2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
58 #include <trace/events/block.h>
65 #define cpu_to_group(cpu) cpu_to_node(cpu)
66 #define ANY_GROUP NUMA_NO_NODE
68 static bool devices_handle_discard_safely = false;
69 module_param(devices_handle_discard_safely, bool, 0644);
70 MODULE_PARM_DESC(devices_handle_discard_safely,
71 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
72 static struct workqueue_struct *raid5_wq;
77 #define NR_STRIPES 256
78 #define STRIPE_SIZE PAGE_SIZE
79 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
80 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
81 #define IO_THRESHOLD 1
82 #define BYPASS_THRESHOLD 1
83 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
84 #define HASH_MASK (NR_HASH - 1)
85 #define MAX_STRIPE_BATCH 8
87 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
89 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
90 return &conf->stripe_hashtbl[hash];
93 static inline int stripe_hash_locks_hash(sector_t sect)
95 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
98 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
100 spin_lock_irq(conf->hash_locks + hash);
101 spin_lock(&conf->device_lock);
104 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
106 spin_unlock(&conf->device_lock);
107 spin_unlock_irq(conf->hash_locks + hash);
110 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
114 spin_lock(conf->hash_locks);
115 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
116 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
117 spin_lock(&conf->device_lock);
120 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
123 spin_unlock(&conf->device_lock);
124 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
125 spin_unlock(conf->hash_locks + i - 1);
129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
130 * order without overlap. There may be several bio's per stripe+device, and
131 * a bio could span several devices.
132 * When walking this list for a particular stripe+device, we must never proceed
133 * beyond a bio that extends past this device, as the next bio might no longer
135 * This function is used to determine the 'next' bio in the list, given the sector
136 * of the current stripe+device
138 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
140 int sectors = bio_sectors(bio);
141 if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
148 * We maintain a biased count of active stripes in the bottom 16 bits of
149 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
151 static inline int raid5_bi_processed_stripes(struct bio *bio)
153 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
154 return (atomic_read(segments) >> 16) & 0xffff;
157 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
159 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
160 return atomic_sub_return(1, segments) & 0xffff;
163 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
165 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
166 atomic_inc(segments);
169 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
172 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
176 old = atomic_read(segments);
177 new = (old & 0xffff) | (cnt << 16);
178 } while (atomic_cmpxchg(segments, old, new) != old);
181 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
183 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
184 atomic_set(segments, cnt);
187 /* Find first data disk in a raid6 stripe */
188 static inline int raid6_d0(struct stripe_head *sh)
191 /* ddf always start from first device */
193 /* md starts just after Q block */
194 if (sh->qd_idx == sh->disks - 1)
197 return sh->qd_idx + 1;
199 static inline int raid6_next_disk(int disk, int raid_disks)
202 return (disk < raid_disks) ? disk : 0;
205 /* When walking through the disks in a raid5, starting at raid6_d0,
206 * We need to map each disk to a 'slot', where the data disks are slot
207 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
208 * is raid_disks-1. This help does that mapping.
210 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
211 int *count, int syndrome_disks)
217 if (idx == sh->pd_idx)
218 return syndrome_disks;
219 if (idx == sh->qd_idx)
220 return syndrome_disks + 1;
226 static void return_io(struct bio *return_bi)
228 struct bio *bi = return_bi;
231 return_bi = bi->bi_next;
233 bi->bi_iter.bi_size = 0;
234 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
241 static void print_raid5_conf (struct r5conf *conf);
243 static int stripe_operations_active(struct stripe_head *sh)
245 return sh->check_state || sh->reconstruct_state ||
246 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
247 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
250 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
252 struct r5conf *conf = sh->raid_conf;
253 struct r5worker_group *group;
255 int i, cpu = sh->cpu;
257 if (!cpu_online(cpu)) {
258 cpu = cpumask_any(cpu_online_mask);
262 if (list_empty(&sh->lru)) {
263 struct r5worker_group *group;
264 group = conf->worker_groups + cpu_to_group(cpu);
265 list_add_tail(&sh->lru, &group->handle_list);
266 group->stripes_cnt++;
270 if (conf->worker_cnt_per_group == 0) {
271 md_wakeup_thread(conf->mddev->thread);
275 group = conf->worker_groups + cpu_to_group(sh->cpu);
277 group->workers[0].working = true;
278 /* at least one worker should run to avoid race */
279 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
281 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
282 /* wakeup more workers */
283 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
284 if (group->workers[i].working == false) {
285 group->workers[i].working = true;
286 queue_work_on(sh->cpu, raid5_wq,
287 &group->workers[i].work);
293 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
294 struct list_head *temp_inactive_list)
296 BUG_ON(!list_empty(&sh->lru));
297 BUG_ON(atomic_read(&conf->active_stripes)==0);
298 if (test_bit(STRIPE_HANDLE, &sh->state)) {
299 if (test_bit(STRIPE_DELAYED, &sh->state) &&
300 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
301 list_add_tail(&sh->lru, &conf->delayed_list);
302 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
303 sh->bm_seq - conf->seq_write > 0)
304 list_add_tail(&sh->lru, &conf->bitmap_list);
306 clear_bit(STRIPE_DELAYED, &sh->state);
307 clear_bit(STRIPE_BIT_DELAY, &sh->state);
308 if (conf->worker_cnt_per_group == 0) {
309 list_add_tail(&sh->lru, &conf->handle_list);
311 raid5_wakeup_stripe_thread(sh);
315 md_wakeup_thread(conf->mddev->thread);
317 BUG_ON(stripe_operations_active(sh));
318 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
319 if (atomic_dec_return(&conf->preread_active_stripes)
321 md_wakeup_thread(conf->mddev->thread);
322 atomic_dec(&conf->active_stripes);
323 if (!test_bit(STRIPE_EXPANDING, &sh->state))
324 list_add_tail(&sh->lru, temp_inactive_list);
328 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
329 struct list_head *temp_inactive_list)
331 if (atomic_dec_and_test(&sh->count))
332 do_release_stripe(conf, sh, temp_inactive_list);
336 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
338 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
339 * given time. Adding stripes only takes device lock, while deleting stripes
340 * only takes hash lock.
342 static void release_inactive_stripe_list(struct r5conf *conf,
343 struct list_head *temp_inactive_list,
347 bool do_wakeup = false;
350 if (hash == NR_STRIPE_HASH_LOCKS) {
351 size = NR_STRIPE_HASH_LOCKS;
352 hash = NR_STRIPE_HASH_LOCKS - 1;
356 struct list_head *list = &temp_inactive_list[size - 1];
359 * We don't hold any lock here yet, get_active_stripe() might
360 * remove stripes from the list
362 if (!list_empty_careful(list)) {
363 spin_lock_irqsave(conf->hash_locks + hash, flags);
364 if (list_empty(conf->inactive_list + hash) &&
366 atomic_dec(&conf->empty_inactive_list_nr);
367 list_splice_tail_init(list, conf->inactive_list + hash);
369 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
376 wake_up(&conf->wait_for_stripe);
377 if (conf->retry_read_aligned)
378 md_wakeup_thread(conf->mddev->thread);
382 /* should hold conf->device_lock already */
383 static int release_stripe_list(struct r5conf *conf,
384 struct list_head *temp_inactive_list)
386 struct stripe_head *sh;
388 struct llist_node *head;
390 head = llist_del_all(&conf->released_stripes);
391 head = llist_reverse_order(head);
395 sh = llist_entry(head, struct stripe_head, release_list);
396 head = llist_next(head);
397 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
399 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
401 * Don't worry the bit is set here, because if the bit is set
402 * again, the count is always > 1. This is true for
403 * STRIPE_ON_UNPLUG_LIST bit too.
405 hash = sh->hash_lock_index;
406 __release_stripe(conf, sh, &temp_inactive_list[hash]);
413 static void release_stripe(struct stripe_head *sh)
415 struct r5conf *conf = sh->raid_conf;
417 struct list_head list;
421 /* Avoid release_list until the last reference.
423 if (atomic_add_unless(&sh->count, -1, 1))
426 if (unlikely(!conf->mddev->thread) ||
427 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
429 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
431 md_wakeup_thread(conf->mddev->thread);
434 local_irq_save(flags);
435 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
436 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
437 INIT_LIST_HEAD(&list);
438 hash = sh->hash_lock_index;
439 do_release_stripe(conf, sh, &list);
440 spin_unlock(&conf->device_lock);
441 release_inactive_stripe_list(conf, &list, hash);
443 local_irq_restore(flags);
446 static inline void remove_hash(struct stripe_head *sh)
448 pr_debug("remove_hash(), stripe %llu\n",
449 (unsigned long long)sh->sector);
451 hlist_del_init(&sh->hash);
454 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
456 struct hlist_head *hp = stripe_hash(conf, sh->sector);
458 pr_debug("insert_hash(), stripe %llu\n",
459 (unsigned long long)sh->sector);
461 hlist_add_head(&sh->hash, hp);
464 /* find an idle stripe, make sure it is unhashed, and return it. */
465 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
467 struct stripe_head *sh = NULL;
468 struct list_head *first;
470 if (list_empty(conf->inactive_list + hash))
472 first = (conf->inactive_list + hash)->next;
473 sh = list_entry(first, struct stripe_head, lru);
474 list_del_init(first);
476 atomic_inc(&conf->active_stripes);
477 BUG_ON(hash != sh->hash_lock_index);
478 if (list_empty(conf->inactive_list + hash))
479 atomic_inc(&conf->empty_inactive_list_nr);
484 static void shrink_buffers(struct stripe_head *sh)
488 int num = sh->raid_conf->pool_size;
490 for (i = 0; i < num ; i++) {
491 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
495 sh->dev[i].page = NULL;
500 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
503 int num = sh->raid_conf->pool_size;
505 for (i = 0; i < num; i++) {
508 if (!(page = alloc_page(gfp))) {
511 sh->dev[i].page = page;
512 sh->dev[i].orig_page = page;
517 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
518 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
519 struct stripe_head *sh);
521 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
523 struct r5conf *conf = sh->raid_conf;
526 BUG_ON(atomic_read(&sh->count) != 0);
527 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
528 BUG_ON(stripe_operations_active(sh));
529 BUG_ON(sh->batch_head);
531 pr_debug("init_stripe called, stripe %llu\n",
532 (unsigned long long)sector);
534 seq = read_seqcount_begin(&conf->gen_lock);
535 sh->generation = conf->generation - previous;
536 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
538 stripe_set_idx(sector, conf, previous, sh);
541 for (i = sh->disks; i--; ) {
542 struct r5dev *dev = &sh->dev[i];
544 if (dev->toread || dev->read || dev->towrite || dev->written ||
545 test_bit(R5_LOCKED, &dev->flags)) {
546 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
547 (unsigned long long)sh->sector, i, dev->toread,
548 dev->read, dev->towrite, dev->written,
549 test_bit(R5_LOCKED, &dev->flags));
553 raid5_build_block(sh, i, previous);
555 if (read_seqcount_retry(&conf->gen_lock, seq))
557 sh->overwrite_disks = 0;
558 insert_hash(conf, sh);
559 sh->cpu = smp_processor_id();
560 set_bit(STRIPE_BATCH_READY, &sh->state);
563 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
566 struct stripe_head *sh;
568 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
569 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
570 if (sh->sector == sector && sh->generation == generation)
572 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
577 * Need to check if array has failed when deciding whether to:
579 * - remove non-faulty devices
582 * This determination is simple when no reshape is happening.
583 * However if there is a reshape, we need to carefully check
584 * both the before and after sections.
585 * This is because some failed devices may only affect one
586 * of the two sections, and some non-in_sync devices may
587 * be insync in the section most affected by failed devices.
589 static int calc_degraded(struct r5conf *conf)
591 int degraded, degraded2;
596 for (i = 0; i < conf->previous_raid_disks; i++) {
597 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
598 if (rdev && test_bit(Faulty, &rdev->flags))
599 rdev = rcu_dereference(conf->disks[i].replacement);
600 if (!rdev || test_bit(Faulty, &rdev->flags))
602 else if (test_bit(In_sync, &rdev->flags))
605 /* not in-sync or faulty.
606 * If the reshape increases the number of devices,
607 * this is being recovered by the reshape, so
608 * this 'previous' section is not in_sync.
609 * If the number of devices is being reduced however,
610 * the device can only be part of the array if
611 * we are reverting a reshape, so this section will
614 if (conf->raid_disks >= conf->previous_raid_disks)
618 if (conf->raid_disks == conf->previous_raid_disks)
622 for (i = 0; i < conf->raid_disks; i++) {
623 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
624 if (rdev && test_bit(Faulty, &rdev->flags))
625 rdev = rcu_dereference(conf->disks[i].replacement);
626 if (!rdev || test_bit(Faulty, &rdev->flags))
628 else if (test_bit(In_sync, &rdev->flags))
631 /* not in-sync or faulty.
632 * If reshape increases the number of devices, this
633 * section has already been recovered, else it
634 * almost certainly hasn't.
636 if (conf->raid_disks <= conf->previous_raid_disks)
640 if (degraded2 > degraded)
645 static int has_failed(struct r5conf *conf)
649 if (conf->mddev->reshape_position == MaxSector)
650 return conf->mddev->degraded > conf->max_degraded;
652 degraded = calc_degraded(conf);
653 if (degraded > conf->max_degraded)
658 static struct stripe_head *
659 get_active_stripe(struct r5conf *conf, sector_t sector,
660 int previous, int noblock, int noquiesce)
662 struct stripe_head *sh;
663 int hash = stripe_hash_locks_hash(sector);
665 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
667 spin_lock_irq(conf->hash_locks + hash);
670 wait_event_lock_irq(conf->wait_for_stripe,
671 conf->quiesce == 0 || noquiesce,
672 *(conf->hash_locks + hash));
673 sh = __find_stripe(conf, sector, conf->generation - previous);
675 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
676 sh = get_free_stripe(conf, hash);
677 if (!sh && llist_empty(&conf->released_stripes) &&
678 !test_bit(R5_DID_ALLOC, &conf->cache_state))
679 set_bit(R5_ALLOC_MORE,
682 if (noblock && sh == NULL)
685 set_bit(R5_INACTIVE_BLOCKED,
688 conf->wait_for_stripe,
689 !list_empty(conf->inactive_list + hash) &&
690 (atomic_read(&conf->active_stripes)
691 < (conf->max_nr_stripes * 3 / 4)
692 || !test_bit(R5_INACTIVE_BLOCKED,
693 &conf->cache_state)),
694 *(conf->hash_locks + hash));
695 clear_bit(R5_INACTIVE_BLOCKED,
698 init_stripe(sh, sector, previous);
699 atomic_inc(&sh->count);
701 } else if (!atomic_inc_not_zero(&sh->count)) {
702 spin_lock(&conf->device_lock);
703 if (!atomic_read(&sh->count)) {
704 if (!test_bit(STRIPE_HANDLE, &sh->state))
705 atomic_inc(&conf->active_stripes);
706 BUG_ON(list_empty(&sh->lru) &&
707 !test_bit(STRIPE_EXPANDING, &sh->state));
708 list_del_init(&sh->lru);
710 sh->group->stripes_cnt--;
714 atomic_inc(&sh->count);
715 spin_unlock(&conf->device_lock);
717 } while (sh == NULL);
719 spin_unlock_irq(conf->hash_locks + hash);
723 static bool is_full_stripe_write(struct stripe_head *sh)
725 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
726 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
729 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
733 spin_lock(&sh2->stripe_lock);
734 spin_lock_nested(&sh1->stripe_lock, 1);
736 spin_lock(&sh1->stripe_lock);
737 spin_lock_nested(&sh2->stripe_lock, 1);
741 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
743 spin_unlock(&sh1->stripe_lock);
744 spin_unlock(&sh2->stripe_lock);
748 /* Only freshly new full stripe normal write stripe can be added to a batch list */
749 static bool stripe_can_batch(struct stripe_head *sh)
751 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
752 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
753 is_full_stripe_write(sh);
756 /* we only do back search */
757 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
759 struct stripe_head *head;
760 sector_t head_sector, tmp_sec;
764 if (!stripe_can_batch(sh))
766 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
767 tmp_sec = sh->sector;
768 if (!sector_div(tmp_sec, conf->chunk_sectors))
770 head_sector = sh->sector - STRIPE_SECTORS;
772 hash = stripe_hash_locks_hash(head_sector);
773 spin_lock_irq(conf->hash_locks + hash);
774 head = __find_stripe(conf, head_sector, conf->generation);
775 if (head && !atomic_inc_not_zero(&head->count)) {
776 spin_lock(&conf->device_lock);
777 if (!atomic_read(&head->count)) {
778 if (!test_bit(STRIPE_HANDLE, &head->state))
779 atomic_inc(&conf->active_stripes);
780 BUG_ON(list_empty(&head->lru) &&
781 !test_bit(STRIPE_EXPANDING, &head->state));
782 list_del_init(&head->lru);
784 head->group->stripes_cnt--;
788 atomic_inc(&head->count);
789 spin_unlock(&conf->device_lock);
791 spin_unlock_irq(conf->hash_locks + hash);
795 if (!stripe_can_batch(head))
798 lock_two_stripes(head, sh);
799 /* clear_batch_ready clear the flag */
800 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
807 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
809 if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
812 if (head->batch_head) {
813 spin_lock(&head->batch_head->batch_lock);
814 /* This batch list is already running */
815 if (!stripe_can_batch(head)) {
816 spin_unlock(&head->batch_head->batch_lock);
821 * at this point, head's BATCH_READY could be cleared, but we
822 * can still add the stripe to batch list
824 list_add(&sh->batch_list, &head->batch_list);
825 spin_unlock(&head->batch_head->batch_lock);
827 sh->batch_head = head->batch_head;
829 head->batch_head = head;
830 sh->batch_head = head->batch_head;
831 spin_lock(&head->batch_lock);
832 list_add_tail(&sh->batch_list, &head->batch_list);
833 spin_unlock(&head->batch_lock);
836 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
837 if (atomic_dec_return(&conf->preread_active_stripes)
839 md_wakeup_thread(conf->mddev->thread);
841 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
842 int seq = sh->bm_seq;
843 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
844 sh->batch_head->bm_seq > seq)
845 seq = sh->batch_head->bm_seq;
846 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
847 sh->batch_head->bm_seq = seq;
850 atomic_inc(&sh->count);
852 unlock_two_stripes(head, sh);
854 release_stripe(head);
857 /* Determine if 'data_offset' or 'new_data_offset' should be used
858 * in this stripe_head.
860 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
862 sector_t progress = conf->reshape_progress;
863 /* Need a memory barrier to make sure we see the value
864 * of conf->generation, or ->data_offset that was set before
865 * reshape_progress was updated.
868 if (progress == MaxSector)
870 if (sh->generation == conf->generation - 1)
872 /* We are in a reshape, and this is a new-generation stripe,
873 * so use new_data_offset.
879 raid5_end_read_request(struct bio *bi, int error);
881 raid5_end_write_request(struct bio *bi, int error);
883 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
885 struct r5conf *conf = sh->raid_conf;
886 int i, disks = sh->disks;
887 struct stripe_head *head_sh = sh;
891 for (i = disks; i--; ) {
893 int replace_only = 0;
894 struct bio *bi, *rbi;
895 struct md_rdev *rdev, *rrdev = NULL;
898 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
899 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
903 if (test_bit(R5_Discard, &sh->dev[i].flags))
905 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
907 else if (test_and_clear_bit(R5_WantReplace,
908 &sh->dev[i].flags)) {
913 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
917 bi = &sh->dev[i].req;
918 rbi = &sh->dev[i].rreq; /* For writing to replacement */
921 rrdev = rcu_dereference(conf->disks[i].replacement);
922 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
923 rdev = rcu_dereference(conf->disks[i].rdev);
932 /* We raced and saw duplicates */
935 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
940 if (rdev && test_bit(Faulty, &rdev->flags))
943 atomic_inc(&rdev->nr_pending);
944 if (rrdev && test_bit(Faulty, &rrdev->flags))
947 atomic_inc(&rrdev->nr_pending);
950 /* We have already checked bad blocks for reads. Now
951 * need to check for writes. We never accept write errors
952 * on the replacement, so we don't to check rrdev.
954 while ((rw & WRITE) && rdev &&
955 test_bit(WriteErrorSeen, &rdev->flags)) {
958 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
959 &first_bad, &bad_sectors);
964 set_bit(BlockedBadBlocks, &rdev->flags);
965 if (!conf->mddev->external &&
966 conf->mddev->flags) {
967 /* It is very unlikely, but we might
968 * still need to write out the
969 * bad block log - better give it
971 md_check_recovery(conf->mddev);
974 * Because md_wait_for_blocked_rdev
975 * will dec nr_pending, we must
976 * increment it first.
978 atomic_inc(&rdev->nr_pending);
979 md_wait_for_blocked_rdev(rdev, conf->mddev);
981 /* Acknowledged bad block - skip the write */
982 rdev_dec_pending(rdev, conf->mddev);
988 if (s->syncing || s->expanding || s->expanded
990 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
992 set_bit(STRIPE_IO_STARTED, &sh->state);
995 bi->bi_bdev = rdev->bdev;
997 bi->bi_end_io = (rw & WRITE)
998 ? raid5_end_write_request
999 : raid5_end_read_request;
1000 bi->bi_private = sh;
1002 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1003 __func__, (unsigned long long)sh->sector,
1005 atomic_inc(&sh->count);
1007 atomic_inc(&head_sh->count);
1008 if (use_new_offset(conf, sh))
1009 bi->bi_iter.bi_sector = (sh->sector
1010 + rdev->new_data_offset);
1012 bi->bi_iter.bi_sector = (sh->sector
1013 + rdev->data_offset);
1014 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1015 bi->bi_rw |= REQ_NOMERGE;
1017 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1018 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1019 sh->dev[i].vec.bv_page = sh->dev[i].page;
1021 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1022 bi->bi_io_vec[0].bv_offset = 0;
1023 bi->bi_iter.bi_size = STRIPE_SIZE;
1025 * If this is discard request, set bi_vcnt 0. We don't
1026 * want to confuse SCSI because SCSI will replace payload
1028 if (rw & REQ_DISCARD)
1031 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1033 if (conf->mddev->gendisk)
1034 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1035 bi, disk_devt(conf->mddev->gendisk),
1037 generic_make_request(bi);
1040 if (s->syncing || s->expanding || s->expanded
1042 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1044 set_bit(STRIPE_IO_STARTED, &sh->state);
1047 rbi->bi_bdev = rrdev->bdev;
1049 BUG_ON(!(rw & WRITE));
1050 rbi->bi_end_io = raid5_end_write_request;
1051 rbi->bi_private = sh;
1053 pr_debug("%s: for %llu schedule op %ld on "
1054 "replacement disc %d\n",
1055 __func__, (unsigned long long)sh->sector,
1057 atomic_inc(&sh->count);
1059 atomic_inc(&head_sh->count);
1060 if (use_new_offset(conf, sh))
1061 rbi->bi_iter.bi_sector = (sh->sector
1062 + rrdev->new_data_offset);
1064 rbi->bi_iter.bi_sector = (sh->sector
1065 + rrdev->data_offset);
1066 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1067 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1068 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1070 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1071 rbi->bi_io_vec[0].bv_offset = 0;
1072 rbi->bi_iter.bi_size = STRIPE_SIZE;
1074 * If this is discard request, set bi_vcnt 0. We don't
1075 * want to confuse SCSI because SCSI will replace payload
1077 if (rw & REQ_DISCARD)
1079 if (conf->mddev->gendisk)
1080 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1081 rbi, disk_devt(conf->mddev->gendisk),
1083 generic_make_request(rbi);
1085 if (!rdev && !rrdev) {
1087 set_bit(STRIPE_DEGRADED, &sh->state);
1088 pr_debug("skip op %ld on disc %d for sector %llu\n",
1089 bi->bi_rw, i, (unsigned long long)sh->sector);
1090 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1091 set_bit(STRIPE_HANDLE, &sh->state);
1094 if (!head_sh->batch_head)
1096 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1103 static struct dma_async_tx_descriptor *
1104 async_copy_data(int frombio, struct bio *bio, struct page **page,
1105 sector_t sector, struct dma_async_tx_descriptor *tx,
1106 struct stripe_head *sh)
1109 struct bvec_iter iter;
1110 struct page *bio_page;
1112 struct async_submit_ctl submit;
1113 enum async_tx_flags flags = 0;
1115 if (bio->bi_iter.bi_sector >= sector)
1116 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1118 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1121 flags |= ASYNC_TX_FENCE;
1122 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1124 bio_for_each_segment(bvl, bio, iter) {
1125 int len = bvl.bv_len;
1129 if (page_offset < 0) {
1130 b_offset = -page_offset;
1131 page_offset += b_offset;
1135 if (len > 0 && page_offset + len > STRIPE_SIZE)
1136 clen = STRIPE_SIZE - page_offset;
1141 b_offset += bvl.bv_offset;
1142 bio_page = bvl.bv_page;
1144 if (sh->raid_conf->skip_copy &&
1145 b_offset == 0 && page_offset == 0 &&
1146 clen == STRIPE_SIZE)
1149 tx = async_memcpy(*page, bio_page, page_offset,
1150 b_offset, clen, &submit);
1152 tx = async_memcpy(bio_page, *page, b_offset,
1153 page_offset, clen, &submit);
1155 /* chain the operations */
1156 submit.depend_tx = tx;
1158 if (clen < len) /* hit end of page */
1166 static void ops_complete_biofill(void *stripe_head_ref)
1168 struct stripe_head *sh = stripe_head_ref;
1169 struct bio *return_bi = NULL;
1172 pr_debug("%s: stripe %llu\n", __func__,
1173 (unsigned long long)sh->sector);
1175 /* clear completed biofills */
1176 for (i = sh->disks; i--; ) {
1177 struct r5dev *dev = &sh->dev[i];
1179 /* acknowledge completion of a biofill operation */
1180 /* and check if we need to reply to a read request,
1181 * new R5_Wantfill requests are held off until
1182 * !STRIPE_BIOFILL_RUN
1184 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1185 struct bio *rbi, *rbi2;
1190 while (rbi && rbi->bi_iter.bi_sector <
1191 dev->sector + STRIPE_SECTORS) {
1192 rbi2 = r5_next_bio(rbi, dev->sector);
1193 if (!raid5_dec_bi_active_stripes(rbi)) {
1194 rbi->bi_next = return_bi;
1201 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1203 return_io(return_bi);
1205 set_bit(STRIPE_HANDLE, &sh->state);
1209 static void ops_run_biofill(struct stripe_head *sh)
1211 struct dma_async_tx_descriptor *tx = NULL;
1212 struct async_submit_ctl submit;
1215 BUG_ON(sh->batch_head);
1216 pr_debug("%s: stripe %llu\n", __func__,
1217 (unsigned long long)sh->sector);
1219 for (i = sh->disks; i--; ) {
1220 struct r5dev *dev = &sh->dev[i];
1221 if (test_bit(R5_Wantfill, &dev->flags)) {
1223 spin_lock_irq(&sh->stripe_lock);
1224 dev->read = rbi = dev->toread;
1226 spin_unlock_irq(&sh->stripe_lock);
1227 while (rbi && rbi->bi_iter.bi_sector <
1228 dev->sector + STRIPE_SECTORS) {
1229 tx = async_copy_data(0, rbi, &dev->page,
1230 dev->sector, tx, sh);
1231 rbi = r5_next_bio(rbi, dev->sector);
1236 atomic_inc(&sh->count);
1237 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1238 async_trigger_callback(&submit);
1241 static void mark_target_uptodate(struct stripe_head *sh, int target)
1248 tgt = &sh->dev[target];
1249 set_bit(R5_UPTODATE, &tgt->flags);
1250 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1251 clear_bit(R5_Wantcompute, &tgt->flags);
1254 static void ops_complete_compute(void *stripe_head_ref)
1256 struct stripe_head *sh = stripe_head_ref;
1258 pr_debug("%s: stripe %llu\n", __func__,
1259 (unsigned long long)sh->sector);
1261 /* mark the computed target(s) as uptodate */
1262 mark_target_uptodate(sh, sh->ops.target);
1263 mark_target_uptodate(sh, sh->ops.target2);
1265 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1266 if (sh->check_state == check_state_compute_run)
1267 sh->check_state = check_state_compute_result;
1268 set_bit(STRIPE_HANDLE, &sh->state);
1272 /* return a pointer to the address conversion region of the scribble buffer */
1273 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1274 struct raid5_percpu *percpu, int i)
1278 addr = flex_array_get(percpu->scribble, i);
1279 return addr + sizeof(struct page *) * (sh->disks + 2);
1282 /* return a pointer to the address conversion region of the scribble buffer */
1283 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1287 addr = flex_array_get(percpu->scribble, i);
1291 static struct dma_async_tx_descriptor *
1292 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1294 int disks = sh->disks;
1295 struct page **xor_srcs = to_addr_page(percpu, 0);
1296 int target = sh->ops.target;
1297 struct r5dev *tgt = &sh->dev[target];
1298 struct page *xor_dest = tgt->page;
1300 struct dma_async_tx_descriptor *tx;
1301 struct async_submit_ctl submit;
1304 BUG_ON(sh->batch_head);
1306 pr_debug("%s: stripe %llu block: %d\n",
1307 __func__, (unsigned long long)sh->sector, target);
1308 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1310 for (i = disks; i--; )
1312 xor_srcs[count++] = sh->dev[i].page;
1314 atomic_inc(&sh->count);
1316 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1317 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1318 if (unlikely(count == 1))
1319 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1321 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1326 /* set_syndrome_sources - populate source buffers for gen_syndrome
1327 * @srcs - (struct page *) array of size sh->disks
1328 * @sh - stripe_head to parse
1330 * Populates srcs in proper layout order for the stripe and returns the
1331 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1332 * destination buffer is recorded in srcs[count] and the Q destination
1333 * is recorded in srcs[count+1]].
1335 static int set_syndrome_sources(struct page **srcs,
1336 struct stripe_head *sh,
1339 int disks = sh->disks;
1340 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1341 int d0_idx = raid6_d0(sh);
1345 for (i = 0; i < disks; i++)
1351 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1352 struct r5dev *dev = &sh->dev[i];
1354 if (i == sh->qd_idx || i == sh->pd_idx ||
1355 (srctype == SYNDROME_SRC_ALL) ||
1356 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1357 test_bit(R5_Wantdrain, &dev->flags)) ||
1358 (srctype == SYNDROME_SRC_WRITTEN &&
1360 srcs[slot] = sh->dev[i].page;
1361 i = raid6_next_disk(i, disks);
1362 } while (i != d0_idx);
1364 return syndrome_disks;
1367 static struct dma_async_tx_descriptor *
1368 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1370 int disks = sh->disks;
1371 struct page **blocks = to_addr_page(percpu, 0);
1373 int qd_idx = sh->qd_idx;
1374 struct dma_async_tx_descriptor *tx;
1375 struct async_submit_ctl submit;
1381 BUG_ON(sh->batch_head);
1382 if (sh->ops.target < 0)
1383 target = sh->ops.target2;
1384 else if (sh->ops.target2 < 0)
1385 target = sh->ops.target;
1387 /* we should only have one valid target */
1390 pr_debug("%s: stripe %llu block: %d\n",
1391 __func__, (unsigned long long)sh->sector, target);
1393 tgt = &sh->dev[target];
1394 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1397 atomic_inc(&sh->count);
1399 if (target == qd_idx) {
1400 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1401 blocks[count] = NULL; /* regenerating p is not necessary */
1402 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1403 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1404 ops_complete_compute, sh,
1405 to_addr_conv(sh, percpu, 0));
1406 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1408 /* Compute any data- or p-drive using XOR */
1410 for (i = disks; i-- ; ) {
1411 if (i == target || i == qd_idx)
1413 blocks[count++] = sh->dev[i].page;
1416 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1417 NULL, ops_complete_compute, sh,
1418 to_addr_conv(sh, percpu, 0));
1419 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1425 static struct dma_async_tx_descriptor *
1426 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1428 int i, count, disks = sh->disks;
1429 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1430 int d0_idx = raid6_d0(sh);
1431 int faila = -1, failb = -1;
1432 int target = sh->ops.target;
1433 int target2 = sh->ops.target2;
1434 struct r5dev *tgt = &sh->dev[target];
1435 struct r5dev *tgt2 = &sh->dev[target2];
1436 struct dma_async_tx_descriptor *tx;
1437 struct page **blocks = to_addr_page(percpu, 0);
1438 struct async_submit_ctl submit;
1440 BUG_ON(sh->batch_head);
1441 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1442 __func__, (unsigned long long)sh->sector, target, target2);
1443 BUG_ON(target < 0 || target2 < 0);
1444 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1445 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1447 /* we need to open-code set_syndrome_sources to handle the
1448 * slot number conversion for 'faila' and 'failb'
1450 for (i = 0; i < disks ; i++)
1455 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1457 blocks[slot] = sh->dev[i].page;
1463 i = raid6_next_disk(i, disks);
1464 } while (i != d0_idx);
1466 BUG_ON(faila == failb);
1469 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1470 __func__, (unsigned long long)sh->sector, faila, failb);
1472 atomic_inc(&sh->count);
1474 if (failb == syndrome_disks+1) {
1475 /* Q disk is one of the missing disks */
1476 if (faila == syndrome_disks) {
1477 /* Missing P+Q, just recompute */
1478 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1479 ops_complete_compute, sh,
1480 to_addr_conv(sh, percpu, 0));
1481 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1482 STRIPE_SIZE, &submit);
1486 int qd_idx = sh->qd_idx;
1488 /* Missing D+Q: recompute D from P, then recompute Q */
1489 if (target == qd_idx)
1490 data_target = target2;
1492 data_target = target;
1495 for (i = disks; i-- ; ) {
1496 if (i == data_target || i == qd_idx)
1498 blocks[count++] = sh->dev[i].page;
1500 dest = sh->dev[data_target].page;
1501 init_async_submit(&submit,
1502 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1504 to_addr_conv(sh, percpu, 0));
1505 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1508 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1509 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1510 ops_complete_compute, sh,
1511 to_addr_conv(sh, percpu, 0));
1512 return async_gen_syndrome(blocks, 0, count+2,
1513 STRIPE_SIZE, &submit);
1516 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1517 ops_complete_compute, sh,
1518 to_addr_conv(sh, percpu, 0));
1519 if (failb == syndrome_disks) {
1520 /* We're missing D+P. */
1521 return async_raid6_datap_recov(syndrome_disks+2,
1525 /* We're missing D+D. */
1526 return async_raid6_2data_recov(syndrome_disks+2,
1527 STRIPE_SIZE, faila, failb,
1533 static void ops_complete_prexor(void *stripe_head_ref)
1535 struct stripe_head *sh = stripe_head_ref;
1537 pr_debug("%s: stripe %llu\n", __func__,
1538 (unsigned long long)sh->sector);
1541 static struct dma_async_tx_descriptor *
1542 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1543 struct dma_async_tx_descriptor *tx)
1545 int disks = sh->disks;
1546 struct page **xor_srcs = to_addr_page(percpu, 0);
1547 int count = 0, pd_idx = sh->pd_idx, i;
1548 struct async_submit_ctl submit;
1550 /* existing parity data subtracted */
1551 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1553 BUG_ON(sh->batch_head);
1554 pr_debug("%s: stripe %llu\n", __func__,
1555 (unsigned long long)sh->sector);
1557 for (i = disks; i--; ) {
1558 struct r5dev *dev = &sh->dev[i];
1559 /* Only process blocks that are known to be uptodate */
1560 if (test_bit(R5_Wantdrain, &dev->flags))
1561 xor_srcs[count++] = dev->page;
1564 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1565 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1566 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1571 static struct dma_async_tx_descriptor *
1572 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1573 struct dma_async_tx_descriptor *tx)
1575 struct page **blocks = to_addr_page(percpu, 0);
1577 struct async_submit_ctl submit;
1579 pr_debug("%s: stripe %llu\n", __func__,
1580 (unsigned long long)sh->sector);
1582 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1584 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1585 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1586 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1591 static struct dma_async_tx_descriptor *
1592 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1594 int disks = sh->disks;
1596 struct stripe_head *head_sh = sh;
1598 pr_debug("%s: stripe %llu\n", __func__,
1599 (unsigned long long)sh->sector);
1601 for (i = disks; i--; ) {
1606 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1611 spin_lock_irq(&sh->stripe_lock);
1612 chosen = dev->towrite;
1613 dev->towrite = NULL;
1614 sh->overwrite_disks = 0;
1615 BUG_ON(dev->written);
1616 wbi = dev->written = chosen;
1617 spin_unlock_irq(&sh->stripe_lock);
1618 WARN_ON(dev->page != dev->orig_page);
1620 while (wbi && wbi->bi_iter.bi_sector <
1621 dev->sector + STRIPE_SECTORS) {
1622 if (wbi->bi_rw & REQ_FUA)
1623 set_bit(R5_WantFUA, &dev->flags);
1624 if (wbi->bi_rw & REQ_SYNC)
1625 set_bit(R5_SyncIO, &dev->flags);
1626 if (wbi->bi_rw & REQ_DISCARD)
1627 set_bit(R5_Discard, &dev->flags);
1629 tx = async_copy_data(1, wbi, &dev->page,
1630 dev->sector, tx, sh);
1631 if (dev->page != dev->orig_page) {
1632 set_bit(R5_SkipCopy, &dev->flags);
1633 clear_bit(R5_UPTODATE, &dev->flags);
1634 clear_bit(R5_OVERWRITE, &dev->flags);
1637 wbi = r5_next_bio(wbi, dev->sector);
1640 if (head_sh->batch_head) {
1641 sh = list_first_entry(&sh->batch_list,
1654 static void ops_complete_reconstruct(void *stripe_head_ref)
1656 struct stripe_head *sh = stripe_head_ref;
1657 int disks = sh->disks;
1658 int pd_idx = sh->pd_idx;
1659 int qd_idx = sh->qd_idx;
1661 bool fua = false, sync = false, discard = false;
1663 pr_debug("%s: stripe %llu\n", __func__,
1664 (unsigned long long)sh->sector);
1666 for (i = disks; i--; ) {
1667 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1668 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1669 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1672 for (i = disks; i--; ) {
1673 struct r5dev *dev = &sh->dev[i];
1675 if (dev->written || i == pd_idx || i == qd_idx) {
1676 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1677 set_bit(R5_UPTODATE, &dev->flags);
1679 set_bit(R5_WantFUA, &dev->flags);
1681 set_bit(R5_SyncIO, &dev->flags);
1685 if (sh->reconstruct_state == reconstruct_state_drain_run)
1686 sh->reconstruct_state = reconstruct_state_drain_result;
1687 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1688 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1690 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1691 sh->reconstruct_state = reconstruct_state_result;
1694 set_bit(STRIPE_HANDLE, &sh->state);
1699 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1700 struct dma_async_tx_descriptor *tx)
1702 int disks = sh->disks;
1703 struct page **xor_srcs;
1704 struct async_submit_ctl submit;
1705 int count, pd_idx = sh->pd_idx, i;
1706 struct page *xor_dest;
1708 unsigned long flags;
1710 struct stripe_head *head_sh = sh;
1713 pr_debug("%s: stripe %llu\n", __func__,
1714 (unsigned long long)sh->sector);
1716 for (i = 0; i < sh->disks; i++) {
1719 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1722 if (i >= sh->disks) {
1723 atomic_inc(&sh->count);
1724 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1725 ops_complete_reconstruct(sh);
1730 xor_srcs = to_addr_page(percpu, j);
1731 /* check if prexor is active which means only process blocks
1732 * that are part of a read-modify-write (written)
1734 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1736 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1737 for (i = disks; i--; ) {
1738 struct r5dev *dev = &sh->dev[i];
1739 if (head_sh->dev[i].written)
1740 xor_srcs[count++] = dev->page;
1743 xor_dest = sh->dev[pd_idx].page;
1744 for (i = disks; i--; ) {
1745 struct r5dev *dev = &sh->dev[i];
1747 xor_srcs[count++] = dev->page;
1751 /* 1/ if we prexor'd then the dest is reused as a source
1752 * 2/ if we did not prexor then we are redoing the parity
1753 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1754 * for the synchronous xor case
1756 last_stripe = !head_sh->batch_head ||
1757 list_first_entry(&sh->batch_list,
1758 struct stripe_head, batch_list) == head_sh;
1760 flags = ASYNC_TX_ACK |
1761 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1763 atomic_inc(&head_sh->count);
1764 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1765 to_addr_conv(sh, percpu, j));
1767 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1768 init_async_submit(&submit, flags, tx, NULL, NULL,
1769 to_addr_conv(sh, percpu, j));
1772 if (unlikely(count == 1))
1773 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1775 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1778 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1785 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1786 struct dma_async_tx_descriptor *tx)
1788 struct async_submit_ctl submit;
1789 struct page **blocks;
1790 int count, i, j = 0;
1791 struct stripe_head *head_sh = sh;
1794 unsigned long txflags;
1796 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1798 for (i = 0; i < sh->disks; i++) {
1799 if (sh->pd_idx == i || sh->qd_idx == i)
1801 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1804 if (i >= sh->disks) {
1805 atomic_inc(&sh->count);
1806 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1807 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1808 ops_complete_reconstruct(sh);
1813 blocks = to_addr_page(percpu, j);
1815 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1816 synflags = SYNDROME_SRC_WRITTEN;
1817 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1819 synflags = SYNDROME_SRC_ALL;
1820 txflags = ASYNC_TX_ACK;
1823 count = set_syndrome_sources(blocks, sh, synflags);
1824 last_stripe = !head_sh->batch_head ||
1825 list_first_entry(&sh->batch_list,
1826 struct stripe_head, batch_list) == head_sh;
1829 atomic_inc(&head_sh->count);
1830 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1831 head_sh, to_addr_conv(sh, percpu, j));
1833 init_async_submit(&submit, 0, tx, NULL, NULL,
1834 to_addr_conv(sh, percpu, j));
1835 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1838 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1844 static void ops_complete_check(void *stripe_head_ref)
1846 struct stripe_head *sh = stripe_head_ref;
1848 pr_debug("%s: stripe %llu\n", __func__,
1849 (unsigned long long)sh->sector);
1851 sh->check_state = check_state_check_result;
1852 set_bit(STRIPE_HANDLE, &sh->state);
1856 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1858 int disks = sh->disks;
1859 int pd_idx = sh->pd_idx;
1860 int qd_idx = sh->qd_idx;
1861 struct page *xor_dest;
1862 struct page **xor_srcs = to_addr_page(percpu, 0);
1863 struct dma_async_tx_descriptor *tx;
1864 struct async_submit_ctl submit;
1868 pr_debug("%s: stripe %llu\n", __func__,
1869 (unsigned long long)sh->sector);
1871 BUG_ON(sh->batch_head);
1873 xor_dest = sh->dev[pd_idx].page;
1874 xor_srcs[count++] = xor_dest;
1875 for (i = disks; i--; ) {
1876 if (i == pd_idx || i == qd_idx)
1878 xor_srcs[count++] = sh->dev[i].page;
1881 init_async_submit(&submit, 0, NULL, NULL, NULL,
1882 to_addr_conv(sh, percpu, 0));
1883 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1884 &sh->ops.zero_sum_result, &submit);
1886 atomic_inc(&sh->count);
1887 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1888 tx = async_trigger_callback(&submit);
1891 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1893 struct page **srcs = to_addr_page(percpu, 0);
1894 struct async_submit_ctl submit;
1897 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1898 (unsigned long long)sh->sector, checkp);
1900 BUG_ON(sh->batch_head);
1901 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1905 atomic_inc(&sh->count);
1906 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1907 sh, to_addr_conv(sh, percpu, 0));
1908 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1909 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1912 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1914 int overlap_clear = 0, i, disks = sh->disks;
1915 struct dma_async_tx_descriptor *tx = NULL;
1916 struct r5conf *conf = sh->raid_conf;
1917 int level = conf->level;
1918 struct raid5_percpu *percpu;
1921 cpu = get_cpu_light();
1922 percpu = per_cpu_ptr(conf->percpu, cpu);
1923 spin_lock(&percpu->lock);
1924 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1925 ops_run_biofill(sh);
1929 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1931 tx = ops_run_compute5(sh, percpu);
1933 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1934 tx = ops_run_compute6_1(sh, percpu);
1936 tx = ops_run_compute6_2(sh, percpu);
1938 /* terminate the chain if reconstruct is not set to be run */
1939 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1943 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1945 tx = ops_run_prexor5(sh, percpu, tx);
1947 tx = ops_run_prexor6(sh, percpu, tx);
1950 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1951 tx = ops_run_biodrain(sh, tx);
1955 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1957 ops_run_reconstruct5(sh, percpu, tx);
1959 ops_run_reconstruct6(sh, percpu, tx);
1962 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1963 if (sh->check_state == check_state_run)
1964 ops_run_check_p(sh, percpu);
1965 else if (sh->check_state == check_state_run_q)
1966 ops_run_check_pq(sh, percpu, 0);
1967 else if (sh->check_state == check_state_run_pq)
1968 ops_run_check_pq(sh, percpu, 1);
1973 if (overlap_clear && !sh->batch_head)
1974 for (i = disks; i--; ) {
1975 struct r5dev *dev = &sh->dev[i];
1976 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1977 wake_up(&sh->raid_conf->wait_for_overlap);
1979 spin_unlock(&percpu->lock);
1983 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1985 struct stripe_head *sh;
1987 sh = kmem_cache_zalloc(sc, gfp);
1989 spin_lock_init(&sh->stripe_lock);
1990 spin_lock_init(&sh->batch_lock);
1991 INIT_LIST_HEAD(&sh->batch_list);
1992 INIT_LIST_HEAD(&sh->lru);
1993 atomic_set(&sh->count, 1);
1997 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
1999 struct stripe_head *sh;
2001 sh = alloc_stripe(conf->slab_cache, gfp);
2005 sh->raid_conf = conf;
2007 if (grow_buffers(sh, gfp)) {
2009 kmem_cache_free(conf->slab_cache, sh);
2012 sh->hash_lock_index =
2013 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2014 /* we just created an active stripe so... */
2015 atomic_inc(&conf->active_stripes);
2018 conf->max_nr_stripes++;
2022 static int grow_stripes(struct r5conf *conf, int num)
2024 struct kmem_cache *sc;
2025 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2027 if (conf->mddev->gendisk)
2028 sprintf(conf->cache_name[0],
2029 "raid%d-%s", conf->level, mdname(conf->mddev));
2031 sprintf(conf->cache_name[0],
2032 "raid%d-%p", conf->level, conf->mddev);
2033 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2035 conf->active_name = 0;
2036 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2037 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2041 conf->slab_cache = sc;
2042 conf->pool_size = devs;
2044 if (!grow_one_stripe(conf, GFP_KERNEL))
2051 * scribble_len - return the required size of the scribble region
2052 * @num - total number of disks in the array
2054 * The size must be enough to contain:
2055 * 1/ a struct page pointer for each device in the array +2
2056 * 2/ room to convert each entry in (1) to its corresponding dma
2057 * (dma_map_page()) or page (page_address()) address.
2059 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2060 * calculate over all devices (not just the data blocks), using zeros in place
2061 * of the P and Q blocks.
2063 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2065 struct flex_array *ret;
2068 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2069 ret = flex_array_alloc(len, cnt, flags);
2072 /* always prealloc all elements, so no locking is required */
2073 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2074 flex_array_free(ret);
2080 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2085 mddev_suspend(conf->mddev);
2087 for_each_present_cpu(cpu) {
2088 struct raid5_percpu *percpu;
2089 struct flex_array *scribble;
2091 percpu = per_cpu_ptr(conf->percpu, cpu);
2092 scribble = scribble_alloc(new_disks,
2093 new_sectors / STRIPE_SECTORS,
2097 flex_array_free(percpu->scribble);
2098 percpu->scribble = scribble;
2105 mddev_resume(conf->mddev);
2109 static int resize_stripes(struct r5conf *conf, int newsize)
2111 /* Make all the stripes able to hold 'newsize' devices.
2112 * New slots in each stripe get 'page' set to a new page.
2114 * This happens in stages:
2115 * 1/ create a new kmem_cache and allocate the required number of
2117 * 2/ gather all the old stripe_heads and transfer the pages across
2118 * to the new stripe_heads. This will have the side effect of
2119 * freezing the array as once all stripe_heads have been collected,
2120 * no IO will be possible. Old stripe heads are freed once their
2121 * pages have been transferred over, and the old kmem_cache is
2122 * freed when all stripes are done.
2123 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2124 * we simple return a failre status - no need to clean anything up.
2125 * 4/ allocate new pages for the new slots in the new stripe_heads.
2126 * If this fails, we don't bother trying the shrink the
2127 * stripe_heads down again, we just leave them as they are.
2128 * As each stripe_head is processed the new one is released into
2131 * Once step2 is started, we cannot afford to wait for a write,
2132 * so we use GFP_NOIO allocations.
2134 struct stripe_head *osh, *nsh;
2135 LIST_HEAD(newstripes);
2136 struct disk_info *ndisks;
2138 struct kmem_cache *sc;
2142 if (newsize <= conf->pool_size)
2143 return 0; /* never bother to shrink */
2145 err = md_allow_write(conf->mddev);
2150 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2151 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2156 for (i = conf->max_nr_stripes; i; i--) {
2157 nsh = alloc_stripe(sc, GFP_KERNEL);
2161 nsh->raid_conf = conf;
2162 list_add(&nsh->lru, &newstripes);
2165 /* didn't get enough, give up */
2166 while (!list_empty(&newstripes)) {
2167 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2168 list_del(&nsh->lru);
2169 kmem_cache_free(sc, nsh);
2171 kmem_cache_destroy(sc);
2174 /* Step 2 - Must use GFP_NOIO now.
2175 * OK, we have enough stripes, start collecting inactive
2176 * stripes and copying them over
2180 list_for_each_entry(nsh, &newstripes, lru) {
2181 lock_device_hash_lock(conf, hash);
2182 wait_event_cmd(conf->wait_for_stripe,
2183 !list_empty(conf->inactive_list + hash),
2184 unlock_device_hash_lock(conf, hash),
2185 lock_device_hash_lock(conf, hash));
2186 osh = get_free_stripe(conf, hash);
2187 unlock_device_hash_lock(conf, hash);
2189 for(i=0; i<conf->pool_size; i++) {
2190 nsh->dev[i].page = osh->dev[i].page;
2191 nsh->dev[i].orig_page = osh->dev[i].page;
2193 nsh->hash_lock_index = hash;
2194 kmem_cache_free(conf->slab_cache, osh);
2196 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2197 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2202 kmem_cache_destroy(conf->slab_cache);
2205 * At this point, we are holding all the stripes so the array
2206 * is completely stalled, so now is a good time to resize
2207 * conf->disks and the scribble region
2209 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2211 for (i=0; i<conf->raid_disks; i++)
2212 ndisks[i] = conf->disks[i];
2214 conf->disks = ndisks;
2218 /* Step 4, return new stripes to service */
2219 while(!list_empty(&newstripes)) {
2220 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2221 list_del_init(&nsh->lru);
2223 for (i=conf->raid_disks; i < newsize; i++)
2224 if (nsh->dev[i].page == NULL) {
2225 struct page *p = alloc_page(GFP_NOIO);
2226 nsh->dev[i].page = p;
2227 nsh->dev[i].orig_page = p;
2231 release_stripe(nsh);
2233 /* critical section pass, GFP_NOIO no longer needed */
2235 conf->slab_cache = sc;
2236 conf->active_name = 1-conf->active_name;
2238 conf->pool_size = newsize;
2242 static int drop_one_stripe(struct r5conf *conf)
2244 struct stripe_head *sh;
2245 int hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
2247 spin_lock_irq(conf->hash_locks + hash);
2248 sh = get_free_stripe(conf, hash);
2249 spin_unlock_irq(conf->hash_locks + hash);
2252 BUG_ON(atomic_read(&sh->count));
2254 kmem_cache_free(conf->slab_cache, sh);
2255 atomic_dec(&conf->active_stripes);
2256 conf->max_nr_stripes--;
2260 static void shrink_stripes(struct r5conf *conf)
2262 while (conf->max_nr_stripes &&
2263 drop_one_stripe(conf))
2266 if (conf->slab_cache)
2267 kmem_cache_destroy(conf->slab_cache);
2268 conf->slab_cache = NULL;
2271 static void raid5_end_read_request(struct bio * bi, int error)
2273 struct stripe_head *sh = bi->bi_private;
2274 struct r5conf *conf = sh->raid_conf;
2275 int disks = sh->disks, i;
2276 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2277 char b[BDEVNAME_SIZE];
2278 struct md_rdev *rdev = NULL;
2281 for (i=0 ; i<disks; i++)
2282 if (bi == &sh->dev[i].req)
2285 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2286 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2292 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2293 /* If replacement finished while this request was outstanding,
2294 * 'replacement' might be NULL already.
2295 * In that case it moved down to 'rdev'.
2296 * rdev is not removed until all requests are finished.
2298 rdev = conf->disks[i].replacement;
2300 rdev = conf->disks[i].rdev;
2302 if (use_new_offset(conf, sh))
2303 s = sh->sector + rdev->new_data_offset;
2305 s = sh->sector + rdev->data_offset;
2307 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2308 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2309 /* Note that this cannot happen on a
2310 * replacement device. We just fail those on
2315 "md/raid:%s: read error corrected"
2316 " (%lu sectors at %llu on %s)\n",
2317 mdname(conf->mddev), STRIPE_SECTORS,
2318 (unsigned long long)s,
2319 bdevname(rdev->bdev, b));
2320 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2321 clear_bit(R5_ReadError, &sh->dev[i].flags);
2322 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2323 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2324 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2326 if (atomic_read(&rdev->read_errors))
2327 atomic_set(&rdev->read_errors, 0);
2329 const char *bdn = bdevname(rdev->bdev, b);
2333 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2334 atomic_inc(&rdev->read_errors);
2335 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2338 "md/raid:%s: read error on replacement device "
2339 "(sector %llu on %s).\n",
2340 mdname(conf->mddev),
2341 (unsigned long long)s,
2343 else if (conf->mddev->degraded >= conf->max_degraded) {
2347 "md/raid:%s: read error not correctable "
2348 "(sector %llu on %s).\n",
2349 mdname(conf->mddev),
2350 (unsigned long long)s,
2352 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2357 "md/raid:%s: read error NOT corrected!! "
2358 "(sector %llu on %s).\n",
2359 mdname(conf->mddev),
2360 (unsigned long long)s,
2362 } else if (atomic_read(&rdev->read_errors)
2363 > conf->max_nr_stripes)
2365 "md/raid:%s: Too many read errors, failing device %s.\n",
2366 mdname(conf->mddev), bdn);
2369 if (set_bad && test_bit(In_sync, &rdev->flags)
2370 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2373 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2374 set_bit(R5_ReadError, &sh->dev[i].flags);
2375 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2377 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2379 clear_bit(R5_ReadError, &sh->dev[i].flags);
2380 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2382 && test_bit(In_sync, &rdev->flags)
2383 && rdev_set_badblocks(
2384 rdev, sh->sector, STRIPE_SECTORS, 0)))
2385 md_error(conf->mddev, rdev);
2388 rdev_dec_pending(rdev, conf->mddev);
2389 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2390 set_bit(STRIPE_HANDLE, &sh->state);
2394 static void raid5_end_write_request(struct bio *bi, int error)
2396 struct stripe_head *sh = bi->bi_private;
2397 struct r5conf *conf = sh->raid_conf;
2398 int disks = sh->disks, i;
2399 struct md_rdev *uninitialized_var(rdev);
2400 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2403 int replacement = 0;
2405 for (i = 0 ; i < disks; i++) {
2406 if (bi == &sh->dev[i].req) {
2407 rdev = conf->disks[i].rdev;
2410 if (bi == &sh->dev[i].rreq) {
2411 rdev = conf->disks[i].replacement;
2415 /* rdev was removed and 'replacement'
2416 * replaced it. rdev is not removed
2417 * until all requests are finished.
2419 rdev = conf->disks[i].rdev;
2423 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2424 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2433 md_error(conf->mddev, rdev);
2434 else if (is_badblock(rdev, sh->sector,
2436 &first_bad, &bad_sectors))
2437 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2440 set_bit(STRIPE_DEGRADED, &sh->state);
2441 set_bit(WriteErrorSeen, &rdev->flags);
2442 set_bit(R5_WriteError, &sh->dev[i].flags);
2443 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2444 set_bit(MD_RECOVERY_NEEDED,
2445 &rdev->mddev->recovery);
2446 } else if (is_badblock(rdev, sh->sector,
2448 &first_bad, &bad_sectors)) {
2449 set_bit(R5_MadeGood, &sh->dev[i].flags);
2450 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2451 /* That was a successful write so make
2452 * sure it looks like we already did
2455 set_bit(R5_ReWrite, &sh->dev[i].flags);
2458 rdev_dec_pending(rdev, conf->mddev);
2460 if (sh->batch_head && !uptodate && !replacement)
2461 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2463 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2464 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2465 set_bit(STRIPE_HANDLE, &sh->state);
2468 if (sh->batch_head && sh != sh->batch_head)
2469 release_stripe(sh->batch_head);
2472 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2474 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2476 struct r5dev *dev = &sh->dev[i];
2478 bio_init(&dev->req);
2479 dev->req.bi_io_vec = &dev->vec;
2480 dev->req.bi_max_vecs = 1;
2481 dev->req.bi_private = sh;
2483 bio_init(&dev->rreq);
2484 dev->rreq.bi_io_vec = &dev->rvec;
2485 dev->rreq.bi_max_vecs = 1;
2486 dev->rreq.bi_private = sh;
2489 dev->sector = compute_blocknr(sh, i, previous);
2492 static void error(struct mddev *mddev, struct md_rdev *rdev)
2494 char b[BDEVNAME_SIZE];
2495 struct r5conf *conf = mddev->private;
2496 unsigned long flags;
2497 pr_debug("raid456: error called\n");
2499 spin_lock_irqsave(&conf->device_lock, flags);
2500 clear_bit(In_sync, &rdev->flags);
2501 mddev->degraded = calc_degraded(conf);
2502 spin_unlock_irqrestore(&conf->device_lock, flags);
2503 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2505 set_bit(Blocked, &rdev->flags);
2506 set_bit(Faulty, &rdev->flags);
2507 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2509 "md/raid:%s: Disk failure on %s, disabling device.\n"
2510 "md/raid:%s: Operation continuing on %d devices.\n",
2512 bdevname(rdev->bdev, b),
2514 conf->raid_disks - mddev->degraded);
2518 * Input: a 'big' sector number,
2519 * Output: index of the data and parity disk, and the sector # in them.
2521 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2522 int previous, int *dd_idx,
2523 struct stripe_head *sh)
2525 sector_t stripe, stripe2;
2526 sector_t chunk_number;
2527 unsigned int chunk_offset;
2530 sector_t new_sector;
2531 int algorithm = previous ? conf->prev_algo
2533 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2534 : conf->chunk_sectors;
2535 int raid_disks = previous ? conf->previous_raid_disks
2537 int data_disks = raid_disks - conf->max_degraded;
2539 /* First compute the information on this sector */
2542 * Compute the chunk number and the sector offset inside the chunk
2544 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2545 chunk_number = r_sector;
2548 * Compute the stripe number
2550 stripe = chunk_number;
2551 *dd_idx = sector_div(stripe, data_disks);
2554 * Select the parity disk based on the user selected algorithm.
2556 pd_idx = qd_idx = -1;
2557 switch(conf->level) {
2559 pd_idx = data_disks;
2562 switch (algorithm) {
2563 case ALGORITHM_LEFT_ASYMMETRIC:
2564 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2565 if (*dd_idx >= pd_idx)
2568 case ALGORITHM_RIGHT_ASYMMETRIC:
2569 pd_idx = sector_div(stripe2, raid_disks);
2570 if (*dd_idx >= pd_idx)
2573 case ALGORITHM_LEFT_SYMMETRIC:
2574 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2575 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2577 case ALGORITHM_RIGHT_SYMMETRIC:
2578 pd_idx = sector_div(stripe2, raid_disks);
2579 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2581 case ALGORITHM_PARITY_0:
2585 case ALGORITHM_PARITY_N:
2586 pd_idx = data_disks;
2594 switch (algorithm) {
2595 case ALGORITHM_LEFT_ASYMMETRIC:
2596 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2597 qd_idx = pd_idx + 1;
2598 if (pd_idx == raid_disks-1) {
2599 (*dd_idx)++; /* Q D D D P */
2601 } else if (*dd_idx >= pd_idx)
2602 (*dd_idx) += 2; /* D D P Q D */
2604 case ALGORITHM_RIGHT_ASYMMETRIC:
2605 pd_idx = sector_div(stripe2, raid_disks);
2606 qd_idx = pd_idx + 1;
2607 if (pd_idx == raid_disks-1) {
2608 (*dd_idx)++; /* Q D D D P */
2610 } else if (*dd_idx >= pd_idx)
2611 (*dd_idx) += 2; /* D D P Q D */
2613 case ALGORITHM_LEFT_SYMMETRIC:
2614 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2615 qd_idx = (pd_idx + 1) % raid_disks;
2616 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2618 case ALGORITHM_RIGHT_SYMMETRIC:
2619 pd_idx = sector_div(stripe2, raid_disks);
2620 qd_idx = (pd_idx + 1) % raid_disks;
2621 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2624 case ALGORITHM_PARITY_0:
2629 case ALGORITHM_PARITY_N:
2630 pd_idx = data_disks;
2631 qd_idx = data_disks + 1;
2634 case ALGORITHM_ROTATING_ZERO_RESTART:
2635 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2636 * of blocks for computing Q is different.
2638 pd_idx = sector_div(stripe2, raid_disks);
2639 qd_idx = pd_idx + 1;
2640 if (pd_idx == raid_disks-1) {
2641 (*dd_idx)++; /* Q D D D P */
2643 } else if (*dd_idx >= pd_idx)
2644 (*dd_idx) += 2; /* D D P Q D */
2648 case ALGORITHM_ROTATING_N_RESTART:
2649 /* Same a left_asymmetric, by first stripe is
2650 * D D D P Q rather than
2654 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2655 qd_idx = pd_idx + 1;
2656 if (pd_idx == raid_disks-1) {
2657 (*dd_idx)++; /* Q D D D P */
2659 } else if (*dd_idx >= pd_idx)
2660 (*dd_idx) += 2; /* D D P Q D */
2664 case ALGORITHM_ROTATING_N_CONTINUE:
2665 /* Same as left_symmetric but Q is before P */
2666 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2667 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2668 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2672 case ALGORITHM_LEFT_ASYMMETRIC_6:
2673 /* RAID5 left_asymmetric, with Q on last device */
2674 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2675 if (*dd_idx >= pd_idx)
2677 qd_idx = raid_disks - 1;
2680 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2681 pd_idx = sector_div(stripe2, raid_disks-1);
2682 if (*dd_idx >= pd_idx)
2684 qd_idx = raid_disks - 1;
2687 case ALGORITHM_LEFT_SYMMETRIC_6:
2688 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2689 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2690 qd_idx = raid_disks - 1;
2693 case ALGORITHM_RIGHT_SYMMETRIC_6:
2694 pd_idx = sector_div(stripe2, raid_disks-1);
2695 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2696 qd_idx = raid_disks - 1;
2699 case ALGORITHM_PARITY_0_6:
2702 qd_idx = raid_disks - 1;
2712 sh->pd_idx = pd_idx;
2713 sh->qd_idx = qd_idx;
2714 sh->ddf_layout = ddf_layout;
2717 * Finally, compute the new sector number
2719 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2723 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2725 struct r5conf *conf = sh->raid_conf;
2726 int raid_disks = sh->disks;
2727 int data_disks = raid_disks - conf->max_degraded;
2728 sector_t new_sector = sh->sector, check;
2729 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2730 : conf->chunk_sectors;
2731 int algorithm = previous ? conf->prev_algo
2735 sector_t chunk_number;
2736 int dummy1, dd_idx = i;
2738 struct stripe_head sh2;
2740 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2741 stripe = new_sector;
2743 if (i == sh->pd_idx)
2745 switch(conf->level) {
2748 switch (algorithm) {
2749 case ALGORITHM_LEFT_ASYMMETRIC:
2750 case ALGORITHM_RIGHT_ASYMMETRIC:
2754 case ALGORITHM_LEFT_SYMMETRIC:
2755 case ALGORITHM_RIGHT_SYMMETRIC:
2758 i -= (sh->pd_idx + 1);
2760 case ALGORITHM_PARITY_0:
2763 case ALGORITHM_PARITY_N:
2770 if (i == sh->qd_idx)
2771 return 0; /* It is the Q disk */
2772 switch (algorithm) {
2773 case ALGORITHM_LEFT_ASYMMETRIC:
2774 case ALGORITHM_RIGHT_ASYMMETRIC:
2775 case ALGORITHM_ROTATING_ZERO_RESTART:
2776 case ALGORITHM_ROTATING_N_RESTART:
2777 if (sh->pd_idx == raid_disks-1)
2778 i--; /* Q D D D P */
2779 else if (i > sh->pd_idx)
2780 i -= 2; /* D D P Q D */
2782 case ALGORITHM_LEFT_SYMMETRIC:
2783 case ALGORITHM_RIGHT_SYMMETRIC:
2784 if (sh->pd_idx == raid_disks-1)
2785 i--; /* Q D D D P */
2790 i -= (sh->pd_idx + 2);
2793 case ALGORITHM_PARITY_0:
2796 case ALGORITHM_PARITY_N:
2798 case ALGORITHM_ROTATING_N_CONTINUE:
2799 /* Like left_symmetric, but P is before Q */
2800 if (sh->pd_idx == 0)
2801 i--; /* P D D D Q */
2806 i -= (sh->pd_idx + 1);
2809 case ALGORITHM_LEFT_ASYMMETRIC_6:
2810 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2814 case ALGORITHM_LEFT_SYMMETRIC_6:
2815 case ALGORITHM_RIGHT_SYMMETRIC_6:
2817 i += data_disks + 1;
2818 i -= (sh->pd_idx + 1);
2820 case ALGORITHM_PARITY_0_6:
2829 chunk_number = stripe * data_disks + i;
2830 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2832 check = raid5_compute_sector(conf, r_sector,
2833 previous, &dummy1, &sh2);
2834 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2835 || sh2.qd_idx != sh->qd_idx) {
2836 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2837 mdname(conf->mddev));
2844 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2845 int rcw, int expand)
2847 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2848 struct r5conf *conf = sh->raid_conf;
2849 int level = conf->level;
2853 for (i = disks; i--; ) {
2854 struct r5dev *dev = &sh->dev[i];
2857 set_bit(R5_LOCKED, &dev->flags);
2858 set_bit(R5_Wantdrain, &dev->flags);
2860 clear_bit(R5_UPTODATE, &dev->flags);
2864 /* if we are not expanding this is a proper write request, and
2865 * there will be bios with new data to be drained into the
2870 /* False alarm, nothing to do */
2872 sh->reconstruct_state = reconstruct_state_drain_run;
2873 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2875 sh->reconstruct_state = reconstruct_state_run;
2877 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2879 if (s->locked + conf->max_degraded == disks)
2880 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2881 atomic_inc(&conf->pending_full_writes);
2883 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2884 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2885 BUG_ON(level == 6 &&
2886 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2887 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2889 for (i = disks; i--; ) {
2890 struct r5dev *dev = &sh->dev[i];
2891 if (i == pd_idx || i == qd_idx)
2895 (test_bit(R5_UPTODATE, &dev->flags) ||
2896 test_bit(R5_Wantcompute, &dev->flags))) {
2897 set_bit(R5_Wantdrain, &dev->flags);
2898 set_bit(R5_LOCKED, &dev->flags);
2899 clear_bit(R5_UPTODATE, &dev->flags);
2904 /* False alarm - nothing to do */
2906 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2907 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2908 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2909 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2912 /* keep the parity disk(s) locked while asynchronous operations
2915 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2916 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2920 int qd_idx = sh->qd_idx;
2921 struct r5dev *dev = &sh->dev[qd_idx];
2923 set_bit(R5_LOCKED, &dev->flags);
2924 clear_bit(R5_UPTODATE, &dev->flags);
2928 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2929 __func__, (unsigned long long)sh->sector,
2930 s->locked, s->ops_request);
2934 * Each stripe/dev can have one or more bion attached.
2935 * toread/towrite point to the first in a chain.
2936 * The bi_next chain must be in order.
2938 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2939 int forwrite, int previous)
2942 struct r5conf *conf = sh->raid_conf;
2945 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2946 (unsigned long long)bi->bi_iter.bi_sector,
2947 (unsigned long long)sh->sector);
2950 * If several bio share a stripe. The bio bi_phys_segments acts as a
2951 * reference count to avoid race. The reference count should already be
2952 * increased before this function is called (for example, in
2953 * make_request()), so other bio sharing this stripe will not free the
2954 * stripe. If a stripe is owned by one stripe, the stripe lock will
2957 spin_lock_irq(&sh->stripe_lock);
2958 /* Don't allow new IO added to stripes in batch list */
2962 bip = &sh->dev[dd_idx].towrite;
2966 bip = &sh->dev[dd_idx].toread;
2967 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2968 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2970 bip = & (*bip)->bi_next;
2972 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2975 if (!forwrite || previous)
2976 clear_bit(STRIPE_BATCH_READY, &sh->state);
2978 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2982 raid5_inc_bi_active_stripes(bi);
2985 /* check if page is covered */
2986 sector_t sector = sh->dev[dd_idx].sector;
2987 for (bi=sh->dev[dd_idx].towrite;
2988 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2989 bi && bi->bi_iter.bi_sector <= sector;
2990 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2991 if (bio_end_sector(bi) >= sector)
2992 sector = bio_end_sector(bi);
2994 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2995 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
2996 sh->overwrite_disks++;
2999 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3000 (unsigned long long)(*bip)->bi_iter.bi_sector,
3001 (unsigned long long)sh->sector, dd_idx);
3003 if (conf->mddev->bitmap && firstwrite) {
3004 /* Cannot hold spinlock over bitmap_startwrite,
3005 * but must ensure this isn't added to a batch until
3006 * we have added to the bitmap and set bm_seq.
3007 * So set STRIPE_BITMAP_PENDING to prevent
3009 * If multiple add_stripe_bio() calls race here they
3010 * much all set STRIPE_BITMAP_PENDING. So only the first one
3011 * to complete "bitmap_startwrite" gets to set
3012 * STRIPE_BIT_DELAY. This is important as once a stripe
3013 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3016 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3017 spin_unlock_irq(&sh->stripe_lock);
3018 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3020 spin_lock_irq(&sh->stripe_lock);
3021 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3022 if (!sh->batch_head) {
3023 sh->bm_seq = conf->seq_flush+1;
3024 set_bit(STRIPE_BIT_DELAY, &sh->state);
3027 spin_unlock_irq(&sh->stripe_lock);
3029 if (stripe_can_batch(sh))
3030 stripe_add_to_batch_list(conf, sh);
3034 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3035 spin_unlock_irq(&sh->stripe_lock);
3039 static void end_reshape(struct r5conf *conf);
3041 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3042 struct stripe_head *sh)
3044 int sectors_per_chunk =
3045 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3047 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3048 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3050 raid5_compute_sector(conf,
3051 stripe * (disks - conf->max_degraded)
3052 *sectors_per_chunk + chunk_offset,
3058 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3059 struct stripe_head_state *s, int disks,
3060 struct bio **return_bi)
3063 BUG_ON(sh->batch_head);
3064 for (i = disks; i--; ) {
3068 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3069 struct md_rdev *rdev;
3071 rdev = rcu_dereference(conf->disks[i].rdev);
3072 if (rdev && test_bit(In_sync, &rdev->flags))
3073 atomic_inc(&rdev->nr_pending);
3078 if (!rdev_set_badblocks(
3082 md_error(conf->mddev, rdev);
3083 rdev_dec_pending(rdev, conf->mddev);
3086 spin_lock_irq(&sh->stripe_lock);
3087 /* fail all writes first */
3088 bi = sh->dev[i].towrite;
3089 sh->dev[i].towrite = NULL;
3090 sh->overwrite_disks = 0;
3091 spin_unlock_irq(&sh->stripe_lock);
3095 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3096 wake_up(&conf->wait_for_overlap);
3098 while (bi && bi->bi_iter.bi_sector <
3099 sh->dev[i].sector + STRIPE_SECTORS) {
3100 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3101 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3102 if (!raid5_dec_bi_active_stripes(bi)) {
3103 md_write_end(conf->mddev);
3104 bi->bi_next = *return_bi;
3110 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3111 STRIPE_SECTORS, 0, 0);
3113 /* and fail all 'written' */
3114 bi = sh->dev[i].written;
3115 sh->dev[i].written = NULL;
3116 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3117 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3118 sh->dev[i].page = sh->dev[i].orig_page;
3121 if (bi) bitmap_end = 1;
3122 while (bi && bi->bi_iter.bi_sector <
3123 sh->dev[i].sector + STRIPE_SECTORS) {
3124 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3125 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3126 if (!raid5_dec_bi_active_stripes(bi)) {
3127 md_write_end(conf->mddev);
3128 bi->bi_next = *return_bi;
3134 /* fail any reads if this device is non-operational and
3135 * the data has not reached the cache yet.
3137 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3138 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3139 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3140 spin_lock_irq(&sh->stripe_lock);
3141 bi = sh->dev[i].toread;
3142 sh->dev[i].toread = NULL;
3143 spin_unlock_irq(&sh->stripe_lock);
3144 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3145 wake_up(&conf->wait_for_overlap);
3146 while (bi && bi->bi_iter.bi_sector <
3147 sh->dev[i].sector + STRIPE_SECTORS) {
3148 struct bio *nextbi =
3149 r5_next_bio(bi, sh->dev[i].sector);
3150 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3151 if (!raid5_dec_bi_active_stripes(bi)) {
3152 bi->bi_next = *return_bi;
3159 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3160 STRIPE_SECTORS, 0, 0);
3161 /* If we were in the middle of a write the parity block might
3162 * still be locked - so just clear all R5_LOCKED flags
3164 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3167 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3168 if (atomic_dec_and_test(&conf->pending_full_writes))
3169 md_wakeup_thread(conf->mddev->thread);
3173 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3174 struct stripe_head_state *s)
3179 BUG_ON(sh->batch_head);
3180 clear_bit(STRIPE_SYNCING, &sh->state);
3181 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3182 wake_up(&conf->wait_for_overlap);
3185 /* There is nothing more to do for sync/check/repair.
3186 * Don't even need to abort as that is handled elsewhere
3187 * if needed, and not always wanted e.g. if there is a known
3189 * For recover/replace we need to record a bad block on all
3190 * non-sync devices, or abort the recovery
3192 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3193 /* During recovery devices cannot be removed, so
3194 * locking and refcounting of rdevs is not needed
3196 for (i = 0; i < conf->raid_disks; i++) {
3197 struct md_rdev *rdev = conf->disks[i].rdev;
3199 && !test_bit(Faulty, &rdev->flags)
3200 && !test_bit(In_sync, &rdev->flags)
3201 && !rdev_set_badblocks(rdev, sh->sector,
3204 rdev = conf->disks[i].replacement;
3206 && !test_bit(Faulty, &rdev->flags)
3207 && !test_bit(In_sync, &rdev->flags)
3208 && !rdev_set_badblocks(rdev, sh->sector,
3213 conf->recovery_disabled =
3214 conf->mddev->recovery_disabled;
3216 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3219 static int want_replace(struct stripe_head *sh, int disk_idx)
3221 struct md_rdev *rdev;
3223 /* Doing recovery so rcu locking not required */
3224 rdev = sh->raid_conf->disks[disk_idx].replacement;
3226 && !test_bit(Faulty, &rdev->flags)
3227 && !test_bit(In_sync, &rdev->flags)
3228 && (rdev->recovery_offset <= sh->sector
3229 || rdev->mddev->recovery_cp <= sh->sector))
3235 /* fetch_block - checks the given member device to see if its data needs
3236 * to be read or computed to satisfy a request.
3238 * Returns 1 when no more member devices need to be checked, otherwise returns
3239 * 0 to tell the loop in handle_stripe_fill to continue
3242 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3243 int disk_idx, int disks)
3245 struct r5dev *dev = &sh->dev[disk_idx];
3246 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3247 &sh->dev[s->failed_num[1]] };
3251 if (test_bit(R5_LOCKED, &dev->flags) ||
3252 test_bit(R5_UPTODATE, &dev->flags))
3253 /* No point reading this as we already have it or have
3254 * decided to get it.
3259 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3260 /* We need this block to directly satisfy a request */
3263 if (s->syncing || s->expanding ||
3264 (s->replacing && want_replace(sh, disk_idx)))
3265 /* When syncing, or expanding we read everything.
3266 * When replacing, we need the replaced block.
3270 if ((s->failed >= 1 && fdev[0]->toread) ||
3271 (s->failed >= 2 && fdev[1]->toread))
3272 /* If we want to read from a failed device, then
3273 * we need to actually read every other device.
3277 /* Sometimes neither read-modify-write nor reconstruct-write
3278 * cycles can work. In those cases we read every block we
3279 * can. Then the parity-update is certain to have enough to
3281 * This can only be a problem when we need to write something,
3282 * and some device has failed. If either of those tests
3283 * fail we need look no further.
3285 if (!s->failed || !s->to_write)
3288 if (test_bit(R5_Insync, &dev->flags) &&
3289 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3290 /* Pre-reads at not permitted until after short delay
3291 * to gather multiple requests. However if this
3292 * device is no Insync, the block could only be be computed
3293 * and there is no need to delay that.
3297 for (i = 0; i < s->failed; i++) {
3298 if (fdev[i]->towrite &&
3299 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3300 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3301 /* If we have a partial write to a failed
3302 * device, then we will need to reconstruct
3303 * the content of that device, so all other
3304 * devices must be read.
3309 /* If we are forced to do a reconstruct-write, either because
3310 * the current RAID6 implementation only supports that, or
3311 * or because parity cannot be trusted and we are currently
3312 * recovering it, there is extra need to be careful.
3313 * If one of the devices that we would need to read, because
3314 * it is not being overwritten (and maybe not written at all)
3315 * is missing/faulty, then we need to read everything we can.
3317 if (sh->raid_conf->level != 6 &&
3318 sh->sector < sh->raid_conf->mddev->recovery_cp)
3319 /* reconstruct-write isn't being forced */
3321 for (i = 0; i < s->failed; i++) {
3322 if (s->failed_num[i] != sh->pd_idx &&
3323 s->failed_num[i] != sh->qd_idx &&
3324 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3325 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3332 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3333 int disk_idx, int disks)
3335 struct r5dev *dev = &sh->dev[disk_idx];
3337 /* is the data in this block needed, and can we get it? */
3338 if (need_this_block(sh, s, disk_idx, disks)) {
3339 /* we would like to get this block, possibly by computing it,
3340 * otherwise read it if the backing disk is insync
3342 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3343 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3344 BUG_ON(sh->batch_head);
3345 if ((s->uptodate == disks - 1) &&
3346 (s->failed && (disk_idx == s->failed_num[0] ||
3347 disk_idx == s->failed_num[1]))) {
3348 /* have disk failed, and we're requested to fetch it;
3351 pr_debug("Computing stripe %llu block %d\n",
3352 (unsigned long long)sh->sector, disk_idx);
3353 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3354 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3355 set_bit(R5_Wantcompute, &dev->flags);
3356 sh->ops.target = disk_idx;
3357 sh->ops.target2 = -1; /* no 2nd target */
3359 /* Careful: from this point on 'uptodate' is in the eye
3360 * of raid_run_ops which services 'compute' operations
3361 * before writes. R5_Wantcompute flags a block that will
3362 * be R5_UPTODATE by the time it is needed for a
3363 * subsequent operation.
3367 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3368 /* Computing 2-failure is *very* expensive; only
3369 * do it if failed >= 2
3372 for (other = disks; other--; ) {
3373 if (other == disk_idx)
3375 if (!test_bit(R5_UPTODATE,
3376 &sh->dev[other].flags))
3380 pr_debug("Computing stripe %llu blocks %d,%d\n",
3381 (unsigned long long)sh->sector,
3383 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3384 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3385 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3386 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3387 sh->ops.target = disk_idx;
3388 sh->ops.target2 = other;
3392 } else if (test_bit(R5_Insync, &dev->flags)) {
3393 set_bit(R5_LOCKED, &dev->flags);
3394 set_bit(R5_Wantread, &dev->flags);
3396 pr_debug("Reading block %d (sync=%d)\n",
3397 disk_idx, s->syncing);
3405 * handle_stripe_fill - read or compute data to satisfy pending requests.
3407 static void handle_stripe_fill(struct stripe_head *sh,
3408 struct stripe_head_state *s,
3413 /* look for blocks to read/compute, skip this if a compute
3414 * is already in flight, or if the stripe contents are in the
3415 * midst of changing due to a write
3417 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3418 !sh->reconstruct_state)
3419 for (i = disks; i--; )
3420 if (fetch_block(sh, s, i, disks))
3422 set_bit(STRIPE_HANDLE, &sh->state);
3425 static void break_stripe_batch_list(struct stripe_head *head_sh,
3426 unsigned long handle_flags);
3427 /* handle_stripe_clean_event
3428 * any written block on an uptodate or failed drive can be returned.
3429 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3430 * never LOCKED, so we don't need to test 'failed' directly.
3432 static void handle_stripe_clean_event(struct r5conf *conf,
3433 struct stripe_head *sh, int disks, struct bio **return_bi)
3437 int discard_pending = 0;
3438 struct stripe_head *head_sh = sh;
3439 bool do_endio = false;
3441 for (i = disks; i--; )
3442 if (sh->dev[i].written) {
3444 if (!test_bit(R5_LOCKED, &dev->flags) &&
3445 (test_bit(R5_UPTODATE, &dev->flags) ||
3446 test_bit(R5_Discard, &dev->flags) ||
3447 test_bit(R5_SkipCopy, &dev->flags))) {
3448 /* We can return any write requests */
3449 struct bio *wbi, *wbi2;
3450 pr_debug("Return write for disc %d\n", i);
3451 if (test_and_clear_bit(R5_Discard, &dev->flags))
3452 clear_bit(R5_UPTODATE, &dev->flags);
3453 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3454 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3459 dev->page = dev->orig_page;
3461 dev->written = NULL;
3462 while (wbi && wbi->bi_iter.bi_sector <
3463 dev->sector + STRIPE_SECTORS) {
3464 wbi2 = r5_next_bio(wbi, dev->sector);
3465 if (!raid5_dec_bi_active_stripes(wbi)) {
3466 md_write_end(conf->mddev);
3467 wbi->bi_next = *return_bi;
3472 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3474 !test_bit(STRIPE_DEGRADED, &sh->state),
3476 if (head_sh->batch_head) {
3477 sh = list_first_entry(&sh->batch_list,
3480 if (sh != head_sh) {
3487 } else if (test_bit(R5_Discard, &dev->flags))
3488 discard_pending = 1;
3489 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3490 WARN_ON(dev->page != dev->orig_page);
3492 if (!discard_pending &&
3493 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3494 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3495 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3496 if (sh->qd_idx >= 0) {
3497 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3498 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3500 /* now that discard is done we can proceed with any sync */
3501 clear_bit(STRIPE_DISCARD, &sh->state);
3503 * SCSI discard will change some bio fields and the stripe has
3504 * no updated data, so remove it from hash list and the stripe
3505 * will be reinitialized
3507 spin_lock_irq(&conf->device_lock);
3510 if (head_sh->batch_head) {
3511 sh = list_first_entry(&sh->batch_list,
3512 struct stripe_head, batch_list);
3516 spin_unlock_irq(&conf->device_lock);
3519 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3520 set_bit(STRIPE_HANDLE, &sh->state);
3524 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3525 if (atomic_dec_and_test(&conf->pending_full_writes))
3526 md_wakeup_thread(conf->mddev->thread);
3528 if (head_sh->batch_head && do_endio)
3529 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3532 static void handle_stripe_dirtying(struct r5conf *conf,
3533 struct stripe_head *sh,
3534 struct stripe_head_state *s,
3537 int rmw = 0, rcw = 0, i;
3538 sector_t recovery_cp = conf->mddev->recovery_cp;
3540 /* Check whether resync is now happening or should start.
3541 * If yes, then the array is dirty (after unclean shutdown or
3542 * initial creation), so parity in some stripes might be inconsistent.
3543 * In this case, we need to always do reconstruct-write, to ensure
3544 * that in case of drive failure or read-error correction, we
3545 * generate correct data from the parity.
3547 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3548 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3550 /* Calculate the real rcw later - for now make it
3551 * look like rcw is cheaper
3554 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3555 conf->rmw_level, (unsigned long long)recovery_cp,
3556 (unsigned long long)sh->sector);
3557 } else for (i = disks; i--; ) {
3558 /* would I have to read this buffer for read_modify_write */
3559 struct r5dev *dev = &sh->dev[i];
3560 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3561 !test_bit(R5_LOCKED, &dev->flags) &&
3562 !(test_bit(R5_UPTODATE, &dev->flags) ||
3563 test_bit(R5_Wantcompute, &dev->flags))) {
3564 if (test_bit(R5_Insync, &dev->flags))
3567 rmw += 2*disks; /* cannot read it */
3569 /* Would I have to read this buffer for reconstruct_write */
3570 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3571 i != sh->pd_idx && i != sh->qd_idx &&
3572 !test_bit(R5_LOCKED, &dev->flags) &&
3573 !(test_bit(R5_UPTODATE, &dev->flags) ||
3574 test_bit(R5_Wantcompute, &dev->flags))) {
3575 if (test_bit(R5_Insync, &dev->flags))
3581 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3582 (unsigned long long)sh->sector, rmw, rcw);
3583 set_bit(STRIPE_HANDLE, &sh->state);
3584 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3585 /* prefer read-modify-write, but need to get some data */
3586 if (conf->mddev->queue)
3587 blk_add_trace_msg(conf->mddev->queue,
3588 "raid5 rmw %llu %d",
3589 (unsigned long long)sh->sector, rmw);
3590 for (i = disks; i--; ) {
3591 struct r5dev *dev = &sh->dev[i];
3592 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3593 !test_bit(R5_LOCKED, &dev->flags) &&
3594 !(test_bit(R5_UPTODATE, &dev->flags) ||
3595 test_bit(R5_Wantcompute, &dev->flags)) &&
3596 test_bit(R5_Insync, &dev->flags)) {
3597 if (test_bit(STRIPE_PREREAD_ACTIVE,
3599 pr_debug("Read_old block %d for r-m-w\n",
3601 set_bit(R5_LOCKED, &dev->flags);
3602 set_bit(R5_Wantread, &dev->flags);
3605 set_bit(STRIPE_DELAYED, &sh->state);
3606 set_bit(STRIPE_HANDLE, &sh->state);
3611 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3612 /* want reconstruct write, but need to get some data */
3615 for (i = disks; i--; ) {
3616 struct r5dev *dev = &sh->dev[i];
3617 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3618 i != sh->pd_idx && i != sh->qd_idx &&
3619 !test_bit(R5_LOCKED, &dev->flags) &&
3620 !(test_bit(R5_UPTODATE, &dev->flags) ||
3621 test_bit(R5_Wantcompute, &dev->flags))) {
3623 if (test_bit(R5_Insync, &dev->flags) &&
3624 test_bit(STRIPE_PREREAD_ACTIVE,
3626 pr_debug("Read_old block "
3627 "%d for Reconstruct\n", i);
3628 set_bit(R5_LOCKED, &dev->flags);
3629 set_bit(R5_Wantread, &dev->flags);
3633 set_bit(STRIPE_DELAYED, &sh->state);
3634 set_bit(STRIPE_HANDLE, &sh->state);
3638 if (rcw && conf->mddev->queue)
3639 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3640 (unsigned long long)sh->sector,
3641 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3644 if (rcw > disks && rmw > disks &&
3645 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3646 set_bit(STRIPE_DELAYED, &sh->state);
3648 /* now if nothing is locked, and if we have enough data,
3649 * we can start a write request
3651 /* since handle_stripe can be called at any time we need to handle the
3652 * case where a compute block operation has been submitted and then a
3653 * subsequent call wants to start a write request. raid_run_ops only
3654 * handles the case where compute block and reconstruct are requested
3655 * simultaneously. If this is not the case then new writes need to be
3656 * held off until the compute completes.
3658 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3659 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3660 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3661 schedule_reconstruction(sh, s, rcw == 0, 0);
3664 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3665 struct stripe_head_state *s, int disks)
3667 struct r5dev *dev = NULL;
3669 BUG_ON(sh->batch_head);
3670 set_bit(STRIPE_HANDLE, &sh->state);
3672 switch (sh->check_state) {
3673 case check_state_idle:
3674 /* start a new check operation if there are no failures */
3675 if (s->failed == 0) {
3676 BUG_ON(s->uptodate != disks);
3677 sh->check_state = check_state_run;
3678 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3679 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3683 dev = &sh->dev[s->failed_num[0]];
3685 case check_state_compute_result:
3686 sh->check_state = check_state_idle;
3688 dev = &sh->dev[sh->pd_idx];
3690 /* check that a write has not made the stripe insync */
3691 if (test_bit(STRIPE_INSYNC, &sh->state))
3694 /* either failed parity check, or recovery is happening */
3695 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3696 BUG_ON(s->uptodate != disks);
3698 set_bit(R5_LOCKED, &dev->flags);
3700 set_bit(R5_Wantwrite, &dev->flags);
3702 clear_bit(STRIPE_DEGRADED, &sh->state);
3703 set_bit(STRIPE_INSYNC, &sh->state);
3705 case check_state_run:
3706 break; /* we will be called again upon completion */
3707 case check_state_check_result:
3708 sh->check_state = check_state_idle;
3710 /* if a failure occurred during the check operation, leave
3711 * STRIPE_INSYNC not set and let the stripe be handled again
3716 /* handle a successful check operation, if parity is correct
3717 * we are done. Otherwise update the mismatch count and repair
3718 * parity if !MD_RECOVERY_CHECK
3720 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3721 /* parity is correct (on disc,
3722 * not in buffer any more)
3724 set_bit(STRIPE_INSYNC, &sh->state);
3726 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3727 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3728 /* don't try to repair!! */
3729 set_bit(STRIPE_INSYNC, &sh->state);
3731 sh->check_state = check_state_compute_run;
3732 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3733 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3734 set_bit(R5_Wantcompute,
3735 &sh->dev[sh->pd_idx].flags);
3736 sh->ops.target = sh->pd_idx;
3737 sh->ops.target2 = -1;
3742 case check_state_compute_run:
3745 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3746 __func__, sh->check_state,
3747 (unsigned long long) sh->sector);
3752 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3753 struct stripe_head_state *s,
3756 int pd_idx = sh->pd_idx;
3757 int qd_idx = sh->qd_idx;
3760 BUG_ON(sh->batch_head);
3761 set_bit(STRIPE_HANDLE, &sh->state);
3763 BUG_ON(s->failed > 2);
3765 /* Want to check and possibly repair P and Q.
3766 * However there could be one 'failed' device, in which
3767 * case we can only check one of them, possibly using the
3768 * other to generate missing data
3771 switch (sh->check_state) {
3772 case check_state_idle:
3773 /* start a new check operation if there are < 2 failures */
3774 if (s->failed == s->q_failed) {
3775 /* The only possible failed device holds Q, so it
3776 * makes sense to check P (If anything else were failed,
3777 * we would have used P to recreate it).
3779 sh->check_state = check_state_run;
3781 if (!s->q_failed && s->failed < 2) {
3782 /* Q is not failed, and we didn't use it to generate
3783 * anything, so it makes sense to check it
3785 if (sh->check_state == check_state_run)
3786 sh->check_state = check_state_run_pq;
3788 sh->check_state = check_state_run_q;
3791 /* discard potentially stale zero_sum_result */
3792 sh->ops.zero_sum_result = 0;
3794 if (sh->check_state == check_state_run) {
3795 /* async_xor_zero_sum destroys the contents of P */
3796 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3799 if (sh->check_state >= check_state_run &&
3800 sh->check_state <= check_state_run_pq) {
3801 /* async_syndrome_zero_sum preserves P and Q, so
3802 * no need to mark them !uptodate here
3804 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3808 /* we have 2-disk failure */
3809 BUG_ON(s->failed != 2);
3811 case check_state_compute_result:
3812 sh->check_state = check_state_idle;
3814 /* check that a write has not made the stripe insync */
3815 if (test_bit(STRIPE_INSYNC, &sh->state))
3818 /* now write out any block on a failed drive,
3819 * or P or Q if they were recomputed
3821 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3822 if (s->failed == 2) {
3823 dev = &sh->dev[s->failed_num[1]];
3825 set_bit(R5_LOCKED, &dev->flags);
3826 set_bit(R5_Wantwrite, &dev->flags);
3828 if (s->failed >= 1) {
3829 dev = &sh->dev[s->failed_num[0]];
3831 set_bit(R5_LOCKED, &dev->flags);
3832 set_bit(R5_Wantwrite, &dev->flags);
3834 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3835 dev = &sh->dev[pd_idx];
3837 set_bit(R5_LOCKED, &dev->flags);
3838 set_bit(R5_Wantwrite, &dev->flags);
3840 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3841 dev = &sh->dev[qd_idx];
3843 set_bit(R5_LOCKED, &dev->flags);
3844 set_bit(R5_Wantwrite, &dev->flags);
3846 clear_bit(STRIPE_DEGRADED, &sh->state);
3848 set_bit(STRIPE_INSYNC, &sh->state);
3850 case check_state_run:
3851 case check_state_run_q:
3852 case check_state_run_pq:
3853 break; /* we will be called again upon completion */
3854 case check_state_check_result:
3855 sh->check_state = check_state_idle;
3857 /* handle a successful check operation, if parity is correct
3858 * we are done. Otherwise update the mismatch count and repair
3859 * parity if !MD_RECOVERY_CHECK
3861 if (sh->ops.zero_sum_result == 0) {
3862 /* both parities are correct */
3864 set_bit(STRIPE_INSYNC, &sh->state);
3866 /* in contrast to the raid5 case we can validate
3867 * parity, but still have a failure to write
3870 sh->check_state = check_state_compute_result;
3871 /* Returning at this point means that we may go
3872 * off and bring p and/or q uptodate again so
3873 * we make sure to check zero_sum_result again
3874 * to verify if p or q need writeback
3878 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3879 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3880 /* don't try to repair!! */
3881 set_bit(STRIPE_INSYNC, &sh->state);
3883 int *target = &sh->ops.target;
3885 sh->ops.target = -1;
3886 sh->ops.target2 = -1;
3887 sh->check_state = check_state_compute_run;
3888 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3889 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3890 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3891 set_bit(R5_Wantcompute,
3892 &sh->dev[pd_idx].flags);
3894 target = &sh->ops.target2;
3897 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3898 set_bit(R5_Wantcompute,
3899 &sh->dev[qd_idx].flags);
3906 case check_state_compute_run:
3909 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3910 __func__, sh->check_state,
3911 (unsigned long long) sh->sector);
3916 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3920 /* We have read all the blocks in this stripe and now we need to
3921 * copy some of them into a target stripe for expand.
3923 struct dma_async_tx_descriptor *tx = NULL;
3924 BUG_ON(sh->batch_head);
3925 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3926 for (i = 0; i < sh->disks; i++)
3927 if (i != sh->pd_idx && i != sh->qd_idx) {
3929 struct stripe_head *sh2;
3930 struct async_submit_ctl submit;
3932 sector_t bn = compute_blocknr(sh, i, 1);
3933 sector_t s = raid5_compute_sector(conf, bn, 0,
3935 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3937 /* so far only the early blocks of this stripe
3938 * have been requested. When later blocks
3939 * get requested, we will try again
3942 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3943 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3944 /* must have already done this block */
3945 release_stripe(sh2);
3949 /* place all the copies on one channel */
3950 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3951 tx = async_memcpy(sh2->dev[dd_idx].page,
3952 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3955 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3956 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3957 for (j = 0; j < conf->raid_disks; j++)
3958 if (j != sh2->pd_idx &&
3960 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3962 if (j == conf->raid_disks) {
3963 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3964 set_bit(STRIPE_HANDLE, &sh2->state);
3966 release_stripe(sh2);
3969 /* done submitting copies, wait for them to complete */
3970 async_tx_quiesce(&tx);
3974 * handle_stripe - do things to a stripe.
3976 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3977 * state of various bits to see what needs to be done.
3979 * return some read requests which now have data
3980 * return some write requests which are safely on storage
3981 * schedule a read on some buffers
3982 * schedule a write of some buffers
3983 * return confirmation of parity correctness
3987 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3989 struct r5conf *conf = sh->raid_conf;
3990 int disks = sh->disks;
3993 int do_recovery = 0;
3995 memset(s, 0, sizeof(*s));
3997 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
3998 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
3999 s->failed_num[0] = -1;
4000 s->failed_num[1] = -1;
4002 /* Now to look around and see what can be done */
4004 for (i=disks; i--; ) {
4005 struct md_rdev *rdev;
4012 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4014 dev->toread, dev->towrite, dev->written);
4015 /* maybe we can reply to a read
4017 * new wantfill requests are only permitted while
4018 * ops_complete_biofill is guaranteed to be inactive
4020 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4021 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4022 set_bit(R5_Wantfill, &dev->flags);
4024 /* now count some things */
4025 if (test_bit(R5_LOCKED, &dev->flags))
4027 if (test_bit(R5_UPTODATE, &dev->flags))
4029 if (test_bit(R5_Wantcompute, &dev->flags)) {
4031 BUG_ON(s->compute > 2);
4034 if (test_bit(R5_Wantfill, &dev->flags))
4036 else if (dev->toread)
4040 if (!test_bit(R5_OVERWRITE, &dev->flags))
4045 /* Prefer to use the replacement for reads, but only
4046 * if it is recovered enough and has no bad blocks.
4048 rdev = rcu_dereference(conf->disks[i].replacement);
4049 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4050 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4051 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4052 &first_bad, &bad_sectors))
4053 set_bit(R5_ReadRepl, &dev->flags);
4056 set_bit(R5_NeedReplace, &dev->flags);
4057 rdev = rcu_dereference(conf->disks[i].rdev);
4058 clear_bit(R5_ReadRepl, &dev->flags);
4060 if (rdev && test_bit(Faulty, &rdev->flags))
4063 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4064 &first_bad, &bad_sectors);
4065 if (s->blocked_rdev == NULL
4066 && (test_bit(Blocked, &rdev->flags)
4069 set_bit(BlockedBadBlocks,
4071 s->blocked_rdev = rdev;
4072 atomic_inc(&rdev->nr_pending);
4075 clear_bit(R5_Insync, &dev->flags);
4079 /* also not in-sync */
4080 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4081 test_bit(R5_UPTODATE, &dev->flags)) {
4082 /* treat as in-sync, but with a read error
4083 * which we can now try to correct
4085 set_bit(R5_Insync, &dev->flags);
4086 set_bit(R5_ReadError, &dev->flags);
4088 } else if (test_bit(In_sync, &rdev->flags))
4089 set_bit(R5_Insync, &dev->flags);
4090 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4091 /* in sync if before recovery_offset */
4092 set_bit(R5_Insync, &dev->flags);
4093 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4094 test_bit(R5_Expanded, &dev->flags))
4095 /* If we've reshaped into here, we assume it is Insync.
4096 * We will shortly update recovery_offset to make
4099 set_bit(R5_Insync, &dev->flags);
4101 if (test_bit(R5_WriteError, &dev->flags)) {
4102 /* This flag does not apply to '.replacement'
4103 * only to .rdev, so make sure to check that*/
4104 struct md_rdev *rdev2 = rcu_dereference(
4105 conf->disks[i].rdev);
4107 clear_bit(R5_Insync, &dev->flags);
4108 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4109 s->handle_bad_blocks = 1;
4110 atomic_inc(&rdev2->nr_pending);
4112 clear_bit(R5_WriteError, &dev->flags);
4114 if (test_bit(R5_MadeGood, &dev->flags)) {
4115 /* This flag does not apply to '.replacement'
4116 * only to .rdev, so make sure to check that*/
4117 struct md_rdev *rdev2 = rcu_dereference(
4118 conf->disks[i].rdev);
4119 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4120 s->handle_bad_blocks = 1;
4121 atomic_inc(&rdev2->nr_pending);
4123 clear_bit(R5_MadeGood, &dev->flags);
4125 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4126 struct md_rdev *rdev2 = rcu_dereference(
4127 conf->disks[i].replacement);
4128 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4129 s->handle_bad_blocks = 1;
4130 atomic_inc(&rdev2->nr_pending);
4132 clear_bit(R5_MadeGoodRepl, &dev->flags);
4134 if (!test_bit(R5_Insync, &dev->flags)) {
4135 /* The ReadError flag will just be confusing now */
4136 clear_bit(R5_ReadError, &dev->flags);
4137 clear_bit(R5_ReWrite, &dev->flags);
4139 if (test_bit(R5_ReadError, &dev->flags))
4140 clear_bit(R5_Insync, &dev->flags);
4141 if (!test_bit(R5_Insync, &dev->flags)) {
4143 s->failed_num[s->failed] = i;
4145 if (rdev && !test_bit(Faulty, &rdev->flags))
4149 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4150 /* If there is a failed device being replaced,
4151 * we must be recovering.
4152 * else if we are after recovery_cp, we must be syncing
4153 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4154 * else we can only be replacing
4155 * sync and recovery both need to read all devices, and so
4156 * use the same flag.
4159 sh->sector >= conf->mddev->recovery_cp ||
4160 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4168 static int clear_batch_ready(struct stripe_head *sh)
4170 /* Return '1' if this is a member of batch, or
4171 * '0' if it is a lone stripe or a head which can now be
4174 struct stripe_head *tmp;
4175 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4176 return (sh->batch_head && sh->batch_head != sh);
4177 spin_lock(&sh->stripe_lock);
4178 if (!sh->batch_head) {
4179 spin_unlock(&sh->stripe_lock);
4184 * this stripe could be added to a batch list before we check
4185 * BATCH_READY, skips it
4187 if (sh->batch_head != sh) {
4188 spin_unlock(&sh->stripe_lock);
4191 spin_lock(&sh->batch_lock);
4192 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4193 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4194 spin_unlock(&sh->batch_lock);
4195 spin_unlock(&sh->stripe_lock);
4198 * BATCH_READY is cleared, no new stripes can be added.
4199 * batch_list can be accessed without lock
4204 static void break_stripe_batch_list(struct stripe_head *head_sh,
4205 unsigned long handle_flags)
4207 struct stripe_head *sh, *next;
4211 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4213 list_del_init(&sh->batch_list);
4215 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4216 (1 << STRIPE_SYNCING) |
4217 (1 << STRIPE_REPLACED) |
4218 (1 << STRIPE_PREREAD_ACTIVE) |
4219 (1 << STRIPE_DELAYED) |
4220 (1 << STRIPE_BIT_DELAY) |
4221 (1 << STRIPE_FULL_WRITE) |
4222 (1 << STRIPE_BIOFILL_RUN) |
4223 (1 << STRIPE_COMPUTE_RUN) |
4224 (1 << STRIPE_OPS_REQ_PENDING) |
4225 (1 << STRIPE_DISCARD) |
4226 (1 << STRIPE_BATCH_READY) |
4227 (1 << STRIPE_BATCH_ERR) |
4228 (1 << STRIPE_BITMAP_PENDING)));
4229 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4230 (1 << STRIPE_REPLACED)));
4232 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4233 (1 << STRIPE_DEGRADED)),
4234 head_sh->state & (1 << STRIPE_INSYNC));
4236 sh->check_state = head_sh->check_state;
4237 sh->reconstruct_state = head_sh->reconstruct_state;
4238 for (i = 0; i < sh->disks; i++) {
4239 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4241 sh->dev[i].flags = head_sh->dev[i].flags &
4242 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4244 spin_lock_irq(&sh->stripe_lock);
4245 sh->batch_head = NULL;
4246 spin_unlock_irq(&sh->stripe_lock);
4247 if (handle_flags == 0 ||
4248 sh->state & handle_flags)
4249 set_bit(STRIPE_HANDLE, &sh->state);
4252 spin_lock_irq(&head_sh->stripe_lock);
4253 head_sh->batch_head = NULL;
4254 spin_unlock_irq(&head_sh->stripe_lock);
4255 for (i = 0; i < head_sh->disks; i++)
4256 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4258 if (head_sh->state & handle_flags)
4259 set_bit(STRIPE_HANDLE, &head_sh->state);
4262 wake_up(&head_sh->raid_conf->wait_for_overlap);
4265 static void handle_stripe(struct stripe_head *sh)
4267 struct stripe_head_state s;
4268 struct r5conf *conf = sh->raid_conf;
4271 int disks = sh->disks;
4272 struct r5dev *pdev, *qdev;
4274 clear_bit(STRIPE_HANDLE, &sh->state);
4275 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4276 /* already being handled, ensure it gets handled
4277 * again when current action finishes */
4278 set_bit(STRIPE_HANDLE, &sh->state);
4282 if (clear_batch_ready(sh) ) {
4283 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4287 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4288 break_stripe_batch_list(sh, 0);
4290 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4291 spin_lock(&sh->stripe_lock);
4292 /* Cannot process 'sync' concurrently with 'discard' */
4293 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4294 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4295 set_bit(STRIPE_SYNCING, &sh->state);
4296 clear_bit(STRIPE_INSYNC, &sh->state);
4297 clear_bit(STRIPE_REPLACED, &sh->state);
4299 spin_unlock(&sh->stripe_lock);
4301 clear_bit(STRIPE_DELAYED, &sh->state);
4303 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4304 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4305 (unsigned long long)sh->sector, sh->state,
4306 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4307 sh->check_state, sh->reconstruct_state);
4309 analyse_stripe(sh, &s);
4311 if (s.handle_bad_blocks) {
4312 set_bit(STRIPE_HANDLE, &sh->state);
4316 if (unlikely(s.blocked_rdev)) {
4317 if (s.syncing || s.expanding || s.expanded ||
4318 s.replacing || s.to_write || s.written) {
4319 set_bit(STRIPE_HANDLE, &sh->state);
4322 /* There is nothing for the blocked_rdev to block */
4323 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4324 s.blocked_rdev = NULL;
4327 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4328 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4329 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4332 pr_debug("locked=%d uptodate=%d to_read=%d"
4333 " to_write=%d failed=%d failed_num=%d,%d\n",
4334 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4335 s.failed_num[0], s.failed_num[1]);
4336 /* check if the array has lost more than max_degraded devices and,
4337 * if so, some requests might need to be failed.
4339 if (s.failed > conf->max_degraded) {
4340 sh->check_state = 0;
4341 sh->reconstruct_state = 0;
4342 break_stripe_batch_list(sh, 0);
4343 if (s.to_read+s.to_write+s.written)
4344 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4345 if (s.syncing + s.replacing)
4346 handle_failed_sync(conf, sh, &s);
4349 /* Now we check to see if any write operations have recently
4353 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4355 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4356 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4357 sh->reconstruct_state = reconstruct_state_idle;
4359 /* All the 'written' buffers and the parity block are ready to
4360 * be written back to disk
4362 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4363 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4364 BUG_ON(sh->qd_idx >= 0 &&
4365 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4366 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4367 for (i = disks; i--; ) {
4368 struct r5dev *dev = &sh->dev[i];
4369 if (test_bit(R5_LOCKED, &dev->flags) &&
4370 (i == sh->pd_idx || i == sh->qd_idx ||
4372 pr_debug("Writing block %d\n", i);
4373 set_bit(R5_Wantwrite, &dev->flags);
4378 if (!test_bit(R5_Insync, &dev->flags) ||
4379 ((i == sh->pd_idx || i == sh->qd_idx) &&
4381 set_bit(STRIPE_INSYNC, &sh->state);
4384 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4385 s.dec_preread_active = 1;
4389 * might be able to return some write requests if the parity blocks
4390 * are safe, or on a failed drive
4392 pdev = &sh->dev[sh->pd_idx];
4393 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4394 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4395 qdev = &sh->dev[sh->qd_idx];
4396 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4397 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4401 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4402 && !test_bit(R5_LOCKED, &pdev->flags)
4403 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4404 test_bit(R5_Discard, &pdev->flags))))) &&
4405 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4406 && !test_bit(R5_LOCKED, &qdev->flags)
4407 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4408 test_bit(R5_Discard, &qdev->flags))))))
4409 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4411 /* Now we might consider reading some blocks, either to check/generate
4412 * parity, or to satisfy requests
4413 * or to load a block that is being partially written.
4415 if (s.to_read || s.non_overwrite
4416 || (conf->level == 6 && s.to_write && s.failed)
4417 || (s.syncing && (s.uptodate + s.compute < disks))
4420 handle_stripe_fill(sh, &s, disks);
4422 /* Now to consider new write requests and what else, if anything
4423 * should be read. We do not handle new writes when:
4424 * 1/ A 'write' operation (copy+xor) is already in flight.
4425 * 2/ A 'check' operation is in flight, as it may clobber the parity
4428 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4429 handle_stripe_dirtying(conf, sh, &s, disks);
4431 /* maybe we need to check and possibly fix the parity for this stripe
4432 * Any reads will already have been scheduled, so we just see if enough
4433 * data is available. The parity check is held off while parity
4434 * dependent operations are in flight.
4436 if (sh->check_state ||
4437 (s.syncing && s.locked == 0 &&
4438 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4439 !test_bit(STRIPE_INSYNC, &sh->state))) {
4440 if (conf->level == 6)
4441 handle_parity_checks6(conf, sh, &s, disks);
4443 handle_parity_checks5(conf, sh, &s, disks);
4446 if ((s.replacing || s.syncing) && s.locked == 0
4447 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4448 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4449 /* Write out to replacement devices where possible */
4450 for (i = 0; i < conf->raid_disks; i++)
4451 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4452 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4453 set_bit(R5_WantReplace, &sh->dev[i].flags);
4454 set_bit(R5_LOCKED, &sh->dev[i].flags);
4458 set_bit(STRIPE_INSYNC, &sh->state);
4459 set_bit(STRIPE_REPLACED, &sh->state);
4461 if ((s.syncing || s.replacing) && s.locked == 0 &&
4462 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4463 test_bit(STRIPE_INSYNC, &sh->state)) {
4464 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4465 clear_bit(STRIPE_SYNCING, &sh->state);
4466 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4467 wake_up(&conf->wait_for_overlap);
4470 /* If the failed drives are just a ReadError, then we might need
4471 * to progress the repair/check process
4473 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4474 for (i = 0; i < s.failed; i++) {
4475 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4476 if (test_bit(R5_ReadError, &dev->flags)
4477 && !test_bit(R5_LOCKED, &dev->flags)
4478 && test_bit(R5_UPTODATE, &dev->flags)
4480 if (!test_bit(R5_ReWrite, &dev->flags)) {
4481 set_bit(R5_Wantwrite, &dev->flags);
4482 set_bit(R5_ReWrite, &dev->flags);
4483 set_bit(R5_LOCKED, &dev->flags);
4486 /* let's read it back */
4487 set_bit(R5_Wantread, &dev->flags);
4488 set_bit(R5_LOCKED, &dev->flags);
4494 /* Finish reconstruct operations initiated by the expansion process */
4495 if (sh->reconstruct_state == reconstruct_state_result) {
4496 struct stripe_head *sh_src
4497 = get_active_stripe(conf, sh->sector, 1, 1, 1);
4498 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4499 /* sh cannot be written until sh_src has been read.
4500 * so arrange for sh to be delayed a little
4502 set_bit(STRIPE_DELAYED, &sh->state);
4503 set_bit(STRIPE_HANDLE, &sh->state);
4504 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4506 atomic_inc(&conf->preread_active_stripes);
4507 release_stripe(sh_src);
4511 release_stripe(sh_src);
4513 sh->reconstruct_state = reconstruct_state_idle;
4514 clear_bit(STRIPE_EXPANDING, &sh->state);
4515 for (i = conf->raid_disks; i--; ) {
4516 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4517 set_bit(R5_LOCKED, &sh->dev[i].flags);
4522 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4523 !sh->reconstruct_state) {
4524 /* Need to write out all blocks after computing parity */
4525 sh->disks = conf->raid_disks;
4526 stripe_set_idx(sh->sector, conf, 0, sh);
4527 schedule_reconstruction(sh, &s, 1, 1);
4528 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4529 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4530 atomic_dec(&conf->reshape_stripes);
4531 wake_up(&conf->wait_for_overlap);
4532 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4535 if (s.expanding && s.locked == 0 &&
4536 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4537 handle_stripe_expansion(conf, sh);
4540 /* wait for this device to become unblocked */
4541 if (unlikely(s.blocked_rdev)) {
4542 if (conf->mddev->external)
4543 md_wait_for_blocked_rdev(s.blocked_rdev,
4546 /* Internal metadata will immediately
4547 * be written by raid5d, so we don't
4548 * need to wait here.
4550 rdev_dec_pending(s.blocked_rdev,
4554 if (s.handle_bad_blocks)
4555 for (i = disks; i--; ) {
4556 struct md_rdev *rdev;
4557 struct r5dev *dev = &sh->dev[i];
4558 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4559 /* We own a safe reference to the rdev */
4560 rdev = conf->disks[i].rdev;
4561 if (!rdev_set_badblocks(rdev, sh->sector,
4563 md_error(conf->mddev, rdev);
4564 rdev_dec_pending(rdev, conf->mddev);
4566 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4567 rdev = conf->disks[i].rdev;
4568 rdev_clear_badblocks(rdev, sh->sector,
4570 rdev_dec_pending(rdev, conf->mddev);
4572 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4573 rdev = conf->disks[i].replacement;
4575 /* rdev have been moved down */
4576 rdev = conf->disks[i].rdev;
4577 rdev_clear_badblocks(rdev, sh->sector,
4579 rdev_dec_pending(rdev, conf->mddev);
4584 raid_run_ops(sh, s.ops_request);
4588 if (s.dec_preread_active) {
4589 /* We delay this until after ops_run_io so that if make_request
4590 * is waiting on a flush, it won't continue until the writes
4591 * have actually been submitted.
4593 atomic_dec(&conf->preread_active_stripes);
4594 if (atomic_read(&conf->preread_active_stripes) <
4596 md_wakeup_thread(conf->mddev->thread);
4599 return_io(s.return_bi);
4601 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4604 static void raid5_activate_delayed(struct r5conf *conf)
4606 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4607 while (!list_empty(&conf->delayed_list)) {
4608 struct list_head *l = conf->delayed_list.next;
4609 struct stripe_head *sh;
4610 sh = list_entry(l, struct stripe_head, lru);
4612 clear_bit(STRIPE_DELAYED, &sh->state);
4613 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4614 atomic_inc(&conf->preread_active_stripes);
4615 list_add_tail(&sh->lru, &conf->hold_list);
4616 raid5_wakeup_stripe_thread(sh);
4621 static void activate_bit_delay(struct r5conf *conf,
4622 struct list_head *temp_inactive_list)
4624 /* device_lock is held */
4625 struct list_head head;
4626 list_add(&head, &conf->bitmap_list);
4627 list_del_init(&conf->bitmap_list);
4628 while (!list_empty(&head)) {
4629 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4631 list_del_init(&sh->lru);
4632 atomic_inc(&sh->count);
4633 hash = sh->hash_lock_index;
4634 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4638 static int raid5_congested(struct mddev *mddev, int bits)
4640 struct r5conf *conf = mddev->private;
4642 /* No difference between reads and writes. Just check
4643 * how busy the stripe_cache is
4646 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4650 if (atomic_read(&conf->empty_inactive_list_nr))
4656 /* We want read requests to align with chunks where possible,
4657 * but write requests don't need to.
4659 static int raid5_mergeable_bvec(struct mddev *mddev,
4660 struct bvec_merge_data *bvm,
4661 struct bio_vec *biovec)
4663 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4665 unsigned int chunk_sectors = mddev->chunk_sectors;
4666 unsigned int bio_sectors = bvm->bi_size >> 9;
4669 * always allow writes to be mergeable, read as well if array
4670 * is degraded as we'll go through stripe cache anyway.
4672 if ((bvm->bi_rw & 1) == WRITE || mddev->degraded)
4673 return biovec->bv_len;
4675 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4676 chunk_sectors = mddev->new_chunk_sectors;
4677 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4678 if (max < 0) max = 0;
4679 if (max <= biovec->bv_len && bio_sectors == 0)
4680 return biovec->bv_len;
4685 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4687 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4688 unsigned int chunk_sectors = mddev->chunk_sectors;
4689 unsigned int bio_sectors = bio_sectors(bio);
4691 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4692 chunk_sectors = mddev->new_chunk_sectors;
4693 return chunk_sectors >=
4694 ((sector & (chunk_sectors - 1)) + bio_sectors);
4698 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4699 * later sampled by raid5d.
4701 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4703 unsigned long flags;
4705 spin_lock_irqsave(&conf->device_lock, flags);
4707 bi->bi_next = conf->retry_read_aligned_list;
4708 conf->retry_read_aligned_list = bi;
4710 spin_unlock_irqrestore(&conf->device_lock, flags);
4711 md_wakeup_thread(conf->mddev->thread);
4714 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4718 bi = conf->retry_read_aligned;
4720 conf->retry_read_aligned = NULL;
4723 bi = conf->retry_read_aligned_list;
4725 conf->retry_read_aligned_list = bi->bi_next;
4728 * this sets the active strip count to 1 and the processed
4729 * strip count to zero (upper 8 bits)
4731 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4738 * The "raid5_align_endio" should check if the read succeeded and if it
4739 * did, call bio_endio on the original bio (having bio_put the new bio
4741 * If the read failed..
4743 static void raid5_align_endio(struct bio *bi, int error)
4745 struct bio* raid_bi = bi->bi_private;
4746 struct mddev *mddev;
4747 struct r5conf *conf;
4748 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4749 struct md_rdev *rdev;
4753 rdev = (void*)raid_bi->bi_next;
4754 raid_bi->bi_next = NULL;
4755 mddev = rdev->mddev;
4756 conf = mddev->private;
4758 rdev_dec_pending(rdev, conf->mddev);
4760 if (!error && uptodate) {
4761 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4763 bio_endio(raid_bi, 0);
4764 if (atomic_dec_and_test(&conf->active_aligned_reads))
4765 wake_up(&conf->wait_for_stripe);
4769 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4771 add_bio_to_retry(raid_bi, conf);
4774 static int bio_fits_rdev(struct bio *bi)
4776 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4778 if (bio_sectors(bi) > queue_max_sectors(q))
4780 blk_recount_segments(q, bi);
4781 if (bi->bi_phys_segments > queue_max_segments(q))
4784 if (q->merge_bvec_fn)
4785 /* it's too hard to apply the merge_bvec_fn at this stage,
4793 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4795 struct r5conf *conf = mddev->private;
4797 struct bio* align_bi;
4798 struct md_rdev *rdev;
4799 sector_t end_sector;
4801 if (!in_chunk_boundary(mddev, raid_bio)) {
4802 pr_debug("chunk_aligned_read : non aligned\n");
4806 * use bio_clone_mddev to make a copy of the bio
4808 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4812 * set bi_end_io to a new function, and set bi_private to the
4815 align_bi->bi_end_io = raid5_align_endio;
4816 align_bi->bi_private = raid_bio;
4820 align_bi->bi_iter.bi_sector =
4821 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4824 end_sector = bio_end_sector(align_bi);
4826 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4827 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4828 rdev->recovery_offset < end_sector) {
4829 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4831 (test_bit(Faulty, &rdev->flags) ||
4832 !(test_bit(In_sync, &rdev->flags) ||
4833 rdev->recovery_offset >= end_sector)))
4840 atomic_inc(&rdev->nr_pending);
4842 raid_bio->bi_next = (void*)rdev;
4843 align_bi->bi_bdev = rdev->bdev;
4844 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
4846 if (!bio_fits_rdev(align_bi) ||
4847 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4848 bio_sectors(align_bi),
4849 &first_bad, &bad_sectors)) {
4850 /* too big in some way, or has a known bad block */
4852 rdev_dec_pending(rdev, mddev);
4856 /* No reshape active, so we can trust rdev->data_offset */
4857 align_bi->bi_iter.bi_sector += rdev->data_offset;
4859 spin_lock_irq(&conf->device_lock);
4860 wait_event_lock_irq(conf->wait_for_stripe,
4863 atomic_inc(&conf->active_aligned_reads);
4864 spin_unlock_irq(&conf->device_lock);
4867 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4868 align_bi, disk_devt(mddev->gendisk),
4869 raid_bio->bi_iter.bi_sector);
4870 generic_make_request(align_bi);
4879 /* __get_priority_stripe - get the next stripe to process
4881 * Full stripe writes are allowed to pass preread active stripes up until
4882 * the bypass_threshold is exceeded. In general the bypass_count
4883 * increments when the handle_list is handled before the hold_list; however, it
4884 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4885 * stripe with in flight i/o. The bypass_count will be reset when the
4886 * head of the hold_list has changed, i.e. the head was promoted to the
4889 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4891 struct stripe_head *sh = NULL, *tmp;
4892 struct list_head *handle_list = NULL;
4893 struct r5worker_group *wg = NULL;
4895 if (conf->worker_cnt_per_group == 0) {
4896 handle_list = &conf->handle_list;
4897 } else if (group != ANY_GROUP) {
4898 handle_list = &conf->worker_groups[group].handle_list;
4899 wg = &conf->worker_groups[group];
4902 for (i = 0; i < conf->group_cnt; i++) {
4903 handle_list = &conf->worker_groups[i].handle_list;
4904 wg = &conf->worker_groups[i];
4905 if (!list_empty(handle_list))
4910 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4912 list_empty(handle_list) ? "empty" : "busy",
4913 list_empty(&conf->hold_list) ? "empty" : "busy",
4914 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4916 if (!list_empty(handle_list)) {
4917 sh = list_entry(handle_list->next, typeof(*sh), lru);
4919 if (list_empty(&conf->hold_list))
4920 conf->bypass_count = 0;
4921 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4922 if (conf->hold_list.next == conf->last_hold)
4923 conf->bypass_count++;
4925 conf->last_hold = conf->hold_list.next;
4926 conf->bypass_count -= conf->bypass_threshold;
4927 if (conf->bypass_count < 0)
4928 conf->bypass_count = 0;
4931 } else if (!list_empty(&conf->hold_list) &&
4932 ((conf->bypass_threshold &&
4933 conf->bypass_count > conf->bypass_threshold) ||
4934 atomic_read(&conf->pending_full_writes) == 0)) {
4936 list_for_each_entry(tmp, &conf->hold_list, lru) {
4937 if (conf->worker_cnt_per_group == 0 ||
4938 group == ANY_GROUP ||
4939 !cpu_online(tmp->cpu) ||
4940 cpu_to_group(tmp->cpu) == group) {
4947 conf->bypass_count -= conf->bypass_threshold;
4948 if (conf->bypass_count < 0)
4949 conf->bypass_count = 0;
4961 list_del_init(&sh->lru);
4962 BUG_ON(atomic_inc_return(&sh->count) != 1);
4966 struct raid5_plug_cb {
4967 struct blk_plug_cb cb;
4968 struct list_head list;
4969 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4972 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4974 struct raid5_plug_cb *cb = container_of(
4975 blk_cb, struct raid5_plug_cb, cb);
4976 struct stripe_head *sh;
4977 struct mddev *mddev = cb->cb.data;
4978 struct r5conf *conf = mddev->private;
4982 if (cb->list.next && !list_empty(&cb->list)) {
4983 spin_lock_irq(&conf->device_lock);
4984 while (!list_empty(&cb->list)) {
4985 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4986 list_del_init(&sh->lru);
4988 * avoid race release_stripe_plug() sees
4989 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4990 * is still in our list
4992 smp_mb__before_atomic();
4993 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4995 * STRIPE_ON_RELEASE_LIST could be set here. In that
4996 * case, the count is always > 1 here
4998 hash = sh->hash_lock_index;
4999 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5002 spin_unlock_irq(&conf->device_lock);
5004 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5005 NR_STRIPE_HASH_LOCKS);
5007 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5011 static void release_stripe_plug(struct mddev *mddev,
5012 struct stripe_head *sh)
5014 struct blk_plug_cb *blk_cb = blk_check_plugged(
5015 raid5_unplug, mddev,
5016 sizeof(struct raid5_plug_cb));
5017 struct raid5_plug_cb *cb;
5024 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5026 if (cb->list.next == NULL) {
5028 INIT_LIST_HEAD(&cb->list);
5029 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5030 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5033 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5034 list_add_tail(&sh->lru, &cb->list);
5039 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5041 struct r5conf *conf = mddev->private;
5042 sector_t logical_sector, last_sector;
5043 struct stripe_head *sh;
5047 if (mddev->reshape_position != MaxSector)
5048 /* Skip discard while reshape is happening */
5051 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5052 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5055 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5057 stripe_sectors = conf->chunk_sectors *
5058 (conf->raid_disks - conf->max_degraded);
5059 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5061 sector_div(last_sector, stripe_sectors);
5063 logical_sector *= conf->chunk_sectors;
5064 last_sector *= conf->chunk_sectors;
5066 for (; logical_sector < last_sector;
5067 logical_sector += STRIPE_SECTORS) {
5071 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
5072 prepare_to_wait(&conf->wait_for_overlap, &w,
5073 TASK_UNINTERRUPTIBLE);
5074 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5075 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5080 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5081 spin_lock_irq(&sh->stripe_lock);
5082 for (d = 0; d < conf->raid_disks; d++) {
5083 if (d == sh->pd_idx || d == sh->qd_idx)
5085 if (sh->dev[d].towrite || sh->dev[d].toread) {
5086 set_bit(R5_Overlap, &sh->dev[d].flags);
5087 spin_unlock_irq(&sh->stripe_lock);
5093 set_bit(STRIPE_DISCARD, &sh->state);
5094 finish_wait(&conf->wait_for_overlap, &w);
5095 sh->overwrite_disks = 0;
5096 for (d = 0; d < conf->raid_disks; d++) {
5097 if (d == sh->pd_idx || d == sh->qd_idx)
5099 sh->dev[d].towrite = bi;
5100 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5101 raid5_inc_bi_active_stripes(bi);
5102 sh->overwrite_disks++;
5104 spin_unlock_irq(&sh->stripe_lock);
5105 if (conf->mddev->bitmap) {
5107 d < conf->raid_disks - conf->max_degraded;
5109 bitmap_startwrite(mddev->bitmap,
5113 sh->bm_seq = conf->seq_flush + 1;
5114 set_bit(STRIPE_BIT_DELAY, &sh->state);
5117 set_bit(STRIPE_HANDLE, &sh->state);
5118 clear_bit(STRIPE_DELAYED, &sh->state);
5119 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5120 atomic_inc(&conf->preread_active_stripes);
5121 release_stripe_plug(mddev, sh);
5124 remaining = raid5_dec_bi_active_stripes(bi);
5125 if (remaining == 0) {
5126 md_write_end(mddev);
5131 static void make_request(struct mddev *mddev, struct bio * bi)
5133 struct r5conf *conf = mddev->private;
5135 sector_t new_sector;
5136 sector_t logical_sector, last_sector;
5137 struct stripe_head *sh;
5138 const int rw = bio_data_dir(bi);
5143 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5144 md_flush_request(mddev, bi);
5148 md_write_start(mddev, bi);
5151 * If array is degraded, better not do chunk aligned read because
5152 * later we might have to read it again in order to reconstruct
5153 * data on failed drives.
5155 if (rw == READ && mddev->degraded == 0 &&
5156 mddev->reshape_position == MaxSector &&
5157 chunk_aligned_read(mddev,bi))
5160 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5161 make_discard_request(mddev, bi);
5165 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5166 last_sector = bio_end_sector(bi);
5168 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5170 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5171 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5177 seq = read_seqcount_begin(&conf->gen_lock);
5180 prepare_to_wait(&conf->wait_for_overlap, &w,
5181 TASK_UNINTERRUPTIBLE);
5182 if (unlikely(conf->reshape_progress != MaxSector)) {
5183 /* spinlock is needed as reshape_progress may be
5184 * 64bit on a 32bit platform, and so it might be
5185 * possible to see a half-updated value
5186 * Of course reshape_progress could change after
5187 * the lock is dropped, so once we get a reference
5188 * to the stripe that we think it is, we will have
5191 spin_lock_irq(&conf->device_lock);
5192 if (mddev->reshape_backwards
5193 ? logical_sector < conf->reshape_progress
5194 : logical_sector >= conf->reshape_progress) {
5197 if (mddev->reshape_backwards
5198 ? logical_sector < conf->reshape_safe
5199 : logical_sector >= conf->reshape_safe) {
5200 spin_unlock_irq(&conf->device_lock);
5206 spin_unlock_irq(&conf->device_lock);
5209 new_sector = raid5_compute_sector(conf, logical_sector,
5212 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5213 (unsigned long long)new_sector,
5214 (unsigned long long)logical_sector);
5216 sh = get_active_stripe(conf, new_sector, previous,
5217 (bi->bi_rw&RWA_MASK), 0);
5219 if (unlikely(previous)) {
5220 /* expansion might have moved on while waiting for a
5221 * stripe, so we must do the range check again.
5222 * Expansion could still move past after this
5223 * test, but as we are holding a reference to
5224 * 'sh', we know that if that happens,
5225 * STRIPE_EXPANDING will get set and the expansion
5226 * won't proceed until we finish with the stripe.
5229 spin_lock_irq(&conf->device_lock);
5230 if (mddev->reshape_backwards
5231 ? logical_sector >= conf->reshape_progress
5232 : logical_sector < conf->reshape_progress)
5233 /* mismatch, need to try again */
5235 spin_unlock_irq(&conf->device_lock);
5243 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5244 /* Might have got the wrong stripe_head
5252 logical_sector >= mddev->suspend_lo &&
5253 logical_sector < mddev->suspend_hi) {
5255 /* As the suspend_* range is controlled by
5256 * userspace, we want an interruptible
5259 flush_signals(current);
5260 prepare_to_wait(&conf->wait_for_overlap,
5261 &w, TASK_INTERRUPTIBLE);
5262 if (logical_sector >= mddev->suspend_lo &&
5263 logical_sector < mddev->suspend_hi) {
5270 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5271 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5272 /* Stripe is busy expanding or
5273 * add failed due to overlap. Flush everything
5276 md_wakeup_thread(mddev->thread);
5282 set_bit(STRIPE_HANDLE, &sh->state);
5283 clear_bit(STRIPE_DELAYED, &sh->state);
5284 if ((!sh->batch_head || sh == sh->batch_head) &&
5285 (bi->bi_rw & REQ_SYNC) &&
5286 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5287 atomic_inc(&conf->preread_active_stripes);
5288 release_stripe_plug(mddev, sh);
5290 /* cannot get stripe for read-ahead, just give-up */
5291 clear_bit(BIO_UPTODATE, &bi->bi_flags);
5295 finish_wait(&conf->wait_for_overlap, &w);
5297 remaining = raid5_dec_bi_active_stripes(bi);
5298 if (remaining == 0) {
5301 md_write_end(mddev);
5303 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5309 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5311 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5313 /* reshaping is quite different to recovery/resync so it is
5314 * handled quite separately ... here.
5316 * On each call to sync_request, we gather one chunk worth of
5317 * destination stripes and flag them as expanding.
5318 * Then we find all the source stripes and request reads.
5319 * As the reads complete, handle_stripe will copy the data
5320 * into the destination stripe and release that stripe.
5322 struct r5conf *conf = mddev->private;
5323 struct stripe_head *sh;
5324 sector_t first_sector, last_sector;
5325 int raid_disks = conf->previous_raid_disks;
5326 int data_disks = raid_disks - conf->max_degraded;
5327 int new_data_disks = conf->raid_disks - conf->max_degraded;
5330 sector_t writepos, readpos, safepos;
5331 sector_t stripe_addr;
5332 int reshape_sectors;
5333 struct list_head stripes;
5335 if (sector_nr == 0) {
5336 /* If restarting in the middle, skip the initial sectors */
5337 if (mddev->reshape_backwards &&
5338 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5339 sector_nr = raid5_size(mddev, 0, 0)
5340 - conf->reshape_progress;
5341 } else if (!mddev->reshape_backwards &&
5342 conf->reshape_progress > 0)
5343 sector_nr = conf->reshape_progress;
5344 sector_div(sector_nr, new_data_disks);
5346 mddev->curr_resync_completed = sector_nr;
5347 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5353 /* We need to process a full chunk at a time.
5354 * If old and new chunk sizes differ, we need to process the
5357 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
5358 reshape_sectors = mddev->new_chunk_sectors;
5360 reshape_sectors = mddev->chunk_sectors;
5362 /* We update the metadata at least every 10 seconds, or when
5363 * the data about to be copied would over-write the source of
5364 * the data at the front of the range. i.e. one new_stripe
5365 * along from reshape_progress new_maps to after where
5366 * reshape_safe old_maps to
5368 writepos = conf->reshape_progress;
5369 sector_div(writepos, new_data_disks);
5370 readpos = conf->reshape_progress;
5371 sector_div(readpos, data_disks);
5372 safepos = conf->reshape_safe;
5373 sector_div(safepos, data_disks);
5374 if (mddev->reshape_backwards) {
5375 writepos -= min_t(sector_t, reshape_sectors, writepos);
5376 readpos += reshape_sectors;
5377 safepos += reshape_sectors;
5379 writepos += reshape_sectors;
5380 readpos -= min_t(sector_t, reshape_sectors, readpos);
5381 safepos -= min_t(sector_t, reshape_sectors, safepos);
5384 /* Having calculated the 'writepos' possibly use it
5385 * to set 'stripe_addr' which is where we will write to.
5387 if (mddev->reshape_backwards) {
5388 BUG_ON(conf->reshape_progress == 0);
5389 stripe_addr = writepos;
5390 BUG_ON((mddev->dev_sectors &
5391 ~((sector_t)reshape_sectors - 1))
5392 - reshape_sectors - stripe_addr
5395 BUG_ON(writepos != sector_nr + reshape_sectors);
5396 stripe_addr = sector_nr;
5399 /* 'writepos' is the most advanced device address we might write.
5400 * 'readpos' is the least advanced device address we might read.
5401 * 'safepos' is the least address recorded in the metadata as having
5403 * If there is a min_offset_diff, these are adjusted either by
5404 * increasing the safepos/readpos if diff is negative, or
5405 * increasing writepos if diff is positive.
5406 * If 'readpos' is then behind 'writepos', there is no way that we can
5407 * ensure safety in the face of a crash - that must be done by userspace
5408 * making a backup of the data. So in that case there is no particular
5409 * rush to update metadata.
5410 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5411 * update the metadata to advance 'safepos' to match 'readpos' so that
5412 * we can be safe in the event of a crash.
5413 * So we insist on updating metadata if safepos is behind writepos and
5414 * readpos is beyond writepos.
5415 * In any case, update the metadata every 10 seconds.
5416 * Maybe that number should be configurable, but I'm not sure it is
5417 * worth it.... maybe it could be a multiple of safemode_delay???
5419 if (conf->min_offset_diff < 0) {
5420 safepos += -conf->min_offset_diff;
5421 readpos += -conf->min_offset_diff;
5423 writepos += conf->min_offset_diff;
5425 if ((mddev->reshape_backwards
5426 ? (safepos > writepos && readpos < writepos)
5427 : (safepos < writepos && readpos > writepos)) ||
5428 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5429 /* Cannot proceed until we've updated the superblock... */
5430 wait_event(conf->wait_for_overlap,
5431 atomic_read(&conf->reshape_stripes)==0
5432 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5433 if (atomic_read(&conf->reshape_stripes) != 0)
5435 mddev->reshape_position = conf->reshape_progress;
5436 mddev->curr_resync_completed = sector_nr;
5437 conf->reshape_checkpoint = jiffies;
5438 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5439 md_wakeup_thread(mddev->thread);
5440 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5441 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5442 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5444 spin_lock_irq(&conf->device_lock);
5445 conf->reshape_safe = mddev->reshape_position;
5446 spin_unlock_irq(&conf->device_lock);
5447 wake_up(&conf->wait_for_overlap);
5448 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5451 INIT_LIST_HEAD(&stripes);
5452 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5454 int skipped_disk = 0;
5455 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5456 set_bit(STRIPE_EXPANDING, &sh->state);
5457 atomic_inc(&conf->reshape_stripes);
5458 /* If any of this stripe is beyond the end of the old
5459 * array, then we need to zero those blocks
5461 for (j=sh->disks; j--;) {
5463 if (j == sh->pd_idx)
5465 if (conf->level == 6 &&
5468 s = compute_blocknr(sh, j, 0);
5469 if (s < raid5_size(mddev, 0, 0)) {
5473 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5474 set_bit(R5_Expanded, &sh->dev[j].flags);
5475 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5477 if (!skipped_disk) {
5478 set_bit(STRIPE_EXPAND_READY, &sh->state);
5479 set_bit(STRIPE_HANDLE, &sh->state);
5481 list_add(&sh->lru, &stripes);
5483 spin_lock_irq(&conf->device_lock);
5484 if (mddev->reshape_backwards)
5485 conf->reshape_progress -= reshape_sectors * new_data_disks;
5487 conf->reshape_progress += reshape_sectors * new_data_disks;
5488 spin_unlock_irq(&conf->device_lock);
5489 /* Ok, those stripe are ready. We can start scheduling
5490 * reads on the source stripes.
5491 * The source stripes are determined by mapping the first and last
5492 * block on the destination stripes.
5495 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5498 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5499 * new_data_disks - 1),
5501 if (last_sector >= mddev->dev_sectors)
5502 last_sector = mddev->dev_sectors - 1;
5503 while (first_sector <= last_sector) {
5504 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
5505 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5506 set_bit(STRIPE_HANDLE, &sh->state);
5508 first_sector += STRIPE_SECTORS;
5510 /* Now that the sources are clearly marked, we can release
5511 * the destination stripes
5513 while (!list_empty(&stripes)) {
5514 sh = list_entry(stripes.next, struct stripe_head, lru);
5515 list_del_init(&sh->lru);
5518 /* If this takes us to the resync_max point where we have to pause,
5519 * then we need to write out the superblock.
5521 sector_nr += reshape_sectors;
5522 if ((sector_nr - mddev->curr_resync_completed) * 2
5523 >= mddev->resync_max - mddev->curr_resync_completed) {
5524 /* Cannot proceed until we've updated the superblock... */
5525 wait_event(conf->wait_for_overlap,
5526 atomic_read(&conf->reshape_stripes) == 0
5527 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5528 if (atomic_read(&conf->reshape_stripes) != 0)
5530 mddev->reshape_position = conf->reshape_progress;
5531 mddev->curr_resync_completed = sector_nr;
5532 conf->reshape_checkpoint = jiffies;
5533 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5534 md_wakeup_thread(mddev->thread);
5535 wait_event(mddev->sb_wait,
5536 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5537 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5538 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5540 spin_lock_irq(&conf->device_lock);
5541 conf->reshape_safe = mddev->reshape_position;
5542 spin_unlock_irq(&conf->device_lock);
5543 wake_up(&conf->wait_for_overlap);
5544 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5547 return reshape_sectors;
5550 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5552 struct r5conf *conf = mddev->private;
5553 struct stripe_head *sh;
5554 sector_t max_sector = mddev->dev_sectors;
5555 sector_t sync_blocks;
5556 int still_degraded = 0;
5559 if (sector_nr >= max_sector) {
5560 /* just being told to finish up .. nothing much to do */
5562 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5567 if (mddev->curr_resync < max_sector) /* aborted */
5568 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5570 else /* completed sync */
5572 bitmap_close_sync(mddev->bitmap);
5577 /* Allow raid5_quiesce to complete */
5578 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5580 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5581 return reshape_request(mddev, sector_nr, skipped);
5583 /* No need to check resync_max as we never do more than one
5584 * stripe, and as resync_max will always be on a chunk boundary,
5585 * if the check in md_do_sync didn't fire, there is no chance
5586 * of overstepping resync_max here
5589 /* if there is too many failed drives and we are trying
5590 * to resync, then assert that we are finished, because there is
5591 * nothing we can do.
5593 if (mddev->degraded >= conf->max_degraded &&
5594 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5595 sector_t rv = mddev->dev_sectors - sector_nr;
5599 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5601 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5602 sync_blocks >= STRIPE_SECTORS) {
5603 /* we can skip this block, and probably more */
5604 sync_blocks /= STRIPE_SECTORS;
5606 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5609 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5611 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5613 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5614 /* make sure we don't swamp the stripe cache if someone else
5615 * is trying to get access
5617 schedule_timeout_uninterruptible(1);
5619 /* Need to check if array will still be degraded after recovery/resync
5620 * Note in case of > 1 drive failures it's possible we're rebuilding
5621 * one drive while leaving another faulty drive in array.
5624 for (i = 0; i < conf->raid_disks; i++) {
5625 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5627 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5632 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5634 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5635 set_bit(STRIPE_HANDLE, &sh->state);
5639 return STRIPE_SECTORS;
5642 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5644 /* We may not be able to submit a whole bio at once as there
5645 * may not be enough stripe_heads available.
5646 * We cannot pre-allocate enough stripe_heads as we may need
5647 * more than exist in the cache (if we allow ever large chunks).
5648 * So we do one stripe head at a time and record in
5649 * ->bi_hw_segments how many have been done.
5651 * We *know* that this entire raid_bio is in one chunk, so
5652 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5654 struct stripe_head *sh;
5656 sector_t sector, logical_sector, last_sector;
5661 logical_sector = raid_bio->bi_iter.bi_sector &
5662 ~((sector_t)STRIPE_SECTORS-1);
5663 sector = raid5_compute_sector(conf, logical_sector,
5665 last_sector = bio_end_sector(raid_bio);
5667 for (; logical_sector < last_sector;
5668 logical_sector += STRIPE_SECTORS,
5669 sector += STRIPE_SECTORS,
5672 if (scnt < raid5_bi_processed_stripes(raid_bio))
5673 /* already done this stripe */
5676 sh = get_active_stripe(conf, sector, 0, 1, 1);
5679 /* failed to get a stripe - must wait */
5680 raid5_set_bi_processed_stripes(raid_bio, scnt);
5681 conf->retry_read_aligned = raid_bio;
5685 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5687 raid5_set_bi_processed_stripes(raid_bio, scnt);
5688 conf->retry_read_aligned = raid_bio;
5692 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5697 remaining = raid5_dec_bi_active_stripes(raid_bio);
5698 if (remaining == 0) {
5699 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5701 bio_endio(raid_bio, 0);
5703 if (atomic_dec_and_test(&conf->active_aligned_reads))
5704 wake_up(&conf->wait_for_stripe);
5708 static int handle_active_stripes(struct r5conf *conf, int group,
5709 struct r5worker *worker,
5710 struct list_head *temp_inactive_list)
5712 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5713 int i, batch_size = 0, hash;
5714 bool release_inactive = false;
5716 while (batch_size < MAX_STRIPE_BATCH &&
5717 (sh = __get_priority_stripe(conf, group)) != NULL)
5718 batch[batch_size++] = sh;
5720 if (batch_size == 0) {
5721 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5722 if (!list_empty(temp_inactive_list + i))
5724 if (i == NR_STRIPE_HASH_LOCKS)
5726 release_inactive = true;
5728 spin_unlock_irq(&conf->device_lock);
5730 release_inactive_stripe_list(conf, temp_inactive_list,
5731 NR_STRIPE_HASH_LOCKS);
5733 if (release_inactive) {
5734 spin_lock_irq(&conf->device_lock);
5738 for (i = 0; i < batch_size; i++)
5739 handle_stripe(batch[i]);
5743 spin_lock_irq(&conf->device_lock);
5744 for (i = 0; i < batch_size; i++) {
5745 hash = batch[i]->hash_lock_index;
5746 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5751 static void raid5_do_work(struct work_struct *work)
5753 struct r5worker *worker = container_of(work, struct r5worker, work);
5754 struct r5worker_group *group = worker->group;
5755 struct r5conf *conf = group->conf;
5756 int group_id = group - conf->worker_groups;
5758 struct blk_plug plug;
5760 pr_debug("+++ raid5worker active\n");
5762 blk_start_plug(&plug);
5764 spin_lock_irq(&conf->device_lock);
5766 int batch_size, released;
5768 released = release_stripe_list(conf, worker->temp_inactive_list);
5770 batch_size = handle_active_stripes(conf, group_id, worker,
5771 worker->temp_inactive_list);
5772 worker->working = false;
5773 if (!batch_size && !released)
5775 handled += batch_size;
5777 pr_debug("%d stripes handled\n", handled);
5779 spin_unlock_irq(&conf->device_lock);
5780 blk_finish_plug(&plug);
5782 pr_debug("--- raid5worker inactive\n");
5786 * This is our raid5 kernel thread.
5788 * We scan the hash table for stripes which can be handled now.
5789 * During the scan, completed stripes are saved for us by the interrupt
5790 * handler, so that they will not have to wait for our next wakeup.
5792 static void raid5d(struct md_thread *thread)
5794 struct mddev *mddev = thread->mddev;
5795 struct r5conf *conf = mddev->private;
5797 struct blk_plug plug;
5799 pr_debug("+++ raid5d active\n");
5801 md_check_recovery(mddev);
5803 blk_start_plug(&plug);
5805 spin_lock_irq(&conf->device_lock);
5808 int batch_size, released;
5810 released = release_stripe_list(conf, conf->temp_inactive_list);
5812 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5815 !list_empty(&conf->bitmap_list)) {
5816 /* Now is a good time to flush some bitmap updates */
5818 spin_unlock_irq(&conf->device_lock);
5819 bitmap_unplug(mddev->bitmap);
5820 spin_lock_irq(&conf->device_lock);
5821 conf->seq_write = conf->seq_flush;
5822 activate_bit_delay(conf, conf->temp_inactive_list);
5824 raid5_activate_delayed(conf);
5826 while ((bio = remove_bio_from_retry(conf))) {
5828 spin_unlock_irq(&conf->device_lock);
5829 ok = retry_aligned_read(conf, bio);
5830 spin_lock_irq(&conf->device_lock);
5836 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5837 conf->temp_inactive_list);
5838 if (!batch_size && !released)
5840 handled += batch_size;
5842 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5843 spin_unlock_irq(&conf->device_lock);
5844 md_check_recovery(mddev);
5845 spin_lock_irq(&conf->device_lock);
5848 pr_debug("%d stripes handled\n", handled);
5850 spin_unlock_irq(&conf->device_lock);
5851 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state)) {
5852 grow_one_stripe(conf, __GFP_NOWARN);
5853 /* Set flag even if allocation failed. This helps
5854 * slow down allocation requests when mem is short
5856 set_bit(R5_DID_ALLOC, &conf->cache_state);
5859 async_tx_issue_pending_all();
5860 blk_finish_plug(&plug);
5862 pr_debug("--- raid5d inactive\n");
5866 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5868 struct r5conf *conf;
5870 spin_lock(&mddev->lock);
5871 conf = mddev->private;
5873 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5874 spin_unlock(&mddev->lock);
5879 raid5_set_cache_size(struct mddev *mddev, int size)
5881 struct r5conf *conf = mddev->private;
5884 if (size <= 16 || size > 32768)
5887 conf->min_nr_stripes = size;
5888 while (size < conf->max_nr_stripes &&
5889 drop_one_stripe(conf))
5893 err = md_allow_write(mddev);
5897 while (size > conf->max_nr_stripes)
5898 if (!grow_one_stripe(conf, GFP_KERNEL))
5903 EXPORT_SYMBOL(raid5_set_cache_size);
5906 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5908 struct r5conf *conf;
5912 if (len >= PAGE_SIZE)
5914 if (kstrtoul(page, 10, &new))
5916 err = mddev_lock(mddev);
5919 conf = mddev->private;
5923 err = raid5_set_cache_size(mddev, new);
5924 mddev_unlock(mddev);
5929 static struct md_sysfs_entry
5930 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5931 raid5_show_stripe_cache_size,
5932 raid5_store_stripe_cache_size);
5935 raid5_show_rmw_level(struct mddev *mddev, char *page)
5937 struct r5conf *conf = mddev->private;
5939 return sprintf(page, "%d\n", conf->rmw_level);
5945 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
5947 struct r5conf *conf = mddev->private;
5953 if (len >= PAGE_SIZE)
5956 if (kstrtoul(page, 10, &new))
5959 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
5962 if (new != PARITY_DISABLE_RMW &&
5963 new != PARITY_ENABLE_RMW &&
5964 new != PARITY_PREFER_RMW)
5967 conf->rmw_level = new;
5971 static struct md_sysfs_entry
5972 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
5973 raid5_show_rmw_level,
5974 raid5_store_rmw_level);
5978 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5980 struct r5conf *conf;
5982 spin_lock(&mddev->lock);
5983 conf = mddev->private;
5985 ret = sprintf(page, "%d\n", conf->bypass_threshold);
5986 spin_unlock(&mddev->lock);
5991 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
5993 struct r5conf *conf;
5997 if (len >= PAGE_SIZE)
5999 if (kstrtoul(page, 10, &new))
6002 err = mddev_lock(mddev);
6005 conf = mddev->private;
6008 else if (new > conf->min_nr_stripes)
6011 conf->bypass_threshold = new;
6012 mddev_unlock(mddev);
6016 static struct md_sysfs_entry
6017 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6019 raid5_show_preread_threshold,
6020 raid5_store_preread_threshold);
6023 raid5_show_skip_copy(struct mddev *mddev, char *page)
6025 struct r5conf *conf;
6027 spin_lock(&mddev->lock);
6028 conf = mddev->private;
6030 ret = sprintf(page, "%d\n", conf->skip_copy);
6031 spin_unlock(&mddev->lock);
6036 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6038 struct r5conf *conf;
6042 if (len >= PAGE_SIZE)
6044 if (kstrtoul(page, 10, &new))
6048 err = mddev_lock(mddev);
6051 conf = mddev->private;
6054 else if (new != conf->skip_copy) {
6055 mddev_suspend(mddev);
6056 conf->skip_copy = new;
6058 mddev->queue->backing_dev_info.capabilities |=
6059 BDI_CAP_STABLE_WRITES;
6061 mddev->queue->backing_dev_info.capabilities &=
6062 ~BDI_CAP_STABLE_WRITES;
6063 mddev_resume(mddev);
6065 mddev_unlock(mddev);
6069 static struct md_sysfs_entry
6070 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6071 raid5_show_skip_copy,
6072 raid5_store_skip_copy);
6075 stripe_cache_active_show(struct mddev *mddev, char *page)
6077 struct r5conf *conf = mddev->private;
6079 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6084 static struct md_sysfs_entry
6085 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6088 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6090 struct r5conf *conf;
6092 spin_lock(&mddev->lock);
6093 conf = mddev->private;
6095 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6096 spin_unlock(&mddev->lock);
6100 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6102 int *worker_cnt_per_group,
6103 struct r5worker_group **worker_groups);
6105 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6107 struct r5conf *conf;
6110 struct r5worker_group *new_groups, *old_groups;
6111 int group_cnt, worker_cnt_per_group;
6113 if (len >= PAGE_SIZE)
6115 if (kstrtoul(page, 10, &new))
6118 err = mddev_lock(mddev);
6121 conf = mddev->private;
6124 else if (new != conf->worker_cnt_per_group) {
6125 mddev_suspend(mddev);
6127 old_groups = conf->worker_groups;
6129 flush_workqueue(raid5_wq);
6131 err = alloc_thread_groups(conf, new,
6132 &group_cnt, &worker_cnt_per_group,
6135 spin_lock_irq(&conf->device_lock);
6136 conf->group_cnt = group_cnt;
6137 conf->worker_cnt_per_group = worker_cnt_per_group;
6138 conf->worker_groups = new_groups;
6139 spin_unlock_irq(&conf->device_lock);
6142 kfree(old_groups[0].workers);
6145 mddev_resume(mddev);
6147 mddev_unlock(mddev);
6152 static struct md_sysfs_entry
6153 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6154 raid5_show_group_thread_cnt,
6155 raid5_store_group_thread_cnt);
6157 static struct attribute *raid5_attrs[] = {
6158 &raid5_stripecache_size.attr,
6159 &raid5_stripecache_active.attr,
6160 &raid5_preread_bypass_threshold.attr,
6161 &raid5_group_thread_cnt.attr,
6162 &raid5_skip_copy.attr,
6163 &raid5_rmw_level.attr,
6166 static struct attribute_group raid5_attrs_group = {
6168 .attrs = raid5_attrs,
6171 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6173 int *worker_cnt_per_group,
6174 struct r5worker_group **worker_groups)
6178 struct r5worker *workers;
6180 *worker_cnt_per_group = cnt;
6183 *worker_groups = NULL;
6186 *group_cnt = num_possible_nodes();
6187 size = sizeof(struct r5worker) * cnt;
6188 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6189 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6190 *group_cnt, GFP_NOIO);
6191 if (!*worker_groups || !workers) {
6193 kfree(*worker_groups);
6197 for (i = 0; i < *group_cnt; i++) {
6198 struct r5worker_group *group;
6200 group = &(*worker_groups)[i];
6201 INIT_LIST_HEAD(&group->handle_list);
6203 group->workers = workers + i * cnt;
6205 for (j = 0; j < cnt; j++) {
6206 struct r5worker *worker = group->workers + j;
6207 worker->group = group;
6208 INIT_WORK(&worker->work, raid5_do_work);
6210 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6211 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6218 static void free_thread_groups(struct r5conf *conf)
6220 if (conf->worker_groups)
6221 kfree(conf->worker_groups[0].workers);
6222 kfree(conf->worker_groups);
6223 conf->worker_groups = NULL;
6227 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6229 struct r5conf *conf = mddev->private;
6232 sectors = mddev->dev_sectors;
6234 /* size is defined by the smallest of previous and new size */
6235 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6237 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6238 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
6239 return sectors * (raid_disks - conf->max_degraded);
6242 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6244 safe_put_page(percpu->spare_page);
6245 if (percpu->scribble)
6246 flex_array_free(percpu->scribble);
6247 percpu->spare_page = NULL;
6248 percpu->scribble = NULL;
6251 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6253 if (conf->level == 6 && !percpu->spare_page)
6254 percpu->spare_page = alloc_page(GFP_KERNEL);
6255 if (!percpu->scribble)
6256 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6257 conf->previous_raid_disks),
6258 max(conf->chunk_sectors,
6259 conf->prev_chunk_sectors)
6263 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6264 free_scratch_buffer(conf, percpu);
6271 static void raid5_free_percpu(struct r5conf *conf)
6278 #ifdef CONFIG_HOTPLUG_CPU
6279 unregister_cpu_notifier(&conf->cpu_notify);
6283 for_each_possible_cpu(cpu)
6284 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6287 free_percpu(conf->percpu);
6290 static void free_conf(struct r5conf *conf)
6292 if (conf->shrinker.seeks)
6293 unregister_shrinker(&conf->shrinker);
6294 free_thread_groups(conf);
6295 shrink_stripes(conf);
6296 raid5_free_percpu(conf);
6298 kfree(conf->stripe_hashtbl);
6302 #ifdef CONFIG_HOTPLUG_CPU
6303 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6306 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6307 long cpu = (long)hcpu;
6308 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6311 case CPU_UP_PREPARE:
6312 case CPU_UP_PREPARE_FROZEN:
6313 if (alloc_scratch_buffer(conf, percpu)) {
6314 pr_err("%s: failed memory allocation for cpu%ld\n",
6316 return notifier_from_errno(-ENOMEM);
6320 case CPU_DEAD_FROZEN:
6321 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6330 static int raid5_alloc_percpu(struct r5conf *conf)
6335 conf->percpu = alloc_percpu(struct raid5_percpu);
6339 #ifdef CONFIG_HOTPLUG_CPU
6340 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6341 conf->cpu_notify.priority = 0;
6342 err = register_cpu_notifier(&conf->cpu_notify);
6348 for_each_present_cpu(cpu) {
6349 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6351 pr_err("%s: failed memory allocation for cpu%ld\n",
6355 spin_lock_init(&per_cpu_ptr(conf->percpu, cpu)->lock);
6362 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6363 struct shrink_control *sc)
6365 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6367 while (ret < sc->nr_to_scan) {
6368 if (drop_one_stripe(conf) == 0)
6375 static unsigned long raid5_cache_count(struct shrinker *shrink,
6376 struct shrink_control *sc)
6378 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6380 if (conf->max_nr_stripes < conf->min_nr_stripes)
6381 /* unlikely, but not impossible */
6383 return conf->max_nr_stripes - conf->min_nr_stripes;
6386 static struct r5conf *setup_conf(struct mddev *mddev)
6388 struct r5conf *conf;
6389 int raid_disk, memory, max_disks;
6390 struct md_rdev *rdev;
6391 struct disk_info *disk;
6394 int group_cnt, worker_cnt_per_group;
6395 struct r5worker_group *new_group;
6397 if (mddev->new_level != 5
6398 && mddev->new_level != 4
6399 && mddev->new_level != 6) {
6400 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6401 mdname(mddev), mddev->new_level);
6402 return ERR_PTR(-EIO);
6404 if ((mddev->new_level == 5
6405 && !algorithm_valid_raid5(mddev->new_layout)) ||
6406 (mddev->new_level == 6
6407 && !algorithm_valid_raid6(mddev->new_layout))) {
6408 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6409 mdname(mddev), mddev->new_layout);
6410 return ERR_PTR(-EIO);
6412 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6413 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6414 mdname(mddev), mddev->raid_disks);
6415 return ERR_PTR(-EINVAL);
6418 if (!mddev->new_chunk_sectors ||
6419 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6420 !is_power_of_2(mddev->new_chunk_sectors)) {
6421 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6422 mdname(mddev), mddev->new_chunk_sectors << 9);
6423 return ERR_PTR(-EINVAL);
6426 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6429 /* Don't enable multi-threading by default*/
6430 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6432 conf->group_cnt = group_cnt;
6433 conf->worker_cnt_per_group = worker_cnt_per_group;
6434 conf->worker_groups = new_group;
6437 spin_lock_init(&conf->device_lock);
6438 seqcount_init(&conf->gen_lock);
6439 init_waitqueue_head(&conf->wait_for_stripe);
6440 init_waitqueue_head(&conf->wait_for_overlap);
6441 INIT_LIST_HEAD(&conf->handle_list);
6442 INIT_LIST_HEAD(&conf->hold_list);
6443 INIT_LIST_HEAD(&conf->delayed_list);
6444 INIT_LIST_HEAD(&conf->bitmap_list);
6445 init_llist_head(&conf->released_stripes);
6446 atomic_set(&conf->active_stripes, 0);
6447 atomic_set(&conf->preread_active_stripes, 0);
6448 atomic_set(&conf->active_aligned_reads, 0);
6449 conf->bypass_threshold = BYPASS_THRESHOLD;
6450 conf->recovery_disabled = mddev->recovery_disabled - 1;
6452 conf->raid_disks = mddev->raid_disks;
6453 if (mddev->reshape_position == MaxSector)
6454 conf->previous_raid_disks = mddev->raid_disks;
6456 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6457 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6459 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6464 conf->mddev = mddev;
6466 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6469 /* We init hash_locks[0] separately to that it can be used
6470 * as the reference lock in the spin_lock_nest_lock() call
6471 * in lock_all_device_hash_locks_irq in order to convince
6472 * lockdep that we know what we are doing.
6474 spin_lock_init(conf->hash_locks);
6475 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6476 spin_lock_init(conf->hash_locks + i);
6478 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6479 INIT_LIST_HEAD(conf->inactive_list + i);
6481 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6482 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6484 conf->level = mddev->new_level;
6485 conf->chunk_sectors = mddev->new_chunk_sectors;
6486 if (raid5_alloc_percpu(conf) != 0)
6489 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6491 rdev_for_each(rdev, mddev) {
6492 raid_disk = rdev->raid_disk;
6493 if (raid_disk >= max_disks
6496 disk = conf->disks + raid_disk;
6498 if (test_bit(Replacement, &rdev->flags)) {
6499 if (disk->replacement)
6501 disk->replacement = rdev;
6508 if (test_bit(In_sync, &rdev->flags)) {
6509 char b[BDEVNAME_SIZE];
6510 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6512 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6513 } else if (rdev->saved_raid_disk != raid_disk)
6514 /* Cannot rely on bitmap to complete recovery */
6518 conf->level = mddev->new_level;
6519 if (conf->level == 6) {
6520 conf->max_degraded = 2;
6521 if (raid6_call.xor_syndrome)
6522 conf->rmw_level = PARITY_ENABLE_RMW;
6524 conf->rmw_level = PARITY_DISABLE_RMW;
6526 conf->max_degraded = 1;
6527 conf->rmw_level = PARITY_ENABLE_RMW;
6529 conf->algorithm = mddev->new_layout;
6530 conf->reshape_progress = mddev->reshape_position;
6531 if (conf->reshape_progress != MaxSector) {
6532 conf->prev_chunk_sectors = mddev->chunk_sectors;
6533 conf->prev_algo = mddev->layout;
6536 conf->min_nr_stripes = NR_STRIPES;
6537 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6538 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6539 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6540 if (grow_stripes(conf, conf->min_nr_stripes)) {
6542 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6543 mdname(mddev), memory);
6546 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6547 mdname(mddev), memory);
6549 * Losing a stripe head costs more than the time to refill it,
6550 * it reduces the queue depth and so can hurt throughput.
6551 * So set it rather large, scaled by number of devices.
6553 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6554 conf->shrinker.scan_objects = raid5_cache_scan;
6555 conf->shrinker.count_objects = raid5_cache_count;
6556 conf->shrinker.batch = 128;
6557 conf->shrinker.flags = 0;
6558 register_shrinker(&conf->shrinker);
6560 sprintf(pers_name, "raid%d", mddev->new_level);
6561 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6562 if (!conf->thread) {
6564 "md/raid:%s: couldn't allocate thread.\n",
6574 return ERR_PTR(-EIO);
6576 return ERR_PTR(-ENOMEM);
6579 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6582 case ALGORITHM_PARITY_0:
6583 if (raid_disk < max_degraded)
6586 case ALGORITHM_PARITY_N:
6587 if (raid_disk >= raid_disks - max_degraded)
6590 case ALGORITHM_PARITY_0_6:
6591 if (raid_disk == 0 ||
6592 raid_disk == raid_disks - 1)
6595 case ALGORITHM_LEFT_ASYMMETRIC_6:
6596 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6597 case ALGORITHM_LEFT_SYMMETRIC_6:
6598 case ALGORITHM_RIGHT_SYMMETRIC_6:
6599 if (raid_disk == raid_disks - 1)
6605 static int run(struct mddev *mddev)
6607 struct r5conf *conf;
6608 int working_disks = 0;
6609 int dirty_parity_disks = 0;
6610 struct md_rdev *rdev;
6611 sector_t reshape_offset = 0;
6613 long long min_offset_diff = 0;
6616 if (mddev->recovery_cp != MaxSector)
6617 printk(KERN_NOTICE "md/raid:%s: not clean"
6618 " -- starting background reconstruction\n",
6621 rdev_for_each(rdev, mddev) {
6623 if (rdev->raid_disk < 0)
6625 diff = (rdev->new_data_offset - rdev->data_offset);
6627 min_offset_diff = diff;
6629 } else if (mddev->reshape_backwards &&
6630 diff < min_offset_diff)
6631 min_offset_diff = diff;
6632 else if (!mddev->reshape_backwards &&
6633 diff > min_offset_diff)
6634 min_offset_diff = diff;
6637 if (mddev->reshape_position != MaxSector) {
6638 /* Check that we can continue the reshape.
6639 * Difficulties arise if the stripe we would write to
6640 * next is at or after the stripe we would read from next.
6641 * For a reshape that changes the number of devices, this
6642 * is only possible for a very short time, and mdadm makes
6643 * sure that time appears to have past before assembling
6644 * the array. So we fail if that time hasn't passed.
6645 * For a reshape that keeps the number of devices the same
6646 * mdadm must be monitoring the reshape can keeping the
6647 * critical areas read-only and backed up. It will start
6648 * the array in read-only mode, so we check for that.
6650 sector_t here_new, here_old;
6652 int max_degraded = (mddev->level == 6 ? 2 : 1);
6654 if (mddev->new_level != mddev->level) {
6655 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6656 "required - aborting.\n",
6660 old_disks = mddev->raid_disks - mddev->delta_disks;
6661 /* reshape_position must be on a new-stripe boundary, and one
6662 * further up in new geometry must map after here in old
6665 here_new = mddev->reshape_position;
6666 if (sector_div(here_new, mddev->new_chunk_sectors *
6667 (mddev->raid_disks - max_degraded))) {
6668 printk(KERN_ERR "md/raid:%s: reshape_position not "
6669 "on a stripe boundary\n", mdname(mddev));
6672 reshape_offset = here_new * mddev->new_chunk_sectors;
6673 /* here_new is the stripe we will write to */
6674 here_old = mddev->reshape_position;
6675 sector_div(here_old, mddev->chunk_sectors *
6676 (old_disks-max_degraded));
6677 /* here_old is the first stripe that we might need to read
6679 if (mddev->delta_disks == 0) {
6680 if ((here_new * mddev->new_chunk_sectors !=
6681 here_old * mddev->chunk_sectors)) {
6682 printk(KERN_ERR "md/raid:%s: reshape position is"
6683 " confused - aborting\n", mdname(mddev));
6686 /* We cannot be sure it is safe to start an in-place
6687 * reshape. It is only safe if user-space is monitoring
6688 * and taking constant backups.
6689 * mdadm always starts a situation like this in
6690 * readonly mode so it can take control before
6691 * allowing any writes. So just check for that.
6693 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6694 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6695 /* not really in-place - so OK */;
6696 else if (mddev->ro == 0) {
6697 printk(KERN_ERR "md/raid:%s: in-place reshape "
6698 "must be started in read-only mode "
6703 } else if (mddev->reshape_backwards
6704 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6705 here_old * mddev->chunk_sectors)
6706 : (here_new * mddev->new_chunk_sectors >=
6707 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6708 /* Reading from the same stripe as writing to - bad */
6709 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6710 "auto-recovery - aborting.\n",
6714 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6716 /* OK, we should be able to continue; */
6718 BUG_ON(mddev->level != mddev->new_level);
6719 BUG_ON(mddev->layout != mddev->new_layout);
6720 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6721 BUG_ON(mddev->delta_disks != 0);
6724 if (mddev->private == NULL)
6725 conf = setup_conf(mddev);
6727 conf = mddev->private;
6730 return PTR_ERR(conf);
6732 conf->min_offset_diff = min_offset_diff;
6733 mddev->thread = conf->thread;
6734 conf->thread = NULL;
6735 mddev->private = conf;
6737 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6739 rdev = conf->disks[i].rdev;
6740 if (!rdev && conf->disks[i].replacement) {
6741 /* The replacement is all we have yet */
6742 rdev = conf->disks[i].replacement;
6743 conf->disks[i].replacement = NULL;
6744 clear_bit(Replacement, &rdev->flags);
6745 conf->disks[i].rdev = rdev;
6749 if (conf->disks[i].replacement &&
6750 conf->reshape_progress != MaxSector) {
6751 /* replacements and reshape simply do not mix. */
6752 printk(KERN_ERR "md: cannot handle concurrent "
6753 "replacement and reshape.\n");
6756 if (test_bit(In_sync, &rdev->flags)) {
6760 /* This disc is not fully in-sync. However if it
6761 * just stored parity (beyond the recovery_offset),
6762 * when we don't need to be concerned about the
6763 * array being dirty.
6764 * When reshape goes 'backwards', we never have
6765 * partially completed devices, so we only need
6766 * to worry about reshape going forwards.
6768 /* Hack because v0.91 doesn't store recovery_offset properly. */
6769 if (mddev->major_version == 0 &&
6770 mddev->minor_version > 90)
6771 rdev->recovery_offset = reshape_offset;
6773 if (rdev->recovery_offset < reshape_offset) {
6774 /* We need to check old and new layout */
6775 if (!only_parity(rdev->raid_disk,
6778 conf->max_degraded))
6781 if (!only_parity(rdev->raid_disk,
6783 conf->previous_raid_disks,
6784 conf->max_degraded))
6786 dirty_parity_disks++;
6790 * 0 for a fully functional array, 1 or 2 for a degraded array.
6792 mddev->degraded = calc_degraded(conf);
6794 if (has_failed(conf)) {
6795 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6796 " (%d/%d failed)\n",
6797 mdname(mddev), mddev->degraded, conf->raid_disks);
6801 /* device size must be a multiple of chunk size */
6802 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6803 mddev->resync_max_sectors = mddev->dev_sectors;
6805 if (mddev->degraded > dirty_parity_disks &&
6806 mddev->recovery_cp != MaxSector) {
6807 if (mddev->ok_start_degraded)
6809 "md/raid:%s: starting dirty degraded array"
6810 " - data corruption possible.\n",
6814 "md/raid:%s: cannot start dirty degraded array.\n",
6820 if (mddev->degraded == 0)
6821 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6822 " devices, algorithm %d\n", mdname(mddev), conf->level,
6823 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6826 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6827 " out of %d devices, algorithm %d\n",
6828 mdname(mddev), conf->level,
6829 mddev->raid_disks - mddev->degraded,
6830 mddev->raid_disks, mddev->new_layout);
6832 print_raid5_conf(conf);
6834 if (conf->reshape_progress != MaxSector) {
6835 conf->reshape_safe = conf->reshape_progress;
6836 atomic_set(&conf->reshape_stripes, 0);
6837 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6838 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6839 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6840 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6841 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6845 /* Ok, everything is just fine now */
6846 if (mddev->to_remove == &raid5_attrs_group)
6847 mddev->to_remove = NULL;
6848 else if (mddev->kobj.sd &&
6849 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6851 "raid5: failed to create sysfs attributes for %s\n",
6853 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6857 bool discard_supported = true;
6858 /* read-ahead size must cover two whole stripes, which
6859 * is 2 * (datadisks) * chunksize where 'n' is the
6860 * number of raid devices
6862 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6863 int stripe = data_disks *
6864 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6865 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6866 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6868 chunk_size = mddev->chunk_sectors << 9;
6869 blk_queue_io_min(mddev->queue, chunk_size);
6870 blk_queue_io_opt(mddev->queue, chunk_size *
6871 (conf->raid_disks - conf->max_degraded));
6872 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6874 * We can only discard a whole stripe. It doesn't make sense to
6875 * discard data disk but write parity disk
6877 stripe = stripe * PAGE_SIZE;
6878 /* Round up to power of 2, as discard handling
6879 * currently assumes that */
6880 while ((stripe-1) & stripe)
6881 stripe = (stripe | (stripe-1)) + 1;
6882 mddev->queue->limits.discard_alignment = stripe;
6883 mddev->queue->limits.discard_granularity = stripe;
6885 * unaligned part of discard request will be ignored, so can't
6886 * guarantee discard_zeroes_data
6888 mddev->queue->limits.discard_zeroes_data = 0;
6890 blk_queue_max_write_same_sectors(mddev->queue, 0);
6892 rdev_for_each(rdev, mddev) {
6893 disk_stack_limits(mddev->gendisk, rdev->bdev,
6894 rdev->data_offset << 9);
6895 disk_stack_limits(mddev->gendisk, rdev->bdev,
6896 rdev->new_data_offset << 9);
6898 * discard_zeroes_data is required, otherwise data
6899 * could be lost. Consider a scenario: discard a stripe
6900 * (the stripe could be inconsistent if
6901 * discard_zeroes_data is 0); write one disk of the
6902 * stripe (the stripe could be inconsistent again
6903 * depending on which disks are used to calculate
6904 * parity); the disk is broken; The stripe data of this
6907 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6908 !bdev_get_queue(rdev->bdev)->
6909 limits.discard_zeroes_data)
6910 discard_supported = false;
6911 /* Unfortunately, discard_zeroes_data is not currently
6912 * a guarantee - just a hint. So we only allow DISCARD
6913 * if the sysadmin has confirmed that only safe devices
6914 * are in use by setting a module parameter.
6916 if (!devices_handle_discard_safely) {
6917 if (discard_supported) {
6918 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6919 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6921 discard_supported = false;
6925 if (discard_supported &&
6926 mddev->queue->limits.max_discard_sectors >= stripe &&
6927 mddev->queue->limits.discard_granularity >= stripe)
6928 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6931 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6937 md_unregister_thread(&mddev->thread);
6938 print_raid5_conf(conf);
6940 mddev->private = NULL;
6941 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6945 static void raid5_free(struct mddev *mddev, void *priv)
6947 struct r5conf *conf = priv;
6950 mddev->to_remove = &raid5_attrs_group;
6953 static void status(struct seq_file *seq, struct mddev *mddev)
6955 struct r5conf *conf = mddev->private;
6958 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6959 mddev->chunk_sectors / 2, mddev->layout);
6960 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6961 for (i = 0; i < conf->raid_disks; i++)
6962 seq_printf (seq, "%s",
6963 conf->disks[i].rdev &&
6964 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6965 seq_printf (seq, "]");
6968 static void print_raid5_conf (struct r5conf *conf)
6971 struct disk_info *tmp;
6973 printk(KERN_DEBUG "RAID conf printout:\n");
6975 printk("(conf==NULL)\n");
6978 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6980 conf->raid_disks - conf->mddev->degraded);
6982 for (i = 0; i < conf->raid_disks; i++) {
6983 char b[BDEVNAME_SIZE];
6984 tmp = conf->disks + i;
6986 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6987 i, !test_bit(Faulty, &tmp->rdev->flags),
6988 bdevname(tmp->rdev->bdev, b));
6992 static int raid5_spare_active(struct mddev *mddev)
6995 struct r5conf *conf = mddev->private;
6996 struct disk_info *tmp;
6998 unsigned long flags;
7000 for (i = 0; i < conf->raid_disks; i++) {
7001 tmp = conf->disks + i;
7002 if (tmp->replacement
7003 && tmp->replacement->recovery_offset == MaxSector
7004 && !test_bit(Faulty, &tmp->replacement->flags)
7005 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7006 /* Replacement has just become active. */
7008 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7011 /* Replaced device not technically faulty,
7012 * but we need to be sure it gets removed
7013 * and never re-added.
7015 set_bit(Faulty, &tmp->rdev->flags);
7016 sysfs_notify_dirent_safe(
7017 tmp->rdev->sysfs_state);
7019 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7020 } else if (tmp->rdev
7021 && tmp->rdev->recovery_offset == MaxSector
7022 && !test_bit(Faulty, &tmp->rdev->flags)
7023 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7025 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7028 spin_lock_irqsave(&conf->device_lock, flags);
7029 mddev->degraded = calc_degraded(conf);
7030 spin_unlock_irqrestore(&conf->device_lock, flags);
7031 print_raid5_conf(conf);
7035 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7037 struct r5conf *conf = mddev->private;
7039 int number = rdev->raid_disk;
7040 struct md_rdev **rdevp;
7041 struct disk_info *p = conf->disks + number;
7043 print_raid5_conf(conf);
7044 if (rdev == p->rdev)
7046 else if (rdev == p->replacement)
7047 rdevp = &p->replacement;
7051 if (number >= conf->raid_disks &&
7052 conf->reshape_progress == MaxSector)
7053 clear_bit(In_sync, &rdev->flags);
7055 if (test_bit(In_sync, &rdev->flags) ||
7056 atomic_read(&rdev->nr_pending)) {
7060 /* Only remove non-faulty devices if recovery
7063 if (!test_bit(Faulty, &rdev->flags) &&
7064 mddev->recovery_disabled != conf->recovery_disabled &&
7065 !has_failed(conf) &&
7066 (!p->replacement || p->replacement == rdev) &&
7067 number < conf->raid_disks) {
7073 if (atomic_read(&rdev->nr_pending)) {
7074 /* lost the race, try later */
7077 } else if (p->replacement) {
7078 /* We must have just cleared 'rdev' */
7079 p->rdev = p->replacement;
7080 clear_bit(Replacement, &p->replacement->flags);
7081 smp_mb(); /* Make sure other CPUs may see both as identical
7082 * but will never see neither - if they are careful
7084 p->replacement = NULL;
7085 clear_bit(WantReplacement, &rdev->flags);
7087 /* We might have just removed the Replacement as faulty-
7088 * clear the bit just in case
7090 clear_bit(WantReplacement, &rdev->flags);
7093 print_raid5_conf(conf);
7097 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7099 struct r5conf *conf = mddev->private;
7102 struct disk_info *p;
7104 int last = conf->raid_disks - 1;
7106 if (mddev->recovery_disabled == conf->recovery_disabled)
7109 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7110 /* no point adding a device */
7113 if (rdev->raid_disk >= 0)
7114 first = last = rdev->raid_disk;
7117 * find the disk ... but prefer rdev->saved_raid_disk
7120 if (rdev->saved_raid_disk >= 0 &&
7121 rdev->saved_raid_disk >= first &&
7122 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7123 first = rdev->saved_raid_disk;
7125 for (disk = first; disk <= last; disk++) {
7126 p = conf->disks + disk;
7127 if (p->rdev == NULL) {
7128 clear_bit(In_sync, &rdev->flags);
7129 rdev->raid_disk = disk;
7131 if (rdev->saved_raid_disk != disk)
7133 rcu_assign_pointer(p->rdev, rdev);
7137 for (disk = first; disk <= last; disk++) {
7138 p = conf->disks + disk;
7139 if (test_bit(WantReplacement, &p->rdev->flags) &&
7140 p->replacement == NULL) {
7141 clear_bit(In_sync, &rdev->flags);
7142 set_bit(Replacement, &rdev->flags);
7143 rdev->raid_disk = disk;
7146 rcu_assign_pointer(p->replacement, rdev);
7151 print_raid5_conf(conf);
7155 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7157 /* no resync is happening, and there is enough space
7158 * on all devices, so we can resize.
7159 * We need to make sure resync covers any new space.
7160 * If the array is shrinking we should possibly wait until
7161 * any io in the removed space completes, but it hardly seems
7165 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7166 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7167 if (mddev->external_size &&
7168 mddev->array_sectors > newsize)
7170 if (mddev->bitmap) {
7171 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7175 md_set_array_sectors(mddev, newsize);
7176 set_capacity(mddev->gendisk, mddev->array_sectors);
7177 revalidate_disk(mddev->gendisk);
7178 if (sectors > mddev->dev_sectors &&
7179 mddev->recovery_cp > mddev->dev_sectors) {
7180 mddev->recovery_cp = mddev->dev_sectors;
7181 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7183 mddev->dev_sectors = sectors;
7184 mddev->resync_max_sectors = sectors;
7188 static int check_stripe_cache(struct mddev *mddev)
7190 /* Can only proceed if there are plenty of stripe_heads.
7191 * We need a minimum of one full stripe,, and for sensible progress
7192 * it is best to have about 4 times that.
7193 * If we require 4 times, then the default 256 4K stripe_heads will
7194 * allow for chunk sizes up to 256K, which is probably OK.
7195 * If the chunk size is greater, user-space should request more
7196 * stripe_heads first.
7198 struct r5conf *conf = mddev->private;
7199 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7200 > conf->min_nr_stripes ||
7201 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7202 > conf->min_nr_stripes) {
7203 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7205 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7212 static int check_reshape(struct mddev *mddev)
7214 struct r5conf *conf = mddev->private;
7216 if (mddev->delta_disks == 0 &&
7217 mddev->new_layout == mddev->layout &&
7218 mddev->new_chunk_sectors == mddev->chunk_sectors)
7219 return 0; /* nothing to do */
7220 if (has_failed(conf))
7222 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7223 /* We might be able to shrink, but the devices must
7224 * be made bigger first.
7225 * For raid6, 4 is the minimum size.
7226 * Otherwise 2 is the minimum
7229 if (mddev->level == 6)
7231 if (mddev->raid_disks + mddev->delta_disks < min)
7235 if (!check_stripe_cache(mddev))
7238 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7239 mddev->delta_disks > 0)
7240 if (resize_chunks(conf,
7241 conf->previous_raid_disks
7242 + max(0, mddev->delta_disks),
7243 max(mddev->new_chunk_sectors,
7244 mddev->chunk_sectors)
7247 return resize_stripes(conf, (conf->previous_raid_disks
7248 + mddev->delta_disks));
7251 static int raid5_start_reshape(struct mddev *mddev)
7253 struct r5conf *conf = mddev->private;
7254 struct md_rdev *rdev;
7256 unsigned long flags;
7258 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7261 if (!check_stripe_cache(mddev))
7264 if (has_failed(conf))
7267 rdev_for_each(rdev, mddev) {
7268 if (!test_bit(In_sync, &rdev->flags)
7269 && !test_bit(Faulty, &rdev->flags))
7273 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7274 /* Not enough devices even to make a degraded array
7279 /* Refuse to reduce size of the array. Any reductions in
7280 * array size must be through explicit setting of array_size
7283 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7284 < mddev->array_sectors) {
7285 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7286 "before number of disks\n", mdname(mddev));
7290 atomic_set(&conf->reshape_stripes, 0);
7291 spin_lock_irq(&conf->device_lock);
7292 write_seqcount_begin(&conf->gen_lock);
7293 conf->previous_raid_disks = conf->raid_disks;
7294 conf->raid_disks += mddev->delta_disks;
7295 conf->prev_chunk_sectors = conf->chunk_sectors;
7296 conf->chunk_sectors = mddev->new_chunk_sectors;
7297 conf->prev_algo = conf->algorithm;
7298 conf->algorithm = mddev->new_layout;
7300 /* Code that selects data_offset needs to see the generation update
7301 * if reshape_progress has been set - so a memory barrier needed.
7304 if (mddev->reshape_backwards)
7305 conf->reshape_progress = raid5_size(mddev, 0, 0);
7307 conf->reshape_progress = 0;
7308 conf->reshape_safe = conf->reshape_progress;
7309 write_seqcount_end(&conf->gen_lock);
7310 spin_unlock_irq(&conf->device_lock);
7312 /* Now make sure any requests that proceeded on the assumption
7313 * the reshape wasn't running - like Discard or Read - have
7316 mddev_suspend(mddev);
7317 mddev_resume(mddev);
7319 /* Add some new drives, as many as will fit.
7320 * We know there are enough to make the newly sized array work.
7321 * Don't add devices if we are reducing the number of
7322 * devices in the array. This is because it is not possible
7323 * to correctly record the "partially reconstructed" state of
7324 * such devices during the reshape and confusion could result.
7326 if (mddev->delta_disks >= 0) {
7327 rdev_for_each(rdev, mddev)
7328 if (rdev->raid_disk < 0 &&
7329 !test_bit(Faulty, &rdev->flags)) {
7330 if (raid5_add_disk(mddev, rdev) == 0) {
7332 >= conf->previous_raid_disks)
7333 set_bit(In_sync, &rdev->flags);
7335 rdev->recovery_offset = 0;
7337 if (sysfs_link_rdev(mddev, rdev))
7338 /* Failure here is OK */;
7340 } else if (rdev->raid_disk >= conf->previous_raid_disks
7341 && !test_bit(Faulty, &rdev->flags)) {
7342 /* This is a spare that was manually added */
7343 set_bit(In_sync, &rdev->flags);
7346 /* When a reshape changes the number of devices,
7347 * ->degraded is measured against the larger of the
7348 * pre and post number of devices.
7350 spin_lock_irqsave(&conf->device_lock, flags);
7351 mddev->degraded = calc_degraded(conf);
7352 spin_unlock_irqrestore(&conf->device_lock, flags);
7354 mddev->raid_disks = conf->raid_disks;
7355 mddev->reshape_position = conf->reshape_progress;
7356 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7358 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7359 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7360 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7361 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7362 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7363 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7365 if (!mddev->sync_thread) {
7366 mddev->recovery = 0;
7367 spin_lock_irq(&conf->device_lock);
7368 write_seqcount_begin(&conf->gen_lock);
7369 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7370 mddev->new_chunk_sectors =
7371 conf->chunk_sectors = conf->prev_chunk_sectors;
7372 mddev->new_layout = conf->algorithm = conf->prev_algo;
7373 rdev_for_each(rdev, mddev)
7374 rdev->new_data_offset = rdev->data_offset;
7376 conf->generation --;
7377 conf->reshape_progress = MaxSector;
7378 mddev->reshape_position = MaxSector;
7379 write_seqcount_end(&conf->gen_lock);
7380 spin_unlock_irq(&conf->device_lock);
7383 conf->reshape_checkpoint = jiffies;
7384 md_wakeup_thread(mddev->sync_thread);
7385 md_new_event(mddev);
7389 /* This is called from the reshape thread and should make any
7390 * changes needed in 'conf'
7392 static void end_reshape(struct r5conf *conf)
7395 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7396 struct md_rdev *rdev;
7398 spin_lock_irq(&conf->device_lock);
7399 conf->previous_raid_disks = conf->raid_disks;
7400 rdev_for_each(rdev, conf->mddev)
7401 rdev->data_offset = rdev->new_data_offset;
7403 conf->reshape_progress = MaxSector;
7404 spin_unlock_irq(&conf->device_lock);
7405 wake_up(&conf->wait_for_overlap);
7407 /* read-ahead size must cover two whole stripes, which is
7408 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7410 if (conf->mddev->queue) {
7411 int data_disks = conf->raid_disks - conf->max_degraded;
7412 int stripe = data_disks * ((conf->chunk_sectors << 9)
7414 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7415 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7420 /* This is called from the raid5d thread with mddev_lock held.
7421 * It makes config changes to the device.
7423 static void raid5_finish_reshape(struct mddev *mddev)
7425 struct r5conf *conf = mddev->private;
7427 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7429 if (mddev->delta_disks > 0) {
7430 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7431 set_capacity(mddev->gendisk, mddev->array_sectors);
7432 revalidate_disk(mddev->gendisk);
7435 spin_lock_irq(&conf->device_lock);
7436 mddev->degraded = calc_degraded(conf);
7437 spin_unlock_irq(&conf->device_lock);
7438 for (d = conf->raid_disks ;
7439 d < conf->raid_disks - mddev->delta_disks;
7441 struct md_rdev *rdev = conf->disks[d].rdev;
7443 clear_bit(In_sync, &rdev->flags);
7444 rdev = conf->disks[d].replacement;
7446 clear_bit(In_sync, &rdev->flags);
7449 mddev->layout = conf->algorithm;
7450 mddev->chunk_sectors = conf->chunk_sectors;
7451 mddev->reshape_position = MaxSector;
7452 mddev->delta_disks = 0;
7453 mddev->reshape_backwards = 0;
7457 static void raid5_quiesce(struct mddev *mddev, int state)
7459 struct r5conf *conf = mddev->private;
7462 case 2: /* resume for a suspend */
7463 wake_up(&conf->wait_for_overlap);
7466 case 1: /* stop all writes */
7467 lock_all_device_hash_locks_irq(conf);
7468 /* '2' tells resync/reshape to pause so that all
7469 * active stripes can drain
7472 wait_event_cmd(conf->wait_for_stripe,
7473 atomic_read(&conf->active_stripes) == 0 &&
7474 atomic_read(&conf->active_aligned_reads) == 0,
7475 unlock_all_device_hash_locks_irq(conf),
7476 lock_all_device_hash_locks_irq(conf));
7478 unlock_all_device_hash_locks_irq(conf);
7479 /* allow reshape to continue */
7480 wake_up(&conf->wait_for_overlap);
7483 case 0: /* re-enable writes */
7484 lock_all_device_hash_locks_irq(conf);
7486 wake_up(&conf->wait_for_stripe);
7487 wake_up(&conf->wait_for_overlap);
7488 unlock_all_device_hash_locks_irq(conf);
7493 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7495 struct r0conf *raid0_conf = mddev->private;
7498 /* for raid0 takeover only one zone is supported */
7499 if (raid0_conf->nr_strip_zones > 1) {
7500 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7502 return ERR_PTR(-EINVAL);
7505 sectors = raid0_conf->strip_zone[0].zone_end;
7506 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7507 mddev->dev_sectors = sectors;
7508 mddev->new_level = level;
7509 mddev->new_layout = ALGORITHM_PARITY_N;
7510 mddev->new_chunk_sectors = mddev->chunk_sectors;
7511 mddev->raid_disks += 1;
7512 mddev->delta_disks = 1;
7513 /* make sure it will be not marked as dirty */
7514 mddev->recovery_cp = MaxSector;
7516 return setup_conf(mddev);
7519 static void *raid5_takeover_raid1(struct mddev *mddev)
7523 if (mddev->raid_disks != 2 ||
7524 mddev->degraded > 1)
7525 return ERR_PTR(-EINVAL);
7527 /* Should check if there are write-behind devices? */
7529 chunksect = 64*2; /* 64K by default */
7531 /* The array must be an exact multiple of chunksize */
7532 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7535 if ((chunksect<<9) < STRIPE_SIZE)
7536 /* array size does not allow a suitable chunk size */
7537 return ERR_PTR(-EINVAL);
7539 mddev->new_level = 5;
7540 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7541 mddev->new_chunk_sectors = chunksect;
7543 return setup_conf(mddev);
7546 static void *raid5_takeover_raid6(struct mddev *mddev)
7550 switch (mddev->layout) {
7551 case ALGORITHM_LEFT_ASYMMETRIC_6:
7552 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7554 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7555 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7557 case ALGORITHM_LEFT_SYMMETRIC_6:
7558 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7560 case ALGORITHM_RIGHT_SYMMETRIC_6:
7561 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7563 case ALGORITHM_PARITY_0_6:
7564 new_layout = ALGORITHM_PARITY_0;
7566 case ALGORITHM_PARITY_N:
7567 new_layout = ALGORITHM_PARITY_N;
7570 return ERR_PTR(-EINVAL);
7572 mddev->new_level = 5;
7573 mddev->new_layout = new_layout;
7574 mddev->delta_disks = -1;
7575 mddev->raid_disks -= 1;
7576 return setup_conf(mddev);
7579 static int raid5_check_reshape(struct mddev *mddev)
7581 /* For a 2-drive array, the layout and chunk size can be changed
7582 * immediately as not restriping is needed.
7583 * For larger arrays we record the new value - after validation
7584 * to be used by a reshape pass.
7586 struct r5conf *conf = mddev->private;
7587 int new_chunk = mddev->new_chunk_sectors;
7589 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7591 if (new_chunk > 0) {
7592 if (!is_power_of_2(new_chunk))
7594 if (new_chunk < (PAGE_SIZE>>9))
7596 if (mddev->array_sectors & (new_chunk-1))
7597 /* not factor of array size */
7601 /* They look valid */
7603 if (mddev->raid_disks == 2) {
7604 /* can make the change immediately */
7605 if (mddev->new_layout >= 0) {
7606 conf->algorithm = mddev->new_layout;
7607 mddev->layout = mddev->new_layout;
7609 if (new_chunk > 0) {
7610 conf->chunk_sectors = new_chunk ;
7611 mddev->chunk_sectors = new_chunk;
7613 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7614 md_wakeup_thread(mddev->thread);
7616 return check_reshape(mddev);
7619 static int raid6_check_reshape(struct mddev *mddev)
7621 int new_chunk = mddev->new_chunk_sectors;
7623 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7625 if (new_chunk > 0) {
7626 if (!is_power_of_2(new_chunk))
7628 if (new_chunk < (PAGE_SIZE >> 9))
7630 if (mddev->array_sectors & (new_chunk-1))
7631 /* not factor of array size */
7635 /* They look valid */
7636 return check_reshape(mddev);
7639 static void *raid5_takeover(struct mddev *mddev)
7641 /* raid5 can take over:
7642 * raid0 - if there is only one strip zone - make it a raid4 layout
7643 * raid1 - if there are two drives. We need to know the chunk size
7644 * raid4 - trivial - just use a raid4 layout.
7645 * raid6 - Providing it is a *_6 layout
7647 if (mddev->level == 0)
7648 return raid45_takeover_raid0(mddev, 5);
7649 if (mddev->level == 1)
7650 return raid5_takeover_raid1(mddev);
7651 if (mddev->level == 4) {
7652 mddev->new_layout = ALGORITHM_PARITY_N;
7653 mddev->new_level = 5;
7654 return setup_conf(mddev);
7656 if (mddev->level == 6)
7657 return raid5_takeover_raid6(mddev);
7659 return ERR_PTR(-EINVAL);
7662 static void *raid4_takeover(struct mddev *mddev)
7664 /* raid4 can take over:
7665 * raid0 - if there is only one strip zone
7666 * raid5 - if layout is right
7668 if (mddev->level == 0)
7669 return raid45_takeover_raid0(mddev, 4);
7670 if (mddev->level == 5 &&
7671 mddev->layout == ALGORITHM_PARITY_N) {
7672 mddev->new_layout = 0;
7673 mddev->new_level = 4;
7674 return setup_conf(mddev);
7676 return ERR_PTR(-EINVAL);
7679 static struct md_personality raid5_personality;
7681 static void *raid6_takeover(struct mddev *mddev)
7683 /* Currently can only take over a raid5. We map the
7684 * personality to an equivalent raid6 personality
7685 * with the Q block at the end.
7689 if (mddev->pers != &raid5_personality)
7690 return ERR_PTR(-EINVAL);
7691 if (mddev->degraded > 1)
7692 return ERR_PTR(-EINVAL);
7693 if (mddev->raid_disks > 253)
7694 return ERR_PTR(-EINVAL);
7695 if (mddev->raid_disks < 3)
7696 return ERR_PTR(-EINVAL);
7698 switch (mddev->layout) {
7699 case ALGORITHM_LEFT_ASYMMETRIC:
7700 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7702 case ALGORITHM_RIGHT_ASYMMETRIC:
7703 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7705 case ALGORITHM_LEFT_SYMMETRIC:
7706 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7708 case ALGORITHM_RIGHT_SYMMETRIC:
7709 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7711 case ALGORITHM_PARITY_0:
7712 new_layout = ALGORITHM_PARITY_0_6;
7714 case ALGORITHM_PARITY_N:
7715 new_layout = ALGORITHM_PARITY_N;
7718 return ERR_PTR(-EINVAL);
7720 mddev->new_level = 6;
7721 mddev->new_layout = new_layout;
7722 mddev->delta_disks = 1;
7723 mddev->raid_disks += 1;
7724 return setup_conf(mddev);
7727 static struct md_personality raid6_personality =
7731 .owner = THIS_MODULE,
7732 .make_request = make_request,
7736 .error_handler = error,
7737 .hot_add_disk = raid5_add_disk,
7738 .hot_remove_disk= raid5_remove_disk,
7739 .spare_active = raid5_spare_active,
7740 .sync_request = sync_request,
7741 .resize = raid5_resize,
7743 .check_reshape = raid6_check_reshape,
7744 .start_reshape = raid5_start_reshape,
7745 .finish_reshape = raid5_finish_reshape,
7746 .quiesce = raid5_quiesce,
7747 .takeover = raid6_takeover,
7748 .congested = raid5_congested,
7749 .mergeable_bvec = raid5_mergeable_bvec,
7751 static struct md_personality raid5_personality =
7755 .owner = THIS_MODULE,
7756 .make_request = make_request,
7760 .error_handler = error,
7761 .hot_add_disk = raid5_add_disk,
7762 .hot_remove_disk= raid5_remove_disk,
7763 .spare_active = raid5_spare_active,
7764 .sync_request = sync_request,
7765 .resize = raid5_resize,
7767 .check_reshape = raid5_check_reshape,
7768 .start_reshape = raid5_start_reshape,
7769 .finish_reshape = raid5_finish_reshape,
7770 .quiesce = raid5_quiesce,
7771 .takeover = raid5_takeover,
7772 .congested = raid5_congested,
7773 .mergeable_bvec = raid5_mergeable_bvec,
7776 static struct md_personality raid4_personality =
7780 .owner = THIS_MODULE,
7781 .make_request = make_request,
7785 .error_handler = error,
7786 .hot_add_disk = raid5_add_disk,
7787 .hot_remove_disk= raid5_remove_disk,
7788 .spare_active = raid5_spare_active,
7789 .sync_request = sync_request,
7790 .resize = raid5_resize,
7792 .check_reshape = raid5_check_reshape,
7793 .start_reshape = raid5_start_reshape,
7794 .finish_reshape = raid5_finish_reshape,
7795 .quiesce = raid5_quiesce,
7796 .takeover = raid4_takeover,
7797 .congested = raid5_congested,
7798 .mergeable_bvec = raid5_mergeable_bvec,
7801 static int __init raid5_init(void)
7803 raid5_wq = alloc_workqueue("raid5wq",
7804 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7807 register_md_personality(&raid6_personality);
7808 register_md_personality(&raid5_personality);
7809 register_md_personality(&raid4_personality);
7813 static void raid5_exit(void)
7815 unregister_md_personality(&raid6_personality);
7816 unregister_md_personality(&raid5_personality);
7817 unregister_md_personality(&raid4_personality);
7818 destroy_workqueue(raid5_wq);
7821 module_init(raid5_init);
7822 module_exit(raid5_exit);
7823 MODULE_LICENSE("GPL");
7824 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7825 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7826 MODULE_ALIAS("md-raid5");
7827 MODULE_ALIAS("md-raid4");
7828 MODULE_ALIAS("md-level-5");
7829 MODULE_ALIAS("md-level-4");
7830 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7831 MODULE_ALIAS("md-raid6");
7832 MODULE_ALIAS("md-level-6");
7834 /* This used to be two separate modules, they were: */
7835 MODULE_ALIAS("raid5");
7836 MODULE_ALIAS("raid6");