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;
1922 percpu = per_cpu_ptr(conf->percpu, cpu);
1923 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1924 ops_run_biofill(sh);
1928 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1930 tx = ops_run_compute5(sh, percpu);
1932 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1933 tx = ops_run_compute6_1(sh, percpu);
1935 tx = ops_run_compute6_2(sh, percpu);
1937 /* terminate the chain if reconstruct is not set to be run */
1938 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1942 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1944 tx = ops_run_prexor5(sh, percpu, tx);
1946 tx = ops_run_prexor6(sh, percpu, tx);
1949 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1950 tx = ops_run_biodrain(sh, tx);
1954 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1956 ops_run_reconstruct5(sh, percpu, tx);
1958 ops_run_reconstruct6(sh, percpu, tx);
1961 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1962 if (sh->check_state == check_state_run)
1963 ops_run_check_p(sh, percpu);
1964 else if (sh->check_state == check_state_run_q)
1965 ops_run_check_pq(sh, percpu, 0);
1966 else if (sh->check_state == check_state_run_pq)
1967 ops_run_check_pq(sh, percpu, 1);
1972 if (overlap_clear && !sh->batch_head)
1973 for (i = disks; i--; ) {
1974 struct r5dev *dev = &sh->dev[i];
1975 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1976 wake_up(&sh->raid_conf->wait_for_overlap);
1981 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1983 struct stripe_head *sh;
1985 sh = kmem_cache_zalloc(sc, gfp);
1987 spin_lock_init(&sh->stripe_lock);
1988 spin_lock_init(&sh->batch_lock);
1989 INIT_LIST_HEAD(&sh->batch_list);
1990 INIT_LIST_HEAD(&sh->lru);
1991 atomic_set(&sh->count, 1);
1995 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
1997 struct stripe_head *sh;
1999 sh = alloc_stripe(conf->slab_cache, gfp);
2003 sh->raid_conf = conf;
2005 if (grow_buffers(sh, gfp)) {
2007 kmem_cache_free(conf->slab_cache, sh);
2010 sh->hash_lock_index =
2011 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2012 /* we just created an active stripe so... */
2013 atomic_inc(&conf->active_stripes);
2016 conf->max_nr_stripes++;
2020 static int grow_stripes(struct r5conf *conf, int num)
2022 struct kmem_cache *sc;
2023 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2025 if (conf->mddev->gendisk)
2026 sprintf(conf->cache_name[0],
2027 "raid%d-%s", conf->level, mdname(conf->mddev));
2029 sprintf(conf->cache_name[0],
2030 "raid%d-%p", conf->level, conf->mddev);
2031 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2033 conf->active_name = 0;
2034 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2035 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2039 conf->slab_cache = sc;
2040 conf->pool_size = devs;
2042 if (!grow_one_stripe(conf, GFP_KERNEL))
2049 * scribble_len - return the required size of the scribble region
2050 * @num - total number of disks in the array
2052 * The size must be enough to contain:
2053 * 1/ a struct page pointer for each device in the array +2
2054 * 2/ room to convert each entry in (1) to its corresponding dma
2055 * (dma_map_page()) or page (page_address()) address.
2057 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2058 * calculate over all devices (not just the data blocks), using zeros in place
2059 * of the P and Q blocks.
2061 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2063 struct flex_array *ret;
2066 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2067 ret = flex_array_alloc(len, cnt, flags);
2070 /* always prealloc all elements, so no locking is required */
2071 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2072 flex_array_free(ret);
2078 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2083 mddev_suspend(conf->mddev);
2085 for_each_present_cpu(cpu) {
2086 struct raid5_percpu *percpu;
2087 struct flex_array *scribble;
2089 percpu = per_cpu_ptr(conf->percpu, cpu);
2090 scribble = scribble_alloc(new_disks,
2091 new_sectors / STRIPE_SECTORS,
2095 flex_array_free(percpu->scribble);
2096 percpu->scribble = scribble;
2103 mddev_resume(conf->mddev);
2107 static int resize_stripes(struct r5conf *conf, int newsize)
2109 /* Make all the stripes able to hold 'newsize' devices.
2110 * New slots in each stripe get 'page' set to a new page.
2112 * This happens in stages:
2113 * 1/ create a new kmem_cache and allocate the required number of
2115 * 2/ gather all the old stripe_heads and transfer the pages across
2116 * to the new stripe_heads. This will have the side effect of
2117 * freezing the array as once all stripe_heads have been collected,
2118 * no IO will be possible. Old stripe heads are freed once their
2119 * pages have been transferred over, and the old kmem_cache is
2120 * freed when all stripes are done.
2121 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2122 * we simple return a failre status - no need to clean anything up.
2123 * 4/ allocate new pages for the new slots in the new stripe_heads.
2124 * If this fails, we don't bother trying the shrink the
2125 * stripe_heads down again, we just leave them as they are.
2126 * As each stripe_head is processed the new one is released into
2129 * Once step2 is started, we cannot afford to wait for a write,
2130 * so we use GFP_NOIO allocations.
2132 struct stripe_head *osh, *nsh;
2133 LIST_HEAD(newstripes);
2134 struct disk_info *ndisks;
2136 struct kmem_cache *sc;
2140 if (newsize <= conf->pool_size)
2141 return 0; /* never bother to shrink */
2143 err = md_allow_write(conf->mddev);
2148 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2149 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2154 for (i = conf->max_nr_stripes; i; i--) {
2155 nsh = alloc_stripe(sc, GFP_KERNEL);
2159 nsh->raid_conf = conf;
2160 list_add(&nsh->lru, &newstripes);
2163 /* didn't get enough, give up */
2164 while (!list_empty(&newstripes)) {
2165 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2166 list_del(&nsh->lru);
2167 kmem_cache_free(sc, nsh);
2169 kmem_cache_destroy(sc);
2172 /* Step 2 - Must use GFP_NOIO now.
2173 * OK, we have enough stripes, start collecting inactive
2174 * stripes and copying them over
2178 list_for_each_entry(nsh, &newstripes, lru) {
2179 lock_device_hash_lock(conf, hash);
2180 wait_event_cmd(conf->wait_for_stripe,
2181 !list_empty(conf->inactive_list + hash),
2182 unlock_device_hash_lock(conf, hash),
2183 lock_device_hash_lock(conf, hash));
2184 osh = get_free_stripe(conf, hash);
2185 unlock_device_hash_lock(conf, hash);
2187 for(i=0; i<conf->pool_size; i++) {
2188 nsh->dev[i].page = osh->dev[i].page;
2189 nsh->dev[i].orig_page = osh->dev[i].page;
2191 nsh->hash_lock_index = hash;
2192 kmem_cache_free(conf->slab_cache, osh);
2194 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2195 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2200 kmem_cache_destroy(conf->slab_cache);
2203 * At this point, we are holding all the stripes so the array
2204 * is completely stalled, so now is a good time to resize
2205 * conf->disks and the scribble region
2207 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2209 for (i=0; i<conf->raid_disks; i++)
2210 ndisks[i] = conf->disks[i];
2212 conf->disks = ndisks;
2216 /* Step 4, return new stripes to service */
2217 while(!list_empty(&newstripes)) {
2218 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2219 list_del_init(&nsh->lru);
2221 for (i=conf->raid_disks; i < newsize; i++)
2222 if (nsh->dev[i].page == NULL) {
2223 struct page *p = alloc_page(GFP_NOIO);
2224 nsh->dev[i].page = p;
2225 nsh->dev[i].orig_page = p;
2229 release_stripe(nsh);
2231 /* critical section pass, GFP_NOIO no longer needed */
2233 conf->slab_cache = sc;
2234 conf->active_name = 1-conf->active_name;
2236 conf->pool_size = newsize;
2240 static int drop_one_stripe(struct r5conf *conf)
2242 struct stripe_head *sh;
2243 int hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
2245 spin_lock_irq(conf->hash_locks + hash);
2246 sh = get_free_stripe(conf, hash);
2247 spin_unlock_irq(conf->hash_locks + hash);
2250 BUG_ON(atomic_read(&sh->count));
2252 kmem_cache_free(conf->slab_cache, sh);
2253 atomic_dec(&conf->active_stripes);
2254 conf->max_nr_stripes--;
2258 static void shrink_stripes(struct r5conf *conf)
2260 while (conf->max_nr_stripes &&
2261 drop_one_stripe(conf))
2264 if (conf->slab_cache)
2265 kmem_cache_destroy(conf->slab_cache);
2266 conf->slab_cache = NULL;
2269 static void raid5_end_read_request(struct bio * bi, int error)
2271 struct stripe_head *sh = bi->bi_private;
2272 struct r5conf *conf = sh->raid_conf;
2273 int disks = sh->disks, i;
2274 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2275 char b[BDEVNAME_SIZE];
2276 struct md_rdev *rdev = NULL;
2279 for (i=0 ; i<disks; i++)
2280 if (bi == &sh->dev[i].req)
2283 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2284 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2290 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2291 /* If replacement finished while this request was outstanding,
2292 * 'replacement' might be NULL already.
2293 * In that case it moved down to 'rdev'.
2294 * rdev is not removed until all requests are finished.
2296 rdev = conf->disks[i].replacement;
2298 rdev = conf->disks[i].rdev;
2300 if (use_new_offset(conf, sh))
2301 s = sh->sector + rdev->new_data_offset;
2303 s = sh->sector + rdev->data_offset;
2305 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2306 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2307 /* Note that this cannot happen on a
2308 * replacement device. We just fail those on
2313 "md/raid:%s: read error corrected"
2314 " (%lu sectors at %llu on %s)\n",
2315 mdname(conf->mddev), STRIPE_SECTORS,
2316 (unsigned long long)s,
2317 bdevname(rdev->bdev, b));
2318 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2319 clear_bit(R5_ReadError, &sh->dev[i].flags);
2320 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2321 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2322 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2324 if (atomic_read(&rdev->read_errors))
2325 atomic_set(&rdev->read_errors, 0);
2327 const char *bdn = bdevname(rdev->bdev, b);
2331 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2332 atomic_inc(&rdev->read_errors);
2333 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2336 "md/raid:%s: read error on replacement device "
2337 "(sector %llu on %s).\n",
2338 mdname(conf->mddev),
2339 (unsigned long long)s,
2341 else if (conf->mddev->degraded >= conf->max_degraded) {
2345 "md/raid:%s: read error not correctable "
2346 "(sector %llu on %s).\n",
2347 mdname(conf->mddev),
2348 (unsigned long long)s,
2350 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2355 "md/raid:%s: read error NOT corrected!! "
2356 "(sector %llu on %s).\n",
2357 mdname(conf->mddev),
2358 (unsigned long long)s,
2360 } else if (atomic_read(&rdev->read_errors)
2361 > conf->max_nr_stripes)
2363 "md/raid:%s: Too many read errors, failing device %s.\n",
2364 mdname(conf->mddev), bdn);
2367 if (set_bad && test_bit(In_sync, &rdev->flags)
2368 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2371 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2372 set_bit(R5_ReadError, &sh->dev[i].flags);
2373 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2375 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2377 clear_bit(R5_ReadError, &sh->dev[i].flags);
2378 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2380 && test_bit(In_sync, &rdev->flags)
2381 && rdev_set_badblocks(
2382 rdev, sh->sector, STRIPE_SECTORS, 0)))
2383 md_error(conf->mddev, rdev);
2386 rdev_dec_pending(rdev, conf->mddev);
2387 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2388 set_bit(STRIPE_HANDLE, &sh->state);
2392 static void raid5_end_write_request(struct bio *bi, int error)
2394 struct stripe_head *sh = bi->bi_private;
2395 struct r5conf *conf = sh->raid_conf;
2396 int disks = sh->disks, i;
2397 struct md_rdev *uninitialized_var(rdev);
2398 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2401 int replacement = 0;
2403 for (i = 0 ; i < disks; i++) {
2404 if (bi == &sh->dev[i].req) {
2405 rdev = conf->disks[i].rdev;
2408 if (bi == &sh->dev[i].rreq) {
2409 rdev = conf->disks[i].replacement;
2413 /* rdev was removed and 'replacement'
2414 * replaced it. rdev is not removed
2415 * until all requests are finished.
2417 rdev = conf->disks[i].rdev;
2421 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2422 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2431 md_error(conf->mddev, rdev);
2432 else if (is_badblock(rdev, sh->sector,
2434 &first_bad, &bad_sectors))
2435 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2438 set_bit(STRIPE_DEGRADED, &sh->state);
2439 set_bit(WriteErrorSeen, &rdev->flags);
2440 set_bit(R5_WriteError, &sh->dev[i].flags);
2441 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2442 set_bit(MD_RECOVERY_NEEDED,
2443 &rdev->mddev->recovery);
2444 } else if (is_badblock(rdev, sh->sector,
2446 &first_bad, &bad_sectors)) {
2447 set_bit(R5_MadeGood, &sh->dev[i].flags);
2448 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2449 /* That was a successful write so make
2450 * sure it looks like we already did
2453 set_bit(R5_ReWrite, &sh->dev[i].flags);
2456 rdev_dec_pending(rdev, conf->mddev);
2458 if (sh->batch_head && !uptodate && !replacement)
2459 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2461 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2462 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2463 set_bit(STRIPE_HANDLE, &sh->state);
2466 if (sh->batch_head && sh != sh->batch_head)
2467 release_stripe(sh->batch_head);
2470 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2472 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2474 struct r5dev *dev = &sh->dev[i];
2476 bio_init(&dev->req);
2477 dev->req.bi_io_vec = &dev->vec;
2478 dev->req.bi_max_vecs = 1;
2479 dev->req.bi_private = sh;
2481 bio_init(&dev->rreq);
2482 dev->rreq.bi_io_vec = &dev->rvec;
2483 dev->rreq.bi_max_vecs = 1;
2484 dev->rreq.bi_private = sh;
2487 dev->sector = compute_blocknr(sh, i, previous);
2490 static void error(struct mddev *mddev, struct md_rdev *rdev)
2492 char b[BDEVNAME_SIZE];
2493 struct r5conf *conf = mddev->private;
2494 unsigned long flags;
2495 pr_debug("raid456: error called\n");
2497 spin_lock_irqsave(&conf->device_lock, flags);
2498 clear_bit(In_sync, &rdev->flags);
2499 mddev->degraded = calc_degraded(conf);
2500 spin_unlock_irqrestore(&conf->device_lock, flags);
2501 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2503 set_bit(Blocked, &rdev->flags);
2504 set_bit(Faulty, &rdev->flags);
2505 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2507 "md/raid:%s: Disk failure on %s, disabling device.\n"
2508 "md/raid:%s: Operation continuing on %d devices.\n",
2510 bdevname(rdev->bdev, b),
2512 conf->raid_disks - mddev->degraded);
2516 * Input: a 'big' sector number,
2517 * Output: index of the data and parity disk, and the sector # in them.
2519 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2520 int previous, int *dd_idx,
2521 struct stripe_head *sh)
2523 sector_t stripe, stripe2;
2524 sector_t chunk_number;
2525 unsigned int chunk_offset;
2528 sector_t new_sector;
2529 int algorithm = previous ? conf->prev_algo
2531 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2532 : conf->chunk_sectors;
2533 int raid_disks = previous ? conf->previous_raid_disks
2535 int data_disks = raid_disks - conf->max_degraded;
2537 /* First compute the information on this sector */
2540 * Compute the chunk number and the sector offset inside the chunk
2542 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2543 chunk_number = r_sector;
2546 * Compute the stripe number
2548 stripe = chunk_number;
2549 *dd_idx = sector_div(stripe, data_disks);
2552 * Select the parity disk based on the user selected algorithm.
2554 pd_idx = qd_idx = -1;
2555 switch(conf->level) {
2557 pd_idx = data_disks;
2560 switch (algorithm) {
2561 case ALGORITHM_LEFT_ASYMMETRIC:
2562 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2563 if (*dd_idx >= pd_idx)
2566 case ALGORITHM_RIGHT_ASYMMETRIC:
2567 pd_idx = sector_div(stripe2, raid_disks);
2568 if (*dd_idx >= pd_idx)
2571 case ALGORITHM_LEFT_SYMMETRIC:
2572 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2573 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2575 case ALGORITHM_RIGHT_SYMMETRIC:
2576 pd_idx = sector_div(stripe2, raid_disks);
2577 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2579 case ALGORITHM_PARITY_0:
2583 case ALGORITHM_PARITY_N:
2584 pd_idx = data_disks;
2592 switch (algorithm) {
2593 case ALGORITHM_LEFT_ASYMMETRIC:
2594 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2595 qd_idx = pd_idx + 1;
2596 if (pd_idx == raid_disks-1) {
2597 (*dd_idx)++; /* Q D D D P */
2599 } else if (*dd_idx >= pd_idx)
2600 (*dd_idx) += 2; /* D D P Q D */
2602 case ALGORITHM_RIGHT_ASYMMETRIC:
2603 pd_idx = sector_div(stripe2, raid_disks);
2604 qd_idx = pd_idx + 1;
2605 if (pd_idx == raid_disks-1) {
2606 (*dd_idx)++; /* Q D D D P */
2608 } else if (*dd_idx >= pd_idx)
2609 (*dd_idx) += 2; /* D D P Q D */
2611 case ALGORITHM_LEFT_SYMMETRIC:
2612 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2613 qd_idx = (pd_idx + 1) % raid_disks;
2614 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2616 case ALGORITHM_RIGHT_SYMMETRIC:
2617 pd_idx = sector_div(stripe2, raid_disks);
2618 qd_idx = (pd_idx + 1) % raid_disks;
2619 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2622 case ALGORITHM_PARITY_0:
2627 case ALGORITHM_PARITY_N:
2628 pd_idx = data_disks;
2629 qd_idx = data_disks + 1;
2632 case ALGORITHM_ROTATING_ZERO_RESTART:
2633 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2634 * of blocks for computing Q is different.
2636 pd_idx = sector_div(stripe2, raid_disks);
2637 qd_idx = pd_idx + 1;
2638 if (pd_idx == raid_disks-1) {
2639 (*dd_idx)++; /* Q D D D P */
2641 } else if (*dd_idx >= pd_idx)
2642 (*dd_idx) += 2; /* D D P Q D */
2646 case ALGORITHM_ROTATING_N_RESTART:
2647 /* Same a left_asymmetric, by first stripe is
2648 * D D D P Q rather than
2652 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2653 qd_idx = pd_idx + 1;
2654 if (pd_idx == raid_disks-1) {
2655 (*dd_idx)++; /* Q D D D P */
2657 } else if (*dd_idx >= pd_idx)
2658 (*dd_idx) += 2; /* D D P Q D */
2662 case ALGORITHM_ROTATING_N_CONTINUE:
2663 /* Same as left_symmetric but Q is before P */
2664 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2665 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2666 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2670 case ALGORITHM_LEFT_ASYMMETRIC_6:
2671 /* RAID5 left_asymmetric, with Q on last device */
2672 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2673 if (*dd_idx >= pd_idx)
2675 qd_idx = raid_disks - 1;
2678 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2679 pd_idx = sector_div(stripe2, raid_disks-1);
2680 if (*dd_idx >= pd_idx)
2682 qd_idx = raid_disks - 1;
2685 case ALGORITHM_LEFT_SYMMETRIC_6:
2686 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2687 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2688 qd_idx = raid_disks - 1;
2691 case ALGORITHM_RIGHT_SYMMETRIC_6:
2692 pd_idx = sector_div(stripe2, raid_disks-1);
2693 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2694 qd_idx = raid_disks - 1;
2697 case ALGORITHM_PARITY_0_6:
2700 qd_idx = raid_disks - 1;
2710 sh->pd_idx = pd_idx;
2711 sh->qd_idx = qd_idx;
2712 sh->ddf_layout = ddf_layout;
2715 * Finally, compute the new sector number
2717 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2721 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2723 struct r5conf *conf = sh->raid_conf;
2724 int raid_disks = sh->disks;
2725 int data_disks = raid_disks - conf->max_degraded;
2726 sector_t new_sector = sh->sector, check;
2727 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2728 : conf->chunk_sectors;
2729 int algorithm = previous ? conf->prev_algo
2733 sector_t chunk_number;
2734 int dummy1, dd_idx = i;
2736 struct stripe_head sh2;
2738 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2739 stripe = new_sector;
2741 if (i == sh->pd_idx)
2743 switch(conf->level) {
2746 switch (algorithm) {
2747 case ALGORITHM_LEFT_ASYMMETRIC:
2748 case ALGORITHM_RIGHT_ASYMMETRIC:
2752 case ALGORITHM_LEFT_SYMMETRIC:
2753 case ALGORITHM_RIGHT_SYMMETRIC:
2756 i -= (sh->pd_idx + 1);
2758 case ALGORITHM_PARITY_0:
2761 case ALGORITHM_PARITY_N:
2768 if (i == sh->qd_idx)
2769 return 0; /* It is the Q disk */
2770 switch (algorithm) {
2771 case ALGORITHM_LEFT_ASYMMETRIC:
2772 case ALGORITHM_RIGHT_ASYMMETRIC:
2773 case ALGORITHM_ROTATING_ZERO_RESTART:
2774 case ALGORITHM_ROTATING_N_RESTART:
2775 if (sh->pd_idx == raid_disks-1)
2776 i--; /* Q D D D P */
2777 else if (i > sh->pd_idx)
2778 i -= 2; /* D D P Q D */
2780 case ALGORITHM_LEFT_SYMMETRIC:
2781 case ALGORITHM_RIGHT_SYMMETRIC:
2782 if (sh->pd_idx == raid_disks-1)
2783 i--; /* Q D D D P */
2788 i -= (sh->pd_idx + 2);
2791 case ALGORITHM_PARITY_0:
2794 case ALGORITHM_PARITY_N:
2796 case ALGORITHM_ROTATING_N_CONTINUE:
2797 /* Like left_symmetric, but P is before Q */
2798 if (sh->pd_idx == 0)
2799 i--; /* P D D D Q */
2804 i -= (sh->pd_idx + 1);
2807 case ALGORITHM_LEFT_ASYMMETRIC_6:
2808 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2812 case ALGORITHM_LEFT_SYMMETRIC_6:
2813 case ALGORITHM_RIGHT_SYMMETRIC_6:
2815 i += data_disks + 1;
2816 i -= (sh->pd_idx + 1);
2818 case ALGORITHM_PARITY_0_6:
2827 chunk_number = stripe * data_disks + i;
2828 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2830 check = raid5_compute_sector(conf, r_sector,
2831 previous, &dummy1, &sh2);
2832 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2833 || sh2.qd_idx != sh->qd_idx) {
2834 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2835 mdname(conf->mddev));
2842 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2843 int rcw, int expand)
2845 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2846 struct r5conf *conf = sh->raid_conf;
2847 int level = conf->level;
2851 for (i = disks; i--; ) {
2852 struct r5dev *dev = &sh->dev[i];
2855 set_bit(R5_LOCKED, &dev->flags);
2856 set_bit(R5_Wantdrain, &dev->flags);
2858 clear_bit(R5_UPTODATE, &dev->flags);
2862 /* if we are not expanding this is a proper write request, and
2863 * there will be bios with new data to be drained into the
2868 /* False alarm, nothing to do */
2870 sh->reconstruct_state = reconstruct_state_drain_run;
2871 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2873 sh->reconstruct_state = reconstruct_state_run;
2875 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2877 if (s->locked + conf->max_degraded == disks)
2878 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2879 atomic_inc(&conf->pending_full_writes);
2881 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2882 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2883 BUG_ON(level == 6 &&
2884 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2885 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2887 for (i = disks; i--; ) {
2888 struct r5dev *dev = &sh->dev[i];
2889 if (i == pd_idx || i == qd_idx)
2893 (test_bit(R5_UPTODATE, &dev->flags) ||
2894 test_bit(R5_Wantcompute, &dev->flags))) {
2895 set_bit(R5_Wantdrain, &dev->flags);
2896 set_bit(R5_LOCKED, &dev->flags);
2897 clear_bit(R5_UPTODATE, &dev->flags);
2902 /* False alarm - nothing to do */
2904 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2905 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2906 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2907 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2910 /* keep the parity disk(s) locked while asynchronous operations
2913 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2914 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2918 int qd_idx = sh->qd_idx;
2919 struct r5dev *dev = &sh->dev[qd_idx];
2921 set_bit(R5_LOCKED, &dev->flags);
2922 clear_bit(R5_UPTODATE, &dev->flags);
2926 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2927 __func__, (unsigned long long)sh->sector,
2928 s->locked, s->ops_request);
2932 * Each stripe/dev can have one or more bion attached.
2933 * toread/towrite point to the first in a chain.
2934 * The bi_next chain must be in order.
2936 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2937 int forwrite, int previous)
2940 struct r5conf *conf = sh->raid_conf;
2943 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2944 (unsigned long long)bi->bi_iter.bi_sector,
2945 (unsigned long long)sh->sector);
2948 * If several bio share a stripe. The bio bi_phys_segments acts as a
2949 * reference count to avoid race. The reference count should already be
2950 * increased before this function is called (for example, in
2951 * make_request()), so other bio sharing this stripe will not free the
2952 * stripe. If a stripe is owned by one stripe, the stripe lock will
2955 spin_lock_irq(&sh->stripe_lock);
2956 /* Don't allow new IO added to stripes in batch list */
2960 bip = &sh->dev[dd_idx].towrite;
2964 bip = &sh->dev[dd_idx].toread;
2965 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2966 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2968 bip = & (*bip)->bi_next;
2970 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2973 if (!forwrite || previous)
2974 clear_bit(STRIPE_BATCH_READY, &sh->state);
2976 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2980 raid5_inc_bi_active_stripes(bi);
2983 /* check if page is covered */
2984 sector_t sector = sh->dev[dd_idx].sector;
2985 for (bi=sh->dev[dd_idx].towrite;
2986 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2987 bi && bi->bi_iter.bi_sector <= sector;
2988 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2989 if (bio_end_sector(bi) >= sector)
2990 sector = bio_end_sector(bi);
2992 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2993 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
2994 sh->overwrite_disks++;
2997 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2998 (unsigned long long)(*bip)->bi_iter.bi_sector,
2999 (unsigned long long)sh->sector, dd_idx);
3001 if (conf->mddev->bitmap && firstwrite) {
3002 /* Cannot hold spinlock over bitmap_startwrite,
3003 * but must ensure this isn't added to a batch until
3004 * we have added to the bitmap and set bm_seq.
3005 * So set STRIPE_BITMAP_PENDING to prevent
3007 * If multiple add_stripe_bio() calls race here they
3008 * much all set STRIPE_BITMAP_PENDING. So only the first one
3009 * to complete "bitmap_startwrite" gets to set
3010 * STRIPE_BIT_DELAY. This is important as once a stripe
3011 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3014 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3015 spin_unlock_irq(&sh->stripe_lock);
3016 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3018 spin_lock_irq(&sh->stripe_lock);
3019 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3020 if (!sh->batch_head) {
3021 sh->bm_seq = conf->seq_flush+1;
3022 set_bit(STRIPE_BIT_DELAY, &sh->state);
3025 spin_unlock_irq(&sh->stripe_lock);
3027 if (stripe_can_batch(sh))
3028 stripe_add_to_batch_list(conf, sh);
3032 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3033 spin_unlock_irq(&sh->stripe_lock);
3037 static void end_reshape(struct r5conf *conf);
3039 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3040 struct stripe_head *sh)
3042 int sectors_per_chunk =
3043 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3045 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3046 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3048 raid5_compute_sector(conf,
3049 stripe * (disks - conf->max_degraded)
3050 *sectors_per_chunk + chunk_offset,
3056 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3057 struct stripe_head_state *s, int disks,
3058 struct bio **return_bi)
3061 BUG_ON(sh->batch_head);
3062 for (i = disks; i--; ) {
3066 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3067 struct md_rdev *rdev;
3069 rdev = rcu_dereference(conf->disks[i].rdev);
3070 if (rdev && test_bit(In_sync, &rdev->flags))
3071 atomic_inc(&rdev->nr_pending);
3076 if (!rdev_set_badblocks(
3080 md_error(conf->mddev, rdev);
3081 rdev_dec_pending(rdev, conf->mddev);
3084 spin_lock_irq(&sh->stripe_lock);
3085 /* fail all writes first */
3086 bi = sh->dev[i].towrite;
3087 sh->dev[i].towrite = NULL;
3088 sh->overwrite_disks = 0;
3089 spin_unlock_irq(&sh->stripe_lock);
3093 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3094 wake_up(&conf->wait_for_overlap);
3096 while (bi && bi->bi_iter.bi_sector <
3097 sh->dev[i].sector + STRIPE_SECTORS) {
3098 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3099 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3100 if (!raid5_dec_bi_active_stripes(bi)) {
3101 md_write_end(conf->mddev);
3102 bi->bi_next = *return_bi;
3108 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3109 STRIPE_SECTORS, 0, 0);
3111 /* and fail all 'written' */
3112 bi = sh->dev[i].written;
3113 sh->dev[i].written = NULL;
3114 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3115 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3116 sh->dev[i].page = sh->dev[i].orig_page;
3119 if (bi) bitmap_end = 1;
3120 while (bi && bi->bi_iter.bi_sector <
3121 sh->dev[i].sector + STRIPE_SECTORS) {
3122 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3123 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3124 if (!raid5_dec_bi_active_stripes(bi)) {
3125 md_write_end(conf->mddev);
3126 bi->bi_next = *return_bi;
3132 /* fail any reads if this device is non-operational and
3133 * the data has not reached the cache yet.
3135 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3136 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3137 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3138 spin_lock_irq(&sh->stripe_lock);
3139 bi = sh->dev[i].toread;
3140 sh->dev[i].toread = NULL;
3141 spin_unlock_irq(&sh->stripe_lock);
3142 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3143 wake_up(&conf->wait_for_overlap);
3144 while (bi && bi->bi_iter.bi_sector <
3145 sh->dev[i].sector + STRIPE_SECTORS) {
3146 struct bio *nextbi =
3147 r5_next_bio(bi, sh->dev[i].sector);
3148 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3149 if (!raid5_dec_bi_active_stripes(bi)) {
3150 bi->bi_next = *return_bi;
3157 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3158 STRIPE_SECTORS, 0, 0);
3159 /* If we were in the middle of a write the parity block might
3160 * still be locked - so just clear all R5_LOCKED flags
3162 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3165 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3166 if (atomic_dec_and_test(&conf->pending_full_writes))
3167 md_wakeup_thread(conf->mddev->thread);
3171 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3172 struct stripe_head_state *s)
3177 BUG_ON(sh->batch_head);
3178 clear_bit(STRIPE_SYNCING, &sh->state);
3179 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3180 wake_up(&conf->wait_for_overlap);
3183 /* There is nothing more to do for sync/check/repair.
3184 * Don't even need to abort as that is handled elsewhere
3185 * if needed, and not always wanted e.g. if there is a known
3187 * For recover/replace we need to record a bad block on all
3188 * non-sync devices, or abort the recovery
3190 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3191 /* During recovery devices cannot be removed, so
3192 * locking and refcounting of rdevs is not needed
3194 for (i = 0; i < conf->raid_disks; i++) {
3195 struct md_rdev *rdev = conf->disks[i].rdev;
3197 && !test_bit(Faulty, &rdev->flags)
3198 && !test_bit(In_sync, &rdev->flags)
3199 && !rdev_set_badblocks(rdev, sh->sector,
3202 rdev = conf->disks[i].replacement;
3204 && !test_bit(Faulty, &rdev->flags)
3205 && !test_bit(In_sync, &rdev->flags)
3206 && !rdev_set_badblocks(rdev, sh->sector,
3211 conf->recovery_disabled =
3212 conf->mddev->recovery_disabled;
3214 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3217 static int want_replace(struct stripe_head *sh, int disk_idx)
3219 struct md_rdev *rdev;
3221 /* Doing recovery so rcu locking not required */
3222 rdev = sh->raid_conf->disks[disk_idx].replacement;
3224 && !test_bit(Faulty, &rdev->flags)
3225 && !test_bit(In_sync, &rdev->flags)
3226 && (rdev->recovery_offset <= sh->sector
3227 || rdev->mddev->recovery_cp <= sh->sector))
3233 /* fetch_block - checks the given member device to see if its data needs
3234 * to be read or computed to satisfy a request.
3236 * Returns 1 when no more member devices need to be checked, otherwise returns
3237 * 0 to tell the loop in handle_stripe_fill to continue
3240 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3241 int disk_idx, int disks)
3243 struct r5dev *dev = &sh->dev[disk_idx];
3244 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3245 &sh->dev[s->failed_num[1]] };
3249 if (test_bit(R5_LOCKED, &dev->flags) ||
3250 test_bit(R5_UPTODATE, &dev->flags))
3251 /* No point reading this as we already have it or have
3252 * decided to get it.
3257 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3258 /* We need this block to directly satisfy a request */
3261 if (s->syncing || s->expanding ||
3262 (s->replacing && want_replace(sh, disk_idx)))
3263 /* When syncing, or expanding we read everything.
3264 * When replacing, we need the replaced block.
3268 if ((s->failed >= 1 && fdev[0]->toread) ||
3269 (s->failed >= 2 && fdev[1]->toread))
3270 /* If we want to read from a failed device, then
3271 * we need to actually read every other device.
3275 /* Sometimes neither read-modify-write nor reconstruct-write
3276 * cycles can work. In those cases we read every block we
3277 * can. Then the parity-update is certain to have enough to
3279 * This can only be a problem when we need to write something,
3280 * and some device has failed. If either of those tests
3281 * fail we need look no further.
3283 if (!s->failed || !s->to_write)
3286 if (test_bit(R5_Insync, &dev->flags) &&
3287 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3288 /* Pre-reads at not permitted until after short delay
3289 * to gather multiple requests. However if this
3290 * device is no Insync, the block could only be be computed
3291 * and there is no need to delay that.
3295 for (i = 0; i < s->failed; i++) {
3296 if (fdev[i]->towrite &&
3297 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3298 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3299 /* If we have a partial write to a failed
3300 * device, then we will need to reconstruct
3301 * the content of that device, so all other
3302 * devices must be read.
3307 /* If we are forced to do a reconstruct-write, either because
3308 * the current RAID6 implementation only supports that, or
3309 * or because parity cannot be trusted and we are currently
3310 * recovering it, there is extra need to be careful.
3311 * If one of the devices that we would need to read, because
3312 * it is not being overwritten (and maybe not written at all)
3313 * is missing/faulty, then we need to read everything we can.
3315 if (sh->raid_conf->level != 6 &&
3316 sh->sector < sh->raid_conf->mddev->recovery_cp)
3317 /* reconstruct-write isn't being forced */
3319 for (i = 0; i < s->failed; i++) {
3320 if (s->failed_num[i] != sh->pd_idx &&
3321 s->failed_num[i] != sh->qd_idx &&
3322 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3323 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3330 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3331 int disk_idx, int disks)
3333 struct r5dev *dev = &sh->dev[disk_idx];
3335 /* is the data in this block needed, and can we get it? */
3336 if (need_this_block(sh, s, disk_idx, disks)) {
3337 /* we would like to get this block, possibly by computing it,
3338 * otherwise read it if the backing disk is insync
3340 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3341 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3342 BUG_ON(sh->batch_head);
3343 if ((s->uptodate == disks - 1) &&
3344 (s->failed && (disk_idx == s->failed_num[0] ||
3345 disk_idx == s->failed_num[1]))) {
3346 /* have disk failed, and we're requested to fetch it;
3349 pr_debug("Computing stripe %llu block %d\n",
3350 (unsigned long long)sh->sector, disk_idx);
3351 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3352 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3353 set_bit(R5_Wantcompute, &dev->flags);
3354 sh->ops.target = disk_idx;
3355 sh->ops.target2 = -1; /* no 2nd target */
3357 /* Careful: from this point on 'uptodate' is in the eye
3358 * of raid_run_ops which services 'compute' operations
3359 * before writes. R5_Wantcompute flags a block that will
3360 * be R5_UPTODATE by the time it is needed for a
3361 * subsequent operation.
3365 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3366 /* Computing 2-failure is *very* expensive; only
3367 * do it if failed >= 2
3370 for (other = disks; other--; ) {
3371 if (other == disk_idx)
3373 if (!test_bit(R5_UPTODATE,
3374 &sh->dev[other].flags))
3378 pr_debug("Computing stripe %llu blocks %d,%d\n",
3379 (unsigned long long)sh->sector,
3381 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3382 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3383 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3384 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3385 sh->ops.target = disk_idx;
3386 sh->ops.target2 = other;
3390 } else if (test_bit(R5_Insync, &dev->flags)) {
3391 set_bit(R5_LOCKED, &dev->flags);
3392 set_bit(R5_Wantread, &dev->flags);
3394 pr_debug("Reading block %d (sync=%d)\n",
3395 disk_idx, s->syncing);
3403 * handle_stripe_fill - read or compute data to satisfy pending requests.
3405 static void handle_stripe_fill(struct stripe_head *sh,
3406 struct stripe_head_state *s,
3411 /* look for blocks to read/compute, skip this if a compute
3412 * is already in flight, or if the stripe contents are in the
3413 * midst of changing due to a write
3415 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3416 !sh->reconstruct_state)
3417 for (i = disks; i--; )
3418 if (fetch_block(sh, s, i, disks))
3420 set_bit(STRIPE_HANDLE, &sh->state);
3423 /* handle_stripe_clean_event
3424 * any written block on an uptodate or failed drive can be returned.
3425 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3426 * never LOCKED, so we don't need to test 'failed' directly.
3428 static void handle_stripe_clean_event(struct r5conf *conf,
3429 struct stripe_head *sh, int disks, struct bio **return_bi)
3433 int discard_pending = 0;
3434 struct stripe_head *head_sh = sh;
3435 bool do_endio = false;
3438 for (i = disks; i--; )
3439 if (sh->dev[i].written) {
3441 if (!test_bit(R5_LOCKED, &dev->flags) &&
3442 (test_bit(R5_UPTODATE, &dev->flags) ||
3443 test_bit(R5_Discard, &dev->flags) ||
3444 test_bit(R5_SkipCopy, &dev->flags))) {
3445 /* We can return any write requests */
3446 struct bio *wbi, *wbi2;
3447 pr_debug("Return write for disc %d\n", i);
3448 if (test_and_clear_bit(R5_Discard, &dev->flags))
3449 clear_bit(R5_UPTODATE, &dev->flags);
3450 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3451 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3456 dev->page = dev->orig_page;
3458 dev->written = NULL;
3459 while (wbi && wbi->bi_iter.bi_sector <
3460 dev->sector + STRIPE_SECTORS) {
3461 wbi2 = r5_next_bio(wbi, dev->sector);
3462 if (!raid5_dec_bi_active_stripes(wbi)) {
3463 md_write_end(conf->mddev);
3464 wbi->bi_next = *return_bi;
3469 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3471 !test_bit(STRIPE_DEGRADED, &sh->state),
3473 if (head_sh->batch_head) {
3474 sh = list_first_entry(&sh->batch_list,
3477 if (sh != head_sh) {
3484 } else if (test_bit(R5_Discard, &dev->flags))
3485 discard_pending = 1;
3486 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3487 WARN_ON(dev->page != dev->orig_page);
3489 if (!discard_pending &&
3490 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3491 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3492 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3493 if (sh->qd_idx >= 0) {
3494 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3495 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3497 /* now that discard is done we can proceed with any sync */
3498 clear_bit(STRIPE_DISCARD, &sh->state);
3500 * SCSI discard will change some bio fields and the stripe has
3501 * no updated data, so remove it from hash list and the stripe
3502 * will be reinitialized
3504 spin_lock_irq(&conf->device_lock);
3507 if (head_sh->batch_head) {
3508 sh = list_first_entry(&sh->batch_list,
3509 struct stripe_head, batch_list);
3513 spin_unlock_irq(&conf->device_lock);
3516 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3517 set_bit(STRIPE_HANDLE, &sh->state);
3521 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3522 if (atomic_dec_and_test(&conf->pending_full_writes))
3523 md_wakeup_thread(conf->mddev->thread);
3525 if (!head_sh->batch_head || !do_endio)
3527 for (i = 0; i < head_sh->disks; i++) {
3528 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
3531 while (!list_empty(&head_sh->batch_list)) {
3533 sh = list_first_entry(&head_sh->batch_list,
3534 struct stripe_head, batch_list);
3535 list_del_init(&sh->batch_list);
3537 set_mask_bits(&sh->state, ~STRIPE_EXPAND_SYNC_FLAG,
3538 head_sh->state & ~((1 << STRIPE_ACTIVE) |
3539 (1 << STRIPE_PREREAD_ACTIVE) |
3540 STRIPE_EXPAND_SYNC_FLAG));
3541 sh->check_state = head_sh->check_state;
3542 sh->reconstruct_state = head_sh->reconstruct_state;
3543 for (i = 0; i < sh->disks; i++) {
3544 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3546 sh->dev[i].flags = head_sh->dev[i].flags;
3549 spin_lock_irq(&sh->stripe_lock);
3550 sh->batch_head = NULL;
3551 spin_unlock_irq(&sh->stripe_lock);
3552 if (sh->state & STRIPE_EXPAND_SYNC_FLAG)
3553 set_bit(STRIPE_HANDLE, &sh->state);
3557 spin_lock_irq(&head_sh->stripe_lock);
3558 head_sh->batch_head = NULL;
3559 spin_unlock_irq(&head_sh->stripe_lock);
3561 wake_up(&conf->wait_for_overlap);
3562 if (head_sh->state & STRIPE_EXPAND_SYNC_FLAG)
3563 set_bit(STRIPE_HANDLE, &head_sh->state);
3566 static void handle_stripe_dirtying(struct r5conf *conf,
3567 struct stripe_head *sh,
3568 struct stripe_head_state *s,
3571 int rmw = 0, rcw = 0, i;
3572 sector_t recovery_cp = conf->mddev->recovery_cp;
3574 /* Check whether resync is now happening or should start.
3575 * If yes, then the array is dirty (after unclean shutdown or
3576 * initial creation), so parity in some stripes might be inconsistent.
3577 * In this case, we need to always do reconstruct-write, to ensure
3578 * that in case of drive failure or read-error correction, we
3579 * generate correct data from the parity.
3581 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3582 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3584 /* Calculate the real rcw later - for now make it
3585 * look like rcw is cheaper
3588 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3589 conf->rmw_level, (unsigned long long)recovery_cp,
3590 (unsigned long long)sh->sector);
3591 } else for (i = disks; i--; ) {
3592 /* would I have to read this buffer for read_modify_write */
3593 struct r5dev *dev = &sh->dev[i];
3594 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3595 !test_bit(R5_LOCKED, &dev->flags) &&
3596 !(test_bit(R5_UPTODATE, &dev->flags) ||
3597 test_bit(R5_Wantcompute, &dev->flags))) {
3598 if (test_bit(R5_Insync, &dev->flags))
3601 rmw += 2*disks; /* cannot read it */
3603 /* Would I have to read this buffer for reconstruct_write */
3604 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3605 i != sh->pd_idx && i != sh->qd_idx &&
3606 !test_bit(R5_LOCKED, &dev->flags) &&
3607 !(test_bit(R5_UPTODATE, &dev->flags) ||
3608 test_bit(R5_Wantcompute, &dev->flags))) {
3609 if (test_bit(R5_Insync, &dev->flags))
3615 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3616 (unsigned long long)sh->sector, rmw, rcw);
3617 set_bit(STRIPE_HANDLE, &sh->state);
3618 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3619 /* prefer read-modify-write, but need to get some data */
3620 if (conf->mddev->queue)
3621 blk_add_trace_msg(conf->mddev->queue,
3622 "raid5 rmw %llu %d",
3623 (unsigned long long)sh->sector, rmw);
3624 for (i = disks; i--; ) {
3625 struct r5dev *dev = &sh->dev[i];
3626 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3627 !test_bit(R5_LOCKED, &dev->flags) &&
3628 !(test_bit(R5_UPTODATE, &dev->flags) ||
3629 test_bit(R5_Wantcompute, &dev->flags)) &&
3630 test_bit(R5_Insync, &dev->flags)) {
3631 if (test_bit(STRIPE_PREREAD_ACTIVE,
3633 pr_debug("Read_old block %d for r-m-w\n",
3635 set_bit(R5_LOCKED, &dev->flags);
3636 set_bit(R5_Wantread, &dev->flags);
3639 set_bit(STRIPE_DELAYED, &sh->state);
3640 set_bit(STRIPE_HANDLE, &sh->state);
3645 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3646 /* want reconstruct write, but need to get some data */
3649 for (i = disks; i--; ) {
3650 struct r5dev *dev = &sh->dev[i];
3651 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3652 i != sh->pd_idx && i != sh->qd_idx &&
3653 !test_bit(R5_LOCKED, &dev->flags) &&
3654 !(test_bit(R5_UPTODATE, &dev->flags) ||
3655 test_bit(R5_Wantcompute, &dev->flags))) {
3657 if (test_bit(R5_Insync, &dev->flags) &&
3658 test_bit(STRIPE_PREREAD_ACTIVE,
3660 pr_debug("Read_old block "
3661 "%d for Reconstruct\n", i);
3662 set_bit(R5_LOCKED, &dev->flags);
3663 set_bit(R5_Wantread, &dev->flags);
3667 set_bit(STRIPE_DELAYED, &sh->state);
3668 set_bit(STRIPE_HANDLE, &sh->state);
3672 if (rcw && conf->mddev->queue)
3673 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3674 (unsigned long long)sh->sector,
3675 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3678 if (rcw > disks && rmw > disks &&
3679 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3680 set_bit(STRIPE_DELAYED, &sh->state);
3682 /* now if nothing is locked, and if we have enough data,
3683 * we can start a write request
3685 /* since handle_stripe can be called at any time we need to handle the
3686 * case where a compute block operation has been submitted and then a
3687 * subsequent call wants to start a write request. raid_run_ops only
3688 * handles the case where compute block and reconstruct are requested
3689 * simultaneously. If this is not the case then new writes need to be
3690 * held off until the compute completes.
3692 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3693 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3694 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3695 schedule_reconstruction(sh, s, rcw == 0, 0);
3698 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3699 struct stripe_head_state *s, int disks)
3701 struct r5dev *dev = NULL;
3703 BUG_ON(sh->batch_head);
3704 set_bit(STRIPE_HANDLE, &sh->state);
3706 switch (sh->check_state) {
3707 case check_state_idle:
3708 /* start a new check operation if there are no failures */
3709 if (s->failed == 0) {
3710 BUG_ON(s->uptodate != disks);
3711 sh->check_state = check_state_run;
3712 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3713 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3717 dev = &sh->dev[s->failed_num[0]];
3719 case check_state_compute_result:
3720 sh->check_state = check_state_idle;
3722 dev = &sh->dev[sh->pd_idx];
3724 /* check that a write has not made the stripe insync */
3725 if (test_bit(STRIPE_INSYNC, &sh->state))
3728 /* either failed parity check, or recovery is happening */
3729 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3730 BUG_ON(s->uptodate != disks);
3732 set_bit(R5_LOCKED, &dev->flags);
3734 set_bit(R5_Wantwrite, &dev->flags);
3736 clear_bit(STRIPE_DEGRADED, &sh->state);
3737 set_bit(STRIPE_INSYNC, &sh->state);
3739 case check_state_run:
3740 break; /* we will be called again upon completion */
3741 case check_state_check_result:
3742 sh->check_state = check_state_idle;
3744 /* if a failure occurred during the check operation, leave
3745 * STRIPE_INSYNC not set and let the stripe be handled again
3750 /* handle a successful check operation, if parity is correct
3751 * we are done. Otherwise update the mismatch count and repair
3752 * parity if !MD_RECOVERY_CHECK
3754 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3755 /* parity is correct (on disc,
3756 * not in buffer any more)
3758 set_bit(STRIPE_INSYNC, &sh->state);
3760 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3761 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3762 /* don't try to repair!! */
3763 set_bit(STRIPE_INSYNC, &sh->state);
3765 sh->check_state = check_state_compute_run;
3766 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3767 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3768 set_bit(R5_Wantcompute,
3769 &sh->dev[sh->pd_idx].flags);
3770 sh->ops.target = sh->pd_idx;
3771 sh->ops.target2 = -1;
3776 case check_state_compute_run:
3779 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3780 __func__, sh->check_state,
3781 (unsigned long long) sh->sector);
3786 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3787 struct stripe_head_state *s,
3790 int pd_idx = sh->pd_idx;
3791 int qd_idx = sh->qd_idx;
3794 BUG_ON(sh->batch_head);
3795 set_bit(STRIPE_HANDLE, &sh->state);
3797 BUG_ON(s->failed > 2);
3799 /* Want to check and possibly repair P and Q.
3800 * However there could be one 'failed' device, in which
3801 * case we can only check one of them, possibly using the
3802 * other to generate missing data
3805 switch (sh->check_state) {
3806 case check_state_idle:
3807 /* start a new check operation if there are < 2 failures */
3808 if (s->failed == s->q_failed) {
3809 /* The only possible failed device holds Q, so it
3810 * makes sense to check P (If anything else were failed,
3811 * we would have used P to recreate it).
3813 sh->check_state = check_state_run;
3815 if (!s->q_failed && s->failed < 2) {
3816 /* Q is not failed, and we didn't use it to generate
3817 * anything, so it makes sense to check it
3819 if (sh->check_state == check_state_run)
3820 sh->check_state = check_state_run_pq;
3822 sh->check_state = check_state_run_q;
3825 /* discard potentially stale zero_sum_result */
3826 sh->ops.zero_sum_result = 0;
3828 if (sh->check_state == check_state_run) {
3829 /* async_xor_zero_sum destroys the contents of P */
3830 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3833 if (sh->check_state >= check_state_run &&
3834 sh->check_state <= check_state_run_pq) {
3835 /* async_syndrome_zero_sum preserves P and Q, so
3836 * no need to mark them !uptodate here
3838 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3842 /* we have 2-disk failure */
3843 BUG_ON(s->failed != 2);
3845 case check_state_compute_result:
3846 sh->check_state = check_state_idle;
3848 /* check that a write has not made the stripe insync */
3849 if (test_bit(STRIPE_INSYNC, &sh->state))
3852 /* now write out any block on a failed drive,
3853 * or P or Q if they were recomputed
3855 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3856 if (s->failed == 2) {
3857 dev = &sh->dev[s->failed_num[1]];
3859 set_bit(R5_LOCKED, &dev->flags);
3860 set_bit(R5_Wantwrite, &dev->flags);
3862 if (s->failed >= 1) {
3863 dev = &sh->dev[s->failed_num[0]];
3865 set_bit(R5_LOCKED, &dev->flags);
3866 set_bit(R5_Wantwrite, &dev->flags);
3868 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3869 dev = &sh->dev[pd_idx];
3871 set_bit(R5_LOCKED, &dev->flags);
3872 set_bit(R5_Wantwrite, &dev->flags);
3874 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3875 dev = &sh->dev[qd_idx];
3877 set_bit(R5_LOCKED, &dev->flags);
3878 set_bit(R5_Wantwrite, &dev->flags);
3880 clear_bit(STRIPE_DEGRADED, &sh->state);
3882 set_bit(STRIPE_INSYNC, &sh->state);
3884 case check_state_run:
3885 case check_state_run_q:
3886 case check_state_run_pq:
3887 break; /* we will be called again upon completion */
3888 case check_state_check_result:
3889 sh->check_state = check_state_idle;
3891 /* handle a successful check operation, if parity is correct
3892 * we are done. Otherwise update the mismatch count and repair
3893 * parity if !MD_RECOVERY_CHECK
3895 if (sh->ops.zero_sum_result == 0) {
3896 /* both parities are correct */
3898 set_bit(STRIPE_INSYNC, &sh->state);
3900 /* in contrast to the raid5 case we can validate
3901 * parity, but still have a failure to write
3904 sh->check_state = check_state_compute_result;
3905 /* Returning at this point means that we may go
3906 * off and bring p and/or q uptodate again so
3907 * we make sure to check zero_sum_result again
3908 * to verify if p or q need writeback
3912 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3913 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3914 /* don't try to repair!! */
3915 set_bit(STRIPE_INSYNC, &sh->state);
3917 int *target = &sh->ops.target;
3919 sh->ops.target = -1;
3920 sh->ops.target2 = -1;
3921 sh->check_state = check_state_compute_run;
3922 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3923 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3924 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3925 set_bit(R5_Wantcompute,
3926 &sh->dev[pd_idx].flags);
3928 target = &sh->ops.target2;
3931 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3932 set_bit(R5_Wantcompute,
3933 &sh->dev[qd_idx].flags);
3940 case check_state_compute_run:
3943 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3944 __func__, sh->check_state,
3945 (unsigned long long) sh->sector);
3950 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3954 /* We have read all the blocks in this stripe and now we need to
3955 * copy some of them into a target stripe for expand.
3957 struct dma_async_tx_descriptor *tx = NULL;
3958 BUG_ON(sh->batch_head);
3959 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3960 for (i = 0; i < sh->disks; i++)
3961 if (i != sh->pd_idx && i != sh->qd_idx) {
3963 struct stripe_head *sh2;
3964 struct async_submit_ctl submit;
3966 sector_t bn = compute_blocknr(sh, i, 1);
3967 sector_t s = raid5_compute_sector(conf, bn, 0,
3969 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3971 /* so far only the early blocks of this stripe
3972 * have been requested. When later blocks
3973 * get requested, we will try again
3976 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3977 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3978 /* must have already done this block */
3979 release_stripe(sh2);
3983 /* place all the copies on one channel */
3984 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3985 tx = async_memcpy(sh2->dev[dd_idx].page,
3986 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3989 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3990 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3991 for (j = 0; j < conf->raid_disks; j++)
3992 if (j != sh2->pd_idx &&
3994 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3996 if (j == conf->raid_disks) {
3997 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3998 set_bit(STRIPE_HANDLE, &sh2->state);
4000 release_stripe(sh2);
4003 /* done submitting copies, wait for them to complete */
4004 async_tx_quiesce(&tx);
4008 * handle_stripe - do things to a stripe.
4010 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4011 * state of various bits to see what needs to be done.
4013 * return some read requests which now have data
4014 * return some write requests which are safely on storage
4015 * schedule a read on some buffers
4016 * schedule a write of some buffers
4017 * return confirmation of parity correctness
4021 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4023 struct r5conf *conf = sh->raid_conf;
4024 int disks = sh->disks;
4027 int do_recovery = 0;
4029 memset(s, 0, sizeof(*s));
4031 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4032 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4033 s->failed_num[0] = -1;
4034 s->failed_num[1] = -1;
4036 /* Now to look around and see what can be done */
4038 for (i=disks; i--; ) {
4039 struct md_rdev *rdev;
4046 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4048 dev->toread, dev->towrite, dev->written);
4049 /* maybe we can reply to a read
4051 * new wantfill requests are only permitted while
4052 * ops_complete_biofill is guaranteed to be inactive
4054 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4055 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4056 set_bit(R5_Wantfill, &dev->flags);
4058 /* now count some things */
4059 if (test_bit(R5_LOCKED, &dev->flags))
4061 if (test_bit(R5_UPTODATE, &dev->flags))
4063 if (test_bit(R5_Wantcompute, &dev->flags)) {
4065 BUG_ON(s->compute > 2);
4068 if (test_bit(R5_Wantfill, &dev->flags))
4070 else if (dev->toread)
4074 if (!test_bit(R5_OVERWRITE, &dev->flags))
4079 /* Prefer to use the replacement for reads, but only
4080 * if it is recovered enough and has no bad blocks.
4082 rdev = rcu_dereference(conf->disks[i].replacement);
4083 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4084 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4085 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4086 &first_bad, &bad_sectors))
4087 set_bit(R5_ReadRepl, &dev->flags);
4090 set_bit(R5_NeedReplace, &dev->flags);
4091 rdev = rcu_dereference(conf->disks[i].rdev);
4092 clear_bit(R5_ReadRepl, &dev->flags);
4094 if (rdev && test_bit(Faulty, &rdev->flags))
4097 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4098 &first_bad, &bad_sectors);
4099 if (s->blocked_rdev == NULL
4100 && (test_bit(Blocked, &rdev->flags)
4103 set_bit(BlockedBadBlocks,
4105 s->blocked_rdev = rdev;
4106 atomic_inc(&rdev->nr_pending);
4109 clear_bit(R5_Insync, &dev->flags);
4113 /* also not in-sync */
4114 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4115 test_bit(R5_UPTODATE, &dev->flags)) {
4116 /* treat as in-sync, but with a read error
4117 * which we can now try to correct
4119 set_bit(R5_Insync, &dev->flags);
4120 set_bit(R5_ReadError, &dev->flags);
4122 } else if (test_bit(In_sync, &rdev->flags))
4123 set_bit(R5_Insync, &dev->flags);
4124 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4125 /* in sync if before recovery_offset */
4126 set_bit(R5_Insync, &dev->flags);
4127 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4128 test_bit(R5_Expanded, &dev->flags))
4129 /* If we've reshaped into here, we assume it is Insync.
4130 * We will shortly update recovery_offset to make
4133 set_bit(R5_Insync, &dev->flags);
4135 if (test_bit(R5_WriteError, &dev->flags)) {
4136 /* This flag does not apply to '.replacement'
4137 * only to .rdev, so make sure to check that*/
4138 struct md_rdev *rdev2 = rcu_dereference(
4139 conf->disks[i].rdev);
4141 clear_bit(R5_Insync, &dev->flags);
4142 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4143 s->handle_bad_blocks = 1;
4144 atomic_inc(&rdev2->nr_pending);
4146 clear_bit(R5_WriteError, &dev->flags);
4148 if (test_bit(R5_MadeGood, &dev->flags)) {
4149 /* This flag does not apply to '.replacement'
4150 * only to .rdev, so make sure to check that*/
4151 struct md_rdev *rdev2 = rcu_dereference(
4152 conf->disks[i].rdev);
4153 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4154 s->handle_bad_blocks = 1;
4155 atomic_inc(&rdev2->nr_pending);
4157 clear_bit(R5_MadeGood, &dev->flags);
4159 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4160 struct md_rdev *rdev2 = rcu_dereference(
4161 conf->disks[i].replacement);
4162 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4163 s->handle_bad_blocks = 1;
4164 atomic_inc(&rdev2->nr_pending);
4166 clear_bit(R5_MadeGoodRepl, &dev->flags);
4168 if (!test_bit(R5_Insync, &dev->flags)) {
4169 /* The ReadError flag will just be confusing now */
4170 clear_bit(R5_ReadError, &dev->flags);
4171 clear_bit(R5_ReWrite, &dev->flags);
4173 if (test_bit(R5_ReadError, &dev->flags))
4174 clear_bit(R5_Insync, &dev->flags);
4175 if (!test_bit(R5_Insync, &dev->flags)) {
4177 s->failed_num[s->failed] = i;
4179 if (rdev && !test_bit(Faulty, &rdev->flags))
4183 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4184 /* If there is a failed device being replaced,
4185 * we must be recovering.
4186 * else if we are after recovery_cp, we must be syncing
4187 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4188 * else we can only be replacing
4189 * sync and recovery both need to read all devices, and so
4190 * use the same flag.
4193 sh->sector >= conf->mddev->recovery_cp ||
4194 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4202 static int clear_batch_ready(struct stripe_head *sh)
4204 /* Return '1' if this is a member of batch, or
4205 * '0' if it is a lone stripe or a head which can now be
4208 struct stripe_head *tmp;
4209 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4210 return (sh->batch_head && sh->batch_head != sh);
4211 spin_lock(&sh->stripe_lock);
4212 if (!sh->batch_head) {
4213 spin_unlock(&sh->stripe_lock);
4218 * this stripe could be added to a batch list before we check
4219 * BATCH_READY, skips it
4221 if (sh->batch_head != sh) {
4222 spin_unlock(&sh->stripe_lock);
4225 spin_lock(&sh->batch_lock);
4226 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4227 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4228 spin_unlock(&sh->batch_lock);
4229 spin_unlock(&sh->stripe_lock);
4232 * BATCH_READY is cleared, no new stripes can be added.
4233 * batch_list can be accessed without lock
4238 static void break_stripe_batch_list(struct stripe_head *head_sh)
4240 struct stripe_head *sh, *next;
4244 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4246 list_del_init(&sh->batch_list);
4248 set_mask_bits(&sh->state, ~STRIPE_EXPAND_SYNC_FLAG,
4249 head_sh->state & ~((1 << STRIPE_ACTIVE) |
4250 (1 << STRIPE_PREREAD_ACTIVE) |
4251 (1 << STRIPE_DEGRADED) |
4252 STRIPE_EXPAND_SYNC_FLAG));
4253 sh->check_state = head_sh->check_state;
4254 sh->reconstruct_state = head_sh->reconstruct_state;
4255 for (i = 0; i < sh->disks; i++) {
4256 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4258 sh->dev[i].flags = head_sh->dev[i].flags &
4259 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4261 spin_lock_irq(&sh->stripe_lock);
4262 sh->batch_head = NULL;
4263 spin_unlock_irq(&sh->stripe_lock);
4265 set_bit(STRIPE_HANDLE, &sh->state);
4268 spin_lock_irq(&head_sh->stripe_lock);
4269 head_sh->batch_head = NULL;
4270 spin_unlock_irq(&head_sh->stripe_lock);
4271 for (i = 0; i < head_sh->disks; i++)
4272 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4276 wake_up(&head_sh->raid_conf->wait_for_overlap);
4279 static void handle_stripe(struct stripe_head *sh)
4281 struct stripe_head_state s;
4282 struct r5conf *conf = sh->raid_conf;
4285 int disks = sh->disks;
4286 struct r5dev *pdev, *qdev;
4288 clear_bit(STRIPE_HANDLE, &sh->state);
4289 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4290 /* already being handled, ensure it gets handled
4291 * again when current action finishes */
4292 set_bit(STRIPE_HANDLE, &sh->state);
4296 if (clear_batch_ready(sh) ) {
4297 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4301 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4302 break_stripe_batch_list(sh);
4304 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4305 spin_lock(&sh->stripe_lock);
4306 /* Cannot process 'sync' concurrently with 'discard' */
4307 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4308 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4309 set_bit(STRIPE_SYNCING, &sh->state);
4310 clear_bit(STRIPE_INSYNC, &sh->state);
4311 clear_bit(STRIPE_REPLACED, &sh->state);
4313 spin_unlock(&sh->stripe_lock);
4315 clear_bit(STRIPE_DELAYED, &sh->state);
4317 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4318 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4319 (unsigned long long)sh->sector, sh->state,
4320 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4321 sh->check_state, sh->reconstruct_state);
4323 analyse_stripe(sh, &s);
4325 if (s.handle_bad_blocks) {
4326 set_bit(STRIPE_HANDLE, &sh->state);
4330 if (unlikely(s.blocked_rdev)) {
4331 if (s.syncing || s.expanding || s.expanded ||
4332 s.replacing || s.to_write || s.written) {
4333 set_bit(STRIPE_HANDLE, &sh->state);
4336 /* There is nothing for the blocked_rdev to block */
4337 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4338 s.blocked_rdev = NULL;
4341 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4342 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4343 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4346 pr_debug("locked=%d uptodate=%d to_read=%d"
4347 " to_write=%d failed=%d failed_num=%d,%d\n",
4348 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4349 s.failed_num[0], s.failed_num[1]);
4350 /* check if the array has lost more than max_degraded devices and,
4351 * if so, some requests might need to be failed.
4353 if (s.failed > conf->max_degraded) {
4354 sh->check_state = 0;
4355 sh->reconstruct_state = 0;
4356 if (s.to_read+s.to_write+s.written)
4357 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4358 if (s.syncing + s.replacing)
4359 handle_failed_sync(conf, sh, &s);
4362 /* Now we check to see if any write operations have recently
4366 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4368 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4369 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4370 sh->reconstruct_state = reconstruct_state_idle;
4372 /* All the 'written' buffers and the parity block are ready to
4373 * be written back to disk
4375 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4376 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4377 BUG_ON(sh->qd_idx >= 0 &&
4378 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4379 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4380 for (i = disks; i--; ) {
4381 struct r5dev *dev = &sh->dev[i];
4382 if (test_bit(R5_LOCKED, &dev->flags) &&
4383 (i == sh->pd_idx || i == sh->qd_idx ||
4385 pr_debug("Writing block %d\n", i);
4386 set_bit(R5_Wantwrite, &dev->flags);
4391 if (!test_bit(R5_Insync, &dev->flags) ||
4392 ((i == sh->pd_idx || i == sh->qd_idx) &&
4394 set_bit(STRIPE_INSYNC, &sh->state);
4397 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4398 s.dec_preread_active = 1;
4402 * might be able to return some write requests if the parity blocks
4403 * are safe, or on a failed drive
4405 pdev = &sh->dev[sh->pd_idx];
4406 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4407 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4408 qdev = &sh->dev[sh->qd_idx];
4409 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4410 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4414 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4415 && !test_bit(R5_LOCKED, &pdev->flags)
4416 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4417 test_bit(R5_Discard, &pdev->flags))))) &&
4418 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4419 && !test_bit(R5_LOCKED, &qdev->flags)
4420 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4421 test_bit(R5_Discard, &qdev->flags))))))
4422 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4424 /* Now we might consider reading some blocks, either to check/generate
4425 * parity, or to satisfy requests
4426 * or to load a block that is being partially written.
4428 if (s.to_read || s.non_overwrite
4429 || (conf->level == 6 && s.to_write && s.failed)
4430 || (s.syncing && (s.uptodate + s.compute < disks))
4433 handle_stripe_fill(sh, &s, disks);
4435 /* Now to consider new write requests and what else, if anything
4436 * should be read. We do not handle new writes when:
4437 * 1/ A 'write' operation (copy+xor) is already in flight.
4438 * 2/ A 'check' operation is in flight, as it may clobber the parity
4441 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4442 handle_stripe_dirtying(conf, sh, &s, disks);
4444 /* maybe we need to check and possibly fix the parity for this stripe
4445 * Any reads will already have been scheduled, so we just see if enough
4446 * data is available. The parity check is held off while parity
4447 * dependent operations are in flight.
4449 if (sh->check_state ||
4450 (s.syncing && s.locked == 0 &&
4451 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4452 !test_bit(STRIPE_INSYNC, &sh->state))) {
4453 if (conf->level == 6)
4454 handle_parity_checks6(conf, sh, &s, disks);
4456 handle_parity_checks5(conf, sh, &s, disks);
4459 if ((s.replacing || s.syncing) && s.locked == 0
4460 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4461 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4462 /* Write out to replacement devices where possible */
4463 for (i = 0; i < conf->raid_disks; i++)
4464 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4465 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4466 set_bit(R5_WantReplace, &sh->dev[i].flags);
4467 set_bit(R5_LOCKED, &sh->dev[i].flags);
4471 set_bit(STRIPE_INSYNC, &sh->state);
4472 set_bit(STRIPE_REPLACED, &sh->state);
4474 if ((s.syncing || s.replacing) && s.locked == 0 &&
4475 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4476 test_bit(STRIPE_INSYNC, &sh->state)) {
4477 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4478 clear_bit(STRIPE_SYNCING, &sh->state);
4479 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4480 wake_up(&conf->wait_for_overlap);
4483 /* If the failed drives are just a ReadError, then we might need
4484 * to progress the repair/check process
4486 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4487 for (i = 0; i < s.failed; i++) {
4488 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4489 if (test_bit(R5_ReadError, &dev->flags)
4490 && !test_bit(R5_LOCKED, &dev->flags)
4491 && test_bit(R5_UPTODATE, &dev->flags)
4493 if (!test_bit(R5_ReWrite, &dev->flags)) {
4494 set_bit(R5_Wantwrite, &dev->flags);
4495 set_bit(R5_ReWrite, &dev->flags);
4496 set_bit(R5_LOCKED, &dev->flags);
4499 /* let's read it back */
4500 set_bit(R5_Wantread, &dev->flags);
4501 set_bit(R5_LOCKED, &dev->flags);
4507 /* Finish reconstruct operations initiated by the expansion process */
4508 if (sh->reconstruct_state == reconstruct_state_result) {
4509 struct stripe_head *sh_src
4510 = get_active_stripe(conf, sh->sector, 1, 1, 1);
4511 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4512 /* sh cannot be written until sh_src has been read.
4513 * so arrange for sh to be delayed a little
4515 set_bit(STRIPE_DELAYED, &sh->state);
4516 set_bit(STRIPE_HANDLE, &sh->state);
4517 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4519 atomic_inc(&conf->preread_active_stripes);
4520 release_stripe(sh_src);
4524 release_stripe(sh_src);
4526 sh->reconstruct_state = reconstruct_state_idle;
4527 clear_bit(STRIPE_EXPANDING, &sh->state);
4528 for (i = conf->raid_disks; i--; ) {
4529 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4530 set_bit(R5_LOCKED, &sh->dev[i].flags);
4535 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4536 !sh->reconstruct_state) {
4537 /* Need to write out all blocks after computing parity */
4538 sh->disks = conf->raid_disks;
4539 stripe_set_idx(sh->sector, conf, 0, sh);
4540 schedule_reconstruction(sh, &s, 1, 1);
4541 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4542 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4543 atomic_dec(&conf->reshape_stripes);
4544 wake_up(&conf->wait_for_overlap);
4545 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4548 if (s.expanding && s.locked == 0 &&
4549 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4550 handle_stripe_expansion(conf, sh);
4553 /* wait for this device to become unblocked */
4554 if (unlikely(s.blocked_rdev)) {
4555 if (conf->mddev->external)
4556 md_wait_for_blocked_rdev(s.blocked_rdev,
4559 /* Internal metadata will immediately
4560 * be written by raid5d, so we don't
4561 * need to wait here.
4563 rdev_dec_pending(s.blocked_rdev,
4567 if (s.handle_bad_blocks)
4568 for (i = disks; i--; ) {
4569 struct md_rdev *rdev;
4570 struct r5dev *dev = &sh->dev[i];
4571 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4572 /* We own a safe reference to the rdev */
4573 rdev = conf->disks[i].rdev;
4574 if (!rdev_set_badblocks(rdev, sh->sector,
4576 md_error(conf->mddev, rdev);
4577 rdev_dec_pending(rdev, conf->mddev);
4579 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4580 rdev = conf->disks[i].rdev;
4581 rdev_clear_badblocks(rdev, sh->sector,
4583 rdev_dec_pending(rdev, conf->mddev);
4585 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4586 rdev = conf->disks[i].replacement;
4588 /* rdev have been moved down */
4589 rdev = conf->disks[i].rdev;
4590 rdev_clear_badblocks(rdev, sh->sector,
4592 rdev_dec_pending(rdev, conf->mddev);
4597 raid_run_ops(sh, s.ops_request);
4601 if (s.dec_preread_active) {
4602 /* We delay this until after ops_run_io so that if make_request
4603 * is waiting on a flush, it won't continue until the writes
4604 * have actually been submitted.
4606 atomic_dec(&conf->preread_active_stripes);
4607 if (atomic_read(&conf->preread_active_stripes) <
4609 md_wakeup_thread(conf->mddev->thread);
4612 return_io(s.return_bi);
4614 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4617 static void raid5_activate_delayed(struct r5conf *conf)
4619 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4620 while (!list_empty(&conf->delayed_list)) {
4621 struct list_head *l = conf->delayed_list.next;
4622 struct stripe_head *sh;
4623 sh = list_entry(l, struct stripe_head, lru);
4625 clear_bit(STRIPE_DELAYED, &sh->state);
4626 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4627 atomic_inc(&conf->preread_active_stripes);
4628 list_add_tail(&sh->lru, &conf->hold_list);
4629 raid5_wakeup_stripe_thread(sh);
4634 static void activate_bit_delay(struct r5conf *conf,
4635 struct list_head *temp_inactive_list)
4637 /* device_lock is held */
4638 struct list_head head;
4639 list_add(&head, &conf->bitmap_list);
4640 list_del_init(&conf->bitmap_list);
4641 while (!list_empty(&head)) {
4642 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4644 list_del_init(&sh->lru);
4645 atomic_inc(&sh->count);
4646 hash = sh->hash_lock_index;
4647 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4651 static int raid5_congested(struct mddev *mddev, int bits)
4653 struct r5conf *conf = mddev->private;
4655 /* No difference between reads and writes. Just check
4656 * how busy the stripe_cache is
4659 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4663 if (atomic_read(&conf->empty_inactive_list_nr))
4669 /* We want read requests to align with chunks where possible,
4670 * but write requests don't need to.
4672 static int raid5_mergeable_bvec(struct mddev *mddev,
4673 struct bvec_merge_data *bvm,
4674 struct bio_vec *biovec)
4676 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4678 unsigned int chunk_sectors = mddev->chunk_sectors;
4679 unsigned int bio_sectors = bvm->bi_size >> 9;
4682 * always allow writes to be mergeable, read as well if array
4683 * is degraded as we'll go through stripe cache anyway.
4685 if ((bvm->bi_rw & 1) == WRITE || mddev->degraded)
4686 return biovec->bv_len;
4688 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4689 chunk_sectors = mddev->new_chunk_sectors;
4690 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4691 if (max < 0) max = 0;
4692 if (max <= biovec->bv_len && bio_sectors == 0)
4693 return biovec->bv_len;
4698 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4700 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4701 unsigned int chunk_sectors = mddev->chunk_sectors;
4702 unsigned int bio_sectors = bio_sectors(bio);
4704 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4705 chunk_sectors = mddev->new_chunk_sectors;
4706 return chunk_sectors >=
4707 ((sector & (chunk_sectors - 1)) + bio_sectors);
4711 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4712 * later sampled by raid5d.
4714 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4716 unsigned long flags;
4718 spin_lock_irqsave(&conf->device_lock, flags);
4720 bi->bi_next = conf->retry_read_aligned_list;
4721 conf->retry_read_aligned_list = bi;
4723 spin_unlock_irqrestore(&conf->device_lock, flags);
4724 md_wakeup_thread(conf->mddev->thread);
4727 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4731 bi = conf->retry_read_aligned;
4733 conf->retry_read_aligned = NULL;
4736 bi = conf->retry_read_aligned_list;
4738 conf->retry_read_aligned_list = bi->bi_next;
4741 * this sets the active strip count to 1 and the processed
4742 * strip count to zero (upper 8 bits)
4744 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4751 * The "raid5_align_endio" should check if the read succeeded and if it
4752 * did, call bio_endio on the original bio (having bio_put the new bio
4754 * If the read failed..
4756 static void raid5_align_endio(struct bio *bi, int error)
4758 struct bio* raid_bi = bi->bi_private;
4759 struct mddev *mddev;
4760 struct r5conf *conf;
4761 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4762 struct md_rdev *rdev;
4766 rdev = (void*)raid_bi->bi_next;
4767 raid_bi->bi_next = NULL;
4768 mddev = rdev->mddev;
4769 conf = mddev->private;
4771 rdev_dec_pending(rdev, conf->mddev);
4773 if (!error && uptodate) {
4774 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4776 bio_endio(raid_bi, 0);
4777 if (atomic_dec_and_test(&conf->active_aligned_reads))
4778 wake_up(&conf->wait_for_stripe);
4782 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4784 add_bio_to_retry(raid_bi, conf);
4787 static int bio_fits_rdev(struct bio *bi)
4789 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4791 if (bio_sectors(bi) > queue_max_sectors(q))
4793 blk_recount_segments(q, bi);
4794 if (bi->bi_phys_segments > queue_max_segments(q))
4797 if (q->merge_bvec_fn)
4798 /* it's too hard to apply the merge_bvec_fn at this stage,
4806 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4808 struct r5conf *conf = mddev->private;
4810 struct bio* align_bi;
4811 struct md_rdev *rdev;
4812 sector_t end_sector;
4814 if (!in_chunk_boundary(mddev, raid_bio)) {
4815 pr_debug("chunk_aligned_read : non aligned\n");
4819 * use bio_clone_mddev to make a copy of the bio
4821 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4825 * set bi_end_io to a new function, and set bi_private to the
4828 align_bi->bi_end_io = raid5_align_endio;
4829 align_bi->bi_private = raid_bio;
4833 align_bi->bi_iter.bi_sector =
4834 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4837 end_sector = bio_end_sector(align_bi);
4839 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4840 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4841 rdev->recovery_offset < end_sector) {
4842 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4844 (test_bit(Faulty, &rdev->flags) ||
4845 !(test_bit(In_sync, &rdev->flags) ||
4846 rdev->recovery_offset >= end_sector)))
4853 atomic_inc(&rdev->nr_pending);
4855 raid_bio->bi_next = (void*)rdev;
4856 align_bi->bi_bdev = rdev->bdev;
4857 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
4859 if (!bio_fits_rdev(align_bi) ||
4860 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4861 bio_sectors(align_bi),
4862 &first_bad, &bad_sectors)) {
4863 /* too big in some way, or has a known bad block */
4865 rdev_dec_pending(rdev, mddev);
4869 /* No reshape active, so we can trust rdev->data_offset */
4870 align_bi->bi_iter.bi_sector += rdev->data_offset;
4872 spin_lock_irq(&conf->device_lock);
4873 wait_event_lock_irq(conf->wait_for_stripe,
4876 atomic_inc(&conf->active_aligned_reads);
4877 spin_unlock_irq(&conf->device_lock);
4880 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4881 align_bi, disk_devt(mddev->gendisk),
4882 raid_bio->bi_iter.bi_sector);
4883 generic_make_request(align_bi);
4892 /* __get_priority_stripe - get the next stripe to process
4894 * Full stripe writes are allowed to pass preread active stripes up until
4895 * the bypass_threshold is exceeded. In general the bypass_count
4896 * increments when the handle_list is handled before the hold_list; however, it
4897 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4898 * stripe with in flight i/o. The bypass_count will be reset when the
4899 * head of the hold_list has changed, i.e. the head was promoted to the
4902 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4904 struct stripe_head *sh = NULL, *tmp;
4905 struct list_head *handle_list = NULL;
4906 struct r5worker_group *wg = NULL;
4908 if (conf->worker_cnt_per_group == 0) {
4909 handle_list = &conf->handle_list;
4910 } else if (group != ANY_GROUP) {
4911 handle_list = &conf->worker_groups[group].handle_list;
4912 wg = &conf->worker_groups[group];
4915 for (i = 0; i < conf->group_cnt; i++) {
4916 handle_list = &conf->worker_groups[i].handle_list;
4917 wg = &conf->worker_groups[i];
4918 if (!list_empty(handle_list))
4923 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4925 list_empty(handle_list) ? "empty" : "busy",
4926 list_empty(&conf->hold_list) ? "empty" : "busy",
4927 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4929 if (!list_empty(handle_list)) {
4930 sh = list_entry(handle_list->next, typeof(*sh), lru);
4932 if (list_empty(&conf->hold_list))
4933 conf->bypass_count = 0;
4934 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4935 if (conf->hold_list.next == conf->last_hold)
4936 conf->bypass_count++;
4938 conf->last_hold = conf->hold_list.next;
4939 conf->bypass_count -= conf->bypass_threshold;
4940 if (conf->bypass_count < 0)
4941 conf->bypass_count = 0;
4944 } else if (!list_empty(&conf->hold_list) &&
4945 ((conf->bypass_threshold &&
4946 conf->bypass_count > conf->bypass_threshold) ||
4947 atomic_read(&conf->pending_full_writes) == 0)) {
4949 list_for_each_entry(tmp, &conf->hold_list, lru) {
4950 if (conf->worker_cnt_per_group == 0 ||
4951 group == ANY_GROUP ||
4952 !cpu_online(tmp->cpu) ||
4953 cpu_to_group(tmp->cpu) == group) {
4960 conf->bypass_count -= conf->bypass_threshold;
4961 if (conf->bypass_count < 0)
4962 conf->bypass_count = 0;
4974 list_del_init(&sh->lru);
4975 BUG_ON(atomic_inc_return(&sh->count) != 1);
4979 struct raid5_plug_cb {
4980 struct blk_plug_cb cb;
4981 struct list_head list;
4982 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4985 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4987 struct raid5_plug_cb *cb = container_of(
4988 blk_cb, struct raid5_plug_cb, cb);
4989 struct stripe_head *sh;
4990 struct mddev *mddev = cb->cb.data;
4991 struct r5conf *conf = mddev->private;
4995 if (cb->list.next && !list_empty(&cb->list)) {
4996 spin_lock_irq(&conf->device_lock);
4997 while (!list_empty(&cb->list)) {
4998 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4999 list_del_init(&sh->lru);
5001 * avoid race release_stripe_plug() sees
5002 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5003 * is still in our list
5005 smp_mb__before_atomic();
5006 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5008 * STRIPE_ON_RELEASE_LIST could be set here. In that
5009 * case, the count is always > 1 here
5011 hash = sh->hash_lock_index;
5012 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5015 spin_unlock_irq(&conf->device_lock);
5017 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5018 NR_STRIPE_HASH_LOCKS);
5020 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5024 static void release_stripe_plug(struct mddev *mddev,
5025 struct stripe_head *sh)
5027 struct blk_plug_cb *blk_cb = blk_check_plugged(
5028 raid5_unplug, mddev,
5029 sizeof(struct raid5_plug_cb));
5030 struct raid5_plug_cb *cb;
5037 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5039 if (cb->list.next == NULL) {
5041 INIT_LIST_HEAD(&cb->list);
5042 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5043 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5046 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5047 list_add_tail(&sh->lru, &cb->list);
5052 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5054 struct r5conf *conf = mddev->private;
5055 sector_t logical_sector, last_sector;
5056 struct stripe_head *sh;
5060 if (mddev->reshape_position != MaxSector)
5061 /* Skip discard while reshape is happening */
5064 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5065 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5068 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5070 stripe_sectors = conf->chunk_sectors *
5071 (conf->raid_disks - conf->max_degraded);
5072 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5074 sector_div(last_sector, stripe_sectors);
5076 logical_sector *= conf->chunk_sectors;
5077 last_sector *= conf->chunk_sectors;
5079 for (; logical_sector < last_sector;
5080 logical_sector += STRIPE_SECTORS) {
5084 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
5085 prepare_to_wait(&conf->wait_for_overlap, &w,
5086 TASK_UNINTERRUPTIBLE);
5087 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5088 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5093 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5094 spin_lock_irq(&sh->stripe_lock);
5095 for (d = 0; d < conf->raid_disks; d++) {
5096 if (d == sh->pd_idx || d == sh->qd_idx)
5098 if (sh->dev[d].towrite || sh->dev[d].toread) {
5099 set_bit(R5_Overlap, &sh->dev[d].flags);
5100 spin_unlock_irq(&sh->stripe_lock);
5106 set_bit(STRIPE_DISCARD, &sh->state);
5107 finish_wait(&conf->wait_for_overlap, &w);
5108 sh->overwrite_disks = 0;
5109 for (d = 0; d < conf->raid_disks; d++) {
5110 if (d == sh->pd_idx || d == sh->qd_idx)
5112 sh->dev[d].towrite = bi;
5113 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5114 raid5_inc_bi_active_stripes(bi);
5115 sh->overwrite_disks++;
5117 spin_unlock_irq(&sh->stripe_lock);
5118 if (conf->mddev->bitmap) {
5120 d < conf->raid_disks - conf->max_degraded;
5122 bitmap_startwrite(mddev->bitmap,
5126 sh->bm_seq = conf->seq_flush + 1;
5127 set_bit(STRIPE_BIT_DELAY, &sh->state);
5130 set_bit(STRIPE_HANDLE, &sh->state);
5131 clear_bit(STRIPE_DELAYED, &sh->state);
5132 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5133 atomic_inc(&conf->preread_active_stripes);
5134 release_stripe_plug(mddev, sh);
5137 remaining = raid5_dec_bi_active_stripes(bi);
5138 if (remaining == 0) {
5139 md_write_end(mddev);
5144 static void make_request(struct mddev *mddev, struct bio * bi)
5146 struct r5conf *conf = mddev->private;
5148 sector_t new_sector;
5149 sector_t logical_sector, last_sector;
5150 struct stripe_head *sh;
5151 const int rw = bio_data_dir(bi);
5156 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5157 md_flush_request(mddev, bi);
5161 md_write_start(mddev, bi);
5164 * If array is degraded, better not do chunk aligned read because
5165 * later we might have to read it again in order to reconstruct
5166 * data on failed drives.
5168 if (rw == READ && mddev->degraded == 0 &&
5169 mddev->reshape_position == MaxSector &&
5170 chunk_aligned_read(mddev,bi))
5173 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5174 make_discard_request(mddev, bi);
5178 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5179 last_sector = bio_end_sector(bi);
5181 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5183 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5184 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5190 seq = read_seqcount_begin(&conf->gen_lock);
5193 prepare_to_wait(&conf->wait_for_overlap, &w,
5194 TASK_UNINTERRUPTIBLE);
5195 if (unlikely(conf->reshape_progress != MaxSector)) {
5196 /* spinlock is needed as reshape_progress may be
5197 * 64bit on a 32bit platform, and so it might be
5198 * possible to see a half-updated value
5199 * Of course reshape_progress could change after
5200 * the lock is dropped, so once we get a reference
5201 * to the stripe that we think it is, we will have
5204 spin_lock_irq(&conf->device_lock);
5205 if (mddev->reshape_backwards
5206 ? logical_sector < conf->reshape_progress
5207 : logical_sector >= conf->reshape_progress) {
5210 if (mddev->reshape_backwards
5211 ? logical_sector < conf->reshape_safe
5212 : logical_sector >= conf->reshape_safe) {
5213 spin_unlock_irq(&conf->device_lock);
5219 spin_unlock_irq(&conf->device_lock);
5222 new_sector = raid5_compute_sector(conf, logical_sector,
5225 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5226 (unsigned long long)new_sector,
5227 (unsigned long long)logical_sector);
5229 sh = get_active_stripe(conf, new_sector, previous,
5230 (bi->bi_rw&RWA_MASK), 0);
5232 if (unlikely(previous)) {
5233 /* expansion might have moved on while waiting for a
5234 * stripe, so we must do the range check again.
5235 * Expansion could still move past after this
5236 * test, but as we are holding a reference to
5237 * 'sh', we know that if that happens,
5238 * STRIPE_EXPANDING will get set and the expansion
5239 * won't proceed until we finish with the stripe.
5242 spin_lock_irq(&conf->device_lock);
5243 if (mddev->reshape_backwards
5244 ? logical_sector >= conf->reshape_progress
5245 : logical_sector < conf->reshape_progress)
5246 /* mismatch, need to try again */
5248 spin_unlock_irq(&conf->device_lock);
5256 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5257 /* Might have got the wrong stripe_head
5265 logical_sector >= mddev->suspend_lo &&
5266 logical_sector < mddev->suspend_hi) {
5268 /* As the suspend_* range is controlled by
5269 * userspace, we want an interruptible
5272 flush_signals(current);
5273 prepare_to_wait(&conf->wait_for_overlap,
5274 &w, TASK_INTERRUPTIBLE);
5275 if (logical_sector >= mddev->suspend_lo &&
5276 logical_sector < mddev->suspend_hi) {
5283 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5284 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5285 /* Stripe is busy expanding or
5286 * add failed due to overlap. Flush everything
5289 md_wakeup_thread(mddev->thread);
5295 set_bit(STRIPE_HANDLE, &sh->state);
5296 clear_bit(STRIPE_DELAYED, &sh->state);
5297 if ((!sh->batch_head || sh == sh->batch_head) &&
5298 (bi->bi_rw & REQ_SYNC) &&
5299 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5300 atomic_inc(&conf->preread_active_stripes);
5301 release_stripe_plug(mddev, sh);
5303 /* cannot get stripe for read-ahead, just give-up */
5304 clear_bit(BIO_UPTODATE, &bi->bi_flags);
5308 finish_wait(&conf->wait_for_overlap, &w);
5310 remaining = raid5_dec_bi_active_stripes(bi);
5311 if (remaining == 0) {
5314 md_write_end(mddev);
5316 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5322 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5324 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5326 /* reshaping is quite different to recovery/resync so it is
5327 * handled quite separately ... here.
5329 * On each call to sync_request, we gather one chunk worth of
5330 * destination stripes and flag them as expanding.
5331 * Then we find all the source stripes and request reads.
5332 * As the reads complete, handle_stripe will copy the data
5333 * into the destination stripe and release that stripe.
5335 struct r5conf *conf = mddev->private;
5336 struct stripe_head *sh;
5337 sector_t first_sector, last_sector;
5338 int raid_disks = conf->previous_raid_disks;
5339 int data_disks = raid_disks - conf->max_degraded;
5340 int new_data_disks = conf->raid_disks - conf->max_degraded;
5343 sector_t writepos, readpos, safepos;
5344 sector_t stripe_addr;
5345 int reshape_sectors;
5346 struct list_head stripes;
5348 if (sector_nr == 0) {
5349 /* If restarting in the middle, skip the initial sectors */
5350 if (mddev->reshape_backwards &&
5351 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5352 sector_nr = raid5_size(mddev, 0, 0)
5353 - conf->reshape_progress;
5354 } else if (!mddev->reshape_backwards &&
5355 conf->reshape_progress > 0)
5356 sector_nr = conf->reshape_progress;
5357 sector_div(sector_nr, new_data_disks);
5359 mddev->curr_resync_completed = sector_nr;
5360 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5366 /* We need to process a full chunk at a time.
5367 * If old and new chunk sizes differ, we need to process the
5370 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
5371 reshape_sectors = mddev->new_chunk_sectors;
5373 reshape_sectors = mddev->chunk_sectors;
5375 /* We update the metadata at least every 10 seconds, or when
5376 * the data about to be copied would over-write the source of
5377 * the data at the front of the range. i.e. one new_stripe
5378 * along from reshape_progress new_maps to after where
5379 * reshape_safe old_maps to
5381 writepos = conf->reshape_progress;
5382 sector_div(writepos, new_data_disks);
5383 readpos = conf->reshape_progress;
5384 sector_div(readpos, data_disks);
5385 safepos = conf->reshape_safe;
5386 sector_div(safepos, data_disks);
5387 if (mddev->reshape_backwards) {
5388 writepos -= min_t(sector_t, reshape_sectors, writepos);
5389 readpos += reshape_sectors;
5390 safepos += reshape_sectors;
5392 writepos += reshape_sectors;
5393 readpos -= min_t(sector_t, reshape_sectors, readpos);
5394 safepos -= min_t(sector_t, reshape_sectors, safepos);
5397 /* Having calculated the 'writepos' possibly use it
5398 * to set 'stripe_addr' which is where we will write to.
5400 if (mddev->reshape_backwards) {
5401 BUG_ON(conf->reshape_progress == 0);
5402 stripe_addr = writepos;
5403 BUG_ON((mddev->dev_sectors &
5404 ~((sector_t)reshape_sectors - 1))
5405 - reshape_sectors - stripe_addr
5408 BUG_ON(writepos != sector_nr + reshape_sectors);
5409 stripe_addr = sector_nr;
5412 /* 'writepos' is the most advanced device address we might write.
5413 * 'readpos' is the least advanced device address we might read.
5414 * 'safepos' is the least address recorded in the metadata as having
5416 * If there is a min_offset_diff, these are adjusted either by
5417 * increasing the safepos/readpos if diff is negative, or
5418 * increasing writepos if diff is positive.
5419 * If 'readpos' is then behind 'writepos', there is no way that we can
5420 * ensure safety in the face of a crash - that must be done by userspace
5421 * making a backup of the data. So in that case there is no particular
5422 * rush to update metadata.
5423 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5424 * update the metadata to advance 'safepos' to match 'readpos' so that
5425 * we can be safe in the event of a crash.
5426 * So we insist on updating metadata if safepos is behind writepos and
5427 * readpos is beyond writepos.
5428 * In any case, update the metadata every 10 seconds.
5429 * Maybe that number should be configurable, but I'm not sure it is
5430 * worth it.... maybe it could be a multiple of safemode_delay???
5432 if (conf->min_offset_diff < 0) {
5433 safepos += -conf->min_offset_diff;
5434 readpos += -conf->min_offset_diff;
5436 writepos += conf->min_offset_diff;
5438 if ((mddev->reshape_backwards
5439 ? (safepos > writepos && readpos < writepos)
5440 : (safepos < writepos && readpos > writepos)) ||
5441 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5442 /* Cannot proceed until we've updated the superblock... */
5443 wait_event(conf->wait_for_overlap,
5444 atomic_read(&conf->reshape_stripes)==0
5445 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5446 if (atomic_read(&conf->reshape_stripes) != 0)
5448 mddev->reshape_position = conf->reshape_progress;
5449 mddev->curr_resync_completed = sector_nr;
5450 conf->reshape_checkpoint = jiffies;
5451 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5452 md_wakeup_thread(mddev->thread);
5453 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5454 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5455 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5457 spin_lock_irq(&conf->device_lock);
5458 conf->reshape_safe = mddev->reshape_position;
5459 spin_unlock_irq(&conf->device_lock);
5460 wake_up(&conf->wait_for_overlap);
5461 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5464 INIT_LIST_HEAD(&stripes);
5465 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5467 int skipped_disk = 0;
5468 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5469 set_bit(STRIPE_EXPANDING, &sh->state);
5470 atomic_inc(&conf->reshape_stripes);
5471 /* If any of this stripe is beyond the end of the old
5472 * array, then we need to zero those blocks
5474 for (j=sh->disks; j--;) {
5476 if (j == sh->pd_idx)
5478 if (conf->level == 6 &&
5481 s = compute_blocknr(sh, j, 0);
5482 if (s < raid5_size(mddev, 0, 0)) {
5486 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5487 set_bit(R5_Expanded, &sh->dev[j].flags);
5488 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5490 if (!skipped_disk) {
5491 set_bit(STRIPE_EXPAND_READY, &sh->state);
5492 set_bit(STRIPE_HANDLE, &sh->state);
5494 list_add(&sh->lru, &stripes);
5496 spin_lock_irq(&conf->device_lock);
5497 if (mddev->reshape_backwards)
5498 conf->reshape_progress -= reshape_sectors * new_data_disks;
5500 conf->reshape_progress += reshape_sectors * new_data_disks;
5501 spin_unlock_irq(&conf->device_lock);
5502 /* Ok, those stripe are ready. We can start scheduling
5503 * reads on the source stripes.
5504 * The source stripes are determined by mapping the first and last
5505 * block on the destination stripes.
5508 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5511 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5512 * new_data_disks - 1),
5514 if (last_sector >= mddev->dev_sectors)
5515 last_sector = mddev->dev_sectors - 1;
5516 while (first_sector <= last_sector) {
5517 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
5518 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5519 set_bit(STRIPE_HANDLE, &sh->state);
5521 first_sector += STRIPE_SECTORS;
5523 /* Now that the sources are clearly marked, we can release
5524 * the destination stripes
5526 while (!list_empty(&stripes)) {
5527 sh = list_entry(stripes.next, struct stripe_head, lru);
5528 list_del_init(&sh->lru);
5531 /* If this takes us to the resync_max point where we have to pause,
5532 * then we need to write out the superblock.
5534 sector_nr += reshape_sectors;
5535 if ((sector_nr - mddev->curr_resync_completed) * 2
5536 >= mddev->resync_max - mddev->curr_resync_completed) {
5537 /* Cannot proceed until we've updated the superblock... */
5538 wait_event(conf->wait_for_overlap,
5539 atomic_read(&conf->reshape_stripes) == 0
5540 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5541 if (atomic_read(&conf->reshape_stripes) != 0)
5543 mddev->reshape_position = conf->reshape_progress;
5544 mddev->curr_resync_completed = sector_nr;
5545 conf->reshape_checkpoint = jiffies;
5546 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5547 md_wakeup_thread(mddev->thread);
5548 wait_event(mddev->sb_wait,
5549 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5550 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5551 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5553 spin_lock_irq(&conf->device_lock);
5554 conf->reshape_safe = mddev->reshape_position;
5555 spin_unlock_irq(&conf->device_lock);
5556 wake_up(&conf->wait_for_overlap);
5557 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5560 return reshape_sectors;
5563 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5565 struct r5conf *conf = mddev->private;
5566 struct stripe_head *sh;
5567 sector_t max_sector = mddev->dev_sectors;
5568 sector_t sync_blocks;
5569 int still_degraded = 0;
5572 if (sector_nr >= max_sector) {
5573 /* just being told to finish up .. nothing much to do */
5575 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5580 if (mddev->curr_resync < max_sector) /* aborted */
5581 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5583 else /* completed sync */
5585 bitmap_close_sync(mddev->bitmap);
5590 /* Allow raid5_quiesce to complete */
5591 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5593 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5594 return reshape_request(mddev, sector_nr, skipped);
5596 /* No need to check resync_max as we never do more than one
5597 * stripe, and as resync_max will always be on a chunk boundary,
5598 * if the check in md_do_sync didn't fire, there is no chance
5599 * of overstepping resync_max here
5602 /* if there is too many failed drives and we are trying
5603 * to resync, then assert that we are finished, because there is
5604 * nothing we can do.
5606 if (mddev->degraded >= conf->max_degraded &&
5607 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5608 sector_t rv = mddev->dev_sectors - sector_nr;
5612 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5614 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5615 sync_blocks >= STRIPE_SECTORS) {
5616 /* we can skip this block, and probably more */
5617 sync_blocks /= STRIPE_SECTORS;
5619 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5622 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5624 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5626 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5627 /* make sure we don't swamp the stripe cache if someone else
5628 * is trying to get access
5630 schedule_timeout_uninterruptible(1);
5632 /* Need to check if array will still be degraded after recovery/resync
5633 * Note in case of > 1 drive failures it's possible we're rebuilding
5634 * one drive while leaving another faulty drive in array.
5637 for (i = 0; i < conf->raid_disks; i++) {
5638 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5640 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5645 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5647 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5648 set_bit(STRIPE_HANDLE, &sh->state);
5652 return STRIPE_SECTORS;
5655 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5657 /* We may not be able to submit a whole bio at once as there
5658 * may not be enough stripe_heads available.
5659 * We cannot pre-allocate enough stripe_heads as we may need
5660 * more than exist in the cache (if we allow ever large chunks).
5661 * So we do one stripe head at a time and record in
5662 * ->bi_hw_segments how many have been done.
5664 * We *know* that this entire raid_bio is in one chunk, so
5665 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5667 struct stripe_head *sh;
5669 sector_t sector, logical_sector, last_sector;
5674 logical_sector = raid_bio->bi_iter.bi_sector &
5675 ~((sector_t)STRIPE_SECTORS-1);
5676 sector = raid5_compute_sector(conf, logical_sector,
5678 last_sector = bio_end_sector(raid_bio);
5680 for (; logical_sector < last_sector;
5681 logical_sector += STRIPE_SECTORS,
5682 sector += STRIPE_SECTORS,
5685 if (scnt < raid5_bi_processed_stripes(raid_bio))
5686 /* already done this stripe */
5689 sh = get_active_stripe(conf, sector, 0, 1, 1);
5692 /* failed to get a stripe - must wait */
5693 raid5_set_bi_processed_stripes(raid_bio, scnt);
5694 conf->retry_read_aligned = raid_bio;
5698 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5700 raid5_set_bi_processed_stripes(raid_bio, scnt);
5701 conf->retry_read_aligned = raid_bio;
5705 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5710 remaining = raid5_dec_bi_active_stripes(raid_bio);
5711 if (remaining == 0) {
5712 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5714 bio_endio(raid_bio, 0);
5716 if (atomic_dec_and_test(&conf->active_aligned_reads))
5717 wake_up(&conf->wait_for_stripe);
5721 static int handle_active_stripes(struct r5conf *conf, int group,
5722 struct r5worker *worker,
5723 struct list_head *temp_inactive_list)
5725 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5726 int i, batch_size = 0, hash;
5727 bool release_inactive = false;
5729 while (batch_size < MAX_STRIPE_BATCH &&
5730 (sh = __get_priority_stripe(conf, group)) != NULL)
5731 batch[batch_size++] = sh;
5733 if (batch_size == 0) {
5734 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5735 if (!list_empty(temp_inactive_list + i))
5737 if (i == NR_STRIPE_HASH_LOCKS)
5739 release_inactive = true;
5741 spin_unlock_irq(&conf->device_lock);
5743 release_inactive_stripe_list(conf, temp_inactive_list,
5744 NR_STRIPE_HASH_LOCKS);
5746 if (release_inactive) {
5747 spin_lock_irq(&conf->device_lock);
5751 for (i = 0; i < batch_size; i++)
5752 handle_stripe(batch[i]);
5756 spin_lock_irq(&conf->device_lock);
5757 for (i = 0; i < batch_size; i++) {
5758 hash = batch[i]->hash_lock_index;
5759 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5764 static void raid5_do_work(struct work_struct *work)
5766 struct r5worker *worker = container_of(work, struct r5worker, work);
5767 struct r5worker_group *group = worker->group;
5768 struct r5conf *conf = group->conf;
5769 int group_id = group - conf->worker_groups;
5771 struct blk_plug plug;
5773 pr_debug("+++ raid5worker active\n");
5775 blk_start_plug(&plug);
5777 spin_lock_irq(&conf->device_lock);
5779 int batch_size, released;
5781 released = release_stripe_list(conf, worker->temp_inactive_list);
5783 batch_size = handle_active_stripes(conf, group_id, worker,
5784 worker->temp_inactive_list);
5785 worker->working = false;
5786 if (!batch_size && !released)
5788 handled += batch_size;
5790 pr_debug("%d stripes handled\n", handled);
5792 spin_unlock_irq(&conf->device_lock);
5793 blk_finish_plug(&plug);
5795 pr_debug("--- raid5worker inactive\n");
5799 * This is our raid5 kernel thread.
5801 * We scan the hash table for stripes which can be handled now.
5802 * During the scan, completed stripes are saved for us by the interrupt
5803 * handler, so that they will not have to wait for our next wakeup.
5805 static void raid5d(struct md_thread *thread)
5807 struct mddev *mddev = thread->mddev;
5808 struct r5conf *conf = mddev->private;
5810 struct blk_plug plug;
5812 pr_debug("+++ raid5d active\n");
5814 md_check_recovery(mddev);
5816 blk_start_plug(&plug);
5818 spin_lock_irq(&conf->device_lock);
5821 int batch_size, released;
5823 released = release_stripe_list(conf, conf->temp_inactive_list);
5825 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5828 !list_empty(&conf->bitmap_list)) {
5829 /* Now is a good time to flush some bitmap updates */
5831 spin_unlock_irq(&conf->device_lock);
5832 bitmap_unplug(mddev->bitmap);
5833 spin_lock_irq(&conf->device_lock);
5834 conf->seq_write = conf->seq_flush;
5835 activate_bit_delay(conf, conf->temp_inactive_list);
5837 raid5_activate_delayed(conf);
5839 while ((bio = remove_bio_from_retry(conf))) {
5841 spin_unlock_irq(&conf->device_lock);
5842 ok = retry_aligned_read(conf, bio);
5843 spin_lock_irq(&conf->device_lock);
5849 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5850 conf->temp_inactive_list);
5851 if (!batch_size && !released)
5853 handled += batch_size;
5855 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5856 spin_unlock_irq(&conf->device_lock);
5857 md_check_recovery(mddev);
5858 spin_lock_irq(&conf->device_lock);
5861 pr_debug("%d stripes handled\n", handled);
5863 spin_unlock_irq(&conf->device_lock);
5864 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state)) {
5865 grow_one_stripe(conf, __GFP_NOWARN);
5866 /* Set flag even if allocation failed. This helps
5867 * slow down allocation requests when mem is short
5869 set_bit(R5_DID_ALLOC, &conf->cache_state);
5872 async_tx_issue_pending_all();
5873 blk_finish_plug(&plug);
5875 pr_debug("--- raid5d inactive\n");
5879 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5881 struct r5conf *conf;
5883 spin_lock(&mddev->lock);
5884 conf = mddev->private;
5886 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5887 spin_unlock(&mddev->lock);
5892 raid5_set_cache_size(struct mddev *mddev, int size)
5894 struct r5conf *conf = mddev->private;
5897 if (size <= 16 || size > 32768)
5900 conf->min_nr_stripes = size;
5901 while (size < conf->max_nr_stripes &&
5902 drop_one_stripe(conf))
5906 err = md_allow_write(mddev);
5910 while (size > conf->max_nr_stripes)
5911 if (!grow_one_stripe(conf, GFP_KERNEL))
5916 EXPORT_SYMBOL(raid5_set_cache_size);
5919 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5921 struct r5conf *conf;
5925 if (len >= PAGE_SIZE)
5927 if (kstrtoul(page, 10, &new))
5929 err = mddev_lock(mddev);
5932 conf = mddev->private;
5936 err = raid5_set_cache_size(mddev, new);
5937 mddev_unlock(mddev);
5942 static struct md_sysfs_entry
5943 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5944 raid5_show_stripe_cache_size,
5945 raid5_store_stripe_cache_size);
5948 raid5_show_rmw_level(struct mddev *mddev, char *page)
5950 struct r5conf *conf = mddev->private;
5952 return sprintf(page, "%d\n", conf->rmw_level);
5958 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
5960 struct r5conf *conf = mddev->private;
5966 if (len >= PAGE_SIZE)
5969 if (kstrtoul(page, 10, &new))
5972 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
5975 if (new != PARITY_DISABLE_RMW &&
5976 new != PARITY_ENABLE_RMW &&
5977 new != PARITY_PREFER_RMW)
5980 conf->rmw_level = new;
5984 static struct md_sysfs_entry
5985 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
5986 raid5_show_rmw_level,
5987 raid5_store_rmw_level);
5991 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5993 struct r5conf *conf;
5995 spin_lock(&mddev->lock);
5996 conf = mddev->private;
5998 ret = sprintf(page, "%d\n", conf->bypass_threshold);
5999 spin_unlock(&mddev->lock);
6004 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6006 struct r5conf *conf;
6010 if (len >= PAGE_SIZE)
6012 if (kstrtoul(page, 10, &new))
6015 err = mddev_lock(mddev);
6018 conf = mddev->private;
6021 else if (new > conf->min_nr_stripes)
6024 conf->bypass_threshold = new;
6025 mddev_unlock(mddev);
6029 static struct md_sysfs_entry
6030 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6032 raid5_show_preread_threshold,
6033 raid5_store_preread_threshold);
6036 raid5_show_skip_copy(struct mddev *mddev, char *page)
6038 struct r5conf *conf;
6040 spin_lock(&mddev->lock);
6041 conf = mddev->private;
6043 ret = sprintf(page, "%d\n", conf->skip_copy);
6044 spin_unlock(&mddev->lock);
6049 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6051 struct r5conf *conf;
6055 if (len >= PAGE_SIZE)
6057 if (kstrtoul(page, 10, &new))
6061 err = mddev_lock(mddev);
6064 conf = mddev->private;
6067 else if (new != conf->skip_copy) {
6068 mddev_suspend(mddev);
6069 conf->skip_copy = new;
6071 mddev->queue->backing_dev_info.capabilities |=
6072 BDI_CAP_STABLE_WRITES;
6074 mddev->queue->backing_dev_info.capabilities &=
6075 ~BDI_CAP_STABLE_WRITES;
6076 mddev_resume(mddev);
6078 mddev_unlock(mddev);
6082 static struct md_sysfs_entry
6083 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6084 raid5_show_skip_copy,
6085 raid5_store_skip_copy);
6088 stripe_cache_active_show(struct mddev *mddev, char *page)
6090 struct r5conf *conf = mddev->private;
6092 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6097 static struct md_sysfs_entry
6098 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6101 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6103 struct r5conf *conf;
6105 spin_lock(&mddev->lock);
6106 conf = mddev->private;
6108 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6109 spin_unlock(&mddev->lock);
6113 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6115 int *worker_cnt_per_group,
6116 struct r5worker_group **worker_groups);
6118 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6120 struct r5conf *conf;
6123 struct r5worker_group *new_groups, *old_groups;
6124 int group_cnt, worker_cnt_per_group;
6126 if (len >= PAGE_SIZE)
6128 if (kstrtoul(page, 10, &new))
6131 err = mddev_lock(mddev);
6134 conf = mddev->private;
6137 else if (new != conf->worker_cnt_per_group) {
6138 mddev_suspend(mddev);
6140 old_groups = conf->worker_groups;
6142 flush_workqueue(raid5_wq);
6144 err = alloc_thread_groups(conf, new,
6145 &group_cnt, &worker_cnt_per_group,
6148 spin_lock_irq(&conf->device_lock);
6149 conf->group_cnt = group_cnt;
6150 conf->worker_cnt_per_group = worker_cnt_per_group;
6151 conf->worker_groups = new_groups;
6152 spin_unlock_irq(&conf->device_lock);
6155 kfree(old_groups[0].workers);
6158 mddev_resume(mddev);
6160 mddev_unlock(mddev);
6165 static struct md_sysfs_entry
6166 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6167 raid5_show_group_thread_cnt,
6168 raid5_store_group_thread_cnt);
6170 static struct attribute *raid5_attrs[] = {
6171 &raid5_stripecache_size.attr,
6172 &raid5_stripecache_active.attr,
6173 &raid5_preread_bypass_threshold.attr,
6174 &raid5_group_thread_cnt.attr,
6175 &raid5_skip_copy.attr,
6176 &raid5_rmw_level.attr,
6179 static struct attribute_group raid5_attrs_group = {
6181 .attrs = raid5_attrs,
6184 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6186 int *worker_cnt_per_group,
6187 struct r5worker_group **worker_groups)
6191 struct r5worker *workers;
6193 *worker_cnt_per_group = cnt;
6196 *worker_groups = NULL;
6199 *group_cnt = num_possible_nodes();
6200 size = sizeof(struct r5worker) * cnt;
6201 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6202 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6203 *group_cnt, GFP_NOIO);
6204 if (!*worker_groups || !workers) {
6206 kfree(*worker_groups);
6210 for (i = 0; i < *group_cnt; i++) {
6211 struct r5worker_group *group;
6213 group = &(*worker_groups)[i];
6214 INIT_LIST_HEAD(&group->handle_list);
6216 group->workers = workers + i * cnt;
6218 for (j = 0; j < cnt; j++) {
6219 struct r5worker *worker = group->workers + j;
6220 worker->group = group;
6221 INIT_WORK(&worker->work, raid5_do_work);
6223 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6224 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6231 static void free_thread_groups(struct r5conf *conf)
6233 if (conf->worker_groups)
6234 kfree(conf->worker_groups[0].workers);
6235 kfree(conf->worker_groups);
6236 conf->worker_groups = NULL;
6240 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6242 struct r5conf *conf = mddev->private;
6245 sectors = mddev->dev_sectors;
6247 /* size is defined by the smallest of previous and new size */
6248 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6250 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6251 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
6252 return sectors * (raid_disks - conf->max_degraded);
6255 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6257 safe_put_page(percpu->spare_page);
6258 if (percpu->scribble)
6259 flex_array_free(percpu->scribble);
6260 percpu->spare_page = NULL;
6261 percpu->scribble = NULL;
6264 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6266 if (conf->level == 6 && !percpu->spare_page)
6267 percpu->spare_page = alloc_page(GFP_KERNEL);
6268 if (!percpu->scribble)
6269 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6270 conf->previous_raid_disks),
6271 max(conf->chunk_sectors,
6272 conf->prev_chunk_sectors)
6276 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6277 free_scratch_buffer(conf, percpu);
6284 static void raid5_free_percpu(struct r5conf *conf)
6291 #ifdef CONFIG_HOTPLUG_CPU
6292 unregister_cpu_notifier(&conf->cpu_notify);
6296 for_each_possible_cpu(cpu)
6297 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6300 free_percpu(conf->percpu);
6303 static void free_conf(struct r5conf *conf)
6305 if (conf->shrinker.seeks)
6306 unregister_shrinker(&conf->shrinker);
6307 free_thread_groups(conf);
6308 shrink_stripes(conf);
6309 raid5_free_percpu(conf);
6311 kfree(conf->stripe_hashtbl);
6315 #ifdef CONFIG_HOTPLUG_CPU
6316 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6319 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6320 long cpu = (long)hcpu;
6321 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6324 case CPU_UP_PREPARE:
6325 case CPU_UP_PREPARE_FROZEN:
6326 if (alloc_scratch_buffer(conf, percpu)) {
6327 pr_err("%s: failed memory allocation for cpu%ld\n",
6329 return notifier_from_errno(-ENOMEM);
6333 case CPU_DEAD_FROZEN:
6334 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6343 static int raid5_alloc_percpu(struct r5conf *conf)
6348 conf->percpu = alloc_percpu(struct raid5_percpu);
6352 #ifdef CONFIG_HOTPLUG_CPU
6353 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6354 conf->cpu_notify.priority = 0;
6355 err = register_cpu_notifier(&conf->cpu_notify);
6361 for_each_present_cpu(cpu) {
6362 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6364 pr_err("%s: failed memory allocation for cpu%ld\n",
6374 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6375 struct shrink_control *sc)
6377 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6379 while (ret < sc->nr_to_scan) {
6380 if (drop_one_stripe(conf) == 0)
6387 static unsigned long raid5_cache_count(struct shrinker *shrink,
6388 struct shrink_control *sc)
6390 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6392 if (conf->max_nr_stripes < conf->min_nr_stripes)
6393 /* unlikely, but not impossible */
6395 return conf->max_nr_stripes - conf->min_nr_stripes;
6398 static struct r5conf *setup_conf(struct mddev *mddev)
6400 struct r5conf *conf;
6401 int raid_disk, memory, max_disks;
6402 struct md_rdev *rdev;
6403 struct disk_info *disk;
6406 int group_cnt, worker_cnt_per_group;
6407 struct r5worker_group *new_group;
6409 if (mddev->new_level != 5
6410 && mddev->new_level != 4
6411 && mddev->new_level != 6) {
6412 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6413 mdname(mddev), mddev->new_level);
6414 return ERR_PTR(-EIO);
6416 if ((mddev->new_level == 5
6417 && !algorithm_valid_raid5(mddev->new_layout)) ||
6418 (mddev->new_level == 6
6419 && !algorithm_valid_raid6(mddev->new_layout))) {
6420 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6421 mdname(mddev), mddev->new_layout);
6422 return ERR_PTR(-EIO);
6424 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6425 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6426 mdname(mddev), mddev->raid_disks);
6427 return ERR_PTR(-EINVAL);
6430 if (!mddev->new_chunk_sectors ||
6431 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6432 !is_power_of_2(mddev->new_chunk_sectors)) {
6433 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6434 mdname(mddev), mddev->new_chunk_sectors << 9);
6435 return ERR_PTR(-EINVAL);
6438 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6441 /* Don't enable multi-threading by default*/
6442 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6444 conf->group_cnt = group_cnt;
6445 conf->worker_cnt_per_group = worker_cnt_per_group;
6446 conf->worker_groups = new_group;
6449 spin_lock_init(&conf->device_lock);
6450 seqcount_init(&conf->gen_lock);
6451 init_waitqueue_head(&conf->wait_for_stripe);
6452 init_waitqueue_head(&conf->wait_for_overlap);
6453 INIT_LIST_HEAD(&conf->handle_list);
6454 INIT_LIST_HEAD(&conf->hold_list);
6455 INIT_LIST_HEAD(&conf->delayed_list);
6456 INIT_LIST_HEAD(&conf->bitmap_list);
6457 init_llist_head(&conf->released_stripes);
6458 atomic_set(&conf->active_stripes, 0);
6459 atomic_set(&conf->preread_active_stripes, 0);
6460 atomic_set(&conf->active_aligned_reads, 0);
6461 conf->bypass_threshold = BYPASS_THRESHOLD;
6462 conf->recovery_disabled = mddev->recovery_disabled - 1;
6464 conf->raid_disks = mddev->raid_disks;
6465 if (mddev->reshape_position == MaxSector)
6466 conf->previous_raid_disks = mddev->raid_disks;
6468 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6469 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6471 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6476 conf->mddev = mddev;
6478 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6481 /* We init hash_locks[0] separately to that it can be used
6482 * as the reference lock in the spin_lock_nest_lock() call
6483 * in lock_all_device_hash_locks_irq in order to convince
6484 * lockdep that we know what we are doing.
6486 spin_lock_init(conf->hash_locks);
6487 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6488 spin_lock_init(conf->hash_locks + i);
6490 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6491 INIT_LIST_HEAD(conf->inactive_list + i);
6493 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6494 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6496 conf->level = mddev->new_level;
6497 conf->chunk_sectors = mddev->new_chunk_sectors;
6498 if (raid5_alloc_percpu(conf) != 0)
6501 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6503 rdev_for_each(rdev, mddev) {
6504 raid_disk = rdev->raid_disk;
6505 if (raid_disk >= max_disks
6508 disk = conf->disks + raid_disk;
6510 if (test_bit(Replacement, &rdev->flags)) {
6511 if (disk->replacement)
6513 disk->replacement = rdev;
6520 if (test_bit(In_sync, &rdev->flags)) {
6521 char b[BDEVNAME_SIZE];
6522 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6524 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6525 } else if (rdev->saved_raid_disk != raid_disk)
6526 /* Cannot rely on bitmap to complete recovery */
6530 conf->level = mddev->new_level;
6531 if (conf->level == 6) {
6532 conf->max_degraded = 2;
6533 if (raid6_call.xor_syndrome)
6534 conf->rmw_level = PARITY_ENABLE_RMW;
6536 conf->rmw_level = PARITY_DISABLE_RMW;
6538 conf->max_degraded = 1;
6539 conf->rmw_level = PARITY_ENABLE_RMW;
6541 conf->algorithm = mddev->new_layout;
6542 conf->reshape_progress = mddev->reshape_position;
6543 if (conf->reshape_progress != MaxSector) {
6544 conf->prev_chunk_sectors = mddev->chunk_sectors;
6545 conf->prev_algo = mddev->layout;
6548 conf->min_nr_stripes = NR_STRIPES;
6549 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6550 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6551 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6552 if (grow_stripes(conf, conf->min_nr_stripes)) {
6554 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6555 mdname(mddev), memory);
6558 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6559 mdname(mddev), memory);
6561 * Losing a stripe head costs more than the time to refill it,
6562 * it reduces the queue depth and so can hurt throughput.
6563 * So set it rather large, scaled by number of devices.
6565 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6566 conf->shrinker.scan_objects = raid5_cache_scan;
6567 conf->shrinker.count_objects = raid5_cache_count;
6568 conf->shrinker.batch = 128;
6569 conf->shrinker.flags = 0;
6570 register_shrinker(&conf->shrinker);
6572 sprintf(pers_name, "raid%d", mddev->new_level);
6573 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6574 if (!conf->thread) {
6576 "md/raid:%s: couldn't allocate thread.\n",
6586 return ERR_PTR(-EIO);
6588 return ERR_PTR(-ENOMEM);
6591 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6594 case ALGORITHM_PARITY_0:
6595 if (raid_disk < max_degraded)
6598 case ALGORITHM_PARITY_N:
6599 if (raid_disk >= raid_disks - max_degraded)
6602 case ALGORITHM_PARITY_0_6:
6603 if (raid_disk == 0 ||
6604 raid_disk == raid_disks - 1)
6607 case ALGORITHM_LEFT_ASYMMETRIC_6:
6608 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6609 case ALGORITHM_LEFT_SYMMETRIC_6:
6610 case ALGORITHM_RIGHT_SYMMETRIC_6:
6611 if (raid_disk == raid_disks - 1)
6617 static int run(struct mddev *mddev)
6619 struct r5conf *conf;
6620 int working_disks = 0;
6621 int dirty_parity_disks = 0;
6622 struct md_rdev *rdev;
6623 sector_t reshape_offset = 0;
6625 long long min_offset_diff = 0;
6628 if (mddev->recovery_cp != MaxSector)
6629 printk(KERN_NOTICE "md/raid:%s: not clean"
6630 " -- starting background reconstruction\n",
6633 rdev_for_each(rdev, mddev) {
6635 if (rdev->raid_disk < 0)
6637 diff = (rdev->new_data_offset - rdev->data_offset);
6639 min_offset_diff = diff;
6641 } else if (mddev->reshape_backwards &&
6642 diff < min_offset_diff)
6643 min_offset_diff = diff;
6644 else if (!mddev->reshape_backwards &&
6645 diff > min_offset_diff)
6646 min_offset_diff = diff;
6649 if (mddev->reshape_position != MaxSector) {
6650 /* Check that we can continue the reshape.
6651 * Difficulties arise if the stripe we would write to
6652 * next is at or after the stripe we would read from next.
6653 * For a reshape that changes the number of devices, this
6654 * is only possible for a very short time, and mdadm makes
6655 * sure that time appears to have past before assembling
6656 * the array. So we fail if that time hasn't passed.
6657 * For a reshape that keeps the number of devices the same
6658 * mdadm must be monitoring the reshape can keeping the
6659 * critical areas read-only and backed up. It will start
6660 * the array in read-only mode, so we check for that.
6662 sector_t here_new, here_old;
6664 int max_degraded = (mddev->level == 6 ? 2 : 1);
6666 if (mddev->new_level != mddev->level) {
6667 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6668 "required - aborting.\n",
6672 old_disks = mddev->raid_disks - mddev->delta_disks;
6673 /* reshape_position must be on a new-stripe boundary, and one
6674 * further up in new geometry must map after here in old
6677 here_new = mddev->reshape_position;
6678 if (sector_div(here_new, mddev->new_chunk_sectors *
6679 (mddev->raid_disks - max_degraded))) {
6680 printk(KERN_ERR "md/raid:%s: reshape_position not "
6681 "on a stripe boundary\n", mdname(mddev));
6684 reshape_offset = here_new * mddev->new_chunk_sectors;
6685 /* here_new is the stripe we will write to */
6686 here_old = mddev->reshape_position;
6687 sector_div(here_old, mddev->chunk_sectors *
6688 (old_disks-max_degraded));
6689 /* here_old is the first stripe that we might need to read
6691 if (mddev->delta_disks == 0) {
6692 if ((here_new * mddev->new_chunk_sectors !=
6693 here_old * mddev->chunk_sectors)) {
6694 printk(KERN_ERR "md/raid:%s: reshape position is"
6695 " confused - aborting\n", mdname(mddev));
6698 /* We cannot be sure it is safe to start an in-place
6699 * reshape. It is only safe if user-space is monitoring
6700 * and taking constant backups.
6701 * mdadm always starts a situation like this in
6702 * readonly mode so it can take control before
6703 * allowing any writes. So just check for that.
6705 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6706 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6707 /* not really in-place - so OK */;
6708 else if (mddev->ro == 0) {
6709 printk(KERN_ERR "md/raid:%s: in-place reshape "
6710 "must be started in read-only mode "
6715 } else if (mddev->reshape_backwards
6716 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6717 here_old * mddev->chunk_sectors)
6718 : (here_new * mddev->new_chunk_sectors >=
6719 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6720 /* Reading from the same stripe as writing to - bad */
6721 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6722 "auto-recovery - aborting.\n",
6726 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6728 /* OK, we should be able to continue; */
6730 BUG_ON(mddev->level != mddev->new_level);
6731 BUG_ON(mddev->layout != mddev->new_layout);
6732 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6733 BUG_ON(mddev->delta_disks != 0);
6736 if (mddev->private == NULL)
6737 conf = setup_conf(mddev);
6739 conf = mddev->private;
6742 return PTR_ERR(conf);
6744 conf->min_offset_diff = min_offset_diff;
6745 mddev->thread = conf->thread;
6746 conf->thread = NULL;
6747 mddev->private = conf;
6749 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6751 rdev = conf->disks[i].rdev;
6752 if (!rdev && conf->disks[i].replacement) {
6753 /* The replacement is all we have yet */
6754 rdev = conf->disks[i].replacement;
6755 conf->disks[i].replacement = NULL;
6756 clear_bit(Replacement, &rdev->flags);
6757 conf->disks[i].rdev = rdev;
6761 if (conf->disks[i].replacement &&
6762 conf->reshape_progress != MaxSector) {
6763 /* replacements and reshape simply do not mix. */
6764 printk(KERN_ERR "md: cannot handle concurrent "
6765 "replacement and reshape.\n");
6768 if (test_bit(In_sync, &rdev->flags)) {
6772 /* This disc is not fully in-sync. However if it
6773 * just stored parity (beyond the recovery_offset),
6774 * when we don't need to be concerned about the
6775 * array being dirty.
6776 * When reshape goes 'backwards', we never have
6777 * partially completed devices, so we only need
6778 * to worry about reshape going forwards.
6780 /* Hack because v0.91 doesn't store recovery_offset properly. */
6781 if (mddev->major_version == 0 &&
6782 mddev->minor_version > 90)
6783 rdev->recovery_offset = reshape_offset;
6785 if (rdev->recovery_offset < reshape_offset) {
6786 /* We need to check old and new layout */
6787 if (!only_parity(rdev->raid_disk,
6790 conf->max_degraded))
6793 if (!only_parity(rdev->raid_disk,
6795 conf->previous_raid_disks,
6796 conf->max_degraded))
6798 dirty_parity_disks++;
6802 * 0 for a fully functional array, 1 or 2 for a degraded array.
6804 mddev->degraded = calc_degraded(conf);
6806 if (has_failed(conf)) {
6807 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6808 " (%d/%d failed)\n",
6809 mdname(mddev), mddev->degraded, conf->raid_disks);
6813 /* device size must be a multiple of chunk size */
6814 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6815 mddev->resync_max_sectors = mddev->dev_sectors;
6817 if (mddev->degraded > dirty_parity_disks &&
6818 mddev->recovery_cp != MaxSector) {
6819 if (mddev->ok_start_degraded)
6821 "md/raid:%s: starting dirty degraded array"
6822 " - data corruption possible.\n",
6826 "md/raid:%s: cannot start dirty degraded array.\n",
6832 if (mddev->degraded == 0)
6833 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6834 " devices, algorithm %d\n", mdname(mddev), conf->level,
6835 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6838 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6839 " out of %d devices, algorithm %d\n",
6840 mdname(mddev), conf->level,
6841 mddev->raid_disks - mddev->degraded,
6842 mddev->raid_disks, mddev->new_layout);
6844 print_raid5_conf(conf);
6846 if (conf->reshape_progress != MaxSector) {
6847 conf->reshape_safe = conf->reshape_progress;
6848 atomic_set(&conf->reshape_stripes, 0);
6849 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6850 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6851 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6852 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6853 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6857 /* Ok, everything is just fine now */
6858 if (mddev->to_remove == &raid5_attrs_group)
6859 mddev->to_remove = NULL;
6860 else if (mddev->kobj.sd &&
6861 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6863 "raid5: failed to create sysfs attributes for %s\n",
6865 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6869 bool discard_supported = true;
6870 /* read-ahead size must cover two whole stripes, which
6871 * is 2 * (datadisks) * chunksize where 'n' is the
6872 * number of raid devices
6874 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6875 int stripe = data_disks *
6876 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6877 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6878 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6880 chunk_size = mddev->chunk_sectors << 9;
6881 blk_queue_io_min(mddev->queue, chunk_size);
6882 blk_queue_io_opt(mddev->queue, chunk_size *
6883 (conf->raid_disks - conf->max_degraded));
6884 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6886 * We can only discard a whole stripe. It doesn't make sense to
6887 * discard data disk but write parity disk
6889 stripe = stripe * PAGE_SIZE;
6890 /* Round up to power of 2, as discard handling
6891 * currently assumes that */
6892 while ((stripe-1) & stripe)
6893 stripe = (stripe | (stripe-1)) + 1;
6894 mddev->queue->limits.discard_alignment = stripe;
6895 mddev->queue->limits.discard_granularity = stripe;
6897 * unaligned part of discard request will be ignored, so can't
6898 * guarantee discard_zeroes_data
6900 mddev->queue->limits.discard_zeroes_data = 0;
6902 blk_queue_max_write_same_sectors(mddev->queue, 0);
6904 rdev_for_each(rdev, mddev) {
6905 disk_stack_limits(mddev->gendisk, rdev->bdev,
6906 rdev->data_offset << 9);
6907 disk_stack_limits(mddev->gendisk, rdev->bdev,
6908 rdev->new_data_offset << 9);
6910 * discard_zeroes_data is required, otherwise data
6911 * could be lost. Consider a scenario: discard a stripe
6912 * (the stripe could be inconsistent if
6913 * discard_zeroes_data is 0); write one disk of the
6914 * stripe (the stripe could be inconsistent again
6915 * depending on which disks are used to calculate
6916 * parity); the disk is broken; The stripe data of this
6919 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6920 !bdev_get_queue(rdev->bdev)->
6921 limits.discard_zeroes_data)
6922 discard_supported = false;
6923 /* Unfortunately, discard_zeroes_data is not currently
6924 * a guarantee - just a hint. So we only allow DISCARD
6925 * if the sysadmin has confirmed that only safe devices
6926 * are in use by setting a module parameter.
6928 if (!devices_handle_discard_safely) {
6929 if (discard_supported) {
6930 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6931 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6933 discard_supported = false;
6937 if (discard_supported &&
6938 mddev->queue->limits.max_discard_sectors >= stripe &&
6939 mddev->queue->limits.discard_granularity >= stripe)
6940 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6943 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6949 md_unregister_thread(&mddev->thread);
6950 print_raid5_conf(conf);
6952 mddev->private = NULL;
6953 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6957 static void raid5_free(struct mddev *mddev, void *priv)
6959 struct r5conf *conf = priv;
6962 mddev->to_remove = &raid5_attrs_group;
6965 static void status(struct seq_file *seq, struct mddev *mddev)
6967 struct r5conf *conf = mddev->private;
6970 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6971 mddev->chunk_sectors / 2, mddev->layout);
6972 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6973 for (i = 0; i < conf->raid_disks; i++)
6974 seq_printf (seq, "%s",
6975 conf->disks[i].rdev &&
6976 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6977 seq_printf (seq, "]");
6980 static void print_raid5_conf (struct r5conf *conf)
6983 struct disk_info *tmp;
6985 printk(KERN_DEBUG "RAID conf printout:\n");
6987 printk("(conf==NULL)\n");
6990 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6992 conf->raid_disks - conf->mddev->degraded);
6994 for (i = 0; i < conf->raid_disks; i++) {
6995 char b[BDEVNAME_SIZE];
6996 tmp = conf->disks + i;
6998 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6999 i, !test_bit(Faulty, &tmp->rdev->flags),
7000 bdevname(tmp->rdev->bdev, b));
7004 static int raid5_spare_active(struct mddev *mddev)
7007 struct r5conf *conf = mddev->private;
7008 struct disk_info *tmp;
7010 unsigned long flags;
7012 for (i = 0; i < conf->raid_disks; i++) {
7013 tmp = conf->disks + i;
7014 if (tmp->replacement
7015 && tmp->replacement->recovery_offset == MaxSector
7016 && !test_bit(Faulty, &tmp->replacement->flags)
7017 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7018 /* Replacement has just become active. */
7020 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7023 /* Replaced device not technically faulty,
7024 * but we need to be sure it gets removed
7025 * and never re-added.
7027 set_bit(Faulty, &tmp->rdev->flags);
7028 sysfs_notify_dirent_safe(
7029 tmp->rdev->sysfs_state);
7031 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7032 } else if (tmp->rdev
7033 && tmp->rdev->recovery_offset == MaxSector
7034 && !test_bit(Faulty, &tmp->rdev->flags)
7035 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7037 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7040 spin_lock_irqsave(&conf->device_lock, flags);
7041 mddev->degraded = calc_degraded(conf);
7042 spin_unlock_irqrestore(&conf->device_lock, flags);
7043 print_raid5_conf(conf);
7047 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7049 struct r5conf *conf = mddev->private;
7051 int number = rdev->raid_disk;
7052 struct md_rdev **rdevp;
7053 struct disk_info *p = conf->disks + number;
7055 print_raid5_conf(conf);
7056 if (rdev == p->rdev)
7058 else if (rdev == p->replacement)
7059 rdevp = &p->replacement;
7063 if (number >= conf->raid_disks &&
7064 conf->reshape_progress == MaxSector)
7065 clear_bit(In_sync, &rdev->flags);
7067 if (test_bit(In_sync, &rdev->flags) ||
7068 atomic_read(&rdev->nr_pending)) {
7072 /* Only remove non-faulty devices if recovery
7075 if (!test_bit(Faulty, &rdev->flags) &&
7076 mddev->recovery_disabled != conf->recovery_disabled &&
7077 !has_failed(conf) &&
7078 (!p->replacement || p->replacement == rdev) &&
7079 number < conf->raid_disks) {
7085 if (atomic_read(&rdev->nr_pending)) {
7086 /* lost the race, try later */
7089 } else if (p->replacement) {
7090 /* We must have just cleared 'rdev' */
7091 p->rdev = p->replacement;
7092 clear_bit(Replacement, &p->replacement->flags);
7093 smp_mb(); /* Make sure other CPUs may see both as identical
7094 * but will never see neither - if they are careful
7096 p->replacement = NULL;
7097 clear_bit(WantReplacement, &rdev->flags);
7099 /* We might have just removed the Replacement as faulty-
7100 * clear the bit just in case
7102 clear_bit(WantReplacement, &rdev->flags);
7105 print_raid5_conf(conf);
7109 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7111 struct r5conf *conf = mddev->private;
7114 struct disk_info *p;
7116 int last = conf->raid_disks - 1;
7118 if (mddev->recovery_disabled == conf->recovery_disabled)
7121 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7122 /* no point adding a device */
7125 if (rdev->raid_disk >= 0)
7126 first = last = rdev->raid_disk;
7129 * find the disk ... but prefer rdev->saved_raid_disk
7132 if (rdev->saved_raid_disk >= 0 &&
7133 rdev->saved_raid_disk >= first &&
7134 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7135 first = rdev->saved_raid_disk;
7137 for (disk = first; disk <= last; disk++) {
7138 p = conf->disks + disk;
7139 if (p->rdev == NULL) {
7140 clear_bit(In_sync, &rdev->flags);
7141 rdev->raid_disk = disk;
7143 if (rdev->saved_raid_disk != disk)
7145 rcu_assign_pointer(p->rdev, rdev);
7149 for (disk = first; disk <= last; disk++) {
7150 p = conf->disks + disk;
7151 if (test_bit(WantReplacement, &p->rdev->flags) &&
7152 p->replacement == NULL) {
7153 clear_bit(In_sync, &rdev->flags);
7154 set_bit(Replacement, &rdev->flags);
7155 rdev->raid_disk = disk;
7158 rcu_assign_pointer(p->replacement, rdev);
7163 print_raid5_conf(conf);
7167 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7169 /* no resync is happening, and there is enough space
7170 * on all devices, so we can resize.
7171 * We need to make sure resync covers any new space.
7172 * If the array is shrinking we should possibly wait until
7173 * any io in the removed space completes, but it hardly seems
7177 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7178 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7179 if (mddev->external_size &&
7180 mddev->array_sectors > newsize)
7182 if (mddev->bitmap) {
7183 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7187 md_set_array_sectors(mddev, newsize);
7188 set_capacity(mddev->gendisk, mddev->array_sectors);
7189 revalidate_disk(mddev->gendisk);
7190 if (sectors > mddev->dev_sectors &&
7191 mddev->recovery_cp > mddev->dev_sectors) {
7192 mddev->recovery_cp = mddev->dev_sectors;
7193 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7195 mddev->dev_sectors = sectors;
7196 mddev->resync_max_sectors = sectors;
7200 static int check_stripe_cache(struct mddev *mddev)
7202 /* Can only proceed if there are plenty of stripe_heads.
7203 * We need a minimum of one full stripe,, and for sensible progress
7204 * it is best to have about 4 times that.
7205 * If we require 4 times, then the default 256 4K stripe_heads will
7206 * allow for chunk sizes up to 256K, which is probably OK.
7207 * If the chunk size is greater, user-space should request more
7208 * stripe_heads first.
7210 struct r5conf *conf = mddev->private;
7211 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7212 > conf->min_nr_stripes ||
7213 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7214 > conf->min_nr_stripes) {
7215 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7217 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7224 static int check_reshape(struct mddev *mddev)
7226 struct r5conf *conf = mddev->private;
7228 if (mddev->delta_disks == 0 &&
7229 mddev->new_layout == mddev->layout &&
7230 mddev->new_chunk_sectors == mddev->chunk_sectors)
7231 return 0; /* nothing to do */
7232 if (has_failed(conf))
7234 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7235 /* We might be able to shrink, but the devices must
7236 * be made bigger first.
7237 * For raid6, 4 is the minimum size.
7238 * Otherwise 2 is the minimum
7241 if (mddev->level == 6)
7243 if (mddev->raid_disks + mddev->delta_disks < min)
7247 if (!check_stripe_cache(mddev))
7250 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7251 mddev->delta_disks > 0)
7252 if (resize_chunks(conf,
7253 conf->previous_raid_disks
7254 + max(0, mddev->delta_disks),
7255 max(mddev->new_chunk_sectors,
7256 mddev->chunk_sectors)
7259 return resize_stripes(conf, (conf->previous_raid_disks
7260 + mddev->delta_disks));
7263 static int raid5_start_reshape(struct mddev *mddev)
7265 struct r5conf *conf = mddev->private;
7266 struct md_rdev *rdev;
7268 unsigned long flags;
7270 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7273 if (!check_stripe_cache(mddev))
7276 if (has_failed(conf))
7279 rdev_for_each(rdev, mddev) {
7280 if (!test_bit(In_sync, &rdev->flags)
7281 && !test_bit(Faulty, &rdev->flags))
7285 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7286 /* Not enough devices even to make a degraded array
7291 /* Refuse to reduce size of the array. Any reductions in
7292 * array size must be through explicit setting of array_size
7295 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7296 < mddev->array_sectors) {
7297 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7298 "before number of disks\n", mdname(mddev));
7302 atomic_set(&conf->reshape_stripes, 0);
7303 spin_lock_irq(&conf->device_lock);
7304 write_seqcount_begin(&conf->gen_lock);
7305 conf->previous_raid_disks = conf->raid_disks;
7306 conf->raid_disks += mddev->delta_disks;
7307 conf->prev_chunk_sectors = conf->chunk_sectors;
7308 conf->chunk_sectors = mddev->new_chunk_sectors;
7309 conf->prev_algo = conf->algorithm;
7310 conf->algorithm = mddev->new_layout;
7312 /* Code that selects data_offset needs to see the generation update
7313 * if reshape_progress has been set - so a memory barrier needed.
7316 if (mddev->reshape_backwards)
7317 conf->reshape_progress = raid5_size(mddev, 0, 0);
7319 conf->reshape_progress = 0;
7320 conf->reshape_safe = conf->reshape_progress;
7321 write_seqcount_end(&conf->gen_lock);
7322 spin_unlock_irq(&conf->device_lock);
7324 /* Now make sure any requests that proceeded on the assumption
7325 * the reshape wasn't running - like Discard or Read - have
7328 mddev_suspend(mddev);
7329 mddev_resume(mddev);
7331 /* Add some new drives, as many as will fit.
7332 * We know there are enough to make the newly sized array work.
7333 * Don't add devices if we are reducing the number of
7334 * devices in the array. This is because it is not possible
7335 * to correctly record the "partially reconstructed" state of
7336 * such devices during the reshape and confusion could result.
7338 if (mddev->delta_disks >= 0) {
7339 rdev_for_each(rdev, mddev)
7340 if (rdev->raid_disk < 0 &&
7341 !test_bit(Faulty, &rdev->flags)) {
7342 if (raid5_add_disk(mddev, rdev) == 0) {
7344 >= conf->previous_raid_disks)
7345 set_bit(In_sync, &rdev->flags);
7347 rdev->recovery_offset = 0;
7349 if (sysfs_link_rdev(mddev, rdev))
7350 /* Failure here is OK */;
7352 } else if (rdev->raid_disk >= conf->previous_raid_disks
7353 && !test_bit(Faulty, &rdev->flags)) {
7354 /* This is a spare that was manually added */
7355 set_bit(In_sync, &rdev->flags);
7358 /* When a reshape changes the number of devices,
7359 * ->degraded is measured against the larger of the
7360 * pre and post number of devices.
7362 spin_lock_irqsave(&conf->device_lock, flags);
7363 mddev->degraded = calc_degraded(conf);
7364 spin_unlock_irqrestore(&conf->device_lock, flags);
7366 mddev->raid_disks = conf->raid_disks;
7367 mddev->reshape_position = conf->reshape_progress;
7368 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7370 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7371 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7372 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7373 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7374 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7376 if (!mddev->sync_thread) {
7377 mddev->recovery = 0;
7378 spin_lock_irq(&conf->device_lock);
7379 write_seqcount_begin(&conf->gen_lock);
7380 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7381 mddev->new_chunk_sectors =
7382 conf->chunk_sectors = conf->prev_chunk_sectors;
7383 mddev->new_layout = conf->algorithm = conf->prev_algo;
7384 rdev_for_each(rdev, mddev)
7385 rdev->new_data_offset = rdev->data_offset;
7387 conf->generation --;
7388 conf->reshape_progress = MaxSector;
7389 mddev->reshape_position = MaxSector;
7390 write_seqcount_end(&conf->gen_lock);
7391 spin_unlock_irq(&conf->device_lock);
7394 conf->reshape_checkpoint = jiffies;
7395 md_wakeup_thread(mddev->sync_thread);
7396 md_new_event(mddev);
7400 /* This is called from the reshape thread and should make any
7401 * changes needed in 'conf'
7403 static void end_reshape(struct r5conf *conf)
7406 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7407 struct md_rdev *rdev;
7409 spin_lock_irq(&conf->device_lock);
7410 conf->previous_raid_disks = conf->raid_disks;
7411 rdev_for_each(rdev, conf->mddev)
7412 rdev->data_offset = rdev->new_data_offset;
7414 conf->reshape_progress = MaxSector;
7415 spin_unlock_irq(&conf->device_lock);
7416 wake_up(&conf->wait_for_overlap);
7418 /* read-ahead size must cover two whole stripes, which is
7419 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7421 if (conf->mddev->queue) {
7422 int data_disks = conf->raid_disks - conf->max_degraded;
7423 int stripe = data_disks * ((conf->chunk_sectors << 9)
7425 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7426 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7431 /* This is called from the raid5d thread with mddev_lock held.
7432 * It makes config changes to the device.
7434 static void raid5_finish_reshape(struct mddev *mddev)
7436 struct r5conf *conf = mddev->private;
7438 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7440 if (mddev->delta_disks > 0) {
7441 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7442 set_capacity(mddev->gendisk, mddev->array_sectors);
7443 revalidate_disk(mddev->gendisk);
7446 spin_lock_irq(&conf->device_lock);
7447 mddev->degraded = calc_degraded(conf);
7448 spin_unlock_irq(&conf->device_lock);
7449 for (d = conf->raid_disks ;
7450 d < conf->raid_disks - mddev->delta_disks;
7452 struct md_rdev *rdev = conf->disks[d].rdev;
7454 clear_bit(In_sync, &rdev->flags);
7455 rdev = conf->disks[d].replacement;
7457 clear_bit(In_sync, &rdev->flags);
7460 mddev->layout = conf->algorithm;
7461 mddev->chunk_sectors = conf->chunk_sectors;
7462 mddev->reshape_position = MaxSector;
7463 mddev->delta_disks = 0;
7464 mddev->reshape_backwards = 0;
7468 static void raid5_quiesce(struct mddev *mddev, int state)
7470 struct r5conf *conf = mddev->private;
7473 case 2: /* resume for a suspend */
7474 wake_up(&conf->wait_for_overlap);
7477 case 1: /* stop all writes */
7478 lock_all_device_hash_locks_irq(conf);
7479 /* '2' tells resync/reshape to pause so that all
7480 * active stripes can drain
7483 wait_event_cmd(conf->wait_for_stripe,
7484 atomic_read(&conf->active_stripes) == 0 &&
7485 atomic_read(&conf->active_aligned_reads) == 0,
7486 unlock_all_device_hash_locks_irq(conf),
7487 lock_all_device_hash_locks_irq(conf));
7489 unlock_all_device_hash_locks_irq(conf);
7490 /* allow reshape to continue */
7491 wake_up(&conf->wait_for_overlap);
7494 case 0: /* re-enable writes */
7495 lock_all_device_hash_locks_irq(conf);
7497 wake_up(&conf->wait_for_stripe);
7498 wake_up(&conf->wait_for_overlap);
7499 unlock_all_device_hash_locks_irq(conf);
7504 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7506 struct r0conf *raid0_conf = mddev->private;
7509 /* for raid0 takeover only one zone is supported */
7510 if (raid0_conf->nr_strip_zones > 1) {
7511 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7513 return ERR_PTR(-EINVAL);
7516 sectors = raid0_conf->strip_zone[0].zone_end;
7517 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7518 mddev->dev_sectors = sectors;
7519 mddev->new_level = level;
7520 mddev->new_layout = ALGORITHM_PARITY_N;
7521 mddev->new_chunk_sectors = mddev->chunk_sectors;
7522 mddev->raid_disks += 1;
7523 mddev->delta_disks = 1;
7524 /* make sure it will be not marked as dirty */
7525 mddev->recovery_cp = MaxSector;
7527 return setup_conf(mddev);
7530 static void *raid5_takeover_raid1(struct mddev *mddev)
7534 if (mddev->raid_disks != 2 ||
7535 mddev->degraded > 1)
7536 return ERR_PTR(-EINVAL);
7538 /* Should check if there are write-behind devices? */
7540 chunksect = 64*2; /* 64K by default */
7542 /* The array must be an exact multiple of chunksize */
7543 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7546 if ((chunksect<<9) < STRIPE_SIZE)
7547 /* array size does not allow a suitable chunk size */
7548 return ERR_PTR(-EINVAL);
7550 mddev->new_level = 5;
7551 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7552 mddev->new_chunk_sectors = chunksect;
7554 return setup_conf(mddev);
7557 static void *raid5_takeover_raid6(struct mddev *mddev)
7561 switch (mddev->layout) {
7562 case ALGORITHM_LEFT_ASYMMETRIC_6:
7563 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7565 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7566 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7568 case ALGORITHM_LEFT_SYMMETRIC_6:
7569 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7571 case ALGORITHM_RIGHT_SYMMETRIC_6:
7572 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7574 case ALGORITHM_PARITY_0_6:
7575 new_layout = ALGORITHM_PARITY_0;
7577 case ALGORITHM_PARITY_N:
7578 new_layout = ALGORITHM_PARITY_N;
7581 return ERR_PTR(-EINVAL);
7583 mddev->new_level = 5;
7584 mddev->new_layout = new_layout;
7585 mddev->delta_disks = -1;
7586 mddev->raid_disks -= 1;
7587 return setup_conf(mddev);
7590 static int raid5_check_reshape(struct mddev *mddev)
7592 /* For a 2-drive array, the layout and chunk size can be changed
7593 * immediately as not restriping is needed.
7594 * For larger arrays we record the new value - after validation
7595 * to be used by a reshape pass.
7597 struct r5conf *conf = mddev->private;
7598 int new_chunk = mddev->new_chunk_sectors;
7600 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7602 if (new_chunk > 0) {
7603 if (!is_power_of_2(new_chunk))
7605 if (new_chunk < (PAGE_SIZE>>9))
7607 if (mddev->array_sectors & (new_chunk-1))
7608 /* not factor of array size */
7612 /* They look valid */
7614 if (mddev->raid_disks == 2) {
7615 /* can make the change immediately */
7616 if (mddev->new_layout >= 0) {
7617 conf->algorithm = mddev->new_layout;
7618 mddev->layout = mddev->new_layout;
7620 if (new_chunk > 0) {
7621 conf->chunk_sectors = new_chunk ;
7622 mddev->chunk_sectors = new_chunk;
7624 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7625 md_wakeup_thread(mddev->thread);
7627 return check_reshape(mddev);
7630 static int raid6_check_reshape(struct mddev *mddev)
7632 int new_chunk = mddev->new_chunk_sectors;
7634 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7636 if (new_chunk > 0) {
7637 if (!is_power_of_2(new_chunk))
7639 if (new_chunk < (PAGE_SIZE >> 9))
7641 if (mddev->array_sectors & (new_chunk-1))
7642 /* not factor of array size */
7646 /* They look valid */
7647 return check_reshape(mddev);
7650 static void *raid5_takeover(struct mddev *mddev)
7652 /* raid5 can take over:
7653 * raid0 - if there is only one strip zone - make it a raid4 layout
7654 * raid1 - if there are two drives. We need to know the chunk size
7655 * raid4 - trivial - just use a raid4 layout.
7656 * raid6 - Providing it is a *_6 layout
7658 if (mddev->level == 0)
7659 return raid45_takeover_raid0(mddev, 5);
7660 if (mddev->level == 1)
7661 return raid5_takeover_raid1(mddev);
7662 if (mddev->level == 4) {
7663 mddev->new_layout = ALGORITHM_PARITY_N;
7664 mddev->new_level = 5;
7665 return setup_conf(mddev);
7667 if (mddev->level == 6)
7668 return raid5_takeover_raid6(mddev);
7670 return ERR_PTR(-EINVAL);
7673 static void *raid4_takeover(struct mddev *mddev)
7675 /* raid4 can take over:
7676 * raid0 - if there is only one strip zone
7677 * raid5 - if layout is right
7679 if (mddev->level == 0)
7680 return raid45_takeover_raid0(mddev, 4);
7681 if (mddev->level == 5 &&
7682 mddev->layout == ALGORITHM_PARITY_N) {
7683 mddev->new_layout = 0;
7684 mddev->new_level = 4;
7685 return setup_conf(mddev);
7687 return ERR_PTR(-EINVAL);
7690 static struct md_personality raid5_personality;
7692 static void *raid6_takeover(struct mddev *mddev)
7694 /* Currently can only take over a raid5. We map the
7695 * personality to an equivalent raid6 personality
7696 * with the Q block at the end.
7700 if (mddev->pers != &raid5_personality)
7701 return ERR_PTR(-EINVAL);
7702 if (mddev->degraded > 1)
7703 return ERR_PTR(-EINVAL);
7704 if (mddev->raid_disks > 253)
7705 return ERR_PTR(-EINVAL);
7706 if (mddev->raid_disks < 3)
7707 return ERR_PTR(-EINVAL);
7709 switch (mddev->layout) {
7710 case ALGORITHM_LEFT_ASYMMETRIC:
7711 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7713 case ALGORITHM_RIGHT_ASYMMETRIC:
7714 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7716 case ALGORITHM_LEFT_SYMMETRIC:
7717 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7719 case ALGORITHM_RIGHT_SYMMETRIC:
7720 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7722 case ALGORITHM_PARITY_0:
7723 new_layout = ALGORITHM_PARITY_0_6;
7725 case ALGORITHM_PARITY_N:
7726 new_layout = ALGORITHM_PARITY_N;
7729 return ERR_PTR(-EINVAL);
7731 mddev->new_level = 6;
7732 mddev->new_layout = new_layout;
7733 mddev->delta_disks = 1;
7734 mddev->raid_disks += 1;
7735 return setup_conf(mddev);
7738 static struct md_personality raid6_personality =
7742 .owner = THIS_MODULE,
7743 .make_request = make_request,
7747 .error_handler = error,
7748 .hot_add_disk = raid5_add_disk,
7749 .hot_remove_disk= raid5_remove_disk,
7750 .spare_active = raid5_spare_active,
7751 .sync_request = sync_request,
7752 .resize = raid5_resize,
7754 .check_reshape = raid6_check_reshape,
7755 .start_reshape = raid5_start_reshape,
7756 .finish_reshape = raid5_finish_reshape,
7757 .quiesce = raid5_quiesce,
7758 .takeover = raid6_takeover,
7759 .congested = raid5_congested,
7760 .mergeable_bvec = raid5_mergeable_bvec,
7762 static struct md_personality raid5_personality =
7766 .owner = THIS_MODULE,
7767 .make_request = make_request,
7771 .error_handler = error,
7772 .hot_add_disk = raid5_add_disk,
7773 .hot_remove_disk= raid5_remove_disk,
7774 .spare_active = raid5_spare_active,
7775 .sync_request = sync_request,
7776 .resize = raid5_resize,
7778 .check_reshape = raid5_check_reshape,
7779 .start_reshape = raid5_start_reshape,
7780 .finish_reshape = raid5_finish_reshape,
7781 .quiesce = raid5_quiesce,
7782 .takeover = raid5_takeover,
7783 .congested = raid5_congested,
7784 .mergeable_bvec = raid5_mergeable_bvec,
7787 static struct md_personality raid4_personality =
7791 .owner = THIS_MODULE,
7792 .make_request = make_request,
7796 .error_handler = error,
7797 .hot_add_disk = raid5_add_disk,
7798 .hot_remove_disk= raid5_remove_disk,
7799 .spare_active = raid5_spare_active,
7800 .sync_request = sync_request,
7801 .resize = raid5_resize,
7803 .check_reshape = raid5_check_reshape,
7804 .start_reshape = raid5_start_reshape,
7805 .finish_reshape = raid5_finish_reshape,
7806 .quiesce = raid5_quiesce,
7807 .takeover = raid4_takeover,
7808 .congested = raid5_congested,
7809 .mergeable_bvec = raid5_mergeable_bvec,
7812 static int __init raid5_init(void)
7814 raid5_wq = alloc_workqueue("raid5wq",
7815 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7818 register_md_personality(&raid6_personality);
7819 register_md_personality(&raid5_personality);
7820 register_md_personality(&raid4_personality);
7824 static void raid5_exit(void)
7826 unregister_md_personality(&raid6_personality);
7827 unregister_md_personality(&raid5_personality);
7828 unregister_md_personality(&raid4_personality);
7829 destroy_workqueue(raid5_wq);
7832 module_init(raid5_init);
7833 module_exit(raid5_exit);
7834 MODULE_LICENSE("GPL");
7835 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7836 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7837 MODULE_ALIAS("md-raid5");
7838 MODULE_ALIAS("md-raid4");
7839 MODULE_ALIAS("md-level-5");
7840 MODULE_ALIAS("md-level-4");
7841 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7842 MODULE_ALIAS("md-raid6");
7843 MODULE_ALIAS("md-level-6");
7845 /* This used to be two separate modules, they were: */
7846 MODULE_ALIAS("raid5");
7847 MODULE_ALIAS("raid6");