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 (!conf->inactive_blocked)
676 sh = get_free_stripe(conf, hash);
677 if (noblock && sh == NULL)
680 conf->inactive_blocked = 1;
682 conf->wait_for_stripe,
683 !list_empty(conf->inactive_list + hash) &&
684 (atomic_read(&conf->active_stripes)
685 < (conf->max_nr_stripes * 3 / 4)
686 || !conf->inactive_blocked),
687 *(conf->hash_locks + hash));
688 conf->inactive_blocked = 0;
690 init_stripe(sh, sector, previous);
691 atomic_inc(&sh->count);
693 } else if (!atomic_inc_not_zero(&sh->count)) {
694 spin_lock(&conf->device_lock);
695 if (!atomic_read(&sh->count)) {
696 if (!test_bit(STRIPE_HANDLE, &sh->state))
697 atomic_inc(&conf->active_stripes);
698 BUG_ON(list_empty(&sh->lru) &&
699 !test_bit(STRIPE_EXPANDING, &sh->state));
700 list_del_init(&sh->lru);
702 sh->group->stripes_cnt--;
706 atomic_inc(&sh->count);
707 spin_unlock(&conf->device_lock);
709 } while (sh == NULL);
711 spin_unlock_irq(conf->hash_locks + hash);
715 static bool is_full_stripe_write(struct stripe_head *sh)
717 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
718 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
721 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
725 spin_lock(&sh2->stripe_lock);
726 spin_lock_nested(&sh1->stripe_lock, 1);
728 spin_lock(&sh1->stripe_lock);
729 spin_lock_nested(&sh2->stripe_lock, 1);
733 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
735 spin_unlock(&sh1->stripe_lock);
736 spin_unlock(&sh2->stripe_lock);
740 /* Only freshly new full stripe normal write stripe can be added to a batch list */
741 static bool stripe_can_batch(struct stripe_head *sh)
743 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
744 is_full_stripe_write(sh);
747 /* we only do back search */
748 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
750 struct stripe_head *head;
751 sector_t head_sector, tmp_sec;
755 if (!stripe_can_batch(sh))
757 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
758 tmp_sec = sh->sector;
759 if (!sector_div(tmp_sec, conf->chunk_sectors))
761 head_sector = sh->sector - STRIPE_SECTORS;
763 hash = stripe_hash_locks_hash(head_sector);
764 spin_lock_irq(conf->hash_locks + hash);
765 head = __find_stripe(conf, head_sector, conf->generation);
766 if (head && !atomic_inc_not_zero(&head->count)) {
767 spin_lock(&conf->device_lock);
768 if (!atomic_read(&head->count)) {
769 if (!test_bit(STRIPE_HANDLE, &head->state))
770 atomic_inc(&conf->active_stripes);
771 BUG_ON(list_empty(&head->lru) &&
772 !test_bit(STRIPE_EXPANDING, &head->state));
773 list_del_init(&head->lru);
775 head->group->stripes_cnt--;
779 atomic_inc(&head->count);
780 spin_unlock(&conf->device_lock);
782 spin_unlock_irq(conf->hash_locks + hash);
786 if (!stripe_can_batch(head))
789 lock_two_stripes(head, sh);
790 /* clear_batch_ready clear the flag */
791 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
798 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
800 if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
803 if (head->batch_head) {
804 spin_lock(&head->batch_head->batch_lock);
805 /* This batch list is already running */
806 if (!stripe_can_batch(head)) {
807 spin_unlock(&head->batch_head->batch_lock);
812 * at this point, head's BATCH_READY could be cleared, but we
813 * can still add the stripe to batch list
815 list_add(&sh->batch_list, &head->batch_list);
816 spin_unlock(&head->batch_head->batch_lock);
818 sh->batch_head = head->batch_head;
820 head->batch_head = head;
821 sh->batch_head = head->batch_head;
822 spin_lock(&head->batch_lock);
823 list_add_tail(&sh->batch_list, &head->batch_list);
824 spin_unlock(&head->batch_lock);
827 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
828 if (atomic_dec_return(&conf->preread_active_stripes)
830 md_wakeup_thread(conf->mddev->thread);
832 atomic_inc(&sh->count);
834 unlock_two_stripes(head, sh);
836 release_stripe(head);
839 /* Determine if 'data_offset' or 'new_data_offset' should be used
840 * in this stripe_head.
842 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
844 sector_t progress = conf->reshape_progress;
845 /* Need a memory barrier to make sure we see the value
846 * of conf->generation, or ->data_offset that was set before
847 * reshape_progress was updated.
850 if (progress == MaxSector)
852 if (sh->generation == conf->generation - 1)
854 /* We are in a reshape, and this is a new-generation stripe,
855 * so use new_data_offset.
861 raid5_end_read_request(struct bio *bi, int error);
863 raid5_end_write_request(struct bio *bi, int error);
865 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
867 struct r5conf *conf = sh->raid_conf;
868 int i, disks = sh->disks;
869 struct stripe_head *head_sh = sh;
873 for (i = disks; i--; ) {
875 int replace_only = 0;
876 struct bio *bi, *rbi;
877 struct md_rdev *rdev, *rrdev = NULL;
880 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
881 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
885 if (test_bit(R5_Discard, &sh->dev[i].flags))
887 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
889 else if (test_and_clear_bit(R5_WantReplace,
890 &sh->dev[i].flags)) {
895 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
899 bi = &sh->dev[i].req;
900 rbi = &sh->dev[i].rreq; /* For writing to replacement */
903 rrdev = rcu_dereference(conf->disks[i].replacement);
904 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
905 rdev = rcu_dereference(conf->disks[i].rdev);
914 /* We raced and saw duplicates */
917 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
922 if (rdev && test_bit(Faulty, &rdev->flags))
925 atomic_inc(&rdev->nr_pending);
926 if (rrdev && test_bit(Faulty, &rrdev->flags))
929 atomic_inc(&rrdev->nr_pending);
932 /* We have already checked bad blocks for reads. Now
933 * need to check for writes. We never accept write errors
934 * on the replacement, so we don't to check rrdev.
936 while ((rw & WRITE) && rdev &&
937 test_bit(WriteErrorSeen, &rdev->flags)) {
940 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
941 &first_bad, &bad_sectors);
946 set_bit(BlockedBadBlocks, &rdev->flags);
947 if (!conf->mddev->external &&
948 conf->mddev->flags) {
949 /* It is very unlikely, but we might
950 * still need to write out the
951 * bad block log - better give it
953 md_check_recovery(conf->mddev);
956 * Because md_wait_for_blocked_rdev
957 * will dec nr_pending, we must
958 * increment it first.
960 atomic_inc(&rdev->nr_pending);
961 md_wait_for_blocked_rdev(rdev, conf->mddev);
963 /* Acknowledged bad block - skip the write */
964 rdev_dec_pending(rdev, conf->mddev);
970 if (s->syncing || s->expanding || s->expanded
972 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
974 set_bit(STRIPE_IO_STARTED, &sh->state);
977 bi->bi_bdev = rdev->bdev;
979 bi->bi_end_io = (rw & WRITE)
980 ? raid5_end_write_request
981 : raid5_end_read_request;
984 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
985 __func__, (unsigned long long)sh->sector,
987 atomic_inc(&sh->count);
989 atomic_inc(&head_sh->count);
990 if (use_new_offset(conf, sh))
991 bi->bi_iter.bi_sector = (sh->sector
992 + rdev->new_data_offset);
994 bi->bi_iter.bi_sector = (sh->sector
995 + rdev->data_offset);
996 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
997 bi->bi_rw |= REQ_NOMERGE;
999 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1000 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1001 sh->dev[i].vec.bv_page = sh->dev[i].page;
1003 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1004 bi->bi_io_vec[0].bv_offset = 0;
1005 bi->bi_iter.bi_size = STRIPE_SIZE;
1007 * If this is discard request, set bi_vcnt 0. We don't
1008 * want to confuse SCSI because SCSI will replace payload
1010 if (rw & REQ_DISCARD)
1013 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1015 if (conf->mddev->gendisk)
1016 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1017 bi, disk_devt(conf->mddev->gendisk),
1019 generic_make_request(bi);
1022 if (s->syncing || s->expanding || s->expanded
1024 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1026 set_bit(STRIPE_IO_STARTED, &sh->state);
1029 rbi->bi_bdev = rrdev->bdev;
1031 BUG_ON(!(rw & WRITE));
1032 rbi->bi_end_io = raid5_end_write_request;
1033 rbi->bi_private = sh;
1035 pr_debug("%s: for %llu schedule op %ld on "
1036 "replacement disc %d\n",
1037 __func__, (unsigned long long)sh->sector,
1039 atomic_inc(&sh->count);
1041 atomic_inc(&head_sh->count);
1042 if (use_new_offset(conf, sh))
1043 rbi->bi_iter.bi_sector = (sh->sector
1044 + rrdev->new_data_offset);
1046 rbi->bi_iter.bi_sector = (sh->sector
1047 + rrdev->data_offset);
1048 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1049 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1050 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1052 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1053 rbi->bi_io_vec[0].bv_offset = 0;
1054 rbi->bi_iter.bi_size = STRIPE_SIZE;
1056 * If this is discard request, set bi_vcnt 0. We don't
1057 * want to confuse SCSI because SCSI will replace payload
1059 if (rw & REQ_DISCARD)
1061 if (conf->mddev->gendisk)
1062 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1063 rbi, disk_devt(conf->mddev->gendisk),
1065 generic_make_request(rbi);
1067 if (!rdev && !rrdev) {
1069 set_bit(STRIPE_DEGRADED, &sh->state);
1070 pr_debug("skip op %ld on disc %d for sector %llu\n",
1071 bi->bi_rw, i, (unsigned long long)sh->sector);
1072 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1074 set_bit(STRIPE_BATCH_ERR,
1075 &sh->batch_head->state);
1076 set_bit(STRIPE_HANDLE, &sh->state);
1079 if (!head_sh->batch_head)
1081 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1088 static struct dma_async_tx_descriptor *
1089 async_copy_data(int frombio, struct bio *bio, struct page **page,
1090 sector_t sector, struct dma_async_tx_descriptor *tx,
1091 struct stripe_head *sh)
1094 struct bvec_iter iter;
1095 struct page *bio_page;
1097 struct async_submit_ctl submit;
1098 enum async_tx_flags flags = 0;
1100 if (bio->bi_iter.bi_sector >= sector)
1101 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1103 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1106 flags |= ASYNC_TX_FENCE;
1107 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1109 bio_for_each_segment(bvl, bio, iter) {
1110 int len = bvl.bv_len;
1114 if (page_offset < 0) {
1115 b_offset = -page_offset;
1116 page_offset += b_offset;
1120 if (len > 0 && page_offset + len > STRIPE_SIZE)
1121 clen = STRIPE_SIZE - page_offset;
1126 b_offset += bvl.bv_offset;
1127 bio_page = bvl.bv_page;
1129 if (sh->raid_conf->skip_copy &&
1130 b_offset == 0 && page_offset == 0 &&
1131 clen == STRIPE_SIZE)
1134 tx = async_memcpy(*page, bio_page, page_offset,
1135 b_offset, clen, &submit);
1137 tx = async_memcpy(bio_page, *page, b_offset,
1138 page_offset, clen, &submit);
1140 /* chain the operations */
1141 submit.depend_tx = tx;
1143 if (clen < len) /* hit end of page */
1151 static void ops_complete_biofill(void *stripe_head_ref)
1153 struct stripe_head *sh = stripe_head_ref;
1154 struct bio *return_bi = NULL;
1157 pr_debug("%s: stripe %llu\n", __func__,
1158 (unsigned long long)sh->sector);
1160 /* clear completed biofills */
1161 for (i = sh->disks; i--; ) {
1162 struct r5dev *dev = &sh->dev[i];
1164 /* acknowledge completion of a biofill operation */
1165 /* and check if we need to reply to a read request,
1166 * new R5_Wantfill requests are held off until
1167 * !STRIPE_BIOFILL_RUN
1169 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1170 struct bio *rbi, *rbi2;
1175 while (rbi && rbi->bi_iter.bi_sector <
1176 dev->sector + STRIPE_SECTORS) {
1177 rbi2 = r5_next_bio(rbi, dev->sector);
1178 if (!raid5_dec_bi_active_stripes(rbi)) {
1179 rbi->bi_next = return_bi;
1186 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1188 return_io(return_bi);
1190 set_bit(STRIPE_HANDLE, &sh->state);
1194 static void ops_run_biofill(struct stripe_head *sh)
1196 struct dma_async_tx_descriptor *tx = NULL;
1197 struct async_submit_ctl submit;
1200 BUG_ON(sh->batch_head);
1201 pr_debug("%s: stripe %llu\n", __func__,
1202 (unsigned long long)sh->sector);
1204 for (i = sh->disks; i--; ) {
1205 struct r5dev *dev = &sh->dev[i];
1206 if (test_bit(R5_Wantfill, &dev->flags)) {
1208 spin_lock_irq(&sh->stripe_lock);
1209 dev->read = rbi = dev->toread;
1211 spin_unlock_irq(&sh->stripe_lock);
1212 while (rbi && rbi->bi_iter.bi_sector <
1213 dev->sector + STRIPE_SECTORS) {
1214 tx = async_copy_data(0, rbi, &dev->page,
1215 dev->sector, tx, sh);
1216 rbi = r5_next_bio(rbi, dev->sector);
1221 atomic_inc(&sh->count);
1222 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1223 async_trigger_callback(&submit);
1226 static void mark_target_uptodate(struct stripe_head *sh, int target)
1233 tgt = &sh->dev[target];
1234 set_bit(R5_UPTODATE, &tgt->flags);
1235 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1236 clear_bit(R5_Wantcompute, &tgt->flags);
1239 static void ops_complete_compute(void *stripe_head_ref)
1241 struct stripe_head *sh = stripe_head_ref;
1243 pr_debug("%s: stripe %llu\n", __func__,
1244 (unsigned long long)sh->sector);
1246 /* mark the computed target(s) as uptodate */
1247 mark_target_uptodate(sh, sh->ops.target);
1248 mark_target_uptodate(sh, sh->ops.target2);
1250 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1251 if (sh->check_state == check_state_compute_run)
1252 sh->check_state = check_state_compute_result;
1253 set_bit(STRIPE_HANDLE, &sh->state);
1257 /* return a pointer to the address conversion region of the scribble buffer */
1258 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1259 struct raid5_percpu *percpu, int i)
1263 addr = flex_array_get(percpu->scribble, i);
1264 return addr + sizeof(struct page *) * (sh->disks + 2);
1267 /* return a pointer to the address conversion region of the scribble buffer */
1268 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1272 addr = flex_array_get(percpu->scribble, i);
1276 static struct dma_async_tx_descriptor *
1277 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1279 int disks = sh->disks;
1280 struct page **xor_srcs = to_addr_page(percpu, 0);
1281 int target = sh->ops.target;
1282 struct r5dev *tgt = &sh->dev[target];
1283 struct page *xor_dest = tgt->page;
1285 struct dma_async_tx_descriptor *tx;
1286 struct async_submit_ctl submit;
1289 BUG_ON(sh->batch_head);
1291 pr_debug("%s: stripe %llu block: %d\n",
1292 __func__, (unsigned long long)sh->sector, target);
1293 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1295 for (i = disks; i--; )
1297 xor_srcs[count++] = sh->dev[i].page;
1299 atomic_inc(&sh->count);
1301 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1302 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1303 if (unlikely(count == 1))
1304 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1306 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1311 /* set_syndrome_sources - populate source buffers for gen_syndrome
1312 * @srcs - (struct page *) array of size sh->disks
1313 * @sh - stripe_head to parse
1315 * Populates srcs in proper layout order for the stripe and returns the
1316 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1317 * destination buffer is recorded in srcs[count] and the Q destination
1318 * is recorded in srcs[count+1]].
1320 static int set_syndrome_sources(struct page **srcs,
1321 struct stripe_head *sh,
1324 int disks = sh->disks;
1325 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1326 int d0_idx = raid6_d0(sh);
1330 for (i = 0; i < disks; i++)
1336 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1337 struct r5dev *dev = &sh->dev[i];
1339 if (i == sh->qd_idx || i == sh->pd_idx ||
1340 (srctype == SYNDROME_SRC_ALL) ||
1341 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1342 test_bit(R5_Wantdrain, &dev->flags)) ||
1343 (srctype == SYNDROME_SRC_WRITTEN &&
1345 srcs[slot] = sh->dev[i].page;
1346 i = raid6_next_disk(i, disks);
1347 } while (i != d0_idx);
1349 return syndrome_disks;
1352 static struct dma_async_tx_descriptor *
1353 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1355 int disks = sh->disks;
1356 struct page **blocks = to_addr_page(percpu, 0);
1358 int qd_idx = sh->qd_idx;
1359 struct dma_async_tx_descriptor *tx;
1360 struct async_submit_ctl submit;
1366 BUG_ON(sh->batch_head);
1367 if (sh->ops.target < 0)
1368 target = sh->ops.target2;
1369 else if (sh->ops.target2 < 0)
1370 target = sh->ops.target;
1372 /* we should only have one valid target */
1375 pr_debug("%s: stripe %llu block: %d\n",
1376 __func__, (unsigned long long)sh->sector, target);
1378 tgt = &sh->dev[target];
1379 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1382 atomic_inc(&sh->count);
1384 if (target == qd_idx) {
1385 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1386 blocks[count] = NULL; /* regenerating p is not necessary */
1387 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1388 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1389 ops_complete_compute, sh,
1390 to_addr_conv(sh, percpu, 0));
1391 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1393 /* Compute any data- or p-drive using XOR */
1395 for (i = disks; i-- ; ) {
1396 if (i == target || i == qd_idx)
1398 blocks[count++] = sh->dev[i].page;
1401 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1402 NULL, ops_complete_compute, sh,
1403 to_addr_conv(sh, percpu, 0));
1404 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1410 static struct dma_async_tx_descriptor *
1411 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1413 int i, count, disks = sh->disks;
1414 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1415 int d0_idx = raid6_d0(sh);
1416 int faila = -1, failb = -1;
1417 int target = sh->ops.target;
1418 int target2 = sh->ops.target2;
1419 struct r5dev *tgt = &sh->dev[target];
1420 struct r5dev *tgt2 = &sh->dev[target2];
1421 struct dma_async_tx_descriptor *tx;
1422 struct page **blocks = to_addr_page(percpu, 0);
1423 struct async_submit_ctl submit;
1425 BUG_ON(sh->batch_head);
1426 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1427 __func__, (unsigned long long)sh->sector, target, target2);
1428 BUG_ON(target < 0 || target2 < 0);
1429 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1430 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1432 /* we need to open-code set_syndrome_sources to handle the
1433 * slot number conversion for 'faila' and 'failb'
1435 for (i = 0; i < disks ; i++)
1440 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1442 blocks[slot] = sh->dev[i].page;
1448 i = raid6_next_disk(i, disks);
1449 } while (i != d0_idx);
1451 BUG_ON(faila == failb);
1454 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1455 __func__, (unsigned long long)sh->sector, faila, failb);
1457 atomic_inc(&sh->count);
1459 if (failb == syndrome_disks+1) {
1460 /* Q disk is one of the missing disks */
1461 if (faila == syndrome_disks) {
1462 /* Missing P+Q, just recompute */
1463 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1464 ops_complete_compute, sh,
1465 to_addr_conv(sh, percpu, 0));
1466 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1467 STRIPE_SIZE, &submit);
1471 int qd_idx = sh->qd_idx;
1473 /* Missing D+Q: recompute D from P, then recompute Q */
1474 if (target == qd_idx)
1475 data_target = target2;
1477 data_target = target;
1480 for (i = disks; i-- ; ) {
1481 if (i == data_target || i == qd_idx)
1483 blocks[count++] = sh->dev[i].page;
1485 dest = sh->dev[data_target].page;
1486 init_async_submit(&submit,
1487 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1489 to_addr_conv(sh, percpu, 0));
1490 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1493 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1494 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1495 ops_complete_compute, sh,
1496 to_addr_conv(sh, percpu, 0));
1497 return async_gen_syndrome(blocks, 0, count+2,
1498 STRIPE_SIZE, &submit);
1501 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1502 ops_complete_compute, sh,
1503 to_addr_conv(sh, percpu, 0));
1504 if (failb == syndrome_disks) {
1505 /* We're missing D+P. */
1506 return async_raid6_datap_recov(syndrome_disks+2,
1510 /* We're missing D+D. */
1511 return async_raid6_2data_recov(syndrome_disks+2,
1512 STRIPE_SIZE, faila, failb,
1518 static void ops_complete_prexor(void *stripe_head_ref)
1520 struct stripe_head *sh = stripe_head_ref;
1522 pr_debug("%s: stripe %llu\n", __func__,
1523 (unsigned long long)sh->sector);
1526 static struct dma_async_tx_descriptor *
1527 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1528 struct dma_async_tx_descriptor *tx)
1530 int disks = sh->disks;
1531 struct page **xor_srcs = to_addr_page(percpu, 0);
1532 int count = 0, pd_idx = sh->pd_idx, i;
1533 struct async_submit_ctl submit;
1535 /* existing parity data subtracted */
1536 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1538 BUG_ON(sh->batch_head);
1539 pr_debug("%s: stripe %llu\n", __func__,
1540 (unsigned long long)sh->sector);
1542 for (i = disks; i--; ) {
1543 struct r5dev *dev = &sh->dev[i];
1544 /* Only process blocks that are known to be uptodate */
1545 if (test_bit(R5_Wantdrain, &dev->flags))
1546 xor_srcs[count++] = dev->page;
1549 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1550 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1551 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1556 static struct dma_async_tx_descriptor *
1557 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1558 struct dma_async_tx_descriptor *tx)
1560 struct page **blocks = to_addr_page(percpu, 0);
1562 struct async_submit_ctl submit;
1564 pr_debug("%s: stripe %llu\n", __func__,
1565 (unsigned long long)sh->sector);
1567 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1569 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1570 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1571 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1576 static struct dma_async_tx_descriptor *
1577 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1579 int disks = sh->disks;
1581 struct stripe_head *head_sh = sh;
1583 pr_debug("%s: stripe %llu\n", __func__,
1584 (unsigned long long)sh->sector);
1586 for (i = disks; i--; ) {
1591 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1596 spin_lock_irq(&sh->stripe_lock);
1597 chosen = dev->towrite;
1598 dev->towrite = NULL;
1599 sh->overwrite_disks = 0;
1600 BUG_ON(dev->written);
1601 wbi = dev->written = chosen;
1602 spin_unlock_irq(&sh->stripe_lock);
1603 WARN_ON(dev->page != dev->orig_page);
1605 while (wbi && wbi->bi_iter.bi_sector <
1606 dev->sector + STRIPE_SECTORS) {
1607 if (wbi->bi_rw & REQ_FUA)
1608 set_bit(R5_WantFUA, &dev->flags);
1609 if (wbi->bi_rw & REQ_SYNC)
1610 set_bit(R5_SyncIO, &dev->flags);
1611 if (wbi->bi_rw & REQ_DISCARD)
1612 set_bit(R5_Discard, &dev->flags);
1614 tx = async_copy_data(1, wbi, &dev->page,
1615 dev->sector, tx, sh);
1616 if (dev->page != dev->orig_page) {
1617 set_bit(R5_SkipCopy, &dev->flags);
1618 clear_bit(R5_UPTODATE, &dev->flags);
1619 clear_bit(R5_OVERWRITE, &dev->flags);
1622 wbi = r5_next_bio(wbi, dev->sector);
1625 if (head_sh->batch_head) {
1626 sh = list_first_entry(&sh->batch_list,
1639 static void ops_complete_reconstruct(void *stripe_head_ref)
1641 struct stripe_head *sh = stripe_head_ref;
1642 int disks = sh->disks;
1643 int pd_idx = sh->pd_idx;
1644 int qd_idx = sh->qd_idx;
1646 bool fua = false, sync = false, discard = false;
1648 pr_debug("%s: stripe %llu\n", __func__,
1649 (unsigned long long)sh->sector);
1651 for (i = disks; i--; ) {
1652 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1653 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1654 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1657 for (i = disks; i--; ) {
1658 struct r5dev *dev = &sh->dev[i];
1660 if (dev->written || i == pd_idx || i == qd_idx) {
1661 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1662 set_bit(R5_UPTODATE, &dev->flags);
1664 set_bit(R5_WantFUA, &dev->flags);
1666 set_bit(R5_SyncIO, &dev->flags);
1670 if (sh->reconstruct_state == reconstruct_state_drain_run)
1671 sh->reconstruct_state = reconstruct_state_drain_result;
1672 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1673 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1675 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1676 sh->reconstruct_state = reconstruct_state_result;
1679 set_bit(STRIPE_HANDLE, &sh->state);
1684 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1685 struct dma_async_tx_descriptor *tx)
1687 int disks = sh->disks;
1688 struct page **xor_srcs;
1689 struct async_submit_ctl submit;
1690 int count, pd_idx = sh->pd_idx, i;
1691 struct page *xor_dest;
1693 unsigned long flags;
1695 struct stripe_head *head_sh = sh;
1698 pr_debug("%s: stripe %llu\n", __func__,
1699 (unsigned long long)sh->sector);
1701 for (i = 0; i < sh->disks; i++) {
1704 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1707 if (i >= sh->disks) {
1708 atomic_inc(&sh->count);
1709 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1710 ops_complete_reconstruct(sh);
1715 xor_srcs = to_addr_page(percpu, j);
1716 /* check if prexor is active which means only process blocks
1717 * that are part of a read-modify-write (written)
1719 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1721 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1722 for (i = disks; i--; ) {
1723 struct r5dev *dev = &sh->dev[i];
1724 if (head_sh->dev[i].written)
1725 xor_srcs[count++] = dev->page;
1728 xor_dest = sh->dev[pd_idx].page;
1729 for (i = disks; i--; ) {
1730 struct r5dev *dev = &sh->dev[i];
1732 xor_srcs[count++] = dev->page;
1736 /* 1/ if we prexor'd then the dest is reused as a source
1737 * 2/ if we did not prexor then we are redoing the parity
1738 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1739 * for the synchronous xor case
1741 last_stripe = !head_sh->batch_head ||
1742 list_first_entry(&sh->batch_list,
1743 struct stripe_head, batch_list) == head_sh;
1745 flags = ASYNC_TX_ACK |
1746 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1748 atomic_inc(&head_sh->count);
1749 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1750 to_addr_conv(sh, percpu, j));
1752 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1753 init_async_submit(&submit, flags, tx, NULL, NULL,
1754 to_addr_conv(sh, percpu, j));
1757 if (unlikely(count == 1))
1758 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1760 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1763 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1770 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1771 struct dma_async_tx_descriptor *tx)
1773 struct async_submit_ctl submit;
1774 struct page **blocks;
1775 int count, i, j = 0;
1776 struct stripe_head *head_sh = sh;
1779 unsigned long txflags;
1781 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1783 for (i = 0; i < sh->disks; i++) {
1784 if (sh->pd_idx == i || sh->qd_idx == i)
1786 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1789 if (i >= sh->disks) {
1790 atomic_inc(&sh->count);
1791 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1792 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1793 ops_complete_reconstruct(sh);
1798 blocks = to_addr_page(percpu, j);
1800 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1801 synflags = SYNDROME_SRC_WRITTEN;
1802 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1804 synflags = SYNDROME_SRC_ALL;
1805 txflags = ASYNC_TX_ACK;
1808 count = set_syndrome_sources(blocks, sh, synflags);
1809 last_stripe = !head_sh->batch_head ||
1810 list_first_entry(&sh->batch_list,
1811 struct stripe_head, batch_list) == head_sh;
1814 atomic_inc(&head_sh->count);
1815 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1816 head_sh, to_addr_conv(sh, percpu, j));
1818 init_async_submit(&submit, 0, tx, NULL, NULL,
1819 to_addr_conv(sh, percpu, j));
1820 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1823 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1829 static void ops_complete_check(void *stripe_head_ref)
1831 struct stripe_head *sh = stripe_head_ref;
1833 pr_debug("%s: stripe %llu\n", __func__,
1834 (unsigned long long)sh->sector);
1836 sh->check_state = check_state_check_result;
1837 set_bit(STRIPE_HANDLE, &sh->state);
1841 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1843 int disks = sh->disks;
1844 int pd_idx = sh->pd_idx;
1845 int qd_idx = sh->qd_idx;
1846 struct page *xor_dest;
1847 struct page **xor_srcs = to_addr_page(percpu, 0);
1848 struct dma_async_tx_descriptor *tx;
1849 struct async_submit_ctl submit;
1853 pr_debug("%s: stripe %llu\n", __func__,
1854 (unsigned long long)sh->sector);
1856 BUG_ON(sh->batch_head);
1858 xor_dest = sh->dev[pd_idx].page;
1859 xor_srcs[count++] = xor_dest;
1860 for (i = disks; i--; ) {
1861 if (i == pd_idx || i == qd_idx)
1863 xor_srcs[count++] = sh->dev[i].page;
1866 init_async_submit(&submit, 0, NULL, NULL, NULL,
1867 to_addr_conv(sh, percpu, 0));
1868 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1869 &sh->ops.zero_sum_result, &submit);
1871 atomic_inc(&sh->count);
1872 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1873 tx = async_trigger_callback(&submit);
1876 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1878 struct page **srcs = to_addr_page(percpu, 0);
1879 struct async_submit_ctl submit;
1882 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1883 (unsigned long long)sh->sector, checkp);
1885 BUG_ON(sh->batch_head);
1886 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1890 atomic_inc(&sh->count);
1891 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1892 sh, to_addr_conv(sh, percpu, 0));
1893 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1894 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1897 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1899 int overlap_clear = 0, i, disks = sh->disks;
1900 struct dma_async_tx_descriptor *tx = NULL;
1901 struct r5conf *conf = sh->raid_conf;
1902 int level = conf->level;
1903 struct raid5_percpu *percpu;
1907 percpu = per_cpu_ptr(conf->percpu, cpu);
1908 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1909 ops_run_biofill(sh);
1913 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1915 tx = ops_run_compute5(sh, percpu);
1917 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1918 tx = ops_run_compute6_1(sh, percpu);
1920 tx = ops_run_compute6_2(sh, percpu);
1922 /* terminate the chain if reconstruct is not set to be run */
1923 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1927 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1929 tx = ops_run_prexor5(sh, percpu, tx);
1931 tx = ops_run_prexor6(sh, percpu, tx);
1934 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1935 tx = ops_run_biodrain(sh, tx);
1939 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1941 ops_run_reconstruct5(sh, percpu, tx);
1943 ops_run_reconstruct6(sh, percpu, tx);
1946 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1947 if (sh->check_state == check_state_run)
1948 ops_run_check_p(sh, percpu);
1949 else if (sh->check_state == check_state_run_q)
1950 ops_run_check_pq(sh, percpu, 0);
1951 else if (sh->check_state == check_state_run_pq)
1952 ops_run_check_pq(sh, percpu, 1);
1957 if (overlap_clear && !sh->batch_head)
1958 for (i = disks; i--; ) {
1959 struct r5dev *dev = &sh->dev[i];
1960 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1961 wake_up(&sh->raid_conf->wait_for_overlap);
1966 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
1968 struct stripe_head *sh;
1969 sh = kmem_cache_zalloc(conf->slab_cache, gfp);
1973 sh->raid_conf = conf;
1975 spin_lock_init(&sh->stripe_lock);
1977 if (grow_buffers(sh, gfp)) {
1979 kmem_cache_free(conf->slab_cache, sh);
1982 sh->hash_lock_index =
1983 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
1984 /* we just created an active stripe so... */
1985 atomic_set(&sh->count, 1);
1986 atomic_inc(&conf->active_stripes);
1987 INIT_LIST_HEAD(&sh->lru);
1989 spin_lock_init(&sh->batch_lock);
1990 INIT_LIST_HEAD(&sh->batch_list);
1991 sh->batch_head = NULL;
1993 conf->max_nr_stripes++;
1997 static int grow_stripes(struct r5conf *conf, int num)
1999 struct kmem_cache *sc;
2000 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2002 if (conf->mddev->gendisk)
2003 sprintf(conf->cache_name[0],
2004 "raid%d-%s", conf->level, mdname(conf->mddev));
2006 sprintf(conf->cache_name[0],
2007 "raid%d-%p", conf->level, conf->mddev);
2008 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2010 conf->active_name = 0;
2011 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2012 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2016 conf->slab_cache = sc;
2017 conf->pool_size = devs;
2019 if (!grow_one_stripe(conf, GFP_KERNEL))
2026 * scribble_len - return the required size of the scribble region
2027 * @num - total number of disks in the array
2029 * The size must be enough to contain:
2030 * 1/ a struct page pointer for each device in the array +2
2031 * 2/ room to convert each entry in (1) to its corresponding dma
2032 * (dma_map_page()) or page (page_address()) address.
2034 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2035 * calculate over all devices (not just the data blocks), using zeros in place
2036 * of the P and Q blocks.
2038 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2040 struct flex_array *ret;
2043 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2044 ret = flex_array_alloc(len, cnt, flags);
2047 /* always prealloc all elements, so no locking is required */
2048 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2049 flex_array_free(ret);
2055 static int resize_stripes(struct r5conf *conf, int newsize)
2057 /* Make all the stripes able to hold 'newsize' devices.
2058 * New slots in each stripe get 'page' set to a new page.
2060 * This happens in stages:
2061 * 1/ create a new kmem_cache and allocate the required number of
2063 * 2/ gather all the old stripe_heads and transfer the pages across
2064 * to the new stripe_heads. This will have the side effect of
2065 * freezing the array as once all stripe_heads have been collected,
2066 * no IO will be possible. Old stripe heads are freed once their
2067 * pages have been transferred over, and the old kmem_cache is
2068 * freed when all stripes are done.
2069 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2070 * we simple return a failre status - no need to clean anything up.
2071 * 4/ allocate new pages for the new slots in the new stripe_heads.
2072 * If this fails, we don't bother trying the shrink the
2073 * stripe_heads down again, we just leave them as they are.
2074 * As each stripe_head is processed the new one is released into
2077 * Once step2 is started, we cannot afford to wait for a write,
2078 * so we use GFP_NOIO allocations.
2080 struct stripe_head *osh, *nsh;
2081 LIST_HEAD(newstripes);
2082 struct disk_info *ndisks;
2085 struct kmem_cache *sc;
2089 if (newsize <= conf->pool_size)
2090 return 0; /* never bother to shrink */
2092 err = md_allow_write(conf->mddev);
2097 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2098 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2103 for (i = conf->max_nr_stripes; i; i--) {
2104 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
2108 nsh->raid_conf = conf;
2109 spin_lock_init(&nsh->stripe_lock);
2111 list_add(&nsh->lru, &newstripes);
2114 /* didn't get enough, give up */
2115 while (!list_empty(&newstripes)) {
2116 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2117 list_del(&nsh->lru);
2118 kmem_cache_free(sc, nsh);
2120 kmem_cache_destroy(sc);
2123 /* Step 2 - Must use GFP_NOIO now.
2124 * OK, we have enough stripes, start collecting inactive
2125 * stripes and copying them over
2129 list_for_each_entry(nsh, &newstripes, lru) {
2130 lock_device_hash_lock(conf, hash);
2131 wait_event_cmd(conf->wait_for_stripe,
2132 !list_empty(conf->inactive_list + hash),
2133 unlock_device_hash_lock(conf, hash),
2134 lock_device_hash_lock(conf, hash));
2135 osh = get_free_stripe(conf, hash);
2136 unlock_device_hash_lock(conf, hash);
2137 atomic_set(&nsh->count, 1);
2138 for(i=0; i<conf->pool_size; i++) {
2139 nsh->dev[i].page = osh->dev[i].page;
2140 nsh->dev[i].orig_page = osh->dev[i].page;
2142 for( ; i<newsize; i++)
2143 nsh->dev[i].page = NULL;
2144 nsh->hash_lock_index = hash;
2145 kmem_cache_free(conf->slab_cache, osh);
2147 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2148 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2153 kmem_cache_destroy(conf->slab_cache);
2156 * At this point, we are holding all the stripes so the array
2157 * is completely stalled, so now is a good time to resize
2158 * conf->disks and the scribble region
2160 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2162 for (i=0; i<conf->raid_disks; i++)
2163 ndisks[i] = conf->disks[i];
2165 conf->disks = ndisks;
2170 for_each_present_cpu(cpu) {
2171 struct raid5_percpu *percpu;
2172 struct flex_array *scribble;
2174 percpu = per_cpu_ptr(conf->percpu, cpu);
2175 scribble = scribble_alloc(newsize, conf->chunk_sectors /
2176 STRIPE_SECTORS, GFP_NOIO);
2179 flex_array_free(percpu->scribble);
2180 percpu->scribble = scribble;
2188 /* Step 4, return new stripes to service */
2189 while(!list_empty(&newstripes)) {
2190 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2191 list_del_init(&nsh->lru);
2193 for (i=conf->raid_disks; i < newsize; i++)
2194 if (nsh->dev[i].page == NULL) {
2195 struct page *p = alloc_page(GFP_NOIO);
2196 nsh->dev[i].page = p;
2197 nsh->dev[i].orig_page = p;
2201 release_stripe(nsh);
2203 /* critical section pass, GFP_NOIO no longer needed */
2205 conf->slab_cache = sc;
2206 conf->active_name = 1-conf->active_name;
2207 conf->pool_size = newsize;
2211 static int drop_one_stripe(struct r5conf *conf)
2213 struct stripe_head *sh;
2214 int hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
2216 spin_lock_irq(conf->hash_locks + hash);
2217 sh = get_free_stripe(conf, hash);
2218 spin_unlock_irq(conf->hash_locks + hash);
2221 BUG_ON(atomic_read(&sh->count));
2223 kmem_cache_free(conf->slab_cache, sh);
2224 atomic_dec(&conf->active_stripes);
2225 conf->max_nr_stripes--;
2229 static void shrink_stripes(struct r5conf *conf)
2231 while (conf->max_nr_stripes &&
2232 drop_one_stripe(conf))
2235 if (conf->slab_cache)
2236 kmem_cache_destroy(conf->slab_cache);
2237 conf->slab_cache = NULL;
2240 static void raid5_end_read_request(struct bio * bi, int error)
2242 struct stripe_head *sh = bi->bi_private;
2243 struct r5conf *conf = sh->raid_conf;
2244 int disks = sh->disks, i;
2245 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2246 char b[BDEVNAME_SIZE];
2247 struct md_rdev *rdev = NULL;
2250 for (i=0 ; i<disks; i++)
2251 if (bi == &sh->dev[i].req)
2254 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2255 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2261 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2262 /* If replacement finished while this request was outstanding,
2263 * 'replacement' might be NULL already.
2264 * In that case it moved down to 'rdev'.
2265 * rdev is not removed until all requests are finished.
2267 rdev = conf->disks[i].replacement;
2269 rdev = conf->disks[i].rdev;
2271 if (use_new_offset(conf, sh))
2272 s = sh->sector + rdev->new_data_offset;
2274 s = sh->sector + rdev->data_offset;
2276 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2277 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2278 /* Note that this cannot happen on a
2279 * replacement device. We just fail those on
2284 "md/raid:%s: read error corrected"
2285 " (%lu sectors at %llu on %s)\n",
2286 mdname(conf->mddev), STRIPE_SECTORS,
2287 (unsigned long long)s,
2288 bdevname(rdev->bdev, b));
2289 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2290 clear_bit(R5_ReadError, &sh->dev[i].flags);
2291 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2292 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2293 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2295 if (atomic_read(&rdev->read_errors))
2296 atomic_set(&rdev->read_errors, 0);
2298 const char *bdn = bdevname(rdev->bdev, b);
2302 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2303 atomic_inc(&rdev->read_errors);
2304 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2307 "md/raid:%s: read error on replacement device "
2308 "(sector %llu on %s).\n",
2309 mdname(conf->mddev),
2310 (unsigned long long)s,
2312 else if (conf->mddev->degraded >= conf->max_degraded) {
2316 "md/raid:%s: read error not correctable "
2317 "(sector %llu on %s).\n",
2318 mdname(conf->mddev),
2319 (unsigned long long)s,
2321 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2326 "md/raid:%s: read error NOT corrected!! "
2327 "(sector %llu on %s).\n",
2328 mdname(conf->mddev),
2329 (unsigned long long)s,
2331 } else if (atomic_read(&rdev->read_errors)
2332 > conf->max_nr_stripes)
2334 "md/raid:%s: Too many read errors, failing device %s.\n",
2335 mdname(conf->mddev), bdn);
2338 if (set_bad && test_bit(In_sync, &rdev->flags)
2339 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2342 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2343 set_bit(R5_ReadError, &sh->dev[i].flags);
2344 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2346 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2348 clear_bit(R5_ReadError, &sh->dev[i].flags);
2349 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2351 && test_bit(In_sync, &rdev->flags)
2352 && rdev_set_badblocks(
2353 rdev, sh->sector, STRIPE_SECTORS, 0)))
2354 md_error(conf->mddev, rdev);
2357 rdev_dec_pending(rdev, conf->mddev);
2358 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2359 set_bit(STRIPE_HANDLE, &sh->state);
2363 static void raid5_end_write_request(struct bio *bi, int error)
2365 struct stripe_head *sh = bi->bi_private;
2366 struct r5conf *conf = sh->raid_conf;
2367 int disks = sh->disks, i;
2368 struct md_rdev *uninitialized_var(rdev);
2369 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2372 int replacement = 0;
2374 for (i = 0 ; i < disks; i++) {
2375 if (bi == &sh->dev[i].req) {
2376 rdev = conf->disks[i].rdev;
2379 if (bi == &sh->dev[i].rreq) {
2380 rdev = conf->disks[i].replacement;
2384 /* rdev was removed and 'replacement'
2385 * replaced it. rdev is not removed
2386 * until all requests are finished.
2388 rdev = conf->disks[i].rdev;
2392 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2393 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2402 md_error(conf->mddev, rdev);
2403 else if (is_badblock(rdev, sh->sector,
2405 &first_bad, &bad_sectors))
2406 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2409 set_bit(STRIPE_DEGRADED, &sh->state);
2410 set_bit(WriteErrorSeen, &rdev->flags);
2411 set_bit(R5_WriteError, &sh->dev[i].flags);
2412 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2413 set_bit(MD_RECOVERY_NEEDED,
2414 &rdev->mddev->recovery);
2415 } else if (is_badblock(rdev, sh->sector,
2417 &first_bad, &bad_sectors)) {
2418 set_bit(R5_MadeGood, &sh->dev[i].flags);
2419 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2420 /* That was a successful write so make
2421 * sure it looks like we already did
2424 set_bit(R5_ReWrite, &sh->dev[i].flags);
2427 rdev_dec_pending(rdev, conf->mddev);
2429 if (sh->batch_head && !uptodate)
2430 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2432 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2433 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2434 set_bit(STRIPE_HANDLE, &sh->state);
2437 if (sh->batch_head && sh != sh->batch_head)
2438 release_stripe(sh->batch_head);
2441 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2443 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2445 struct r5dev *dev = &sh->dev[i];
2447 bio_init(&dev->req);
2448 dev->req.bi_io_vec = &dev->vec;
2449 dev->req.bi_max_vecs = 1;
2450 dev->req.bi_private = sh;
2452 bio_init(&dev->rreq);
2453 dev->rreq.bi_io_vec = &dev->rvec;
2454 dev->rreq.bi_max_vecs = 1;
2455 dev->rreq.bi_private = sh;
2458 dev->sector = compute_blocknr(sh, i, previous);
2461 static void error(struct mddev *mddev, struct md_rdev *rdev)
2463 char b[BDEVNAME_SIZE];
2464 struct r5conf *conf = mddev->private;
2465 unsigned long flags;
2466 pr_debug("raid456: error called\n");
2468 spin_lock_irqsave(&conf->device_lock, flags);
2469 clear_bit(In_sync, &rdev->flags);
2470 mddev->degraded = calc_degraded(conf);
2471 spin_unlock_irqrestore(&conf->device_lock, flags);
2472 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2474 set_bit(Blocked, &rdev->flags);
2475 set_bit(Faulty, &rdev->flags);
2476 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2478 "md/raid:%s: Disk failure on %s, disabling device.\n"
2479 "md/raid:%s: Operation continuing on %d devices.\n",
2481 bdevname(rdev->bdev, b),
2483 conf->raid_disks - mddev->degraded);
2487 * Input: a 'big' sector number,
2488 * Output: index of the data and parity disk, and the sector # in them.
2490 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2491 int previous, int *dd_idx,
2492 struct stripe_head *sh)
2494 sector_t stripe, stripe2;
2495 sector_t chunk_number;
2496 unsigned int chunk_offset;
2499 sector_t new_sector;
2500 int algorithm = previous ? conf->prev_algo
2502 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2503 : conf->chunk_sectors;
2504 int raid_disks = previous ? conf->previous_raid_disks
2506 int data_disks = raid_disks - conf->max_degraded;
2508 /* First compute the information on this sector */
2511 * Compute the chunk number and the sector offset inside the chunk
2513 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2514 chunk_number = r_sector;
2517 * Compute the stripe number
2519 stripe = chunk_number;
2520 *dd_idx = sector_div(stripe, data_disks);
2523 * Select the parity disk based on the user selected algorithm.
2525 pd_idx = qd_idx = -1;
2526 switch(conf->level) {
2528 pd_idx = data_disks;
2531 switch (algorithm) {
2532 case ALGORITHM_LEFT_ASYMMETRIC:
2533 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2534 if (*dd_idx >= pd_idx)
2537 case ALGORITHM_RIGHT_ASYMMETRIC:
2538 pd_idx = sector_div(stripe2, raid_disks);
2539 if (*dd_idx >= pd_idx)
2542 case ALGORITHM_LEFT_SYMMETRIC:
2543 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2544 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2546 case ALGORITHM_RIGHT_SYMMETRIC:
2547 pd_idx = sector_div(stripe2, raid_disks);
2548 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2550 case ALGORITHM_PARITY_0:
2554 case ALGORITHM_PARITY_N:
2555 pd_idx = data_disks;
2563 switch (algorithm) {
2564 case ALGORITHM_LEFT_ASYMMETRIC:
2565 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2566 qd_idx = pd_idx + 1;
2567 if (pd_idx == raid_disks-1) {
2568 (*dd_idx)++; /* Q D D D P */
2570 } else if (*dd_idx >= pd_idx)
2571 (*dd_idx) += 2; /* D D P Q D */
2573 case ALGORITHM_RIGHT_ASYMMETRIC:
2574 pd_idx = sector_div(stripe2, raid_disks);
2575 qd_idx = pd_idx + 1;
2576 if (pd_idx == raid_disks-1) {
2577 (*dd_idx)++; /* Q D D D P */
2579 } else if (*dd_idx >= pd_idx)
2580 (*dd_idx) += 2; /* D D P Q D */
2582 case ALGORITHM_LEFT_SYMMETRIC:
2583 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2584 qd_idx = (pd_idx + 1) % raid_disks;
2585 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2587 case ALGORITHM_RIGHT_SYMMETRIC:
2588 pd_idx = sector_div(stripe2, raid_disks);
2589 qd_idx = (pd_idx + 1) % raid_disks;
2590 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2593 case ALGORITHM_PARITY_0:
2598 case ALGORITHM_PARITY_N:
2599 pd_idx = data_disks;
2600 qd_idx = data_disks + 1;
2603 case ALGORITHM_ROTATING_ZERO_RESTART:
2604 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2605 * of blocks for computing Q is different.
2607 pd_idx = sector_div(stripe2, raid_disks);
2608 qd_idx = pd_idx + 1;
2609 if (pd_idx == raid_disks-1) {
2610 (*dd_idx)++; /* Q D D D P */
2612 } else if (*dd_idx >= pd_idx)
2613 (*dd_idx) += 2; /* D D P Q D */
2617 case ALGORITHM_ROTATING_N_RESTART:
2618 /* Same a left_asymmetric, by first stripe is
2619 * D D D P Q rather than
2623 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2624 qd_idx = pd_idx + 1;
2625 if (pd_idx == raid_disks-1) {
2626 (*dd_idx)++; /* Q D D D P */
2628 } else if (*dd_idx >= pd_idx)
2629 (*dd_idx) += 2; /* D D P Q D */
2633 case ALGORITHM_ROTATING_N_CONTINUE:
2634 /* Same as left_symmetric but Q is before P */
2635 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2636 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2637 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2641 case ALGORITHM_LEFT_ASYMMETRIC_6:
2642 /* RAID5 left_asymmetric, with Q on last device */
2643 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2644 if (*dd_idx >= pd_idx)
2646 qd_idx = raid_disks - 1;
2649 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2650 pd_idx = sector_div(stripe2, raid_disks-1);
2651 if (*dd_idx >= pd_idx)
2653 qd_idx = raid_disks - 1;
2656 case ALGORITHM_LEFT_SYMMETRIC_6:
2657 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2658 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2659 qd_idx = raid_disks - 1;
2662 case ALGORITHM_RIGHT_SYMMETRIC_6:
2663 pd_idx = sector_div(stripe2, raid_disks-1);
2664 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2665 qd_idx = raid_disks - 1;
2668 case ALGORITHM_PARITY_0_6:
2671 qd_idx = raid_disks - 1;
2681 sh->pd_idx = pd_idx;
2682 sh->qd_idx = qd_idx;
2683 sh->ddf_layout = ddf_layout;
2686 * Finally, compute the new sector number
2688 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2692 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2694 struct r5conf *conf = sh->raid_conf;
2695 int raid_disks = sh->disks;
2696 int data_disks = raid_disks - conf->max_degraded;
2697 sector_t new_sector = sh->sector, check;
2698 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2699 : conf->chunk_sectors;
2700 int algorithm = previous ? conf->prev_algo
2704 sector_t chunk_number;
2705 int dummy1, dd_idx = i;
2707 struct stripe_head sh2;
2709 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2710 stripe = new_sector;
2712 if (i == sh->pd_idx)
2714 switch(conf->level) {
2717 switch (algorithm) {
2718 case ALGORITHM_LEFT_ASYMMETRIC:
2719 case ALGORITHM_RIGHT_ASYMMETRIC:
2723 case ALGORITHM_LEFT_SYMMETRIC:
2724 case ALGORITHM_RIGHT_SYMMETRIC:
2727 i -= (sh->pd_idx + 1);
2729 case ALGORITHM_PARITY_0:
2732 case ALGORITHM_PARITY_N:
2739 if (i == sh->qd_idx)
2740 return 0; /* It is the Q disk */
2741 switch (algorithm) {
2742 case ALGORITHM_LEFT_ASYMMETRIC:
2743 case ALGORITHM_RIGHT_ASYMMETRIC:
2744 case ALGORITHM_ROTATING_ZERO_RESTART:
2745 case ALGORITHM_ROTATING_N_RESTART:
2746 if (sh->pd_idx == raid_disks-1)
2747 i--; /* Q D D D P */
2748 else if (i > sh->pd_idx)
2749 i -= 2; /* D D P Q D */
2751 case ALGORITHM_LEFT_SYMMETRIC:
2752 case ALGORITHM_RIGHT_SYMMETRIC:
2753 if (sh->pd_idx == raid_disks-1)
2754 i--; /* Q D D D P */
2759 i -= (sh->pd_idx + 2);
2762 case ALGORITHM_PARITY_0:
2765 case ALGORITHM_PARITY_N:
2767 case ALGORITHM_ROTATING_N_CONTINUE:
2768 /* Like left_symmetric, but P is before Q */
2769 if (sh->pd_idx == 0)
2770 i--; /* P D D D Q */
2775 i -= (sh->pd_idx + 1);
2778 case ALGORITHM_LEFT_ASYMMETRIC_6:
2779 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2783 case ALGORITHM_LEFT_SYMMETRIC_6:
2784 case ALGORITHM_RIGHT_SYMMETRIC_6:
2786 i += data_disks + 1;
2787 i -= (sh->pd_idx + 1);
2789 case ALGORITHM_PARITY_0_6:
2798 chunk_number = stripe * data_disks + i;
2799 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2801 check = raid5_compute_sector(conf, r_sector,
2802 previous, &dummy1, &sh2);
2803 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2804 || sh2.qd_idx != sh->qd_idx) {
2805 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2806 mdname(conf->mddev));
2813 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2814 int rcw, int expand)
2816 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2817 struct r5conf *conf = sh->raid_conf;
2818 int level = conf->level;
2822 for (i = disks; i--; ) {
2823 struct r5dev *dev = &sh->dev[i];
2826 set_bit(R5_LOCKED, &dev->flags);
2827 set_bit(R5_Wantdrain, &dev->flags);
2829 clear_bit(R5_UPTODATE, &dev->flags);
2833 /* if we are not expanding this is a proper write request, and
2834 * there will be bios with new data to be drained into the
2839 /* False alarm, nothing to do */
2841 sh->reconstruct_state = reconstruct_state_drain_run;
2842 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2844 sh->reconstruct_state = reconstruct_state_run;
2846 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2848 if (s->locked + conf->max_degraded == disks)
2849 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2850 atomic_inc(&conf->pending_full_writes);
2852 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2853 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2854 BUG_ON(level == 6 &&
2855 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2856 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2858 for (i = disks; i--; ) {
2859 struct r5dev *dev = &sh->dev[i];
2860 if (i == pd_idx || i == qd_idx)
2864 (test_bit(R5_UPTODATE, &dev->flags) ||
2865 test_bit(R5_Wantcompute, &dev->flags))) {
2866 set_bit(R5_Wantdrain, &dev->flags);
2867 set_bit(R5_LOCKED, &dev->flags);
2868 clear_bit(R5_UPTODATE, &dev->flags);
2873 /* False alarm - nothing to do */
2875 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2876 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2877 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2878 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2881 /* keep the parity disk(s) locked while asynchronous operations
2884 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2885 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2889 int qd_idx = sh->qd_idx;
2890 struct r5dev *dev = &sh->dev[qd_idx];
2892 set_bit(R5_LOCKED, &dev->flags);
2893 clear_bit(R5_UPTODATE, &dev->flags);
2897 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2898 __func__, (unsigned long long)sh->sector,
2899 s->locked, s->ops_request);
2903 * Each stripe/dev can have one or more bion attached.
2904 * toread/towrite point to the first in a chain.
2905 * The bi_next chain must be in order.
2907 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2908 int forwrite, int previous)
2911 struct r5conf *conf = sh->raid_conf;
2914 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2915 (unsigned long long)bi->bi_iter.bi_sector,
2916 (unsigned long long)sh->sector);
2919 * If several bio share a stripe. The bio bi_phys_segments acts as a
2920 * reference count to avoid race. The reference count should already be
2921 * increased before this function is called (for example, in
2922 * make_request()), so other bio sharing this stripe will not free the
2923 * stripe. If a stripe is owned by one stripe, the stripe lock will
2926 spin_lock_irq(&sh->stripe_lock);
2927 /* Don't allow new IO added to stripes in batch list */
2931 bip = &sh->dev[dd_idx].towrite;
2935 bip = &sh->dev[dd_idx].toread;
2936 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2937 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2939 bip = & (*bip)->bi_next;
2941 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2944 if (!forwrite || previous)
2945 clear_bit(STRIPE_BATCH_READY, &sh->state);
2947 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2951 raid5_inc_bi_active_stripes(bi);
2954 /* check if page is covered */
2955 sector_t sector = sh->dev[dd_idx].sector;
2956 for (bi=sh->dev[dd_idx].towrite;
2957 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2958 bi && bi->bi_iter.bi_sector <= sector;
2959 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2960 if (bio_end_sector(bi) >= sector)
2961 sector = bio_end_sector(bi);
2963 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2964 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
2965 sh->overwrite_disks++;
2968 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2969 (unsigned long long)(*bip)->bi_iter.bi_sector,
2970 (unsigned long long)sh->sector, dd_idx);
2971 spin_unlock_irq(&sh->stripe_lock);
2973 if (conf->mddev->bitmap && firstwrite) {
2974 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2976 sh->bm_seq = conf->seq_flush+1;
2977 set_bit(STRIPE_BIT_DELAY, &sh->state);
2980 if (stripe_can_batch(sh))
2981 stripe_add_to_batch_list(conf, sh);
2985 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2986 spin_unlock_irq(&sh->stripe_lock);
2990 static void end_reshape(struct r5conf *conf);
2992 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2993 struct stripe_head *sh)
2995 int sectors_per_chunk =
2996 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2998 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2999 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3001 raid5_compute_sector(conf,
3002 stripe * (disks - conf->max_degraded)
3003 *sectors_per_chunk + chunk_offset,
3009 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3010 struct stripe_head_state *s, int disks,
3011 struct bio **return_bi)
3014 BUG_ON(sh->batch_head);
3015 for (i = disks; i--; ) {
3019 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3020 struct md_rdev *rdev;
3022 rdev = rcu_dereference(conf->disks[i].rdev);
3023 if (rdev && test_bit(In_sync, &rdev->flags))
3024 atomic_inc(&rdev->nr_pending);
3029 if (!rdev_set_badblocks(
3033 md_error(conf->mddev, rdev);
3034 rdev_dec_pending(rdev, conf->mddev);
3037 spin_lock_irq(&sh->stripe_lock);
3038 /* fail all writes first */
3039 bi = sh->dev[i].towrite;
3040 sh->dev[i].towrite = NULL;
3041 sh->overwrite_disks = 0;
3042 spin_unlock_irq(&sh->stripe_lock);
3046 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3047 wake_up(&conf->wait_for_overlap);
3049 while (bi && bi->bi_iter.bi_sector <
3050 sh->dev[i].sector + STRIPE_SECTORS) {
3051 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3052 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3053 if (!raid5_dec_bi_active_stripes(bi)) {
3054 md_write_end(conf->mddev);
3055 bi->bi_next = *return_bi;
3061 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3062 STRIPE_SECTORS, 0, 0);
3064 /* and fail all 'written' */
3065 bi = sh->dev[i].written;
3066 sh->dev[i].written = NULL;
3067 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3068 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3069 sh->dev[i].page = sh->dev[i].orig_page;
3072 if (bi) bitmap_end = 1;
3073 while (bi && bi->bi_iter.bi_sector <
3074 sh->dev[i].sector + STRIPE_SECTORS) {
3075 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3076 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3077 if (!raid5_dec_bi_active_stripes(bi)) {
3078 md_write_end(conf->mddev);
3079 bi->bi_next = *return_bi;
3085 /* fail any reads if this device is non-operational and
3086 * the data has not reached the cache yet.
3088 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3089 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3090 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3091 spin_lock_irq(&sh->stripe_lock);
3092 bi = sh->dev[i].toread;
3093 sh->dev[i].toread = NULL;
3094 spin_unlock_irq(&sh->stripe_lock);
3095 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3096 wake_up(&conf->wait_for_overlap);
3097 while (bi && bi->bi_iter.bi_sector <
3098 sh->dev[i].sector + STRIPE_SECTORS) {
3099 struct bio *nextbi =
3100 r5_next_bio(bi, sh->dev[i].sector);
3101 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3102 if (!raid5_dec_bi_active_stripes(bi)) {
3103 bi->bi_next = *return_bi;
3110 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3111 STRIPE_SECTORS, 0, 0);
3112 /* If we were in the middle of a write the parity block might
3113 * still be locked - so just clear all R5_LOCKED flags
3115 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3118 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3119 if (atomic_dec_and_test(&conf->pending_full_writes))
3120 md_wakeup_thread(conf->mddev->thread);
3124 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3125 struct stripe_head_state *s)
3130 BUG_ON(sh->batch_head);
3131 clear_bit(STRIPE_SYNCING, &sh->state);
3132 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3133 wake_up(&conf->wait_for_overlap);
3136 /* There is nothing more to do for sync/check/repair.
3137 * Don't even need to abort as that is handled elsewhere
3138 * if needed, and not always wanted e.g. if there is a known
3140 * For recover/replace we need to record a bad block on all
3141 * non-sync devices, or abort the recovery
3143 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3144 /* During recovery devices cannot be removed, so
3145 * locking and refcounting of rdevs is not needed
3147 for (i = 0; i < conf->raid_disks; i++) {
3148 struct md_rdev *rdev = conf->disks[i].rdev;
3150 && !test_bit(Faulty, &rdev->flags)
3151 && !test_bit(In_sync, &rdev->flags)
3152 && !rdev_set_badblocks(rdev, sh->sector,
3155 rdev = conf->disks[i].replacement;
3157 && !test_bit(Faulty, &rdev->flags)
3158 && !test_bit(In_sync, &rdev->flags)
3159 && !rdev_set_badblocks(rdev, sh->sector,
3164 conf->recovery_disabled =
3165 conf->mddev->recovery_disabled;
3167 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3170 static int want_replace(struct stripe_head *sh, int disk_idx)
3172 struct md_rdev *rdev;
3174 /* Doing recovery so rcu locking not required */
3175 rdev = sh->raid_conf->disks[disk_idx].replacement;
3177 && !test_bit(Faulty, &rdev->flags)
3178 && !test_bit(In_sync, &rdev->flags)
3179 && (rdev->recovery_offset <= sh->sector
3180 || rdev->mddev->recovery_cp <= sh->sector))
3186 /* fetch_block - checks the given member device to see if its data needs
3187 * to be read or computed to satisfy a request.
3189 * Returns 1 when no more member devices need to be checked, otherwise returns
3190 * 0 to tell the loop in handle_stripe_fill to continue
3193 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3194 int disk_idx, int disks)
3196 struct r5dev *dev = &sh->dev[disk_idx];
3197 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3198 &sh->dev[s->failed_num[1]] };
3202 if (test_bit(R5_LOCKED, &dev->flags) ||
3203 test_bit(R5_UPTODATE, &dev->flags))
3204 /* No point reading this as we already have it or have
3205 * decided to get it.
3210 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3211 /* We need this block to directly satisfy a request */
3214 if (s->syncing || s->expanding ||
3215 (s->replacing && want_replace(sh, disk_idx)))
3216 /* When syncing, or expanding we read everything.
3217 * When replacing, we need the replaced block.
3221 if ((s->failed >= 1 && fdev[0]->toread) ||
3222 (s->failed >= 2 && fdev[1]->toread))
3223 /* If we want to read from a failed device, then
3224 * we need to actually read every other device.
3228 /* Sometimes neither read-modify-write nor reconstruct-write
3229 * cycles can work. In those cases we read every block we
3230 * can. Then the parity-update is certain to have enough to
3232 * This can only be a problem when we need to write something,
3233 * and some device has failed. If either of those tests
3234 * fail we need look no further.
3236 if (!s->failed || !s->to_write)
3239 if (test_bit(R5_Insync, &dev->flags) &&
3240 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3241 /* Pre-reads at not permitted until after short delay
3242 * to gather multiple requests. However if this
3243 * device is no Insync, the block could only be be computed
3244 * and there is no need to delay that.
3248 for (i = 0; i < s->failed; i++) {
3249 if (fdev[i]->towrite &&
3250 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3251 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3252 /* If we have a partial write to a failed
3253 * device, then we will need to reconstruct
3254 * the content of that device, so all other
3255 * devices must be read.
3260 /* If we are forced to do a reconstruct-write, either because
3261 * the current RAID6 implementation only supports that, or
3262 * or because parity cannot be trusted and we are currently
3263 * recovering it, there is extra need to be careful.
3264 * If one of the devices that we would need to read, because
3265 * it is not being overwritten (and maybe not written at all)
3266 * is missing/faulty, then we need to read everything we can.
3268 if (sh->raid_conf->level != 6 &&
3269 sh->sector < sh->raid_conf->mddev->recovery_cp)
3270 /* reconstruct-write isn't being forced */
3272 for (i = 0; i < s->failed; i++) {
3273 if (!test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3274 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3281 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3282 int disk_idx, int disks)
3284 struct r5dev *dev = &sh->dev[disk_idx];
3286 /* is the data in this block needed, and can we get it? */
3287 if (need_this_block(sh, s, disk_idx, disks)) {
3288 /* we would like to get this block, possibly by computing it,
3289 * otherwise read it if the backing disk is insync
3291 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3292 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3293 if ((s->uptodate == disks - 1) &&
3294 (s->failed && (disk_idx == s->failed_num[0] ||
3295 disk_idx == s->failed_num[1]))) {
3296 /* have disk failed, and we're requested to fetch it;
3299 pr_debug("Computing stripe %llu block %d\n",
3300 (unsigned long long)sh->sector, disk_idx);
3301 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3302 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3303 set_bit(R5_Wantcompute, &dev->flags);
3304 sh->ops.target = disk_idx;
3305 sh->ops.target2 = -1; /* no 2nd target */
3307 /* Careful: from this point on 'uptodate' is in the eye
3308 * of raid_run_ops which services 'compute' operations
3309 * before writes. R5_Wantcompute flags a block that will
3310 * be R5_UPTODATE by the time it is needed for a
3311 * subsequent operation.
3315 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3316 /* Computing 2-failure is *very* expensive; only
3317 * do it if failed >= 2
3320 for (other = disks; other--; ) {
3321 if (other == disk_idx)
3323 if (!test_bit(R5_UPTODATE,
3324 &sh->dev[other].flags))
3328 pr_debug("Computing stripe %llu blocks %d,%d\n",
3329 (unsigned long long)sh->sector,
3331 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3332 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3333 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3334 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3335 sh->ops.target = disk_idx;
3336 sh->ops.target2 = other;
3340 } else if (test_bit(R5_Insync, &dev->flags)) {
3341 set_bit(R5_LOCKED, &dev->flags);
3342 set_bit(R5_Wantread, &dev->flags);
3344 pr_debug("Reading block %d (sync=%d)\n",
3345 disk_idx, s->syncing);
3353 * handle_stripe_fill - read or compute data to satisfy pending requests.
3355 static void handle_stripe_fill(struct stripe_head *sh,
3356 struct stripe_head_state *s,
3361 BUG_ON(sh->batch_head);
3362 /* look for blocks to read/compute, skip this if a compute
3363 * is already in flight, or if the stripe contents are in the
3364 * midst of changing due to a write
3366 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3367 !sh->reconstruct_state)
3368 for (i = disks; i--; )
3369 if (fetch_block(sh, s, i, disks))
3371 set_bit(STRIPE_HANDLE, &sh->state);
3374 /* handle_stripe_clean_event
3375 * any written block on an uptodate or failed drive can be returned.
3376 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3377 * never LOCKED, so we don't need to test 'failed' directly.
3379 static void handle_stripe_clean_event(struct r5conf *conf,
3380 struct stripe_head *sh, int disks, struct bio **return_bi)
3384 int discard_pending = 0;
3385 struct stripe_head *head_sh = sh;
3386 bool do_endio = false;
3389 for (i = disks; i--; )
3390 if (sh->dev[i].written) {
3392 if (!test_bit(R5_LOCKED, &dev->flags) &&
3393 (test_bit(R5_UPTODATE, &dev->flags) ||
3394 test_bit(R5_Discard, &dev->flags) ||
3395 test_bit(R5_SkipCopy, &dev->flags))) {
3396 /* We can return any write requests */
3397 struct bio *wbi, *wbi2;
3398 pr_debug("Return write for disc %d\n", i);
3399 if (test_and_clear_bit(R5_Discard, &dev->flags))
3400 clear_bit(R5_UPTODATE, &dev->flags);
3401 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3402 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3407 dev->page = dev->orig_page;
3409 dev->written = NULL;
3410 while (wbi && wbi->bi_iter.bi_sector <
3411 dev->sector + STRIPE_SECTORS) {
3412 wbi2 = r5_next_bio(wbi, dev->sector);
3413 if (!raid5_dec_bi_active_stripes(wbi)) {
3414 md_write_end(conf->mddev);
3415 wbi->bi_next = *return_bi;
3420 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3422 !test_bit(STRIPE_DEGRADED, &sh->state),
3424 if (head_sh->batch_head) {
3425 sh = list_first_entry(&sh->batch_list,
3428 if (sh != head_sh) {
3435 } else if (test_bit(R5_Discard, &dev->flags))
3436 discard_pending = 1;
3437 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3438 WARN_ON(dev->page != dev->orig_page);
3440 if (!discard_pending &&
3441 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3442 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3443 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3444 if (sh->qd_idx >= 0) {
3445 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3446 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3448 /* now that discard is done we can proceed with any sync */
3449 clear_bit(STRIPE_DISCARD, &sh->state);
3451 * SCSI discard will change some bio fields and the stripe has
3452 * no updated data, so remove it from hash list and the stripe
3453 * will be reinitialized
3455 spin_lock_irq(&conf->device_lock);
3458 if (head_sh->batch_head) {
3459 sh = list_first_entry(&sh->batch_list,
3460 struct stripe_head, batch_list);
3464 spin_unlock_irq(&conf->device_lock);
3467 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3468 set_bit(STRIPE_HANDLE, &sh->state);
3472 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3473 if (atomic_dec_and_test(&conf->pending_full_writes))
3474 md_wakeup_thread(conf->mddev->thread);
3476 if (!head_sh->batch_head || !do_endio)
3478 for (i = 0; i < head_sh->disks; i++) {
3479 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
3482 while (!list_empty(&head_sh->batch_list)) {
3484 sh = list_first_entry(&head_sh->batch_list,
3485 struct stripe_head, batch_list);
3486 list_del_init(&sh->batch_list);
3488 set_mask_bits(&sh->state, ~STRIPE_EXPAND_SYNC_FLAG,
3489 head_sh->state & ~((1 << STRIPE_ACTIVE) |
3490 (1 << STRIPE_PREREAD_ACTIVE) |
3491 STRIPE_EXPAND_SYNC_FLAG));
3492 sh->check_state = head_sh->check_state;
3493 sh->reconstruct_state = head_sh->reconstruct_state;
3494 for (i = 0; i < sh->disks; i++) {
3495 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3497 sh->dev[i].flags = head_sh->dev[i].flags;
3500 spin_lock_irq(&sh->stripe_lock);
3501 sh->batch_head = NULL;
3502 spin_unlock_irq(&sh->stripe_lock);
3503 if (sh->state & STRIPE_EXPAND_SYNC_FLAG)
3504 set_bit(STRIPE_HANDLE, &sh->state);
3508 spin_lock_irq(&head_sh->stripe_lock);
3509 head_sh->batch_head = NULL;
3510 spin_unlock_irq(&head_sh->stripe_lock);
3511 wake_up_nr(&conf->wait_for_overlap, wakeup_nr);
3512 if (head_sh->state & STRIPE_EXPAND_SYNC_FLAG)
3513 set_bit(STRIPE_HANDLE, &head_sh->state);
3516 static void handle_stripe_dirtying(struct r5conf *conf,
3517 struct stripe_head *sh,
3518 struct stripe_head_state *s,
3521 int rmw = 0, rcw = 0, i;
3522 sector_t recovery_cp = conf->mddev->recovery_cp;
3524 /* Check whether resync is now happening or should start.
3525 * If yes, then the array is dirty (after unclean shutdown or
3526 * initial creation), so parity in some stripes might be inconsistent.
3527 * In this case, we need to always do reconstruct-write, to ensure
3528 * that in case of drive failure or read-error correction, we
3529 * generate correct data from the parity.
3531 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3532 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3534 /* Calculate the real rcw later - for now make it
3535 * look like rcw is cheaper
3538 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3539 conf->rmw_level, (unsigned long long)recovery_cp,
3540 (unsigned long long)sh->sector);
3541 } else for (i = disks; i--; ) {
3542 /* would I have to read this buffer for read_modify_write */
3543 struct r5dev *dev = &sh->dev[i];
3544 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3545 !test_bit(R5_LOCKED, &dev->flags) &&
3546 !(test_bit(R5_UPTODATE, &dev->flags) ||
3547 test_bit(R5_Wantcompute, &dev->flags))) {
3548 if (test_bit(R5_Insync, &dev->flags))
3551 rmw += 2*disks; /* cannot read it */
3553 /* Would I have to read this buffer for reconstruct_write */
3554 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3555 i != sh->pd_idx && i != sh->qd_idx &&
3556 !test_bit(R5_LOCKED, &dev->flags) &&
3557 !(test_bit(R5_UPTODATE, &dev->flags) ||
3558 test_bit(R5_Wantcompute, &dev->flags))) {
3559 if (test_bit(R5_Insync, &dev->flags))
3565 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3566 (unsigned long long)sh->sector, rmw, rcw);
3567 set_bit(STRIPE_HANDLE, &sh->state);
3568 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3569 /* prefer read-modify-write, but need to get some data */
3570 if (conf->mddev->queue)
3571 blk_add_trace_msg(conf->mddev->queue,
3572 "raid5 rmw %llu %d",
3573 (unsigned long long)sh->sector, rmw);
3574 for (i = disks; i--; ) {
3575 struct r5dev *dev = &sh->dev[i];
3576 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3577 !test_bit(R5_LOCKED, &dev->flags) &&
3578 !(test_bit(R5_UPTODATE, &dev->flags) ||
3579 test_bit(R5_Wantcompute, &dev->flags)) &&
3580 test_bit(R5_Insync, &dev->flags)) {
3581 if (test_bit(STRIPE_PREREAD_ACTIVE,
3583 pr_debug("Read_old block %d for r-m-w\n",
3585 set_bit(R5_LOCKED, &dev->flags);
3586 set_bit(R5_Wantread, &dev->flags);
3589 set_bit(STRIPE_DELAYED, &sh->state);
3590 set_bit(STRIPE_HANDLE, &sh->state);
3595 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3596 /* want reconstruct write, but need to get some data */
3599 for (i = disks; i--; ) {
3600 struct r5dev *dev = &sh->dev[i];
3601 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3602 i != sh->pd_idx && i != sh->qd_idx &&
3603 !test_bit(R5_LOCKED, &dev->flags) &&
3604 !(test_bit(R5_UPTODATE, &dev->flags) ||
3605 test_bit(R5_Wantcompute, &dev->flags))) {
3607 if (test_bit(R5_Insync, &dev->flags) &&
3608 test_bit(STRIPE_PREREAD_ACTIVE,
3610 pr_debug("Read_old block "
3611 "%d for Reconstruct\n", i);
3612 set_bit(R5_LOCKED, &dev->flags);
3613 set_bit(R5_Wantread, &dev->flags);
3617 set_bit(STRIPE_DELAYED, &sh->state);
3618 set_bit(STRIPE_HANDLE, &sh->state);
3622 if (rcw && conf->mddev->queue)
3623 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3624 (unsigned long long)sh->sector,
3625 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3628 if (rcw > disks && rmw > disks &&
3629 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3630 set_bit(STRIPE_DELAYED, &sh->state);
3632 /* now if nothing is locked, and if we have enough data,
3633 * we can start a write request
3635 /* since handle_stripe can be called at any time we need to handle the
3636 * case where a compute block operation has been submitted and then a
3637 * subsequent call wants to start a write request. raid_run_ops only
3638 * handles the case where compute block and reconstruct are requested
3639 * simultaneously. If this is not the case then new writes need to be
3640 * held off until the compute completes.
3642 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3643 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3644 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3645 schedule_reconstruction(sh, s, rcw == 0, 0);
3648 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3649 struct stripe_head_state *s, int disks)
3651 struct r5dev *dev = NULL;
3653 BUG_ON(sh->batch_head);
3654 set_bit(STRIPE_HANDLE, &sh->state);
3656 switch (sh->check_state) {
3657 case check_state_idle:
3658 /* start a new check operation if there are no failures */
3659 if (s->failed == 0) {
3660 BUG_ON(s->uptodate != disks);
3661 sh->check_state = check_state_run;
3662 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3663 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3667 dev = &sh->dev[s->failed_num[0]];
3669 case check_state_compute_result:
3670 sh->check_state = check_state_idle;
3672 dev = &sh->dev[sh->pd_idx];
3674 /* check that a write has not made the stripe insync */
3675 if (test_bit(STRIPE_INSYNC, &sh->state))
3678 /* either failed parity check, or recovery is happening */
3679 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3680 BUG_ON(s->uptodate != disks);
3682 set_bit(R5_LOCKED, &dev->flags);
3684 set_bit(R5_Wantwrite, &dev->flags);
3686 clear_bit(STRIPE_DEGRADED, &sh->state);
3687 set_bit(STRIPE_INSYNC, &sh->state);
3689 case check_state_run:
3690 break; /* we will be called again upon completion */
3691 case check_state_check_result:
3692 sh->check_state = check_state_idle;
3694 /* if a failure occurred during the check operation, leave
3695 * STRIPE_INSYNC not set and let the stripe be handled again
3700 /* handle a successful check operation, if parity is correct
3701 * we are done. Otherwise update the mismatch count and repair
3702 * parity if !MD_RECOVERY_CHECK
3704 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3705 /* parity is correct (on disc,
3706 * not in buffer any more)
3708 set_bit(STRIPE_INSYNC, &sh->state);
3710 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3711 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3712 /* don't try to repair!! */
3713 set_bit(STRIPE_INSYNC, &sh->state);
3715 sh->check_state = check_state_compute_run;
3716 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3717 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3718 set_bit(R5_Wantcompute,
3719 &sh->dev[sh->pd_idx].flags);
3720 sh->ops.target = sh->pd_idx;
3721 sh->ops.target2 = -1;
3726 case check_state_compute_run:
3729 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3730 __func__, sh->check_state,
3731 (unsigned long long) sh->sector);
3736 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3737 struct stripe_head_state *s,
3740 int pd_idx = sh->pd_idx;
3741 int qd_idx = sh->qd_idx;
3744 BUG_ON(sh->batch_head);
3745 set_bit(STRIPE_HANDLE, &sh->state);
3747 BUG_ON(s->failed > 2);
3749 /* Want to check and possibly repair P and Q.
3750 * However there could be one 'failed' device, in which
3751 * case we can only check one of them, possibly using the
3752 * other to generate missing data
3755 switch (sh->check_state) {
3756 case check_state_idle:
3757 /* start a new check operation if there are < 2 failures */
3758 if (s->failed == s->q_failed) {
3759 /* The only possible failed device holds Q, so it
3760 * makes sense to check P (If anything else were failed,
3761 * we would have used P to recreate it).
3763 sh->check_state = check_state_run;
3765 if (!s->q_failed && s->failed < 2) {
3766 /* Q is not failed, and we didn't use it to generate
3767 * anything, so it makes sense to check it
3769 if (sh->check_state == check_state_run)
3770 sh->check_state = check_state_run_pq;
3772 sh->check_state = check_state_run_q;
3775 /* discard potentially stale zero_sum_result */
3776 sh->ops.zero_sum_result = 0;
3778 if (sh->check_state == check_state_run) {
3779 /* async_xor_zero_sum destroys the contents of P */
3780 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3783 if (sh->check_state >= check_state_run &&
3784 sh->check_state <= check_state_run_pq) {
3785 /* async_syndrome_zero_sum preserves P and Q, so
3786 * no need to mark them !uptodate here
3788 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3792 /* we have 2-disk failure */
3793 BUG_ON(s->failed != 2);
3795 case check_state_compute_result:
3796 sh->check_state = check_state_idle;
3798 /* check that a write has not made the stripe insync */
3799 if (test_bit(STRIPE_INSYNC, &sh->state))
3802 /* now write out any block on a failed drive,
3803 * or P or Q if they were recomputed
3805 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3806 if (s->failed == 2) {
3807 dev = &sh->dev[s->failed_num[1]];
3809 set_bit(R5_LOCKED, &dev->flags);
3810 set_bit(R5_Wantwrite, &dev->flags);
3812 if (s->failed >= 1) {
3813 dev = &sh->dev[s->failed_num[0]];
3815 set_bit(R5_LOCKED, &dev->flags);
3816 set_bit(R5_Wantwrite, &dev->flags);
3818 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3819 dev = &sh->dev[pd_idx];
3821 set_bit(R5_LOCKED, &dev->flags);
3822 set_bit(R5_Wantwrite, &dev->flags);
3824 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3825 dev = &sh->dev[qd_idx];
3827 set_bit(R5_LOCKED, &dev->flags);
3828 set_bit(R5_Wantwrite, &dev->flags);
3830 clear_bit(STRIPE_DEGRADED, &sh->state);
3832 set_bit(STRIPE_INSYNC, &sh->state);
3834 case check_state_run:
3835 case check_state_run_q:
3836 case check_state_run_pq:
3837 break; /* we will be called again upon completion */
3838 case check_state_check_result:
3839 sh->check_state = check_state_idle;
3841 /* handle a successful check operation, if parity is correct
3842 * we are done. Otherwise update the mismatch count and repair
3843 * parity if !MD_RECOVERY_CHECK
3845 if (sh->ops.zero_sum_result == 0) {
3846 /* both parities are correct */
3848 set_bit(STRIPE_INSYNC, &sh->state);
3850 /* in contrast to the raid5 case we can validate
3851 * parity, but still have a failure to write
3854 sh->check_state = check_state_compute_result;
3855 /* Returning at this point means that we may go
3856 * off and bring p and/or q uptodate again so
3857 * we make sure to check zero_sum_result again
3858 * to verify if p or q need writeback
3862 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3863 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3864 /* don't try to repair!! */
3865 set_bit(STRIPE_INSYNC, &sh->state);
3867 int *target = &sh->ops.target;
3869 sh->ops.target = -1;
3870 sh->ops.target2 = -1;
3871 sh->check_state = check_state_compute_run;
3872 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3873 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3874 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3875 set_bit(R5_Wantcompute,
3876 &sh->dev[pd_idx].flags);
3878 target = &sh->ops.target2;
3881 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3882 set_bit(R5_Wantcompute,
3883 &sh->dev[qd_idx].flags);
3890 case check_state_compute_run:
3893 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3894 __func__, sh->check_state,
3895 (unsigned long long) sh->sector);
3900 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3904 /* We have read all the blocks in this stripe and now we need to
3905 * copy some of them into a target stripe for expand.
3907 struct dma_async_tx_descriptor *tx = NULL;
3908 BUG_ON(sh->batch_head);
3909 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3910 for (i = 0; i < sh->disks; i++)
3911 if (i != sh->pd_idx && i != sh->qd_idx) {
3913 struct stripe_head *sh2;
3914 struct async_submit_ctl submit;
3916 sector_t bn = compute_blocknr(sh, i, 1);
3917 sector_t s = raid5_compute_sector(conf, bn, 0,
3919 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3921 /* so far only the early blocks of this stripe
3922 * have been requested. When later blocks
3923 * get requested, we will try again
3926 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3927 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3928 /* must have already done this block */
3929 release_stripe(sh2);
3933 /* place all the copies on one channel */
3934 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3935 tx = async_memcpy(sh2->dev[dd_idx].page,
3936 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3939 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3940 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3941 for (j = 0; j < conf->raid_disks; j++)
3942 if (j != sh2->pd_idx &&
3944 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3946 if (j == conf->raid_disks) {
3947 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3948 set_bit(STRIPE_HANDLE, &sh2->state);
3950 release_stripe(sh2);
3953 /* done submitting copies, wait for them to complete */
3954 async_tx_quiesce(&tx);
3958 * handle_stripe - do things to a stripe.
3960 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3961 * state of various bits to see what needs to be done.
3963 * return some read requests which now have data
3964 * return some write requests which are safely on storage
3965 * schedule a read on some buffers
3966 * schedule a write of some buffers
3967 * return confirmation of parity correctness
3971 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3973 struct r5conf *conf = sh->raid_conf;
3974 int disks = sh->disks;
3977 int do_recovery = 0;
3979 memset(s, 0, sizeof(*s));
3981 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
3982 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
3983 s->failed_num[0] = -1;
3984 s->failed_num[1] = -1;
3986 /* Now to look around and see what can be done */
3988 for (i=disks; i--; ) {
3989 struct md_rdev *rdev;
3996 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3998 dev->toread, dev->towrite, dev->written);
3999 /* maybe we can reply to a read
4001 * new wantfill requests are only permitted while
4002 * ops_complete_biofill is guaranteed to be inactive
4004 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4005 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4006 set_bit(R5_Wantfill, &dev->flags);
4008 /* now count some things */
4009 if (test_bit(R5_LOCKED, &dev->flags))
4011 if (test_bit(R5_UPTODATE, &dev->flags))
4013 if (test_bit(R5_Wantcompute, &dev->flags)) {
4015 BUG_ON(s->compute > 2);
4018 if (test_bit(R5_Wantfill, &dev->flags))
4020 else if (dev->toread)
4024 if (!test_bit(R5_OVERWRITE, &dev->flags))
4029 /* Prefer to use the replacement for reads, but only
4030 * if it is recovered enough and has no bad blocks.
4032 rdev = rcu_dereference(conf->disks[i].replacement);
4033 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4034 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4035 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4036 &first_bad, &bad_sectors))
4037 set_bit(R5_ReadRepl, &dev->flags);
4040 set_bit(R5_NeedReplace, &dev->flags);
4041 rdev = rcu_dereference(conf->disks[i].rdev);
4042 clear_bit(R5_ReadRepl, &dev->flags);
4044 if (rdev && test_bit(Faulty, &rdev->flags))
4047 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4048 &first_bad, &bad_sectors);
4049 if (s->blocked_rdev == NULL
4050 && (test_bit(Blocked, &rdev->flags)
4053 set_bit(BlockedBadBlocks,
4055 s->blocked_rdev = rdev;
4056 atomic_inc(&rdev->nr_pending);
4059 clear_bit(R5_Insync, &dev->flags);
4063 /* also not in-sync */
4064 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4065 test_bit(R5_UPTODATE, &dev->flags)) {
4066 /* treat as in-sync, but with a read error
4067 * which we can now try to correct
4069 set_bit(R5_Insync, &dev->flags);
4070 set_bit(R5_ReadError, &dev->flags);
4072 } else if (test_bit(In_sync, &rdev->flags))
4073 set_bit(R5_Insync, &dev->flags);
4074 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4075 /* in sync if before recovery_offset */
4076 set_bit(R5_Insync, &dev->flags);
4077 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4078 test_bit(R5_Expanded, &dev->flags))
4079 /* If we've reshaped into here, we assume it is Insync.
4080 * We will shortly update recovery_offset to make
4083 set_bit(R5_Insync, &dev->flags);
4085 if (test_bit(R5_WriteError, &dev->flags)) {
4086 /* This flag does not apply to '.replacement'
4087 * only to .rdev, so make sure to check that*/
4088 struct md_rdev *rdev2 = rcu_dereference(
4089 conf->disks[i].rdev);
4091 clear_bit(R5_Insync, &dev->flags);
4092 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4093 s->handle_bad_blocks = 1;
4094 atomic_inc(&rdev2->nr_pending);
4096 clear_bit(R5_WriteError, &dev->flags);
4098 if (test_bit(R5_MadeGood, &dev->flags)) {
4099 /* This flag does not apply to '.replacement'
4100 * only to .rdev, so make sure to check that*/
4101 struct md_rdev *rdev2 = rcu_dereference(
4102 conf->disks[i].rdev);
4103 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4104 s->handle_bad_blocks = 1;
4105 atomic_inc(&rdev2->nr_pending);
4107 clear_bit(R5_MadeGood, &dev->flags);
4109 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4110 struct md_rdev *rdev2 = rcu_dereference(
4111 conf->disks[i].replacement);
4112 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4113 s->handle_bad_blocks = 1;
4114 atomic_inc(&rdev2->nr_pending);
4116 clear_bit(R5_MadeGoodRepl, &dev->flags);
4118 if (!test_bit(R5_Insync, &dev->flags)) {
4119 /* The ReadError flag will just be confusing now */
4120 clear_bit(R5_ReadError, &dev->flags);
4121 clear_bit(R5_ReWrite, &dev->flags);
4123 if (test_bit(R5_ReadError, &dev->flags))
4124 clear_bit(R5_Insync, &dev->flags);
4125 if (!test_bit(R5_Insync, &dev->flags)) {
4127 s->failed_num[s->failed] = i;
4129 if (rdev && !test_bit(Faulty, &rdev->flags))
4133 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4134 /* If there is a failed device being replaced,
4135 * we must be recovering.
4136 * else if we are after recovery_cp, we must be syncing
4137 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4138 * else we can only be replacing
4139 * sync and recovery both need to read all devices, and so
4140 * use the same flag.
4143 sh->sector >= conf->mddev->recovery_cp ||
4144 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4152 static int clear_batch_ready(struct stripe_head *sh)
4154 struct stripe_head *tmp;
4155 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4157 spin_lock(&sh->stripe_lock);
4158 if (!sh->batch_head) {
4159 spin_unlock(&sh->stripe_lock);
4164 * this stripe could be added to a batch list before we check
4165 * BATCH_READY, skips it
4167 if (sh->batch_head != sh) {
4168 spin_unlock(&sh->stripe_lock);
4171 spin_lock(&sh->batch_lock);
4172 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4173 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4174 spin_unlock(&sh->batch_lock);
4175 spin_unlock(&sh->stripe_lock);
4178 * BATCH_READY is cleared, no new stripes can be added.
4179 * batch_list can be accessed without lock
4184 static void check_break_stripe_batch_list(struct stripe_head *sh)
4186 struct stripe_head *head_sh, *next;
4189 if (!test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4194 sh = list_first_entry(&sh->batch_list,
4195 struct stripe_head, batch_list);
4196 BUG_ON(sh == head_sh);
4197 } while (!test_bit(STRIPE_DEGRADED, &sh->state));
4199 while (sh != head_sh) {
4200 next = list_first_entry(&sh->batch_list,
4201 struct stripe_head, batch_list);
4202 list_del_init(&sh->batch_list);
4204 set_mask_bits(&sh->state, ~STRIPE_EXPAND_SYNC_FLAG,
4205 head_sh->state & ~((1 << STRIPE_ACTIVE) |
4206 (1 << STRIPE_PREREAD_ACTIVE) |
4207 (1 << STRIPE_DEGRADED) |
4208 STRIPE_EXPAND_SYNC_FLAG));
4209 sh->check_state = head_sh->check_state;
4210 sh->reconstruct_state = head_sh->reconstruct_state;
4211 for (i = 0; i < sh->disks; i++)
4212 sh->dev[i].flags = head_sh->dev[i].flags &
4213 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4215 spin_lock_irq(&sh->stripe_lock);
4216 sh->batch_head = NULL;
4217 spin_unlock_irq(&sh->stripe_lock);
4219 set_bit(STRIPE_HANDLE, &sh->state);
4226 static void handle_stripe(struct stripe_head *sh)
4228 struct stripe_head_state s;
4229 struct r5conf *conf = sh->raid_conf;
4232 int disks = sh->disks;
4233 struct r5dev *pdev, *qdev;
4235 clear_bit(STRIPE_HANDLE, &sh->state);
4236 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4237 /* already being handled, ensure it gets handled
4238 * again when current action finishes */
4239 set_bit(STRIPE_HANDLE, &sh->state);
4243 if (clear_batch_ready(sh) ) {
4244 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4248 check_break_stripe_batch_list(sh);
4250 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4251 spin_lock(&sh->stripe_lock);
4252 /* Cannot process 'sync' concurrently with 'discard' */
4253 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4254 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4255 set_bit(STRIPE_SYNCING, &sh->state);
4256 clear_bit(STRIPE_INSYNC, &sh->state);
4257 clear_bit(STRIPE_REPLACED, &sh->state);
4259 spin_unlock(&sh->stripe_lock);
4261 clear_bit(STRIPE_DELAYED, &sh->state);
4263 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4264 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4265 (unsigned long long)sh->sector, sh->state,
4266 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4267 sh->check_state, sh->reconstruct_state);
4269 analyse_stripe(sh, &s);
4271 if (s.handle_bad_blocks) {
4272 set_bit(STRIPE_HANDLE, &sh->state);
4276 if (unlikely(s.blocked_rdev)) {
4277 if (s.syncing || s.expanding || s.expanded ||
4278 s.replacing || s.to_write || s.written) {
4279 set_bit(STRIPE_HANDLE, &sh->state);
4282 /* There is nothing for the blocked_rdev to block */
4283 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4284 s.blocked_rdev = NULL;
4287 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4288 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4289 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4292 pr_debug("locked=%d uptodate=%d to_read=%d"
4293 " to_write=%d failed=%d failed_num=%d,%d\n",
4294 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4295 s.failed_num[0], s.failed_num[1]);
4296 /* check if the array has lost more than max_degraded devices and,
4297 * if so, some requests might need to be failed.
4299 if (s.failed > conf->max_degraded) {
4300 sh->check_state = 0;
4301 sh->reconstruct_state = 0;
4302 if (s.to_read+s.to_write+s.written)
4303 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4304 if (s.syncing + s.replacing)
4305 handle_failed_sync(conf, sh, &s);
4308 /* Now we check to see if any write operations have recently
4312 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4314 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4315 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4316 sh->reconstruct_state = reconstruct_state_idle;
4318 /* All the 'written' buffers and the parity block are ready to
4319 * be written back to disk
4321 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4322 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4323 BUG_ON(sh->qd_idx >= 0 &&
4324 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4325 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4326 for (i = disks; i--; ) {
4327 struct r5dev *dev = &sh->dev[i];
4328 if (test_bit(R5_LOCKED, &dev->flags) &&
4329 (i == sh->pd_idx || i == sh->qd_idx ||
4331 pr_debug("Writing block %d\n", i);
4332 set_bit(R5_Wantwrite, &dev->flags);
4337 if (!test_bit(R5_Insync, &dev->flags) ||
4338 ((i == sh->pd_idx || i == sh->qd_idx) &&
4340 set_bit(STRIPE_INSYNC, &sh->state);
4343 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4344 s.dec_preread_active = 1;
4348 * might be able to return some write requests if the parity blocks
4349 * are safe, or on a failed drive
4351 pdev = &sh->dev[sh->pd_idx];
4352 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4353 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4354 qdev = &sh->dev[sh->qd_idx];
4355 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4356 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4360 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4361 && !test_bit(R5_LOCKED, &pdev->flags)
4362 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4363 test_bit(R5_Discard, &pdev->flags))))) &&
4364 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4365 && !test_bit(R5_LOCKED, &qdev->flags)
4366 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4367 test_bit(R5_Discard, &qdev->flags))))))
4368 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4370 /* Now we might consider reading some blocks, either to check/generate
4371 * parity, or to satisfy requests
4372 * or to load a block that is being partially written.
4374 if (s.to_read || s.non_overwrite
4375 || (conf->level == 6 && s.to_write && s.failed)
4376 || (s.syncing && (s.uptodate + s.compute < disks))
4379 handle_stripe_fill(sh, &s, disks);
4381 /* Now to consider new write requests and what else, if anything
4382 * should be read. We do not handle new writes when:
4383 * 1/ A 'write' operation (copy+xor) is already in flight.
4384 * 2/ A 'check' operation is in flight, as it may clobber the parity
4387 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4388 handle_stripe_dirtying(conf, sh, &s, disks);
4390 /* maybe we need to check and possibly fix the parity for this stripe
4391 * Any reads will already have been scheduled, so we just see if enough
4392 * data is available. The parity check is held off while parity
4393 * dependent operations are in flight.
4395 if (sh->check_state ||
4396 (s.syncing && s.locked == 0 &&
4397 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4398 !test_bit(STRIPE_INSYNC, &sh->state))) {
4399 if (conf->level == 6)
4400 handle_parity_checks6(conf, sh, &s, disks);
4402 handle_parity_checks5(conf, sh, &s, disks);
4405 if ((s.replacing || s.syncing) && s.locked == 0
4406 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4407 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4408 /* Write out to replacement devices where possible */
4409 for (i = 0; i < conf->raid_disks; i++)
4410 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4411 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4412 set_bit(R5_WantReplace, &sh->dev[i].flags);
4413 set_bit(R5_LOCKED, &sh->dev[i].flags);
4417 set_bit(STRIPE_INSYNC, &sh->state);
4418 set_bit(STRIPE_REPLACED, &sh->state);
4420 if ((s.syncing || s.replacing) && s.locked == 0 &&
4421 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4422 test_bit(STRIPE_INSYNC, &sh->state)) {
4423 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4424 clear_bit(STRIPE_SYNCING, &sh->state);
4425 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4426 wake_up(&conf->wait_for_overlap);
4429 /* If the failed drives are just a ReadError, then we might need
4430 * to progress the repair/check process
4432 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4433 for (i = 0; i < s.failed; i++) {
4434 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4435 if (test_bit(R5_ReadError, &dev->flags)
4436 && !test_bit(R5_LOCKED, &dev->flags)
4437 && test_bit(R5_UPTODATE, &dev->flags)
4439 if (!test_bit(R5_ReWrite, &dev->flags)) {
4440 set_bit(R5_Wantwrite, &dev->flags);
4441 set_bit(R5_ReWrite, &dev->flags);
4442 set_bit(R5_LOCKED, &dev->flags);
4445 /* let's read it back */
4446 set_bit(R5_Wantread, &dev->flags);
4447 set_bit(R5_LOCKED, &dev->flags);
4453 /* Finish reconstruct operations initiated by the expansion process */
4454 if (sh->reconstruct_state == reconstruct_state_result) {
4455 struct stripe_head *sh_src
4456 = get_active_stripe(conf, sh->sector, 1, 1, 1);
4457 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4458 /* sh cannot be written until sh_src has been read.
4459 * so arrange for sh to be delayed a little
4461 set_bit(STRIPE_DELAYED, &sh->state);
4462 set_bit(STRIPE_HANDLE, &sh->state);
4463 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4465 atomic_inc(&conf->preread_active_stripes);
4466 release_stripe(sh_src);
4470 release_stripe(sh_src);
4472 sh->reconstruct_state = reconstruct_state_idle;
4473 clear_bit(STRIPE_EXPANDING, &sh->state);
4474 for (i = conf->raid_disks; i--; ) {
4475 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4476 set_bit(R5_LOCKED, &sh->dev[i].flags);
4481 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4482 !sh->reconstruct_state) {
4483 /* Need to write out all blocks after computing parity */
4484 sh->disks = conf->raid_disks;
4485 stripe_set_idx(sh->sector, conf, 0, sh);
4486 schedule_reconstruction(sh, &s, 1, 1);
4487 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4488 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4489 atomic_dec(&conf->reshape_stripes);
4490 wake_up(&conf->wait_for_overlap);
4491 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4494 if (s.expanding && s.locked == 0 &&
4495 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4496 handle_stripe_expansion(conf, sh);
4499 /* wait for this device to become unblocked */
4500 if (unlikely(s.blocked_rdev)) {
4501 if (conf->mddev->external)
4502 md_wait_for_blocked_rdev(s.blocked_rdev,
4505 /* Internal metadata will immediately
4506 * be written by raid5d, so we don't
4507 * need to wait here.
4509 rdev_dec_pending(s.blocked_rdev,
4513 if (s.handle_bad_blocks)
4514 for (i = disks; i--; ) {
4515 struct md_rdev *rdev;
4516 struct r5dev *dev = &sh->dev[i];
4517 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4518 /* We own a safe reference to the rdev */
4519 rdev = conf->disks[i].rdev;
4520 if (!rdev_set_badblocks(rdev, sh->sector,
4522 md_error(conf->mddev, rdev);
4523 rdev_dec_pending(rdev, conf->mddev);
4525 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4526 rdev = conf->disks[i].rdev;
4527 rdev_clear_badblocks(rdev, sh->sector,
4529 rdev_dec_pending(rdev, conf->mddev);
4531 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4532 rdev = conf->disks[i].replacement;
4534 /* rdev have been moved down */
4535 rdev = conf->disks[i].rdev;
4536 rdev_clear_badblocks(rdev, sh->sector,
4538 rdev_dec_pending(rdev, conf->mddev);
4543 raid_run_ops(sh, s.ops_request);
4547 if (s.dec_preread_active) {
4548 /* We delay this until after ops_run_io so that if make_request
4549 * is waiting on a flush, it won't continue until the writes
4550 * have actually been submitted.
4552 atomic_dec(&conf->preread_active_stripes);
4553 if (atomic_read(&conf->preread_active_stripes) <
4555 md_wakeup_thread(conf->mddev->thread);
4558 return_io(s.return_bi);
4560 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4563 static void raid5_activate_delayed(struct r5conf *conf)
4565 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4566 while (!list_empty(&conf->delayed_list)) {
4567 struct list_head *l = conf->delayed_list.next;
4568 struct stripe_head *sh;
4569 sh = list_entry(l, struct stripe_head, lru);
4571 clear_bit(STRIPE_DELAYED, &sh->state);
4572 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4573 atomic_inc(&conf->preread_active_stripes);
4574 list_add_tail(&sh->lru, &conf->hold_list);
4575 raid5_wakeup_stripe_thread(sh);
4580 static void activate_bit_delay(struct r5conf *conf,
4581 struct list_head *temp_inactive_list)
4583 /* device_lock is held */
4584 struct list_head head;
4585 list_add(&head, &conf->bitmap_list);
4586 list_del_init(&conf->bitmap_list);
4587 while (!list_empty(&head)) {
4588 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4590 list_del_init(&sh->lru);
4591 atomic_inc(&sh->count);
4592 hash = sh->hash_lock_index;
4593 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4597 static int raid5_congested(struct mddev *mddev, int bits)
4599 struct r5conf *conf = mddev->private;
4601 /* No difference between reads and writes. Just check
4602 * how busy the stripe_cache is
4605 if (conf->inactive_blocked)
4609 if (atomic_read(&conf->empty_inactive_list_nr))
4615 /* We want read requests to align with chunks where possible,
4616 * but write requests don't need to.
4618 static int raid5_mergeable_bvec(struct mddev *mddev,
4619 struct bvec_merge_data *bvm,
4620 struct bio_vec *biovec)
4622 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4624 unsigned int chunk_sectors = mddev->chunk_sectors;
4625 unsigned int bio_sectors = bvm->bi_size >> 9;
4627 if ((bvm->bi_rw & 1) == WRITE)
4628 return biovec->bv_len; /* always allow writes to be mergeable */
4630 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4631 chunk_sectors = mddev->new_chunk_sectors;
4632 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4633 if (max < 0) max = 0;
4634 if (max <= biovec->bv_len && bio_sectors == 0)
4635 return biovec->bv_len;
4640 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4642 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4643 unsigned int chunk_sectors = mddev->chunk_sectors;
4644 unsigned int bio_sectors = bio_sectors(bio);
4646 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4647 chunk_sectors = mddev->new_chunk_sectors;
4648 return chunk_sectors >=
4649 ((sector & (chunk_sectors - 1)) + bio_sectors);
4653 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4654 * later sampled by raid5d.
4656 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4658 unsigned long flags;
4660 spin_lock_irqsave(&conf->device_lock, flags);
4662 bi->bi_next = conf->retry_read_aligned_list;
4663 conf->retry_read_aligned_list = bi;
4665 spin_unlock_irqrestore(&conf->device_lock, flags);
4666 md_wakeup_thread(conf->mddev->thread);
4669 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4673 bi = conf->retry_read_aligned;
4675 conf->retry_read_aligned = NULL;
4678 bi = conf->retry_read_aligned_list;
4680 conf->retry_read_aligned_list = bi->bi_next;
4683 * this sets the active strip count to 1 and the processed
4684 * strip count to zero (upper 8 bits)
4686 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4693 * The "raid5_align_endio" should check if the read succeeded and if it
4694 * did, call bio_endio on the original bio (having bio_put the new bio
4696 * If the read failed..
4698 static void raid5_align_endio(struct bio *bi, int error)
4700 struct bio* raid_bi = bi->bi_private;
4701 struct mddev *mddev;
4702 struct r5conf *conf;
4703 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4704 struct md_rdev *rdev;
4708 rdev = (void*)raid_bi->bi_next;
4709 raid_bi->bi_next = NULL;
4710 mddev = rdev->mddev;
4711 conf = mddev->private;
4713 rdev_dec_pending(rdev, conf->mddev);
4715 if (!error && uptodate) {
4716 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4718 bio_endio(raid_bi, 0);
4719 if (atomic_dec_and_test(&conf->active_aligned_reads))
4720 wake_up(&conf->wait_for_stripe);
4724 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4726 add_bio_to_retry(raid_bi, conf);
4729 static int bio_fits_rdev(struct bio *bi)
4731 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4733 if (bio_sectors(bi) > queue_max_sectors(q))
4735 blk_recount_segments(q, bi);
4736 if (bi->bi_phys_segments > queue_max_segments(q))
4739 if (q->merge_bvec_fn)
4740 /* it's too hard to apply the merge_bvec_fn at this stage,
4748 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4750 struct r5conf *conf = mddev->private;
4752 struct bio* align_bi;
4753 struct md_rdev *rdev;
4754 sector_t end_sector;
4756 if (!in_chunk_boundary(mddev, raid_bio)) {
4757 pr_debug("chunk_aligned_read : non aligned\n");
4761 * use bio_clone_mddev to make a copy of the bio
4763 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4767 * set bi_end_io to a new function, and set bi_private to the
4770 align_bi->bi_end_io = raid5_align_endio;
4771 align_bi->bi_private = raid_bio;
4775 align_bi->bi_iter.bi_sector =
4776 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4779 end_sector = bio_end_sector(align_bi);
4781 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4782 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4783 rdev->recovery_offset < end_sector) {
4784 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4786 (test_bit(Faulty, &rdev->flags) ||
4787 !(test_bit(In_sync, &rdev->flags) ||
4788 rdev->recovery_offset >= end_sector)))
4795 atomic_inc(&rdev->nr_pending);
4797 raid_bio->bi_next = (void*)rdev;
4798 align_bi->bi_bdev = rdev->bdev;
4799 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
4801 if (!bio_fits_rdev(align_bi) ||
4802 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4803 bio_sectors(align_bi),
4804 &first_bad, &bad_sectors)) {
4805 /* too big in some way, or has a known bad block */
4807 rdev_dec_pending(rdev, mddev);
4811 /* No reshape active, so we can trust rdev->data_offset */
4812 align_bi->bi_iter.bi_sector += rdev->data_offset;
4814 spin_lock_irq(&conf->device_lock);
4815 wait_event_lock_irq(conf->wait_for_stripe,
4818 atomic_inc(&conf->active_aligned_reads);
4819 spin_unlock_irq(&conf->device_lock);
4822 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4823 align_bi, disk_devt(mddev->gendisk),
4824 raid_bio->bi_iter.bi_sector);
4825 generic_make_request(align_bi);
4834 /* __get_priority_stripe - get the next stripe to process
4836 * Full stripe writes are allowed to pass preread active stripes up until
4837 * the bypass_threshold is exceeded. In general the bypass_count
4838 * increments when the handle_list is handled before the hold_list; however, it
4839 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4840 * stripe with in flight i/o. The bypass_count will be reset when the
4841 * head of the hold_list has changed, i.e. the head was promoted to the
4844 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4846 struct stripe_head *sh = NULL, *tmp;
4847 struct list_head *handle_list = NULL;
4848 struct r5worker_group *wg = NULL;
4850 if (conf->worker_cnt_per_group == 0) {
4851 handle_list = &conf->handle_list;
4852 } else if (group != ANY_GROUP) {
4853 handle_list = &conf->worker_groups[group].handle_list;
4854 wg = &conf->worker_groups[group];
4857 for (i = 0; i < conf->group_cnt; i++) {
4858 handle_list = &conf->worker_groups[i].handle_list;
4859 wg = &conf->worker_groups[i];
4860 if (!list_empty(handle_list))
4865 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4867 list_empty(handle_list) ? "empty" : "busy",
4868 list_empty(&conf->hold_list) ? "empty" : "busy",
4869 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4871 if (!list_empty(handle_list)) {
4872 sh = list_entry(handle_list->next, typeof(*sh), lru);
4874 if (list_empty(&conf->hold_list))
4875 conf->bypass_count = 0;
4876 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4877 if (conf->hold_list.next == conf->last_hold)
4878 conf->bypass_count++;
4880 conf->last_hold = conf->hold_list.next;
4881 conf->bypass_count -= conf->bypass_threshold;
4882 if (conf->bypass_count < 0)
4883 conf->bypass_count = 0;
4886 } else if (!list_empty(&conf->hold_list) &&
4887 ((conf->bypass_threshold &&
4888 conf->bypass_count > conf->bypass_threshold) ||
4889 atomic_read(&conf->pending_full_writes) == 0)) {
4891 list_for_each_entry(tmp, &conf->hold_list, lru) {
4892 if (conf->worker_cnt_per_group == 0 ||
4893 group == ANY_GROUP ||
4894 !cpu_online(tmp->cpu) ||
4895 cpu_to_group(tmp->cpu) == group) {
4902 conf->bypass_count -= conf->bypass_threshold;
4903 if (conf->bypass_count < 0)
4904 conf->bypass_count = 0;
4916 list_del_init(&sh->lru);
4917 BUG_ON(atomic_inc_return(&sh->count) != 1);
4921 struct raid5_plug_cb {
4922 struct blk_plug_cb cb;
4923 struct list_head list;
4924 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4927 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4929 struct raid5_plug_cb *cb = container_of(
4930 blk_cb, struct raid5_plug_cb, cb);
4931 struct stripe_head *sh;
4932 struct mddev *mddev = cb->cb.data;
4933 struct r5conf *conf = mddev->private;
4937 if (cb->list.next && !list_empty(&cb->list)) {
4938 spin_lock_irq(&conf->device_lock);
4939 while (!list_empty(&cb->list)) {
4940 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4941 list_del_init(&sh->lru);
4943 * avoid race release_stripe_plug() sees
4944 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4945 * is still in our list
4947 smp_mb__before_atomic();
4948 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4950 * STRIPE_ON_RELEASE_LIST could be set here. In that
4951 * case, the count is always > 1 here
4953 hash = sh->hash_lock_index;
4954 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
4957 spin_unlock_irq(&conf->device_lock);
4959 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4960 NR_STRIPE_HASH_LOCKS);
4962 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4966 static void release_stripe_plug(struct mddev *mddev,
4967 struct stripe_head *sh)
4969 struct blk_plug_cb *blk_cb = blk_check_plugged(
4970 raid5_unplug, mddev,
4971 sizeof(struct raid5_plug_cb));
4972 struct raid5_plug_cb *cb;
4979 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4981 if (cb->list.next == NULL) {
4983 INIT_LIST_HEAD(&cb->list);
4984 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
4985 INIT_LIST_HEAD(cb->temp_inactive_list + i);
4988 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4989 list_add_tail(&sh->lru, &cb->list);
4994 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4996 struct r5conf *conf = mddev->private;
4997 sector_t logical_sector, last_sector;
4998 struct stripe_head *sh;
5002 if (mddev->reshape_position != MaxSector)
5003 /* Skip discard while reshape is happening */
5006 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5007 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5010 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5012 stripe_sectors = conf->chunk_sectors *
5013 (conf->raid_disks - conf->max_degraded);
5014 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5016 sector_div(last_sector, stripe_sectors);
5018 logical_sector *= conf->chunk_sectors;
5019 last_sector *= conf->chunk_sectors;
5021 for (; logical_sector < last_sector;
5022 logical_sector += STRIPE_SECTORS) {
5026 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
5027 prepare_to_wait(&conf->wait_for_overlap, &w,
5028 TASK_UNINTERRUPTIBLE);
5029 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5030 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5035 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5036 spin_lock_irq(&sh->stripe_lock);
5037 for (d = 0; d < conf->raid_disks; d++) {
5038 if (d == sh->pd_idx || d == sh->qd_idx)
5040 if (sh->dev[d].towrite || sh->dev[d].toread) {
5041 set_bit(R5_Overlap, &sh->dev[d].flags);
5042 spin_unlock_irq(&sh->stripe_lock);
5048 set_bit(STRIPE_DISCARD, &sh->state);
5049 finish_wait(&conf->wait_for_overlap, &w);
5050 sh->overwrite_disks = 0;
5051 for (d = 0; d < conf->raid_disks; d++) {
5052 if (d == sh->pd_idx || d == sh->qd_idx)
5054 sh->dev[d].towrite = bi;
5055 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5056 raid5_inc_bi_active_stripes(bi);
5057 sh->overwrite_disks++;
5059 spin_unlock_irq(&sh->stripe_lock);
5060 if (conf->mddev->bitmap) {
5062 d < conf->raid_disks - conf->max_degraded;
5064 bitmap_startwrite(mddev->bitmap,
5068 sh->bm_seq = conf->seq_flush + 1;
5069 set_bit(STRIPE_BIT_DELAY, &sh->state);
5072 set_bit(STRIPE_HANDLE, &sh->state);
5073 clear_bit(STRIPE_DELAYED, &sh->state);
5074 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5075 atomic_inc(&conf->preread_active_stripes);
5076 release_stripe_plug(mddev, sh);
5079 remaining = raid5_dec_bi_active_stripes(bi);
5080 if (remaining == 0) {
5081 md_write_end(mddev);
5086 static void make_request(struct mddev *mddev, struct bio * bi)
5088 struct r5conf *conf = mddev->private;
5090 sector_t new_sector;
5091 sector_t logical_sector, last_sector;
5092 struct stripe_head *sh;
5093 const int rw = bio_data_dir(bi);
5098 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5099 md_flush_request(mddev, bi);
5103 md_write_start(mddev, bi);
5106 mddev->reshape_position == MaxSector &&
5107 chunk_aligned_read(mddev,bi))
5110 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5111 make_discard_request(mddev, bi);
5115 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5116 last_sector = bio_end_sector(bi);
5118 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5120 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5121 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5127 seq = read_seqcount_begin(&conf->gen_lock);
5130 prepare_to_wait(&conf->wait_for_overlap, &w,
5131 TASK_UNINTERRUPTIBLE);
5132 if (unlikely(conf->reshape_progress != MaxSector)) {
5133 /* spinlock is needed as reshape_progress may be
5134 * 64bit on a 32bit platform, and so it might be
5135 * possible to see a half-updated value
5136 * Of course reshape_progress could change after
5137 * the lock is dropped, so once we get a reference
5138 * to the stripe that we think it is, we will have
5141 spin_lock_irq(&conf->device_lock);
5142 if (mddev->reshape_backwards
5143 ? logical_sector < conf->reshape_progress
5144 : logical_sector >= conf->reshape_progress) {
5147 if (mddev->reshape_backwards
5148 ? logical_sector < conf->reshape_safe
5149 : logical_sector >= conf->reshape_safe) {
5150 spin_unlock_irq(&conf->device_lock);
5156 spin_unlock_irq(&conf->device_lock);
5159 new_sector = raid5_compute_sector(conf, logical_sector,
5162 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5163 (unsigned long long)new_sector,
5164 (unsigned long long)logical_sector);
5166 sh = get_active_stripe(conf, new_sector, previous,
5167 (bi->bi_rw&RWA_MASK), 0);
5169 if (unlikely(previous)) {
5170 /* expansion might have moved on while waiting for a
5171 * stripe, so we must do the range check again.
5172 * Expansion could still move past after this
5173 * test, but as we are holding a reference to
5174 * 'sh', we know that if that happens,
5175 * STRIPE_EXPANDING will get set and the expansion
5176 * won't proceed until we finish with the stripe.
5179 spin_lock_irq(&conf->device_lock);
5180 if (mddev->reshape_backwards
5181 ? logical_sector >= conf->reshape_progress
5182 : logical_sector < conf->reshape_progress)
5183 /* mismatch, need to try again */
5185 spin_unlock_irq(&conf->device_lock);
5193 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5194 /* Might have got the wrong stripe_head
5202 logical_sector >= mddev->suspend_lo &&
5203 logical_sector < mddev->suspend_hi) {
5205 /* As the suspend_* range is controlled by
5206 * userspace, we want an interruptible
5209 flush_signals(current);
5210 prepare_to_wait(&conf->wait_for_overlap,
5211 &w, TASK_INTERRUPTIBLE);
5212 if (logical_sector >= mddev->suspend_lo &&
5213 logical_sector < mddev->suspend_hi) {
5220 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5221 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5222 /* Stripe is busy expanding or
5223 * add failed due to overlap. Flush everything
5226 md_wakeup_thread(mddev->thread);
5232 set_bit(STRIPE_HANDLE, &sh->state);
5233 clear_bit(STRIPE_DELAYED, &sh->state);
5234 if ((!sh->batch_head || sh == sh->batch_head) &&
5235 (bi->bi_rw & REQ_SYNC) &&
5236 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5237 atomic_inc(&conf->preread_active_stripes);
5238 release_stripe_plug(mddev, sh);
5240 /* cannot get stripe for read-ahead, just give-up */
5241 clear_bit(BIO_UPTODATE, &bi->bi_flags);
5245 finish_wait(&conf->wait_for_overlap, &w);
5247 remaining = raid5_dec_bi_active_stripes(bi);
5248 if (remaining == 0) {
5251 md_write_end(mddev);
5253 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5259 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5261 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5263 /* reshaping is quite different to recovery/resync so it is
5264 * handled quite separately ... here.
5266 * On each call to sync_request, we gather one chunk worth of
5267 * destination stripes and flag them as expanding.
5268 * Then we find all the source stripes and request reads.
5269 * As the reads complete, handle_stripe will copy the data
5270 * into the destination stripe and release that stripe.
5272 struct r5conf *conf = mddev->private;
5273 struct stripe_head *sh;
5274 sector_t first_sector, last_sector;
5275 int raid_disks = conf->previous_raid_disks;
5276 int data_disks = raid_disks - conf->max_degraded;
5277 int new_data_disks = conf->raid_disks - conf->max_degraded;
5280 sector_t writepos, readpos, safepos;
5281 sector_t stripe_addr;
5282 int reshape_sectors;
5283 struct list_head stripes;
5285 if (sector_nr == 0) {
5286 /* If restarting in the middle, skip the initial sectors */
5287 if (mddev->reshape_backwards &&
5288 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5289 sector_nr = raid5_size(mddev, 0, 0)
5290 - conf->reshape_progress;
5291 } else if (!mddev->reshape_backwards &&
5292 conf->reshape_progress > 0)
5293 sector_nr = conf->reshape_progress;
5294 sector_div(sector_nr, new_data_disks);
5296 mddev->curr_resync_completed = sector_nr;
5297 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5303 /* We need to process a full chunk at a time.
5304 * If old and new chunk sizes differ, we need to process the
5307 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
5308 reshape_sectors = mddev->new_chunk_sectors;
5310 reshape_sectors = mddev->chunk_sectors;
5312 /* We update the metadata at least every 10 seconds, or when
5313 * the data about to be copied would over-write the source of
5314 * the data at the front of the range. i.e. one new_stripe
5315 * along from reshape_progress new_maps to after where
5316 * reshape_safe old_maps to
5318 writepos = conf->reshape_progress;
5319 sector_div(writepos, new_data_disks);
5320 readpos = conf->reshape_progress;
5321 sector_div(readpos, data_disks);
5322 safepos = conf->reshape_safe;
5323 sector_div(safepos, data_disks);
5324 if (mddev->reshape_backwards) {
5325 writepos -= min_t(sector_t, reshape_sectors, writepos);
5326 readpos += reshape_sectors;
5327 safepos += reshape_sectors;
5329 writepos += reshape_sectors;
5330 readpos -= min_t(sector_t, reshape_sectors, readpos);
5331 safepos -= min_t(sector_t, reshape_sectors, safepos);
5334 /* Having calculated the 'writepos' possibly use it
5335 * to set 'stripe_addr' which is where we will write to.
5337 if (mddev->reshape_backwards) {
5338 BUG_ON(conf->reshape_progress == 0);
5339 stripe_addr = writepos;
5340 BUG_ON((mddev->dev_sectors &
5341 ~((sector_t)reshape_sectors - 1))
5342 - reshape_sectors - stripe_addr
5345 BUG_ON(writepos != sector_nr + reshape_sectors);
5346 stripe_addr = sector_nr;
5349 /* 'writepos' is the most advanced device address we might write.
5350 * 'readpos' is the least advanced device address we might read.
5351 * 'safepos' is the least address recorded in the metadata as having
5353 * If there is a min_offset_diff, these are adjusted either by
5354 * increasing the safepos/readpos if diff is negative, or
5355 * increasing writepos if diff is positive.
5356 * If 'readpos' is then behind 'writepos', there is no way that we can
5357 * ensure safety in the face of a crash - that must be done by userspace
5358 * making a backup of the data. So in that case there is no particular
5359 * rush to update metadata.
5360 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5361 * update the metadata to advance 'safepos' to match 'readpos' so that
5362 * we can be safe in the event of a crash.
5363 * So we insist on updating metadata if safepos is behind writepos and
5364 * readpos is beyond writepos.
5365 * In any case, update the metadata every 10 seconds.
5366 * Maybe that number should be configurable, but I'm not sure it is
5367 * worth it.... maybe it could be a multiple of safemode_delay???
5369 if (conf->min_offset_diff < 0) {
5370 safepos += -conf->min_offset_diff;
5371 readpos += -conf->min_offset_diff;
5373 writepos += conf->min_offset_diff;
5375 if ((mddev->reshape_backwards
5376 ? (safepos > writepos && readpos < writepos)
5377 : (safepos < writepos && readpos > writepos)) ||
5378 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5379 /* Cannot proceed until we've updated the superblock... */
5380 wait_event(conf->wait_for_overlap,
5381 atomic_read(&conf->reshape_stripes)==0
5382 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5383 if (atomic_read(&conf->reshape_stripes) != 0)
5385 mddev->reshape_position = conf->reshape_progress;
5386 mddev->curr_resync_completed = sector_nr;
5387 conf->reshape_checkpoint = jiffies;
5388 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5389 md_wakeup_thread(mddev->thread);
5390 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5391 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5392 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5394 spin_lock_irq(&conf->device_lock);
5395 conf->reshape_safe = mddev->reshape_position;
5396 spin_unlock_irq(&conf->device_lock);
5397 wake_up(&conf->wait_for_overlap);
5398 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5401 INIT_LIST_HEAD(&stripes);
5402 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5404 int skipped_disk = 0;
5405 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5406 set_bit(STRIPE_EXPANDING, &sh->state);
5407 atomic_inc(&conf->reshape_stripes);
5408 /* If any of this stripe is beyond the end of the old
5409 * array, then we need to zero those blocks
5411 for (j=sh->disks; j--;) {
5413 if (j == sh->pd_idx)
5415 if (conf->level == 6 &&
5418 s = compute_blocknr(sh, j, 0);
5419 if (s < raid5_size(mddev, 0, 0)) {
5423 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5424 set_bit(R5_Expanded, &sh->dev[j].flags);
5425 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5427 if (!skipped_disk) {
5428 set_bit(STRIPE_EXPAND_READY, &sh->state);
5429 set_bit(STRIPE_HANDLE, &sh->state);
5431 list_add(&sh->lru, &stripes);
5433 spin_lock_irq(&conf->device_lock);
5434 if (mddev->reshape_backwards)
5435 conf->reshape_progress -= reshape_sectors * new_data_disks;
5437 conf->reshape_progress += reshape_sectors * new_data_disks;
5438 spin_unlock_irq(&conf->device_lock);
5439 /* Ok, those stripe are ready. We can start scheduling
5440 * reads on the source stripes.
5441 * The source stripes are determined by mapping the first and last
5442 * block on the destination stripes.
5445 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5448 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5449 * new_data_disks - 1),
5451 if (last_sector >= mddev->dev_sectors)
5452 last_sector = mddev->dev_sectors - 1;
5453 while (first_sector <= last_sector) {
5454 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
5455 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5456 set_bit(STRIPE_HANDLE, &sh->state);
5458 first_sector += STRIPE_SECTORS;
5460 /* Now that the sources are clearly marked, we can release
5461 * the destination stripes
5463 while (!list_empty(&stripes)) {
5464 sh = list_entry(stripes.next, struct stripe_head, lru);
5465 list_del_init(&sh->lru);
5468 /* If this takes us to the resync_max point where we have to pause,
5469 * then we need to write out the superblock.
5471 sector_nr += reshape_sectors;
5472 if ((sector_nr - mddev->curr_resync_completed) * 2
5473 >= mddev->resync_max - mddev->curr_resync_completed) {
5474 /* Cannot proceed until we've updated the superblock... */
5475 wait_event(conf->wait_for_overlap,
5476 atomic_read(&conf->reshape_stripes) == 0
5477 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5478 if (atomic_read(&conf->reshape_stripes) != 0)
5480 mddev->reshape_position = conf->reshape_progress;
5481 mddev->curr_resync_completed = sector_nr;
5482 conf->reshape_checkpoint = jiffies;
5483 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5484 md_wakeup_thread(mddev->thread);
5485 wait_event(mddev->sb_wait,
5486 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5487 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5488 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5490 spin_lock_irq(&conf->device_lock);
5491 conf->reshape_safe = mddev->reshape_position;
5492 spin_unlock_irq(&conf->device_lock);
5493 wake_up(&conf->wait_for_overlap);
5494 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5497 return reshape_sectors;
5500 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5502 struct r5conf *conf = mddev->private;
5503 struct stripe_head *sh;
5504 sector_t max_sector = mddev->dev_sectors;
5505 sector_t sync_blocks;
5506 int still_degraded = 0;
5509 if (sector_nr >= max_sector) {
5510 /* just being told to finish up .. nothing much to do */
5512 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5517 if (mddev->curr_resync < max_sector) /* aborted */
5518 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5520 else /* completed sync */
5522 bitmap_close_sync(mddev->bitmap);
5527 /* Allow raid5_quiesce to complete */
5528 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5530 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5531 return reshape_request(mddev, sector_nr, skipped);
5533 /* No need to check resync_max as we never do more than one
5534 * stripe, and as resync_max will always be on a chunk boundary,
5535 * if the check in md_do_sync didn't fire, there is no chance
5536 * of overstepping resync_max here
5539 /* if there is too many failed drives and we are trying
5540 * to resync, then assert that we are finished, because there is
5541 * nothing we can do.
5543 if (mddev->degraded >= conf->max_degraded &&
5544 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5545 sector_t rv = mddev->dev_sectors - sector_nr;
5549 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5551 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5552 sync_blocks >= STRIPE_SECTORS) {
5553 /* we can skip this block, and probably more */
5554 sync_blocks /= STRIPE_SECTORS;
5556 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5559 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5561 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5563 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5564 /* make sure we don't swamp the stripe cache if someone else
5565 * is trying to get access
5567 schedule_timeout_uninterruptible(1);
5569 /* Need to check if array will still be degraded after recovery/resync
5570 * Note in case of > 1 drive failures it's possible we're rebuilding
5571 * one drive while leaving another faulty drive in array.
5574 for (i = 0; i < conf->raid_disks; i++) {
5575 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5577 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5582 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5584 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5585 set_bit(STRIPE_HANDLE, &sh->state);
5589 return STRIPE_SECTORS;
5592 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5594 /* We may not be able to submit a whole bio at once as there
5595 * may not be enough stripe_heads available.
5596 * We cannot pre-allocate enough stripe_heads as we may need
5597 * more than exist in the cache (if we allow ever large chunks).
5598 * So we do one stripe head at a time and record in
5599 * ->bi_hw_segments how many have been done.
5601 * We *know* that this entire raid_bio is in one chunk, so
5602 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5604 struct stripe_head *sh;
5606 sector_t sector, logical_sector, last_sector;
5611 logical_sector = raid_bio->bi_iter.bi_sector &
5612 ~((sector_t)STRIPE_SECTORS-1);
5613 sector = raid5_compute_sector(conf, logical_sector,
5615 last_sector = bio_end_sector(raid_bio);
5617 for (; logical_sector < last_sector;
5618 logical_sector += STRIPE_SECTORS,
5619 sector += STRIPE_SECTORS,
5622 if (scnt < raid5_bi_processed_stripes(raid_bio))
5623 /* already done this stripe */
5626 sh = get_active_stripe(conf, sector, 0, 1, 1);
5629 /* failed to get a stripe - must wait */
5630 raid5_set_bi_processed_stripes(raid_bio, scnt);
5631 conf->retry_read_aligned = raid_bio;
5635 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5637 raid5_set_bi_processed_stripes(raid_bio, scnt);
5638 conf->retry_read_aligned = raid_bio;
5642 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5647 remaining = raid5_dec_bi_active_stripes(raid_bio);
5648 if (remaining == 0) {
5649 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5651 bio_endio(raid_bio, 0);
5653 if (atomic_dec_and_test(&conf->active_aligned_reads))
5654 wake_up(&conf->wait_for_stripe);
5658 static int handle_active_stripes(struct r5conf *conf, int group,
5659 struct r5worker *worker,
5660 struct list_head *temp_inactive_list)
5662 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5663 int i, batch_size = 0, hash;
5664 bool release_inactive = false;
5666 while (batch_size < MAX_STRIPE_BATCH &&
5667 (sh = __get_priority_stripe(conf, group)) != NULL)
5668 batch[batch_size++] = sh;
5670 if (batch_size == 0) {
5671 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5672 if (!list_empty(temp_inactive_list + i))
5674 if (i == NR_STRIPE_HASH_LOCKS)
5676 release_inactive = true;
5678 spin_unlock_irq(&conf->device_lock);
5680 release_inactive_stripe_list(conf, temp_inactive_list,
5681 NR_STRIPE_HASH_LOCKS);
5683 if (release_inactive) {
5684 spin_lock_irq(&conf->device_lock);
5688 for (i = 0; i < batch_size; i++)
5689 handle_stripe(batch[i]);
5693 spin_lock_irq(&conf->device_lock);
5694 for (i = 0; i < batch_size; i++) {
5695 hash = batch[i]->hash_lock_index;
5696 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5701 static void raid5_do_work(struct work_struct *work)
5703 struct r5worker *worker = container_of(work, struct r5worker, work);
5704 struct r5worker_group *group = worker->group;
5705 struct r5conf *conf = group->conf;
5706 int group_id = group - conf->worker_groups;
5708 struct blk_plug plug;
5710 pr_debug("+++ raid5worker active\n");
5712 blk_start_plug(&plug);
5714 spin_lock_irq(&conf->device_lock);
5716 int batch_size, released;
5718 released = release_stripe_list(conf, worker->temp_inactive_list);
5720 batch_size = handle_active_stripes(conf, group_id, worker,
5721 worker->temp_inactive_list);
5722 worker->working = false;
5723 if (!batch_size && !released)
5725 handled += batch_size;
5727 pr_debug("%d stripes handled\n", handled);
5729 spin_unlock_irq(&conf->device_lock);
5730 blk_finish_plug(&plug);
5732 pr_debug("--- raid5worker inactive\n");
5736 * This is our raid5 kernel thread.
5738 * We scan the hash table for stripes which can be handled now.
5739 * During the scan, completed stripes are saved for us by the interrupt
5740 * handler, so that they will not have to wait for our next wakeup.
5742 static void raid5d(struct md_thread *thread)
5744 struct mddev *mddev = thread->mddev;
5745 struct r5conf *conf = mddev->private;
5747 struct blk_plug plug;
5749 pr_debug("+++ raid5d active\n");
5751 md_check_recovery(mddev);
5753 blk_start_plug(&plug);
5755 spin_lock_irq(&conf->device_lock);
5758 int batch_size, released;
5760 released = release_stripe_list(conf, conf->temp_inactive_list);
5763 !list_empty(&conf->bitmap_list)) {
5764 /* Now is a good time to flush some bitmap updates */
5766 spin_unlock_irq(&conf->device_lock);
5767 bitmap_unplug(mddev->bitmap);
5768 spin_lock_irq(&conf->device_lock);
5769 conf->seq_write = conf->seq_flush;
5770 activate_bit_delay(conf, conf->temp_inactive_list);
5772 raid5_activate_delayed(conf);
5774 while ((bio = remove_bio_from_retry(conf))) {
5776 spin_unlock_irq(&conf->device_lock);
5777 ok = retry_aligned_read(conf, bio);
5778 spin_lock_irq(&conf->device_lock);
5784 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5785 conf->temp_inactive_list);
5786 if (!batch_size && !released)
5788 handled += batch_size;
5790 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5791 spin_unlock_irq(&conf->device_lock);
5792 md_check_recovery(mddev);
5793 spin_lock_irq(&conf->device_lock);
5796 pr_debug("%d stripes handled\n", handled);
5798 spin_unlock_irq(&conf->device_lock);
5800 async_tx_issue_pending_all();
5801 blk_finish_plug(&plug);
5803 pr_debug("--- raid5d inactive\n");
5807 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5809 struct r5conf *conf;
5811 spin_lock(&mddev->lock);
5812 conf = mddev->private;
5814 ret = sprintf(page, "%d\n", conf->max_nr_stripes);
5815 spin_unlock(&mddev->lock);
5820 raid5_set_cache_size(struct mddev *mddev, int size)
5822 struct r5conf *conf = mddev->private;
5825 if (size <= 16 || size > 32768)
5828 while (size < conf->max_nr_stripes &&
5829 drop_one_stripe(conf))
5832 err = md_allow_write(mddev);
5836 while (size > conf->max_nr_stripes)
5837 if (!grow_one_stripe(conf, GFP_KERNEL))
5842 EXPORT_SYMBOL(raid5_set_cache_size);
5845 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5847 struct r5conf *conf;
5851 if (len >= PAGE_SIZE)
5853 if (kstrtoul(page, 10, &new))
5855 err = mddev_lock(mddev);
5858 conf = mddev->private;
5862 err = raid5_set_cache_size(mddev, new);
5863 mddev_unlock(mddev);
5868 static struct md_sysfs_entry
5869 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5870 raid5_show_stripe_cache_size,
5871 raid5_store_stripe_cache_size);
5874 raid5_show_rmw_level(struct mddev *mddev, char *page)
5876 struct r5conf *conf = mddev->private;
5878 return sprintf(page, "%d\n", conf->rmw_level);
5884 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
5886 struct r5conf *conf = mddev->private;
5892 if (len >= PAGE_SIZE)
5895 if (kstrtoul(page, 10, &new))
5898 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
5901 if (new != PARITY_DISABLE_RMW &&
5902 new != PARITY_ENABLE_RMW &&
5903 new != PARITY_PREFER_RMW)
5906 conf->rmw_level = new;
5910 static struct md_sysfs_entry
5911 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
5912 raid5_show_rmw_level,
5913 raid5_store_rmw_level);
5917 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5919 struct r5conf *conf;
5921 spin_lock(&mddev->lock);
5922 conf = mddev->private;
5924 ret = sprintf(page, "%d\n", conf->bypass_threshold);
5925 spin_unlock(&mddev->lock);
5930 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
5932 struct r5conf *conf;
5936 if (len >= PAGE_SIZE)
5938 if (kstrtoul(page, 10, &new))
5941 err = mddev_lock(mddev);
5944 conf = mddev->private;
5947 else if (new > conf->max_nr_stripes)
5950 conf->bypass_threshold = new;
5951 mddev_unlock(mddev);
5955 static struct md_sysfs_entry
5956 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5958 raid5_show_preread_threshold,
5959 raid5_store_preread_threshold);
5962 raid5_show_skip_copy(struct mddev *mddev, char *page)
5964 struct r5conf *conf;
5966 spin_lock(&mddev->lock);
5967 conf = mddev->private;
5969 ret = sprintf(page, "%d\n", conf->skip_copy);
5970 spin_unlock(&mddev->lock);
5975 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
5977 struct r5conf *conf;
5981 if (len >= PAGE_SIZE)
5983 if (kstrtoul(page, 10, &new))
5987 err = mddev_lock(mddev);
5990 conf = mddev->private;
5993 else if (new != conf->skip_copy) {
5994 mddev_suspend(mddev);
5995 conf->skip_copy = new;
5997 mddev->queue->backing_dev_info.capabilities |=
5998 BDI_CAP_STABLE_WRITES;
6000 mddev->queue->backing_dev_info.capabilities &=
6001 ~BDI_CAP_STABLE_WRITES;
6002 mddev_resume(mddev);
6004 mddev_unlock(mddev);
6008 static struct md_sysfs_entry
6009 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6010 raid5_show_skip_copy,
6011 raid5_store_skip_copy);
6014 stripe_cache_active_show(struct mddev *mddev, char *page)
6016 struct r5conf *conf = mddev->private;
6018 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6023 static struct md_sysfs_entry
6024 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6027 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6029 struct r5conf *conf;
6031 spin_lock(&mddev->lock);
6032 conf = mddev->private;
6034 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6035 spin_unlock(&mddev->lock);
6039 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6041 int *worker_cnt_per_group,
6042 struct r5worker_group **worker_groups);
6044 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6046 struct r5conf *conf;
6049 struct r5worker_group *new_groups, *old_groups;
6050 int group_cnt, worker_cnt_per_group;
6052 if (len >= PAGE_SIZE)
6054 if (kstrtoul(page, 10, &new))
6057 err = mddev_lock(mddev);
6060 conf = mddev->private;
6063 else if (new != conf->worker_cnt_per_group) {
6064 mddev_suspend(mddev);
6066 old_groups = conf->worker_groups;
6068 flush_workqueue(raid5_wq);
6070 err = alloc_thread_groups(conf, new,
6071 &group_cnt, &worker_cnt_per_group,
6074 spin_lock_irq(&conf->device_lock);
6075 conf->group_cnt = group_cnt;
6076 conf->worker_cnt_per_group = worker_cnt_per_group;
6077 conf->worker_groups = new_groups;
6078 spin_unlock_irq(&conf->device_lock);
6081 kfree(old_groups[0].workers);
6084 mddev_resume(mddev);
6086 mddev_unlock(mddev);
6091 static struct md_sysfs_entry
6092 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6093 raid5_show_group_thread_cnt,
6094 raid5_store_group_thread_cnt);
6096 static struct attribute *raid5_attrs[] = {
6097 &raid5_stripecache_size.attr,
6098 &raid5_stripecache_active.attr,
6099 &raid5_preread_bypass_threshold.attr,
6100 &raid5_group_thread_cnt.attr,
6101 &raid5_skip_copy.attr,
6102 &raid5_rmw_level.attr,
6105 static struct attribute_group raid5_attrs_group = {
6107 .attrs = raid5_attrs,
6110 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6112 int *worker_cnt_per_group,
6113 struct r5worker_group **worker_groups)
6117 struct r5worker *workers;
6119 *worker_cnt_per_group = cnt;
6122 *worker_groups = NULL;
6125 *group_cnt = num_possible_nodes();
6126 size = sizeof(struct r5worker) * cnt;
6127 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6128 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6129 *group_cnt, GFP_NOIO);
6130 if (!*worker_groups || !workers) {
6132 kfree(*worker_groups);
6136 for (i = 0; i < *group_cnt; i++) {
6137 struct r5worker_group *group;
6139 group = &(*worker_groups)[i];
6140 INIT_LIST_HEAD(&group->handle_list);
6142 group->workers = workers + i * cnt;
6144 for (j = 0; j < cnt; j++) {
6145 struct r5worker *worker = group->workers + j;
6146 worker->group = group;
6147 INIT_WORK(&worker->work, raid5_do_work);
6149 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6150 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6157 static void free_thread_groups(struct r5conf *conf)
6159 if (conf->worker_groups)
6160 kfree(conf->worker_groups[0].workers);
6161 kfree(conf->worker_groups);
6162 conf->worker_groups = NULL;
6166 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6168 struct r5conf *conf = mddev->private;
6171 sectors = mddev->dev_sectors;
6173 /* size is defined by the smallest of previous and new size */
6174 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6176 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6177 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
6178 return sectors * (raid_disks - conf->max_degraded);
6181 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6183 safe_put_page(percpu->spare_page);
6184 if (percpu->scribble)
6185 flex_array_free(percpu->scribble);
6186 percpu->spare_page = NULL;
6187 percpu->scribble = NULL;
6190 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6192 if (conf->level == 6 && !percpu->spare_page)
6193 percpu->spare_page = alloc_page(GFP_KERNEL);
6194 if (!percpu->scribble)
6195 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6196 conf->previous_raid_disks), conf->chunk_sectors /
6197 STRIPE_SECTORS, GFP_KERNEL);
6199 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6200 free_scratch_buffer(conf, percpu);
6207 static void raid5_free_percpu(struct r5conf *conf)
6214 #ifdef CONFIG_HOTPLUG_CPU
6215 unregister_cpu_notifier(&conf->cpu_notify);
6219 for_each_possible_cpu(cpu)
6220 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6223 free_percpu(conf->percpu);
6226 static void free_conf(struct r5conf *conf)
6228 free_thread_groups(conf);
6229 shrink_stripes(conf);
6230 raid5_free_percpu(conf);
6232 kfree(conf->stripe_hashtbl);
6236 #ifdef CONFIG_HOTPLUG_CPU
6237 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6240 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6241 long cpu = (long)hcpu;
6242 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6245 case CPU_UP_PREPARE:
6246 case CPU_UP_PREPARE_FROZEN:
6247 if (alloc_scratch_buffer(conf, percpu)) {
6248 pr_err("%s: failed memory allocation for cpu%ld\n",
6250 return notifier_from_errno(-ENOMEM);
6254 case CPU_DEAD_FROZEN:
6255 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6264 static int raid5_alloc_percpu(struct r5conf *conf)
6269 conf->percpu = alloc_percpu(struct raid5_percpu);
6273 #ifdef CONFIG_HOTPLUG_CPU
6274 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6275 conf->cpu_notify.priority = 0;
6276 err = register_cpu_notifier(&conf->cpu_notify);
6282 for_each_present_cpu(cpu) {
6283 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6285 pr_err("%s: failed memory allocation for cpu%ld\n",
6295 static struct r5conf *setup_conf(struct mddev *mddev)
6297 struct r5conf *conf;
6298 int raid_disk, memory, max_disks;
6299 struct md_rdev *rdev;
6300 struct disk_info *disk;
6303 int group_cnt, worker_cnt_per_group;
6304 struct r5worker_group *new_group;
6306 if (mddev->new_level != 5
6307 && mddev->new_level != 4
6308 && mddev->new_level != 6) {
6309 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6310 mdname(mddev), mddev->new_level);
6311 return ERR_PTR(-EIO);
6313 if ((mddev->new_level == 5
6314 && !algorithm_valid_raid5(mddev->new_layout)) ||
6315 (mddev->new_level == 6
6316 && !algorithm_valid_raid6(mddev->new_layout))) {
6317 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6318 mdname(mddev), mddev->new_layout);
6319 return ERR_PTR(-EIO);
6321 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6322 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6323 mdname(mddev), mddev->raid_disks);
6324 return ERR_PTR(-EINVAL);
6327 if (!mddev->new_chunk_sectors ||
6328 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6329 !is_power_of_2(mddev->new_chunk_sectors)) {
6330 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6331 mdname(mddev), mddev->new_chunk_sectors << 9);
6332 return ERR_PTR(-EINVAL);
6335 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6338 /* Don't enable multi-threading by default*/
6339 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6341 conf->group_cnt = group_cnt;
6342 conf->worker_cnt_per_group = worker_cnt_per_group;
6343 conf->worker_groups = new_group;
6346 spin_lock_init(&conf->device_lock);
6347 seqcount_init(&conf->gen_lock);
6348 init_waitqueue_head(&conf->wait_for_stripe);
6349 init_waitqueue_head(&conf->wait_for_overlap);
6350 INIT_LIST_HEAD(&conf->handle_list);
6351 INIT_LIST_HEAD(&conf->hold_list);
6352 INIT_LIST_HEAD(&conf->delayed_list);
6353 INIT_LIST_HEAD(&conf->bitmap_list);
6354 init_llist_head(&conf->released_stripes);
6355 atomic_set(&conf->active_stripes, 0);
6356 atomic_set(&conf->preread_active_stripes, 0);
6357 atomic_set(&conf->active_aligned_reads, 0);
6358 conf->bypass_threshold = BYPASS_THRESHOLD;
6359 conf->recovery_disabled = mddev->recovery_disabled - 1;
6361 conf->raid_disks = mddev->raid_disks;
6362 if (mddev->reshape_position == MaxSector)
6363 conf->previous_raid_disks = mddev->raid_disks;
6365 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6366 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6368 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6373 conf->mddev = mddev;
6375 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6378 /* We init hash_locks[0] separately to that it can be used
6379 * as the reference lock in the spin_lock_nest_lock() call
6380 * in lock_all_device_hash_locks_irq in order to convince
6381 * lockdep that we know what we are doing.
6383 spin_lock_init(conf->hash_locks);
6384 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6385 spin_lock_init(conf->hash_locks + i);
6387 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6388 INIT_LIST_HEAD(conf->inactive_list + i);
6390 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6391 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6393 conf->level = mddev->new_level;
6394 conf->chunk_sectors = mddev->new_chunk_sectors;
6395 if (raid5_alloc_percpu(conf) != 0)
6398 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6400 rdev_for_each(rdev, mddev) {
6401 raid_disk = rdev->raid_disk;
6402 if (raid_disk >= max_disks
6405 disk = conf->disks + raid_disk;
6407 if (test_bit(Replacement, &rdev->flags)) {
6408 if (disk->replacement)
6410 disk->replacement = rdev;
6417 if (test_bit(In_sync, &rdev->flags)) {
6418 char b[BDEVNAME_SIZE];
6419 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6421 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6422 } else if (rdev->saved_raid_disk != raid_disk)
6423 /* Cannot rely on bitmap to complete recovery */
6427 conf->level = mddev->new_level;
6428 if (conf->level == 6) {
6429 conf->max_degraded = 2;
6430 if (raid6_call.xor_syndrome)
6431 conf->rmw_level = PARITY_ENABLE_RMW;
6433 conf->rmw_level = PARITY_DISABLE_RMW;
6435 conf->max_degraded = 1;
6436 conf->rmw_level = PARITY_ENABLE_RMW;
6438 conf->algorithm = mddev->new_layout;
6439 conf->reshape_progress = mddev->reshape_position;
6440 if (conf->reshape_progress != MaxSector) {
6441 conf->prev_chunk_sectors = mddev->chunk_sectors;
6442 conf->prev_algo = mddev->layout;
6445 memory = NR_STRIPES * (sizeof(struct stripe_head) +
6446 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6447 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6448 if (grow_stripes(conf, NR_STRIPES)) {
6450 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6451 mdname(mddev), memory);
6454 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6455 mdname(mddev), memory);
6457 sprintf(pers_name, "raid%d", mddev->new_level);
6458 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6459 if (!conf->thread) {
6461 "md/raid:%s: couldn't allocate thread.\n",
6471 return ERR_PTR(-EIO);
6473 return ERR_PTR(-ENOMEM);
6476 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6479 case ALGORITHM_PARITY_0:
6480 if (raid_disk < max_degraded)
6483 case ALGORITHM_PARITY_N:
6484 if (raid_disk >= raid_disks - max_degraded)
6487 case ALGORITHM_PARITY_0_6:
6488 if (raid_disk == 0 ||
6489 raid_disk == raid_disks - 1)
6492 case ALGORITHM_LEFT_ASYMMETRIC_6:
6493 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6494 case ALGORITHM_LEFT_SYMMETRIC_6:
6495 case ALGORITHM_RIGHT_SYMMETRIC_6:
6496 if (raid_disk == raid_disks - 1)
6502 static int run(struct mddev *mddev)
6504 struct r5conf *conf;
6505 int working_disks = 0;
6506 int dirty_parity_disks = 0;
6507 struct md_rdev *rdev;
6508 sector_t reshape_offset = 0;
6510 long long min_offset_diff = 0;
6513 if (mddev->recovery_cp != MaxSector)
6514 printk(KERN_NOTICE "md/raid:%s: not clean"
6515 " -- starting background reconstruction\n",
6518 rdev_for_each(rdev, mddev) {
6520 if (rdev->raid_disk < 0)
6522 diff = (rdev->new_data_offset - rdev->data_offset);
6524 min_offset_diff = diff;
6526 } else if (mddev->reshape_backwards &&
6527 diff < min_offset_diff)
6528 min_offset_diff = diff;
6529 else if (!mddev->reshape_backwards &&
6530 diff > min_offset_diff)
6531 min_offset_diff = diff;
6534 if (mddev->reshape_position != MaxSector) {
6535 /* Check that we can continue the reshape.
6536 * Difficulties arise if the stripe we would write to
6537 * next is at or after the stripe we would read from next.
6538 * For a reshape that changes the number of devices, this
6539 * is only possible for a very short time, and mdadm makes
6540 * sure that time appears to have past before assembling
6541 * the array. So we fail if that time hasn't passed.
6542 * For a reshape that keeps the number of devices the same
6543 * mdadm must be monitoring the reshape can keeping the
6544 * critical areas read-only and backed up. It will start
6545 * the array in read-only mode, so we check for that.
6547 sector_t here_new, here_old;
6549 int max_degraded = (mddev->level == 6 ? 2 : 1);
6551 if (mddev->new_level != mddev->level) {
6552 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6553 "required - aborting.\n",
6557 old_disks = mddev->raid_disks - mddev->delta_disks;
6558 /* reshape_position must be on a new-stripe boundary, and one
6559 * further up in new geometry must map after here in old
6562 here_new = mddev->reshape_position;
6563 if (sector_div(here_new, mddev->new_chunk_sectors *
6564 (mddev->raid_disks - max_degraded))) {
6565 printk(KERN_ERR "md/raid:%s: reshape_position not "
6566 "on a stripe boundary\n", mdname(mddev));
6569 reshape_offset = here_new * mddev->new_chunk_sectors;
6570 /* here_new is the stripe we will write to */
6571 here_old = mddev->reshape_position;
6572 sector_div(here_old, mddev->chunk_sectors *
6573 (old_disks-max_degraded));
6574 /* here_old is the first stripe that we might need to read
6576 if (mddev->delta_disks == 0) {
6577 if ((here_new * mddev->new_chunk_sectors !=
6578 here_old * mddev->chunk_sectors)) {
6579 printk(KERN_ERR "md/raid:%s: reshape position is"
6580 " confused - aborting\n", mdname(mddev));
6583 /* We cannot be sure it is safe to start an in-place
6584 * reshape. It is only safe if user-space is monitoring
6585 * and taking constant backups.
6586 * mdadm always starts a situation like this in
6587 * readonly mode so it can take control before
6588 * allowing any writes. So just check for that.
6590 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6591 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6592 /* not really in-place - so OK */;
6593 else if (mddev->ro == 0) {
6594 printk(KERN_ERR "md/raid:%s: in-place reshape "
6595 "must be started in read-only mode "
6600 } else if (mddev->reshape_backwards
6601 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6602 here_old * mddev->chunk_sectors)
6603 : (here_new * mddev->new_chunk_sectors >=
6604 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6605 /* Reading from the same stripe as writing to - bad */
6606 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6607 "auto-recovery - aborting.\n",
6611 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6613 /* OK, we should be able to continue; */
6615 BUG_ON(mddev->level != mddev->new_level);
6616 BUG_ON(mddev->layout != mddev->new_layout);
6617 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6618 BUG_ON(mddev->delta_disks != 0);
6621 if (mddev->private == NULL)
6622 conf = setup_conf(mddev);
6624 conf = mddev->private;
6627 return PTR_ERR(conf);
6629 conf->min_offset_diff = min_offset_diff;
6630 mddev->thread = conf->thread;
6631 conf->thread = NULL;
6632 mddev->private = conf;
6634 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6636 rdev = conf->disks[i].rdev;
6637 if (!rdev && conf->disks[i].replacement) {
6638 /* The replacement is all we have yet */
6639 rdev = conf->disks[i].replacement;
6640 conf->disks[i].replacement = NULL;
6641 clear_bit(Replacement, &rdev->flags);
6642 conf->disks[i].rdev = rdev;
6646 if (conf->disks[i].replacement &&
6647 conf->reshape_progress != MaxSector) {
6648 /* replacements and reshape simply do not mix. */
6649 printk(KERN_ERR "md: cannot handle concurrent "
6650 "replacement and reshape.\n");
6653 if (test_bit(In_sync, &rdev->flags)) {
6657 /* This disc is not fully in-sync. However if it
6658 * just stored parity (beyond the recovery_offset),
6659 * when we don't need to be concerned about the
6660 * array being dirty.
6661 * When reshape goes 'backwards', we never have
6662 * partially completed devices, so we only need
6663 * to worry about reshape going forwards.
6665 /* Hack because v0.91 doesn't store recovery_offset properly. */
6666 if (mddev->major_version == 0 &&
6667 mddev->minor_version > 90)
6668 rdev->recovery_offset = reshape_offset;
6670 if (rdev->recovery_offset < reshape_offset) {
6671 /* We need to check old and new layout */
6672 if (!only_parity(rdev->raid_disk,
6675 conf->max_degraded))
6678 if (!only_parity(rdev->raid_disk,
6680 conf->previous_raid_disks,
6681 conf->max_degraded))
6683 dirty_parity_disks++;
6687 * 0 for a fully functional array, 1 or 2 for a degraded array.
6689 mddev->degraded = calc_degraded(conf);
6691 if (has_failed(conf)) {
6692 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6693 " (%d/%d failed)\n",
6694 mdname(mddev), mddev->degraded, conf->raid_disks);
6698 /* device size must be a multiple of chunk size */
6699 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6700 mddev->resync_max_sectors = mddev->dev_sectors;
6702 if (mddev->degraded > dirty_parity_disks &&
6703 mddev->recovery_cp != MaxSector) {
6704 if (mddev->ok_start_degraded)
6706 "md/raid:%s: starting dirty degraded array"
6707 " - data corruption possible.\n",
6711 "md/raid:%s: cannot start dirty degraded array.\n",
6717 if (mddev->degraded == 0)
6718 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6719 " devices, algorithm %d\n", mdname(mddev), conf->level,
6720 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6723 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6724 " out of %d devices, algorithm %d\n",
6725 mdname(mddev), conf->level,
6726 mddev->raid_disks - mddev->degraded,
6727 mddev->raid_disks, mddev->new_layout);
6729 print_raid5_conf(conf);
6731 if (conf->reshape_progress != MaxSector) {
6732 conf->reshape_safe = conf->reshape_progress;
6733 atomic_set(&conf->reshape_stripes, 0);
6734 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6735 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6736 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6737 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6738 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6742 /* Ok, everything is just fine now */
6743 if (mddev->to_remove == &raid5_attrs_group)
6744 mddev->to_remove = NULL;
6745 else if (mddev->kobj.sd &&
6746 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6748 "raid5: failed to create sysfs attributes for %s\n",
6750 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6754 bool discard_supported = true;
6755 /* read-ahead size must cover two whole stripes, which
6756 * is 2 * (datadisks) * chunksize where 'n' is the
6757 * number of raid devices
6759 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6760 int stripe = data_disks *
6761 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6762 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6763 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6765 chunk_size = mddev->chunk_sectors << 9;
6766 blk_queue_io_min(mddev->queue, chunk_size);
6767 blk_queue_io_opt(mddev->queue, chunk_size *
6768 (conf->raid_disks - conf->max_degraded));
6769 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6771 * We can only discard a whole stripe. It doesn't make sense to
6772 * discard data disk but write parity disk
6774 stripe = stripe * PAGE_SIZE;
6775 /* Round up to power of 2, as discard handling
6776 * currently assumes that */
6777 while ((stripe-1) & stripe)
6778 stripe = (stripe | (stripe-1)) + 1;
6779 mddev->queue->limits.discard_alignment = stripe;
6780 mddev->queue->limits.discard_granularity = stripe;
6782 * unaligned part of discard request will be ignored, so can't
6783 * guarantee discard_zeroes_data
6785 mddev->queue->limits.discard_zeroes_data = 0;
6787 blk_queue_max_write_same_sectors(mddev->queue, 0);
6789 rdev_for_each(rdev, mddev) {
6790 disk_stack_limits(mddev->gendisk, rdev->bdev,
6791 rdev->data_offset << 9);
6792 disk_stack_limits(mddev->gendisk, rdev->bdev,
6793 rdev->new_data_offset << 9);
6795 * discard_zeroes_data is required, otherwise data
6796 * could be lost. Consider a scenario: discard a stripe
6797 * (the stripe could be inconsistent if
6798 * discard_zeroes_data is 0); write one disk of the
6799 * stripe (the stripe could be inconsistent again
6800 * depending on which disks are used to calculate
6801 * parity); the disk is broken; The stripe data of this
6804 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6805 !bdev_get_queue(rdev->bdev)->
6806 limits.discard_zeroes_data)
6807 discard_supported = false;
6808 /* Unfortunately, discard_zeroes_data is not currently
6809 * a guarantee - just a hint. So we only allow DISCARD
6810 * if the sysadmin has confirmed that only safe devices
6811 * are in use by setting a module parameter.
6813 if (!devices_handle_discard_safely) {
6814 if (discard_supported) {
6815 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6816 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6818 discard_supported = false;
6822 if (discard_supported &&
6823 mddev->queue->limits.max_discard_sectors >= stripe &&
6824 mddev->queue->limits.discard_granularity >= stripe)
6825 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6828 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6834 md_unregister_thread(&mddev->thread);
6835 print_raid5_conf(conf);
6837 mddev->private = NULL;
6838 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6842 static void raid5_free(struct mddev *mddev, void *priv)
6844 struct r5conf *conf = priv;
6847 mddev->to_remove = &raid5_attrs_group;
6850 static void status(struct seq_file *seq, struct mddev *mddev)
6852 struct r5conf *conf = mddev->private;
6855 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6856 mddev->chunk_sectors / 2, mddev->layout);
6857 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6858 for (i = 0; i < conf->raid_disks; i++)
6859 seq_printf (seq, "%s",
6860 conf->disks[i].rdev &&
6861 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6862 seq_printf (seq, "]");
6865 static void print_raid5_conf (struct r5conf *conf)
6868 struct disk_info *tmp;
6870 printk(KERN_DEBUG "RAID conf printout:\n");
6872 printk("(conf==NULL)\n");
6875 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6877 conf->raid_disks - conf->mddev->degraded);
6879 for (i = 0; i < conf->raid_disks; i++) {
6880 char b[BDEVNAME_SIZE];
6881 tmp = conf->disks + i;
6883 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6884 i, !test_bit(Faulty, &tmp->rdev->flags),
6885 bdevname(tmp->rdev->bdev, b));
6889 static int raid5_spare_active(struct mddev *mddev)
6892 struct r5conf *conf = mddev->private;
6893 struct disk_info *tmp;
6895 unsigned long flags;
6897 for (i = 0; i < conf->raid_disks; i++) {
6898 tmp = conf->disks + i;
6899 if (tmp->replacement
6900 && tmp->replacement->recovery_offset == MaxSector
6901 && !test_bit(Faulty, &tmp->replacement->flags)
6902 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
6903 /* Replacement has just become active. */
6905 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
6908 /* Replaced device not technically faulty,
6909 * but we need to be sure it gets removed
6910 * and never re-added.
6912 set_bit(Faulty, &tmp->rdev->flags);
6913 sysfs_notify_dirent_safe(
6914 tmp->rdev->sysfs_state);
6916 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
6917 } else if (tmp->rdev
6918 && tmp->rdev->recovery_offset == MaxSector
6919 && !test_bit(Faulty, &tmp->rdev->flags)
6920 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
6922 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
6925 spin_lock_irqsave(&conf->device_lock, flags);
6926 mddev->degraded = calc_degraded(conf);
6927 spin_unlock_irqrestore(&conf->device_lock, flags);
6928 print_raid5_conf(conf);
6932 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
6934 struct r5conf *conf = mddev->private;
6936 int number = rdev->raid_disk;
6937 struct md_rdev **rdevp;
6938 struct disk_info *p = conf->disks + number;
6940 print_raid5_conf(conf);
6941 if (rdev == p->rdev)
6943 else if (rdev == p->replacement)
6944 rdevp = &p->replacement;
6948 if (number >= conf->raid_disks &&
6949 conf->reshape_progress == MaxSector)
6950 clear_bit(In_sync, &rdev->flags);
6952 if (test_bit(In_sync, &rdev->flags) ||
6953 atomic_read(&rdev->nr_pending)) {
6957 /* Only remove non-faulty devices if recovery
6960 if (!test_bit(Faulty, &rdev->flags) &&
6961 mddev->recovery_disabled != conf->recovery_disabled &&
6962 !has_failed(conf) &&
6963 (!p->replacement || p->replacement == rdev) &&
6964 number < conf->raid_disks) {
6970 if (atomic_read(&rdev->nr_pending)) {
6971 /* lost the race, try later */
6974 } else if (p->replacement) {
6975 /* We must have just cleared 'rdev' */
6976 p->rdev = p->replacement;
6977 clear_bit(Replacement, &p->replacement->flags);
6978 smp_mb(); /* Make sure other CPUs may see both as identical
6979 * but will never see neither - if they are careful
6981 p->replacement = NULL;
6982 clear_bit(WantReplacement, &rdev->flags);
6984 /* We might have just removed the Replacement as faulty-
6985 * clear the bit just in case
6987 clear_bit(WantReplacement, &rdev->flags);
6990 print_raid5_conf(conf);
6994 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
6996 struct r5conf *conf = mddev->private;
6999 struct disk_info *p;
7001 int last = conf->raid_disks - 1;
7003 if (mddev->recovery_disabled == conf->recovery_disabled)
7006 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7007 /* no point adding a device */
7010 if (rdev->raid_disk >= 0)
7011 first = last = rdev->raid_disk;
7014 * find the disk ... but prefer rdev->saved_raid_disk
7017 if (rdev->saved_raid_disk >= 0 &&
7018 rdev->saved_raid_disk >= first &&
7019 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7020 first = rdev->saved_raid_disk;
7022 for (disk = first; disk <= last; disk++) {
7023 p = conf->disks + disk;
7024 if (p->rdev == NULL) {
7025 clear_bit(In_sync, &rdev->flags);
7026 rdev->raid_disk = disk;
7028 if (rdev->saved_raid_disk != disk)
7030 rcu_assign_pointer(p->rdev, rdev);
7034 for (disk = first; disk <= last; disk++) {
7035 p = conf->disks + disk;
7036 if (test_bit(WantReplacement, &p->rdev->flags) &&
7037 p->replacement == NULL) {
7038 clear_bit(In_sync, &rdev->flags);
7039 set_bit(Replacement, &rdev->flags);
7040 rdev->raid_disk = disk;
7043 rcu_assign_pointer(p->replacement, rdev);
7048 print_raid5_conf(conf);
7052 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7054 /* no resync is happening, and there is enough space
7055 * on all devices, so we can resize.
7056 * We need to make sure resync covers any new space.
7057 * If the array is shrinking we should possibly wait until
7058 * any io in the removed space completes, but it hardly seems
7062 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7063 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7064 if (mddev->external_size &&
7065 mddev->array_sectors > newsize)
7067 if (mddev->bitmap) {
7068 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7072 md_set_array_sectors(mddev, newsize);
7073 set_capacity(mddev->gendisk, mddev->array_sectors);
7074 revalidate_disk(mddev->gendisk);
7075 if (sectors > mddev->dev_sectors &&
7076 mddev->recovery_cp > mddev->dev_sectors) {
7077 mddev->recovery_cp = mddev->dev_sectors;
7078 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7080 mddev->dev_sectors = sectors;
7081 mddev->resync_max_sectors = sectors;
7085 static int check_stripe_cache(struct mddev *mddev)
7087 /* Can only proceed if there are plenty of stripe_heads.
7088 * We need a minimum of one full stripe,, and for sensible progress
7089 * it is best to have about 4 times that.
7090 * If we require 4 times, then the default 256 4K stripe_heads will
7091 * allow for chunk sizes up to 256K, which is probably OK.
7092 * If the chunk size is greater, user-space should request more
7093 * stripe_heads first.
7095 struct r5conf *conf = mddev->private;
7096 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7097 > conf->max_nr_stripes ||
7098 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7099 > conf->max_nr_stripes) {
7100 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7102 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7109 static int check_reshape(struct mddev *mddev)
7111 struct r5conf *conf = mddev->private;
7113 if (mddev->delta_disks == 0 &&
7114 mddev->new_layout == mddev->layout &&
7115 mddev->new_chunk_sectors == mddev->chunk_sectors)
7116 return 0; /* nothing to do */
7117 if (has_failed(conf))
7119 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7120 /* We might be able to shrink, but the devices must
7121 * be made bigger first.
7122 * For raid6, 4 is the minimum size.
7123 * Otherwise 2 is the minimum
7126 if (mddev->level == 6)
7128 if (mddev->raid_disks + mddev->delta_disks < min)
7132 if (!check_stripe_cache(mddev))
7135 return resize_stripes(conf, (conf->previous_raid_disks
7136 + mddev->delta_disks));
7139 static int raid5_start_reshape(struct mddev *mddev)
7141 struct r5conf *conf = mddev->private;
7142 struct md_rdev *rdev;
7144 unsigned long flags;
7146 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7149 if (!check_stripe_cache(mddev))
7152 if (has_failed(conf))
7155 rdev_for_each(rdev, mddev) {
7156 if (!test_bit(In_sync, &rdev->flags)
7157 && !test_bit(Faulty, &rdev->flags))
7161 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7162 /* Not enough devices even to make a degraded array
7167 /* Refuse to reduce size of the array. Any reductions in
7168 * array size must be through explicit setting of array_size
7171 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7172 < mddev->array_sectors) {
7173 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7174 "before number of disks\n", mdname(mddev));
7178 atomic_set(&conf->reshape_stripes, 0);
7179 spin_lock_irq(&conf->device_lock);
7180 write_seqcount_begin(&conf->gen_lock);
7181 conf->previous_raid_disks = conf->raid_disks;
7182 conf->raid_disks += mddev->delta_disks;
7183 conf->prev_chunk_sectors = conf->chunk_sectors;
7184 conf->chunk_sectors = mddev->new_chunk_sectors;
7185 conf->prev_algo = conf->algorithm;
7186 conf->algorithm = mddev->new_layout;
7188 /* Code that selects data_offset needs to see the generation update
7189 * if reshape_progress has been set - so a memory barrier needed.
7192 if (mddev->reshape_backwards)
7193 conf->reshape_progress = raid5_size(mddev, 0, 0);
7195 conf->reshape_progress = 0;
7196 conf->reshape_safe = conf->reshape_progress;
7197 write_seqcount_end(&conf->gen_lock);
7198 spin_unlock_irq(&conf->device_lock);
7200 /* Now make sure any requests that proceeded on the assumption
7201 * the reshape wasn't running - like Discard or Read - have
7204 mddev_suspend(mddev);
7205 mddev_resume(mddev);
7207 /* Add some new drives, as many as will fit.
7208 * We know there are enough to make the newly sized array work.
7209 * Don't add devices if we are reducing the number of
7210 * devices in the array. This is because it is not possible
7211 * to correctly record the "partially reconstructed" state of
7212 * such devices during the reshape and confusion could result.
7214 if (mddev->delta_disks >= 0) {
7215 rdev_for_each(rdev, mddev)
7216 if (rdev->raid_disk < 0 &&
7217 !test_bit(Faulty, &rdev->flags)) {
7218 if (raid5_add_disk(mddev, rdev) == 0) {
7220 >= conf->previous_raid_disks)
7221 set_bit(In_sync, &rdev->flags);
7223 rdev->recovery_offset = 0;
7225 if (sysfs_link_rdev(mddev, rdev))
7226 /* Failure here is OK */;
7228 } else if (rdev->raid_disk >= conf->previous_raid_disks
7229 && !test_bit(Faulty, &rdev->flags)) {
7230 /* This is a spare that was manually added */
7231 set_bit(In_sync, &rdev->flags);
7234 /* When a reshape changes the number of devices,
7235 * ->degraded is measured against the larger of the
7236 * pre and post number of devices.
7238 spin_lock_irqsave(&conf->device_lock, flags);
7239 mddev->degraded = calc_degraded(conf);
7240 spin_unlock_irqrestore(&conf->device_lock, flags);
7242 mddev->raid_disks = conf->raid_disks;
7243 mddev->reshape_position = conf->reshape_progress;
7244 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7246 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7247 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7248 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7249 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7250 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7252 if (!mddev->sync_thread) {
7253 mddev->recovery = 0;
7254 spin_lock_irq(&conf->device_lock);
7255 write_seqcount_begin(&conf->gen_lock);
7256 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7257 mddev->new_chunk_sectors =
7258 conf->chunk_sectors = conf->prev_chunk_sectors;
7259 mddev->new_layout = conf->algorithm = conf->prev_algo;
7260 rdev_for_each(rdev, mddev)
7261 rdev->new_data_offset = rdev->data_offset;
7263 conf->generation --;
7264 conf->reshape_progress = MaxSector;
7265 mddev->reshape_position = MaxSector;
7266 write_seqcount_end(&conf->gen_lock);
7267 spin_unlock_irq(&conf->device_lock);
7270 conf->reshape_checkpoint = jiffies;
7271 md_wakeup_thread(mddev->sync_thread);
7272 md_new_event(mddev);
7276 /* This is called from the reshape thread and should make any
7277 * changes needed in 'conf'
7279 static void end_reshape(struct r5conf *conf)
7282 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7283 struct md_rdev *rdev;
7285 spin_lock_irq(&conf->device_lock);
7286 conf->previous_raid_disks = conf->raid_disks;
7287 rdev_for_each(rdev, conf->mddev)
7288 rdev->data_offset = rdev->new_data_offset;
7290 conf->reshape_progress = MaxSector;
7291 spin_unlock_irq(&conf->device_lock);
7292 wake_up(&conf->wait_for_overlap);
7294 /* read-ahead size must cover two whole stripes, which is
7295 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7297 if (conf->mddev->queue) {
7298 int data_disks = conf->raid_disks - conf->max_degraded;
7299 int stripe = data_disks * ((conf->chunk_sectors << 9)
7301 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7302 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7307 /* This is called from the raid5d thread with mddev_lock held.
7308 * It makes config changes to the device.
7310 static void raid5_finish_reshape(struct mddev *mddev)
7312 struct r5conf *conf = mddev->private;
7314 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7316 if (mddev->delta_disks > 0) {
7317 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7318 set_capacity(mddev->gendisk, mddev->array_sectors);
7319 revalidate_disk(mddev->gendisk);
7322 spin_lock_irq(&conf->device_lock);
7323 mddev->degraded = calc_degraded(conf);
7324 spin_unlock_irq(&conf->device_lock);
7325 for (d = conf->raid_disks ;
7326 d < conf->raid_disks - mddev->delta_disks;
7328 struct md_rdev *rdev = conf->disks[d].rdev;
7330 clear_bit(In_sync, &rdev->flags);
7331 rdev = conf->disks[d].replacement;
7333 clear_bit(In_sync, &rdev->flags);
7336 mddev->layout = conf->algorithm;
7337 mddev->chunk_sectors = conf->chunk_sectors;
7338 mddev->reshape_position = MaxSector;
7339 mddev->delta_disks = 0;
7340 mddev->reshape_backwards = 0;
7344 static void raid5_quiesce(struct mddev *mddev, int state)
7346 struct r5conf *conf = mddev->private;
7349 case 2: /* resume for a suspend */
7350 wake_up(&conf->wait_for_overlap);
7353 case 1: /* stop all writes */
7354 lock_all_device_hash_locks_irq(conf);
7355 /* '2' tells resync/reshape to pause so that all
7356 * active stripes can drain
7359 wait_event_cmd(conf->wait_for_stripe,
7360 atomic_read(&conf->active_stripes) == 0 &&
7361 atomic_read(&conf->active_aligned_reads) == 0,
7362 unlock_all_device_hash_locks_irq(conf),
7363 lock_all_device_hash_locks_irq(conf));
7365 unlock_all_device_hash_locks_irq(conf);
7366 /* allow reshape to continue */
7367 wake_up(&conf->wait_for_overlap);
7370 case 0: /* re-enable writes */
7371 lock_all_device_hash_locks_irq(conf);
7373 wake_up(&conf->wait_for_stripe);
7374 wake_up(&conf->wait_for_overlap);
7375 unlock_all_device_hash_locks_irq(conf);
7380 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7382 struct r0conf *raid0_conf = mddev->private;
7385 /* for raid0 takeover only one zone is supported */
7386 if (raid0_conf->nr_strip_zones > 1) {
7387 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7389 return ERR_PTR(-EINVAL);
7392 sectors = raid0_conf->strip_zone[0].zone_end;
7393 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7394 mddev->dev_sectors = sectors;
7395 mddev->new_level = level;
7396 mddev->new_layout = ALGORITHM_PARITY_N;
7397 mddev->new_chunk_sectors = mddev->chunk_sectors;
7398 mddev->raid_disks += 1;
7399 mddev->delta_disks = 1;
7400 /* make sure it will be not marked as dirty */
7401 mddev->recovery_cp = MaxSector;
7403 return setup_conf(mddev);
7406 static void *raid5_takeover_raid1(struct mddev *mddev)
7410 if (mddev->raid_disks != 2 ||
7411 mddev->degraded > 1)
7412 return ERR_PTR(-EINVAL);
7414 /* Should check if there are write-behind devices? */
7416 chunksect = 64*2; /* 64K by default */
7418 /* The array must be an exact multiple of chunksize */
7419 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7422 if ((chunksect<<9) < STRIPE_SIZE)
7423 /* array size does not allow a suitable chunk size */
7424 return ERR_PTR(-EINVAL);
7426 mddev->new_level = 5;
7427 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7428 mddev->new_chunk_sectors = chunksect;
7430 return setup_conf(mddev);
7433 static void *raid5_takeover_raid6(struct mddev *mddev)
7437 switch (mddev->layout) {
7438 case ALGORITHM_LEFT_ASYMMETRIC_6:
7439 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7441 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7442 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7444 case ALGORITHM_LEFT_SYMMETRIC_6:
7445 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7447 case ALGORITHM_RIGHT_SYMMETRIC_6:
7448 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7450 case ALGORITHM_PARITY_0_6:
7451 new_layout = ALGORITHM_PARITY_0;
7453 case ALGORITHM_PARITY_N:
7454 new_layout = ALGORITHM_PARITY_N;
7457 return ERR_PTR(-EINVAL);
7459 mddev->new_level = 5;
7460 mddev->new_layout = new_layout;
7461 mddev->delta_disks = -1;
7462 mddev->raid_disks -= 1;
7463 return setup_conf(mddev);
7466 static int raid5_check_reshape(struct mddev *mddev)
7468 /* For a 2-drive array, the layout and chunk size can be changed
7469 * immediately as not restriping is needed.
7470 * For larger arrays we record the new value - after validation
7471 * to be used by a reshape pass.
7473 struct r5conf *conf = mddev->private;
7474 int new_chunk = mddev->new_chunk_sectors;
7476 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7478 if (new_chunk > 0) {
7479 if (!is_power_of_2(new_chunk))
7481 if (new_chunk < (PAGE_SIZE>>9))
7483 if (mddev->array_sectors & (new_chunk-1))
7484 /* not factor of array size */
7488 /* They look valid */
7490 if (mddev->raid_disks == 2) {
7491 /* can make the change immediately */
7492 if (mddev->new_layout >= 0) {
7493 conf->algorithm = mddev->new_layout;
7494 mddev->layout = mddev->new_layout;
7496 if (new_chunk > 0) {
7497 conf->chunk_sectors = new_chunk ;
7498 mddev->chunk_sectors = new_chunk;
7500 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7501 md_wakeup_thread(mddev->thread);
7503 return check_reshape(mddev);
7506 static int raid6_check_reshape(struct mddev *mddev)
7508 int new_chunk = mddev->new_chunk_sectors;
7510 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7512 if (new_chunk > 0) {
7513 if (!is_power_of_2(new_chunk))
7515 if (new_chunk < (PAGE_SIZE >> 9))
7517 if (mddev->array_sectors & (new_chunk-1))
7518 /* not factor of array size */
7522 /* They look valid */
7523 return check_reshape(mddev);
7526 static void *raid5_takeover(struct mddev *mddev)
7528 /* raid5 can take over:
7529 * raid0 - if there is only one strip zone - make it a raid4 layout
7530 * raid1 - if there are two drives. We need to know the chunk size
7531 * raid4 - trivial - just use a raid4 layout.
7532 * raid6 - Providing it is a *_6 layout
7534 if (mddev->level == 0)
7535 return raid45_takeover_raid0(mddev, 5);
7536 if (mddev->level == 1)
7537 return raid5_takeover_raid1(mddev);
7538 if (mddev->level == 4) {
7539 mddev->new_layout = ALGORITHM_PARITY_N;
7540 mddev->new_level = 5;
7541 return setup_conf(mddev);
7543 if (mddev->level == 6)
7544 return raid5_takeover_raid6(mddev);
7546 return ERR_PTR(-EINVAL);
7549 static void *raid4_takeover(struct mddev *mddev)
7551 /* raid4 can take over:
7552 * raid0 - if there is only one strip zone
7553 * raid5 - if layout is right
7555 if (mddev->level == 0)
7556 return raid45_takeover_raid0(mddev, 4);
7557 if (mddev->level == 5 &&
7558 mddev->layout == ALGORITHM_PARITY_N) {
7559 mddev->new_layout = 0;
7560 mddev->new_level = 4;
7561 return setup_conf(mddev);
7563 return ERR_PTR(-EINVAL);
7566 static struct md_personality raid5_personality;
7568 static void *raid6_takeover(struct mddev *mddev)
7570 /* Currently can only take over a raid5. We map the
7571 * personality to an equivalent raid6 personality
7572 * with the Q block at the end.
7576 if (mddev->pers != &raid5_personality)
7577 return ERR_PTR(-EINVAL);
7578 if (mddev->degraded > 1)
7579 return ERR_PTR(-EINVAL);
7580 if (mddev->raid_disks > 253)
7581 return ERR_PTR(-EINVAL);
7582 if (mddev->raid_disks < 3)
7583 return ERR_PTR(-EINVAL);
7585 switch (mddev->layout) {
7586 case ALGORITHM_LEFT_ASYMMETRIC:
7587 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7589 case ALGORITHM_RIGHT_ASYMMETRIC:
7590 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7592 case ALGORITHM_LEFT_SYMMETRIC:
7593 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7595 case ALGORITHM_RIGHT_SYMMETRIC:
7596 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7598 case ALGORITHM_PARITY_0:
7599 new_layout = ALGORITHM_PARITY_0_6;
7601 case ALGORITHM_PARITY_N:
7602 new_layout = ALGORITHM_PARITY_N;
7605 return ERR_PTR(-EINVAL);
7607 mddev->new_level = 6;
7608 mddev->new_layout = new_layout;
7609 mddev->delta_disks = 1;
7610 mddev->raid_disks += 1;
7611 return setup_conf(mddev);
7614 static struct md_personality raid6_personality =
7618 .owner = THIS_MODULE,
7619 .make_request = make_request,
7623 .error_handler = error,
7624 .hot_add_disk = raid5_add_disk,
7625 .hot_remove_disk= raid5_remove_disk,
7626 .spare_active = raid5_spare_active,
7627 .sync_request = sync_request,
7628 .resize = raid5_resize,
7630 .check_reshape = raid6_check_reshape,
7631 .start_reshape = raid5_start_reshape,
7632 .finish_reshape = raid5_finish_reshape,
7633 .quiesce = raid5_quiesce,
7634 .takeover = raid6_takeover,
7635 .congested = raid5_congested,
7636 .mergeable_bvec = raid5_mergeable_bvec,
7638 static struct md_personality raid5_personality =
7642 .owner = THIS_MODULE,
7643 .make_request = make_request,
7647 .error_handler = error,
7648 .hot_add_disk = raid5_add_disk,
7649 .hot_remove_disk= raid5_remove_disk,
7650 .spare_active = raid5_spare_active,
7651 .sync_request = sync_request,
7652 .resize = raid5_resize,
7654 .check_reshape = raid5_check_reshape,
7655 .start_reshape = raid5_start_reshape,
7656 .finish_reshape = raid5_finish_reshape,
7657 .quiesce = raid5_quiesce,
7658 .takeover = raid5_takeover,
7659 .congested = raid5_congested,
7660 .mergeable_bvec = raid5_mergeable_bvec,
7663 static struct md_personality raid4_personality =
7667 .owner = THIS_MODULE,
7668 .make_request = make_request,
7672 .error_handler = error,
7673 .hot_add_disk = raid5_add_disk,
7674 .hot_remove_disk= raid5_remove_disk,
7675 .spare_active = raid5_spare_active,
7676 .sync_request = sync_request,
7677 .resize = raid5_resize,
7679 .check_reshape = raid5_check_reshape,
7680 .start_reshape = raid5_start_reshape,
7681 .finish_reshape = raid5_finish_reshape,
7682 .quiesce = raid5_quiesce,
7683 .takeover = raid4_takeover,
7684 .congested = raid5_congested,
7685 .mergeable_bvec = raid5_mergeable_bvec,
7688 static int __init raid5_init(void)
7690 raid5_wq = alloc_workqueue("raid5wq",
7691 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7694 register_md_personality(&raid6_personality);
7695 register_md_personality(&raid5_personality);
7696 register_md_personality(&raid4_personality);
7700 static void raid5_exit(void)
7702 unregister_md_personality(&raid6_personality);
7703 unregister_md_personality(&raid5_personality);
7704 unregister_md_personality(&raid4_personality);
7705 destroy_workqueue(raid5_wq);
7708 module_init(raid5_init);
7709 module_exit(raid5_exit);
7710 MODULE_LICENSE("GPL");
7711 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7712 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7713 MODULE_ALIAS("md-raid5");
7714 MODULE_ALIAS("md-raid4");
7715 MODULE_ALIAS("md-level-5");
7716 MODULE_ALIAS("md-level-4");
7717 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7718 MODULE_ALIAS("md-raid6");
7719 MODULE_ALIAS("md-level-6");
7721 /* This used to be two separate modules, they were: */
7722 MODULE_ALIAS("raid5");
7723 MODULE_ALIAS("raid6");