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)
503 int num = sh->raid_conf->pool_size;
505 for (i = 0; i < num; i++) {
508 if (!(page = alloc_page(GFP_KERNEL))) {
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);
1073 set_bit(STRIPE_HANDLE, &sh->state);
1076 if (!head_sh->batch_head)
1078 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1085 static struct dma_async_tx_descriptor *
1086 async_copy_data(int frombio, struct bio *bio, struct page **page,
1087 sector_t sector, struct dma_async_tx_descriptor *tx,
1088 struct stripe_head *sh)
1091 struct bvec_iter iter;
1092 struct page *bio_page;
1094 struct async_submit_ctl submit;
1095 enum async_tx_flags flags = 0;
1097 if (bio->bi_iter.bi_sector >= sector)
1098 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1100 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1103 flags |= ASYNC_TX_FENCE;
1104 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1106 bio_for_each_segment(bvl, bio, iter) {
1107 int len = bvl.bv_len;
1111 if (page_offset < 0) {
1112 b_offset = -page_offset;
1113 page_offset += b_offset;
1117 if (len > 0 && page_offset + len > STRIPE_SIZE)
1118 clen = STRIPE_SIZE - page_offset;
1123 b_offset += bvl.bv_offset;
1124 bio_page = bvl.bv_page;
1126 if (sh->raid_conf->skip_copy &&
1127 b_offset == 0 && page_offset == 0 &&
1128 clen == STRIPE_SIZE)
1131 tx = async_memcpy(*page, bio_page, page_offset,
1132 b_offset, clen, &submit);
1134 tx = async_memcpy(bio_page, *page, b_offset,
1135 page_offset, clen, &submit);
1137 /* chain the operations */
1138 submit.depend_tx = tx;
1140 if (clen < len) /* hit end of page */
1148 static void ops_complete_biofill(void *stripe_head_ref)
1150 struct stripe_head *sh = stripe_head_ref;
1151 struct bio *return_bi = NULL;
1154 pr_debug("%s: stripe %llu\n", __func__,
1155 (unsigned long long)sh->sector);
1157 /* clear completed biofills */
1158 for (i = sh->disks; i--; ) {
1159 struct r5dev *dev = &sh->dev[i];
1161 /* acknowledge completion of a biofill operation */
1162 /* and check if we need to reply to a read request,
1163 * new R5_Wantfill requests are held off until
1164 * !STRIPE_BIOFILL_RUN
1166 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1167 struct bio *rbi, *rbi2;
1172 while (rbi && rbi->bi_iter.bi_sector <
1173 dev->sector + STRIPE_SECTORS) {
1174 rbi2 = r5_next_bio(rbi, dev->sector);
1175 if (!raid5_dec_bi_active_stripes(rbi)) {
1176 rbi->bi_next = return_bi;
1183 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1185 return_io(return_bi);
1187 set_bit(STRIPE_HANDLE, &sh->state);
1191 static void ops_run_biofill(struct stripe_head *sh)
1193 struct dma_async_tx_descriptor *tx = NULL;
1194 struct async_submit_ctl submit;
1197 BUG_ON(sh->batch_head);
1198 pr_debug("%s: stripe %llu\n", __func__,
1199 (unsigned long long)sh->sector);
1201 for (i = sh->disks; i--; ) {
1202 struct r5dev *dev = &sh->dev[i];
1203 if (test_bit(R5_Wantfill, &dev->flags)) {
1205 spin_lock_irq(&sh->stripe_lock);
1206 dev->read = rbi = dev->toread;
1208 spin_unlock_irq(&sh->stripe_lock);
1209 while (rbi && rbi->bi_iter.bi_sector <
1210 dev->sector + STRIPE_SECTORS) {
1211 tx = async_copy_data(0, rbi, &dev->page,
1212 dev->sector, tx, sh);
1213 rbi = r5_next_bio(rbi, dev->sector);
1218 atomic_inc(&sh->count);
1219 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1220 async_trigger_callback(&submit);
1223 static void mark_target_uptodate(struct stripe_head *sh, int target)
1230 tgt = &sh->dev[target];
1231 set_bit(R5_UPTODATE, &tgt->flags);
1232 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1233 clear_bit(R5_Wantcompute, &tgt->flags);
1236 static void ops_complete_compute(void *stripe_head_ref)
1238 struct stripe_head *sh = stripe_head_ref;
1240 pr_debug("%s: stripe %llu\n", __func__,
1241 (unsigned long long)sh->sector);
1243 /* mark the computed target(s) as uptodate */
1244 mark_target_uptodate(sh, sh->ops.target);
1245 mark_target_uptodate(sh, sh->ops.target2);
1247 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1248 if (sh->check_state == check_state_compute_run)
1249 sh->check_state = check_state_compute_result;
1250 set_bit(STRIPE_HANDLE, &sh->state);
1254 /* return a pointer to the address conversion region of the scribble buffer */
1255 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1256 struct raid5_percpu *percpu, int i)
1260 addr = flex_array_get(percpu->scribble, i);
1261 return addr + sizeof(struct page *) * (sh->disks + 2);
1264 /* return a pointer to the address conversion region of the scribble buffer */
1265 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1269 addr = flex_array_get(percpu->scribble, i);
1273 static struct dma_async_tx_descriptor *
1274 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1276 int disks = sh->disks;
1277 struct page **xor_srcs = to_addr_page(percpu, 0);
1278 int target = sh->ops.target;
1279 struct r5dev *tgt = &sh->dev[target];
1280 struct page *xor_dest = tgt->page;
1282 struct dma_async_tx_descriptor *tx;
1283 struct async_submit_ctl submit;
1286 BUG_ON(sh->batch_head);
1288 pr_debug("%s: stripe %llu block: %d\n",
1289 __func__, (unsigned long long)sh->sector, target);
1290 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1292 for (i = disks; i--; )
1294 xor_srcs[count++] = sh->dev[i].page;
1296 atomic_inc(&sh->count);
1298 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1299 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1300 if (unlikely(count == 1))
1301 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1303 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1308 /* set_syndrome_sources - populate source buffers for gen_syndrome
1309 * @srcs - (struct page *) array of size sh->disks
1310 * @sh - stripe_head to parse
1312 * Populates srcs in proper layout order for the stripe and returns the
1313 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1314 * destination buffer is recorded in srcs[count] and the Q destination
1315 * is recorded in srcs[count+1]].
1317 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
1319 int disks = sh->disks;
1320 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1321 int d0_idx = raid6_d0(sh);
1325 for (i = 0; i < disks; i++)
1331 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1333 srcs[slot] = sh->dev[i].page;
1334 i = raid6_next_disk(i, disks);
1335 } while (i != d0_idx);
1337 return syndrome_disks;
1340 static struct dma_async_tx_descriptor *
1341 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1343 int disks = sh->disks;
1344 struct page **blocks = to_addr_page(percpu, 0);
1346 int qd_idx = sh->qd_idx;
1347 struct dma_async_tx_descriptor *tx;
1348 struct async_submit_ctl submit;
1354 BUG_ON(sh->batch_head);
1355 if (sh->ops.target < 0)
1356 target = sh->ops.target2;
1357 else if (sh->ops.target2 < 0)
1358 target = sh->ops.target;
1360 /* we should only have one valid target */
1363 pr_debug("%s: stripe %llu block: %d\n",
1364 __func__, (unsigned long long)sh->sector, target);
1366 tgt = &sh->dev[target];
1367 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1370 atomic_inc(&sh->count);
1372 if (target == qd_idx) {
1373 count = set_syndrome_sources(blocks, sh);
1374 blocks[count] = NULL; /* regenerating p is not necessary */
1375 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1376 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1377 ops_complete_compute, sh,
1378 to_addr_conv(sh, percpu, 0));
1379 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1381 /* Compute any data- or p-drive using XOR */
1383 for (i = disks; i-- ; ) {
1384 if (i == target || i == qd_idx)
1386 blocks[count++] = sh->dev[i].page;
1389 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1390 NULL, ops_complete_compute, sh,
1391 to_addr_conv(sh, percpu, 0));
1392 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1398 static struct dma_async_tx_descriptor *
1399 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1401 int i, count, disks = sh->disks;
1402 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1403 int d0_idx = raid6_d0(sh);
1404 int faila = -1, failb = -1;
1405 int target = sh->ops.target;
1406 int target2 = sh->ops.target2;
1407 struct r5dev *tgt = &sh->dev[target];
1408 struct r5dev *tgt2 = &sh->dev[target2];
1409 struct dma_async_tx_descriptor *tx;
1410 struct page **blocks = to_addr_page(percpu, 0);
1411 struct async_submit_ctl submit;
1413 BUG_ON(sh->batch_head);
1414 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1415 __func__, (unsigned long long)sh->sector, target, target2);
1416 BUG_ON(target < 0 || target2 < 0);
1417 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1418 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1420 /* we need to open-code set_syndrome_sources to handle the
1421 * slot number conversion for 'faila' and 'failb'
1423 for (i = 0; i < disks ; i++)
1428 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1430 blocks[slot] = sh->dev[i].page;
1436 i = raid6_next_disk(i, disks);
1437 } while (i != d0_idx);
1439 BUG_ON(faila == failb);
1442 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1443 __func__, (unsigned long long)sh->sector, faila, failb);
1445 atomic_inc(&sh->count);
1447 if (failb == syndrome_disks+1) {
1448 /* Q disk is one of the missing disks */
1449 if (faila == syndrome_disks) {
1450 /* Missing P+Q, just recompute */
1451 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1452 ops_complete_compute, sh,
1453 to_addr_conv(sh, percpu, 0));
1454 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1455 STRIPE_SIZE, &submit);
1459 int qd_idx = sh->qd_idx;
1461 /* Missing D+Q: recompute D from P, then recompute Q */
1462 if (target == qd_idx)
1463 data_target = target2;
1465 data_target = target;
1468 for (i = disks; i-- ; ) {
1469 if (i == data_target || i == qd_idx)
1471 blocks[count++] = sh->dev[i].page;
1473 dest = sh->dev[data_target].page;
1474 init_async_submit(&submit,
1475 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1477 to_addr_conv(sh, percpu, 0));
1478 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1481 count = set_syndrome_sources(blocks, sh);
1482 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1483 ops_complete_compute, sh,
1484 to_addr_conv(sh, percpu, 0));
1485 return async_gen_syndrome(blocks, 0, count+2,
1486 STRIPE_SIZE, &submit);
1489 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1490 ops_complete_compute, sh,
1491 to_addr_conv(sh, percpu, 0));
1492 if (failb == syndrome_disks) {
1493 /* We're missing D+P. */
1494 return async_raid6_datap_recov(syndrome_disks+2,
1498 /* We're missing D+D. */
1499 return async_raid6_2data_recov(syndrome_disks+2,
1500 STRIPE_SIZE, faila, failb,
1506 static void ops_complete_prexor(void *stripe_head_ref)
1508 struct stripe_head *sh = stripe_head_ref;
1510 pr_debug("%s: stripe %llu\n", __func__,
1511 (unsigned long long)sh->sector);
1514 static struct dma_async_tx_descriptor *
1515 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1516 struct dma_async_tx_descriptor *tx)
1518 int disks = sh->disks;
1519 struct page **xor_srcs = to_addr_page(percpu, 0);
1520 int count = 0, pd_idx = sh->pd_idx, i;
1521 struct async_submit_ctl submit;
1523 /* existing parity data subtracted */
1524 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1526 BUG_ON(sh->batch_head);
1527 pr_debug("%s: stripe %llu\n", __func__,
1528 (unsigned long long)sh->sector);
1530 for (i = disks; i--; ) {
1531 struct r5dev *dev = &sh->dev[i];
1532 /* Only process blocks that are known to be uptodate */
1533 if (test_bit(R5_Wantdrain, &dev->flags))
1534 xor_srcs[count++] = dev->page;
1537 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1538 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1539 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1544 static struct dma_async_tx_descriptor *
1545 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1547 int disks = sh->disks;
1549 struct stripe_head *head_sh = sh;
1551 pr_debug("%s: stripe %llu\n", __func__,
1552 (unsigned long long)sh->sector);
1554 for (i = disks; i--; ) {
1559 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1564 spin_lock_irq(&sh->stripe_lock);
1565 chosen = dev->towrite;
1566 dev->towrite = NULL;
1567 sh->overwrite_disks = 0;
1568 BUG_ON(dev->written);
1569 wbi = dev->written = chosen;
1570 spin_unlock_irq(&sh->stripe_lock);
1571 WARN_ON(dev->page != dev->orig_page);
1573 while (wbi && wbi->bi_iter.bi_sector <
1574 dev->sector + STRIPE_SECTORS) {
1575 if (wbi->bi_rw & REQ_FUA)
1576 set_bit(R5_WantFUA, &dev->flags);
1577 if (wbi->bi_rw & REQ_SYNC)
1578 set_bit(R5_SyncIO, &dev->flags);
1579 if (wbi->bi_rw & REQ_DISCARD)
1580 set_bit(R5_Discard, &dev->flags);
1582 tx = async_copy_data(1, wbi, &dev->page,
1583 dev->sector, tx, sh);
1584 if (dev->page != dev->orig_page) {
1585 set_bit(R5_SkipCopy, &dev->flags);
1586 clear_bit(R5_UPTODATE, &dev->flags);
1587 clear_bit(R5_OVERWRITE, &dev->flags);
1590 wbi = r5_next_bio(wbi, dev->sector);
1593 if (head_sh->batch_head) {
1594 sh = list_first_entry(&sh->batch_list,
1607 static void ops_complete_reconstruct(void *stripe_head_ref)
1609 struct stripe_head *sh = stripe_head_ref;
1610 int disks = sh->disks;
1611 int pd_idx = sh->pd_idx;
1612 int qd_idx = sh->qd_idx;
1614 bool fua = false, sync = false, discard = false;
1616 pr_debug("%s: stripe %llu\n", __func__,
1617 (unsigned long long)sh->sector);
1619 for (i = disks; i--; ) {
1620 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1621 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1622 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1625 for (i = disks; i--; ) {
1626 struct r5dev *dev = &sh->dev[i];
1628 if (dev->written || i == pd_idx || i == qd_idx) {
1629 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1630 set_bit(R5_UPTODATE, &dev->flags);
1632 set_bit(R5_WantFUA, &dev->flags);
1634 set_bit(R5_SyncIO, &dev->flags);
1638 if (sh->reconstruct_state == reconstruct_state_drain_run)
1639 sh->reconstruct_state = reconstruct_state_drain_result;
1640 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1641 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1643 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1644 sh->reconstruct_state = reconstruct_state_result;
1647 set_bit(STRIPE_HANDLE, &sh->state);
1652 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1653 struct dma_async_tx_descriptor *tx)
1655 int disks = sh->disks;
1656 struct page **xor_srcs;
1657 struct async_submit_ctl submit;
1658 int count, pd_idx = sh->pd_idx, i;
1659 struct page *xor_dest;
1661 unsigned long flags;
1663 struct stripe_head *head_sh = sh;
1666 pr_debug("%s: stripe %llu\n", __func__,
1667 (unsigned long long)sh->sector);
1669 for (i = 0; i < sh->disks; i++) {
1672 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1675 if (i >= sh->disks) {
1676 atomic_inc(&sh->count);
1677 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1678 ops_complete_reconstruct(sh);
1683 xor_srcs = to_addr_page(percpu, j);
1684 /* check if prexor is active which means only process blocks
1685 * that are part of a read-modify-write (written)
1687 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1689 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1690 for (i = disks; i--; ) {
1691 struct r5dev *dev = &sh->dev[i];
1692 if (head_sh->dev[i].written)
1693 xor_srcs[count++] = dev->page;
1696 xor_dest = sh->dev[pd_idx].page;
1697 for (i = disks; i--; ) {
1698 struct r5dev *dev = &sh->dev[i];
1700 xor_srcs[count++] = dev->page;
1704 /* 1/ if we prexor'd then the dest is reused as a source
1705 * 2/ if we did not prexor then we are redoing the parity
1706 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1707 * for the synchronous xor case
1709 last_stripe = !head_sh->batch_head ||
1710 list_first_entry(&sh->batch_list,
1711 struct stripe_head, batch_list) == head_sh;
1713 flags = ASYNC_TX_ACK |
1714 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1716 atomic_inc(&head_sh->count);
1717 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1718 to_addr_conv(sh, percpu, j));
1720 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1721 init_async_submit(&submit, flags, tx, NULL, NULL,
1722 to_addr_conv(sh, percpu, j));
1725 if (unlikely(count == 1))
1726 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1728 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1731 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1738 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1739 struct dma_async_tx_descriptor *tx)
1741 struct async_submit_ctl submit;
1742 struct page **blocks;
1743 int count, i, j = 0;
1744 struct stripe_head *head_sh = sh;
1747 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1749 for (i = 0; i < sh->disks; i++) {
1750 if (sh->pd_idx == i || sh->qd_idx == i)
1752 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1755 if (i >= sh->disks) {
1756 atomic_inc(&sh->count);
1757 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1758 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1759 ops_complete_reconstruct(sh);
1764 blocks = to_addr_page(percpu, j);
1765 count = set_syndrome_sources(blocks, sh);
1766 last_stripe = !head_sh->batch_head ||
1767 list_first_entry(&sh->batch_list,
1768 struct stripe_head, batch_list) == head_sh;
1771 atomic_inc(&head_sh->count);
1772 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1773 head_sh, to_addr_conv(sh, percpu, j));
1775 init_async_submit(&submit, 0, tx, NULL, NULL,
1776 to_addr_conv(sh, percpu, j));
1777 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1780 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1786 static void ops_complete_check(void *stripe_head_ref)
1788 struct stripe_head *sh = stripe_head_ref;
1790 pr_debug("%s: stripe %llu\n", __func__,
1791 (unsigned long long)sh->sector);
1793 sh->check_state = check_state_check_result;
1794 set_bit(STRIPE_HANDLE, &sh->state);
1798 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1800 int disks = sh->disks;
1801 int pd_idx = sh->pd_idx;
1802 int qd_idx = sh->qd_idx;
1803 struct page *xor_dest;
1804 struct page **xor_srcs = to_addr_page(percpu, 0);
1805 struct dma_async_tx_descriptor *tx;
1806 struct async_submit_ctl submit;
1810 pr_debug("%s: stripe %llu\n", __func__,
1811 (unsigned long long)sh->sector);
1813 BUG_ON(sh->batch_head);
1815 xor_dest = sh->dev[pd_idx].page;
1816 xor_srcs[count++] = xor_dest;
1817 for (i = disks; i--; ) {
1818 if (i == pd_idx || i == qd_idx)
1820 xor_srcs[count++] = sh->dev[i].page;
1823 init_async_submit(&submit, 0, NULL, NULL, NULL,
1824 to_addr_conv(sh, percpu, 0));
1825 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1826 &sh->ops.zero_sum_result, &submit);
1828 atomic_inc(&sh->count);
1829 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1830 tx = async_trigger_callback(&submit);
1833 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1835 struct page **srcs = to_addr_page(percpu, 0);
1836 struct async_submit_ctl submit;
1839 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1840 (unsigned long long)sh->sector, checkp);
1842 BUG_ON(sh->batch_head);
1843 count = set_syndrome_sources(srcs, sh);
1847 atomic_inc(&sh->count);
1848 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1849 sh, to_addr_conv(sh, percpu, 0));
1850 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1851 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1854 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1856 int overlap_clear = 0, i, disks = sh->disks;
1857 struct dma_async_tx_descriptor *tx = NULL;
1858 struct r5conf *conf = sh->raid_conf;
1859 int level = conf->level;
1860 struct raid5_percpu *percpu;
1864 percpu = per_cpu_ptr(conf->percpu, cpu);
1865 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1866 ops_run_biofill(sh);
1870 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1872 tx = ops_run_compute5(sh, percpu);
1874 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1875 tx = ops_run_compute6_1(sh, percpu);
1877 tx = ops_run_compute6_2(sh, percpu);
1879 /* terminate the chain if reconstruct is not set to be run */
1880 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1884 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1885 tx = ops_run_prexor(sh, percpu, tx);
1887 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1888 tx = ops_run_biodrain(sh, tx);
1892 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1894 ops_run_reconstruct5(sh, percpu, tx);
1896 ops_run_reconstruct6(sh, percpu, tx);
1899 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1900 if (sh->check_state == check_state_run)
1901 ops_run_check_p(sh, percpu);
1902 else if (sh->check_state == check_state_run_q)
1903 ops_run_check_pq(sh, percpu, 0);
1904 else if (sh->check_state == check_state_run_pq)
1905 ops_run_check_pq(sh, percpu, 1);
1910 if (overlap_clear && !sh->batch_head)
1911 for (i = disks; i--; ) {
1912 struct r5dev *dev = &sh->dev[i];
1913 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1914 wake_up(&sh->raid_conf->wait_for_overlap);
1919 static int grow_one_stripe(struct r5conf *conf, int hash)
1921 struct stripe_head *sh;
1922 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1926 sh->raid_conf = conf;
1928 spin_lock_init(&sh->stripe_lock);
1930 if (grow_buffers(sh)) {
1932 kmem_cache_free(conf->slab_cache, sh);
1935 sh->hash_lock_index = hash;
1936 /* we just created an active stripe so... */
1937 atomic_set(&sh->count, 1);
1938 atomic_inc(&conf->active_stripes);
1939 INIT_LIST_HEAD(&sh->lru);
1941 spin_lock_init(&sh->batch_lock);
1942 INIT_LIST_HEAD(&sh->batch_list);
1943 sh->batch_head = NULL;
1948 static int grow_stripes(struct r5conf *conf, int num)
1950 struct kmem_cache *sc;
1951 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1954 if (conf->mddev->gendisk)
1955 sprintf(conf->cache_name[0],
1956 "raid%d-%s", conf->level, mdname(conf->mddev));
1958 sprintf(conf->cache_name[0],
1959 "raid%d-%p", conf->level, conf->mddev);
1960 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1962 conf->active_name = 0;
1963 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1964 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1968 conf->slab_cache = sc;
1969 conf->pool_size = devs;
1970 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
1972 if (!grow_one_stripe(conf, hash))
1974 conf->max_nr_stripes++;
1975 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
1981 * scribble_len - return the required size of the scribble region
1982 * @num - total number of disks in the array
1984 * The size must be enough to contain:
1985 * 1/ a struct page pointer for each device in the array +2
1986 * 2/ room to convert each entry in (1) to its corresponding dma
1987 * (dma_map_page()) or page (page_address()) address.
1989 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1990 * calculate over all devices (not just the data blocks), using zeros in place
1991 * of the P and Q blocks.
1993 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
1995 struct flex_array *ret;
1998 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1999 ret = flex_array_alloc(len, cnt, flags);
2002 /* always prealloc all elements, so no locking is required */
2003 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2004 flex_array_free(ret);
2010 static int resize_stripes(struct r5conf *conf, int newsize)
2012 /* Make all the stripes able to hold 'newsize' devices.
2013 * New slots in each stripe get 'page' set to a new page.
2015 * This happens in stages:
2016 * 1/ create a new kmem_cache and allocate the required number of
2018 * 2/ gather all the old stripe_heads and transfer the pages across
2019 * to the new stripe_heads. This will have the side effect of
2020 * freezing the array as once all stripe_heads have been collected,
2021 * no IO will be possible. Old stripe heads are freed once their
2022 * pages have been transferred over, and the old kmem_cache is
2023 * freed when all stripes are done.
2024 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2025 * we simple return a failre status - no need to clean anything up.
2026 * 4/ allocate new pages for the new slots in the new stripe_heads.
2027 * If this fails, we don't bother trying the shrink the
2028 * stripe_heads down again, we just leave them as they are.
2029 * As each stripe_head is processed the new one is released into
2032 * Once step2 is started, we cannot afford to wait for a write,
2033 * so we use GFP_NOIO allocations.
2035 struct stripe_head *osh, *nsh;
2036 LIST_HEAD(newstripes);
2037 struct disk_info *ndisks;
2040 struct kmem_cache *sc;
2044 if (newsize <= conf->pool_size)
2045 return 0; /* never bother to shrink */
2047 err = md_allow_write(conf->mddev);
2052 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2053 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2058 for (i = conf->max_nr_stripes; i; i--) {
2059 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
2063 nsh->raid_conf = conf;
2064 spin_lock_init(&nsh->stripe_lock);
2066 list_add(&nsh->lru, &newstripes);
2069 /* didn't get enough, give up */
2070 while (!list_empty(&newstripes)) {
2071 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2072 list_del(&nsh->lru);
2073 kmem_cache_free(sc, nsh);
2075 kmem_cache_destroy(sc);
2078 /* Step 2 - Must use GFP_NOIO now.
2079 * OK, we have enough stripes, start collecting inactive
2080 * stripes and copying them over
2084 list_for_each_entry(nsh, &newstripes, lru) {
2085 lock_device_hash_lock(conf, hash);
2086 wait_event_cmd(conf->wait_for_stripe,
2087 !list_empty(conf->inactive_list + hash),
2088 unlock_device_hash_lock(conf, hash),
2089 lock_device_hash_lock(conf, hash));
2090 osh = get_free_stripe(conf, hash);
2091 unlock_device_hash_lock(conf, hash);
2092 atomic_set(&nsh->count, 1);
2093 for(i=0; i<conf->pool_size; i++) {
2094 nsh->dev[i].page = osh->dev[i].page;
2095 nsh->dev[i].orig_page = osh->dev[i].page;
2097 for( ; i<newsize; i++)
2098 nsh->dev[i].page = NULL;
2099 nsh->hash_lock_index = hash;
2100 kmem_cache_free(conf->slab_cache, osh);
2102 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2103 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2108 kmem_cache_destroy(conf->slab_cache);
2111 * At this point, we are holding all the stripes so the array
2112 * is completely stalled, so now is a good time to resize
2113 * conf->disks and the scribble region
2115 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2117 for (i=0; i<conf->raid_disks; i++)
2118 ndisks[i] = conf->disks[i];
2120 conf->disks = ndisks;
2125 for_each_present_cpu(cpu) {
2126 struct raid5_percpu *percpu;
2127 struct flex_array *scribble;
2129 percpu = per_cpu_ptr(conf->percpu, cpu);
2130 scribble = scribble_alloc(newsize, conf->chunk_sectors /
2131 STRIPE_SECTORS, GFP_NOIO);
2134 flex_array_free(percpu->scribble);
2135 percpu->scribble = scribble;
2143 /* Step 4, return new stripes to service */
2144 while(!list_empty(&newstripes)) {
2145 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2146 list_del_init(&nsh->lru);
2148 for (i=conf->raid_disks; i < newsize; i++)
2149 if (nsh->dev[i].page == NULL) {
2150 struct page *p = alloc_page(GFP_NOIO);
2151 nsh->dev[i].page = p;
2152 nsh->dev[i].orig_page = p;
2156 release_stripe(nsh);
2158 /* critical section pass, GFP_NOIO no longer needed */
2160 conf->slab_cache = sc;
2161 conf->active_name = 1-conf->active_name;
2162 conf->pool_size = newsize;
2166 static int drop_one_stripe(struct r5conf *conf, int hash)
2168 struct stripe_head *sh;
2170 spin_lock_irq(conf->hash_locks + hash);
2171 sh = get_free_stripe(conf, hash);
2172 spin_unlock_irq(conf->hash_locks + hash);
2175 BUG_ON(atomic_read(&sh->count));
2177 kmem_cache_free(conf->slab_cache, sh);
2178 atomic_dec(&conf->active_stripes);
2182 static void shrink_stripes(struct r5conf *conf)
2185 for (hash = 0; hash < NR_STRIPE_HASH_LOCKS; hash++)
2186 while (drop_one_stripe(conf, hash))
2189 if (conf->slab_cache)
2190 kmem_cache_destroy(conf->slab_cache);
2191 conf->slab_cache = NULL;
2194 static void raid5_end_read_request(struct bio * bi, int error)
2196 struct stripe_head *sh = bi->bi_private;
2197 struct r5conf *conf = sh->raid_conf;
2198 int disks = sh->disks, i;
2199 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2200 char b[BDEVNAME_SIZE];
2201 struct md_rdev *rdev = NULL;
2204 for (i=0 ; i<disks; i++)
2205 if (bi == &sh->dev[i].req)
2208 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2209 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2215 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2216 /* If replacement finished while this request was outstanding,
2217 * 'replacement' might be NULL already.
2218 * In that case it moved down to 'rdev'.
2219 * rdev is not removed until all requests are finished.
2221 rdev = conf->disks[i].replacement;
2223 rdev = conf->disks[i].rdev;
2225 if (use_new_offset(conf, sh))
2226 s = sh->sector + rdev->new_data_offset;
2228 s = sh->sector + rdev->data_offset;
2230 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2231 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2232 /* Note that this cannot happen on a
2233 * replacement device. We just fail those on
2238 "md/raid:%s: read error corrected"
2239 " (%lu sectors at %llu on %s)\n",
2240 mdname(conf->mddev), STRIPE_SECTORS,
2241 (unsigned long long)s,
2242 bdevname(rdev->bdev, b));
2243 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2244 clear_bit(R5_ReadError, &sh->dev[i].flags);
2245 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2246 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2247 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2249 if (atomic_read(&rdev->read_errors))
2250 atomic_set(&rdev->read_errors, 0);
2252 const char *bdn = bdevname(rdev->bdev, b);
2256 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2257 atomic_inc(&rdev->read_errors);
2258 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2261 "md/raid:%s: read error on replacement device "
2262 "(sector %llu on %s).\n",
2263 mdname(conf->mddev),
2264 (unsigned long long)s,
2266 else if (conf->mddev->degraded >= conf->max_degraded) {
2270 "md/raid:%s: read error not correctable "
2271 "(sector %llu on %s).\n",
2272 mdname(conf->mddev),
2273 (unsigned long long)s,
2275 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2280 "md/raid:%s: read error NOT corrected!! "
2281 "(sector %llu on %s).\n",
2282 mdname(conf->mddev),
2283 (unsigned long long)s,
2285 } else if (atomic_read(&rdev->read_errors)
2286 > conf->max_nr_stripes)
2288 "md/raid:%s: Too many read errors, failing device %s.\n",
2289 mdname(conf->mddev), bdn);
2292 if (set_bad && test_bit(In_sync, &rdev->flags)
2293 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2296 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2297 set_bit(R5_ReadError, &sh->dev[i].flags);
2298 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2300 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2302 clear_bit(R5_ReadError, &sh->dev[i].flags);
2303 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2305 && test_bit(In_sync, &rdev->flags)
2306 && rdev_set_badblocks(
2307 rdev, sh->sector, STRIPE_SECTORS, 0)))
2308 md_error(conf->mddev, rdev);
2311 rdev_dec_pending(rdev, conf->mddev);
2312 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2313 set_bit(STRIPE_HANDLE, &sh->state);
2317 static void raid5_end_write_request(struct bio *bi, int error)
2319 struct stripe_head *sh = bi->bi_private;
2320 struct r5conf *conf = sh->raid_conf;
2321 int disks = sh->disks, i;
2322 struct md_rdev *uninitialized_var(rdev);
2323 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2326 int replacement = 0;
2328 for (i = 0 ; i < disks; i++) {
2329 if (bi == &sh->dev[i].req) {
2330 rdev = conf->disks[i].rdev;
2333 if (bi == &sh->dev[i].rreq) {
2334 rdev = conf->disks[i].replacement;
2338 /* rdev was removed and 'replacement'
2339 * replaced it. rdev is not removed
2340 * until all requests are finished.
2342 rdev = conf->disks[i].rdev;
2346 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2347 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2356 md_error(conf->mddev, rdev);
2357 else if (is_badblock(rdev, sh->sector,
2359 &first_bad, &bad_sectors))
2360 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2363 set_bit(STRIPE_DEGRADED, &sh->state);
2364 set_bit(WriteErrorSeen, &rdev->flags);
2365 set_bit(R5_WriteError, &sh->dev[i].flags);
2366 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2367 set_bit(MD_RECOVERY_NEEDED,
2368 &rdev->mddev->recovery);
2369 } else if (is_badblock(rdev, sh->sector,
2371 &first_bad, &bad_sectors)) {
2372 set_bit(R5_MadeGood, &sh->dev[i].flags);
2373 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2374 /* That was a successful write so make
2375 * sure it looks like we already did
2378 set_bit(R5_ReWrite, &sh->dev[i].flags);
2381 rdev_dec_pending(rdev, conf->mddev);
2383 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2384 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2385 set_bit(STRIPE_HANDLE, &sh->state);
2388 if (sh->batch_head && sh != sh->batch_head)
2389 release_stripe(sh->batch_head);
2392 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2394 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2396 struct r5dev *dev = &sh->dev[i];
2398 bio_init(&dev->req);
2399 dev->req.bi_io_vec = &dev->vec;
2400 dev->req.bi_max_vecs = 1;
2401 dev->req.bi_private = sh;
2403 bio_init(&dev->rreq);
2404 dev->rreq.bi_io_vec = &dev->rvec;
2405 dev->rreq.bi_max_vecs = 1;
2406 dev->rreq.bi_private = sh;
2409 dev->sector = compute_blocknr(sh, i, previous);
2412 static void error(struct mddev *mddev, struct md_rdev *rdev)
2414 char b[BDEVNAME_SIZE];
2415 struct r5conf *conf = mddev->private;
2416 unsigned long flags;
2417 pr_debug("raid456: error called\n");
2419 spin_lock_irqsave(&conf->device_lock, flags);
2420 clear_bit(In_sync, &rdev->flags);
2421 mddev->degraded = calc_degraded(conf);
2422 spin_unlock_irqrestore(&conf->device_lock, flags);
2423 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2425 set_bit(Blocked, &rdev->flags);
2426 set_bit(Faulty, &rdev->flags);
2427 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2429 "md/raid:%s: Disk failure on %s, disabling device.\n"
2430 "md/raid:%s: Operation continuing on %d devices.\n",
2432 bdevname(rdev->bdev, b),
2434 conf->raid_disks - mddev->degraded);
2438 * Input: a 'big' sector number,
2439 * Output: index of the data and parity disk, and the sector # in them.
2441 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2442 int previous, int *dd_idx,
2443 struct stripe_head *sh)
2445 sector_t stripe, stripe2;
2446 sector_t chunk_number;
2447 unsigned int chunk_offset;
2450 sector_t new_sector;
2451 int algorithm = previous ? conf->prev_algo
2453 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2454 : conf->chunk_sectors;
2455 int raid_disks = previous ? conf->previous_raid_disks
2457 int data_disks = raid_disks - conf->max_degraded;
2459 /* First compute the information on this sector */
2462 * Compute the chunk number and the sector offset inside the chunk
2464 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2465 chunk_number = r_sector;
2468 * Compute the stripe number
2470 stripe = chunk_number;
2471 *dd_idx = sector_div(stripe, data_disks);
2474 * Select the parity disk based on the user selected algorithm.
2476 pd_idx = qd_idx = -1;
2477 switch(conf->level) {
2479 pd_idx = data_disks;
2482 switch (algorithm) {
2483 case ALGORITHM_LEFT_ASYMMETRIC:
2484 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2485 if (*dd_idx >= pd_idx)
2488 case ALGORITHM_RIGHT_ASYMMETRIC:
2489 pd_idx = sector_div(stripe2, raid_disks);
2490 if (*dd_idx >= pd_idx)
2493 case ALGORITHM_LEFT_SYMMETRIC:
2494 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2495 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2497 case ALGORITHM_RIGHT_SYMMETRIC:
2498 pd_idx = sector_div(stripe2, raid_disks);
2499 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2501 case ALGORITHM_PARITY_0:
2505 case ALGORITHM_PARITY_N:
2506 pd_idx = data_disks;
2514 switch (algorithm) {
2515 case ALGORITHM_LEFT_ASYMMETRIC:
2516 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2517 qd_idx = pd_idx + 1;
2518 if (pd_idx == raid_disks-1) {
2519 (*dd_idx)++; /* Q D D D P */
2521 } else if (*dd_idx >= pd_idx)
2522 (*dd_idx) += 2; /* D D P Q D */
2524 case ALGORITHM_RIGHT_ASYMMETRIC:
2525 pd_idx = sector_div(stripe2, raid_disks);
2526 qd_idx = pd_idx + 1;
2527 if (pd_idx == raid_disks-1) {
2528 (*dd_idx)++; /* Q D D D P */
2530 } else if (*dd_idx >= pd_idx)
2531 (*dd_idx) += 2; /* D D P Q D */
2533 case ALGORITHM_LEFT_SYMMETRIC:
2534 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2535 qd_idx = (pd_idx + 1) % raid_disks;
2536 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2538 case ALGORITHM_RIGHT_SYMMETRIC:
2539 pd_idx = sector_div(stripe2, raid_disks);
2540 qd_idx = (pd_idx + 1) % raid_disks;
2541 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2544 case ALGORITHM_PARITY_0:
2549 case ALGORITHM_PARITY_N:
2550 pd_idx = data_disks;
2551 qd_idx = data_disks + 1;
2554 case ALGORITHM_ROTATING_ZERO_RESTART:
2555 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2556 * of blocks for computing Q is different.
2558 pd_idx = sector_div(stripe2, raid_disks);
2559 qd_idx = pd_idx + 1;
2560 if (pd_idx == raid_disks-1) {
2561 (*dd_idx)++; /* Q D D D P */
2563 } else if (*dd_idx >= pd_idx)
2564 (*dd_idx) += 2; /* D D P Q D */
2568 case ALGORITHM_ROTATING_N_RESTART:
2569 /* Same a left_asymmetric, by first stripe is
2570 * D D D P Q rather than
2574 pd_idx = raid_disks - 1 - 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 */
2584 case ALGORITHM_ROTATING_N_CONTINUE:
2585 /* Same as left_symmetric but Q is before P */
2586 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2587 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2588 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2592 case ALGORITHM_LEFT_ASYMMETRIC_6:
2593 /* RAID5 left_asymmetric, with Q on last device */
2594 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2595 if (*dd_idx >= pd_idx)
2597 qd_idx = raid_disks - 1;
2600 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2601 pd_idx = sector_div(stripe2, raid_disks-1);
2602 if (*dd_idx >= pd_idx)
2604 qd_idx = raid_disks - 1;
2607 case ALGORITHM_LEFT_SYMMETRIC_6:
2608 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2609 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2610 qd_idx = raid_disks - 1;
2613 case ALGORITHM_RIGHT_SYMMETRIC_6:
2614 pd_idx = sector_div(stripe2, raid_disks-1);
2615 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2616 qd_idx = raid_disks - 1;
2619 case ALGORITHM_PARITY_0_6:
2622 qd_idx = raid_disks - 1;
2632 sh->pd_idx = pd_idx;
2633 sh->qd_idx = qd_idx;
2634 sh->ddf_layout = ddf_layout;
2637 * Finally, compute the new sector number
2639 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2643 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2645 struct r5conf *conf = sh->raid_conf;
2646 int raid_disks = sh->disks;
2647 int data_disks = raid_disks - conf->max_degraded;
2648 sector_t new_sector = sh->sector, check;
2649 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2650 : conf->chunk_sectors;
2651 int algorithm = previous ? conf->prev_algo
2655 sector_t chunk_number;
2656 int dummy1, dd_idx = i;
2658 struct stripe_head sh2;
2660 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2661 stripe = new_sector;
2663 if (i == sh->pd_idx)
2665 switch(conf->level) {
2668 switch (algorithm) {
2669 case ALGORITHM_LEFT_ASYMMETRIC:
2670 case ALGORITHM_RIGHT_ASYMMETRIC:
2674 case ALGORITHM_LEFT_SYMMETRIC:
2675 case ALGORITHM_RIGHT_SYMMETRIC:
2678 i -= (sh->pd_idx + 1);
2680 case ALGORITHM_PARITY_0:
2683 case ALGORITHM_PARITY_N:
2690 if (i == sh->qd_idx)
2691 return 0; /* It is the Q disk */
2692 switch (algorithm) {
2693 case ALGORITHM_LEFT_ASYMMETRIC:
2694 case ALGORITHM_RIGHT_ASYMMETRIC:
2695 case ALGORITHM_ROTATING_ZERO_RESTART:
2696 case ALGORITHM_ROTATING_N_RESTART:
2697 if (sh->pd_idx == raid_disks-1)
2698 i--; /* Q D D D P */
2699 else if (i > sh->pd_idx)
2700 i -= 2; /* D D P Q D */
2702 case ALGORITHM_LEFT_SYMMETRIC:
2703 case ALGORITHM_RIGHT_SYMMETRIC:
2704 if (sh->pd_idx == raid_disks-1)
2705 i--; /* Q D D D P */
2710 i -= (sh->pd_idx + 2);
2713 case ALGORITHM_PARITY_0:
2716 case ALGORITHM_PARITY_N:
2718 case ALGORITHM_ROTATING_N_CONTINUE:
2719 /* Like left_symmetric, but P is before Q */
2720 if (sh->pd_idx == 0)
2721 i--; /* P D D D Q */
2726 i -= (sh->pd_idx + 1);
2729 case ALGORITHM_LEFT_ASYMMETRIC_6:
2730 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2734 case ALGORITHM_LEFT_SYMMETRIC_6:
2735 case ALGORITHM_RIGHT_SYMMETRIC_6:
2737 i += data_disks + 1;
2738 i -= (sh->pd_idx + 1);
2740 case ALGORITHM_PARITY_0_6:
2749 chunk_number = stripe * data_disks + i;
2750 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2752 check = raid5_compute_sector(conf, r_sector,
2753 previous, &dummy1, &sh2);
2754 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2755 || sh2.qd_idx != sh->qd_idx) {
2756 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2757 mdname(conf->mddev));
2764 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2765 int rcw, int expand)
2767 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2768 struct r5conf *conf = sh->raid_conf;
2769 int level = conf->level;
2773 for (i = disks; i--; ) {
2774 struct r5dev *dev = &sh->dev[i];
2777 set_bit(R5_LOCKED, &dev->flags);
2778 set_bit(R5_Wantdrain, &dev->flags);
2780 clear_bit(R5_UPTODATE, &dev->flags);
2784 /* if we are not expanding this is a proper write request, and
2785 * there will be bios with new data to be drained into the
2790 /* False alarm, nothing to do */
2792 sh->reconstruct_state = reconstruct_state_drain_run;
2793 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2795 sh->reconstruct_state = reconstruct_state_run;
2797 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2799 if (s->locked + conf->max_degraded == disks)
2800 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2801 atomic_inc(&conf->pending_full_writes);
2804 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2805 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2807 for (i = disks; i--; ) {
2808 struct r5dev *dev = &sh->dev[i];
2813 (test_bit(R5_UPTODATE, &dev->flags) ||
2814 test_bit(R5_Wantcompute, &dev->flags))) {
2815 set_bit(R5_Wantdrain, &dev->flags);
2816 set_bit(R5_LOCKED, &dev->flags);
2817 clear_bit(R5_UPTODATE, &dev->flags);
2822 /* False alarm - nothing to do */
2824 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2825 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2826 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2827 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2830 /* keep the parity disk(s) locked while asynchronous operations
2833 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2834 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2838 int qd_idx = sh->qd_idx;
2839 struct r5dev *dev = &sh->dev[qd_idx];
2841 set_bit(R5_LOCKED, &dev->flags);
2842 clear_bit(R5_UPTODATE, &dev->flags);
2846 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2847 __func__, (unsigned long long)sh->sector,
2848 s->locked, s->ops_request);
2852 * Each stripe/dev can have one or more bion attached.
2853 * toread/towrite point to the first in a chain.
2854 * The bi_next chain must be in order.
2856 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2857 int forwrite, int previous)
2860 struct r5conf *conf = sh->raid_conf;
2863 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2864 (unsigned long long)bi->bi_iter.bi_sector,
2865 (unsigned long long)sh->sector);
2868 * If several bio share a stripe. The bio bi_phys_segments acts as a
2869 * reference count to avoid race. The reference count should already be
2870 * increased before this function is called (for example, in
2871 * make_request()), so other bio sharing this stripe will not free the
2872 * stripe. If a stripe is owned by one stripe, the stripe lock will
2875 spin_lock_irq(&sh->stripe_lock);
2876 /* Don't allow new IO added to stripes in batch list */
2880 bip = &sh->dev[dd_idx].towrite;
2884 bip = &sh->dev[dd_idx].toread;
2885 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2886 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2888 bip = & (*bip)->bi_next;
2890 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2893 if (!forwrite || previous)
2894 clear_bit(STRIPE_BATCH_READY, &sh->state);
2896 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2900 raid5_inc_bi_active_stripes(bi);
2903 /* check if page is covered */
2904 sector_t sector = sh->dev[dd_idx].sector;
2905 for (bi=sh->dev[dd_idx].towrite;
2906 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2907 bi && bi->bi_iter.bi_sector <= sector;
2908 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2909 if (bio_end_sector(bi) >= sector)
2910 sector = bio_end_sector(bi);
2912 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2913 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
2914 sh->overwrite_disks++;
2917 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2918 (unsigned long long)(*bip)->bi_iter.bi_sector,
2919 (unsigned long long)sh->sector, dd_idx);
2920 spin_unlock_irq(&sh->stripe_lock);
2922 if (conf->mddev->bitmap && firstwrite) {
2923 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2925 sh->bm_seq = conf->seq_flush+1;
2926 set_bit(STRIPE_BIT_DELAY, &sh->state);
2929 if (stripe_can_batch(sh))
2930 stripe_add_to_batch_list(conf, sh);
2934 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2935 spin_unlock_irq(&sh->stripe_lock);
2939 static void end_reshape(struct r5conf *conf);
2941 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2942 struct stripe_head *sh)
2944 int sectors_per_chunk =
2945 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2947 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2948 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2950 raid5_compute_sector(conf,
2951 stripe * (disks - conf->max_degraded)
2952 *sectors_per_chunk + chunk_offset,
2958 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2959 struct stripe_head_state *s, int disks,
2960 struct bio **return_bi)
2963 BUG_ON(sh->batch_head);
2964 for (i = disks; i--; ) {
2968 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2969 struct md_rdev *rdev;
2971 rdev = rcu_dereference(conf->disks[i].rdev);
2972 if (rdev && test_bit(In_sync, &rdev->flags))
2973 atomic_inc(&rdev->nr_pending);
2978 if (!rdev_set_badblocks(
2982 md_error(conf->mddev, rdev);
2983 rdev_dec_pending(rdev, conf->mddev);
2986 spin_lock_irq(&sh->stripe_lock);
2987 /* fail all writes first */
2988 bi = sh->dev[i].towrite;
2989 sh->dev[i].towrite = NULL;
2990 sh->overwrite_disks = 0;
2991 spin_unlock_irq(&sh->stripe_lock);
2995 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2996 wake_up(&conf->wait_for_overlap);
2998 while (bi && bi->bi_iter.bi_sector <
2999 sh->dev[i].sector + STRIPE_SECTORS) {
3000 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3001 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3002 if (!raid5_dec_bi_active_stripes(bi)) {
3003 md_write_end(conf->mddev);
3004 bi->bi_next = *return_bi;
3010 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3011 STRIPE_SECTORS, 0, 0);
3013 /* and fail all 'written' */
3014 bi = sh->dev[i].written;
3015 sh->dev[i].written = NULL;
3016 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3017 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3018 sh->dev[i].page = sh->dev[i].orig_page;
3021 if (bi) bitmap_end = 1;
3022 while (bi && bi->bi_iter.bi_sector <
3023 sh->dev[i].sector + STRIPE_SECTORS) {
3024 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3025 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3026 if (!raid5_dec_bi_active_stripes(bi)) {
3027 md_write_end(conf->mddev);
3028 bi->bi_next = *return_bi;
3034 /* fail any reads if this device is non-operational and
3035 * the data has not reached the cache yet.
3037 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3038 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3039 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3040 spin_lock_irq(&sh->stripe_lock);
3041 bi = sh->dev[i].toread;
3042 sh->dev[i].toread = NULL;
3043 spin_unlock_irq(&sh->stripe_lock);
3044 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3045 wake_up(&conf->wait_for_overlap);
3046 while (bi && bi->bi_iter.bi_sector <
3047 sh->dev[i].sector + STRIPE_SECTORS) {
3048 struct bio *nextbi =
3049 r5_next_bio(bi, sh->dev[i].sector);
3050 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3051 if (!raid5_dec_bi_active_stripes(bi)) {
3052 bi->bi_next = *return_bi;
3059 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3060 STRIPE_SECTORS, 0, 0);
3061 /* If we were in the middle of a write the parity block might
3062 * still be locked - so just clear all R5_LOCKED flags
3064 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3067 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3068 if (atomic_dec_and_test(&conf->pending_full_writes))
3069 md_wakeup_thread(conf->mddev->thread);
3073 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3074 struct stripe_head_state *s)
3079 BUG_ON(sh->batch_head);
3080 clear_bit(STRIPE_SYNCING, &sh->state);
3081 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3082 wake_up(&conf->wait_for_overlap);
3085 /* There is nothing more to do for sync/check/repair.
3086 * Don't even need to abort as that is handled elsewhere
3087 * if needed, and not always wanted e.g. if there is a known
3089 * For recover/replace we need to record a bad block on all
3090 * non-sync devices, or abort the recovery
3092 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3093 /* During recovery devices cannot be removed, so
3094 * locking and refcounting of rdevs is not needed
3096 for (i = 0; i < conf->raid_disks; i++) {
3097 struct md_rdev *rdev = conf->disks[i].rdev;
3099 && !test_bit(Faulty, &rdev->flags)
3100 && !test_bit(In_sync, &rdev->flags)
3101 && !rdev_set_badblocks(rdev, sh->sector,
3104 rdev = conf->disks[i].replacement;
3106 && !test_bit(Faulty, &rdev->flags)
3107 && !test_bit(In_sync, &rdev->flags)
3108 && !rdev_set_badblocks(rdev, sh->sector,
3113 conf->recovery_disabled =
3114 conf->mddev->recovery_disabled;
3116 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3119 static int want_replace(struct stripe_head *sh, int disk_idx)
3121 struct md_rdev *rdev;
3123 /* Doing recovery so rcu locking not required */
3124 rdev = sh->raid_conf->disks[disk_idx].replacement;
3126 && !test_bit(Faulty, &rdev->flags)
3127 && !test_bit(In_sync, &rdev->flags)
3128 && (rdev->recovery_offset <= sh->sector
3129 || rdev->mddev->recovery_cp <= sh->sector))
3135 /* fetch_block - checks the given member device to see if its data needs
3136 * to be read or computed to satisfy a request.
3138 * Returns 1 when no more member devices need to be checked, otherwise returns
3139 * 0 to tell the loop in handle_stripe_fill to continue
3142 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3143 int disk_idx, int disks)
3145 struct r5dev *dev = &sh->dev[disk_idx];
3146 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3147 &sh->dev[s->failed_num[1]] };
3151 if (test_bit(R5_LOCKED, &dev->flags) ||
3152 test_bit(R5_UPTODATE, &dev->flags))
3153 /* No point reading this as we already have it or have
3154 * decided to get it.
3159 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3160 /* We need this block to directly satisfy a request */
3163 if (s->syncing || s->expanding ||
3164 (s->replacing && want_replace(sh, disk_idx)))
3165 /* When syncing, or expanding we read everything.
3166 * When replacing, we need the replaced block.
3170 if ((s->failed >= 1 && fdev[0]->toread) ||
3171 (s->failed >= 2 && fdev[1]->toread))
3172 /* If we want to read from a failed device, then
3173 * we need to actually read every other device.
3177 /* Sometimes neither read-modify-write nor reconstruct-write
3178 * cycles can work. In those cases we read every block we
3179 * can. Then the parity-update is certain to have enough to
3181 * This can only be a problem when we need to write something,
3182 * and some device has failed. If either of those tests
3183 * fail we need look no further.
3185 if (!s->failed || !s->to_write)
3188 if (test_bit(R5_Insync, &dev->flags) &&
3189 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3190 /* Pre-reads at not permitted until after short delay
3191 * to gather multiple requests. However if this
3192 * device is no Insync, the block could only be be computed
3193 * and there is no need to delay that.
3197 for (i = 0; i < s->failed; i++) {
3198 if (fdev[i]->towrite &&
3199 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3200 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3201 /* If we have a partial write to a failed
3202 * device, then we will need to reconstruct
3203 * the content of that device, so all other
3204 * devices must be read.
3209 /* If we are forced to do a reconstruct-write, either because
3210 * the current RAID6 implementation only supports that, or
3211 * or because parity cannot be trusted and we are currently
3212 * recovering it, there is extra need to be careful.
3213 * If one of the devices that we would need to read, because
3214 * it is not being overwritten (and maybe not written at all)
3215 * is missing/faulty, then we need to read everything we can.
3217 if (sh->raid_conf->level != 6 &&
3218 sh->sector < sh->raid_conf->mddev->recovery_cp)
3219 /* reconstruct-write isn't being forced */
3221 for (i = 0; i < s->failed; i++) {
3222 if (!test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3223 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3230 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3231 int disk_idx, int disks)
3233 struct r5dev *dev = &sh->dev[disk_idx];
3235 /* is the data in this block needed, and can we get it? */
3236 if (need_this_block(sh, s, disk_idx, disks)) {
3237 /* we would like to get this block, possibly by computing it,
3238 * otherwise read it if the backing disk is insync
3240 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3241 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3242 if ((s->uptodate == disks - 1) &&
3243 (s->failed && (disk_idx == s->failed_num[0] ||
3244 disk_idx == s->failed_num[1]))) {
3245 /* have disk failed, and we're requested to fetch it;
3248 pr_debug("Computing stripe %llu block %d\n",
3249 (unsigned long long)sh->sector, disk_idx);
3250 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3251 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3252 set_bit(R5_Wantcompute, &dev->flags);
3253 sh->ops.target = disk_idx;
3254 sh->ops.target2 = -1; /* no 2nd target */
3256 /* Careful: from this point on 'uptodate' is in the eye
3257 * of raid_run_ops which services 'compute' operations
3258 * before writes. R5_Wantcompute flags a block that will
3259 * be R5_UPTODATE by the time it is needed for a
3260 * subsequent operation.
3264 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3265 /* Computing 2-failure is *very* expensive; only
3266 * do it if failed >= 2
3269 for (other = disks; other--; ) {
3270 if (other == disk_idx)
3272 if (!test_bit(R5_UPTODATE,
3273 &sh->dev[other].flags))
3277 pr_debug("Computing stripe %llu blocks %d,%d\n",
3278 (unsigned long long)sh->sector,
3280 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3281 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3282 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3283 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3284 sh->ops.target = disk_idx;
3285 sh->ops.target2 = other;
3289 } else if (test_bit(R5_Insync, &dev->flags)) {
3290 set_bit(R5_LOCKED, &dev->flags);
3291 set_bit(R5_Wantread, &dev->flags);
3293 pr_debug("Reading block %d (sync=%d)\n",
3294 disk_idx, s->syncing);
3302 * handle_stripe_fill - read or compute data to satisfy pending requests.
3304 static void handle_stripe_fill(struct stripe_head *sh,
3305 struct stripe_head_state *s,
3310 BUG_ON(sh->batch_head);
3311 /* look for blocks to read/compute, skip this if a compute
3312 * is already in flight, or if the stripe contents are in the
3313 * midst of changing due to a write
3315 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3316 !sh->reconstruct_state)
3317 for (i = disks; i--; )
3318 if (fetch_block(sh, s, i, disks))
3320 set_bit(STRIPE_HANDLE, &sh->state);
3323 /* handle_stripe_clean_event
3324 * any written block on an uptodate or failed drive can be returned.
3325 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3326 * never LOCKED, so we don't need to test 'failed' directly.
3328 static void handle_stripe_clean_event(struct r5conf *conf,
3329 struct stripe_head *sh, int disks, struct bio **return_bi)
3333 int discard_pending = 0;
3334 struct stripe_head *head_sh = sh;
3335 bool do_endio = false;
3338 for (i = disks; i--; )
3339 if (sh->dev[i].written) {
3341 if (!test_bit(R5_LOCKED, &dev->flags) &&
3342 (test_bit(R5_UPTODATE, &dev->flags) ||
3343 test_bit(R5_Discard, &dev->flags) ||
3344 test_bit(R5_SkipCopy, &dev->flags))) {
3345 /* We can return any write requests */
3346 struct bio *wbi, *wbi2;
3347 pr_debug("Return write for disc %d\n", i);
3348 if (test_and_clear_bit(R5_Discard, &dev->flags))
3349 clear_bit(R5_UPTODATE, &dev->flags);
3350 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3351 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3356 dev->page = dev->orig_page;
3358 dev->written = NULL;
3359 while (wbi && wbi->bi_iter.bi_sector <
3360 dev->sector + STRIPE_SECTORS) {
3361 wbi2 = r5_next_bio(wbi, dev->sector);
3362 if (!raid5_dec_bi_active_stripes(wbi)) {
3363 md_write_end(conf->mddev);
3364 wbi->bi_next = *return_bi;
3369 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3371 !test_bit(STRIPE_DEGRADED, &sh->state),
3373 if (head_sh->batch_head) {
3374 sh = list_first_entry(&sh->batch_list,
3377 if (sh != head_sh) {
3384 } else if (test_bit(R5_Discard, &dev->flags))
3385 discard_pending = 1;
3386 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3387 WARN_ON(dev->page != dev->orig_page);
3389 if (!discard_pending &&
3390 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3391 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3392 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3393 if (sh->qd_idx >= 0) {
3394 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3395 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3397 /* now that discard is done we can proceed with any sync */
3398 clear_bit(STRIPE_DISCARD, &sh->state);
3400 * SCSI discard will change some bio fields and the stripe has
3401 * no updated data, so remove it from hash list and the stripe
3402 * will be reinitialized
3404 spin_lock_irq(&conf->device_lock);
3407 if (head_sh->batch_head) {
3408 sh = list_first_entry(&sh->batch_list,
3409 struct stripe_head, batch_list);
3413 spin_unlock_irq(&conf->device_lock);
3416 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3417 set_bit(STRIPE_HANDLE, &sh->state);
3421 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3422 if (atomic_dec_and_test(&conf->pending_full_writes))
3423 md_wakeup_thread(conf->mddev->thread);
3425 if (!head_sh->batch_head || !do_endio)
3427 for (i = 0; i < head_sh->disks; i++) {
3428 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
3431 while (!list_empty(&head_sh->batch_list)) {
3433 sh = list_first_entry(&head_sh->batch_list,
3434 struct stripe_head, batch_list);
3435 list_del_init(&sh->batch_list);
3437 sh->state = head_sh->state & (~((1 << STRIPE_ACTIVE) |
3438 (1 << STRIPE_PREREAD_ACTIVE)));
3439 sh->check_state = head_sh->check_state;
3440 sh->reconstruct_state = head_sh->reconstruct_state;
3441 for (i = 0; i < sh->disks; i++) {
3442 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3444 sh->dev[i].flags = head_sh->dev[i].flags;
3447 spin_lock_irq(&sh->stripe_lock);
3448 sh->batch_head = NULL;
3449 spin_unlock_irq(&sh->stripe_lock);
3453 spin_lock_irq(&head_sh->stripe_lock);
3454 head_sh->batch_head = NULL;
3455 spin_unlock_irq(&head_sh->stripe_lock);
3456 wake_up_nr(&conf->wait_for_overlap, wakeup_nr);
3459 static void handle_stripe_dirtying(struct r5conf *conf,
3460 struct stripe_head *sh,
3461 struct stripe_head_state *s,
3464 int rmw = 0, rcw = 0, i;
3465 sector_t recovery_cp = conf->mddev->recovery_cp;
3467 /* RAID6 requires 'rcw' in current implementation.
3468 * Otherwise, check whether resync is now happening or should start.
3469 * If yes, then the array is dirty (after unclean shutdown or
3470 * initial creation), so parity in some stripes might be inconsistent.
3471 * In this case, we need to always do reconstruct-write, to ensure
3472 * that in case of drive failure or read-error correction, we
3473 * generate correct data from the parity.
3475 if (conf->max_degraded == 2 ||
3476 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3478 /* Calculate the real rcw later - for now make it
3479 * look like rcw is cheaper
3482 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
3483 conf->max_degraded, (unsigned long long)recovery_cp,
3484 (unsigned long long)sh->sector);
3485 } else for (i = disks; i--; ) {
3486 /* would I have to read this buffer for read_modify_write */
3487 struct r5dev *dev = &sh->dev[i];
3488 if ((dev->towrite || i == sh->pd_idx) &&
3489 !test_bit(R5_LOCKED, &dev->flags) &&
3490 !(test_bit(R5_UPTODATE, &dev->flags) ||
3491 test_bit(R5_Wantcompute, &dev->flags))) {
3492 if (test_bit(R5_Insync, &dev->flags))
3495 rmw += 2*disks; /* cannot read it */
3497 /* Would I have to read this buffer for reconstruct_write */
3498 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
3499 !test_bit(R5_LOCKED, &dev->flags) &&
3500 !(test_bit(R5_UPTODATE, &dev->flags) ||
3501 test_bit(R5_Wantcompute, &dev->flags))) {
3502 if (test_bit(R5_Insync, &dev->flags))
3508 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3509 (unsigned long long)sh->sector, rmw, rcw);
3510 set_bit(STRIPE_HANDLE, &sh->state);
3511 if (rmw < rcw && rmw > 0) {
3512 /* prefer read-modify-write, but need to get some data */
3513 if (conf->mddev->queue)
3514 blk_add_trace_msg(conf->mddev->queue,
3515 "raid5 rmw %llu %d",
3516 (unsigned long long)sh->sector, rmw);
3517 for (i = disks; i--; ) {
3518 struct r5dev *dev = &sh->dev[i];
3519 if ((dev->towrite || i == sh->pd_idx) &&
3520 !test_bit(R5_LOCKED, &dev->flags) &&
3521 !(test_bit(R5_UPTODATE, &dev->flags) ||
3522 test_bit(R5_Wantcompute, &dev->flags)) &&
3523 test_bit(R5_Insync, &dev->flags)) {
3524 if (test_bit(STRIPE_PREREAD_ACTIVE,
3526 pr_debug("Read_old block %d for r-m-w\n",
3528 set_bit(R5_LOCKED, &dev->flags);
3529 set_bit(R5_Wantread, &dev->flags);
3532 set_bit(STRIPE_DELAYED, &sh->state);
3533 set_bit(STRIPE_HANDLE, &sh->state);
3538 if (rcw <= rmw && rcw > 0) {
3539 /* want reconstruct write, but need to get some data */
3542 for (i = disks; i--; ) {
3543 struct r5dev *dev = &sh->dev[i];
3544 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3545 i != sh->pd_idx && i != sh->qd_idx &&
3546 !test_bit(R5_LOCKED, &dev->flags) &&
3547 !(test_bit(R5_UPTODATE, &dev->flags) ||
3548 test_bit(R5_Wantcompute, &dev->flags))) {
3550 if (test_bit(R5_Insync, &dev->flags) &&
3551 test_bit(STRIPE_PREREAD_ACTIVE,
3553 pr_debug("Read_old block "
3554 "%d for Reconstruct\n", i);
3555 set_bit(R5_LOCKED, &dev->flags);
3556 set_bit(R5_Wantread, &dev->flags);
3560 set_bit(STRIPE_DELAYED, &sh->state);
3561 set_bit(STRIPE_HANDLE, &sh->state);
3565 if (rcw && conf->mddev->queue)
3566 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3567 (unsigned long long)sh->sector,
3568 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3571 if (rcw > disks && rmw > disks &&
3572 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3573 set_bit(STRIPE_DELAYED, &sh->state);
3575 /* now if nothing is locked, and if we have enough data,
3576 * we can start a write request
3578 /* since handle_stripe can be called at any time we need to handle the
3579 * case where a compute block operation has been submitted and then a
3580 * subsequent call wants to start a write request. raid_run_ops only
3581 * handles the case where compute block and reconstruct are requested
3582 * simultaneously. If this is not the case then new writes need to be
3583 * held off until the compute completes.
3585 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3586 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3587 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3588 schedule_reconstruction(sh, s, rcw == 0, 0);
3591 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3592 struct stripe_head_state *s, int disks)
3594 struct r5dev *dev = NULL;
3596 BUG_ON(sh->batch_head);
3597 set_bit(STRIPE_HANDLE, &sh->state);
3599 switch (sh->check_state) {
3600 case check_state_idle:
3601 /* start a new check operation if there are no failures */
3602 if (s->failed == 0) {
3603 BUG_ON(s->uptodate != disks);
3604 sh->check_state = check_state_run;
3605 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3606 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3610 dev = &sh->dev[s->failed_num[0]];
3612 case check_state_compute_result:
3613 sh->check_state = check_state_idle;
3615 dev = &sh->dev[sh->pd_idx];
3617 /* check that a write has not made the stripe insync */
3618 if (test_bit(STRIPE_INSYNC, &sh->state))
3621 /* either failed parity check, or recovery is happening */
3622 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3623 BUG_ON(s->uptodate != disks);
3625 set_bit(R5_LOCKED, &dev->flags);
3627 set_bit(R5_Wantwrite, &dev->flags);
3629 clear_bit(STRIPE_DEGRADED, &sh->state);
3630 set_bit(STRIPE_INSYNC, &sh->state);
3632 case check_state_run:
3633 break; /* we will be called again upon completion */
3634 case check_state_check_result:
3635 sh->check_state = check_state_idle;
3637 /* if a failure occurred during the check operation, leave
3638 * STRIPE_INSYNC not set and let the stripe be handled again
3643 /* handle a successful check operation, if parity is correct
3644 * we are done. Otherwise update the mismatch count and repair
3645 * parity if !MD_RECOVERY_CHECK
3647 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3648 /* parity is correct (on disc,
3649 * not in buffer any more)
3651 set_bit(STRIPE_INSYNC, &sh->state);
3653 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3654 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3655 /* don't try to repair!! */
3656 set_bit(STRIPE_INSYNC, &sh->state);
3658 sh->check_state = check_state_compute_run;
3659 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3660 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3661 set_bit(R5_Wantcompute,
3662 &sh->dev[sh->pd_idx].flags);
3663 sh->ops.target = sh->pd_idx;
3664 sh->ops.target2 = -1;
3669 case check_state_compute_run:
3672 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3673 __func__, sh->check_state,
3674 (unsigned long long) sh->sector);
3679 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3680 struct stripe_head_state *s,
3683 int pd_idx = sh->pd_idx;
3684 int qd_idx = sh->qd_idx;
3687 BUG_ON(sh->batch_head);
3688 set_bit(STRIPE_HANDLE, &sh->state);
3690 BUG_ON(s->failed > 2);
3692 /* Want to check and possibly repair P and Q.
3693 * However there could be one 'failed' device, in which
3694 * case we can only check one of them, possibly using the
3695 * other to generate missing data
3698 switch (sh->check_state) {
3699 case check_state_idle:
3700 /* start a new check operation if there are < 2 failures */
3701 if (s->failed == s->q_failed) {
3702 /* The only possible failed device holds Q, so it
3703 * makes sense to check P (If anything else were failed,
3704 * we would have used P to recreate it).
3706 sh->check_state = check_state_run;
3708 if (!s->q_failed && s->failed < 2) {
3709 /* Q is not failed, and we didn't use it to generate
3710 * anything, so it makes sense to check it
3712 if (sh->check_state == check_state_run)
3713 sh->check_state = check_state_run_pq;
3715 sh->check_state = check_state_run_q;
3718 /* discard potentially stale zero_sum_result */
3719 sh->ops.zero_sum_result = 0;
3721 if (sh->check_state == check_state_run) {
3722 /* async_xor_zero_sum destroys the contents of P */
3723 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3726 if (sh->check_state >= check_state_run &&
3727 sh->check_state <= check_state_run_pq) {
3728 /* async_syndrome_zero_sum preserves P and Q, so
3729 * no need to mark them !uptodate here
3731 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3735 /* we have 2-disk failure */
3736 BUG_ON(s->failed != 2);
3738 case check_state_compute_result:
3739 sh->check_state = check_state_idle;
3741 /* check that a write has not made the stripe insync */
3742 if (test_bit(STRIPE_INSYNC, &sh->state))
3745 /* now write out any block on a failed drive,
3746 * or P or Q if they were recomputed
3748 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3749 if (s->failed == 2) {
3750 dev = &sh->dev[s->failed_num[1]];
3752 set_bit(R5_LOCKED, &dev->flags);
3753 set_bit(R5_Wantwrite, &dev->flags);
3755 if (s->failed >= 1) {
3756 dev = &sh->dev[s->failed_num[0]];
3758 set_bit(R5_LOCKED, &dev->flags);
3759 set_bit(R5_Wantwrite, &dev->flags);
3761 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3762 dev = &sh->dev[pd_idx];
3764 set_bit(R5_LOCKED, &dev->flags);
3765 set_bit(R5_Wantwrite, &dev->flags);
3767 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3768 dev = &sh->dev[qd_idx];
3770 set_bit(R5_LOCKED, &dev->flags);
3771 set_bit(R5_Wantwrite, &dev->flags);
3773 clear_bit(STRIPE_DEGRADED, &sh->state);
3775 set_bit(STRIPE_INSYNC, &sh->state);
3777 case check_state_run:
3778 case check_state_run_q:
3779 case check_state_run_pq:
3780 break; /* we will be called again upon completion */
3781 case check_state_check_result:
3782 sh->check_state = check_state_idle;
3784 /* handle a successful check operation, if parity is correct
3785 * we are done. Otherwise update the mismatch count and repair
3786 * parity if !MD_RECOVERY_CHECK
3788 if (sh->ops.zero_sum_result == 0) {
3789 /* both parities are correct */
3791 set_bit(STRIPE_INSYNC, &sh->state);
3793 /* in contrast to the raid5 case we can validate
3794 * parity, but still have a failure to write
3797 sh->check_state = check_state_compute_result;
3798 /* Returning at this point means that we may go
3799 * off and bring p and/or q uptodate again so
3800 * we make sure to check zero_sum_result again
3801 * to verify if p or q need writeback
3805 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3806 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3807 /* don't try to repair!! */
3808 set_bit(STRIPE_INSYNC, &sh->state);
3810 int *target = &sh->ops.target;
3812 sh->ops.target = -1;
3813 sh->ops.target2 = -1;
3814 sh->check_state = check_state_compute_run;
3815 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3816 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3817 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3818 set_bit(R5_Wantcompute,
3819 &sh->dev[pd_idx].flags);
3821 target = &sh->ops.target2;
3824 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3825 set_bit(R5_Wantcompute,
3826 &sh->dev[qd_idx].flags);
3833 case check_state_compute_run:
3836 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3837 __func__, sh->check_state,
3838 (unsigned long long) sh->sector);
3843 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3847 /* We have read all the blocks in this stripe and now we need to
3848 * copy some of them into a target stripe for expand.
3850 struct dma_async_tx_descriptor *tx = NULL;
3851 BUG_ON(sh->batch_head);
3852 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3853 for (i = 0; i < sh->disks; i++)
3854 if (i != sh->pd_idx && i != sh->qd_idx) {
3856 struct stripe_head *sh2;
3857 struct async_submit_ctl submit;
3859 sector_t bn = compute_blocknr(sh, i, 1);
3860 sector_t s = raid5_compute_sector(conf, bn, 0,
3862 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3864 /* so far only the early blocks of this stripe
3865 * have been requested. When later blocks
3866 * get requested, we will try again
3869 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3870 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3871 /* must have already done this block */
3872 release_stripe(sh2);
3876 /* place all the copies on one channel */
3877 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3878 tx = async_memcpy(sh2->dev[dd_idx].page,
3879 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3882 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3883 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3884 for (j = 0; j < conf->raid_disks; j++)
3885 if (j != sh2->pd_idx &&
3887 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3889 if (j == conf->raid_disks) {
3890 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3891 set_bit(STRIPE_HANDLE, &sh2->state);
3893 release_stripe(sh2);
3896 /* done submitting copies, wait for them to complete */
3897 async_tx_quiesce(&tx);
3901 * handle_stripe - do things to a stripe.
3903 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3904 * state of various bits to see what needs to be done.
3906 * return some read requests which now have data
3907 * return some write requests which are safely on storage
3908 * schedule a read on some buffers
3909 * schedule a write of some buffers
3910 * return confirmation of parity correctness
3914 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3916 struct r5conf *conf = sh->raid_conf;
3917 int disks = sh->disks;
3920 int do_recovery = 0;
3922 memset(s, 0, sizeof(*s));
3924 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3925 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3926 s->failed_num[0] = -1;
3927 s->failed_num[1] = -1;
3929 /* Now to look around and see what can be done */
3931 for (i=disks; i--; ) {
3932 struct md_rdev *rdev;
3939 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3941 dev->toread, dev->towrite, dev->written);
3942 /* maybe we can reply to a read
3944 * new wantfill requests are only permitted while
3945 * ops_complete_biofill is guaranteed to be inactive
3947 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3948 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3949 set_bit(R5_Wantfill, &dev->flags);
3951 /* now count some things */
3952 if (test_bit(R5_LOCKED, &dev->flags))
3954 if (test_bit(R5_UPTODATE, &dev->flags))
3956 if (test_bit(R5_Wantcompute, &dev->flags)) {
3958 BUG_ON(s->compute > 2);
3961 if (test_bit(R5_Wantfill, &dev->flags))
3963 else if (dev->toread)
3967 if (!test_bit(R5_OVERWRITE, &dev->flags))
3972 /* Prefer to use the replacement for reads, but only
3973 * if it is recovered enough and has no bad blocks.
3975 rdev = rcu_dereference(conf->disks[i].replacement);
3976 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3977 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3978 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3979 &first_bad, &bad_sectors))
3980 set_bit(R5_ReadRepl, &dev->flags);
3983 set_bit(R5_NeedReplace, &dev->flags);
3984 rdev = rcu_dereference(conf->disks[i].rdev);
3985 clear_bit(R5_ReadRepl, &dev->flags);
3987 if (rdev && test_bit(Faulty, &rdev->flags))
3990 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3991 &first_bad, &bad_sectors);
3992 if (s->blocked_rdev == NULL
3993 && (test_bit(Blocked, &rdev->flags)
3996 set_bit(BlockedBadBlocks,
3998 s->blocked_rdev = rdev;
3999 atomic_inc(&rdev->nr_pending);
4002 clear_bit(R5_Insync, &dev->flags);
4006 /* also not in-sync */
4007 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4008 test_bit(R5_UPTODATE, &dev->flags)) {
4009 /* treat as in-sync, but with a read error
4010 * which we can now try to correct
4012 set_bit(R5_Insync, &dev->flags);
4013 set_bit(R5_ReadError, &dev->flags);
4015 } else if (test_bit(In_sync, &rdev->flags))
4016 set_bit(R5_Insync, &dev->flags);
4017 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4018 /* in sync if before recovery_offset */
4019 set_bit(R5_Insync, &dev->flags);
4020 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4021 test_bit(R5_Expanded, &dev->flags))
4022 /* If we've reshaped into here, we assume it is Insync.
4023 * We will shortly update recovery_offset to make
4026 set_bit(R5_Insync, &dev->flags);
4028 if (test_bit(R5_WriteError, &dev->flags)) {
4029 /* This flag does not apply to '.replacement'
4030 * only to .rdev, so make sure to check that*/
4031 struct md_rdev *rdev2 = rcu_dereference(
4032 conf->disks[i].rdev);
4034 clear_bit(R5_Insync, &dev->flags);
4035 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4036 s->handle_bad_blocks = 1;
4037 atomic_inc(&rdev2->nr_pending);
4039 clear_bit(R5_WriteError, &dev->flags);
4041 if (test_bit(R5_MadeGood, &dev->flags)) {
4042 /* This flag does not apply to '.replacement'
4043 * only to .rdev, so make sure to check that*/
4044 struct md_rdev *rdev2 = rcu_dereference(
4045 conf->disks[i].rdev);
4046 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4047 s->handle_bad_blocks = 1;
4048 atomic_inc(&rdev2->nr_pending);
4050 clear_bit(R5_MadeGood, &dev->flags);
4052 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4053 struct md_rdev *rdev2 = rcu_dereference(
4054 conf->disks[i].replacement);
4055 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4056 s->handle_bad_blocks = 1;
4057 atomic_inc(&rdev2->nr_pending);
4059 clear_bit(R5_MadeGoodRepl, &dev->flags);
4061 if (!test_bit(R5_Insync, &dev->flags)) {
4062 /* The ReadError flag will just be confusing now */
4063 clear_bit(R5_ReadError, &dev->flags);
4064 clear_bit(R5_ReWrite, &dev->flags);
4066 if (test_bit(R5_ReadError, &dev->flags))
4067 clear_bit(R5_Insync, &dev->flags);
4068 if (!test_bit(R5_Insync, &dev->flags)) {
4070 s->failed_num[s->failed] = i;
4072 if (rdev && !test_bit(Faulty, &rdev->flags))
4076 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4077 /* If there is a failed device being replaced,
4078 * we must be recovering.
4079 * else if we are after recovery_cp, we must be syncing
4080 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4081 * else we can only be replacing
4082 * sync and recovery both need to read all devices, and so
4083 * use the same flag.
4086 sh->sector >= conf->mddev->recovery_cp ||
4087 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4095 static int clear_batch_ready(struct stripe_head *sh)
4097 struct stripe_head *tmp;
4098 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4100 spin_lock(&sh->stripe_lock);
4101 if (!sh->batch_head) {
4102 spin_unlock(&sh->stripe_lock);
4107 * this stripe could be added to a batch list before we check
4108 * BATCH_READY, skips it
4110 if (sh->batch_head != sh) {
4111 spin_unlock(&sh->stripe_lock);
4114 spin_lock(&sh->batch_lock);
4115 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4116 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4117 spin_unlock(&sh->batch_lock);
4118 spin_unlock(&sh->stripe_lock);
4121 * BATCH_READY is cleared, no new stripes can be added.
4122 * batch_list can be accessed without lock
4127 static void handle_stripe(struct stripe_head *sh)
4129 struct stripe_head_state s;
4130 struct r5conf *conf = sh->raid_conf;
4133 int disks = sh->disks;
4134 struct r5dev *pdev, *qdev;
4136 clear_bit(STRIPE_HANDLE, &sh->state);
4137 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4138 /* already being handled, ensure it gets handled
4139 * again when current action finishes */
4140 set_bit(STRIPE_HANDLE, &sh->state);
4144 if (clear_batch_ready(sh) ) {
4145 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4149 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4150 spin_lock(&sh->stripe_lock);
4151 /* Cannot process 'sync' concurrently with 'discard' */
4152 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4153 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4154 set_bit(STRIPE_SYNCING, &sh->state);
4155 clear_bit(STRIPE_INSYNC, &sh->state);
4156 clear_bit(STRIPE_REPLACED, &sh->state);
4158 spin_unlock(&sh->stripe_lock);
4160 clear_bit(STRIPE_DELAYED, &sh->state);
4162 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4163 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4164 (unsigned long long)sh->sector, sh->state,
4165 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4166 sh->check_state, sh->reconstruct_state);
4168 analyse_stripe(sh, &s);
4170 if (s.handle_bad_blocks) {
4171 set_bit(STRIPE_HANDLE, &sh->state);
4175 if (unlikely(s.blocked_rdev)) {
4176 if (s.syncing || s.expanding || s.expanded ||
4177 s.replacing || s.to_write || s.written) {
4178 set_bit(STRIPE_HANDLE, &sh->state);
4181 /* There is nothing for the blocked_rdev to block */
4182 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4183 s.blocked_rdev = NULL;
4186 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4187 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4188 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4191 pr_debug("locked=%d uptodate=%d to_read=%d"
4192 " to_write=%d failed=%d failed_num=%d,%d\n",
4193 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4194 s.failed_num[0], s.failed_num[1]);
4195 /* check if the array has lost more than max_degraded devices and,
4196 * if so, some requests might need to be failed.
4198 if (s.failed > conf->max_degraded) {
4199 sh->check_state = 0;
4200 sh->reconstruct_state = 0;
4201 if (s.to_read+s.to_write+s.written)
4202 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4203 if (s.syncing + s.replacing)
4204 handle_failed_sync(conf, sh, &s);
4207 /* Now we check to see if any write operations have recently
4211 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4213 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4214 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4215 sh->reconstruct_state = reconstruct_state_idle;
4217 /* All the 'written' buffers and the parity block are ready to
4218 * be written back to disk
4220 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4221 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4222 BUG_ON(sh->qd_idx >= 0 &&
4223 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4224 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4225 for (i = disks; i--; ) {
4226 struct r5dev *dev = &sh->dev[i];
4227 if (test_bit(R5_LOCKED, &dev->flags) &&
4228 (i == sh->pd_idx || i == sh->qd_idx ||
4230 pr_debug("Writing block %d\n", i);
4231 set_bit(R5_Wantwrite, &dev->flags);
4236 if (!test_bit(R5_Insync, &dev->flags) ||
4237 ((i == sh->pd_idx || i == sh->qd_idx) &&
4239 set_bit(STRIPE_INSYNC, &sh->state);
4242 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4243 s.dec_preread_active = 1;
4247 * might be able to return some write requests if the parity blocks
4248 * are safe, or on a failed drive
4250 pdev = &sh->dev[sh->pd_idx];
4251 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4252 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4253 qdev = &sh->dev[sh->qd_idx];
4254 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4255 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4259 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4260 && !test_bit(R5_LOCKED, &pdev->flags)
4261 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4262 test_bit(R5_Discard, &pdev->flags))))) &&
4263 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4264 && !test_bit(R5_LOCKED, &qdev->flags)
4265 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4266 test_bit(R5_Discard, &qdev->flags))))))
4267 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4269 /* Now we might consider reading some blocks, either to check/generate
4270 * parity, or to satisfy requests
4271 * or to load a block that is being partially written.
4273 if (s.to_read || s.non_overwrite
4274 || (conf->level == 6 && s.to_write && s.failed)
4275 || (s.syncing && (s.uptodate + s.compute < disks))
4278 handle_stripe_fill(sh, &s, disks);
4280 /* Now to consider new write requests and what else, if anything
4281 * should be read. We do not handle new writes when:
4282 * 1/ A 'write' operation (copy+xor) is already in flight.
4283 * 2/ A 'check' operation is in flight, as it may clobber the parity
4286 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4287 handle_stripe_dirtying(conf, sh, &s, disks);
4289 /* maybe we need to check and possibly fix the parity for this stripe
4290 * Any reads will already have been scheduled, so we just see if enough
4291 * data is available. The parity check is held off while parity
4292 * dependent operations are in flight.
4294 if (sh->check_state ||
4295 (s.syncing && s.locked == 0 &&
4296 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4297 !test_bit(STRIPE_INSYNC, &sh->state))) {
4298 if (conf->level == 6)
4299 handle_parity_checks6(conf, sh, &s, disks);
4301 handle_parity_checks5(conf, sh, &s, disks);
4304 if ((s.replacing || s.syncing) && s.locked == 0
4305 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4306 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4307 /* Write out to replacement devices where possible */
4308 for (i = 0; i < conf->raid_disks; i++)
4309 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4310 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4311 set_bit(R5_WantReplace, &sh->dev[i].flags);
4312 set_bit(R5_LOCKED, &sh->dev[i].flags);
4316 set_bit(STRIPE_INSYNC, &sh->state);
4317 set_bit(STRIPE_REPLACED, &sh->state);
4319 if ((s.syncing || s.replacing) && s.locked == 0 &&
4320 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4321 test_bit(STRIPE_INSYNC, &sh->state)) {
4322 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4323 clear_bit(STRIPE_SYNCING, &sh->state);
4324 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4325 wake_up(&conf->wait_for_overlap);
4328 /* If the failed drives are just a ReadError, then we might need
4329 * to progress the repair/check process
4331 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4332 for (i = 0; i < s.failed; i++) {
4333 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4334 if (test_bit(R5_ReadError, &dev->flags)
4335 && !test_bit(R5_LOCKED, &dev->flags)
4336 && test_bit(R5_UPTODATE, &dev->flags)
4338 if (!test_bit(R5_ReWrite, &dev->flags)) {
4339 set_bit(R5_Wantwrite, &dev->flags);
4340 set_bit(R5_ReWrite, &dev->flags);
4341 set_bit(R5_LOCKED, &dev->flags);
4344 /* let's read it back */
4345 set_bit(R5_Wantread, &dev->flags);
4346 set_bit(R5_LOCKED, &dev->flags);
4352 /* Finish reconstruct operations initiated by the expansion process */
4353 if (sh->reconstruct_state == reconstruct_state_result) {
4354 struct stripe_head *sh_src
4355 = get_active_stripe(conf, sh->sector, 1, 1, 1);
4356 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4357 /* sh cannot be written until sh_src has been read.
4358 * so arrange for sh to be delayed a little
4360 set_bit(STRIPE_DELAYED, &sh->state);
4361 set_bit(STRIPE_HANDLE, &sh->state);
4362 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4364 atomic_inc(&conf->preread_active_stripes);
4365 release_stripe(sh_src);
4369 release_stripe(sh_src);
4371 sh->reconstruct_state = reconstruct_state_idle;
4372 clear_bit(STRIPE_EXPANDING, &sh->state);
4373 for (i = conf->raid_disks; i--; ) {
4374 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4375 set_bit(R5_LOCKED, &sh->dev[i].flags);
4380 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4381 !sh->reconstruct_state) {
4382 /* Need to write out all blocks after computing parity */
4383 sh->disks = conf->raid_disks;
4384 stripe_set_idx(sh->sector, conf, 0, sh);
4385 schedule_reconstruction(sh, &s, 1, 1);
4386 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4387 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4388 atomic_dec(&conf->reshape_stripes);
4389 wake_up(&conf->wait_for_overlap);
4390 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4393 if (s.expanding && s.locked == 0 &&
4394 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4395 handle_stripe_expansion(conf, sh);
4398 /* wait for this device to become unblocked */
4399 if (unlikely(s.blocked_rdev)) {
4400 if (conf->mddev->external)
4401 md_wait_for_blocked_rdev(s.blocked_rdev,
4404 /* Internal metadata will immediately
4405 * be written by raid5d, so we don't
4406 * need to wait here.
4408 rdev_dec_pending(s.blocked_rdev,
4412 if (s.handle_bad_blocks)
4413 for (i = disks; i--; ) {
4414 struct md_rdev *rdev;
4415 struct r5dev *dev = &sh->dev[i];
4416 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4417 /* We own a safe reference to the rdev */
4418 rdev = conf->disks[i].rdev;
4419 if (!rdev_set_badblocks(rdev, sh->sector,
4421 md_error(conf->mddev, rdev);
4422 rdev_dec_pending(rdev, conf->mddev);
4424 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4425 rdev = conf->disks[i].rdev;
4426 rdev_clear_badblocks(rdev, sh->sector,
4428 rdev_dec_pending(rdev, conf->mddev);
4430 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4431 rdev = conf->disks[i].replacement;
4433 /* rdev have been moved down */
4434 rdev = conf->disks[i].rdev;
4435 rdev_clear_badblocks(rdev, sh->sector,
4437 rdev_dec_pending(rdev, conf->mddev);
4442 raid_run_ops(sh, s.ops_request);
4446 if (s.dec_preread_active) {
4447 /* We delay this until after ops_run_io so that if make_request
4448 * is waiting on a flush, it won't continue until the writes
4449 * have actually been submitted.
4451 atomic_dec(&conf->preread_active_stripes);
4452 if (atomic_read(&conf->preread_active_stripes) <
4454 md_wakeup_thread(conf->mddev->thread);
4457 return_io(s.return_bi);
4459 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4462 static void raid5_activate_delayed(struct r5conf *conf)
4464 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4465 while (!list_empty(&conf->delayed_list)) {
4466 struct list_head *l = conf->delayed_list.next;
4467 struct stripe_head *sh;
4468 sh = list_entry(l, struct stripe_head, lru);
4470 clear_bit(STRIPE_DELAYED, &sh->state);
4471 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4472 atomic_inc(&conf->preread_active_stripes);
4473 list_add_tail(&sh->lru, &conf->hold_list);
4474 raid5_wakeup_stripe_thread(sh);
4479 static void activate_bit_delay(struct r5conf *conf,
4480 struct list_head *temp_inactive_list)
4482 /* device_lock is held */
4483 struct list_head head;
4484 list_add(&head, &conf->bitmap_list);
4485 list_del_init(&conf->bitmap_list);
4486 while (!list_empty(&head)) {
4487 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4489 list_del_init(&sh->lru);
4490 atomic_inc(&sh->count);
4491 hash = sh->hash_lock_index;
4492 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4496 static int raid5_congested(struct mddev *mddev, int bits)
4498 struct r5conf *conf = mddev->private;
4500 /* No difference between reads and writes. Just check
4501 * how busy the stripe_cache is
4504 if (conf->inactive_blocked)
4508 if (atomic_read(&conf->empty_inactive_list_nr))
4514 /* We want read requests to align with chunks where possible,
4515 * but write requests don't need to.
4517 static int raid5_mergeable_bvec(struct mddev *mddev,
4518 struct bvec_merge_data *bvm,
4519 struct bio_vec *biovec)
4521 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4523 unsigned int chunk_sectors = mddev->chunk_sectors;
4524 unsigned int bio_sectors = bvm->bi_size >> 9;
4526 if ((bvm->bi_rw & 1) == WRITE)
4527 return biovec->bv_len; /* always allow writes to be mergeable */
4529 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4530 chunk_sectors = mddev->new_chunk_sectors;
4531 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4532 if (max < 0) max = 0;
4533 if (max <= biovec->bv_len && bio_sectors == 0)
4534 return biovec->bv_len;
4539 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4541 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4542 unsigned int chunk_sectors = mddev->chunk_sectors;
4543 unsigned int bio_sectors = bio_sectors(bio);
4545 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4546 chunk_sectors = mddev->new_chunk_sectors;
4547 return chunk_sectors >=
4548 ((sector & (chunk_sectors - 1)) + bio_sectors);
4552 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4553 * later sampled by raid5d.
4555 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4557 unsigned long flags;
4559 spin_lock_irqsave(&conf->device_lock, flags);
4561 bi->bi_next = conf->retry_read_aligned_list;
4562 conf->retry_read_aligned_list = bi;
4564 spin_unlock_irqrestore(&conf->device_lock, flags);
4565 md_wakeup_thread(conf->mddev->thread);
4568 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4572 bi = conf->retry_read_aligned;
4574 conf->retry_read_aligned = NULL;
4577 bi = conf->retry_read_aligned_list;
4579 conf->retry_read_aligned_list = bi->bi_next;
4582 * this sets the active strip count to 1 and the processed
4583 * strip count to zero (upper 8 bits)
4585 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4592 * The "raid5_align_endio" should check if the read succeeded and if it
4593 * did, call bio_endio on the original bio (having bio_put the new bio
4595 * If the read failed..
4597 static void raid5_align_endio(struct bio *bi, int error)
4599 struct bio* raid_bi = bi->bi_private;
4600 struct mddev *mddev;
4601 struct r5conf *conf;
4602 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4603 struct md_rdev *rdev;
4607 rdev = (void*)raid_bi->bi_next;
4608 raid_bi->bi_next = NULL;
4609 mddev = rdev->mddev;
4610 conf = mddev->private;
4612 rdev_dec_pending(rdev, conf->mddev);
4614 if (!error && uptodate) {
4615 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4617 bio_endio(raid_bi, 0);
4618 if (atomic_dec_and_test(&conf->active_aligned_reads))
4619 wake_up(&conf->wait_for_stripe);
4623 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4625 add_bio_to_retry(raid_bi, conf);
4628 static int bio_fits_rdev(struct bio *bi)
4630 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4632 if (bio_sectors(bi) > queue_max_sectors(q))
4634 blk_recount_segments(q, bi);
4635 if (bi->bi_phys_segments > queue_max_segments(q))
4638 if (q->merge_bvec_fn)
4639 /* it's too hard to apply the merge_bvec_fn at this stage,
4647 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4649 struct r5conf *conf = mddev->private;
4651 struct bio* align_bi;
4652 struct md_rdev *rdev;
4653 sector_t end_sector;
4655 if (!in_chunk_boundary(mddev, raid_bio)) {
4656 pr_debug("chunk_aligned_read : non aligned\n");
4660 * use bio_clone_mddev to make a copy of the bio
4662 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4666 * set bi_end_io to a new function, and set bi_private to the
4669 align_bi->bi_end_io = raid5_align_endio;
4670 align_bi->bi_private = raid_bio;
4674 align_bi->bi_iter.bi_sector =
4675 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4678 end_sector = bio_end_sector(align_bi);
4680 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4681 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4682 rdev->recovery_offset < end_sector) {
4683 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4685 (test_bit(Faulty, &rdev->flags) ||
4686 !(test_bit(In_sync, &rdev->flags) ||
4687 rdev->recovery_offset >= end_sector)))
4694 atomic_inc(&rdev->nr_pending);
4696 raid_bio->bi_next = (void*)rdev;
4697 align_bi->bi_bdev = rdev->bdev;
4698 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
4700 if (!bio_fits_rdev(align_bi) ||
4701 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4702 bio_sectors(align_bi),
4703 &first_bad, &bad_sectors)) {
4704 /* too big in some way, or has a known bad block */
4706 rdev_dec_pending(rdev, mddev);
4710 /* No reshape active, so we can trust rdev->data_offset */
4711 align_bi->bi_iter.bi_sector += rdev->data_offset;
4713 spin_lock_irq(&conf->device_lock);
4714 wait_event_lock_irq(conf->wait_for_stripe,
4717 atomic_inc(&conf->active_aligned_reads);
4718 spin_unlock_irq(&conf->device_lock);
4721 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4722 align_bi, disk_devt(mddev->gendisk),
4723 raid_bio->bi_iter.bi_sector);
4724 generic_make_request(align_bi);
4733 /* __get_priority_stripe - get the next stripe to process
4735 * Full stripe writes are allowed to pass preread active stripes up until
4736 * the bypass_threshold is exceeded. In general the bypass_count
4737 * increments when the handle_list is handled before the hold_list; however, it
4738 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4739 * stripe with in flight i/o. The bypass_count will be reset when the
4740 * head of the hold_list has changed, i.e. the head was promoted to the
4743 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4745 struct stripe_head *sh = NULL, *tmp;
4746 struct list_head *handle_list = NULL;
4747 struct r5worker_group *wg = NULL;
4749 if (conf->worker_cnt_per_group == 0) {
4750 handle_list = &conf->handle_list;
4751 } else if (group != ANY_GROUP) {
4752 handle_list = &conf->worker_groups[group].handle_list;
4753 wg = &conf->worker_groups[group];
4756 for (i = 0; i < conf->group_cnt; i++) {
4757 handle_list = &conf->worker_groups[i].handle_list;
4758 wg = &conf->worker_groups[i];
4759 if (!list_empty(handle_list))
4764 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4766 list_empty(handle_list) ? "empty" : "busy",
4767 list_empty(&conf->hold_list) ? "empty" : "busy",
4768 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4770 if (!list_empty(handle_list)) {
4771 sh = list_entry(handle_list->next, typeof(*sh), lru);
4773 if (list_empty(&conf->hold_list))
4774 conf->bypass_count = 0;
4775 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4776 if (conf->hold_list.next == conf->last_hold)
4777 conf->bypass_count++;
4779 conf->last_hold = conf->hold_list.next;
4780 conf->bypass_count -= conf->bypass_threshold;
4781 if (conf->bypass_count < 0)
4782 conf->bypass_count = 0;
4785 } else if (!list_empty(&conf->hold_list) &&
4786 ((conf->bypass_threshold &&
4787 conf->bypass_count > conf->bypass_threshold) ||
4788 atomic_read(&conf->pending_full_writes) == 0)) {
4790 list_for_each_entry(tmp, &conf->hold_list, lru) {
4791 if (conf->worker_cnt_per_group == 0 ||
4792 group == ANY_GROUP ||
4793 !cpu_online(tmp->cpu) ||
4794 cpu_to_group(tmp->cpu) == group) {
4801 conf->bypass_count -= conf->bypass_threshold;
4802 if (conf->bypass_count < 0)
4803 conf->bypass_count = 0;
4815 list_del_init(&sh->lru);
4816 BUG_ON(atomic_inc_return(&sh->count) != 1);
4820 struct raid5_plug_cb {
4821 struct blk_plug_cb cb;
4822 struct list_head list;
4823 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4826 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4828 struct raid5_plug_cb *cb = container_of(
4829 blk_cb, struct raid5_plug_cb, cb);
4830 struct stripe_head *sh;
4831 struct mddev *mddev = cb->cb.data;
4832 struct r5conf *conf = mddev->private;
4836 if (cb->list.next && !list_empty(&cb->list)) {
4837 spin_lock_irq(&conf->device_lock);
4838 while (!list_empty(&cb->list)) {
4839 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4840 list_del_init(&sh->lru);
4842 * avoid race release_stripe_plug() sees
4843 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4844 * is still in our list
4846 smp_mb__before_atomic();
4847 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4849 * STRIPE_ON_RELEASE_LIST could be set here. In that
4850 * case, the count is always > 1 here
4852 hash = sh->hash_lock_index;
4853 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
4856 spin_unlock_irq(&conf->device_lock);
4858 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4859 NR_STRIPE_HASH_LOCKS);
4861 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4865 static void release_stripe_plug(struct mddev *mddev,
4866 struct stripe_head *sh)
4868 struct blk_plug_cb *blk_cb = blk_check_plugged(
4869 raid5_unplug, mddev,
4870 sizeof(struct raid5_plug_cb));
4871 struct raid5_plug_cb *cb;
4878 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4880 if (cb->list.next == NULL) {
4882 INIT_LIST_HEAD(&cb->list);
4883 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
4884 INIT_LIST_HEAD(cb->temp_inactive_list + i);
4887 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4888 list_add_tail(&sh->lru, &cb->list);
4893 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4895 struct r5conf *conf = mddev->private;
4896 sector_t logical_sector, last_sector;
4897 struct stripe_head *sh;
4901 if (mddev->reshape_position != MaxSector)
4902 /* Skip discard while reshape is happening */
4905 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4906 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
4909 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4911 stripe_sectors = conf->chunk_sectors *
4912 (conf->raid_disks - conf->max_degraded);
4913 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4915 sector_div(last_sector, stripe_sectors);
4917 logical_sector *= conf->chunk_sectors;
4918 last_sector *= conf->chunk_sectors;
4920 for (; logical_sector < last_sector;
4921 logical_sector += STRIPE_SECTORS) {
4925 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4926 prepare_to_wait(&conf->wait_for_overlap, &w,
4927 TASK_UNINTERRUPTIBLE);
4928 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4929 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4934 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4935 spin_lock_irq(&sh->stripe_lock);
4936 for (d = 0; d < conf->raid_disks; d++) {
4937 if (d == sh->pd_idx || d == sh->qd_idx)
4939 if (sh->dev[d].towrite || sh->dev[d].toread) {
4940 set_bit(R5_Overlap, &sh->dev[d].flags);
4941 spin_unlock_irq(&sh->stripe_lock);
4947 set_bit(STRIPE_DISCARD, &sh->state);
4948 finish_wait(&conf->wait_for_overlap, &w);
4949 sh->overwrite_disks = 0;
4950 for (d = 0; d < conf->raid_disks; d++) {
4951 if (d == sh->pd_idx || d == sh->qd_idx)
4953 sh->dev[d].towrite = bi;
4954 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4955 raid5_inc_bi_active_stripes(bi);
4956 sh->overwrite_disks++;
4958 spin_unlock_irq(&sh->stripe_lock);
4959 if (conf->mddev->bitmap) {
4961 d < conf->raid_disks - conf->max_degraded;
4963 bitmap_startwrite(mddev->bitmap,
4967 sh->bm_seq = conf->seq_flush + 1;
4968 set_bit(STRIPE_BIT_DELAY, &sh->state);
4971 set_bit(STRIPE_HANDLE, &sh->state);
4972 clear_bit(STRIPE_DELAYED, &sh->state);
4973 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4974 atomic_inc(&conf->preread_active_stripes);
4975 release_stripe_plug(mddev, sh);
4978 remaining = raid5_dec_bi_active_stripes(bi);
4979 if (remaining == 0) {
4980 md_write_end(mddev);
4985 static void make_request(struct mddev *mddev, struct bio * bi)
4987 struct r5conf *conf = mddev->private;
4989 sector_t new_sector;
4990 sector_t logical_sector, last_sector;
4991 struct stripe_head *sh;
4992 const int rw = bio_data_dir(bi);
4997 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4998 md_flush_request(mddev, bi);
5002 md_write_start(mddev, bi);
5005 mddev->reshape_position == MaxSector &&
5006 chunk_aligned_read(mddev,bi))
5009 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5010 make_discard_request(mddev, bi);
5014 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5015 last_sector = bio_end_sector(bi);
5017 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5019 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5020 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5026 seq = read_seqcount_begin(&conf->gen_lock);
5029 prepare_to_wait(&conf->wait_for_overlap, &w,
5030 TASK_UNINTERRUPTIBLE);
5031 if (unlikely(conf->reshape_progress != MaxSector)) {
5032 /* spinlock is needed as reshape_progress may be
5033 * 64bit on a 32bit platform, and so it might be
5034 * possible to see a half-updated value
5035 * Of course reshape_progress could change after
5036 * the lock is dropped, so once we get a reference
5037 * to the stripe that we think it is, we will have
5040 spin_lock_irq(&conf->device_lock);
5041 if (mddev->reshape_backwards
5042 ? logical_sector < conf->reshape_progress
5043 : logical_sector >= conf->reshape_progress) {
5046 if (mddev->reshape_backwards
5047 ? logical_sector < conf->reshape_safe
5048 : logical_sector >= conf->reshape_safe) {
5049 spin_unlock_irq(&conf->device_lock);
5055 spin_unlock_irq(&conf->device_lock);
5058 new_sector = raid5_compute_sector(conf, logical_sector,
5061 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5062 (unsigned long long)new_sector,
5063 (unsigned long long)logical_sector);
5065 sh = get_active_stripe(conf, new_sector, previous,
5066 (bi->bi_rw&RWA_MASK), 0);
5068 if (unlikely(previous)) {
5069 /* expansion might have moved on while waiting for a
5070 * stripe, so we must do the range check again.
5071 * Expansion could still move past after this
5072 * test, but as we are holding a reference to
5073 * 'sh', we know that if that happens,
5074 * STRIPE_EXPANDING will get set and the expansion
5075 * won't proceed until we finish with the stripe.
5078 spin_lock_irq(&conf->device_lock);
5079 if (mddev->reshape_backwards
5080 ? logical_sector >= conf->reshape_progress
5081 : logical_sector < conf->reshape_progress)
5082 /* mismatch, need to try again */
5084 spin_unlock_irq(&conf->device_lock);
5092 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5093 /* Might have got the wrong stripe_head
5101 logical_sector >= mddev->suspend_lo &&
5102 logical_sector < mddev->suspend_hi) {
5104 /* As the suspend_* range is controlled by
5105 * userspace, we want an interruptible
5108 flush_signals(current);
5109 prepare_to_wait(&conf->wait_for_overlap,
5110 &w, TASK_INTERRUPTIBLE);
5111 if (logical_sector >= mddev->suspend_lo &&
5112 logical_sector < mddev->suspend_hi) {
5119 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5120 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5121 /* Stripe is busy expanding or
5122 * add failed due to overlap. Flush everything
5125 md_wakeup_thread(mddev->thread);
5131 set_bit(STRIPE_HANDLE, &sh->state);
5132 clear_bit(STRIPE_DELAYED, &sh->state);
5133 if ((!sh->batch_head || sh == sh->batch_head) &&
5134 (bi->bi_rw & REQ_SYNC) &&
5135 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5136 atomic_inc(&conf->preread_active_stripes);
5137 release_stripe_plug(mddev, sh);
5139 /* cannot get stripe for read-ahead, just give-up */
5140 clear_bit(BIO_UPTODATE, &bi->bi_flags);
5144 finish_wait(&conf->wait_for_overlap, &w);
5146 remaining = raid5_dec_bi_active_stripes(bi);
5147 if (remaining == 0) {
5150 md_write_end(mddev);
5152 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5158 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5160 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5162 /* reshaping is quite different to recovery/resync so it is
5163 * handled quite separately ... here.
5165 * On each call to sync_request, we gather one chunk worth of
5166 * destination stripes and flag them as expanding.
5167 * Then we find all the source stripes and request reads.
5168 * As the reads complete, handle_stripe will copy the data
5169 * into the destination stripe and release that stripe.
5171 struct r5conf *conf = mddev->private;
5172 struct stripe_head *sh;
5173 sector_t first_sector, last_sector;
5174 int raid_disks = conf->previous_raid_disks;
5175 int data_disks = raid_disks - conf->max_degraded;
5176 int new_data_disks = conf->raid_disks - conf->max_degraded;
5179 sector_t writepos, readpos, safepos;
5180 sector_t stripe_addr;
5181 int reshape_sectors;
5182 struct list_head stripes;
5184 if (sector_nr == 0) {
5185 /* If restarting in the middle, skip the initial sectors */
5186 if (mddev->reshape_backwards &&
5187 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5188 sector_nr = raid5_size(mddev, 0, 0)
5189 - conf->reshape_progress;
5190 } else if (!mddev->reshape_backwards &&
5191 conf->reshape_progress > 0)
5192 sector_nr = conf->reshape_progress;
5193 sector_div(sector_nr, new_data_disks);
5195 mddev->curr_resync_completed = sector_nr;
5196 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5202 /* We need to process a full chunk at a time.
5203 * If old and new chunk sizes differ, we need to process the
5206 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
5207 reshape_sectors = mddev->new_chunk_sectors;
5209 reshape_sectors = mddev->chunk_sectors;
5211 /* We update the metadata at least every 10 seconds, or when
5212 * the data about to be copied would over-write the source of
5213 * the data at the front of the range. i.e. one new_stripe
5214 * along from reshape_progress new_maps to after where
5215 * reshape_safe old_maps to
5217 writepos = conf->reshape_progress;
5218 sector_div(writepos, new_data_disks);
5219 readpos = conf->reshape_progress;
5220 sector_div(readpos, data_disks);
5221 safepos = conf->reshape_safe;
5222 sector_div(safepos, data_disks);
5223 if (mddev->reshape_backwards) {
5224 writepos -= min_t(sector_t, reshape_sectors, writepos);
5225 readpos += reshape_sectors;
5226 safepos += reshape_sectors;
5228 writepos += reshape_sectors;
5229 readpos -= min_t(sector_t, reshape_sectors, readpos);
5230 safepos -= min_t(sector_t, reshape_sectors, safepos);
5233 /* Having calculated the 'writepos' possibly use it
5234 * to set 'stripe_addr' which is where we will write to.
5236 if (mddev->reshape_backwards) {
5237 BUG_ON(conf->reshape_progress == 0);
5238 stripe_addr = writepos;
5239 BUG_ON((mddev->dev_sectors &
5240 ~((sector_t)reshape_sectors - 1))
5241 - reshape_sectors - stripe_addr
5244 BUG_ON(writepos != sector_nr + reshape_sectors);
5245 stripe_addr = sector_nr;
5248 /* 'writepos' is the most advanced device address we might write.
5249 * 'readpos' is the least advanced device address we might read.
5250 * 'safepos' is the least address recorded in the metadata as having
5252 * If there is a min_offset_diff, these are adjusted either by
5253 * increasing the safepos/readpos if diff is negative, or
5254 * increasing writepos if diff is positive.
5255 * If 'readpos' is then behind 'writepos', there is no way that we can
5256 * ensure safety in the face of a crash - that must be done by userspace
5257 * making a backup of the data. So in that case there is no particular
5258 * rush to update metadata.
5259 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5260 * update the metadata to advance 'safepos' to match 'readpos' so that
5261 * we can be safe in the event of a crash.
5262 * So we insist on updating metadata if safepos is behind writepos and
5263 * readpos is beyond writepos.
5264 * In any case, update the metadata every 10 seconds.
5265 * Maybe that number should be configurable, but I'm not sure it is
5266 * worth it.... maybe it could be a multiple of safemode_delay???
5268 if (conf->min_offset_diff < 0) {
5269 safepos += -conf->min_offset_diff;
5270 readpos += -conf->min_offset_diff;
5272 writepos += conf->min_offset_diff;
5274 if ((mddev->reshape_backwards
5275 ? (safepos > writepos && readpos < writepos)
5276 : (safepos < writepos && readpos > writepos)) ||
5277 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5278 /* Cannot proceed until we've updated the superblock... */
5279 wait_event(conf->wait_for_overlap,
5280 atomic_read(&conf->reshape_stripes)==0
5281 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5282 if (atomic_read(&conf->reshape_stripes) != 0)
5284 mddev->reshape_position = conf->reshape_progress;
5285 mddev->curr_resync_completed = sector_nr;
5286 conf->reshape_checkpoint = jiffies;
5287 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5288 md_wakeup_thread(mddev->thread);
5289 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5290 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5291 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5293 spin_lock_irq(&conf->device_lock);
5294 conf->reshape_safe = mddev->reshape_position;
5295 spin_unlock_irq(&conf->device_lock);
5296 wake_up(&conf->wait_for_overlap);
5297 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5300 INIT_LIST_HEAD(&stripes);
5301 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5303 int skipped_disk = 0;
5304 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5305 set_bit(STRIPE_EXPANDING, &sh->state);
5306 atomic_inc(&conf->reshape_stripes);
5307 /* If any of this stripe is beyond the end of the old
5308 * array, then we need to zero those blocks
5310 for (j=sh->disks; j--;) {
5312 if (j == sh->pd_idx)
5314 if (conf->level == 6 &&
5317 s = compute_blocknr(sh, j, 0);
5318 if (s < raid5_size(mddev, 0, 0)) {
5322 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5323 set_bit(R5_Expanded, &sh->dev[j].flags);
5324 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5326 if (!skipped_disk) {
5327 set_bit(STRIPE_EXPAND_READY, &sh->state);
5328 set_bit(STRIPE_HANDLE, &sh->state);
5330 list_add(&sh->lru, &stripes);
5332 spin_lock_irq(&conf->device_lock);
5333 if (mddev->reshape_backwards)
5334 conf->reshape_progress -= reshape_sectors * new_data_disks;
5336 conf->reshape_progress += reshape_sectors * new_data_disks;
5337 spin_unlock_irq(&conf->device_lock);
5338 /* Ok, those stripe are ready. We can start scheduling
5339 * reads on the source stripes.
5340 * The source stripes are determined by mapping the first and last
5341 * block on the destination stripes.
5344 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5347 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5348 * new_data_disks - 1),
5350 if (last_sector >= mddev->dev_sectors)
5351 last_sector = mddev->dev_sectors - 1;
5352 while (first_sector <= last_sector) {
5353 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
5354 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5355 set_bit(STRIPE_HANDLE, &sh->state);
5357 first_sector += STRIPE_SECTORS;
5359 /* Now that the sources are clearly marked, we can release
5360 * the destination stripes
5362 while (!list_empty(&stripes)) {
5363 sh = list_entry(stripes.next, struct stripe_head, lru);
5364 list_del_init(&sh->lru);
5367 /* If this takes us to the resync_max point where we have to pause,
5368 * then we need to write out the superblock.
5370 sector_nr += reshape_sectors;
5371 if ((sector_nr - mddev->curr_resync_completed) * 2
5372 >= mddev->resync_max - mddev->curr_resync_completed) {
5373 /* Cannot proceed until we've updated the superblock... */
5374 wait_event(conf->wait_for_overlap,
5375 atomic_read(&conf->reshape_stripes) == 0
5376 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5377 if (atomic_read(&conf->reshape_stripes) != 0)
5379 mddev->reshape_position = conf->reshape_progress;
5380 mddev->curr_resync_completed = sector_nr;
5381 conf->reshape_checkpoint = jiffies;
5382 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5383 md_wakeup_thread(mddev->thread);
5384 wait_event(mddev->sb_wait,
5385 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5386 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5387 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5389 spin_lock_irq(&conf->device_lock);
5390 conf->reshape_safe = mddev->reshape_position;
5391 spin_unlock_irq(&conf->device_lock);
5392 wake_up(&conf->wait_for_overlap);
5393 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5396 return reshape_sectors;
5399 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5401 struct r5conf *conf = mddev->private;
5402 struct stripe_head *sh;
5403 sector_t max_sector = mddev->dev_sectors;
5404 sector_t sync_blocks;
5405 int still_degraded = 0;
5408 if (sector_nr >= max_sector) {
5409 /* just being told to finish up .. nothing much to do */
5411 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5416 if (mddev->curr_resync < max_sector) /* aborted */
5417 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5419 else /* completed sync */
5421 bitmap_close_sync(mddev->bitmap);
5426 /* Allow raid5_quiesce to complete */
5427 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5429 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5430 return reshape_request(mddev, sector_nr, skipped);
5432 /* No need to check resync_max as we never do more than one
5433 * stripe, and as resync_max will always be on a chunk boundary,
5434 * if the check in md_do_sync didn't fire, there is no chance
5435 * of overstepping resync_max here
5438 /* if there is too many failed drives and we are trying
5439 * to resync, then assert that we are finished, because there is
5440 * nothing we can do.
5442 if (mddev->degraded >= conf->max_degraded &&
5443 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5444 sector_t rv = mddev->dev_sectors - sector_nr;
5448 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5450 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5451 sync_blocks >= STRIPE_SECTORS) {
5452 /* we can skip this block, and probably more */
5453 sync_blocks /= STRIPE_SECTORS;
5455 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5458 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5460 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5462 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5463 /* make sure we don't swamp the stripe cache if someone else
5464 * is trying to get access
5466 schedule_timeout_uninterruptible(1);
5468 /* Need to check if array will still be degraded after recovery/resync
5469 * Note in case of > 1 drive failures it's possible we're rebuilding
5470 * one drive while leaving another faulty drive in array.
5473 for (i = 0; i < conf->raid_disks; i++) {
5474 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5476 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5481 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5483 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5484 set_bit(STRIPE_HANDLE, &sh->state);
5488 return STRIPE_SECTORS;
5491 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5493 /* We may not be able to submit a whole bio at once as there
5494 * may not be enough stripe_heads available.
5495 * We cannot pre-allocate enough stripe_heads as we may need
5496 * more than exist in the cache (if we allow ever large chunks).
5497 * So we do one stripe head at a time and record in
5498 * ->bi_hw_segments how many have been done.
5500 * We *know* that this entire raid_bio is in one chunk, so
5501 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5503 struct stripe_head *sh;
5505 sector_t sector, logical_sector, last_sector;
5510 logical_sector = raid_bio->bi_iter.bi_sector &
5511 ~((sector_t)STRIPE_SECTORS-1);
5512 sector = raid5_compute_sector(conf, logical_sector,
5514 last_sector = bio_end_sector(raid_bio);
5516 for (; logical_sector < last_sector;
5517 logical_sector += STRIPE_SECTORS,
5518 sector += STRIPE_SECTORS,
5521 if (scnt < raid5_bi_processed_stripes(raid_bio))
5522 /* already done this stripe */
5525 sh = get_active_stripe(conf, sector, 0, 1, 1);
5528 /* failed to get a stripe - must wait */
5529 raid5_set_bi_processed_stripes(raid_bio, scnt);
5530 conf->retry_read_aligned = raid_bio;
5534 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5536 raid5_set_bi_processed_stripes(raid_bio, scnt);
5537 conf->retry_read_aligned = raid_bio;
5541 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5546 remaining = raid5_dec_bi_active_stripes(raid_bio);
5547 if (remaining == 0) {
5548 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5550 bio_endio(raid_bio, 0);
5552 if (atomic_dec_and_test(&conf->active_aligned_reads))
5553 wake_up(&conf->wait_for_stripe);
5557 static int handle_active_stripes(struct r5conf *conf, int group,
5558 struct r5worker *worker,
5559 struct list_head *temp_inactive_list)
5561 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5562 int i, batch_size = 0, hash;
5563 bool release_inactive = false;
5565 while (batch_size < MAX_STRIPE_BATCH &&
5566 (sh = __get_priority_stripe(conf, group)) != NULL)
5567 batch[batch_size++] = sh;
5569 if (batch_size == 0) {
5570 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5571 if (!list_empty(temp_inactive_list + i))
5573 if (i == NR_STRIPE_HASH_LOCKS)
5575 release_inactive = true;
5577 spin_unlock_irq(&conf->device_lock);
5579 release_inactive_stripe_list(conf, temp_inactive_list,
5580 NR_STRIPE_HASH_LOCKS);
5582 if (release_inactive) {
5583 spin_lock_irq(&conf->device_lock);
5587 for (i = 0; i < batch_size; i++)
5588 handle_stripe(batch[i]);
5592 spin_lock_irq(&conf->device_lock);
5593 for (i = 0; i < batch_size; i++) {
5594 hash = batch[i]->hash_lock_index;
5595 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5600 static void raid5_do_work(struct work_struct *work)
5602 struct r5worker *worker = container_of(work, struct r5worker, work);
5603 struct r5worker_group *group = worker->group;
5604 struct r5conf *conf = group->conf;
5605 int group_id = group - conf->worker_groups;
5607 struct blk_plug plug;
5609 pr_debug("+++ raid5worker active\n");
5611 blk_start_plug(&plug);
5613 spin_lock_irq(&conf->device_lock);
5615 int batch_size, released;
5617 released = release_stripe_list(conf, worker->temp_inactive_list);
5619 batch_size = handle_active_stripes(conf, group_id, worker,
5620 worker->temp_inactive_list);
5621 worker->working = false;
5622 if (!batch_size && !released)
5624 handled += batch_size;
5626 pr_debug("%d stripes handled\n", handled);
5628 spin_unlock_irq(&conf->device_lock);
5629 blk_finish_plug(&plug);
5631 pr_debug("--- raid5worker inactive\n");
5635 * This is our raid5 kernel thread.
5637 * We scan the hash table for stripes which can be handled now.
5638 * During the scan, completed stripes are saved for us by the interrupt
5639 * handler, so that they will not have to wait for our next wakeup.
5641 static void raid5d(struct md_thread *thread)
5643 struct mddev *mddev = thread->mddev;
5644 struct r5conf *conf = mddev->private;
5646 struct blk_plug plug;
5648 pr_debug("+++ raid5d active\n");
5650 md_check_recovery(mddev);
5652 blk_start_plug(&plug);
5654 spin_lock_irq(&conf->device_lock);
5657 int batch_size, released;
5659 released = release_stripe_list(conf, conf->temp_inactive_list);
5662 !list_empty(&conf->bitmap_list)) {
5663 /* Now is a good time to flush some bitmap updates */
5665 spin_unlock_irq(&conf->device_lock);
5666 bitmap_unplug(mddev->bitmap);
5667 spin_lock_irq(&conf->device_lock);
5668 conf->seq_write = conf->seq_flush;
5669 activate_bit_delay(conf, conf->temp_inactive_list);
5671 raid5_activate_delayed(conf);
5673 while ((bio = remove_bio_from_retry(conf))) {
5675 spin_unlock_irq(&conf->device_lock);
5676 ok = retry_aligned_read(conf, bio);
5677 spin_lock_irq(&conf->device_lock);
5683 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5684 conf->temp_inactive_list);
5685 if (!batch_size && !released)
5687 handled += batch_size;
5689 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5690 spin_unlock_irq(&conf->device_lock);
5691 md_check_recovery(mddev);
5692 spin_lock_irq(&conf->device_lock);
5695 pr_debug("%d stripes handled\n", handled);
5697 spin_unlock_irq(&conf->device_lock);
5699 async_tx_issue_pending_all();
5700 blk_finish_plug(&plug);
5702 pr_debug("--- raid5d inactive\n");
5706 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5708 struct r5conf *conf;
5710 spin_lock(&mddev->lock);
5711 conf = mddev->private;
5713 ret = sprintf(page, "%d\n", conf->max_nr_stripes);
5714 spin_unlock(&mddev->lock);
5719 raid5_set_cache_size(struct mddev *mddev, int size)
5721 struct r5conf *conf = mddev->private;
5725 if (size <= 16 || size > 32768)
5727 hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
5728 while (size < conf->max_nr_stripes) {
5729 if (drop_one_stripe(conf, hash))
5730 conf->max_nr_stripes--;
5735 hash = NR_STRIPE_HASH_LOCKS - 1;
5737 err = md_allow_write(mddev);
5740 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
5741 while (size > conf->max_nr_stripes) {
5742 if (grow_one_stripe(conf, hash))
5743 conf->max_nr_stripes++;
5745 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
5749 EXPORT_SYMBOL(raid5_set_cache_size);
5752 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5754 struct r5conf *conf;
5758 if (len >= PAGE_SIZE)
5760 if (kstrtoul(page, 10, &new))
5762 err = mddev_lock(mddev);
5765 conf = mddev->private;
5769 err = raid5_set_cache_size(mddev, new);
5770 mddev_unlock(mddev);
5775 static struct md_sysfs_entry
5776 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5777 raid5_show_stripe_cache_size,
5778 raid5_store_stripe_cache_size);
5781 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5783 struct r5conf *conf;
5785 spin_lock(&mddev->lock);
5786 conf = mddev->private;
5788 ret = sprintf(page, "%d\n", conf->bypass_threshold);
5789 spin_unlock(&mddev->lock);
5794 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
5796 struct r5conf *conf;
5800 if (len >= PAGE_SIZE)
5802 if (kstrtoul(page, 10, &new))
5805 err = mddev_lock(mddev);
5808 conf = mddev->private;
5811 else if (new > conf->max_nr_stripes)
5814 conf->bypass_threshold = new;
5815 mddev_unlock(mddev);
5819 static struct md_sysfs_entry
5820 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5822 raid5_show_preread_threshold,
5823 raid5_store_preread_threshold);
5826 raid5_show_skip_copy(struct mddev *mddev, char *page)
5828 struct r5conf *conf;
5830 spin_lock(&mddev->lock);
5831 conf = mddev->private;
5833 ret = sprintf(page, "%d\n", conf->skip_copy);
5834 spin_unlock(&mddev->lock);
5839 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
5841 struct r5conf *conf;
5845 if (len >= PAGE_SIZE)
5847 if (kstrtoul(page, 10, &new))
5851 err = mddev_lock(mddev);
5854 conf = mddev->private;
5857 else if (new != conf->skip_copy) {
5858 mddev_suspend(mddev);
5859 conf->skip_copy = new;
5861 mddev->queue->backing_dev_info.capabilities |=
5862 BDI_CAP_STABLE_WRITES;
5864 mddev->queue->backing_dev_info.capabilities &=
5865 ~BDI_CAP_STABLE_WRITES;
5866 mddev_resume(mddev);
5868 mddev_unlock(mddev);
5872 static struct md_sysfs_entry
5873 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
5874 raid5_show_skip_copy,
5875 raid5_store_skip_copy);
5878 stripe_cache_active_show(struct mddev *mddev, char *page)
5880 struct r5conf *conf = mddev->private;
5882 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5887 static struct md_sysfs_entry
5888 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
5891 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
5893 struct r5conf *conf;
5895 spin_lock(&mddev->lock);
5896 conf = mddev->private;
5898 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
5899 spin_unlock(&mddev->lock);
5903 static int alloc_thread_groups(struct r5conf *conf, int cnt,
5905 int *worker_cnt_per_group,
5906 struct r5worker_group **worker_groups);
5908 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
5910 struct r5conf *conf;
5913 struct r5worker_group *new_groups, *old_groups;
5914 int group_cnt, worker_cnt_per_group;
5916 if (len >= PAGE_SIZE)
5918 if (kstrtoul(page, 10, &new))
5921 err = mddev_lock(mddev);
5924 conf = mddev->private;
5927 else if (new != conf->worker_cnt_per_group) {
5928 mddev_suspend(mddev);
5930 old_groups = conf->worker_groups;
5932 flush_workqueue(raid5_wq);
5934 err = alloc_thread_groups(conf, new,
5935 &group_cnt, &worker_cnt_per_group,
5938 spin_lock_irq(&conf->device_lock);
5939 conf->group_cnt = group_cnt;
5940 conf->worker_cnt_per_group = worker_cnt_per_group;
5941 conf->worker_groups = new_groups;
5942 spin_unlock_irq(&conf->device_lock);
5945 kfree(old_groups[0].workers);
5948 mddev_resume(mddev);
5950 mddev_unlock(mddev);
5955 static struct md_sysfs_entry
5956 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
5957 raid5_show_group_thread_cnt,
5958 raid5_store_group_thread_cnt);
5960 static struct attribute *raid5_attrs[] = {
5961 &raid5_stripecache_size.attr,
5962 &raid5_stripecache_active.attr,
5963 &raid5_preread_bypass_threshold.attr,
5964 &raid5_group_thread_cnt.attr,
5965 &raid5_skip_copy.attr,
5968 static struct attribute_group raid5_attrs_group = {
5970 .attrs = raid5_attrs,
5973 static int alloc_thread_groups(struct r5conf *conf, int cnt,
5975 int *worker_cnt_per_group,
5976 struct r5worker_group **worker_groups)
5980 struct r5worker *workers;
5982 *worker_cnt_per_group = cnt;
5985 *worker_groups = NULL;
5988 *group_cnt = num_possible_nodes();
5989 size = sizeof(struct r5worker) * cnt;
5990 workers = kzalloc(size * *group_cnt, GFP_NOIO);
5991 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
5992 *group_cnt, GFP_NOIO);
5993 if (!*worker_groups || !workers) {
5995 kfree(*worker_groups);
5999 for (i = 0; i < *group_cnt; i++) {
6000 struct r5worker_group *group;
6002 group = &(*worker_groups)[i];
6003 INIT_LIST_HEAD(&group->handle_list);
6005 group->workers = workers + i * cnt;
6007 for (j = 0; j < cnt; j++) {
6008 struct r5worker *worker = group->workers + j;
6009 worker->group = group;
6010 INIT_WORK(&worker->work, raid5_do_work);
6012 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6013 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6020 static void free_thread_groups(struct r5conf *conf)
6022 if (conf->worker_groups)
6023 kfree(conf->worker_groups[0].workers);
6024 kfree(conf->worker_groups);
6025 conf->worker_groups = NULL;
6029 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6031 struct r5conf *conf = mddev->private;
6034 sectors = mddev->dev_sectors;
6036 /* size is defined by the smallest of previous and new size */
6037 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6039 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6040 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
6041 return sectors * (raid_disks - conf->max_degraded);
6044 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6046 safe_put_page(percpu->spare_page);
6047 if (percpu->scribble)
6048 flex_array_free(percpu->scribble);
6049 percpu->spare_page = NULL;
6050 percpu->scribble = NULL;
6053 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6055 if (conf->level == 6 && !percpu->spare_page)
6056 percpu->spare_page = alloc_page(GFP_KERNEL);
6057 if (!percpu->scribble)
6058 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6059 conf->previous_raid_disks), conf->chunk_sectors /
6060 STRIPE_SECTORS, GFP_KERNEL);
6062 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6063 free_scratch_buffer(conf, percpu);
6070 static void raid5_free_percpu(struct r5conf *conf)
6077 #ifdef CONFIG_HOTPLUG_CPU
6078 unregister_cpu_notifier(&conf->cpu_notify);
6082 for_each_possible_cpu(cpu)
6083 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6086 free_percpu(conf->percpu);
6089 static void free_conf(struct r5conf *conf)
6091 free_thread_groups(conf);
6092 shrink_stripes(conf);
6093 raid5_free_percpu(conf);
6095 kfree(conf->stripe_hashtbl);
6099 #ifdef CONFIG_HOTPLUG_CPU
6100 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6103 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6104 long cpu = (long)hcpu;
6105 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6108 case CPU_UP_PREPARE:
6109 case CPU_UP_PREPARE_FROZEN:
6110 if (alloc_scratch_buffer(conf, percpu)) {
6111 pr_err("%s: failed memory allocation for cpu%ld\n",
6113 return notifier_from_errno(-ENOMEM);
6117 case CPU_DEAD_FROZEN:
6118 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6127 static int raid5_alloc_percpu(struct r5conf *conf)
6132 conf->percpu = alloc_percpu(struct raid5_percpu);
6136 #ifdef CONFIG_HOTPLUG_CPU
6137 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6138 conf->cpu_notify.priority = 0;
6139 err = register_cpu_notifier(&conf->cpu_notify);
6145 for_each_present_cpu(cpu) {
6146 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6148 pr_err("%s: failed memory allocation for cpu%ld\n",
6158 static struct r5conf *setup_conf(struct mddev *mddev)
6160 struct r5conf *conf;
6161 int raid_disk, memory, max_disks;
6162 struct md_rdev *rdev;
6163 struct disk_info *disk;
6166 int group_cnt, worker_cnt_per_group;
6167 struct r5worker_group *new_group;
6169 if (mddev->new_level != 5
6170 && mddev->new_level != 4
6171 && mddev->new_level != 6) {
6172 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6173 mdname(mddev), mddev->new_level);
6174 return ERR_PTR(-EIO);
6176 if ((mddev->new_level == 5
6177 && !algorithm_valid_raid5(mddev->new_layout)) ||
6178 (mddev->new_level == 6
6179 && !algorithm_valid_raid6(mddev->new_layout))) {
6180 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6181 mdname(mddev), mddev->new_layout);
6182 return ERR_PTR(-EIO);
6184 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6185 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6186 mdname(mddev), mddev->raid_disks);
6187 return ERR_PTR(-EINVAL);
6190 if (!mddev->new_chunk_sectors ||
6191 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6192 !is_power_of_2(mddev->new_chunk_sectors)) {
6193 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6194 mdname(mddev), mddev->new_chunk_sectors << 9);
6195 return ERR_PTR(-EINVAL);
6198 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6201 /* Don't enable multi-threading by default*/
6202 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6204 conf->group_cnt = group_cnt;
6205 conf->worker_cnt_per_group = worker_cnt_per_group;
6206 conf->worker_groups = new_group;
6209 spin_lock_init(&conf->device_lock);
6210 seqcount_init(&conf->gen_lock);
6211 init_waitqueue_head(&conf->wait_for_stripe);
6212 init_waitqueue_head(&conf->wait_for_overlap);
6213 INIT_LIST_HEAD(&conf->handle_list);
6214 INIT_LIST_HEAD(&conf->hold_list);
6215 INIT_LIST_HEAD(&conf->delayed_list);
6216 INIT_LIST_HEAD(&conf->bitmap_list);
6217 init_llist_head(&conf->released_stripes);
6218 atomic_set(&conf->active_stripes, 0);
6219 atomic_set(&conf->preread_active_stripes, 0);
6220 atomic_set(&conf->active_aligned_reads, 0);
6221 conf->bypass_threshold = BYPASS_THRESHOLD;
6222 conf->recovery_disabled = mddev->recovery_disabled - 1;
6224 conf->raid_disks = mddev->raid_disks;
6225 if (mddev->reshape_position == MaxSector)
6226 conf->previous_raid_disks = mddev->raid_disks;
6228 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6229 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6231 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6236 conf->mddev = mddev;
6238 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6241 /* We init hash_locks[0] separately to that it can be used
6242 * as the reference lock in the spin_lock_nest_lock() call
6243 * in lock_all_device_hash_locks_irq in order to convince
6244 * lockdep that we know what we are doing.
6246 spin_lock_init(conf->hash_locks);
6247 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6248 spin_lock_init(conf->hash_locks + i);
6250 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6251 INIT_LIST_HEAD(conf->inactive_list + i);
6253 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6254 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6256 conf->level = mddev->new_level;
6257 conf->chunk_sectors = mddev->new_chunk_sectors;
6258 if (raid5_alloc_percpu(conf) != 0)
6261 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6263 rdev_for_each(rdev, mddev) {
6264 raid_disk = rdev->raid_disk;
6265 if (raid_disk >= max_disks
6268 disk = conf->disks + raid_disk;
6270 if (test_bit(Replacement, &rdev->flags)) {
6271 if (disk->replacement)
6273 disk->replacement = rdev;
6280 if (test_bit(In_sync, &rdev->flags)) {
6281 char b[BDEVNAME_SIZE];
6282 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6284 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6285 } else if (rdev->saved_raid_disk != raid_disk)
6286 /* Cannot rely on bitmap to complete recovery */
6290 conf->level = mddev->new_level;
6291 if (conf->level == 6)
6292 conf->max_degraded = 2;
6294 conf->max_degraded = 1;
6295 conf->algorithm = mddev->new_layout;
6296 conf->reshape_progress = mddev->reshape_position;
6297 if (conf->reshape_progress != MaxSector) {
6298 conf->prev_chunk_sectors = mddev->chunk_sectors;
6299 conf->prev_algo = mddev->layout;
6302 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
6303 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6304 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6305 if (grow_stripes(conf, NR_STRIPES)) {
6307 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6308 mdname(mddev), memory);
6311 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6312 mdname(mddev), memory);
6314 sprintf(pers_name, "raid%d", mddev->new_level);
6315 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6316 if (!conf->thread) {
6318 "md/raid:%s: couldn't allocate thread.\n",
6328 return ERR_PTR(-EIO);
6330 return ERR_PTR(-ENOMEM);
6333 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6336 case ALGORITHM_PARITY_0:
6337 if (raid_disk < max_degraded)
6340 case ALGORITHM_PARITY_N:
6341 if (raid_disk >= raid_disks - max_degraded)
6344 case ALGORITHM_PARITY_0_6:
6345 if (raid_disk == 0 ||
6346 raid_disk == raid_disks - 1)
6349 case ALGORITHM_LEFT_ASYMMETRIC_6:
6350 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6351 case ALGORITHM_LEFT_SYMMETRIC_6:
6352 case ALGORITHM_RIGHT_SYMMETRIC_6:
6353 if (raid_disk == raid_disks - 1)
6359 static int run(struct mddev *mddev)
6361 struct r5conf *conf;
6362 int working_disks = 0;
6363 int dirty_parity_disks = 0;
6364 struct md_rdev *rdev;
6365 sector_t reshape_offset = 0;
6367 long long min_offset_diff = 0;
6370 if (mddev->recovery_cp != MaxSector)
6371 printk(KERN_NOTICE "md/raid:%s: not clean"
6372 " -- starting background reconstruction\n",
6375 rdev_for_each(rdev, mddev) {
6377 if (rdev->raid_disk < 0)
6379 diff = (rdev->new_data_offset - rdev->data_offset);
6381 min_offset_diff = diff;
6383 } else if (mddev->reshape_backwards &&
6384 diff < min_offset_diff)
6385 min_offset_diff = diff;
6386 else if (!mddev->reshape_backwards &&
6387 diff > min_offset_diff)
6388 min_offset_diff = diff;
6391 if (mddev->reshape_position != MaxSector) {
6392 /* Check that we can continue the reshape.
6393 * Difficulties arise if the stripe we would write to
6394 * next is at or after the stripe we would read from next.
6395 * For a reshape that changes the number of devices, this
6396 * is only possible for a very short time, and mdadm makes
6397 * sure that time appears to have past before assembling
6398 * the array. So we fail if that time hasn't passed.
6399 * For a reshape that keeps the number of devices the same
6400 * mdadm must be monitoring the reshape can keeping the
6401 * critical areas read-only and backed up. It will start
6402 * the array in read-only mode, so we check for that.
6404 sector_t here_new, here_old;
6406 int max_degraded = (mddev->level == 6 ? 2 : 1);
6408 if (mddev->new_level != mddev->level) {
6409 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6410 "required - aborting.\n",
6414 old_disks = mddev->raid_disks - mddev->delta_disks;
6415 /* reshape_position must be on a new-stripe boundary, and one
6416 * further up in new geometry must map after here in old
6419 here_new = mddev->reshape_position;
6420 if (sector_div(here_new, mddev->new_chunk_sectors *
6421 (mddev->raid_disks - max_degraded))) {
6422 printk(KERN_ERR "md/raid:%s: reshape_position not "
6423 "on a stripe boundary\n", mdname(mddev));
6426 reshape_offset = here_new * mddev->new_chunk_sectors;
6427 /* here_new is the stripe we will write to */
6428 here_old = mddev->reshape_position;
6429 sector_div(here_old, mddev->chunk_sectors *
6430 (old_disks-max_degraded));
6431 /* here_old is the first stripe that we might need to read
6433 if (mddev->delta_disks == 0) {
6434 if ((here_new * mddev->new_chunk_sectors !=
6435 here_old * mddev->chunk_sectors)) {
6436 printk(KERN_ERR "md/raid:%s: reshape position is"
6437 " confused - aborting\n", mdname(mddev));
6440 /* We cannot be sure it is safe to start an in-place
6441 * reshape. It is only safe if user-space is monitoring
6442 * and taking constant backups.
6443 * mdadm always starts a situation like this in
6444 * readonly mode so it can take control before
6445 * allowing any writes. So just check for that.
6447 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6448 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6449 /* not really in-place - so OK */;
6450 else if (mddev->ro == 0) {
6451 printk(KERN_ERR "md/raid:%s: in-place reshape "
6452 "must be started in read-only mode "
6457 } else if (mddev->reshape_backwards
6458 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6459 here_old * mddev->chunk_sectors)
6460 : (here_new * mddev->new_chunk_sectors >=
6461 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6462 /* Reading from the same stripe as writing to - bad */
6463 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6464 "auto-recovery - aborting.\n",
6468 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6470 /* OK, we should be able to continue; */
6472 BUG_ON(mddev->level != mddev->new_level);
6473 BUG_ON(mddev->layout != mddev->new_layout);
6474 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6475 BUG_ON(mddev->delta_disks != 0);
6478 if (mddev->private == NULL)
6479 conf = setup_conf(mddev);
6481 conf = mddev->private;
6484 return PTR_ERR(conf);
6486 conf->min_offset_diff = min_offset_diff;
6487 mddev->thread = conf->thread;
6488 conf->thread = NULL;
6489 mddev->private = conf;
6491 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6493 rdev = conf->disks[i].rdev;
6494 if (!rdev && conf->disks[i].replacement) {
6495 /* The replacement is all we have yet */
6496 rdev = conf->disks[i].replacement;
6497 conf->disks[i].replacement = NULL;
6498 clear_bit(Replacement, &rdev->flags);
6499 conf->disks[i].rdev = rdev;
6503 if (conf->disks[i].replacement &&
6504 conf->reshape_progress != MaxSector) {
6505 /* replacements and reshape simply do not mix. */
6506 printk(KERN_ERR "md: cannot handle concurrent "
6507 "replacement and reshape.\n");
6510 if (test_bit(In_sync, &rdev->flags)) {
6514 /* This disc is not fully in-sync. However if it
6515 * just stored parity (beyond the recovery_offset),
6516 * when we don't need to be concerned about the
6517 * array being dirty.
6518 * When reshape goes 'backwards', we never have
6519 * partially completed devices, so we only need
6520 * to worry about reshape going forwards.
6522 /* Hack because v0.91 doesn't store recovery_offset properly. */
6523 if (mddev->major_version == 0 &&
6524 mddev->minor_version > 90)
6525 rdev->recovery_offset = reshape_offset;
6527 if (rdev->recovery_offset < reshape_offset) {
6528 /* We need to check old and new layout */
6529 if (!only_parity(rdev->raid_disk,
6532 conf->max_degraded))
6535 if (!only_parity(rdev->raid_disk,
6537 conf->previous_raid_disks,
6538 conf->max_degraded))
6540 dirty_parity_disks++;
6544 * 0 for a fully functional array, 1 or 2 for a degraded array.
6546 mddev->degraded = calc_degraded(conf);
6548 if (has_failed(conf)) {
6549 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6550 " (%d/%d failed)\n",
6551 mdname(mddev), mddev->degraded, conf->raid_disks);
6555 /* device size must be a multiple of chunk size */
6556 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6557 mddev->resync_max_sectors = mddev->dev_sectors;
6559 if (mddev->degraded > dirty_parity_disks &&
6560 mddev->recovery_cp != MaxSector) {
6561 if (mddev->ok_start_degraded)
6563 "md/raid:%s: starting dirty degraded array"
6564 " - data corruption possible.\n",
6568 "md/raid:%s: cannot start dirty degraded array.\n",
6574 if (mddev->degraded == 0)
6575 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6576 " devices, algorithm %d\n", mdname(mddev), conf->level,
6577 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6580 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6581 " out of %d devices, algorithm %d\n",
6582 mdname(mddev), conf->level,
6583 mddev->raid_disks - mddev->degraded,
6584 mddev->raid_disks, mddev->new_layout);
6586 print_raid5_conf(conf);
6588 if (conf->reshape_progress != MaxSector) {
6589 conf->reshape_safe = conf->reshape_progress;
6590 atomic_set(&conf->reshape_stripes, 0);
6591 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6592 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6593 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6594 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6595 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6599 /* Ok, everything is just fine now */
6600 if (mddev->to_remove == &raid5_attrs_group)
6601 mddev->to_remove = NULL;
6602 else if (mddev->kobj.sd &&
6603 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6605 "raid5: failed to create sysfs attributes for %s\n",
6607 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6611 bool discard_supported = true;
6612 /* read-ahead size must cover two whole stripes, which
6613 * is 2 * (datadisks) * chunksize where 'n' is the
6614 * number of raid devices
6616 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6617 int stripe = data_disks *
6618 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6619 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6620 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6622 chunk_size = mddev->chunk_sectors << 9;
6623 blk_queue_io_min(mddev->queue, chunk_size);
6624 blk_queue_io_opt(mddev->queue, chunk_size *
6625 (conf->raid_disks - conf->max_degraded));
6626 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6628 * We can only discard a whole stripe. It doesn't make sense to
6629 * discard data disk but write parity disk
6631 stripe = stripe * PAGE_SIZE;
6632 /* Round up to power of 2, as discard handling
6633 * currently assumes that */
6634 while ((stripe-1) & stripe)
6635 stripe = (stripe | (stripe-1)) + 1;
6636 mddev->queue->limits.discard_alignment = stripe;
6637 mddev->queue->limits.discard_granularity = stripe;
6639 * unaligned part of discard request will be ignored, so can't
6640 * guarantee discard_zeroes_data
6642 mddev->queue->limits.discard_zeroes_data = 0;
6644 blk_queue_max_write_same_sectors(mddev->queue, 0);
6646 rdev_for_each(rdev, mddev) {
6647 disk_stack_limits(mddev->gendisk, rdev->bdev,
6648 rdev->data_offset << 9);
6649 disk_stack_limits(mddev->gendisk, rdev->bdev,
6650 rdev->new_data_offset << 9);
6652 * discard_zeroes_data is required, otherwise data
6653 * could be lost. Consider a scenario: discard a stripe
6654 * (the stripe could be inconsistent if
6655 * discard_zeroes_data is 0); write one disk of the
6656 * stripe (the stripe could be inconsistent again
6657 * depending on which disks are used to calculate
6658 * parity); the disk is broken; The stripe data of this
6661 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6662 !bdev_get_queue(rdev->bdev)->
6663 limits.discard_zeroes_data)
6664 discard_supported = false;
6665 /* Unfortunately, discard_zeroes_data is not currently
6666 * a guarantee - just a hint. So we only allow DISCARD
6667 * if the sysadmin has confirmed that only safe devices
6668 * are in use by setting a module parameter.
6670 if (!devices_handle_discard_safely) {
6671 if (discard_supported) {
6672 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6673 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6675 discard_supported = false;
6679 if (discard_supported &&
6680 mddev->queue->limits.max_discard_sectors >= stripe &&
6681 mddev->queue->limits.discard_granularity >= stripe)
6682 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6685 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6691 md_unregister_thread(&mddev->thread);
6692 print_raid5_conf(conf);
6694 mddev->private = NULL;
6695 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6699 static void raid5_free(struct mddev *mddev, void *priv)
6701 struct r5conf *conf = priv;
6704 mddev->to_remove = &raid5_attrs_group;
6707 static void status(struct seq_file *seq, struct mddev *mddev)
6709 struct r5conf *conf = mddev->private;
6712 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6713 mddev->chunk_sectors / 2, mddev->layout);
6714 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6715 for (i = 0; i < conf->raid_disks; i++)
6716 seq_printf (seq, "%s",
6717 conf->disks[i].rdev &&
6718 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6719 seq_printf (seq, "]");
6722 static void print_raid5_conf (struct r5conf *conf)
6725 struct disk_info *tmp;
6727 printk(KERN_DEBUG "RAID conf printout:\n");
6729 printk("(conf==NULL)\n");
6732 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6734 conf->raid_disks - conf->mddev->degraded);
6736 for (i = 0; i < conf->raid_disks; i++) {
6737 char b[BDEVNAME_SIZE];
6738 tmp = conf->disks + i;
6740 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6741 i, !test_bit(Faulty, &tmp->rdev->flags),
6742 bdevname(tmp->rdev->bdev, b));
6746 static int raid5_spare_active(struct mddev *mddev)
6749 struct r5conf *conf = mddev->private;
6750 struct disk_info *tmp;
6752 unsigned long flags;
6754 for (i = 0; i < conf->raid_disks; i++) {
6755 tmp = conf->disks + i;
6756 if (tmp->replacement
6757 && tmp->replacement->recovery_offset == MaxSector
6758 && !test_bit(Faulty, &tmp->replacement->flags)
6759 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
6760 /* Replacement has just become active. */
6762 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
6765 /* Replaced device not technically faulty,
6766 * but we need to be sure it gets removed
6767 * and never re-added.
6769 set_bit(Faulty, &tmp->rdev->flags);
6770 sysfs_notify_dirent_safe(
6771 tmp->rdev->sysfs_state);
6773 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
6774 } else if (tmp->rdev
6775 && tmp->rdev->recovery_offset == MaxSector
6776 && !test_bit(Faulty, &tmp->rdev->flags)
6777 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
6779 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
6782 spin_lock_irqsave(&conf->device_lock, flags);
6783 mddev->degraded = calc_degraded(conf);
6784 spin_unlock_irqrestore(&conf->device_lock, flags);
6785 print_raid5_conf(conf);
6789 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
6791 struct r5conf *conf = mddev->private;
6793 int number = rdev->raid_disk;
6794 struct md_rdev **rdevp;
6795 struct disk_info *p = conf->disks + number;
6797 print_raid5_conf(conf);
6798 if (rdev == p->rdev)
6800 else if (rdev == p->replacement)
6801 rdevp = &p->replacement;
6805 if (number >= conf->raid_disks &&
6806 conf->reshape_progress == MaxSector)
6807 clear_bit(In_sync, &rdev->flags);
6809 if (test_bit(In_sync, &rdev->flags) ||
6810 atomic_read(&rdev->nr_pending)) {
6814 /* Only remove non-faulty devices if recovery
6817 if (!test_bit(Faulty, &rdev->flags) &&
6818 mddev->recovery_disabled != conf->recovery_disabled &&
6819 !has_failed(conf) &&
6820 (!p->replacement || p->replacement == rdev) &&
6821 number < conf->raid_disks) {
6827 if (atomic_read(&rdev->nr_pending)) {
6828 /* lost the race, try later */
6831 } else if (p->replacement) {
6832 /* We must have just cleared 'rdev' */
6833 p->rdev = p->replacement;
6834 clear_bit(Replacement, &p->replacement->flags);
6835 smp_mb(); /* Make sure other CPUs may see both as identical
6836 * but will never see neither - if they are careful
6838 p->replacement = NULL;
6839 clear_bit(WantReplacement, &rdev->flags);
6841 /* We might have just removed the Replacement as faulty-
6842 * clear the bit just in case
6844 clear_bit(WantReplacement, &rdev->flags);
6847 print_raid5_conf(conf);
6851 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
6853 struct r5conf *conf = mddev->private;
6856 struct disk_info *p;
6858 int last = conf->raid_disks - 1;
6860 if (mddev->recovery_disabled == conf->recovery_disabled)
6863 if (rdev->saved_raid_disk < 0 && has_failed(conf))
6864 /* no point adding a device */
6867 if (rdev->raid_disk >= 0)
6868 first = last = rdev->raid_disk;
6871 * find the disk ... but prefer rdev->saved_raid_disk
6874 if (rdev->saved_raid_disk >= 0 &&
6875 rdev->saved_raid_disk >= first &&
6876 conf->disks[rdev->saved_raid_disk].rdev == NULL)
6877 first = rdev->saved_raid_disk;
6879 for (disk = first; disk <= last; disk++) {
6880 p = conf->disks + disk;
6881 if (p->rdev == NULL) {
6882 clear_bit(In_sync, &rdev->flags);
6883 rdev->raid_disk = disk;
6885 if (rdev->saved_raid_disk != disk)
6887 rcu_assign_pointer(p->rdev, rdev);
6891 for (disk = first; disk <= last; disk++) {
6892 p = conf->disks + disk;
6893 if (test_bit(WantReplacement, &p->rdev->flags) &&
6894 p->replacement == NULL) {
6895 clear_bit(In_sync, &rdev->flags);
6896 set_bit(Replacement, &rdev->flags);
6897 rdev->raid_disk = disk;
6900 rcu_assign_pointer(p->replacement, rdev);
6905 print_raid5_conf(conf);
6909 static int raid5_resize(struct mddev *mddev, sector_t sectors)
6911 /* no resync is happening, and there is enough space
6912 * on all devices, so we can resize.
6913 * We need to make sure resync covers any new space.
6914 * If the array is shrinking we should possibly wait until
6915 * any io in the removed space completes, but it hardly seems
6919 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6920 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
6921 if (mddev->external_size &&
6922 mddev->array_sectors > newsize)
6924 if (mddev->bitmap) {
6925 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
6929 md_set_array_sectors(mddev, newsize);
6930 set_capacity(mddev->gendisk, mddev->array_sectors);
6931 revalidate_disk(mddev->gendisk);
6932 if (sectors > mddev->dev_sectors &&
6933 mddev->recovery_cp > mddev->dev_sectors) {
6934 mddev->recovery_cp = mddev->dev_sectors;
6935 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
6937 mddev->dev_sectors = sectors;
6938 mddev->resync_max_sectors = sectors;
6942 static int check_stripe_cache(struct mddev *mddev)
6944 /* Can only proceed if there are plenty of stripe_heads.
6945 * We need a minimum of one full stripe,, and for sensible progress
6946 * it is best to have about 4 times that.
6947 * If we require 4 times, then the default 256 4K stripe_heads will
6948 * allow for chunk sizes up to 256K, which is probably OK.
6949 * If the chunk size is greater, user-space should request more
6950 * stripe_heads first.
6952 struct r5conf *conf = mddev->private;
6953 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
6954 > conf->max_nr_stripes ||
6955 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
6956 > conf->max_nr_stripes) {
6957 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
6959 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
6966 static int check_reshape(struct mddev *mddev)
6968 struct r5conf *conf = mddev->private;
6970 if (mddev->delta_disks == 0 &&
6971 mddev->new_layout == mddev->layout &&
6972 mddev->new_chunk_sectors == mddev->chunk_sectors)
6973 return 0; /* nothing to do */
6974 if (has_failed(conf))
6976 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
6977 /* We might be able to shrink, but the devices must
6978 * be made bigger first.
6979 * For raid6, 4 is the minimum size.
6980 * Otherwise 2 is the minimum
6983 if (mddev->level == 6)
6985 if (mddev->raid_disks + mddev->delta_disks < min)
6989 if (!check_stripe_cache(mddev))
6992 return resize_stripes(conf, (conf->previous_raid_disks
6993 + mddev->delta_disks));
6996 static int raid5_start_reshape(struct mddev *mddev)
6998 struct r5conf *conf = mddev->private;
6999 struct md_rdev *rdev;
7001 unsigned long flags;
7003 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7006 if (!check_stripe_cache(mddev))
7009 if (has_failed(conf))
7012 rdev_for_each(rdev, mddev) {
7013 if (!test_bit(In_sync, &rdev->flags)
7014 && !test_bit(Faulty, &rdev->flags))
7018 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7019 /* Not enough devices even to make a degraded array
7024 /* Refuse to reduce size of the array. Any reductions in
7025 * array size must be through explicit setting of array_size
7028 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7029 < mddev->array_sectors) {
7030 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7031 "before number of disks\n", mdname(mddev));
7035 atomic_set(&conf->reshape_stripes, 0);
7036 spin_lock_irq(&conf->device_lock);
7037 write_seqcount_begin(&conf->gen_lock);
7038 conf->previous_raid_disks = conf->raid_disks;
7039 conf->raid_disks += mddev->delta_disks;
7040 conf->prev_chunk_sectors = conf->chunk_sectors;
7041 conf->chunk_sectors = mddev->new_chunk_sectors;
7042 conf->prev_algo = conf->algorithm;
7043 conf->algorithm = mddev->new_layout;
7045 /* Code that selects data_offset needs to see the generation update
7046 * if reshape_progress has been set - so a memory barrier needed.
7049 if (mddev->reshape_backwards)
7050 conf->reshape_progress = raid5_size(mddev, 0, 0);
7052 conf->reshape_progress = 0;
7053 conf->reshape_safe = conf->reshape_progress;
7054 write_seqcount_end(&conf->gen_lock);
7055 spin_unlock_irq(&conf->device_lock);
7057 /* Now make sure any requests that proceeded on the assumption
7058 * the reshape wasn't running - like Discard or Read - have
7061 mddev_suspend(mddev);
7062 mddev_resume(mddev);
7064 /* Add some new drives, as many as will fit.
7065 * We know there are enough to make the newly sized array work.
7066 * Don't add devices if we are reducing the number of
7067 * devices in the array. This is because it is not possible
7068 * to correctly record the "partially reconstructed" state of
7069 * such devices during the reshape and confusion could result.
7071 if (mddev->delta_disks >= 0) {
7072 rdev_for_each(rdev, mddev)
7073 if (rdev->raid_disk < 0 &&
7074 !test_bit(Faulty, &rdev->flags)) {
7075 if (raid5_add_disk(mddev, rdev) == 0) {
7077 >= conf->previous_raid_disks)
7078 set_bit(In_sync, &rdev->flags);
7080 rdev->recovery_offset = 0;
7082 if (sysfs_link_rdev(mddev, rdev))
7083 /* Failure here is OK */;
7085 } else if (rdev->raid_disk >= conf->previous_raid_disks
7086 && !test_bit(Faulty, &rdev->flags)) {
7087 /* This is a spare that was manually added */
7088 set_bit(In_sync, &rdev->flags);
7091 /* When a reshape changes the number of devices,
7092 * ->degraded is measured against the larger of the
7093 * pre and post number of devices.
7095 spin_lock_irqsave(&conf->device_lock, flags);
7096 mddev->degraded = calc_degraded(conf);
7097 spin_unlock_irqrestore(&conf->device_lock, flags);
7099 mddev->raid_disks = conf->raid_disks;
7100 mddev->reshape_position = conf->reshape_progress;
7101 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7103 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7104 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7105 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7106 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7107 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7109 if (!mddev->sync_thread) {
7110 mddev->recovery = 0;
7111 spin_lock_irq(&conf->device_lock);
7112 write_seqcount_begin(&conf->gen_lock);
7113 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7114 mddev->new_chunk_sectors =
7115 conf->chunk_sectors = conf->prev_chunk_sectors;
7116 mddev->new_layout = conf->algorithm = conf->prev_algo;
7117 rdev_for_each(rdev, mddev)
7118 rdev->new_data_offset = rdev->data_offset;
7120 conf->generation --;
7121 conf->reshape_progress = MaxSector;
7122 mddev->reshape_position = MaxSector;
7123 write_seqcount_end(&conf->gen_lock);
7124 spin_unlock_irq(&conf->device_lock);
7127 conf->reshape_checkpoint = jiffies;
7128 md_wakeup_thread(mddev->sync_thread);
7129 md_new_event(mddev);
7133 /* This is called from the reshape thread and should make any
7134 * changes needed in 'conf'
7136 static void end_reshape(struct r5conf *conf)
7139 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7140 struct md_rdev *rdev;
7142 spin_lock_irq(&conf->device_lock);
7143 conf->previous_raid_disks = conf->raid_disks;
7144 rdev_for_each(rdev, conf->mddev)
7145 rdev->data_offset = rdev->new_data_offset;
7147 conf->reshape_progress = MaxSector;
7148 spin_unlock_irq(&conf->device_lock);
7149 wake_up(&conf->wait_for_overlap);
7151 /* read-ahead size must cover two whole stripes, which is
7152 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7154 if (conf->mddev->queue) {
7155 int data_disks = conf->raid_disks - conf->max_degraded;
7156 int stripe = data_disks * ((conf->chunk_sectors << 9)
7158 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7159 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7164 /* This is called from the raid5d thread with mddev_lock held.
7165 * It makes config changes to the device.
7167 static void raid5_finish_reshape(struct mddev *mddev)
7169 struct r5conf *conf = mddev->private;
7171 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7173 if (mddev->delta_disks > 0) {
7174 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7175 set_capacity(mddev->gendisk, mddev->array_sectors);
7176 revalidate_disk(mddev->gendisk);
7179 spin_lock_irq(&conf->device_lock);
7180 mddev->degraded = calc_degraded(conf);
7181 spin_unlock_irq(&conf->device_lock);
7182 for (d = conf->raid_disks ;
7183 d < conf->raid_disks - mddev->delta_disks;
7185 struct md_rdev *rdev = conf->disks[d].rdev;
7187 clear_bit(In_sync, &rdev->flags);
7188 rdev = conf->disks[d].replacement;
7190 clear_bit(In_sync, &rdev->flags);
7193 mddev->layout = conf->algorithm;
7194 mddev->chunk_sectors = conf->chunk_sectors;
7195 mddev->reshape_position = MaxSector;
7196 mddev->delta_disks = 0;
7197 mddev->reshape_backwards = 0;
7201 static void raid5_quiesce(struct mddev *mddev, int state)
7203 struct r5conf *conf = mddev->private;
7206 case 2: /* resume for a suspend */
7207 wake_up(&conf->wait_for_overlap);
7210 case 1: /* stop all writes */
7211 lock_all_device_hash_locks_irq(conf);
7212 /* '2' tells resync/reshape to pause so that all
7213 * active stripes can drain
7216 wait_event_cmd(conf->wait_for_stripe,
7217 atomic_read(&conf->active_stripes) == 0 &&
7218 atomic_read(&conf->active_aligned_reads) == 0,
7219 unlock_all_device_hash_locks_irq(conf),
7220 lock_all_device_hash_locks_irq(conf));
7222 unlock_all_device_hash_locks_irq(conf);
7223 /* allow reshape to continue */
7224 wake_up(&conf->wait_for_overlap);
7227 case 0: /* re-enable writes */
7228 lock_all_device_hash_locks_irq(conf);
7230 wake_up(&conf->wait_for_stripe);
7231 wake_up(&conf->wait_for_overlap);
7232 unlock_all_device_hash_locks_irq(conf);
7237 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7239 struct r0conf *raid0_conf = mddev->private;
7242 /* for raid0 takeover only one zone is supported */
7243 if (raid0_conf->nr_strip_zones > 1) {
7244 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7246 return ERR_PTR(-EINVAL);
7249 sectors = raid0_conf->strip_zone[0].zone_end;
7250 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7251 mddev->dev_sectors = sectors;
7252 mddev->new_level = level;
7253 mddev->new_layout = ALGORITHM_PARITY_N;
7254 mddev->new_chunk_sectors = mddev->chunk_sectors;
7255 mddev->raid_disks += 1;
7256 mddev->delta_disks = 1;
7257 /* make sure it will be not marked as dirty */
7258 mddev->recovery_cp = MaxSector;
7260 return setup_conf(mddev);
7263 static void *raid5_takeover_raid1(struct mddev *mddev)
7267 if (mddev->raid_disks != 2 ||
7268 mddev->degraded > 1)
7269 return ERR_PTR(-EINVAL);
7271 /* Should check if there are write-behind devices? */
7273 chunksect = 64*2; /* 64K by default */
7275 /* The array must be an exact multiple of chunksize */
7276 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7279 if ((chunksect<<9) < STRIPE_SIZE)
7280 /* array size does not allow a suitable chunk size */
7281 return ERR_PTR(-EINVAL);
7283 mddev->new_level = 5;
7284 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7285 mddev->new_chunk_sectors = chunksect;
7287 return setup_conf(mddev);
7290 static void *raid5_takeover_raid6(struct mddev *mddev)
7294 switch (mddev->layout) {
7295 case ALGORITHM_LEFT_ASYMMETRIC_6:
7296 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7298 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7299 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7301 case ALGORITHM_LEFT_SYMMETRIC_6:
7302 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7304 case ALGORITHM_RIGHT_SYMMETRIC_6:
7305 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7307 case ALGORITHM_PARITY_0_6:
7308 new_layout = ALGORITHM_PARITY_0;
7310 case ALGORITHM_PARITY_N:
7311 new_layout = ALGORITHM_PARITY_N;
7314 return ERR_PTR(-EINVAL);
7316 mddev->new_level = 5;
7317 mddev->new_layout = new_layout;
7318 mddev->delta_disks = -1;
7319 mddev->raid_disks -= 1;
7320 return setup_conf(mddev);
7323 static int raid5_check_reshape(struct mddev *mddev)
7325 /* For a 2-drive array, the layout and chunk size can be changed
7326 * immediately as not restriping is needed.
7327 * For larger arrays we record the new value - after validation
7328 * to be used by a reshape pass.
7330 struct r5conf *conf = mddev->private;
7331 int new_chunk = mddev->new_chunk_sectors;
7333 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7335 if (new_chunk > 0) {
7336 if (!is_power_of_2(new_chunk))
7338 if (new_chunk < (PAGE_SIZE>>9))
7340 if (mddev->array_sectors & (new_chunk-1))
7341 /* not factor of array size */
7345 /* They look valid */
7347 if (mddev->raid_disks == 2) {
7348 /* can make the change immediately */
7349 if (mddev->new_layout >= 0) {
7350 conf->algorithm = mddev->new_layout;
7351 mddev->layout = mddev->new_layout;
7353 if (new_chunk > 0) {
7354 conf->chunk_sectors = new_chunk ;
7355 mddev->chunk_sectors = new_chunk;
7357 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7358 md_wakeup_thread(mddev->thread);
7360 return check_reshape(mddev);
7363 static int raid6_check_reshape(struct mddev *mddev)
7365 int new_chunk = mddev->new_chunk_sectors;
7367 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7369 if (new_chunk > 0) {
7370 if (!is_power_of_2(new_chunk))
7372 if (new_chunk < (PAGE_SIZE >> 9))
7374 if (mddev->array_sectors & (new_chunk-1))
7375 /* not factor of array size */
7379 /* They look valid */
7380 return check_reshape(mddev);
7383 static void *raid5_takeover(struct mddev *mddev)
7385 /* raid5 can take over:
7386 * raid0 - if there is only one strip zone - make it a raid4 layout
7387 * raid1 - if there are two drives. We need to know the chunk size
7388 * raid4 - trivial - just use a raid4 layout.
7389 * raid6 - Providing it is a *_6 layout
7391 if (mddev->level == 0)
7392 return raid45_takeover_raid0(mddev, 5);
7393 if (mddev->level == 1)
7394 return raid5_takeover_raid1(mddev);
7395 if (mddev->level == 4) {
7396 mddev->new_layout = ALGORITHM_PARITY_N;
7397 mddev->new_level = 5;
7398 return setup_conf(mddev);
7400 if (mddev->level == 6)
7401 return raid5_takeover_raid6(mddev);
7403 return ERR_PTR(-EINVAL);
7406 static void *raid4_takeover(struct mddev *mddev)
7408 /* raid4 can take over:
7409 * raid0 - if there is only one strip zone
7410 * raid5 - if layout is right
7412 if (mddev->level == 0)
7413 return raid45_takeover_raid0(mddev, 4);
7414 if (mddev->level == 5 &&
7415 mddev->layout == ALGORITHM_PARITY_N) {
7416 mddev->new_layout = 0;
7417 mddev->new_level = 4;
7418 return setup_conf(mddev);
7420 return ERR_PTR(-EINVAL);
7423 static struct md_personality raid5_personality;
7425 static void *raid6_takeover(struct mddev *mddev)
7427 /* Currently can only take over a raid5. We map the
7428 * personality to an equivalent raid6 personality
7429 * with the Q block at the end.
7433 if (mddev->pers != &raid5_personality)
7434 return ERR_PTR(-EINVAL);
7435 if (mddev->degraded > 1)
7436 return ERR_PTR(-EINVAL);
7437 if (mddev->raid_disks > 253)
7438 return ERR_PTR(-EINVAL);
7439 if (mddev->raid_disks < 3)
7440 return ERR_PTR(-EINVAL);
7442 switch (mddev->layout) {
7443 case ALGORITHM_LEFT_ASYMMETRIC:
7444 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7446 case ALGORITHM_RIGHT_ASYMMETRIC:
7447 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7449 case ALGORITHM_LEFT_SYMMETRIC:
7450 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7452 case ALGORITHM_RIGHT_SYMMETRIC:
7453 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7455 case ALGORITHM_PARITY_0:
7456 new_layout = ALGORITHM_PARITY_0_6;
7458 case ALGORITHM_PARITY_N:
7459 new_layout = ALGORITHM_PARITY_N;
7462 return ERR_PTR(-EINVAL);
7464 mddev->new_level = 6;
7465 mddev->new_layout = new_layout;
7466 mddev->delta_disks = 1;
7467 mddev->raid_disks += 1;
7468 return setup_conf(mddev);
7471 static struct md_personality raid6_personality =
7475 .owner = THIS_MODULE,
7476 .make_request = make_request,
7480 .error_handler = error,
7481 .hot_add_disk = raid5_add_disk,
7482 .hot_remove_disk= raid5_remove_disk,
7483 .spare_active = raid5_spare_active,
7484 .sync_request = sync_request,
7485 .resize = raid5_resize,
7487 .check_reshape = raid6_check_reshape,
7488 .start_reshape = raid5_start_reshape,
7489 .finish_reshape = raid5_finish_reshape,
7490 .quiesce = raid5_quiesce,
7491 .takeover = raid6_takeover,
7492 .congested = raid5_congested,
7493 .mergeable_bvec = raid5_mergeable_bvec,
7495 static struct md_personality raid5_personality =
7499 .owner = THIS_MODULE,
7500 .make_request = make_request,
7504 .error_handler = error,
7505 .hot_add_disk = raid5_add_disk,
7506 .hot_remove_disk= raid5_remove_disk,
7507 .spare_active = raid5_spare_active,
7508 .sync_request = sync_request,
7509 .resize = raid5_resize,
7511 .check_reshape = raid5_check_reshape,
7512 .start_reshape = raid5_start_reshape,
7513 .finish_reshape = raid5_finish_reshape,
7514 .quiesce = raid5_quiesce,
7515 .takeover = raid5_takeover,
7516 .congested = raid5_congested,
7517 .mergeable_bvec = raid5_mergeable_bvec,
7520 static struct md_personality raid4_personality =
7524 .owner = THIS_MODULE,
7525 .make_request = make_request,
7529 .error_handler = error,
7530 .hot_add_disk = raid5_add_disk,
7531 .hot_remove_disk= raid5_remove_disk,
7532 .spare_active = raid5_spare_active,
7533 .sync_request = sync_request,
7534 .resize = raid5_resize,
7536 .check_reshape = raid5_check_reshape,
7537 .start_reshape = raid5_start_reshape,
7538 .finish_reshape = raid5_finish_reshape,
7539 .quiesce = raid5_quiesce,
7540 .takeover = raid4_takeover,
7541 .congested = raid5_congested,
7542 .mergeable_bvec = raid5_mergeable_bvec,
7545 static int __init raid5_init(void)
7547 raid5_wq = alloc_workqueue("raid5wq",
7548 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7551 register_md_personality(&raid6_personality);
7552 register_md_personality(&raid5_personality);
7553 register_md_personality(&raid4_personality);
7557 static void raid5_exit(void)
7559 unregister_md_personality(&raid6_personality);
7560 unregister_md_personality(&raid5_personality);
7561 unregister_md_personality(&raid4_personality);
7562 destroy_workqueue(raid5_wq);
7565 module_init(raid5_init);
7566 module_exit(raid5_exit);
7567 MODULE_LICENSE("GPL");
7568 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7569 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7570 MODULE_ALIAS("md-raid5");
7571 MODULE_ALIAS("md-raid4");
7572 MODULE_ALIAS("md-level-5");
7573 MODULE_ALIAS("md-level-4");
7574 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7575 MODULE_ALIAS("md-raid6");
7576 MODULE_ALIAS("md-level-6");
7578 /* This used to be two separate modules, they were: */
7579 MODULE_ALIAS("raid5");
7580 MODULE_ALIAS("raid6");