
6689 "on a stripe boundary\n", mdname(mddev));
6692 reshape_offset = here_new * mddev->new_chunk_sectors;
6693 /* here_new is the stripe we will write to */
6694 here_old = mddev->reshape_position;
6695 sector_div(here_old, mddev->chunk_sectors *
6696 (old_disks-max_degraded));
6697 /* here_old is the first stripe that we might need to read
6699 if (mddev->delta_disks == 0) {
6700 if ((here_new * mddev->new_chunk_sectors !=
6701 here_old * mddev->chunk_sectors)) {
6702 printk(KERN_ERR "md/raid:%s: reshape position is"
6703 " confused - aborting\n", mdname(mddev));
6706 /* We cannot be sure it is safe to start an in-place
6707 * reshape. It is only safe if user-space is monitoring
6708 * and taking constant backups.
6709 * mdadm always starts a situation like this in
6710 * readonly mode so it can take control before
6711 * allowing any writes. So just check for that.
6713 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6714 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6715 /* not really in-place - so OK */;
6716 else if (mddev->ro == 0) {
6717 printk(KERN_ERR "md/raid:%s: in-place reshape "
6718 "must be started in read-only mode "
6723 } else if (mddev->reshape_backwards
6724 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6725 here_old * mddev->chunk_sectors)
6726 : (here_new * mddev->new_chunk_sectors >=
6727 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6728 /* Reading from the same stripe as writing to - bad */
6729 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6730 "auto-recovery - aborting.\n",
6734 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6736 /* OK, we should be able to continue; */
6738 BUG_ON(mddev->level != mddev->new_level);
6739 BUG_ON(mddev->layout != mddev->new_layout);
6740 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6741 BUG_ON(mddev->delta_disks != 0);
6744 if (mddev->private == NULL)
6745 conf = setup_conf(mddev);
6747 conf = mddev->private;
6750 return PTR_ERR(conf);
6752 conf->min_offset_diff = min_offset_diff;
6753 mddev->thread = conf->thread;
6754 conf->thread = NULL;
6755 mddev->private = conf;
6757 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6759 rdev = conf->disks[i].rdev;
6760 if (!rdev && conf->disks[i].replacement) {
6761 /* The replacement is all we have yet */
6762 rdev = conf->disks[i].replacement;
6763 conf->disks[i].replacement = NULL;
6764 clear_bit(Replacement, &rdev->flags);
6765 conf->disks[i].rdev = rdev;
6769 if (conf->disks[i].replacement &&
6770 conf->reshape_progress != MaxSector) {
6771 /* replacements and reshape simply do not mix. */
6772 printk(KERN_ERR "md: cannot handle concurrent "
6773 "replacement and reshape.\n");
6776 if (test_bit(In_sync, &rdev->flags)) {
6780 /* This disc is not fully in-sync. However if it
6781 * just stored parity (beyond the recovery_offset),
6782 * when we don't need to be concerned about the
6783 * array being dirty.
6784 * When reshape goes 'backwards', we never have
6785 * partially completed devices, so we only need
6786 * to worry about reshape going forwards.
6788 /* Hack because v0.91 doesn't store recovery_offset properly. */
6789 if (mddev->major_version == 0 &&
6790 mddev->minor_version > 90)
6791 rdev->recovery_offset = reshape_offset;
6793 if (rdev->recovery_offset < reshape_offset) {
6794 /* We need to check old and new layout */
6795 if (!only_parity(rdev->raid_disk,
6798 conf->max_degraded))
6801 if (!only_parity(rdev->raid_disk,
6803 conf->previous_raid_disks,
6804 conf->max_degraded))
6806 dirty_parity_disks++;
6810 * 0 for a fully functional array, 1 or 2 for a degraded array.
6812 mddev->degraded = calc_degraded(conf);
6814 if (has_failed(conf)) {
6815 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6816 " (%d/%d failed)\n",
6817 mdname(mddev), mddev->degraded, conf->raid_disks);
6821 /* device size must be a multiple of chunk size */
6822 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6823 mddev->resync_max_sectors = mddev->dev_sectors;
6825 if (mddev->degraded > dirty_parity_disks &&
6826 mddev->recovery_cp != MaxSector) {
6827 if (mddev->ok_start_degraded)
6829 "md/raid:%s: starting dirty degraded array"
6830 " - data corruption possible.\n",
6834 "md/raid:%s: cannot start dirty degraded array.\n",
6840 if (mddev->degraded == 0)
6841 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6842 " devices, algorithm %d\n", mdname(mddev), conf->level,
6843 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6846 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6847 " out of %d devices, algorithm %d\n",
6848 mdname(mddev), conf->level,
6849 mddev->raid_disks - mddev->degraded,
6850 mddev->raid_disks, mddev->new_layout);
6852 print_raid5_conf(conf);
6854 if (conf->reshape_progress != MaxSector) {
6855 conf->reshape_safe = conf->reshape_progress;
6856 atomic_set(&conf->reshape_stripes, 0);
6857 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6858 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6859 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6860 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6861 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6865 /* Ok, everything is just fine now */
6866 if (mddev->to_remove == &raid5_attrs_group)
6867 mddev->to_remove = NULL;
6868 else if (mddev->kobj.sd &&
6869 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6871 "raid5: failed to create sysfs attributes for %s\n",
6873 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6877 bool discard_supported = true;
6878 /* read-ahead size must cover two whole stripes, which
6879 * is 2 * (datadisks) * chunksize where 'n' is the
6880 * number of raid devices
6882 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6883 int stripe = data_disks *
6884 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6885 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6886 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6888 chunk_size = mddev->chunk_sectors << 9;
6889 blk_queue_io_min(mddev->queue, chunk_size);
6890 blk_queue_io_opt(mddev->queue, chunk_size *
6891 (conf->raid_disks - conf->max_degraded));
6892 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6894 * We can only discard a whole stripe. It doesn't make sense to
6895 * discard data disk but write parity disk
6897 stripe = stripe * PAGE_SIZE;
6898 /* Round up to power of 2, as discard handling
6899 * currently assumes that */
6900 while ((stripe-1) & stripe)
6901 stripe = (stripe | (stripe-1)) + 1;
6902 mddev->queue->limits.discard_alignment = stripe;
6903 mddev->queue->limits.discard_granularity = stripe;
6905 * unaligned part of discard request will be ignored, so can't
6906 * guarantee discard_zeroes_data
6908 mddev->queue->limits.discard_zeroes_data = 0;
6910 blk_queue_max_write_same_sectors(mddev->queue, 0);
6912 rdev_for_each(rdev, mddev) {
6913 disk_stack_limits(mddev->gendisk, rdev->bdev,
6914 rdev->data_offset << 9);
6915 disk_stack_limits(mddev->gendisk, rdev->bdev,
6916 rdev->new_data_offset << 9);
6918 * discard_zeroes_data is required, otherwise data
6919 * could be lost. Consider a scenario: discard a stripe
6920 * (the stripe could be inconsistent if
6921 * discard_zeroes_data is 0); write one disk of the
6922 * stripe (the stripe could be inconsistent again
6923 * depending on which disks are used to calculate
6924 * parity); the disk is broken; The stripe data of this
6927 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6928 !bdev_get_queue(rdev->bdev)->
6929 limits.discard_zeroes_data)
6930 discard_supported = false;
6931 /* Unfortunately, discard_zeroes_data is not currently
6932 * a guarantee - just a hint. So we only allow DISCARD
6933 * if the sysadmin has confirmed that only safe devices
6934 * are in use by setting a module parameter.
6936 if (!devices_handle_discard_safely) {
6937 if (discard_supported) {
6938 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6939 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6941 discard_supported = false;
6945 if (discard_supported &&
6946 mddev->queue->limits.max_discard_sectors >= stripe &&
6947 mddev->queue->limits.discard_granularity >= stripe)
6948 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6951 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6957 md_unregister_thread(&mddev->thread);
6958 print_raid5_conf(conf);
6960 mddev->private = NULL;
6961 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6965 static void raid5_free(struct mddev *mddev, void *priv)
6967 struct r5conf *conf = priv;
6970 mddev->to_remove = &raid5_attrs_group;
6973 static void status(struct seq_file *seq, struct mddev *mddev)
6975 struct r5conf *conf = mddev->private;
6978 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6979 mddev->chunk_sectors / 2, mddev->layout);
6980 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6981 for (i = 0; i < conf->raid_disks; i++)
6982 seq_printf (seq, "%s",
6983 conf->disks[i].rdev &&
6984 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6985 seq_printf (seq, "]");
6988 static void print_raid5_conf (struct r5conf *conf)
6991 struct disk_info *tmp;
6993 printk(KERN_DEBUG "RAID conf printout:\n");
6995 printk("(conf==NULL)\n");
6998 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7000 conf->raid_disks - conf->mddev->degraded);
7002 for (i = 0; i < conf->raid_disks; i++) {
7003 char b[BDEVNAME_SIZE];
7004 tmp = conf->disks + i;
7006 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7007 i, !test_bit(Faulty, &tmp->rdev->flags),
7008 bdevname(tmp->rdev->bdev, b));
7012 static int raid5_spare_active(struct mddev *mddev)
7015 struct r5conf *conf = mddev->private;
7016 struct disk_info *tmp;
7018 unsigned long flags;
7020 for (i = 0; i < conf->raid_disks; i++) {
7021 tmp = conf->disks + i;
7022 if (tmp->replacement
7023 && tmp->replacement->recovery_offset == MaxSector
7024 && !test_bit(Faulty, &tmp->replacement->flags)
7025 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7026 /* Replacement has just become active. */
7028 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7031 /* Replaced device not technically faulty,
7032 * but we need to be sure it gets removed
7033 * and never re-added.
7035 set_bit(Faulty, &tmp->rdev->flags);
7036 sysfs_notify_dirent_safe(
7037 tmp->rdev->sysfs_state);
7039 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7040 } else if (tmp->rdev
7041 && tmp->rdev->recovery_offset == MaxSector
7042 && !test_bit(Faulty, &tmp->rdev->flags)
7043 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7045 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7048 spin_lock_irqsave(&conf->device_lock, flags);
7049 mddev->degraded = calc_degraded(conf);
7050 spin_unlock_irqrestore(&conf->device_lock, flags);
7051 print_raid5_conf(conf);
7055 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7057 struct r5conf *conf = mddev->private;
7059 int number = rdev->raid_disk;
7060 struct md_rdev **rdevp;
7061 struct disk_info *p = conf->disks + number;
7063 print_raid5_conf(conf);
7064 if (rdev == p->rdev)
7066 else if (rdev == p->replacement)
7067 rdevp = &p->replacement;
7071 if (number >= conf->raid_disks &&
7072 conf->reshape_progress == MaxSector)
7073 clear_bit(In_sync, &rdev->flags);
7075 if (test_bit(In_sync, &rdev->flags) ||
7076 atomic_read(&rdev->nr_pending)) {
7080 /* Only remove non-faulty devices if recovery
7083 if (!test_bit(Faulty, &rdev->flags) &&
7084 mddev->recovery_disabled != conf->recovery_disabled &&
7085 !has_failed(conf) &&
7086 (!p->replacement || p->replacement == rdev) &&
7087 number < conf->raid_disks) {
7093 if (atomic_read(&rdev->nr_pending)) {
7094 /* lost the race, try later */
7097 } else if (p->replacement) {
7098 /* We must have just cleared 'rdev' */
7099 p->rdev = p->replacement;
7100 clear_bit(Replacement, &p->replacement->flags);
7101 smp_mb(); /* Make sure other CPUs may see both as identical
7102 * but will never see neither - if they are careful
7104 p->replacement = NULL;
7105 clear_bit(WantReplacement, &rdev->flags);
7107 /* We might have just removed the Replacement as faulty-
7108 * clear the bit just in case
7110 clear_bit(WantReplacement, &rdev->flags);
7113 print_raid5_conf(conf);
7117 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7119 struct r5conf *conf = mddev->private;
7122 struct disk_info *p;
7124 int last = conf->raid_disks - 1;
7126 if (mddev->recovery_disabled == conf->recovery_disabled)
7129 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7130 /* no point adding a device */
7133 if (rdev->raid_disk >= 0)
7134 first = last = rdev->raid_disk;
7137 * find the disk ... but prefer rdev->saved_raid_disk
7140 if (rdev->saved_raid_disk >= 0 &&
7141 rdev->saved_raid_disk >= first &&
7142 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7143 first = rdev->saved_raid_disk;
7145 for (disk = first; disk <= last; disk++) {
7146 p = conf->disks + disk;
7147 if (p->rdev == NULL) {
7148 clear_bit(In_sync, &rdev->flags);
7149 rdev->raid_disk = disk;
7151 if (rdev->saved_raid_disk != disk)
7153 rcu_assign_pointer(p->rdev, rdev);
7157 for (disk = first; disk <= last; disk++) {
7158 p = conf->disks + disk;
7159 if (test_bit(WantReplacement, &p->rdev->flags) &&
7160 p->replacement == NULL) {
7161 clear_bit(In_sync, &rdev->flags);
7162 set_bit(Replacement, &rdev->flags);
7163 rdev->raid_disk = disk;
7166 rcu_assign_pointer(p->replacement, rdev);
7171 print_raid5_conf(conf);
7175 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7177 /* no resync is happening, and there is enough space
7178 * on all devices, so we can resize.
7179 * We need to make sure resync covers any new space.
7180 * If the array is shrinking we should possibly wait until
7181 * any io in the removed space completes, but it hardly seems
7185 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7186 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7187 if (mddev->external_size &&
7188 mddev->array_sectors > newsize)
7190 if (mddev->bitmap) {
7191 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7195 md_set_array_sectors(mddev, newsize);
7196 set_capacity(mddev->gendisk, mddev->array_sectors);
7197 revalidate_disk(mddev->gendisk);
7198 if (sectors > mddev->dev_sectors &&
7199 mddev->recovery_cp > mddev->dev_sectors) {
7200 mddev->recovery_cp = mddev->dev_sectors;
7201 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7203 mddev->dev_sectors = sectors;
7204 mddev->resync_max_sectors = sectors;
7208 static int check_stripe_cache(struct mddev *mddev)
7210 /* Can only proceed if there are plenty of stripe_heads.
7211 * We need a minimum of one full stripe,, and for sensible progress
7212 * it is best to have about 4 times that.
7213 * If we require 4 times, then the default 256 4K stripe_heads will
7214 * allow for chunk sizes up to 256K, which is probably OK.
7215 * If the chunk size is greater, user-space should request more
7216 * stripe_heads first.
7218 struct r5conf *conf = mddev->private;
7219 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7220 > conf->min_nr_stripes ||
7221 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7222 > conf->min_nr_stripes) {
7223 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7225 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7232 static int check_reshape(struct mddev *mddev)
7234 struct r5conf *conf = mddev->private;
7236 if (mddev->delta_disks == 0 &&
7237 mddev->new_layout == mddev->layout &&
7238 mddev->new_chunk_sectors == mddev->chunk_sectors)
7239 return 0; /* nothing to do */
7240 if (has_failed(conf))
7242 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7243 /* We might be able to shrink, but the devices must
7244 * be made bigger first.
7245 * For raid6, 4 is the minimum size.
7246 * Otherwise 2 is the minimum
7249 if (mddev->level == 6)
7251 if (mddev->raid_disks + mddev->delta_disks < min)
7255 if (!check_stripe_cache(mddev))
7258 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7259 mddev->delta_disks > 0)
7260 if (resize_chunks(conf,
7261 conf->previous_raid_disks
7262 + max(0, mddev->delta_disks),
7263 max(mddev->new_chunk_sectors,
7264 mddev->chunk_sectors)
7267 return resize_stripes(conf, (conf->previous_raid_disks
7268 + mddev->delta_disks));
7271 static int raid5_start_reshape(struct mddev *mddev)
7273 struct r5conf *conf = mddev->private;
7274 struct md_rdev *rdev;
7276 unsigned long flags;
7278 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7281 if (!check_stripe_cache(mddev))
7284 if (has_failed(conf))
7287 rdev_for_each(rdev, mddev) {
7288 if (!test_bit(In_sync, &rdev->flags)
7289 && !test_bit(Faulty, &rdev->flags))
7293 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7294 /* Not enough devices even to make a degraded array
7299 /* Refuse to reduce size of the array. Any reductions in
7300 * array size must be through explicit setting of array_size
7303 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7304 < mddev->array_sectors) {
7305 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7306 "before number of disks\n", mdname(mddev));
7310 atomic_set(&conf->reshape_stripes, 0);
7311 spin_lock_irq(&conf->device_lock);
7312 write_seqcount_begin(&conf->gen_lock);
7313 conf->previous_raid_disks = conf->raid_disks;
7314 conf->raid_disks += mddev->delta_disks;
7315 conf->prev_chunk_sectors = conf->chunk_sectors;
7316 conf->chunk_sectors = mddev->new_chunk_sectors;
7317 conf->prev_algo = conf->algorithm;
7318 conf->algorithm = mddev->new_layout;
7320 /* Code that selects data_offset needs to see the generation update
7321 * if reshape_progress has been set - so a memory barrier needed.
7324 if (mddev->reshape_backwards)
7325 conf->reshape_progress = raid5_size(mddev, 0, 0);
7327 conf->reshape_progress = 0;
7328 conf->reshape_safe = conf->reshape_progress;
7329 write_seqcount_end(&conf->gen_lock);
7330 spin_unlock_irq(&conf->device_lock);
7332 /* Now make sure any requests that proceeded on the assumption
7333 * the reshape wasn't running - like Discard or Read - have
7336 mddev_suspend(mddev);
7337 mddev_resume(mddev);
7339 /* Add some new drives, as many as will fit.
7340 * We know there are enough to make the newly sized array work.
7341 * Don't add devices if we are reducing the number of
7342 * devices in the array. This is because it is not possible
7343 * to correctly record the "partially reconstructed" state of
7344 * such devices during the reshape and confusion could result.
7346 if (mddev->delta_disks >= 0) {
7347 rdev_for_each(rdev, mddev)
7348 if (rdev->raid_disk < 0 &&
7349 !test_bit(Faulty, &rdev->flags)) {
7350 if (raid5_add_disk(mddev, rdev) == 0) {
7352 >= conf->previous_raid_disks)
7353 set_bit(In_sync, &rdev->flags);
7355 rdev->recovery_offset = 0;
7357 if (sysfs_link_rdev(mddev, rdev))
7358 /* Failure here is OK */;
7360 } else if (rdev->raid_disk >= conf->previous_raid_disks
7361 && !test_bit(Faulty, &rdev->flags)) {
7362 /* This is a spare that was manually added */
7363 set_bit(In_sync, &rdev->flags);
7366 /* When a reshape changes the number of devices,
7367 * ->degraded is measured against the larger of the
7368 * pre and post number of devices.
7370 spin_lock_irqsave(&conf->device_lock, flags);
7371 mddev->degraded = calc_degraded(conf);
7372 spin_unlock_irqrestore(&conf->device_lock, flags);
7374 mddev->raid_disks = conf->raid_disks;
7375 mddev->reshape_position = conf->reshape_progress;
7376 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7378 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7379 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7380 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7381 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7382 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7383 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7385 if (!mddev->sync_thread) {
7386 mddev->recovery = 0;
7387 spin_lock_irq(&conf->device_lock);
7388 write_seqcount_begin(&conf->gen_lock);
7389 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7390 mddev->new_chunk_sectors =
7391 conf->chunk_sectors = conf->prev_chunk_sectors;
7392 mddev->new_layout = conf->algorithm = conf->prev_algo;
7393 rdev_for_each(rdev, mddev)
7394 rdev->new_data_offset = rdev->data_offset;
7396 conf->generation --;
7397 conf->reshape_progress = MaxSector;
7398 mddev->reshape_position = MaxSector;
7399 write_seqcount_end(&conf->gen_lock);
7400 spin_unlock_irq(&conf->device_lock);
7403 conf->reshape_checkpoint = jiffies;
7404 md_wakeup_thread(mddev->sync_thread);
7405 md_new_event(mddev);
7409 /* This is called from the reshape thread and should make any
7410 * changes needed in 'conf'
7412 static void end_reshape(struct r5conf *conf)
7415 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7416 struct md_rdev *rdev;
7418 spin_lock_irq(&conf->device_lock);
7419 conf->previous_raid_disks = conf->raid_disks;
7420 rdev_for_each(rdev, conf->mddev)
7421 rdev->data_offset = rdev->new_data_offset;
7423 conf->reshape_progress = MaxSector;
7424 spin_unlock_irq(&conf->device_lock);
7425 wake_up(&conf->wait_for_overlap);
7427 /* read-ahead size must cover two whole stripes, which is
7428 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7430 if (conf->mddev->queue) {
7431 int data_disks = conf->raid_disks - conf->max_degraded;
7432 int stripe = data_disks * ((conf->chunk_sectors << 9)
7434 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7435 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7440 /* This is called from the raid5d thread with mddev_lock held.
7441 * It makes config changes to the device.
7443 static void raid5_finish_reshape(struct mddev *mddev)
7445 struct r5conf *conf = mddev->private;
7447 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7449 if (mddev->delta_disks > 0) {
7450 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7451 set_capacity(mddev->gendisk, mddev->array_sectors);
7452 revalidate_disk(mddev->gendisk);
7455 spin_lock_irq(&conf->device_lock);
7456 mddev->degraded = calc_degraded(conf);
7457 spin_unlock_irq(&conf->device_lock);
7458 for (d = conf->raid_disks ;
7459 d < conf->raid_disks - mddev->delta_disks;
7461 struct md_rdev *rdev = conf->disks[d].rdev;
7463 clear_bit(In_sync, &rdev->flags);
7464 rdev = conf->disks[d].replacement;
7466 clear_bit(In_sync, &rdev->flags);
7469 mddev->layout = conf->algorithm;
7470 mddev->chunk_sectors = conf->chunk_sectors;
7471 mddev->reshape_position = MaxSector;
7472 mddev->delta_disks = 0;
7473 mddev->reshape_backwards = 0;
7477 static void raid5_quiesce(struct mddev *mddev, int state)
7479 struct r5conf *conf = mddev->private;
7482 case 2: /* resume for a suspend */
7483 wake_up(&conf->wait_for_overlap);
7486 case 1: /* stop all writes */
7487 lock_all_device_hash_locks_irq(conf);
7488 /* '2' tells resync/reshape to pause so that all
7489 * active stripes can drain
7492 wait_event_cmd(conf->wait_for_stripe,
7493 atomic_read(&conf->active_stripes) == 0 &&
7494 atomic_read(&conf->active_aligned_reads) == 0,
7495 unlock_all_device_hash_locks_irq(conf),
7496 lock_all_device_hash_locks_irq(conf));
7498 unlock_all_device_hash_locks_irq(conf);
7499 /* allow reshape to continue */
7500 wake_up(&conf->wait_for_overlap);
7503 case 0: /* re-enable writes */
7504 lock_all_device_hash_locks_irq(conf);
7506 wake_up(&conf->wait_for_stripe);
7507 wake_up(&conf->wait_for_overlap);
7508 unlock_all_device_hash_locks_irq(conf);
7513 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7515 struct r0conf *raid0_conf = mddev->private;
7518 /* for raid0 takeover only one zone is supported */
7519 if (raid0_conf->nr_strip_zones > 1) {
7520 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7522 return ERR_PTR(-EINVAL);
7525 sectors = raid0_conf->strip_zone[0].zone_end;
7526 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7527 mddev->dev_sectors = sectors;
7528 mddev->new_level = level;
7529 mddev->new_layout = ALGORITHM_PARITY_N;
7530 mddev->new_chunk_sectors = mddev->chunk_sectors;
7531 mddev->raid_disks += 1;
7532 mddev->delta_disks = 1;
7533 /* make sure it will be not marked as dirty */
7534 mddev->recovery_cp = MaxSector;
7536 return setup_conf(mddev);
7539 static void *raid5_takeover_raid1(struct mddev *mddev)
7543 if (mddev->raid_disks != 2 ||
7544 mddev->degraded > 1)
7545 return ERR_PTR(-EINVAL);
7547 /* Should check if there are write-behind devices? */
7549 chunksect = 64*2; /* 64K by default */
7551 /* The array must be an exact multiple of chunksize */
7552 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7555 if ((chunksect<<9) < STRIPE_SIZE)
7556 /* array size does not allow a suitable chunk size */
7557 return ERR_PTR(-EINVAL);
7559 mddev->new_level = 5;
7560 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7561 mddev->new_chunk_sectors = chunksect;
7563 return setup_conf(mddev);
7566 static void *raid5_takeover_raid6(struct mddev *mddev)
7570 switch (mddev->layout) {
7571 case ALGORITHM_LEFT_ASYMMETRIC_6:
7572 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7574 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7575 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7577 case ALGORITHM_LEFT_SYMMETRIC_6:
7578 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7580 case ALGORITHM_RIGHT_SYMMETRIC_6:
7581 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7583 case ALGORITHM_PARITY_0_6:
7584 new_layout = ALGORITHM_PARITY_0;
7586 case ALGORITHM_PARITY_N:
7587 new_layout = ALGORITHM_PARITY_N;
7590 return ERR_PTR(-EINVAL);
7592 mddev->new_level = 5;
7593 mddev->new_layout = new_layout;
7594 mddev->delta_disks = -1;
7595 mddev->raid_disks -= 1;
7596 return setup_conf(mddev);
7599 static int raid5_check_reshape(struct mddev *mddev)
7601 /* For a 2-drive array, the layout and chunk size can be changed
7602 * immediately as not restriping is needed.
7603 * For larger arrays we record the new value - after validation
7604 * to be used by a reshape pass.
7606 struct r5conf *conf = mddev->private;
7607 int new_chunk = mddev->new_chunk_sectors;
7609 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7611 if (new_chunk > 0) {
7612 if (!is_power_of_2(new_chunk))
7614 if (new_chunk < (PAGE_SIZE>>9))
7616 if (mddev->array_sectors & (new_chunk-1))
7617 /* not factor of array size */
7621 /* They look valid */
7623 if (mddev->raid_disks == 2) {
7624 /* can make the change immediately */
7625 if (mddev->new_layout >= 0) {
7626 conf->algorithm = mddev->new_layout;
7627 mddev->layout = mddev->new_layout;
7629 if (new_chunk > 0) {
7630 conf->chunk_sectors = new_chunk ;
7631 mddev->chunk_sectors = new_chunk;
7633 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7634 md_wakeup_thread(mddev->thread);
7636 return check_reshape(mddev);
7639 static int raid6_check_reshape(struct mddev *mddev)
7641 int new_chunk = mddev->new_chunk_sectors;
7643 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7645 if (new_chunk > 0) {
7646 if (!is_power_of_2(new_chunk))
7648 if (new_chunk < (PAGE_SIZE >> 9))
7650 if (mddev->array_sectors & (new_chunk-1))
7651 /* not factor of array size */
7655 /* They look valid */
7656 return check_reshape(mddev);
7659 static void *raid5_takeover(struct mddev *mddev)
7661 /* raid5 can take over:
7662 * raid0 - if there is only one strip zone - make it a raid4 layout
7663 * raid1 - if there are two drives. We need to know the chunk size
7664 * raid4 - trivial - just use a raid4 layout.
7665 * raid6 - Providing it is a *_6 layout
7667 if (mddev->level == 0)
7668 return raid45_takeover_raid0(mddev, 5);
7669 if (mddev->level == 1)
7670 return raid5_takeover_raid1(mddev);
7671 if (mddev->level == 4) {
7672 mddev->new_layout = ALGORITHM_PARITY_N;
7673 mddev->new_level = 5;
7674 return setup_conf(mddev);
7676 if (mddev->level == 6)
7677 return raid5_takeover_raid6(mddev);
7679 return ERR_PTR(-EINVAL);
7682 static void *raid4_takeover(struct mddev *mddev)
7684 /* raid4 can take over:
7685 * raid0 - if there is only one strip zone
7686 * raid5 - if layout is right
7688 if (mddev->level == 0)
7689 return raid45_takeover_raid0(mddev, 4);
7690 if (mddev->level == 5 &&
7691 mddev->layout == ALGORITHM_PARITY_N) {
7692 mddev->new_layout = 0;
7693 mddev->new_level = 4;
7694 return setup_conf(mddev);
7696 return ERR_PTR(-EINVAL);
7699 static struct md_personality raid5_personality;
7701 static void *raid6_takeover(struct mddev *mddev)
7703 /* Currently can only take over a raid5. We map the
7704 * personality to an equivalent raid6 personality
7705 * with the Q block at the end.
7709 if (mddev->pers != &raid5_personality)
7710 return ERR_PTR(-EINVAL);
7711 if (mddev->degraded > 1)
7712 return ERR_PTR(-EINVAL);
7713 if (mddev->raid_disks > 253)
7714 return ERR_PTR(-EINVAL);
7715 if (mddev->raid_disks < 3)
7716 return ERR_PTR(-EINVAL);
7718 switch (mddev->layout) {
7719 case ALGORITHM_LEFT_ASYMMETRIC:
7720 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7722 case ALGORITHM_RIGHT_ASYMMETRIC:
7723 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7725 case ALGORITHM_LEFT_SYMMETRIC:
7726 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7728 case ALGORITHM_RIGHT_SYMMETRIC:
7729 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7731 case ALGORITHM_PARITY_0:
7732 new_layout = ALGORITHM_PARITY_0_6;
7734 case ALGORITHM_PARITY_N:
7735 new_layout = ALGORITHM_PARITY_N;
7738 return ERR_PTR(-EINVAL);
7740 mddev->new_level = 6;
7741 mddev->new_layout = new_layout;
7742 mddev->delta_disks = 1;
7743 mddev->raid_disks += 1;
7744 return setup_conf(mddev);
7747 static struct md_personality raid6_personality =
7751 .owner = THIS_MODULE,
7752 .make_request = make_request,
7756 .error_handler = error,
7757 .hot_add_disk = raid5_add_disk,
7758 .hot_remove_disk= raid5_remove_disk,
7759 .spare_active = raid5_spare_active,
7760 .sync_request = sync_request,
7761 .resize = raid5_resize,
7763 .check_reshape = raid6_check_reshape,
7764 .start_reshape = raid5_start_reshape,
7765 .finish_reshape = raid5_finish_reshape,
7766 .quiesce = raid5_quiesce,
7767 .takeover = raid6_takeover,
7768 .congested = raid5_congested,
7769 .mergeable_bvec = raid5_mergeable_bvec,
7771 static struct md_personality raid5_personality =
7775 .owner = THIS_MODULE,
7776 .make_request = make_request,
7780 .error_handler = error,
7781 .hot_add_disk = raid5_add_disk,
7782 .hot_remove_disk= raid5_remove_disk,
7783 .spare_active = raid5_spare_active,
7784 .sync_request = sync_request,
7785 .resize = raid5_resize,
7787 .check_reshape = raid5_check_reshape,
7788 .start_reshape = raid5_start_reshape,
7789 .finish_reshape = raid5_finish_reshape,
7790 .quiesce = raid5_quiesce,
7791 .takeover = raid5_takeover,
7792 .congested = raid5_congested,
7793 .mergeable_bvec = raid5_mergeable_bvec,
7796 static struct md_personality raid4_personality =
7800 .owner = THIS_MODULE,
7801 .make_request = make_request,
7805 .error_handler = error,
7806 .hot_add_disk = raid5_add_disk,
7807 .hot_remove_disk= raid5_remove_disk,
7808 .spare_active = raid5_spare_active,
7809 .sync_request = sync_request,
7810 .resize = raid5_resize,
7812 .check_reshape = raid5_check_reshape,
7813 .start_reshape = raid5_start_reshape,
7814 .finish_reshape = raid5_finish_reshape,
7815 .quiesce = raid5_quiesce,
7816 .takeover = raid4_takeover,
7817 .congested = raid5_congested,
7818 .mergeable_bvec = raid5_mergeable_bvec,
7821 static int __init raid5_init(void)
7823 raid5_wq = alloc_workqueue("raid5wq",
7824 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7827 register_md_personality(&raid6_personality);
7828 register_md_personality(&raid5_personality);
7829 register_md_personality(&raid4_personality);
7833 static void raid5_exit(void)
7835 unregister_md_personality(&raid6_personality);
7836 unregister_md_personality(&raid5_personality);
7837 unregister_md_personality(&raid4_personality);
7838 destroy_workqueue(raid5_wq);
7841 module_init(raid5_init);
7842 module_exit(raid5_exit);
7843 MODULE_LICENSE("GPL");
7844 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7845 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7846 MODULE_ALIAS("md-raid5");
7847 MODULE_ALIAS("md-raid4");
7848 MODULE_ALIAS("md-level-5");
7849 MODULE_ALIAS("md-level-4");
7850 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7851 MODULE_ALIAS("md-raid6");
7852 MODULE_ALIAS("md-level-6");
7854 /* This used to be two separate modules, they were: */
7855 MODULE_ALIAS("raid5");
7856 MODULE_ALIAS("raid6");