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 unsigned long do_wakeup = 0;
351 if (hash == NR_STRIPE_HASH_LOCKS) {
352 size = NR_STRIPE_HASH_LOCKS;
353 hash = NR_STRIPE_HASH_LOCKS - 1;
357 struct list_head *list = &temp_inactive_list[size - 1];
360 * We don't hold any lock here yet, get_active_stripe() might
361 * remove stripes from the list
363 if (!list_empty_careful(list)) {
364 spin_lock_irqsave(conf->hash_locks + hash, flags);
365 if (list_empty(conf->inactive_list + hash) &&
367 atomic_dec(&conf->empty_inactive_list_nr);
368 list_splice_tail_init(list, conf->inactive_list + hash);
369 do_wakeup |= 1 << hash;
370 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
376 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
377 if (do_wakeup & (1 << i))
378 wake_up(&conf->wait_for_stripe[i]);
382 if (atomic_read(&conf->active_stripes) == 0)
383 wake_up(&conf->wait_for_quiescent);
384 if (conf->retry_read_aligned)
385 md_wakeup_thread(conf->mddev->thread);
389 /* should hold conf->device_lock already */
390 static int release_stripe_list(struct r5conf *conf,
391 struct list_head *temp_inactive_list)
393 struct stripe_head *sh;
395 struct llist_node *head;
397 head = llist_del_all(&conf->released_stripes);
398 head = llist_reverse_order(head);
402 sh = llist_entry(head, struct stripe_head, release_list);
403 head = llist_next(head);
404 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
406 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
408 * Don't worry the bit is set here, because if the bit is set
409 * again, the count is always > 1. This is true for
410 * STRIPE_ON_UNPLUG_LIST bit too.
412 hash = sh->hash_lock_index;
413 __release_stripe(conf, sh, &temp_inactive_list[hash]);
420 static void release_stripe(struct stripe_head *sh)
422 struct r5conf *conf = sh->raid_conf;
424 struct list_head list;
428 /* Avoid release_list until the last reference.
430 if (atomic_add_unless(&sh->count, -1, 1))
433 if (unlikely(!conf->mddev->thread) ||
434 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
436 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
438 md_wakeup_thread(conf->mddev->thread);
441 local_irq_save(flags);
442 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
443 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
444 INIT_LIST_HEAD(&list);
445 hash = sh->hash_lock_index;
446 do_release_stripe(conf, sh, &list);
447 spin_unlock(&conf->device_lock);
448 release_inactive_stripe_list(conf, &list, hash);
450 local_irq_restore(flags);
453 static inline void remove_hash(struct stripe_head *sh)
455 pr_debug("remove_hash(), stripe %llu\n",
456 (unsigned long long)sh->sector);
458 hlist_del_init(&sh->hash);
461 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
463 struct hlist_head *hp = stripe_hash(conf, sh->sector);
465 pr_debug("insert_hash(), stripe %llu\n",
466 (unsigned long long)sh->sector);
468 hlist_add_head(&sh->hash, hp);
471 /* find an idle stripe, make sure it is unhashed, and return it. */
472 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
474 struct stripe_head *sh = NULL;
475 struct list_head *first;
477 if (list_empty(conf->inactive_list + hash))
479 first = (conf->inactive_list + hash)->next;
480 sh = list_entry(first, struct stripe_head, lru);
481 list_del_init(first);
483 atomic_inc(&conf->active_stripes);
484 BUG_ON(hash != sh->hash_lock_index);
485 if (list_empty(conf->inactive_list + hash))
486 atomic_inc(&conf->empty_inactive_list_nr);
491 static void shrink_buffers(struct stripe_head *sh)
495 int num = sh->raid_conf->pool_size;
497 for (i = 0; i < num ; i++) {
498 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
502 sh->dev[i].page = NULL;
507 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
510 int num = sh->raid_conf->pool_size;
512 for (i = 0; i < num; i++) {
515 if (!(page = alloc_page(gfp))) {
518 sh->dev[i].page = page;
519 sh->dev[i].orig_page = page;
524 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
525 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
526 struct stripe_head *sh);
528 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
530 struct r5conf *conf = sh->raid_conf;
533 BUG_ON(atomic_read(&sh->count) != 0);
534 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
535 BUG_ON(stripe_operations_active(sh));
536 BUG_ON(sh->batch_head);
538 pr_debug("init_stripe called, stripe %llu\n",
539 (unsigned long long)sector);
541 seq = read_seqcount_begin(&conf->gen_lock);
542 sh->generation = conf->generation - previous;
543 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
545 stripe_set_idx(sector, conf, previous, sh);
548 for (i = sh->disks; i--; ) {
549 struct r5dev *dev = &sh->dev[i];
551 if (dev->toread || dev->read || dev->towrite || dev->written ||
552 test_bit(R5_LOCKED, &dev->flags)) {
553 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
554 (unsigned long long)sh->sector, i, dev->toread,
555 dev->read, dev->towrite, dev->written,
556 test_bit(R5_LOCKED, &dev->flags));
560 raid5_build_block(sh, i, previous);
562 if (read_seqcount_retry(&conf->gen_lock, seq))
564 sh->overwrite_disks = 0;
565 insert_hash(conf, sh);
566 sh->cpu = smp_processor_id();
567 set_bit(STRIPE_BATCH_READY, &sh->state);
570 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
573 struct stripe_head *sh;
575 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
576 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
577 if (sh->sector == sector && sh->generation == generation)
579 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
584 * Need to check if array has failed when deciding whether to:
586 * - remove non-faulty devices
589 * This determination is simple when no reshape is happening.
590 * However if there is a reshape, we need to carefully check
591 * both the before and after sections.
592 * This is because some failed devices may only affect one
593 * of the two sections, and some non-in_sync devices may
594 * be insync in the section most affected by failed devices.
596 static int calc_degraded(struct r5conf *conf)
598 int degraded, degraded2;
603 for (i = 0; i < conf->previous_raid_disks; i++) {
604 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
605 if (rdev && test_bit(Faulty, &rdev->flags))
606 rdev = rcu_dereference(conf->disks[i].replacement);
607 if (!rdev || test_bit(Faulty, &rdev->flags))
609 else if (test_bit(In_sync, &rdev->flags))
612 /* not in-sync or faulty.
613 * If the reshape increases the number of devices,
614 * this is being recovered by the reshape, so
615 * this 'previous' section is not in_sync.
616 * If the number of devices is being reduced however,
617 * the device can only be part of the array if
618 * we are reverting a reshape, so this section will
621 if (conf->raid_disks >= conf->previous_raid_disks)
625 if (conf->raid_disks == conf->previous_raid_disks)
629 for (i = 0; i < conf->raid_disks; i++) {
630 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
631 if (rdev && test_bit(Faulty, &rdev->flags))
632 rdev = rcu_dereference(conf->disks[i].replacement);
633 if (!rdev || test_bit(Faulty, &rdev->flags))
635 else if (test_bit(In_sync, &rdev->flags))
638 /* not in-sync or faulty.
639 * If reshape increases the number of devices, this
640 * section has already been recovered, else it
641 * almost certainly hasn't.
643 if (conf->raid_disks <= conf->previous_raid_disks)
647 if (degraded2 > degraded)
652 static int has_failed(struct r5conf *conf)
656 if (conf->mddev->reshape_position == MaxSector)
657 return conf->mddev->degraded > conf->max_degraded;
659 degraded = calc_degraded(conf);
660 if (degraded > conf->max_degraded)
665 static struct stripe_head *
666 get_active_stripe(struct r5conf *conf, sector_t sector,
667 int previous, int noblock, int noquiesce)
669 struct stripe_head *sh;
670 int hash = stripe_hash_locks_hash(sector);
672 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
674 spin_lock_irq(conf->hash_locks + hash);
677 wait_event_lock_irq(conf->wait_for_quiescent,
678 conf->quiesce == 0 || noquiesce,
679 *(conf->hash_locks + hash));
680 sh = __find_stripe(conf, sector, conf->generation - previous);
682 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
683 sh = get_free_stripe(conf, hash);
684 if (!sh && !test_bit(R5_DID_ALLOC,
686 set_bit(R5_ALLOC_MORE,
689 if (noblock && sh == NULL)
692 set_bit(R5_INACTIVE_BLOCKED,
694 wait_event_exclusive_cmd(
695 conf->wait_for_stripe[hash],
696 !list_empty(conf->inactive_list + hash) &&
697 (atomic_read(&conf->active_stripes)
698 < (conf->max_nr_stripes * 3 / 4)
699 || !test_bit(R5_INACTIVE_BLOCKED,
700 &conf->cache_state)),
701 spin_unlock_irq(conf->hash_locks + hash),
702 spin_lock_irq(conf->hash_locks + hash));
703 clear_bit(R5_INACTIVE_BLOCKED,
706 init_stripe(sh, sector, previous);
707 atomic_inc(&sh->count);
709 } else if (!atomic_inc_not_zero(&sh->count)) {
710 spin_lock(&conf->device_lock);
711 if (!atomic_read(&sh->count)) {
712 if (!test_bit(STRIPE_HANDLE, &sh->state))
713 atomic_inc(&conf->active_stripes);
714 BUG_ON(list_empty(&sh->lru) &&
715 !test_bit(STRIPE_EXPANDING, &sh->state));
716 list_del_init(&sh->lru);
718 sh->group->stripes_cnt--;
722 atomic_inc(&sh->count);
723 spin_unlock(&conf->device_lock);
725 } while (sh == NULL);
727 if (!list_empty(conf->inactive_list + hash))
728 wake_up(&conf->wait_for_stripe[hash]);
730 spin_unlock_irq(conf->hash_locks + hash);
734 static bool is_full_stripe_write(struct stripe_head *sh)
736 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
737 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
740 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
744 spin_lock(&sh2->stripe_lock);
745 spin_lock_nested(&sh1->stripe_lock, 1);
747 spin_lock(&sh1->stripe_lock);
748 spin_lock_nested(&sh2->stripe_lock, 1);
752 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
754 spin_unlock(&sh1->stripe_lock);
755 spin_unlock(&sh2->stripe_lock);
759 /* Only freshly new full stripe normal write stripe can be added to a batch list */
760 static bool stripe_can_batch(struct stripe_head *sh)
762 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
763 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
764 is_full_stripe_write(sh);
767 /* we only do back search */
768 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
770 struct stripe_head *head;
771 sector_t head_sector, tmp_sec;
775 if (!stripe_can_batch(sh))
777 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
778 tmp_sec = sh->sector;
779 if (!sector_div(tmp_sec, conf->chunk_sectors))
781 head_sector = sh->sector - STRIPE_SECTORS;
783 hash = stripe_hash_locks_hash(head_sector);
784 spin_lock_irq(conf->hash_locks + hash);
785 head = __find_stripe(conf, head_sector, conf->generation);
786 if (head && !atomic_inc_not_zero(&head->count)) {
787 spin_lock(&conf->device_lock);
788 if (!atomic_read(&head->count)) {
789 if (!test_bit(STRIPE_HANDLE, &head->state))
790 atomic_inc(&conf->active_stripes);
791 BUG_ON(list_empty(&head->lru) &&
792 !test_bit(STRIPE_EXPANDING, &head->state));
793 list_del_init(&head->lru);
795 head->group->stripes_cnt--;
799 atomic_inc(&head->count);
800 spin_unlock(&conf->device_lock);
802 spin_unlock_irq(conf->hash_locks + hash);
806 if (!stripe_can_batch(head))
809 lock_two_stripes(head, sh);
810 /* clear_batch_ready clear the flag */
811 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
818 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
820 if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
823 if (head->batch_head) {
824 spin_lock(&head->batch_head->batch_lock);
825 /* This batch list is already running */
826 if (!stripe_can_batch(head)) {
827 spin_unlock(&head->batch_head->batch_lock);
832 * at this point, head's BATCH_READY could be cleared, but we
833 * can still add the stripe to batch list
835 list_add(&sh->batch_list, &head->batch_list);
836 spin_unlock(&head->batch_head->batch_lock);
838 sh->batch_head = head->batch_head;
840 head->batch_head = head;
841 sh->batch_head = head->batch_head;
842 spin_lock(&head->batch_lock);
843 list_add_tail(&sh->batch_list, &head->batch_list);
844 spin_unlock(&head->batch_lock);
847 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
848 if (atomic_dec_return(&conf->preread_active_stripes)
850 md_wakeup_thread(conf->mddev->thread);
852 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
853 int seq = sh->bm_seq;
854 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
855 sh->batch_head->bm_seq > seq)
856 seq = sh->batch_head->bm_seq;
857 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
858 sh->batch_head->bm_seq = seq;
861 atomic_inc(&sh->count);
863 unlock_two_stripes(head, sh);
865 release_stripe(head);
868 /* Determine if 'data_offset' or 'new_data_offset' should be used
869 * in this stripe_head.
871 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
873 sector_t progress = conf->reshape_progress;
874 /* Need a memory barrier to make sure we see the value
875 * of conf->generation, or ->data_offset that was set before
876 * reshape_progress was updated.
879 if (progress == MaxSector)
881 if (sh->generation == conf->generation - 1)
883 /* We are in a reshape, and this is a new-generation stripe,
884 * so use new_data_offset.
890 raid5_end_read_request(struct bio *bi, int error);
892 raid5_end_write_request(struct bio *bi, int error);
894 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
896 struct r5conf *conf = sh->raid_conf;
897 int i, disks = sh->disks;
898 struct stripe_head *head_sh = sh;
902 for (i = disks; i--; ) {
904 int replace_only = 0;
905 struct bio *bi, *rbi;
906 struct md_rdev *rdev, *rrdev = NULL;
909 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
910 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
914 if (test_bit(R5_Discard, &sh->dev[i].flags))
916 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
918 else if (test_and_clear_bit(R5_WantReplace,
919 &sh->dev[i].flags)) {
924 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
928 bi = &sh->dev[i].req;
929 rbi = &sh->dev[i].rreq; /* For writing to replacement */
932 rrdev = rcu_dereference(conf->disks[i].replacement);
933 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
934 rdev = rcu_dereference(conf->disks[i].rdev);
943 /* We raced and saw duplicates */
946 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
951 if (rdev && test_bit(Faulty, &rdev->flags))
954 atomic_inc(&rdev->nr_pending);
955 if (rrdev && test_bit(Faulty, &rrdev->flags))
958 atomic_inc(&rrdev->nr_pending);
961 /* We have already checked bad blocks for reads. Now
962 * need to check for writes. We never accept write errors
963 * on the replacement, so we don't to check rrdev.
965 while ((rw & WRITE) && rdev &&
966 test_bit(WriteErrorSeen, &rdev->flags)) {
969 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
970 &first_bad, &bad_sectors);
975 set_bit(BlockedBadBlocks, &rdev->flags);
976 if (!conf->mddev->external &&
977 conf->mddev->flags) {
978 /* It is very unlikely, but we might
979 * still need to write out the
980 * bad block log - better give it
982 md_check_recovery(conf->mddev);
985 * Because md_wait_for_blocked_rdev
986 * will dec nr_pending, we must
987 * increment it first.
989 atomic_inc(&rdev->nr_pending);
990 md_wait_for_blocked_rdev(rdev, conf->mddev);
992 /* Acknowledged bad block - skip the write */
993 rdev_dec_pending(rdev, conf->mddev);
999 if (s->syncing || s->expanding || s->expanded
1001 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1003 set_bit(STRIPE_IO_STARTED, &sh->state);
1006 bi->bi_bdev = rdev->bdev;
1008 bi->bi_end_io = (rw & WRITE)
1009 ? raid5_end_write_request
1010 : raid5_end_read_request;
1011 bi->bi_private = sh;
1013 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1014 __func__, (unsigned long long)sh->sector,
1016 atomic_inc(&sh->count);
1018 atomic_inc(&head_sh->count);
1019 if (use_new_offset(conf, sh))
1020 bi->bi_iter.bi_sector = (sh->sector
1021 + rdev->new_data_offset);
1023 bi->bi_iter.bi_sector = (sh->sector
1024 + rdev->data_offset);
1025 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1026 bi->bi_rw |= REQ_NOMERGE;
1028 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1029 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1030 sh->dev[i].vec.bv_page = sh->dev[i].page;
1032 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1033 bi->bi_io_vec[0].bv_offset = 0;
1034 bi->bi_iter.bi_size = STRIPE_SIZE;
1036 * If this is discard request, set bi_vcnt 0. We don't
1037 * want to confuse SCSI because SCSI will replace payload
1039 if (rw & REQ_DISCARD)
1042 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1044 if (conf->mddev->gendisk)
1045 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1046 bi, disk_devt(conf->mddev->gendisk),
1048 generic_make_request(bi);
1051 if (s->syncing || s->expanding || s->expanded
1053 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1055 set_bit(STRIPE_IO_STARTED, &sh->state);
1058 rbi->bi_bdev = rrdev->bdev;
1060 BUG_ON(!(rw & WRITE));
1061 rbi->bi_end_io = raid5_end_write_request;
1062 rbi->bi_private = sh;
1064 pr_debug("%s: for %llu schedule op %ld on "
1065 "replacement disc %d\n",
1066 __func__, (unsigned long long)sh->sector,
1068 atomic_inc(&sh->count);
1070 atomic_inc(&head_sh->count);
1071 if (use_new_offset(conf, sh))
1072 rbi->bi_iter.bi_sector = (sh->sector
1073 + rrdev->new_data_offset);
1075 rbi->bi_iter.bi_sector = (sh->sector
1076 + rrdev->data_offset);
1077 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1078 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1079 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1081 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1082 rbi->bi_io_vec[0].bv_offset = 0;
1083 rbi->bi_iter.bi_size = STRIPE_SIZE;
1085 * If this is discard request, set bi_vcnt 0. We don't
1086 * want to confuse SCSI because SCSI will replace payload
1088 if (rw & REQ_DISCARD)
1090 if (conf->mddev->gendisk)
1091 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1092 rbi, disk_devt(conf->mddev->gendisk),
1094 generic_make_request(rbi);
1096 if (!rdev && !rrdev) {
1098 set_bit(STRIPE_DEGRADED, &sh->state);
1099 pr_debug("skip op %ld on disc %d for sector %llu\n",
1100 bi->bi_rw, i, (unsigned long long)sh->sector);
1101 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1102 set_bit(STRIPE_HANDLE, &sh->state);
1105 if (!head_sh->batch_head)
1107 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1114 static struct dma_async_tx_descriptor *
1115 async_copy_data(int frombio, struct bio *bio, struct page **page,
1116 sector_t sector, struct dma_async_tx_descriptor *tx,
1117 struct stripe_head *sh)
1120 struct bvec_iter iter;
1121 struct page *bio_page;
1123 struct async_submit_ctl submit;
1124 enum async_tx_flags flags = 0;
1126 if (bio->bi_iter.bi_sector >= sector)
1127 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1129 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1132 flags |= ASYNC_TX_FENCE;
1133 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1135 bio_for_each_segment(bvl, bio, iter) {
1136 int len = bvl.bv_len;
1140 if (page_offset < 0) {
1141 b_offset = -page_offset;
1142 page_offset += b_offset;
1146 if (len > 0 && page_offset + len > STRIPE_SIZE)
1147 clen = STRIPE_SIZE - page_offset;
1152 b_offset += bvl.bv_offset;
1153 bio_page = bvl.bv_page;
1155 if (sh->raid_conf->skip_copy &&
1156 b_offset == 0 && page_offset == 0 &&
1157 clen == STRIPE_SIZE)
1160 tx = async_memcpy(*page, bio_page, page_offset,
1161 b_offset, clen, &submit);
1163 tx = async_memcpy(bio_page, *page, b_offset,
1164 page_offset, clen, &submit);
1166 /* chain the operations */
1167 submit.depend_tx = tx;
1169 if (clen < len) /* hit end of page */
1177 static void ops_complete_biofill(void *stripe_head_ref)
1179 struct stripe_head *sh = stripe_head_ref;
1180 struct bio *return_bi = NULL;
1183 pr_debug("%s: stripe %llu\n", __func__,
1184 (unsigned long long)sh->sector);
1186 /* clear completed biofills */
1187 for (i = sh->disks; i--; ) {
1188 struct r5dev *dev = &sh->dev[i];
1190 /* acknowledge completion of a biofill operation */
1191 /* and check if we need to reply to a read request,
1192 * new R5_Wantfill requests are held off until
1193 * !STRIPE_BIOFILL_RUN
1195 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1196 struct bio *rbi, *rbi2;
1201 while (rbi && rbi->bi_iter.bi_sector <
1202 dev->sector + STRIPE_SECTORS) {
1203 rbi2 = r5_next_bio(rbi, dev->sector);
1204 if (!raid5_dec_bi_active_stripes(rbi)) {
1205 rbi->bi_next = return_bi;
1212 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1214 return_io(return_bi);
1216 set_bit(STRIPE_HANDLE, &sh->state);
1220 static void ops_run_biofill(struct stripe_head *sh)
1222 struct dma_async_tx_descriptor *tx = NULL;
1223 struct async_submit_ctl submit;
1226 BUG_ON(sh->batch_head);
1227 pr_debug("%s: stripe %llu\n", __func__,
1228 (unsigned long long)sh->sector);
1230 for (i = sh->disks; i--; ) {
1231 struct r5dev *dev = &sh->dev[i];
1232 if (test_bit(R5_Wantfill, &dev->flags)) {
1234 spin_lock_irq(&sh->stripe_lock);
1235 dev->read = rbi = dev->toread;
1237 spin_unlock_irq(&sh->stripe_lock);
1238 while (rbi && rbi->bi_iter.bi_sector <
1239 dev->sector + STRIPE_SECTORS) {
1240 tx = async_copy_data(0, rbi, &dev->page,
1241 dev->sector, tx, sh);
1242 rbi = r5_next_bio(rbi, dev->sector);
1247 atomic_inc(&sh->count);
1248 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1249 async_trigger_callback(&submit);
1252 static void mark_target_uptodate(struct stripe_head *sh, int target)
1259 tgt = &sh->dev[target];
1260 set_bit(R5_UPTODATE, &tgt->flags);
1261 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1262 clear_bit(R5_Wantcompute, &tgt->flags);
1265 static void ops_complete_compute(void *stripe_head_ref)
1267 struct stripe_head *sh = stripe_head_ref;
1269 pr_debug("%s: stripe %llu\n", __func__,
1270 (unsigned long long)sh->sector);
1272 /* mark the computed target(s) as uptodate */
1273 mark_target_uptodate(sh, sh->ops.target);
1274 mark_target_uptodate(sh, sh->ops.target2);
1276 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1277 if (sh->check_state == check_state_compute_run)
1278 sh->check_state = check_state_compute_result;
1279 set_bit(STRIPE_HANDLE, &sh->state);
1283 /* return a pointer to the address conversion region of the scribble buffer */
1284 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1285 struct raid5_percpu *percpu, int i)
1289 addr = flex_array_get(percpu->scribble, i);
1290 return addr + sizeof(struct page *) * (sh->disks + 2);
1293 /* return a pointer to the address conversion region of the scribble buffer */
1294 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1298 addr = flex_array_get(percpu->scribble, i);
1302 static struct dma_async_tx_descriptor *
1303 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1305 int disks = sh->disks;
1306 struct page **xor_srcs = to_addr_page(percpu, 0);
1307 int target = sh->ops.target;
1308 struct r5dev *tgt = &sh->dev[target];
1309 struct page *xor_dest = tgt->page;
1311 struct dma_async_tx_descriptor *tx;
1312 struct async_submit_ctl submit;
1315 BUG_ON(sh->batch_head);
1317 pr_debug("%s: stripe %llu block: %d\n",
1318 __func__, (unsigned long long)sh->sector, target);
1319 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1321 for (i = disks; i--; )
1323 xor_srcs[count++] = sh->dev[i].page;
1325 atomic_inc(&sh->count);
1327 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1328 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1329 if (unlikely(count == 1))
1330 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1332 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1337 /* set_syndrome_sources - populate source buffers for gen_syndrome
1338 * @srcs - (struct page *) array of size sh->disks
1339 * @sh - stripe_head to parse
1341 * Populates srcs in proper layout order for the stripe and returns the
1342 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1343 * destination buffer is recorded in srcs[count] and the Q destination
1344 * is recorded in srcs[count+1]].
1346 static int set_syndrome_sources(struct page **srcs,
1347 struct stripe_head *sh,
1350 int disks = sh->disks;
1351 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1352 int d0_idx = raid6_d0(sh);
1356 for (i = 0; i < disks; i++)
1362 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1363 struct r5dev *dev = &sh->dev[i];
1365 if (i == sh->qd_idx || i == sh->pd_idx ||
1366 (srctype == SYNDROME_SRC_ALL) ||
1367 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1368 test_bit(R5_Wantdrain, &dev->flags)) ||
1369 (srctype == SYNDROME_SRC_WRITTEN &&
1371 srcs[slot] = sh->dev[i].page;
1372 i = raid6_next_disk(i, disks);
1373 } while (i != d0_idx);
1375 return syndrome_disks;
1378 static struct dma_async_tx_descriptor *
1379 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1381 int disks = sh->disks;
1382 struct page **blocks = to_addr_page(percpu, 0);
1384 int qd_idx = sh->qd_idx;
1385 struct dma_async_tx_descriptor *tx;
1386 struct async_submit_ctl submit;
1392 BUG_ON(sh->batch_head);
1393 if (sh->ops.target < 0)
1394 target = sh->ops.target2;
1395 else if (sh->ops.target2 < 0)
1396 target = sh->ops.target;
1398 /* we should only have one valid target */
1401 pr_debug("%s: stripe %llu block: %d\n",
1402 __func__, (unsigned long long)sh->sector, target);
1404 tgt = &sh->dev[target];
1405 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1408 atomic_inc(&sh->count);
1410 if (target == qd_idx) {
1411 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1412 blocks[count] = NULL; /* regenerating p is not necessary */
1413 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1414 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1415 ops_complete_compute, sh,
1416 to_addr_conv(sh, percpu, 0));
1417 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1419 /* Compute any data- or p-drive using XOR */
1421 for (i = disks; i-- ; ) {
1422 if (i == target || i == qd_idx)
1424 blocks[count++] = sh->dev[i].page;
1427 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1428 NULL, ops_complete_compute, sh,
1429 to_addr_conv(sh, percpu, 0));
1430 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1436 static struct dma_async_tx_descriptor *
1437 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1439 int i, count, disks = sh->disks;
1440 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1441 int d0_idx = raid6_d0(sh);
1442 int faila = -1, failb = -1;
1443 int target = sh->ops.target;
1444 int target2 = sh->ops.target2;
1445 struct r5dev *tgt = &sh->dev[target];
1446 struct r5dev *tgt2 = &sh->dev[target2];
1447 struct dma_async_tx_descriptor *tx;
1448 struct page **blocks = to_addr_page(percpu, 0);
1449 struct async_submit_ctl submit;
1451 BUG_ON(sh->batch_head);
1452 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1453 __func__, (unsigned long long)sh->sector, target, target2);
1454 BUG_ON(target < 0 || target2 < 0);
1455 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1456 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1458 /* we need to open-code set_syndrome_sources to handle the
1459 * slot number conversion for 'faila' and 'failb'
1461 for (i = 0; i < disks ; i++)
1466 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1468 blocks[slot] = sh->dev[i].page;
1474 i = raid6_next_disk(i, disks);
1475 } while (i != d0_idx);
1477 BUG_ON(faila == failb);
1480 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1481 __func__, (unsigned long long)sh->sector, faila, failb);
1483 atomic_inc(&sh->count);
1485 if (failb == syndrome_disks+1) {
1486 /* Q disk is one of the missing disks */
1487 if (faila == syndrome_disks) {
1488 /* Missing P+Q, just recompute */
1489 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1490 ops_complete_compute, sh,
1491 to_addr_conv(sh, percpu, 0));
1492 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1493 STRIPE_SIZE, &submit);
1497 int qd_idx = sh->qd_idx;
1499 /* Missing D+Q: recompute D from P, then recompute Q */
1500 if (target == qd_idx)
1501 data_target = target2;
1503 data_target = target;
1506 for (i = disks; i-- ; ) {
1507 if (i == data_target || i == qd_idx)
1509 blocks[count++] = sh->dev[i].page;
1511 dest = sh->dev[data_target].page;
1512 init_async_submit(&submit,
1513 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1515 to_addr_conv(sh, percpu, 0));
1516 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1519 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1520 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1521 ops_complete_compute, sh,
1522 to_addr_conv(sh, percpu, 0));
1523 return async_gen_syndrome(blocks, 0, count+2,
1524 STRIPE_SIZE, &submit);
1527 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1528 ops_complete_compute, sh,
1529 to_addr_conv(sh, percpu, 0));
1530 if (failb == syndrome_disks) {
1531 /* We're missing D+P. */
1532 return async_raid6_datap_recov(syndrome_disks+2,
1536 /* We're missing D+D. */
1537 return async_raid6_2data_recov(syndrome_disks+2,
1538 STRIPE_SIZE, faila, failb,
1544 static void ops_complete_prexor(void *stripe_head_ref)
1546 struct stripe_head *sh = stripe_head_ref;
1548 pr_debug("%s: stripe %llu\n", __func__,
1549 (unsigned long long)sh->sector);
1552 static struct dma_async_tx_descriptor *
1553 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1554 struct dma_async_tx_descriptor *tx)
1556 int disks = sh->disks;
1557 struct page **xor_srcs = to_addr_page(percpu, 0);
1558 int count = 0, pd_idx = sh->pd_idx, i;
1559 struct async_submit_ctl submit;
1561 /* existing parity data subtracted */
1562 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1564 BUG_ON(sh->batch_head);
1565 pr_debug("%s: stripe %llu\n", __func__,
1566 (unsigned long long)sh->sector);
1568 for (i = disks; i--; ) {
1569 struct r5dev *dev = &sh->dev[i];
1570 /* Only process blocks that are known to be uptodate */
1571 if (test_bit(R5_Wantdrain, &dev->flags))
1572 xor_srcs[count++] = dev->page;
1575 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1576 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1577 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1582 static struct dma_async_tx_descriptor *
1583 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1584 struct dma_async_tx_descriptor *tx)
1586 struct page **blocks = to_addr_page(percpu, 0);
1588 struct async_submit_ctl submit;
1590 pr_debug("%s: stripe %llu\n", __func__,
1591 (unsigned long long)sh->sector);
1593 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1595 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1596 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1597 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1602 static struct dma_async_tx_descriptor *
1603 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1605 int disks = sh->disks;
1607 struct stripe_head *head_sh = sh;
1609 pr_debug("%s: stripe %llu\n", __func__,
1610 (unsigned long long)sh->sector);
1612 for (i = disks; i--; ) {
1617 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1622 spin_lock_irq(&sh->stripe_lock);
1623 chosen = dev->towrite;
1624 dev->towrite = NULL;
1625 sh->overwrite_disks = 0;
1626 BUG_ON(dev->written);
1627 wbi = dev->written = chosen;
1628 spin_unlock_irq(&sh->stripe_lock);
1629 WARN_ON(dev->page != dev->orig_page);
1631 while (wbi && wbi->bi_iter.bi_sector <
1632 dev->sector + STRIPE_SECTORS) {
1633 if (wbi->bi_rw & REQ_FUA)
1634 set_bit(R5_WantFUA, &dev->flags);
1635 if (wbi->bi_rw & REQ_SYNC)
1636 set_bit(R5_SyncIO, &dev->flags);
1637 if (wbi->bi_rw & REQ_DISCARD)
1638 set_bit(R5_Discard, &dev->flags);
1640 tx = async_copy_data(1, wbi, &dev->page,
1641 dev->sector, tx, sh);
1642 if (dev->page != dev->orig_page) {
1643 set_bit(R5_SkipCopy, &dev->flags);
1644 clear_bit(R5_UPTODATE, &dev->flags);
1645 clear_bit(R5_OVERWRITE, &dev->flags);
1648 wbi = r5_next_bio(wbi, dev->sector);
1651 if (head_sh->batch_head) {
1652 sh = list_first_entry(&sh->batch_list,
1665 static void ops_complete_reconstruct(void *stripe_head_ref)
1667 struct stripe_head *sh = stripe_head_ref;
1668 int disks = sh->disks;
1669 int pd_idx = sh->pd_idx;
1670 int qd_idx = sh->qd_idx;
1672 bool fua = false, sync = false, discard = false;
1674 pr_debug("%s: stripe %llu\n", __func__,
1675 (unsigned long long)sh->sector);
1677 for (i = disks; i--; ) {
1678 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1679 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1680 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1683 for (i = disks; i--; ) {
1684 struct r5dev *dev = &sh->dev[i];
1686 if (dev->written || i == pd_idx || i == qd_idx) {
1687 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1688 set_bit(R5_UPTODATE, &dev->flags);
1690 set_bit(R5_WantFUA, &dev->flags);
1692 set_bit(R5_SyncIO, &dev->flags);
1696 if (sh->reconstruct_state == reconstruct_state_drain_run)
1697 sh->reconstruct_state = reconstruct_state_drain_result;
1698 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1699 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1701 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1702 sh->reconstruct_state = reconstruct_state_result;
1705 set_bit(STRIPE_HANDLE, &sh->state);
1710 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1711 struct dma_async_tx_descriptor *tx)
1713 int disks = sh->disks;
1714 struct page **xor_srcs;
1715 struct async_submit_ctl submit;
1716 int count, pd_idx = sh->pd_idx, i;
1717 struct page *xor_dest;
1719 unsigned long flags;
1721 struct stripe_head *head_sh = sh;
1724 pr_debug("%s: stripe %llu\n", __func__,
1725 (unsigned long long)sh->sector);
1727 for (i = 0; i < sh->disks; i++) {
1730 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1733 if (i >= sh->disks) {
1734 atomic_inc(&sh->count);
1735 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1736 ops_complete_reconstruct(sh);
1741 xor_srcs = to_addr_page(percpu, j);
1742 /* check if prexor is active which means only process blocks
1743 * that are part of a read-modify-write (written)
1745 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1747 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1748 for (i = disks; i--; ) {
1749 struct r5dev *dev = &sh->dev[i];
1750 if (head_sh->dev[i].written)
1751 xor_srcs[count++] = dev->page;
1754 xor_dest = sh->dev[pd_idx].page;
1755 for (i = disks; i--; ) {
1756 struct r5dev *dev = &sh->dev[i];
1758 xor_srcs[count++] = dev->page;
1762 /* 1/ if we prexor'd then the dest is reused as a source
1763 * 2/ if we did not prexor then we are redoing the parity
1764 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1765 * for the synchronous xor case
1767 last_stripe = !head_sh->batch_head ||
1768 list_first_entry(&sh->batch_list,
1769 struct stripe_head, batch_list) == head_sh;
1771 flags = ASYNC_TX_ACK |
1772 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1774 atomic_inc(&head_sh->count);
1775 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1776 to_addr_conv(sh, percpu, j));
1778 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1779 init_async_submit(&submit, flags, tx, NULL, NULL,
1780 to_addr_conv(sh, percpu, j));
1783 if (unlikely(count == 1))
1784 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1786 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1789 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1796 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1797 struct dma_async_tx_descriptor *tx)
1799 struct async_submit_ctl submit;
1800 struct page **blocks;
1801 int count, i, j = 0;
1802 struct stripe_head *head_sh = sh;
1805 unsigned long txflags;
1807 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1809 for (i = 0; i < sh->disks; i++) {
1810 if (sh->pd_idx == i || sh->qd_idx == i)
1812 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1815 if (i >= sh->disks) {
1816 atomic_inc(&sh->count);
1817 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1818 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1819 ops_complete_reconstruct(sh);
1824 blocks = to_addr_page(percpu, j);
1826 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1827 synflags = SYNDROME_SRC_WRITTEN;
1828 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1830 synflags = SYNDROME_SRC_ALL;
1831 txflags = ASYNC_TX_ACK;
1834 count = set_syndrome_sources(blocks, sh, synflags);
1835 last_stripe = !head_sh->batch_head ||
1836 list_first_entry(&sh->batch_list,
1837 struct stripe_head, batch_list) == head_sh;
1840 atomic_inc(&head_sh->count);
1841 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1842 head_sh, to_addr_conv(sh, percpu, j));
1844 init_async_submit(&submit, 0, tx, NULL, NULL,
1845 to_addr_conv(sh, percpu, j));
1846 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1849 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1855 static void ops_complete_check(void *stripe_head_ref)
1857 struct stripe_head *sh = stripe_head_ref;
1859 pr_debug("%s: stripe %llu\n", __func__,
1860 (unsigned long long)sh->sector);
1862 sh->check_state = check_state_check_result;
1863 set_bit(STRIPE_HANDLE, &sh->state);
1867 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1869 int disks = sh->disks;
1870 int pd_idx = sh->pd_idx;
1871 int qd_idx = sh->qd_idx;
1872 struct page *xor_dest;
1873 struct page **xor_srcs = to_addr_page(percpu, 0);
1874 struct dma_async_tx_descriptor *tx;
1875 struct async_submit_ctl submit;
1879 pr_debug("%s: stripe %llu\n", __func__,
1880 (unsigned long long)sh->sector);
1882 BUG_ON(sh->batch_head);
1884 xor_dest = sh->dev[pd_idx].page;
1885 xor_srcs[count++] = xor_dest;
1886 for (i = disks; i--; ) {
1887 if (i == pd_idx || i == qd_idx)
1889 xor_srcs[count++] = sh->dev[i].page;
1892 init_async_submit(&submit, 0, NULL, NULL, NULL,
1893 to_addr_conv(sh, percpu, 0));
1894 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1895 &sh->ops.zero_sum_result, &submit);
1897 atomic_inc(&sh->count);
1898 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1899 tx = async_trigger_callback(&submit);
1902 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1904 struct page **srcs = to_addr_page(percpu, 0);
1905 struct async_submit_ctl submit;
1908 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1909 (unsigned long long)sh->sector, checkp);
1911 BUG_ON(sh->batch_head);
1912 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1916 atomic_inc(&sh->count);
1917 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1918 sh, to_addr_conv(sh, percpu, 0));
1919 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1920 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1923 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1925 int overlap_clear = 0, i, disks = sh->disks;
1926 struct dma_async_tx_descriptor *tx = NULL;
1927 struct r5conf *conf = sh->raid_conf;
1928 int level = conf->level;
1929 struct raid5_percpu *percpu;
1933 percpu = per_cpu_ptr(conf->percpu, cpu);
1934 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1935 ops_run_biofill(sh);
1939 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1941 tx = ops_run_compute5(sh, percpu);
1943 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1944 tx = ops_run_compute6_1(sh, percpu);
1946 tx = ops_run_compute6_2(sh, percpu);
1948 /* terminate the chain if reconstruct is not set to be run */
1949 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1953 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1955 tx = ops_run_prexor5(sh, percpu, tx);
1957 tx = ops_run_prexor6(sh, percpu, tx);
1960 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1961 tx = ops_run_biodrain(sh, tx);
1965 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1967 ops_run_reconstruct5(sh, percpu, tx);
1969 ops_run_reconstruct6(sh, percpu, tx);
1972 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1973 if (sh->check_state == check_state_run)
1974 ops_run_check_p(sh, percpu);
1975 else if (sh->check_state == check_state_run_q)
1976 ops_run_check_pq(sh, percpu, 0);
1977 else if (sh->check_state == check_state_run_pq)
1978 ops_run_check_pq(sh, percpu, 1);
1983 if (overlap_clear && !sh->batch_head)
1984 for (i = disks; i--; ) {
1985 struct r5dev *dev = &sh->dev[i];
1986 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1987 wake_up(&sh->raid_conf->wait_for_overlap);
1992 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1994 struct stripe_head *sh;
1996 sh = kmem_cache_zalloc(sc, gfp);
1998 spin_lock_init(&sh->stripe_lock);
1999 spin_lock_init(&sh->batch_lock);
2000 INIT_LIST_HEAD(&sh->batch_list);
2001 INIT_LIST_HEAD(&sh->lru);
2002 atomic_set(&sh->count, 1);
2006 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2008 struct stripe_head *sh;
2010 sh = alloc_stripe(conf->slab_cache, gfp);
2014 sh->raid_conf = conf;
2016 if (grow_buffers(sh, gfp)) {
2018 kmem_cache_free(conf->slab_cache, sh);
2021 sh->hash_lock_index =
2022 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2023 /* we just created an active stripe so... */
2024 atomic_inc(&conf->active_stripes);
2027 conf->max_nr_stripes++;
2031 static int grow_stripes(struct r5conf *conf, int num)
2033 struct kmem_cache *sc;
2034 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2036 if (conf->mddev->gendisk)
2037 sprintf(conf->cache_name[0],
2038 "raid%d-%s", conf->level, mdname(conf->mddev));
2040 sprintf(conf->cache_name[0],
2041 "raid%d-%p", conf->level, conf->mddev);
2042 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2044 conf->active_name = 0;
2045 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2046 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2050 conf->slab_cache = sc;
2051 conf->pool_size = devs;
2053 if (!grow_one_stripe(conf, GFP_KERNEL))
2060 * scribble_len - return the required size of the scribble region
2061 * @num - total number of disks in the array
2063 * The size must be enough to contain:
2064 * 1/ a struct page pointer for each device in the array +2
2065 * 2/ room to convert each entry in (1) to its corresponding dma
2066 * (dma_map_page()) or page (page_address()) address.
2068 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2069 * calculate over all devices (not just the data blocks), using zeros in place
2070 * of the P and Q blocks.
2072 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2074 struct flex_array *ret;
2077 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2078 ret = flex_array_alloc(len, cnt, flags);
2081 /* always prealloc all elements, so no locking is required */
2082 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2083 flex_array_free(ret);
2089 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2094 mddev_suspend(conf->mddev);
2096 for_each_present_cpu(cpu) {
2097 struct raid5_percpu *percpu;
2098 struct flex_array *scribble;
2100 percpu = per_cpu_ptr(conf->percpu, cpu);
2101 scribble = scribble_alloc(new_disks,
2102 new_sectors / STRIPE_SECTORS,
2106 flex_array_free(percpu->scribble);
2107 percpu->scribble = scribble;
2114 mddev_resume(conf->mddev);
2118 static int resize_stripes(struct r5conf *conf, int newsize)
2120 /* Make all the stripes able to hold 'newsize' devices.
2121 * New slots in each stripe get 'page' set to a new page.
2123 * This happens in stages:
2124 * 1/ create a new kmem_cache and allocate the required number of
2126 * 2/ gather all the old stripe_heads and transfer the pages across
2127 * to the new stripe_heads. This will have the side effect of
2128 * freezing the array as once all stripe_heads have been collected,
2129 * no IO will be possible. Old stripe heads are freed once their
2130 * pages have been transferred over, and the old kmem_cache is
2131 * freed when all stripes are done.
2132 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2133 * we simple return a failre status - no need to clean anything up.
2134 * 4/ allocate new pages for the new slots in the new stripe_heads.
2135 * If this fails, we don't bother trying the shrink the
2136 * stripe_heads down again, we just leave them as they are.
2137 * As each stripe_head is processed the new one is released into
2140 * Once step2 is started, we cannot afford to wait for a write,
2141 * so we use GFP_NOIO allocations.
2143 struct stripe_head *osh, *nsh;
2144 LIST_HEAD(newstripes);
2145 struct disk_info *ndisks;
2147 struct kmem_cache *sc;
2151 if (newsize <= conf->pool_size)
2152 return 0; /* never bother to shrink */
2154 err = md_allow_write(conf->mddev);
2159 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2160 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2165 /* Need to ensure auto-resizing doesn't interfere */
2166 mutex_lock(&conf->cache_size_mutex);
2168 for (i = conf->max_nr_stripes; i; i--) {
2169 nsh = alloc_stripe(sc, GFP_KERNEL);
2173 nsh->raid_conf = conf;
2174 list_add(&nsh->lru, &newstripes);
2177 /* didn't get enough, give up */
2178 while (!list_empty(&newstripes)) {
2179 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2180 list_del(&nsh->lru);
2181 kmem_cache_free(sc, nsh);
2183 kmem_cache_destroy(sc);
2184 mutex_unlock(&conf->cache_size_mutex);
2187 /* Step 2 - Must use GFP_NOIO now.
2188 * OK, we have enough stripes, start collecting inactive
2189 * stripes and copying them over
2193 list_for_each_entry(nsh, &newstripes, lru) {
2194 lock_device_hash_lock(conf, hash);
2195 wait_event_exclusive_cmd(conf->wait_for_stripe[hash],
2196 !list_empty(conf->inactive_list + hash),
2197 unlock_device_hash_lock(conf, hash),
2198 lock_device_hash_lock(conf, hash));
2199 osh = get_free_stripe(conf, hash);
2200 unlock_device_hash_lock(conf, hash);
2202 for(i=0; i<conf->pool_size; i++) {
2203 nsh->dev[i].page = osh->dev[i].page;
2204 nsh->dev[i].orig_page = osh->dev[i].page;
2206 nsh->hash_lock_index = hash;
2207 kmem_cache_free(conf->slab_cache, osh);
2209 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2210 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2215 kmem_cache_destroy(conf->slab_cache);
2218 * At this point, we are holding all the stripes so the array
2219 * is completely stalled, so now is a good time to resize
2220 * conf->disks and the scribble region
2222 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2224 for (i=0; i<conf->raid_disks; i++)
2225 ndisks[i] = conf->disks[i];
2227 conf->disks = ndisks;
2231 mutex_unlock(&conf->cache_size_mutex);
2232 /* Step 4, return new stripes to service */
2233 while(!list_empty(&newstripes)) {
2234 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2235 list_del_init(&nsh->lru);
2237 for (i=conf->raid_disks; i < newsize; i++)
2238 if (nsh->dev[i].page == NULL) {
2239 struct page *p = alloc_page(GFP_NOIO);
2240 nsh->dev[i].page = p;
2241 nsh->dev[i].orig_page = p;
2245 release_stripe(nsh);
2247 /* critical section pass, GFP_NOIO no longer needed */
2249 conf->slab_cache = sc;
2250 conf->active_name = 1-conf->active_name;
2252 conf->pool_size = newsize;
2256 static int drop_one_stripe(struct r5conf *conf)
2258 struct stripe_head *sh;
2259 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2261 spin_lock_irq(conf->hash_locks + hash);
2262 sh = get_free_stripe(conf, hash);
2263 spin_unlock_irq(conf->hash_locks + hash);
2266 BUG_ON(atomic_read(&sh->count));
2268 kmem_cache_free(conf->slab_cache, sh);
2269 atomic_dec(&conf->active_stripes);
2270 conf->max_nr_stripes--;
2274 static void shrink_stripes(struct r5conf *conf)
2276 while (conf->max_nr_stripes &&
2277 drop_one_stripe(conf))
2280 if (conf->slab_cache)
2281 kmem_cache_destroy(conf->slab_cache);
2282 conf->slab_cache = NULL;
2285 static void raid5_end_read_request(struct bio * bi, int error)
2287 struct stripe_head *sh = bi->bi_private;
2288 struct r5conf *conf = sh->raid_conf;
2289 int disks = sh->disks, i;
2290 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2291 char b[BDEVNAME_SIZE];
2292 struct md_rdev *rdev = NULL;
2295 for (i=0 ; i<disks; i++)
2296 if (bi == &sh->dev[i].req)
2299 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2300 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2306 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2307 /* If replacement finished while this request was outstanding,
2308 * 'replacement' might be NULL already.
2309 * In that case it moved down to 'rdev'.
2310 * rdev is not removed until all requests are finished.
2312 rdev = conf->disks[i].replacement;
2314 rdev = conf->disks[i].rdev;
2316 if (use_new_offset(conf, sh))
2317 s = sh->sector + rdev->new_data_offset;
2319 s = sh->sector + rdev->data_offset;
2321 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2322 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2323 /* Note that this cannot happen on a
2324 * replacement device. We just fail those on
2329 "md/raid:%s: read error corrected"
2330 " (%lu sectors at %llu on %s)\n",
2331 mdname(conf->mddev), STRIPE_SECTORS,
2332 (unsigned long long)s,
2333 bdevname(rdev->bdev, b));
2334 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2335 clear_bit(R5_ReadError, &sh->dev[i].flags);
2336 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2337 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2338 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2340 if (atomic_read(&rdev->read_errors))
2341 atomic_set(&rdev->read_errors, 0);
2343 const char *bdn = bdevname(rdev->bdev, b);
2347 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2348 atomic_inc(&rdev->read_errors);
2349 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2352 "md/raid:%s: read error on replacement device "
2353 "(sector %llu on %s).\n",
2354 mdname(conf->mddev),
2355 (unsigned long long)s,
2357 else if (conf->mddev->degraded >= conf->max_degraded) {
2361 "md/raid:%s: read error not correctable "
2362 "(sector %llu on %s).\n",
2363 mdname(conf->mddev),
2364 (unsigned long long)s,
2366 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2371 "md/raid:%s: read error NOT corrected!! "
2372 "(sector %llu on %s).\n",
2373 mdname(conf->mddev),
2374 (unsigned long long)s,
2376 } else if (atomic_read(&rdev->read_errors)
2377 > conf->max_nr_stripes)
2379 "md/raid:%s: Too many read errors, failing device %s.\n",
2380 mdname(conf->mddev), bdn);
2383 if (set_bad && test_bit(In_sync, &rdev->flags)
2384 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2387 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2388 set_bit(R5_ReadError, &sh->dev[i].flags);
2389 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2391 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2393 clear_bit(R5_ReadError, &sh->dev[i].flags);
2394 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2396 && test_bit(In_sync, &rdev->flags)
2397 && rdev_set_badblocks(
2398 rdev, sh->sector, STRIPE_SECTORS, 0)))
2399 md_error(conf->mddev, rdev);
2402 rdev_dec_pending(rdev, conf->mddev);
2403 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2404 set_bit(STRIPE_HANDLE, &sh->state);
2408 static void raid5_end_write_request(struct bio *bi, int error)
2410 struct stripe_head *sh = bi->bi_private;
2411 struct r5conf *conf = sh->raid_conf;
2412 int disks = sh->disks, i;
2413 struct md_rdev *uninitialized_var(rdev);
2414 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2417 int replacement = 0;
2419 for (i = 0 ; i < disks; i++) {
2420 if (bi == &sh->dev[i].req) {
2421 rdev = conf->disks[i].rdev;
2424 if (bi == &sh->dev[i].rreq) {
2425 rdev = conf->disks[i].replacement;
2429 /* rdev was removed and 'replacement'
2430 * replaced it. rdev is not removed
2431 * until all requests are finished.
2433 rdev = conf->disks[i].rdev;
2437 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2438 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2447 md_error(conf->mddev, rdev);
2448 else if (is_badblock(rdev, sh->sector,
2450 &first_bad, &bad_sectors))
2451 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2454 set_bit(STRIPE_DEGRADED, &sh->state);
2455 set_bit(WriteErrorSeen, &rdev->flags);
2456 set_bit(R5_WriteError, &sh->dev[i].flags);
2457 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2458 set_bit(MD_RECOVERY_NEEDED,
2459 &rdev->mddev->recovery);
2460 } else if (is_badblock(rdev, sh->sector,
2462 &first_bad, &bad_sectors)) {
2463 set_bit(R5_MadeGood, &sh->dev[i].flags);
2464 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2465 /* That was a successful write so make
2466 * sure it looks like we already did
2469 set_bit(R5_ReWrite, &sh->dev[i].flags);
2472 rdev_dec_pending(rdev, conf->mddev);
2474 if (sh->batch_head && !uptodate && !replacement)
2475 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2477 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2478 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2479 set_bit(STRIPE_HANDLE, &sh->state);
2482 if (sh->batch_head && sh != sh->batch_head)
2483 release_stripe(sh->batch_head);
2486 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2488 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2490 struct r5dev *dev = &sh->dev[i];
2492 bio_init(&dev->req);
2493 dev->req.bi_io_vec = &dev->vec;
2494 dev->req.bi_max_vecs = 1;
2495 dev->req.bi_private = sh;
2497 bio_init(&dev->rreq);
2498 dev->rreq.bi_io_vec = &dev->rvec;
2499 dev->rreq.bi_max_vecs = 1;
2500 dev->rreq.bi_private = sh;
2503 dev->sector = compute_blocknr(sh, i, previous);
2506 static void error(struct mddev *mddev, struct md_rdev *rdev)
2508 char b[BDEVNAME_SIZE];
2509 struct r5conf *conf = mddev->private;
2510 unsigned long flags;
2511 pr_debug("raid456: error called\n");
2513 spin_lock_irqsave(&conf->device_lock, flags);
2514 clear_bit(In_sync, &rdev->flags);
2515 mddev->degraded = calc_degraded(conf);
2516 spin_unlock_irqrestore(&conf->device_lock, flags);
2517 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2519 set_bit(Blocked, &rdev->flags);
2520 set_bit(Faulty, &rdev->flags);
2521 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2523 "md/raid:%s: Disk failure on %s, disabling device.\n"
2524 "md/raid:%s: Operation continuing on %d devices.\n",
2526 bdevname(rdev->bdev, b),
2528 conf->raid_disks - mddev->degraded);
2532 * Input: a 'big' sector number,
2533 * Output: index of the data and parity disk, and the sector # in them.
2535 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2536 int previous, int *dd_idx,
2537 struct stripe_head *sh)
2539 sector_t stripe, stripe2;
2540 sector_t chunk_number;
2541 unsigned int chunk_offset;
2544 sector_t new_sector;
2545 int algorithm = previous ? conf->prev_algo
2547 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2548 : conf->chunk_sectors;
2549 int raid_disks = previous ? conf->previous_raid_disks
2551 int data_disks = raid_disks - conf->max_degraded;
2553 /* First compute the information on this sector */
2556 * Compute the chunk number and the sector offset inside the chunk
2558 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2559 chunk_number = r_sector;
2562 * Compute the stripe number
2564 stripe = chunk_number;
2565 *dd_idx = sector_div(stripe, data_disks);
2568 * Select the parity disk based on the user selected algorithm.
2570 pd_idx = qd_idx = -1;
2571 switch(conf->level) {
2573 pd_idx = data_disks;
2576 switch (algorithm) {
2577 case ALGORITHM_LEFT_ASYMMETRIC:
2578 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2579 if (*dd_idx >= pd_idx)
2582 case ALGORITHM_RIGHT_ASYMMETRIC:
2583 pd_idx = sector_div(stripe2, raid_disks);
2584 if (*dd_idx >= pd_idx)
2587 case ALGORITHM_LEFT_SYMMETRIC:
2588 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2589 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2591 case ALGORITHM_RIGHT_SYMMETRIC:
2592 pd_idx = sector_div(stripe2, raid_disks);
2593 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2595 case ALGORITHM_PARITY_0:
2599 case ALGORITHM_PARITY_N:
2600 pd_idx = data_disks;
2608 switch (algorithm) {
2609 case ALGORITHM_LEFT_ASYMMETRIC:
2610 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2611 qd_idx = pd_idx + 1;
2612 if (pd_idx == raid_disks-1) {
2613 (*dd_idx)++; /* Q D D D P */
2615 } else if (*dd_idx >= pd_idx)
2616 (*dd_idx) += 2; /* D D P Q D */
2618 case ALGORITHM_RIGHT_ASYMMETRIC:
2619 pd_idx = sector_div(stripe2, raid_disks);
2620 qd_idx = pd_idx + 1;
2621 if (pd_idx == raid_disks-1) {
2622 (*dd_idx)++; /* Q D D D P */
2624 } else if (*dd_idx >= pd_idx)
2625 (*dd_idx) += 2; /* D D P Q D */
2627 case ALGORITHM_LEFT_SYMMETRIC:
2628 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2629 qd_idx = (pd_idx + 1) % raid_disks;
2630 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2632 case ALGORITHM_RIGHT_SYMMETRIC:
2633 pd_idx = sector_div(stripe2, raid_disks);
2634 qd_idx = (pd_idx + 1) % raid_disks;
2635 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2638 case ALGORITHM_PARITY_0:
2643 case ALGORITHM_PARITY_N:
2644 pd_idx = data_disks;
2645 qd_idx = data_disks + 1;
2648 case ALGORITHM_ROTATING_ZERO_RESTART:
2649 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2650 * of blocks for computing Q is different.
2652 pd_idx = sector_div(stripe2, raid_disks);
2653 qd_idx = pd_idx + 1;
2654 if (pd_idx == raid_disks-1) {
2655 (*dd_idx)++; /* Q D D D P */
2657 } else if (*dd_idx >= pd_idx)
2658 (*dd_idx) += 2; /* D D P Q D */
2662 case ALGORITHM_ROTATING_N_RESTART:
2663 /* Same a left_asymmetric, by first stripe is
2664 * D D D P Q rather than
2668 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2669 qd_idx = pd_idx + 1;
2670 if (pd_idx == raid_disks-1) {
2671 (*dd_idx)++; /* Q D D D P */
2673 } else if (*dd_idx >= pd_idx)
2674 (*dd_idx) += 2; /* D D P Q D */
2678 case ALGORITHM_ROTATING_N_CONTINUE:
2679 /* Same as left_symmetric but Q is before P */
2680 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2681 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2682 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2686 case ALGORITHM_LEFT_ASYMMETRIC_6:
2687 /* RAID5 left_asymmetric, with Q on last device */
2688 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2689 if (*dd_idx >= pd_idx)
2691 qd_idx = raid_disks - 1;
2694 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2695 pd_idx = sector_div(stripe2, raid_disks-1);
2696 if (*dd_idx >= pd_idx)
2698 qd_idx = raid_disks - 1;
2701 case ALGORITHM_LEFT_SYMMETRIC_6:
2702 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2703 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2704 qd_idx = raid_disks - 1;
2707 case ALGORITHM_RIGHT_SYMMETRIC_6:
2708 pd_idx = sector_div(stripe2, raid_disks-1);
2709 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2710 qd_idx = raid_disks - 1;
2713 case ALGORITHM_PARITY_0_6:
2716 qd_idx = raid_disks - 1;
2726 sh->pd_idx = pd_idx;
2727 sh->qd_idx = qd_idx;
2728 sh->ddf_layout = ddf_layout;
2731 * Finally, compute the new sector number
2733 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2737 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2739 struct r5conf *conf = sh->raid_conf;
2740 int raid_disks = sh->disks;
2741 int data_disks = raid_disks - conf->max_degraded;
2742 sector_t new_sector = sh->sector, check;
2743 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2744 : conf->chunk_sectors;
2745 int algorithm = previous ? conf->prev_algo
2749 sector_t chunk_number;
2750 int dummy1, dd_idx = i;
2752 struct stripe_head sh2;
2754 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2755 stripe = new_sector;
2757 if (i == sh->pd_idx)
2759 switch(conf->level) {
2762 switch (algorithm) {
2763 case ALGORITHM_LEFT_ASYMMETRIC:
2764 case ALGORITHM_RIGHT_ASYMMETRIC:
2768 case ALGORITHM_LEFT_SYMMETRIC:
2769 case ALGORITHM_RIGHT_SYMMETRIC:
2772 i -= (sh->pd_idx + 1);
2774 case ALGORITHM_PARITY_0:
2777 case ALGORITHM_PARITY_N:
2784 if (i == sh->qd_idx)
2785 return 0; /* It is the Q disk */
2786 switch (algorithm) {
2787 case ALGORITHM_LEFT_ASYMMETRIC:
2788 case ALGORITHM_RIGHT_ASYMMETRIC:
2789 case ALGORITHM_ROTATING_ZERO_RESTART:
2790 case ALGORITHM_ROTATING_N_RESTART:
2791 if (sh->pd_idx == raid_disks-1)
2792 i--; /* Q D D D P */
2793 else if (i > sh->pd_idx)
2794 i -= 2; /* D D P Q D */
2796 case ALGORITHM_LEFT_SYMMETRIC:
2797 case ALGORITHM_RIGHT_SYMMETRIC:
2798 if (sh->pd_idx == raid_disks-1)
2799 i--; /* Q D D D P */
2804 i -= (sh->pd_idx + 2);
2807 case ALGORITHM_PARITY_0:
2810 case ALGORITHM_PARITY_N:
2812 case ALGORITHM_ROTATING_N_CONTINUE:
2813 /* Like left_symmetric, but P is before Q */
2814 if (sh->pd_idx == 0)
2815 i--; /* P D D D Q */
2820 i -= (sh->pd_idx + 1);
2823 case ALGORITHM_LEFT_ASYMMETRIC_6:
2824 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2828 case ALGORITHM_LEFT_SYMMETRIC_6:
2829 case ALGORITHM_RIGHT_SYMMETRIC_6:
2831 i += data_disks + 1;
2832 i -= (sh->pd_idx + 1);
2834 case ALGORITHM_PARITY_0_6:
2843 chunk_number = stripe * data_disks + i;
2844 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2846 check = raid5_compute_sector(conf, r_sector,
2847 previous, &dummy1, &sh2);
2848 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2849 || sh2.qd_idx != sh->qd_idx) {
2850 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2851 mdname(conf->mddev));
2858 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2859 int rcw, int expand)
2861 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2862 struct r5conf *conf = sh->raid_conf;
2863 int level = conf->level;
2867 for (i = disks; i--; ) {
2868 struct r5dev *dev = &sh->dev[i];
2871 set_bit(R5_LOCKED, &dev->flags);
2872 set_bit(R5_Wantdrain, &dev->flags);
2874 clear_bit(R5_UPTODATE, &dev->flags);
2878 /* if we are not expanding this is a proper write request, and
2879 * there will be bios with new data to be drained into the
2884 /* False alarm, nothing to do */
2886 sh->reconstruct_state = reconstruct_state_drain_run;
2887 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2889 sh->reconstruct_state = reconstruct_state_run;
2891 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2893 if (s->locked + conf->max_degraded == disks)
2894 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2895 atomic_inc(&conf->pending_full_writes);
2897 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2898 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2899 BUG_ON(level == 6 &&
2900 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2901 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2903 for (i = disks; i--; ) {
2904 struct r5dev *dev = &sh->dev[i];
2905 if (i == pd_idx || i == qd_idx)
2909 (test_bit(R5_UPTODATE, &dev->flags) ||
2910 test_bit(R5_Wantcompute, &dev->flags))) {
2911 set_bit(R5_Wantdrain, &dev->flags);
2912 set_bit(R5_LOCKED, &dev->flags);
2913 clear_bit(R5_UPTODATE, &dev->flags);
2918 /* False alarm - nothing to do */
2920 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2921 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2922 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2923 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2926 /* keep the parity disk(s) locked while asynchronous operations
2929 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2930 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2934 int qd_idx = sh->qd_idx;
2935 struct r5dev *dev = &sh->dev[qd_idx];
2937 set_bit(R5_LOCKED, &dev->flags);
2938 clear_bit(R5_UPTODATE, &dev->flags);
2942 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2943 __func__, (unsigned long long)sh->sector,
2944 s->locked, s->ops_request);
2948 * Each stripe/dev can have one or more bion attached.
2949 * toread/towrite point to the first in a chain.
2950 * The bi_next chain must be in order.
2952 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2953 int forwrite, int previous)
2956 struct r5conf *conf = sh->raid_conf;
2959 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2960 (unsigned long long)bi->bi_iter.bi_sector,
2961 (unsigned long long)sh->sector);
2964 * If several bio share a stripe. The bio bi_phys_segments acts as a
2965 * reference count to avoid race. The reference count should already be
2966 * increased before this function is called (for example, in
2967 * make_request()), so other bio sharing this stripe will not free the
2968 * stripe. If a stripe is owned by one stripe, the stripe lock will
2971 spin_lock_irq(&sh->stripe_lock);
2972 /* Don't allow new IO added to stripes in batch list */
2976 bip = &sh->dev[dd_idx].towrite;
2980 bip = &sh->dev[dd_idx].toread;
2981 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2982 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2984 bip = & (*bip)->bi_next;
2986 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2989 if (!forwrite || previous)
2990 clear_bit(STRIPE_BATCH_READY, &sh->state);
2992 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2996 raid5_inc_bi_active_stripes(bi);
2999 /* check if page is covered */
3000 sector_t sector = sh->dev[dd_idx].sector;
3001 for (bi=sh->dev[dd_idx].towrite;
3002 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3003 bi && bi->bi_iter.bi_sector <= sector;
3004 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3005 if (bio_end_sector(bi) >= sector)
3006 sector = bio_end_sector(bi);
3008 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3009 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3010 sh->overwrite_disks++;
3013 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3014 (unsigned long long)(*bip)->bi_iter.bi_sector,
3015 (unsigned long long)sh->sector, dd_idx);
3017 if (conf->mddev->bitmap && firstwrite) {
3018 /* Cannot hold spinlock over bitmap_startwrite,
3019 * but must ensure this isn't added to a batch until
3020 * we have added to the bitmap and set bm_seq.
3021 * So set STRIPE_BITMAP_PENDING to prevent
3023 * If multiple add_stripe_bio() calls race here they
3024 * much all set STRIPE_BITMAP_PENDING. So only the first one
3025 * to complete "bitmap_startwrite" gets to set
3026 * STRIPE_BIT_DELAY. This is important as once a stripe
3027 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3030 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3031 spin_unlock_irq(&sh->stripe_lock);
3032 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3034 spin_lock_irq(&sh->stripe_lock);
3035 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3036 if (!sh->batch_head) {
3037 sh->bm_seq = conf->seq_flush+1;
3038 set_bit(STRIPE_BIT_DELAY, &sh->state);
3041 spin_unlock_irq(&sh->stripe_lock);
3043 if (stripe_can_batch(sh))
3044 stripe_add_to_batch_list(conf, sh);
3048 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3049 spin_unlock_irq(&sh->stripe_lock);
3053 static void end_reshape(struct r5conf *conf);
3055 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3056 struct stripe_head *sh)
3058 int sectors_per_chunk =
3059 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3061 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3062 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3064 raid5_compute_sector(conf,
3065 stripe * (disks - conf->max_degraded)
3066 *sectors_per_chunk + chunk_offset,
3072 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3073 struct stripe_head_state *s, int disks,
3074 struct bio **return_bi)
3077 BUG_ON(sh->batch_head);
3078 for (i = disks; i--; ) {
3082 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3083 struct md_rdev *rdev;
3085 rdev = rcu_dereference(conf->disks[i].rdev);
3086 if (rdev && test_bit(In_sync, &rdev->flags))
3087 atomic_inc(&rdev->nr_pending);
3092 if (!rdev_set_badblocks(
3096 md_error(conf->mddev, rdev);
3097 rdev_dec_pending(rdev, conf->mddev);
3100 spin_lock_irq(&sh->stripe_lock);
3101 /* fail all writes first */
3102 bi = sh->dev[i].towrite;
3103 sh->dev[i].towrite = NULL;
3104 sh->overwrite_disks = 0;
3105 spin_unlock_irq(&sh->stripe_lock);
3109 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3110 wake_up(&conf->wait_for_overlap);
3112 while (bi && bi->bi_iter.bi_sector <
3113 sh->dev[i].sector + STRIPE_SECTORS) {
3114 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3115 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3116 if (!raid5_dec_bi_active_stripes(bi)) {
3117 md_write_end(conf->mddev);
3118 bi->bi_next = *return_bi;
3124 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3125 STRIPE_SECTORS, 0, 0);
3127 /* and fail all 'written' */
3128 bi = sh->dev[i].written;
3129 sh->dev[i].written = NULL;
3130 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3131 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3132 sh->dev[i].page = sh->dev[i].orig_page;
3135 if (bi) bitmap_end = 1;
3136 while (bi && bi->bi_iter.bi_sector <
3137 sh->dev[i].sector + STRIPE_SECTORS) {
3138 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3139 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3140 if (!raid5_dec_bi_active_stripes(bi)) {
3141 md_write_end(conf->mddev);
3142 bi->bi_next = *return_bi;
3148 /* fail any reads if this device is non-operational and
3149 * the data has not reached the cache yet.
3151 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3152 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3153 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3154 spin_lock_irq(&sh->stripe_lock);
3155 bi = sh->dev[i].toread;
3156 sh->dev[i].toread = NULL;
3157 spin_unlock_irq(&sh->stripe_lock);
3158 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3159 wake_up(&conf->wait_for_overlap);
3160 while (bi && bi->bi_iter.bi_sector <
3161 sh->dev[i].sector + STRIPE_SECTORS) {
3162 struct bio *nextbi =
3163 r5_next_bio(bi, sh->dev[i].sector);
3164 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3165 if (!raid5_dec_bi_active_stripes(bi)) {
3166 bi->bi_next = *return_bi;
3173 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3174 STRIPE_SECTORS, 0, 0);
3175 /* If we were in the middle of a write the parity block might
3176 * still be locked - so just clear all R5_LOCKED flags
3178 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3181 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3182 if (atomic_dec_and_test(&conf->pending_full_writes))
3183 md_wakeup_thread(conf->mddev->thread);
3187 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3188 struct stripe_head_state *s)
3193 BUG_ON(sh->batch_head);
3194 clear_bit(STRIPE_SYNCING, &sh->state);
3195 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3196 wake_up(&conf->wait_for_overlap);
3199 /* There is nothing more to do for sync/check/repair.
3200 * Don't even need to abort as that is handled elsewhere
3201 * if needed, and not always wanted e.g. if there is a known
3203 * For recover/replace we need to record a bad block on all
3204 * non-sync devices, or abort the recovery
3206 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3207 /* During recovery devices cannot be removed, so
3208 * locking and refcounting of rdevs is not needed
3210 for (i = 0; i < conf->raid_disks; i++) {
3211 struct md_rdev *rdev = conf->disks[i].rdev;
3213 && !test_bit(Faulty, &rdev->flags)
3214 && !test_bit(In_sync, &rdev->flags)
3215 && !rdev_set_badblocks(rdev, sh->sector,
3218 rdev = conf->disks[i].replacement;
3220 && !test_bit(Faulty, &rdev->flags)
3221 && !test_bit(In_sync, &rdev->flags)
3222 && !rdev_set_badblocks(rdev, sh->sector,
3227 conf->recovery_disabled =
3228 conf->mddev->recovery_disabled;
3230 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3233 static int want_replace(struct stripe_head *sh, int disk_idx)
3235 struct md_rdev *rdev;
3237 /* Doing recovery so rcu locking not required */
3238 rdev = sh->raid_conf->disks[disk_idx].replacement;
3240 && !test_bit(Faulty, &rdev->flags)
3241 && !test_bit(In_sync, &rdev->flags)
3242 && (rdev->recovery_offset <= sh->sector
3243 || rdev->mddev->recovery_cp <= sh->sector))
3249 /* fetch_block - checks the given member device to see if its data needs
3250 * to be read or computed to satisfy a request.
3252 * Returns 1 when no more member devices need to be checked, otherwise returns
3253 * 0 to tell the loop in handle_stripe_fill to continue
3256 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3257 int disk_idx, int disks)
3259 struct r5dev *dev = &sh->dev[disk_idx];
3260 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3261 &sh->dev[s->failed_num[1]] };
3265 if (test_bit(R5_LOCKED, &dev->flags) ||
3266 test_bit(R5_UPTODATE, &dev->flags))
3267 /* No point reading this as we already have it or have
3268 * decided to get it.
3273 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3274 /* We need this block to directly satisfy a request */
3277 if (s->syncing || s->expanding ||
3278 (s->replacing && want_replace(sh, disk_idx)))
3279 /* When syncing, or expanding we read everything.
3280 * When replacing, we need the replaced block.
3284 if ((s->failed >= 1 && fdev[0]->toread) ||
3285 (s->failed >= 2 && fdev[1]->toread))
3286 /* If we want to read from a failed device, then
3287 * we need to actually read every other device.
3291 /* Sometimes neither read-modify-write nor reconstruct-write
3292 * cycles can work. In those cases we read every block we
3293 * can. Then the parity-update is certain to have enough to
3295 * This can only be a problem when we need to write something,
3296 * and some device has failed. If either of those tests
3297 * fail we need look no further.
3299 if (!s->failed || !s->to_write)
3302 if (test_bit(R5_Insync, &dev->flags) &&
3303 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3304 /* Pre-reads at not permitted until after short delay
3305 * to gather multiple requests. However if this
3306 * device is no Insync, the block could only be be computed
3307 * and there is no need to delay that.
3311 for (i = 0; i < s->failed; i++) {
3312 if (fdev[i]->towrite &&
3313 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3314 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3315 /* If we have a partial write to a failed
3316 * device, then we will need to reconstruct
3317 * the content of that device, so all other
3318 * devices must be read.
3323 /* If we are forced to do a reconstruct-write, either because
3324 * the current RAID6 implementation only supports that, or
3325 * or because parity cannot be trusted and we are currently
3326 * recovering it, there is extra need to be careful.
3327 * If one of the devices that we would need to read, because
3328 * it is not being overwritten (and maybe not written at all)
3329 * is missing/faulty, then we need to read everything we can.
3331 if (sh->raid_conf->level != 6 &&
3332 sh->sector < sh->raid_conf->mddev->recovery_cp)
3333 /* reconstruct-write isn't being forced */
3335 for (i = 0; i < s->failed; i++) {
3336 if (s->failed_num[i] != sh->pd_idx &&
3337 s->failed_num[i] != sh->qd_idx &&
3338 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3339 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3346 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3347 int disk_idx, int disks)
3349 struct r5dev *dev = &sh->dev[disk_idx];
3351 /* is the data in this block needed, and can we get it? */
3352 if (need_this_block(sh, s, disk_idx, disks)) {
3353 /* we would like to get this block, possibly by computing it,
3354 * otherwise read it if the backing disk is insync
3356 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3357 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3358 BUG_ON(sh->batch_head);
3359 if ((s->uptodate == disks - 1) &&
3360 (s->failed && (disk_idx == s->failed_num[0] ||
3361 disk_idx == s->failed_num[1]))) {
3362 /* have disk failed, and we're requested to fetch it;
3365 pr_debug("Computing stripe %llu block %d\n",
3366 (unsigned long long)sh->sector, disk_idx);
3367 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3368 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3369 set_bit(R5_Wantcompute, &dev->flags);
3370 sh->ops.target = disk_idx;
3371 sh->ops.target2 = -1; /* no 2nd target */
3373 /* Careful: from this point on 'uptodate' is in the eye
3374 * of raid_run_ops which services 'compute' operations
3375 * before writes. R5_Wantcompute flags a block that will
3376 * be R5_UPTODATE by the time it is needed for a
3377 * subsequent operation.
3381 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3382 /* Computing 2-failure is *very* expensive; only
3383 * do it if failed >= 2
3386 for (other = disks; other--; ) {
3387 if (other == disk_idx)
3389 if (!test_bit(R5_UPTODATE,
3390 &sh->dev[other].flags))
3394 pr_debug("Computing stripe %llu blocks %d,%d\n",
3395 (unsigned long long)sh->sector,
3397 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3398 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3399 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3400 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3401 sh->ops.target = disk_idx;
3402 sh->ops.target2 = other;
3406 } else if (test_bit(R5_Insync, &dev->flags)) {
3407 set_bit(R5_LOCKED, &dev->flags);
3408 set_bit(R5_Wantread, &dev->flags);
3410 pr_debug("Reading block %d (sync=%d)\n",
3411 disk_idx, s->syncing);
3419 * handle_stripe_fill - read or compute data to satisfy pending requests.
3421 static void handle_stripe_fill(struct stripe_head *sh,
3422 struct stripe_head_state *s,
3427 /* look for blocks to read/compute, skip this if a compute
3428 * is already in flight, or if the stripe contents are in the
3429 * midst of changing due to a write
3431 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3432 !sh->reconstruct_state)
3433 for (i = disks; i--; )
3434 if (fetch_block(sh, s, i, disks))
3436 set_bit(STRIPE_HANDLE, &sh->state);
3439 static void break_stripe_batch_list(struct stripe_head *head_sh,
3440 unsigned long handle_flags);
3441 /* handle_stripe_clean_event
3442 * any written block on an uptodate or failed drive can be returned.
3443 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3444 * never LOCKED, so we don't need to test 'failed' directly.
3446 static void handle_stripe_clean_event(struct r5conf *conf,
3447 struct stripe_head *sh, int disks, struct bio **return_bi)
3451 int discard_pending = 0;
3452 struct stripe_head *head_sh = sh;
3453 bool do_endio = false;
3455 for (i = disks; i--; )
3456 if (sh->dev[i].written) {
3458 if (!test_bit(R5_LOCKED, &dev->flags) &&
3459 (test_bit(R5_UPTODATE, &dev->flags) ||
3460 test_bit(R5_Discard, &dev->flags) ||
3461 test_bit(R5_SkipCopy, &dev->flags))) {
3462 /* We can return any write requests */
3463 struct bio *wbi, *wbi2;
3464 pr_debug("Return write for disc %d\n", i);
3465 if (test_and_clear_bit(R5_Discard, &dev->flags))
3466 clear_bit(R5_UPTODATE, &dev->flags);
3467 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3468 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3473 dev->page = dev->orig_page;
3475 dev->written = NULL;
3476 while (wbi && wbi->bi_iter.bi_sector <
3477 dev->sector + STRIPE_SECTORS) {
3478 wbi2 = r5_next_bio(wbi, dev->sector);
3479 if (!raid5_dec_bi_active_stripes(wbi)) {
3480 md_write_end(conf->mddev);
3481 wbi->bi_next = *return_bi;
3486 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3488 !test_bit(STRIPE_DEGRADED, &sh->state),
3490 if (head_sh->batch_head) {
3491 sh = list_first_entry(&sh->batch_list,
3494 if (sh != head_sh) {
3501 } else if (test_bit(R5_Discard, &dev->flags))
3502 discard_pending = 1;
3503 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3504 WARN_ON(dev->page != dev->orig_page);
3506 if (!discard_pending &&
3507 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3508 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3509 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3510 if (sh->qd_idx >= 0) {
3511 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3512 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3514 /* now that discard is done we can proceed with any sync */
3515 clear_bit(STRIPE_DISCARD, &sh->state);
3517 * SCSI discard will change some bio fields and the stripe has
3518 * no updated data, so remove it from hash list and the stripe
3519 * will be reinitialized
3521 spin_lock_irq(&conf->device_lock);
3524 if (head_sh->batch_head) {
3525 sh = list_first_entry(&sh->batch_list,
3526 struct stripe_head, batch_list);
3530 spin_unlock_irq(&conf->device_lock);
3533 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3534 set_bit(STRIPE_HANDLE, &sh->state);
3538 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3539 if (atomic_dec_and_test(&conf->pending_full_writes))
3540 md_wakeup_thread(conf->mddev->thread);
3542 if (head_sh->batch_head && do_endio)
3543 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3546 static void handle_stripe_dirtying(struct r5conf *conf,
3547 struct stripe_head *sh,
3548 struct stripe_head_state *s,
3551 int rmw = 0, rcw = 0, i;
3552 sector_t recovery_cp = conf->mddev->recovery_cp;
3554 /* Check whether resync is now happening or should start.
3555 * If yes, then the array is dirty (after unclean shutdown or
3556 * initial creation), so parity in some stripes might be inconsistent.
3557 * In this case, we need to always do reconstruct-write, to ensure
3558 * that in case of drive failure or read-error correction, we
3559 * generate correct data from the parity.
3561 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3562 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3564 /* Calculate the real rcw later - for now make it
3565 * look like rcw is cheaper
3568 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3569 conf->rmw_level, (unsigned long long)recovery_cp,
3570 (unsigned long long)sh->sector);
3571 } else for (i = disks; i--; ) {
3572 /* would I have to read this buffer for read_modify_write */
3573 struct r5dev *dev = &sh->dev[i];
3574 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3575 !test_bit(R5_LOCKED, &dev->flags) &&
3576 !(test_bit(R5_UPTODATE, &dev->flags) ||
3577 test_bit(R5_Wantcompute, &dev->flags))) {
3578 if (test_bit(R5_Insync, &dev->flags))
3581 rmw += 2*disks; /* cannot read it */
3583 /* Would I have to read this buffer for reconstruct_write */
3584 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3585 i != sh->pd_idx && i != sh->qd_idx &&
3586 !test_bit(R5_LOCKED, &dev->flags) &&
3587 !(test_bit(R5_UPTODATE, &dev->flags) ||
3588 test_bit(R5_Wantcompute, &dev->flags))) {
3589 if (test_bit(R5_Insync, &dev->flags))
3595 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3596 (unsigned long long)sh->sector, rmw, rcw);
3597 set_bit(STRIPE_HANDLE, &sh->state);
3598 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3599 /* prefer read-modify-write, but need to get some data */
3600 if (conf->mddev->queue)
3601 blk_add_trace_msg(conf->mddev->queue,
3602 "raid5 rmw %llu %d",
3603 (unsigned long long)sh->sector, rmw);
3604 for (i = disks; i--; ) {
3605 struct r5dev *dev = &sh->dev[i];
3606 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3607 !test_bit(R5_LOCKED, &dev->flags) &&
3608 !(test_bit(R5_UPTODATE, &dev->flags) ||
3609 test_bit(R5_Wantcompute, &dev->flags)) &&
3610 test_bit(R5_Insync, &dev->flags)) {
3611 if (test_bit(STRIPE_PREREAD_ACTIVE,
3613 pr_debug("Read_old block %d for r-m-w\n",
3615 set_bit(R5_LOCKED, &dev->flags);
3616 set_bit(R5_Wantread, &dev->flags);
3619 set_bit(STRIPE_DELAYED, &sh->state);
3620 set_bit(STRIPE_HANDLE, &sh->state);
3625 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3626 /* want reconstruct write, but need to get some data */
3629 for (i = disks; i--; ) {
3630 struct r5dev *dev = &sh->dev[i];
3631 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3632 i != sh->pd_idx && i != sh->qd_idx &&
3633 !test_bit(R5_LOCKED, &dev->flags) &&
3634 !(test_bit(R5_UPTODATE, &dev->flags) ||
3635 test_bit(R5_Wantcompute, &dev->flags))) {
3637 if (test_bit(R5_Insync, &dev->flags) &&
3638 test_bit(STRIPE_PREREAD_ACTIVE,
3640 pr_debug("Read_old block "
3641 "%d for Reconstruct\n", i);
3642 set_bit(R5_LOCKED, &dev->flags);
3643 set_bit(R5_Wantread, &dev->flags);
3647 set_bit(STRIPE_DELAYED, &sh->state);
3648 set_bit(STRIPE_HANDLE, &sh->state);
3652 if (rcw && conf->mddev->queue)
3653 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3654 (unsigned long long)sh->sector,
3655 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3658 if (rcw > disks && rmw > disks &&
3659 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3660 set_bit(STRIPE_DELAYED, &sh->state);
3662 /* now if nothing is locked, and if we have enough data,
3663 * we can start a write request
3665 /* since handle_stripe can be called at any time we need to handle the
3666 * case where a compute block operation has been submitted and then a
3667 * subsequent call wants to start a write request. raid_run_ops only
3668 * handles the case where compute block and reconstruct are requested
3669 * simultaneously. If this is not the case then new writes need to be
3670 * held off until the compute completes.
3672 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3673 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3674 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3675 schedule_reconstruction(sh, s, rcw == 0, 0);
3678 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3679 struct stripe_head_state *s, int disks)
3681 struct r5dev *dev = NULL;
3683 BUG_ON(sh->batch_head);
3684 set_bit(STRIPE_HANDLE, &sh->state);
3686 switch (sh->check_state) {
3687 case check_state_idle:
3688 /* start a new check operation if there are no failures */
3689 if (s->failed == 0) {
3690 BUG_ON(s->uptodate != disks);
3691 sh->check_state = check_state_run;
3692 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3693 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3697 dev = &sh->dev[s->failed_num[0]];
3699 case check_state_compute_result:
3700 sh->check_state = check_state_idle;
3702 dev = &sh->dev[sh->pd_idx];
3704 /* check that a write has not made the stripe insync */
3705 if (test_bit(STRIPE_INSYNC, &sh->state))
3708 /* either failed parity check, or recovery is happening */
3709 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3710 BUG_ON(s->uptodate != disks);
3712 set_bit(R5_LOCKED, &dev->flags);
3714 set_bit(R5_Wantwrite, &dev->flags);
3716 clear_bit(STRIPE_DEGRADED, &sh->state);
3717 set_bit(STRIPE_INSYNC, &sh->state);
3719 case check_state_run:
3720 break; /* we will be called again upon completion */
3721 case check_state_check_result:
3722 sh->check_state = check_state_idle;
3724 /* if a failure occurred during the check operation, leave
3725 * STRIPE_INSYNC not set and let the stripe be handled again
3730 /* handle a successful check operation, if parity is correct
3731 * we are done. Otherwise update the mismatch count and repair
3732 * parity if !MD_RECOVERY_CHECK
3734 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3735 /* parity is correct (on disc,
3736 * not in buffer any more)
3738 set_bit(STRIPE_INSYNC, &sh->state);
3740 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3741 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3742 /* don't try to repair!! */
3743 set_bit(STRIPE_INSYNC, &sh->state);
3745 sh->check_state = check_state_compute_run;
3746 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3747 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3748 set_bit(R5_Wantcompute,
3749 &sh->dev[sh->pd_idx].flags);
3750 sh->ops.target = sh->pd_idx;
3751 sh->ops.target2 = -1;
3756 case check_state_compute_run:
3759 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3760 __func__, sh->check_state,
3761 (unsigned long long) sh->sector);
3766 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3767 struct stripe_head_state *s,
3770 int pd_idx = sh->pd_idx;
3771 int qd_idx = sh->qd_idx;
3774 BUG_ON(sh->batch_head);
3775 set_bit(STRIPE_HANDLE, &sh->state);
3777 BUG_ON(s->failed > 2);
3779 /* Want to check and possibly repair P and Q.
3780 * However there could be one 'failed' device, in which
3781 * case we can only check one of them, possibly using the
3782 * other to generate missing data
3785 switch (sh->check_state) {
3786 case check_state_idle:
3787 /* start a new check operation if there are < 2 failures */
3788 if (s->failed == s->q_failed) {
3789 /* The only possible failed device holds Q, so it
3790 * makes sense to check P (If anything else were failed,
3791 * we would have used P to recreate it).
3793 sh->check_state = check_state_run;
3795 if (!s->q_failed && s->failed < 2) {
3796 /* Q is not failed, and we didn't use it to generate
3797 * anything, so it makes sense to check it
3799 if (sh->check_state == check_state_run)
3800 sh->check_state = check_state_run_pq;
3802 sh->check_state = check_state_run_q;
3805 /* discard potentially stale zero_sum_result */
3806 sh->ops.zero_sum_result = 0;
3808 if (sh->check_state == check_state_run) {
3809 /* async_xor_zero_sum destroys the contents of P */
3810 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3813 if (sh->check_state >= check_state_run &&
3814 sh->check_state <= check_state_run_pq) {
3815 /* async_syndrome_zero_sum preserves P and Q, so
3816 * no need to mark them !uptodate here
3818 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3822 /* we have 2-disk failure */
3823 BUG_ON(s->failed != 2);
3825 case check_state_compute_result:
3826 sh->check_state = check_state_idle;
3828 /* check that a write has not made the stripe insync */
3829 if (test_bit(STRIPE_INSYNC, &sh->state))
3832 /* now write out any block on a failed drive,
3833 * or P or Q if they were recomputed
3835 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3836 if (s->failed == 2) {
3837 dev = &sh->dev[s->failed_num[1]];
3839 set_bit(R5_LOCKED, &dev->flags);
3840 set_bit(R5_Wantwrite, &dev->flags);
3842 if (s->failed >= 1) {
3843 dev = &sh->dev[s->failed_num[0]];
3845 set_bit(R5_LOCKED, &dev->flags);
3846 set_bit(R5_Wantwrite, &dev->flags);
3848 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3849 dev = &sh->dev[pd_idx];
3851 set_bit(R5_LOCKED, &dev->flags);
3852 set_bit(R5_Wantwrite, &dev->flags);
3854 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3855 dev = &sh->dev[qd_idx];
3857 set_bit(R5_LOCKED, &dev->flags);
3858 set_bit(R5_Wantwrite, &dev->flags);
3860 clear_bit(STRIPE_DEGRADED, &sh->state);
3862 set_bit(STRIPE_INSYNC, &sh->state);
3864 case check_state_run:
3865 case check_state_run_q:
3866 case check_state_run_pq:
3867 break; /* we will be called again upon completion */
3868 case check_state_check_result:
3869 sh->check_state = check_state_idle;
3871 /* handle a successful check operation, if parity is correct
3872 * we are done. Otherwise update the mismatch count and repair
3873 * parity if !MD_RECOVERY_CHECK
3875 if (sh->ops.zero_sum_result == 0) {
3876 /* both parities are correct */
3878 set_bit(STRIPE_INSYNC, &sh->state);
3880 /* in contrast to the raid5 case we can validate
3881 * parity, but still have a failure to write
3884 sh->check_state = check_state_compute_result;
3885 /* Returning at this point means that we may go
3886 * off and bring p and/or q uptodate again so
3887 * we make sure to check zero_sum_result again
3888 * to verify if p or q need writeback
3892 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3893 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3894 /* don't try to repair!! */
3895 set_bit(STRIPE_INSYNC, &sh->state);
3897 int *target = &sh->ops.target;
3899 sh->ops.target = -1;
3900 sh->ops.target2 = -1;
3901 sh->check_state = check_state_compute_run;
3902 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3903 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3904 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3905 set_bit(R5_Wantcompute,
3906 &sh->dev[pd_idx].flags);
3908 target = &sh->ops.target2;
3911 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3912 set_bit(R5_Wantcompute,
3913 &sh->dev[qd_idx].flags);
3920 case check_state_compute_run:
3923 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3924 __func__, sh->check_state,
3925 (unsigned long long) sh->sector);
3930 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3934 /* We have read all the blocks in this stripe and now we need to
3935 * copy some of them into a target stripe for expand.
3937 struct dma_async_tx_descriptor *tx = NULL;
3938 BUG_ON(sh->batch_head);
3939 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3940 for (i = 0; i < sh->disks; i++)
3941 if (i != sh->pd_idx && i != sh->qd_idx) {
3943 struct stripe_head *sh2;
3944 struct async_submit_ctl submit;
3946 sector_t bn = compute_blocknr(sh, i, 1);
3947 sector_t s = raid5_compute_sector(conf, bn, 0,
3949 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3951 /* so far only the early blocks of this stripe
3952 * have been requested. When later blocks
3953 * get requested, we will try again
3956 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3957 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3958 /* must have already done this block */
3959 release_stripe(sh2);
3963 /* place all the copies on one channel */
3964 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3965 tx = async_memcpy(sh2->dev[dd_idx].page,
3966 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3969 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3970 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3971 for (j = 0; j < conf->raid_disks; j++)
3972 if (j != sh2->pd_idx &&
3974 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3976 if (j == conf->raid_disks) {
3977 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3978 set_bit(STRIPE_HANDLE, &sh2->state);
3980 release_stripe(sh2);
3983 /* done submitting copies, wait for them to complete */
3984 async_tx_quiesce(&tx);
3988 * handle_stripe - do things to a stripe.
3990 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3991 * state of various bits to see what needs to be done.
3993 * return some read requests which now have data
3994 * return some write requests which are safely on storage
3995 * schedule a read on some buffers
3996 * schedule a write of some buffers
3997 * return confirmation of parity correctness
4001 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4003 struct r5conf *conf = sh->raid_conf;
4004 int disks = sh->disks;
4007 int do_recovery = 0;
4009 memset(s, 0, sizeof(*s));
4011 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4012 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4013 s->failed_num[0] = -1;
4014 s->failed_num[1] = -1;
4016 /* Now to look around and see what can be done */
4018 for (i=disks; i--; ) {
4019 struct md_rdev *rdev;
4026 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4028 dev->toread, dev->towrite, dev->written);
4029 /* maybe we can reply to a read
4031 * new wantfill requests are only permitted while
4032 * ops_complete_biofill is guaranteed to be inactive
4034 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4035 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4036 set_bit(R5_Wantfill, &dev->flags);
4038 /* now count some things */
4039 if (test_bit(R5_LOCKED, &dev->flags))
4041 if (test_bit(R5_UPTODATE, &dev->flags))
4043 if (test_bit(R5_Wantcompute, &dev->flags)) {
4045 BUG_ON(s->compute > 2);
4048 if (test_bit(R5_Wantfill, &dev->flags))
4050 else if (dev->toread)
4054 if (!test_bit(R5_OVERWRITE, &dev->flags))
4059 /* Prefer to use the replacement for reads, but only
4060 * if it is recovered enough and has no bad blocks.
4062 rdev = rcu_dereference(conf->disks[i].replacement);
4063 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4064 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4065 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4066 &first_bad, &bad_sectors))
4067 set_bit(R5_ReadRepl, &dev->flags);
4069 if (rdev && !test_bit(Faulty, &rdev->flags))
4070 set_bit(R5_NeedReplace, &dev->flags);
4072 clear_bit(R5_NeedReplace, &dev->flags);
4073 rdev = rcu_dereference(conf->disks[i].rdev);
4074 clear_bit(R5_ReadRepl, &dev->flags);
4076 if (rdev && test_bit(Faulty, &rdev->flags))
4079 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4080 &first_bad, &bad_sectors);
4081 if (s->blocked_rdev == NULL
4082 && (test_bit(Blocked, &rdev->flags)
4085 set_bit(BlockedBadBlocks,
4087 s->blocked_rdev = rdev;
4088 atomic_inc(&rdev->nr_pending);
4091 clear_bit(R5_Insync, &dev->flags);
4095 /* also not in-sync */
4096 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4097 test_bit(R5_UPTODATE, &dev->flags)) {
4098 /* treat as in-sync, but with a read error
4099 * which we can now try to correct
4101 set_bit(R5_Insync, &dev->flags);
4102 set_bit(R5_ReadError, &dev->flags);
4104 } else if (test_bit(In_sync, &rdev->flags))
4105 set_bit(R5_Insync, &dev->flags);
4106 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4107 /* in sync if before recovery_offset */
4108 set_bit(R5_Insync, &dev->flags);
4109 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4110 test_bit(R5_Expanded, &dev->flags))
4111 /* If we've reshaped into here, we assume it is Insync.
4112 * We will shortly update recovery_offset to make
4115 set_bit(R5_Insync, &dev->flags);
4117 if (test_bit(R5_WriteError, &dev->flags)) {
4118 /* This flag does not apply to '.replacement'
4119 * only to .rdev, so make sure to check that*/
4120 struct md_rdev *rdev2 = rcu_dereference(
4121 conf->disks[i].rdev);
4123 clear_bit(R5_Insync, &dev->flags);
4124 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4125 s->handle_bad_blocks = 1;
4126 atomic_inc(&rdev2->nr_pending);
4128 clear_bit(R5_WriteError, &dev->flags);
4130 if (test_bit(R5_MadeGood, &dev->flags)) {
4131 /* This flag does not apply to '.replacement'
4132 * only to .rdev, so make sure to check that*/
4133 struct md_rdev *rdev2 = rcu_dereference(
4134 conf->disks[i].rdev);
4135 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4136 s->handle_bad_blocks = 1;
4137 atomic_inc(&rdev2->nr_pending);
4139 clear_bit(R5_MadeGood, &dev->flags);
4141 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4142 struct md_rdev *rdev2 = rcu_dereference(
4143 conf->disks[i].replacement);
4144 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4145 s->handle_bad_blocks = 1;
4146 atomic_inc(&rdev2->nr_pending);
4148 clear_bit(R5_MadeGoodRepl, &dev->flags);
4150 if (!test_bit(R5_Insync, &dev->flags)) {
4151 /* The ReadError flag will just be confusing now */
4152 clear_bit(R5_ReadError, &dev->flags);
4153 clear_bit(R5_ReWrite, &dev->flags);
4155 if (test_bit(R5_ReadError, &dev->flags))
4156 clear_bit(R5_Insync, &dev->flags);
4157 if (!test_bit(R5_Insync, &dev->flags)) {
4159 s->failed_num[s->failed] = i;
4161 if (rdev && !test_bit(Faulty, &rdev->flags))
4165 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4166 /* If there is a failed device being replaced,
4167 * we must be recovering.
4168 * else if we are after recovery_cp, we must be syncing
4169 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4170 * else we can only be replacing
4171 * sync and recovery both need to read all devices, and so
4172 * use the same flag.
4175 sh->sector >= conf->mddev->recovery_cp ||
4176 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4184 static int clear_batch_ready(struct stripe_head *sh)
4186 /* Return '1' if this is a member of batch, or
4187 * '0' if it is a lone stripe or a head which can now be
4190 struct stripe_head *tmp;
4191 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4192 return (sh->batch_head && sh->batch_head != sh);
4193 spin_lock(&sh->stripe_lock);
4194 if (!sh->batch_head) {
4195 spin_unlock(&sh->stripe_lock);
4200 * this stripe could be added to a batch list before we check
4201 * BATCH_READY, skips it
4203 if (sh->batch_head != sh) {
4204 spin_unlock(&sh->stripe_lock);
4207 spin_lock(&sh->batch_lock);
4208 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4209 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4210 spin_unlock(&sh->batch_lock);
4211 spin_unlock(&sh->stripe_lock);
4214 * BATCH_READY is cleared, no new stripes can be added.
4215 * batch_list can be accessed without lock
4220 static void break_stripe_batch_list(struct stripe_head *head_sh,
4221 unsigned long handle_flags)
4223 struct stripe_head *sh, *next;
4227 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4229 list_del_init(&sh->batch_list);
4231 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4232 (1 << STRIPE_SYNCING) |
4233 (1 << STRIPE_REPLACED) |
4234 (1 << STRIPE_PREREAD_ACTIVE) |
4235 (1 << STRIPE_DELAYED) |
4236 (1 << STRIPE_BIT_DELAY) |
4237 (1 << STRIPE_FULL_WRITE) |
4238 (1 << STRIPE_BIOFILL_RUN) |
4239 (1 << STRIPE_COMPUTE_RUN) |
4240 (1 << STRIPE_OPS_REQ_PENDING) |
4241 (1 << STRIPE_DISCARD) |
4242 (1 << STRIPE_BATCH_READY) |
4243 (1 << STRIPE_BATCH_ERR) |
4244 (1 << STRIPE_BITMAP_PENDING)));
4245 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4246 (1 << STRIPE_REPLACED)));
4248 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4249 (1 << STRIPE_DEGRADED)),
4250 head_sh->state & (1 << STRIPE_INSYNC));
4252 sh->check_state = head_sh->check_state;
4253 sh->reconstruct_state = head_sh->reconstruct_state;
4254 for (i = 0; i < sh->disks; i++) {
4255 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4257 sh->dev[i].flags = head_sh->dev[i].flags &
4258 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4260 spin_lock_irq(&sh->stripe_lock);
4261 sh->batch_head = NULL;
4262 spin_unlock_irq(&sh->stripe_lock);
4263 if (handle_flags == 0 ||
4264 sh->state & handle_flags)
4265 set_bit(STRIPE_HANDLE, &sh->state);
4268 spin_lock_irq(&head_sh->stripe_lock);
4269 head_sh->batch_head = NULL;
4270 spin_unlock_irq(&head_sh->stripe_lock);
4271 for (i = 0; i < head_sh->disks; i++)
4272 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4274 if (head_sh->state & handle_flags)
4275 set_bit(STRIPE_HANDLE, &head_sh->state);
4278 wake_up(&head_sh->raid_conf->wait_for_overlap);
4281 static void handle_stripe(struct stripe_head *sh)
4283 struct stripe_head_state s;
4284 struct r5conf *conf = sh->raid_conf;
4287 int disks = sh->disks;
4288 struct r5dev *pdev, *qdev;
4290 clear_bit(STRIPE_HANDLE, &sh->state);
4291 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4292 /* already being handled, ensure it gets handled
4293 * again when current action finishes */
4294 set_bit(STRIPE_HANDLE, &sh->state);
4298 if (clear_batch_ready(sh) ) {
4299 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4303 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4304 break_stripe_batch_list(sh, 0);
4306 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4307 spin_lock(&sh->stripe_lock);
4308 /* Cannot process 'sync' concurrently with 'discard' */
4309 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4310 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4311 set_bit(STRIPE_SYNCING, &sh->state);
4312 clear_bit(STRIPE_INSYNC, &sh->state);
4313 clear_bit(STRIPE_REPLACED, &sh->state);
4315 spin_unlock(&sh->stripe_lock);
4317 clear_bit(STRIPE_DELAYED, &sh->state);
4319 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4320 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4321 (unsigned long long)sh->sector, sh->state,
4322 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4323 sh->check_state, sh->reconstruct_state);
4325 analyse_stripe(sh, &s);
4327 if (s.handle_bad_blocks) {
4328 set_bit(STRIPE_HANDLE, &sh->state);
4332 if (unlikely(s.blocked_rdev)) {
4333 if (s.syncing || s.expanding || s.expanded ||
4334 s.replacing || s.to_write || s.written) {
4335 set_bit(STRIPE_HANDLE, &sh->state);
4338 /* There is nothing for the blocked_rdev to block */
4339 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4340 s.blocked_rdev = NULL;
4343 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4344 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4345 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4348 pr_debug("locked=%d uptodate=%d to_read=%d"
4349 " to_write=%d failed=%d failed_num=%d,%d\n",
4350 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4351 s.failed_num[0], s.failed_num[1]);
4352 /* check if the array has lost more than max_degraded devices and,
4353 * if so, some requests might need to be failed.
4355 if (s.failed > conf->max_degraded) {
4356 sh->check_state = 0;
4357 sh->reconstruct_state = 0;
4358 break_stripe_batch_list(sh, 0);
4359 if (s.to_read+s.to_write+s.written)
4360 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4361 if (s.syncing + s.replacing)
4362 handle_failed_sync(conf, sh, &s);
4365 /* Now we check to see if any write operations have recently
4369 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4371 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4372 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4373 sh->reconstruct_state = reconstruct_state_idle;
4375 /* All the 'written' buffers and the parity block are ready to
4376 * be written back to disk
4378 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4379 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4380 BUG_ON(sh->qd_idx >= 0 &&
4381 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4382 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4383 for (i = disks; i--; ) {
4384 struct r5dev *dev = &sh->dev[i];
4385 if (test_bit(R5_LOCKED, &dev->flags) &&
4386 (i == sh->pd_idx || i == sh->qd_idx ||
4388 pr_debug("Writing block %d\n", i);
4389 set_bit(R5_Wantwrite, &dev->flags);
4394 if (!test_bit(R5_Insync, &dev->flags) ||
4395 ((i == sh->pd_idx || i == sh->qd_idx) &&
4397 set_bit(STRIPE_INSYNC, &sh->state);
4400 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4401 s.dec_preread_active = 1;
4405 * might be able to return some write requests if the parity blocks
4406 * are safe, or on a failed drive
4408 pdev = &sh->dev[sh->pd_idx];
4409 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4410 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4411 qdev = &sh->dev[sh->qd_idx];
4412 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4413 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4417 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4418 && !test_bit(R5_LOCKED, &pdev->flags)
4419 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4420 test_bit(R5_Discard, &pdev->flags))))) &&
4421 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4422 && !test_bit(R5_LOCKED, &qdev->flags)
4423 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4424 test_bit(R5_Discard, &qdev->flags))))))
4425 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4427 /* Now we might consider reading some blocks, either to check/generate
4428 * parity, or to satisfy requests
4429 * or to load a block that is being partially written.
4431 if (s.to_read || s.non_overwrite
4432 || (conf->level == 6 && s.to_write && s.failed)
4433 || (s.syncing && (s.uptodate + s.compute < disks))
4436 handle_stripe_fill(sh, &s, disks);
4438 /* Now to consider new write requests and what else, if anything
4439 * should be read. We do not handle new writes when:
4440 * 1/ A 'write' operation (copy+xor) is already in flight.
4441 * 2/ A 'check' operation is in flight, as it may clobber the parity
4444 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4445 handle_stripe_dirtying(conf, sh, &s, disks);
4447 /* maybe we need to check and possibly fix the parity for this stripe
4448 * Any reads will already have been scheduled, so we just see if enough
4449 * data is available. The parity check is held off while parity
4450 * dependent operations are in flight.
4452 if (sh->check_state ||
4453 (s.syncing && s.locked == 0 &&
4454 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4455 !test_bit(STRIPE_INSYNC, &sh->state))) {
4456 if (conf->level == 6)
4457 handle_parity_checks6(conf, sh, &s, disks);
4459 handle_parity_checks5(conf, sh, &s, disks);
4462 if ((s.replacing || s.syncing) && s.locked == 0
4463 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4464 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4465 /* Write out to replacement devices where possible */
4466 for (i = 0; i < conf->raid_disks; i++)
4467 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4468 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4469 set_bit(R5_WantReplace, &sh->dev[i].flags);
4470 set_bit(R5_LOCKED, &sh->dev[i].flags);
4474 set_bit(STRIPE_INSYNC, &sh->state);
4475 set_bit(STRIPE_REPLACED, &sh->state);
4477 if ((s.syncing || s.replacing) && s.locked == 0 &&
4478 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4479 test_bit(STRIPE_INSYNC, &sh->state)) {
4480 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4481 clear_bit(STRIPE_SYNCING, &sh->state);
4482 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4483 wake_up(&conf->wait_for_overlap);
4486 /* If the failed drives are just a ReadError, then we might need
4487 * to progress the repair/check process
4489 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4490 for (i = 0; i < s.failed; i++) {
4491 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4492 if (test_bit(R5_ReadError, &dev->flags)
4493 && !test_bit(R5_LOCKED, &dev->flags)
4494 && test_bit(R5_UPTODATE, &dev->flags)
4496 if (!test_bit(R5_ReWrite, &dev->flags)) {
4497 set_bit(R5_Wantwrite, &dev->flags);
4498 set_bit(R5_ReWrite, &dev->flags);
4499 set_bit(R5_LOCKED, &dev->flags);
4502 /* let's read it back */
4503 set_bit(R5_Wantread, &dev->flags);
4504 set_bit(R5_LOCKED, &dev->flags);
4510 /* Finish reconstruct operations initiated by the expansion process */
4511 if (sh->reconstruct_state == reconstruct_state_result) {
4512 struct stripe_head *sh_src
4513 = get_active_stripe(conf, sh->sector, 1, 1, 1);
4514 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4515 /* sh cannot be written until sh_src has been read.
4516 * so arrange for sh to be delayed a little
4518 set_bit(STRIPE_DELAYED, &sh->state);
4519 set_bit(STRIPE_HANDLE, &sh->state);
4520 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4522 atomic_inc(&conf->preread_active_stripes);
4523 release_stripe(sh_src);
4527 release_stripe(sh_src);
4529 sh->reconstruct_state = reconstruct_state_idle;
4530 clear_bit(STRIPE_EXPANDING, &sh->state);
4531 for (i = conf->raid_disks; i--; ) {
4532 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4533 set_bit(R5_LOCKED, &sh->dev[i].flags);
4538 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4539 !sh->reconstruct_state) {
4540 /* Need to write out all blocks after computing parity */
4541 sh->disks = conf->raid_disks;
4542 stripe_set_idx(sh->sector, conf, 0, sh);
4543 schedule_reconstruction(sh, &s, 1, 1);
4544 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4545 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4546 atomic_dec(&conf->reshape_stripes);
4547 wake_up(&conf->wait_for_overlap);
4548 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4551 if (s.expanding && s.locked == 0 &&
4552 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4553 handle_stripe_expansion(conf, sh);
4556 /* wait for this device to become unblocked */
4557 if (unlikely(s.blocked_rdev)) {
4558 if (conf->mddev->external)
4559 md_wait_for_blocked_rdev(s.blocked_rdev,
4562 /* Internal metadata will immediately
4563 * be written by raid5d, so we don't
4564 * need to wait here.
4566 rdev_dec_pending(s.blocked_rdev,
4570 if (s.handle_bad_blocks)
4571 for (i = disks; i--; ) {
4572 struct md_rdev *rdev;
4573 struct r5dev *dev = &sh->dev[i];
4574 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4575 /* We own a safe reference to the rdev */
4576 rdev = conf->disks[i].rdev;
4577 if (!rdev_set_badblocks(rdev, sh->sector,
4579 md_error(conf->mddev, rdev);
4580 rdev_dec_pending(rdev, conf->mddev);
4582 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4583 rdev = conf->disks[i].rdev;
4584 rdev_clear_badblocks(rdev, sh->sector,
4586 rdev_dec_pending(rdev, conf->mddev);
4588 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4589 rdev = conf->disks[i].replacement;
4591 /* rdev have been moved down */
4592 rdev = conf->disks[i].rdev;
4593 rdev_clear_badblocks(rdev, sh->sector,
4595 rdev_dec_pending(rdev, conf->mddev);
4600 raid_run_ops(sh, s.ops_request);
4604 if (s.dec_preread_active) {
4605 /* We delay this until after ops_run_io so that if make_request
4606 * is waiting on a flush, it won't continue until the writes
4607 * have actually been submitted.
4609 atomic_dec(&conf->preread_active_stripes);
4610 if (atomic_read(&conf->preread_active_stripes) <
4612 md_wakeup_thread(conf->mddev->thread);
4615 return_io(s.return_bi);
4617 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4620 static void raid5_activate_delayed(struct r5conf *conf)
4622 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4623 while (!list_empty(&conf->delayed_list)) {
4624 struct list_head *l = conf->delayed_list.next;
4625 struct stripe_head *sh;
4626 sh = list_entry(l, struct stripe_head, lru);
4628 clear_bit(STRIPE_DELAYED, &sh->state);
4629 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4630 atomic_inc(&conf->preread_active_stripes);
4631 list_add_tail(&sh->lru, &conf->hold_list);
4632 raid5_wakeup_stripe_thread(sh);
4637 static void activate_bit_delay(struct r5conf *conf,
4638 struct list_head *temp_inactive_list)
4640 /* device_lock is held */
4641 struct list_head head;
4642 list_add(&head, &conf->bitmap_list);
4643 list_del_init(&conf->bitmap_list);
4644 while (!list_empty(&head)) {
4645 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4647 list_del_init(&sh->lru);
4648 atomic_inc(&sh->count);
4649 hash = sh->hash_lock_index;
4650 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4654 static int raid5_congested(struct mddev *mddev, int bits)
4656 struct r5conf *conf = mddev->private;
4658 /* No difference between reads and writes. Just check
4659 * how busy the stripe_cache is
4662 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4666 if (atomic_read(&conf->empty_inactive_list_nr))
4672 /* We want read requests to align with chunks where possible,
4673 * but write requests don't need to.
4675 static int raid5_mergeable_bvec(struct mddev *mddev,
4676 struct bvec_merge_data *bvm,
4677 struct bio_vec *biovec)
4679 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4681 unsigned int chunk_sectors = mddev->chunk_sectors;
4682 unsigned int bio_sectors = bvm->bi_size >> 9;
4685 * always allow writes to be mergeable, read as well if array
4686 * is degraded as we'll go through stripe cache anyway.
4688 if ((bvm->bi_rw & 1) == WRITE || mddev->degraded)
4689 return biovec->bv_len;
4691 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4692 chunk_sectors = mddev->new_chunk_sectors;
4693 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4694 if (max < 0) max = 0;
4695 if (max <= biovec->bv_len && bio_sectors == 0)
4696 return biovec->bv_len;
4701 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4703 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4704 unsigned int chunk_sectors = mddev->chunk_sectors;
4705 unsigned int bio_sectors = bio_sectors(bio);
4707 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4708 chunk_sectors = mddev->new_chunk_sectors;
4709 return chunk_sectors >=
4710 ((sector & (chunk_sectors - 1)) + bio_sectors);
4714 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4715 * later sampled by raid5d.
4717 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4719 unsigned long flags;
4721 spin_lock_irqsave(&conf->device_lock, flags);
4723 bi->bi_next = conf->retry_read_aligned_list;
4724 conf->retry_read_aligned_list = bi;
4726 spin_unlock_irqrestore(&conf->device_lock, flags);
4727 md_wakeup_thread(conf->mddev->thread);
4730 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4734 bi = conf->retry_read_aligned;
4736 conf->retry_read_aligned = NULL;
4739 bi = conf->retry_read_aligned_list;
4741 conf->retry_read_aligned_list = bi->bi_next;
4744 * this sets the active strip count to 1 and the processed
4745 * strip count to zero (upper 8 bits)
4747 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4754 * The "raid5_align_endio" should check if the read succeeded and if it
4755 * did, call bio_endio on the original bio (having bio_put the new bio
4757 * If the read failed..
4759 static void raid5_align_endio(struct bio *bi, int error)
4761 struct bio* raid_bi = bi->bi_private;
4762 struct mddev *mddev;
4763 struct r5conf *conf;
4764 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4765 struct md_rdev *rdev;
4769 rdev = (void*)raid_bi->bi_next;
4770 raid_bi->bi_next = NULL;
4771 mddev = rdev->mddev;
4772 conf = mddev->private;
4774 rdev_dec_pending(rdev, conf->mddev);
4776 if (!error && uptodate) {
4777 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4779 bio_endio(raid_bi, 0);
4780 if (atomic_dec_and_test(&conf->active_aligned_reads))
4781 wake_up(&conf->wait_for_quiescent);
4785 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4787 add_bio_to_retry(raid_bi, conf);
4790 static int bio_fits_rdev(struct bio *bi)
4792 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4794 if (bio_sectors(bi) > queue_max_sectors(q))
4796 blk_recount_segments(q, bi);
4797 if (bi->bi_phys_segments > queue_max_segments(q))
4800 if (q->merge_bvec_fn)
4801 /* it's too hard to apply the merge_bvec_fn at this stage,
4809 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4811 struct r5conf *conf = mddev->private;
4813 struct bio* align_bi;
4814 struct md_rdev *rdev;
4815 sector_t end_sector;
4817 if (!in_chunk_boundary(mddev, raid_bio)) {
4818 pr_debug("chunk_aligned_read : non aligned\n");
4822 * use bio_clone_mddev to make a copy of the bio
4824 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4828 * set bi_end_io to a new function, and set bi_private to the
4831 align_bi->bi_end_io = raid5_align_endio;
4832 align_bi->bi_private = raid_bio;
4836 align_bi->bi_iter.bi_sector =
4837 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4840 end_sector = bio_end_sector(align_bi);
4842 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4843 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4844 rdev->recovery_offset < end_sector) {
4845 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4847 (test_bit(Faulty, &rdev->flags) ||
4848 !(test_bit(In_sync, &rdev->flags) ||
4849 rdev->recovery_offset >= end_sector)))
4856 atomic_inc(&rdev->nr_pending);
4858 raid_bio->bi_next = (void*)rdev;
4859 align_bi->bi_bdev = rdev->bdev;
4860 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
4862 if (!bio_fits_rdev(align_bi) ||
4863 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4864 bio_sectors(align_bi),
4865 &first_bad, &bad_sectors)) {
4866 /* too big in some way, or has a known bad block */
4868 rdev_dec_pending(rdev, mddev);
4872 /* No reshape active, so we can trust rdev->data_offset */
4873 align_bi->bi_iter.bi_sector += rdev->data_offset;
4875 spin_lock_irq(&conf->device_lock);
4876 wait_event_lock_irq(conf->wait_for_quiescent,
4879 atomic_inc(&conf->active_aligned_reads);
4880 spin_unlock_irq(&conf->device_lock);
4883 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4884 align_bi, disk_devt(mddev->gendisk),
4885 raid_bio->bi_iter.bi_sector);
4886 generic_make_request(align_bi);
4895 /* __get_priority_stripe - get the next stripe to process
4897 * Full stripe writes are allowed to pass preread active stripes up until
4898 * the bypass_threshold is exceeded. In general the bypass_count
4899 * increments when the handle_list is handled before the hold_list; however, it
4900 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4901 * stripe with in flight i/o. The bypass_count will be reset when the
4902 * head of the hold_list has changed, i.e. the head was promoted to the
4905 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4907 struct stripe_head *sh = NULL, *tmp;
4908 struct list_head *handle_list = NULL;
4909 struct r5worker_group *wg = NULL;
4911 if (conf->worker_cnt_per_group == 0) {
4912 handle_list = &conf->handle_list;
4913 } else if (group != ANY_GROUP) {
4914 handle_list = &conf->worker_groups[group].handle_list;
4915 wg = &conf->worker_groups[group];
4918 for (i = 0; i < conf->group_cnt; i++) {
4919 handle_list = &conf->worker_groups[i].handle_list;
4920 wg = &conf->worker_groups[i];
4921 if (!list_empty(handle_list))
4926 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4928 list_empty(handle_list) ? "empty" : "busy",
4929 list_empty(&conf->hold_list) ? "empty" : "busy",
4930 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4932 if (!list_empty(handle_list)) {
4933 sh = list_entry(handle_list->next, typeof(*sh), lru);
4935 if (list_empty(&conf->hold_list))
4936 conf->bypass_count = 0;
4937 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4938 if (conf->hold_list.next == conf->last_hold)
4939 conf->bypass_count++;
4941 conf->last_hold = conf->hold_list.next;
4942 conf->bypass_count -= conf->bypass_threshold;
4943 if (conf->bypass_count < 0)
4944 conf->bypass_count = 0;
4947 } else if (!list_empty(&conf->hold_list) &&
4948 ((conf->bypass_threshold &&
4949 conf->bypass_count > conf->bypass_threshold) ||
4950 atomic_read(&conf->pending_full_writes) == 0)) {
4952 list_for_each_entry(tmp, &conf->hold_list, lru) {
4953 if (conf->worker_cnt_per_group == 0 ||
4954 group == ANY_GROUP ||
4955 !cpu_online(tmp->cpu) ||
4956 cpu_to_group(tmp->cpu) == group) {
4963 conf->bypass_count -= conf->bypass_threshold;
4964 if (conf->bypass_count < 0)
4965 conf->bypass_count = 0;
4977 list_del_init(&sh->lru);
4978 BUG_ON(atomic_inc_return(&sh->count) != 1);
4982 struct raid5_plug_cb {
4983 struct blk_plug_cb cb;
4984 struct list_head list;
4985 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4988 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4990 struct raid5_plug_cb *cb = container_of(
4991 blk_cb, struct raid5_plug_cb, cb);
4992 struct stripe_head *sh;
4993 struct mddev *mddev = cb->cb.data;
4994 struct r5conf *conf = mddev->private;
4998 if (cb->list.next && !list_empty(&cb->list)) {
4999 spin_lock_irq(&conf->device_lock);
5000 while (!list_empty(&cb->list)) {
5001 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5002 list_del_init(&sh->lru);
5004 * avoid race release_stripe_plug() sees
5005 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5006 * is still in our list
5008 smp_mb__before_atomic();
5009 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5011 * STRIPE_ON_RELEASE_LIST could be set here. In that
5012 * case, the count is always > 1 here
5014 hash = sh->hash_lock_index;
5015 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5018 spin_unlock_irq(&conf->device_lock);
5020 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5021 NR_STRIPE_HASH_LOCKS);
5023 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5027 static void release_stripe_plug(struct mddev *mddev,
5028 struct stripe_head *sh)
5030 struct blk_plug_cb *blk_cb = blk_check_plugged(
5031 raid5_unplug, mddev,
5032 sizeof(struct raid5_plug_cb));
5033 struct raid5_plug_cb *cb;
5040 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5042 if (cb->list.next == NULL) {
5044 INIT_LIST_HEAD(&cb->list);
5045 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5046 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5049 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5050 list_add_tail(&sh->lru, &cb->list);
5055 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5057 struct r5conf *conf = mddev->private;
5058 sector_t logical_sector, last_sector;
5059 struct stripe_head *sh;
5063 if (mddev->reshape_position != MaxSector)
5064 /* Skip discard while reshape is happening */
5067 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5068 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5071 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5073 stripe_sectors = conf->chunk_sectors *
5074 (conf->raid_disks - conf->max_degraded);
5075 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5077 sector_div(last_sector, stripe_sectors);
5079 logical_sector *= conf->chunk_sectors;
5080 last_sector *= conf->chunk_sectors;
5082 for (; logical_sector < last_sector;
5083 logical_sector += STRIPE_SECTORS) {
5087 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
5088 prepare_to_wait(&conf->wait_for_overlap, &w,
5089 TASK_UNINTERRUPTIBLE);
5090 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5091 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5096 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5097 spin_lock_irq(&sh->stripe_lock);
5098 for (d = 0; d < conf->raid_disks; d++) {
5099 if (d == sh->pd_idx || d == sh->qd_idx)
5101 if (sh->dev[d].towrite || sh->dev[d].toread) {
5102 set_bit(R5_Overlap, &sh->dev[d].flags);
5103 spin_unlock_irq(&sh->stripe_lock);
5109 set_bit(STRIPE_DISCARD, &sh->state);
5110 finish_wait(&conf->wait_for_overlap, &w);
5111 sh->overwrite_disks = 0;
5112 for (d = 0; d < conf->raid_disks; d++) {
5113 if (d == sh->pd_idx || d == sh->qd_idx)
5115 sh->dev[d].towrite = bi;
5116 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5117 raid5_inc_bi_active_stripes(bi);
5118 sh->overwrite_disks++;
5120 spin_unlock_irq(&sh->stripe_lock);
5121 if (conf->mddev->bitmap) {
5123 d < conf->raid_disks - conf->max_degraded;
5125 bitmap_startwrite(mddev->bitmap,
5129 sh->bm_seq = conf->seq_flush + 1;
5130 set_bit(STRIPE_BIT_DELAY, &sh->state);
5133 set_bit(STRIPE_HANDLE, &sh->state);
5134 clear_bit(STRIPE_DELAYED, &sh->state);
5135 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5136 atomic_inc(&conf->preread_active_stripes);
5137 release_stripe_plug(mddev, sh);
5140 remaining = raid5_dec_bi_active_stripes(bi);
5141 if (remaining == 0) {
5142 md_write_end(mddev);
5147 static void make_request(struct mddev *mddev, struct bio * bi)
5149 struct r5conf *conf = mddev->private;
5151 sector_t new_sector;
5152 sector_t logical_sector, last_sector;
5153 struct stripe_head *sh;
5154 const int rw = bio_data_dir(bi);
5159 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5160 md_flush_request(mddev, bi);
5164 md_write_start(mddev, bi);
5167 * If array is degraded, better not do chunk aligned read because
5168 * later we might have to read it again in order to reconstruct
5169 * data on failed drives.
5171 if (rw == READ && mddev->degraded == 0 &&
5172 mddev->reshape_position == MaxSector &&
5173 chunk_aligned_read(mddev,bi))
5176 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5177 make_discard_request(mddev, bi);
5181 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5182 last_sector = bio_end_sector(bi);
5184 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5186 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5187 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5193 seq = read_seqcount_begin(&conf->gen_lock);
5196 prepare_to_wait(&conf->wait_for_overlap, &w,
5197 TASK_UNINTERRUPTIBLE);
5198 if (unlikely(conf->reshape_progress != MaxSector)) {
5199 /* spinlock is needed as reshape_progress may be
5200 * 64bit on a 32bit platform, and so it might be
5201 * possible to see a half-updated value
5202 * Of course reshape_progress could change after
5203 * the lock is dropped, so once we get a reference
5204 * to the stripe that we think it is, we will have
5207 spin_lock_irq(&conf->device_lock);
5208 if (mddev->reshape_backwards
5209 ? logical_sector < conf->reshape_progress
5210 : logical_sector >= conf->reshape_progress) {
5213 if (mddev->reshape_backwards
5214 ? logical_sector < conf->reshape_safe
5215 : logical_sector >= conf->reshape_safe) {
5216 spin_unlock_irq(&conf->device_lock);
5222 spin_unlock_irq(&conf->device_lock);
5225 new_sector = raid5_compute_sector(conf, logical_sector,
5228 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5229 (unsigned long long)new_sector,
5230 (unsigned long long)logical_sector);
5232 sh = get_active_stripe(conf, new_sector, previous,
5233 (bi->bi_rw&RWA_MASK), 0);
5235 if (unlikely(previous)) {
5236 /* expansion might have moved on while waiting for a
5237 * stripe, so we must do the range check again.
5238 * Expansion could still move past after this
5239 * test, but as we are holding a reference to
5240 * 'sh', we know that if that happens,
5241 * STRIPE_EXPANDING will get set and the expansion
5242 * won't proceed until we finish with the stripe.
5245 spin_lock_irq(&conf->device_lock);
5246 if (mddev->reshape_backwards
5247 ? logical_sector >= conf->reshape_progress
5248 : logical_sector < conf->reshape_progress)
5249 /* mismatch, need to try again */
5251 spin_unlock_irq(&conf->device_lock);
5259 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5260 /* Might have got the wrong stripe_head
5268 logical_sector >= mddev->suspend_lo &&
5269 logical_sector < mddev->suspend_hi) {
5271 /* As the suspend_* range is controlled by
5272 * userspace, we want an interruptible
5275 flush_signals(current);
5276 prepare_to_wait(&conf->wait_for_overlap,
5277 &w, TASK_INTERRUPTIBLE);
5278 if (logical_sector >= mddev->suspend_lo &&
5279 logical_sector < mddev->suspend_hi) {
5286 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5287 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5288 /* Stripe is busy expanding or
5289 * add failed due to overlap. Flush everything
5292 md_wakeup_thread(mddev->thread);
5298 set_bit(STRIPE_HANDLE, &sh->state);
5299 clear_bit(STRIPE_DELAYED, &sh->state);
5300 if ((!sh->batch_head || sh == sh->batch_head) &&
5301 (bi->bi_rw & REQ_SYNC) &&
5302 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5303 atomic_inc(&conf->preread_active_stripes);
5304 release_stripe_plug(mddev, sh);
5306 /* cannot get stripe for read-ahead, just give-up */
5307 clear_bit(BIO_UPTODATE, &bi->bi_flags);
5311 finish_wait(&conf->wait_for_overlap, &w);
5313 remaining = raid5_dec_bi_active_stripes(bi);
5314 if (remaining == 0) {
5317 md_write_end(mddev);
5319 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5325 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5327 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5329 /* reshaping is quite different to recovery/resync so it is
5330 * handled quite separately ... here.
5332 * On each call to sync_request, we gather one chunk worth of
5333 * destination stripes and flag them as expanding.
5334 * Then we find all the source stripes and request reads.
5335 * As the reads complete, handle_stripe will copy the data
5336 * into the destination stripe and release that stripe.
5338 struct r5conf *conf = mddev->private;
5339 struct stripe_head *sh;
5340 sector_t first_sector, last_sector;
5341 int raid_disks = conf->previous_raid_disks;
5342 int data_disks = raid_disks - conf->max_degraded;
5343 int new_data_disks = conf->raid_disks - conf->max_degraded;
5346 sector_t writepos, readpos, safepos;
5347 sector_t stripe_addr;
5348 int reshape_sectors;
5349 struct list_head stripes;
5351 if (sector_nr == 0) {
5352 /* If restarting in the middle, skip the initial sectors */
5353 if (mddev->reshape_backwards &&
5354 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5355 sector_nr = raid5_size(mddev, 0, 0)
5356 - conf->reshape_progress;
5357 } else if (!mddev->reshape_backwards &&
5358 conf->reshape_progress > 0)
5359 sector_nr = conf->reshape_progress;
5360 sector_div(sector_nr, new_data_disks);
5362 mddev->curr_resync_completed = sector_nr;
5363 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5369 /* We need to process a full chunk at a time.
5370 * If old and new chunk sizes differ, we need to process the
5373 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
5374 reshape_sectors = mddev->new_chunk_sectors;
5376 reshape_sectors = mddev->chunk_sectors;
5378 /* We update the metadata at least every 10 seconds, or when
5379 * the data about to be copied would over-write the source of
5380 * the data at the front of the range. i.e. one new_stripe
5381 * along from reshape_progress new_maps to after where
5382 * reshape_safe old_maps to
5384 writepos = conf->reshape_progress;
5385 sector_div(writepos, new_data_disks);
5386 readpos = conf->reshape_progress;
5387 sector_div(readpos, data_disks);
5388 safepos = conf->reshape_safe;
5389 sector_div(safepos, data_disks);
5390 if (mddev->reshape_backwards) {
5391 writepos -= min_t(sector_t, reshape_sectors, writepos);
5392 readpos += reshape_sectors;
5393 safepos += reshape_sectors;
5395 writepos += reshape_sectors;
5396 readpos -= min_t(sector_t, reshape_sectors, readpos);
5397 safepos -= min_t(sector_t, reshape_sectors, safepos);
5400 /* Having calculated the 'writepos' possibly use it
5401 * to set 'stripe_addr' which is where we will write to.
5403 if (mddev->reshape_backwards) {
5404 BUG_ON(conf->reshape_progress == 0);
5405 stripe_addr = writepos;
5406 BUG_ON((mddev->dev_sectors &
5407 ~((sector_t)reshape_sectors - 1))
5408 - reshape_sectors - stripe_addr
5411 BUG_ON(writepos != sector_nr + reshape_sectors);
5412 stripe_addr = sector_nr;
5415 /* 'writepos' is the most advanced device address we might write.
5416 * 'readpos' is the least advanced device address we might read.
5417 * 'safepos' is the least address recorded in the metadata as having
5419 * If there is a min_offset_diff, these are adjusted either by
5420 * increasing the safepos/readpos if diff is negative, or
5421 * increasing writepos if diff is positive.
5422 * If 'readpos' is then behind 'writepos', there is no way that we can
5423 * ensure safety in the face of a crash - that must be done by userspace
5424 * making a backup of the data. So in that case there is no particular
5425 * rush to update metadata.
5426 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5427 * update the metadata to advance 'safepos' to match 'readpos' so that
5428 * we can be safe in the event of a crash.
5429 * So we insist on updating metadata if safepos is behind writepos and
5430 * readpos is beyond writepos.
5431 * In any case, update the metadata every 10 seconds.
5432 * Maybe that number should be configurable, but I'm not sure it is
5433 * worth it.... maybe it could be a multiple of safemode_delay???
5435 if (conf->min_offset_diff < 0) {
5436 safepos += -conf->min_offset_diff;
5437 readpos += -conf->min_offset_diff;
5439 writepos += conf->min_offset_diff;
5441 if ((mddev->reshape_backwards
5442 ? (safepos > writepos && readpos < writepos)
5443 : (safepos < writepos && readpos > writepos)) ||
5444 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5445 /* Cannot proceed until we've updated the superblock... */
5446 wait_event(conf->wait_for_overlap,
5447 atomic_read(&conf->reshape_stripes)==0
5448 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5449 if (atomic_read(&conf->reshape_stripes) != 0)
5451 mddev->reshape_position = conf->reshape_progress;
5452 mddev->curr_resync_completed = sector_nr;
5453 conf->reshape_checkpoint = jiffies;
5454 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5455 md_wakeup_thread(mddev->thread);
5456 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5457 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5458 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5460 spin_lock_irq(&conf->device_lock);
5461 conf->reshape_safe = mddev->reshape_position;
5462 spin_unlock_irq(&conf->device_lock);
5463 wake_up(&conf->wait_for_overlap);
5464 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5467 INIT_LIST_HEAD(&stripes);
5468 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5470 int skipped_disk = 0;
5471 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5472 set_bit(STRIPE_EXPANDING, &sh->state);
5473 atomic_inc(&conf->reshape_stripes);
5474 /* If any of this stripe is beyond the end of the old
5475 * array, then we need to zero those blocks
5477 for (j=sh->disks; j--;) {
5479 if (j == sh->pd_idx)
5481 if (conf->level == 6 &&
5484 s = compute_blocknr(sh, j, 0);
5485 if (s < raid5_size(mddev, 0, 0)) {
5489 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5490 set_bit(R5_Expanded, &sh->dev[j].flags);
5491 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5493 if (!skipped_disk) {
5494 set_bit(STRIPE_EXPAND_READY, &sh->state);
5495 set_bit(STRIPE_HANDLE, &sh->state);
5497 list_add(&sh->lru, &stripes);
5499 spin_lock_irq(&conf->device_lock);
5500 if (mddev->reshape_backwards)
5501 conf->reshape_progress -= reshape_sectors * new_data_disks;
5503 conf->reshape_progress += reshape_sectors * new_data_disks;
5504 spin_unlock_irq(&conf->device_lock);
5505 /* Ok, those stripe are ready. We can start scheduling
5506 * reads on the source stripes.
5507 * The source stripes are determined by mapping the first and last
5508 * block on the destination stripes.
5511 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5514 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5515 * new_data_disks - 1),
5517 if (last_sector >= mddev->dev_sectors)
5518 last_sector = mddev->dev_sectors - 1;
5519 while (first_sector <= last_sector) {
5520 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
5521 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5522 set_bit(STRIPE_HANDLE, &sh->state);
5524 first_sector += STRIPE_SECTORS;
5526 /* Now that the sources are clearly marked, we can release
5527 * the destination stripes
5529 while (!list_empty(&stripes)) {
5530 sh = list_entry(stripes.next, struct stripe_head, lru);
5531 list_del_init(&sh->lru);
5534 /* If this takes us to the resync_max point where we have to pause,
5535 * then we need to write out the superblock.
5537 sector_nr += reshape_sectors;
5538 if ((sector_nr - mddev->curr_resync_completed) * 2
5539 >= mddev->resync_max - mddev->curr_resync_completed) {
5540 /* Cannot proceed until we've updated the superblock... */
5541 wait_event(conf->wait_for_overlap,
5542 atomic_read(&conf->reshape_stripes) == 0
5543 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5544 if (atomic_read(&conf->reshape_stripes) != 0)
5546 mddev->reshape_position = conf->reshape_progress;
5547 mddev->curr_resync_completed = sector_nr;
5548 conf->reshape_checkpoint = jiffies;
5549 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5550 md_wakeup_thread(mddev->thread);
5551 wait_event(mddev->sb_wait,
5552 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5553 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5554 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5556 spin_lock_irq(&conf->device_lock);
5557 conf->reshape_safe = mddev->reshape_position;
5558 spin_unlock_irq(&conf->device_lock);
5559 wake_up(&conf->wait_for_overlap);
5560 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5563 return reshape_sectors;
5566 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5568 struct r5conf *conf = mddev->private;
5569 struct stripe_head *sh;
5570 sector_t max_sector = mddev->dev_sectors;
5571 sector_t sync_blocks;
5572 int still_degraded = 0;
5575 if (sector_nr >= max_sector) {
5576 /* just being told to finish up .. nothing much to do */
5578 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5583 if (mddev->curr_resync < max_sector) /* aborted */
5584 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5586 else /* completed sync */
5588 bitmap_close_sync(mddev->bitmap);
5593 /* Allow raid5_quiesce to complete */
5594 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5596 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5597 return reshape_request(mddev, sector_nr, skipped);
5599 /* No need to check resync_max as we never do more than one
5600 * stripe, and as resync_max will always be on a chunk boundary,
5601 * if the check in md_do_sync didn't fire, there is no chance
5602 * of overstepping resync_max here
5605 /* if there is too many failed drives and we are trying
5606 * to resync, then assert that we are finished, because there is
5607 * nothing we can do.
5609 if (mddev->degraded >= conf->max_degraded &&
5610 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5611 sector_t rv = mddev->dev_sectors - sector_nr;
5615 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5617 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5618 sync_blocks >= STRIPE_SECTORS) {
5619 /* we can skip this block, and probably more */
5620 sync_blocks /= STRIPE_SECTORS;
5622 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5625 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5627 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5629 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5630 /* make sure we don't swamp the stripe cache if someone else
5631 * is trying to get access
5633 schedule_timeout_uninterruptible(1);
5635 /* Need to check if array will still be degraded after recovery/resync
5636 * Note in case of > 1 drive failures it's possible we're rebuilding
5637 * one drive while leaving another faulty drive in array.
5640 for (i = 0; i < conf->raid_disks; i++) {
5641 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5643 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5648 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5650 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5651 set_bit(STRIPE_HANDLE, &sh->state);
5655 return STRIPE_SECTORS;
5658 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5660 /* We may not be able to submit a whole bio at once as there
5661 * may not be enough stripe_heads available.
5662 * We cannot pre-allocate enough stripe_heads as we may need
5663 * more than exist in the cache (if we allow ever large chunks).
5664 * So we do one stripe head at a time and record in
5665 * ->bi_hw_segments how many have been done.
5667 * We *know* that this entire raid_bio is in one chunk, so
5668 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5670 struct stripe_head *sh;
5672 sector_t sector, logical_sector, last_sector;
5677 logical_sector = raid_bio->bi_iter.bi_sector &
5678 ~((sector_t)STRIPE_SECTORS-1);
5679 sector = raid5_compute_sector(conf, logical_sector,
5681 last_sector = bio_end_sector(raid_bio);
5683 for (; logical_sector < last_sector;
5684 logical_sector += STRIPE_SECTORS,
5685 sector += STRIPE_SECTORS,
5688 if (scnt < raid5_bi_processed_stripes(raid_bio))
5689 /* already done this stripe */
5692 sh = get_active_stripe(conf, sector, 0, 1, 1);
5695 /* failed to get a stripe - must wait */
5696 raid5_set_bi_processed_stripes(raid_bio, scnt);
5697 conf->retry_read_aligned = raid_bio;
5701 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5703 raid5_set_bi_processed_stripes(raid_bio, scnt);
5704 conf->retry_read_aligned = raid_bio;
5708 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5713 remaining = raid5_dec_bi_active_stripes(raid_bio);
5714 if (remaining == 0) {
5715 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5717 bio_endio(raid_bio, 0);
5719 if (atomic_dec_and_test(&conf->active_aligned_reads))
5720 wake_up(&conf->wait_for_quiescent);
5724 static int handle_active_stripes(struct r5conf *conf, int group,
5725 struct r5worker *worker,
5726 struct list_head *temp_inactive_list)
5728 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5729 int i, batch_size = 0, hash;
5730 bool release_inactive = false;
5732 while (batch_size < MAX_STRIPE_BATCH &&
5733 (sh = __get_priority_stripe(conf, group)) != NULL)
5734 batch[batch_size++] = sh;
5736 if (batch_size == 0) {
5737 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5738 if (!list_empty(temp_inactive_list + i))
5740 if (i == NR_STRIPE_HASH_LOCKS)
5742 release_inactive = true;
5744 spin_unlock_irq(&conf->device_lock);
5746 release_inactive_stripe_list(conf, temp_inactive_list,
5747 NR_STRIPE_HASH_LOCKS);
5749 if (release_inactive) {
5750 spin_lock_irq(&conf->device_lock);
5754 for (i = 0; i < batch_size; i++)
5755 handle_stripe(batch[i]);
5759 spin_lock_irq(&conf->device_lock);
5760 for (i = 0; i < batch_size; i++) {
5761 hash = batch[i]->hash_lock_index;
5762 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5767 static void raid5_do_work(struct work_struct *work)
5769 struct r5worker *worker = container_of(work, struct r5worker, work);
5770 struct r5worker_group *group = worker->group;
5771 struct r5conf *conf = group->conf;
5772 int group_id = group - conf->worker_groups;
5774 struct blk_plug plug;
5776 pr_debug("+++ raid5worker active\n");
5778 blk_start_plug(&plug);
5780 spin_lock_irq(&conf->device_lock);
5782 int batch_size, released;
5784 released = release_stripe_list(conf, worker->temp_inactive_list);
5786 batch_size = handle_active_stripes(conf, group_id, worker,
5787 worker->temp_inactive_list);
5788 worker->working = false;
5789 if (!batch_size && !released)
5791 handled += batch_size;
5793 pr_debug("%d stripes handled\n", handled);
5795 spin_unlock_irq(&conf->device_lock);
5796 blk_finish_plug(&plug);
5798 pr_debug("--- raid5worker inactive\n");
5802 * This is our raid5 kernel thread.
5804 * We scan the hash table for stripes which can be handled now.
5805 * During the scan, completed stripes are saved for us by the interrupt
5806 * handler, so that they will not have to wait for our next wakeup.
5808 static void raid5d(struct md_thread *thread)
5810 struct mddev *mddev = thread->mddev;
5811 struct r5conf *conf = mddev->private;
5813 struct blk_plug plug;
5815 pr_debug("+++ raid5d active\n");
5817 md_check_recovery(mddev);
5819 blk_start_plug(&plug);
5821 spin_lock_irq(&conf->device_lock);
5824 int batch_size, released;
5826 released = release_stripe_list(conf, conf->temp_inactive_list);
5828 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5831 !list_empty(&conf->bitmap_list)) {
5832 /* Now is a good time to flush some bitmap updates */
5834 spin_unlock_irq(&conf->device_lock);
5835 bitmap_unplug(mddev->bitmap);
5836 spin_lock_irq(&conf->device_lock);
5837 conf->seq_write = conf->seq_flush;
5838 activate_bit_delay(conf, conf->temp_inactive_list);
5840 raid5_activate_delayed(conf);
5842 while ((bio = remove_bio_from_retry(conf))) {
5844 spin_unlock_irq(&conf->device_lock);
5845 ok = retry_aligned_read(conf, bio);
5846 spin_lock_irq(&conf->device_lock);
5852 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5853 conf->temp_inactive_list);
5854 if (!batch_size && !released)
5856 handled += batch_size;
5858 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5859 spin_unlock_irq(&conf->device_lock);
5860 md_check_recovery(mddev);
5861 spin_lock_irq(&conf->device_lock);
5864 pr_debug("%d stripes handled\n", handled);
5866 spin_unlock_irq(&conf->device_lock);
5867 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5868 mutex_trylock(&conf->cache_size_mutex)) {
5869 grow_one_stripe(conf, __GFP_NOWARN);
5870 /* Set flag even if allocation failed. This helps
5871 * slow down allocation requests when mem is short
5873 set_bit(R5_DID_ALLOC, &conf->cache_state);
5874 mutex_unlock(&conf->cache_size_mutex);
5877 async_tx_issue_pending_all();
5878 blk_finish_plug(&plug);
5880 pr_debug("--- raid5d inactive\n");
5884 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5886 struct r5conf *conf;
5888 spin_lock(&mddev->lock);
5889 conf = mddev->private;
5891 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5892 spin_unlock(&mddev->lock);
5897 raid5_set_cache_size(struct mddev *mddev, int size)
5899 struct r5conf *conf = mddev->private;
5902 if (size <= 16 || size > 32768)
5905 conf->min_nr_stripes = size;
5906 mutex_lock(&conf->cache_size_mutex);
5907 while (size < conf->max_nr_stripes &&
5908 drop_one_stripe(conf))
5910 mutex_unlock(&conf->cache_size_mutex);
5913 err = md_allow_write(mddev);
5917 mutex_lock(&conf->cache_size_mutex);
5918 while (size > conf->max_nr_stripes)
5919 if (!grow_one_stripe(conf, GFP_KERNEL))
5921 mutex_unlock(&conf->cache_size_mutex);
5925 EXPORT_SYMBOL(raid5_set_cache_size);
5928 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5930 struct r5conf *conf;
5934 if (len >= PAGE_SIZE)
5936 if (kstrtoul(page, 10, &new))
5938 err = mddev_lock(mddev);
5941 conf = mddev->private;
5945 err = raid5_set_cache_size(mddev, new);
5946 mddev_unlock(mddev);
5951 static struct md_sysfs_entry
5952 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5953 raid5_show_stripe_cache_size,
5954 raid5_store_stripe_cache_size);
5957 raid5_show_rmw_level(struct mddev *mddev, char *page)
5959 struct r5conf *conf = mddev->private;
5961 return sprintf(page, "%d\n", conf->rmw_level);
5967 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
5969 struct r5conf *conf = mddev->private;
5975 if (len >= PAGE_SIZE)
5978 if (kstrtoul(page, 10, &new))
5981 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
5984 if (new != PARITY_DISABLE_RMW &&
5985 new != PARITY_ENABLE_RMW &&
5986 new != PARITY_PREFER_RMW)
5989 conf->rmw_level = new;
5993 static struct md_sysfs_entry
5994 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
5995 raid5_show_rmw_level,
5996 raid5_store_rmw_level);
6000 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6002 struct r5conf *conf;
6004 spin_lock(&mddev->lock);
6005 conf = mddev->private;
6007 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6008 spin_unlock(&mddev->lock);
6013 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6015 struct r5conf *conf;
6019 if (len >= PAGE_SIZE)
6021 if (kstrtoul(page, 10, &new))
6024 err = mddev_lock(mddev);
6027 conf = mddev->private;
6030 else if (new > conf->min_nr_stripes)
6033 conf->bypass_threshold = new;
6034 mddev_unlock(mddev);
6038 static struct md_sysfs_entry
6039 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6041 raid5_show_preread_threshold,
6042 raid5_store_preread_threshold);
6045 raid5_show_skip_copy(struct mddev *mddev, char *page)
6047 struct r5conf *conf;
6049 spin_lock(&mddev->lock);
6050 conf = mddev->private;
6052 ret = sprintf(page, "%d\n", conf->skip_copy);
6053 spin_unlock(&mddev->lock);
6058 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6060 struct r5conf *conf;
6064 if (len >= PAGE_SIZE)
6066 if (kstrtoul(page, 10, &new))
6070 err = mddev_lock(mddev);
6073 conf = mddev->private;
6076 else if (new != conf->skip_copy) {
6077 mddev_suspend(mddev);
6078 conf->skip_copy = new;
6080 mddev->queue->backing_dev_info.capabilities |=
6081 BDI_CAP_STABLE_WRITES;
6083 mddev->queue->backing_dev_info.capabilities &=
6084 ~BDI_CAP_STABLE_WRITES;
6085 mddev_resume(mddev);
6087 mddev_unlock(mddev);
6091 static struct md_sysfs_entry
6092 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6093 raid5_show_skip_copy,
6094 raid5_store_skip_copy);
6097 stripe_cache_active_show(struct mddev *mddev, char *page)
6099 struct r5conf *conf = mddev->private;
6101 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6106 static struct md_sysfs_entry
6107 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6110 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6112 struct r5conf *conf;
6114 spin_lock(&mddev->lock);
6115 conf = mddev->private;
6117 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6118 spin_unlock(&mddev->lock);
6122 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6124 int *worker_cnt_per_group,
6125 struct r5worker_group **worker_groups);
6127 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6129 struct r5conf *conf;
6132 struct r5worker_group *new_groups, *old_groups;
6133 int group_cnt, worker_cnt_per_group;
6135 if (len >= PAGE_SIZE)
6137 if (kstrtoul(page, 10, &new))
6140 err = mddev_lock(mddev);
6143 conf = mddev->private;
6146 else if (new != conf->worker_cnt_per_group) {
6147 mddev_suspend(mddev);
6149 old_groups = conf->worker_groups;
6151 flush_workqueue(raid5_wq);
6153 err = alloc_thread_groups(conf, new,
6154 &group_cnt, &worker_cnt_per_group,
6157 spin_lock_irq(&conf->device_lock);
6158 conf->group_cnt = group_cnt;
6159 conf->worker_cnt_per_group = worker_cnt_per_group;
6160 conf->worker_groups = new_groups;
6161 spin_unlock_irq(&conf->device_lock);
6164 kfree(old_groups[0].workers);
6167 mddev_resume(mddev);
6169 mddev_unlock(mddev);
6174 static struct md_sysfs_entry
6175 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6176 raid5_show_group_thread_cnt,
6177 raid5_store_group_thread_cnt);
6179 static struct attribute *raid5_attrs[] = {
6180 &raid5_stripecache_size.attr,
6181 &raid5_stripecache_active.attr,
6182 &raid5_preread_bypass_threshold.attr,
6183 &raid5_group_thread_cnt.attr,
6184 &raid5_skip_copy.attr,
6185 &raid5_rmw_level.attr,
6188 static struct attribute_group raid5_attrs_group = {
6190 .attrs = raid5_attrs,
6193 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6195 int *worker_cnt_per_group,
6196 struct r5worker_group **worker_groups)
6200 struct r5worker *workers;
6202 *worker_cnt_per_group = cnt;
6205 *worker_groups = NULL;
6208 *group_cnt = num_possible_nodes();
6209 size = sizeof(struct r5worker) * cnt;
6210 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6211 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6212 *group_cnt, GFP_NOIO);
6213 if (!*worker_groups || !workers) {
6215 kfree(*worker_groups);
6219 for (i = 0; i < *group_cnt; i++) {
6220 struct r5worker_group *group;
6222 group = &(*worker_groups)[i];
6223 INIT_LIST_HEAD(&group->handle_list);
6225 group->workers = workers + i * cnt;
6227 for (j = 0; j < cnt; j++) {
6228 struct r5worker *worker = group->workers + j;
6229 worker->group = group;
6230 INIT_WORK(&worker->work, raid5_do_work);
6232 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6233 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6240 static void free_thread_groups(struct r5conf *conf)
6242 if (conf->worker_groups)
6243 kfree(conf->worker_groups[0].workers);
6244 kfree(conf->worker_groups);
6245 conf->worker_groups = NULL;
6249 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6251 struct r5conf *conf = mddev->private;
6254 sectors = mddev->dev_sectors;
6256 /* size is defined by the smallest of previous and new size */
6257 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6259 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6260 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
6261 return sectors * (raid_disks - conf->max_degraded);
6264 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6266 safe_put_page(percpu->spare_page);
6267 if (percpu->scribble)
6268 flex_array_free(percpu->scribble);
6269 percpu->spare_page = NULL;
6270 percpu->scribble = NULL;
6273 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6275 if (conf->level == 6 && !percpu->spare_page)
6276 percpu->spare_page = alloc_page(GFP_KERNEL);
6277 if (!percpu->scribble)
6278 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6279 conf->previous_raid_disks),
6280 max(conf->chunk_sectors,
6281 conf->prev_chunk_sectors)
6285 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6286 free_scratch_buffer(conf, percpu);
6293 static void raid5_free_percpu(struct r5conf *conf)
6300 #ifdef CONFIG_HOTPLUG_CPU
6301 unregister_cpu_notifier(&conf->cpu_notify);
6305 for_each_possible_cpu(cpu)
6306 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6309 free_percpu(conf->percpu);
6312 static void free_conf(struct r5conf *conf)
6314 if (conf->shrinker.seeks)
6315 unregister_shrinker(&conf->shrinker);
6316 free_thread_groups(conf);
6317 shrink_stripes(conf);
6318 raid5_free_percpu(conf);
6320 kfree(conf->stripe_hashtbl);
6324 #ifdef CONFIG_HOTPLUG_CPU
6325 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6328 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6329 long cpu = (long)hcpu;
6330 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6333 case CPU_UP_PREPARE:
6334 case CPU_UP_PREPARE_FROZEN:
6335 if (alloc_scratch_buffer(conf, percpu)) {
6336 pr_err("%s: failed memory allocation for cpu%ld\n",
6338 return notifier_from_errno(-ENOMEM);
6342 case CPU_DEAD_FROZEN:
6343 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6352 static int raid5_alloc_percpu(struct r5conf *conf)
6357 conf->percpu = alloc_percpu(struct raid5_percpu);
6361 #ifdef CONFIG_HOTPLUG_CPU
6362 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6363 conf->cpu_notify.priority = 0;
6364 err = register_cpu_notifier(&conf->cpu_notify);
6370 for_each_present_cpu(cpu) {
6371 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6373 pr_err("%s: failed memory allocation for cpu%ld\n",
6383 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6384 struct shrink_control *sc)
6386 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6387 unsigned long ret = SHRINK_STOP;
6389 if (mutex_trylock(&conf->cache_size_mutex)) {
6391 while (ret < sc->nr_to_scan &&
6392 conf->max_nr_stripes > conf->min_nr_stripes) {
6393 if (drop_one_stripe(conf) == 0) {
6399 mutex_unlock(&conf->cache_size_mutex);
6404 static unsigned long raid5_cache_count(struct shrinker *shrink,
6405 struct shrink_control *sc)
6407 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6409 if (conf->max_nr_stripes < conf->min_nr_stripes)
6410 /* unlikely, but not impossible */
6412 return conf->max_nr_stripes - conf->min_nr_stripes;
6415 static struct r5conf *setup_conf(struct mddev *mddev)
6417 struct r5conf *conf;
6418 int raid_disk, memory, max_disks;
6419 struct md_rdev *rdev;
6420 struct disk_info *disk;
6423 int group_cnt, worker_cnt_per_group;
6424 struct r5worker_group *new_group;
6426 if (mddev->new_level != 5
6427 && mddev->new_level != 4
6428 && mddev->new_level != 6) {
6429 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6430 mdname(mddev), mddev->new_level);
6431 return ERR_PTR(-EIO);
6433 if ((mddev->new_level == 5
6434 && !algorithm_valid_raid5(mddev->new_layout)) ||
6435 (mddev->new_level == 6
6436 && !algorithm_valid_raid6(mddev->new_layout))) {
6437 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6438 mdname(mddev), mddev->new_layout);
6439 return ERR_PTR(-EIO);
6441 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6442 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6443 mdname(mddev), mddev->raid_disks);
6444 return ERR_PTR(-EINVAL);
6447 if (!mddev->new_chunk_sectors ||
6448 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6449 !is_power_of_2(mddev->new_chunk_sectors)) {
6450 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6451 mdname(mddev), mddev->new_chunk_sectors << 9);
6452 return ERR_PTR(-EINVAL);
6455 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6458 /* Don't enable multi-threading by default*/
6459 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6461 conf->group_cnt = group_cnt;
6462 conf->worker_cnt_per_group = worker_cnt_per_group;
6463 conf->worker_groups = new_group;
6466 spin_lock_init(&conf->device_lock);
6467 seqcount_init(&conf->gen_lock);
6468 mutex_init(&conf->cache_size_mutex);
6469 init_waitqueue_head(&conf->wait_for_quiescent);
6470 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
6471 init_waitqueue_head(&conf->wait_for_stripe[i]);
6473 init_waitqueue_head(&conf->wait_for_overlap);
6474 INIT_LIST_HEAD(&conf->handle_list);
6475 INIT_LIST_HEAD(&conf->hold_list);
6476 INIT_LIST_HEAD(&conf->delayed_list);
6477 INIT_LIST_HEAD(&conf->bitmap_list);
6478 init_llist_head(&conf->released_stripes);
6479 atomic_set(&conf->active_stripes, 0);
6480 atomic_set(&conf->preread_active_stripes, 0);
6481 atomic_set(&conf->active_aligned_reads, 0);
6482 conf->bypass_threshold = BYPASS_THRESHOLD;
6483 conf->recovery_disabled = mddev->recovery_disabled - 1;
6485 conf->raid_disks = mddev->raid_disks;
6486 if (mddev->reshape_position == MaxSector)
6487 conf->previous_raid_disks = mddev->raid_disks;
6489 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6490 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6492 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6497 conf->mddev = mddev;
6499 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6502 /* We init hash_locks[0] separately to that it can be used
6503 * as the reference lock in the spin_lock_nest_lock() call
6504 * in lock_all_device_hash_locks_irq in order to convince
6505 * lockdep that we know what we are doing.
6507 spin_lock_init(conf->hash_locks);
6508 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6509 spin_lock_init(conf->hash_locks + i);
6511 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6512 INIT_LIST_HEAD(conf->inactive_list + i);
6514 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6515 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6517 conf->level = mddev->new_level;
6518 conf->chunk_sectors = mddev->new_chunk_sectors;
6519 if (raid5_alloc_percpu(conf) != 0)
6522 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6524 rdev_for_each(rdev, mddev) {
6525 raid_disk = rdev->raid_disk;
6526 if (raid_disk >= max_disks
6529 disk = conf->disks + raid_disk;
6531 if (test_bit(Replacement, &rdev->flags)) {
6532 if (disk->replacement)
6534 disk->replacement = rdev;
6541 if (test_bit(In_sync, &rdev->flags)) {
6542 char b[BDEVNAME_SIZE];
6543 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6545 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6546 } else if (rdev->saved_raid_disk != raid_disk)
6547 /* Cannot rely on bitmap to complete recovery */
6551 conf->level = mddev->new_level;
6552 if (conf->level == 6) {
6553 conf->max_degraded = 2;
6554 if (raid6_call.xor_syndrome)
6555 conf->rmw_level = PARITY_ENABLE_RMW;
6557 conf->rmw_level = PARITY_DISABLE_RMW;
6559 conf->max_degraded = 1;
6560 conf->rmw_level = PARITY_ENABLE_RMW;
6562 conf->algorithm = mddev->new_layout;
6563 conf->reshape_progress = mddev->reshape_position;
6564 if (conf->reshape_progress != MaxSector) {
6565 conf->prev_chunk_sectors = mddev->chunk_sectors;
6566 conf->prev_algo = mddev->layout;
6569 conf->min_nr_stripes = NR_STRIPES;
6570 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6571 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6572 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6573 if (grow_stripes(conf, conf->min_nr_stripes)) {
6575 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6576 mdname(mddev), memory);
6579 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6580 mdname(mddev), memory);
6582 * Losing a stripe head costs more than the time to refill it,
6583 * it reduces the queue depth and so can hurt throughput.
6584 * So set it rather large, scaled by number of devices.
6586 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6587 conf->shrinker.scan_objects = raid5_cache_scan;
6588 conf->shrinker.count_objects = raid5_cache_count;
6589 conf->shrinker.batch = 128;
6590 conf->shrinker.flags = 0;
6591 register_shrinker(&conf->shrinker);
6593 sprintf(pers_name, "raid%d", mddev->new_level);
6594 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6595 if (!conf->thread) {
6597 "md/raid:%s: couldn't allocate thread.\n",
6607 return ERR_PTR(-EIO);
6609 return ERR_PTR(-ENOMEM);
6612 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6615 case ALGORITHM_PARITY_0:
6616 if (raid_disk < max_degraded)
6619 case ALGORITHM_PARITY_N:
6620 if (raid_disk >= raid_disks - max_degraded)
6623 case ALGORITHM_PARITY_0_6:
6624 if (raid_disk == 0 ||
6625 raid_disk == raid_disks - 1)
6628 case ALGORITHM_LEFT_ASYMMETRIC_6:
6629 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6630 case ALGORITHM_LEFT_SYMMETRIC_6:
6631 case ALGORITHM_RIGHT_SYMMETRIC_6:
6632 if (raid_disk == raid_disks - 1)
6638 static int run(struct mddev *mddev)
6640 struct r5conf *conf;
6641 int working_disks = 0;
6642 int dirty_parity_disks = 0;
6643 struct md_rdev *rdev;
6644 sector_t reshape_offset = 0;
6646 long long min_offset_diff = 0;
6649 if (mddev->recovery_cp != MaxSector)
6650 printk(KERN_NOTICE "md/raid:%s: not clean"
6651 " -- starting background reconstruction\n",
6654 rdev_for_each(rdev, mddev) {
6656 if (rdev->raid_disk < 0)
6658 diff = (rdev->new_data_offset - rdev->data_offset);
6660 min_offset_diff = diff;
6662 } else if (mddev->reshape_backwards &&
6663 diff < min_offset_diff)
6664 min_offset_diff = diff;
6665 else if (!mddev->reshape_backwards &&
6666 diff > min_offset_diff)
6667 min_offset_diff = diff;
6670 if (mddev->reshape_position != MaxSector) {
6671 /* Check that we can continue the reshape.
6672 * Difficulties arise if the stripe we would write to
6673 * next is at or after the stripe we would read from next.
6674 * For a reshape that changes the number of devices, this
6675 * is only possible for a very short time, and mdadm makes
6676 * sure that time appears to have past before assembling
6677 * the array. So we fail if that time hasn't passed.
6678 * For a reshape that keeps the number of devices the same
6679 * mdadm must be monitoring the reshape can keeping the
6680 * critical areas read-only and backed up. It will start
6681 * the array in read-only mode, so we check for that.
6683 sector_t here_new, here_old;
6685 int max_degraded = (mddev->level == 6 ? 2 : 1);
6687 if (mddev->new_level != mddev->level) {
6688 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6689 "required - aborting.\n",
6693 old_disks = mddev->raid_disks - mddev->delta_disks;
6694 /* reshape_position must be on a new-stripe boundary, and one
6695 * further up in new geometry must map after here in old
6698 here_new = mddev->reshape_position;
6699 if (sector_div(here_new, mddev->new_chunk_sectors *
6700 (mddev->raid_disks - max_degraded))) {
6701 printk(KERN_ERR "md/raid:%s: reshape_position not "
6702 "on a stripe boundary\n", mdname(mddev));
6705 reshape_offset = here_new * mddev->new_chunk_sectors;
6706 /* here_new is the stripe we will write to */
6707 here_old = mddev->reshape_position;
6708 sector_div(here_old, mddev->chunk_sectors *
6709 (old_disks-max_degraded));
6710 /* here_old is the first stripe that we might need to read
6712 if (mddev->delta_disks == 0) {
6713 if ((here_new * mddev->new_chunk_sectors !=
6714 here_old * mddev->chunk_sectors)) {
6715 printk(KERN_ERR "md/raid:%s: reshape position is"
6716 " confused - aborting\n", mdname(mddev));
6719 /* We cannot be sure it is safe to start an in-place
6720 * reshape. It is only safe if user-space is monitoring
6721 * and taking constant backups.
6722 * mdadm always starts a situation like this in
6723 * readonly mode so it can take control before
6724 * allowing any writes. So just check for that.
6726 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6727 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6728 /* not really in-place - so OK */;
6729 else if (mddev->ro == 0) {
6730 printk(KERN_ERR "md/raid:%s: in-place reshape "
6731 "must be started in read-only mode "
6736 } else if (mddev->reshape_backwards
6737 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6738 here_old * mddev->chunk_sectors)
6739 : (here_new * mddev->new_chunk_sectors >=
6740 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6741 /* Reading from the same stripe as writing to - bad */
6742 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6743 "auto-recovery - aborting.\n",
6747 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6749 /* OK, we should be able to continue; */
6751 BUG_ON(mddev->level != mddev->new_level);
6752 BUG_ON(mddev->layout != mddev->new_layout);
6753 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6754 BUG_ON(mddev->delta_disks != 0);
6757 if (mddev->private == NULL)
6758 conf = setup_conf(mddev);
6760 conf = mddev->private;
6763 return PTR_ERR(conf);
6765 conf->min_offset_diff = min_offset_diff;
6766 mddev->thread = conf->thread;
6767 conf->thread = NULL;
6768 mddev->private = conf;
6770 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6772 rdev = conf->disks[i].rdev;
6773 if (!rdev && conf->disks[i].replacement) {
6774 /* The replacement is all we have yet */
6775 rdev = conf->disks[i].replacement;
6776 conf->disks[i].replacement = NULL;
6777 clear_bit(Replacement, &rdev->flags);
6778 conf->disks[i].rdev = rdev;
6782 if (conf->disks[i].replacement &&
6783 conf->reshape_progress != MaxSector) {
6784 /* replacements and reshape simply do not mix. */
6785 printk(KERN_ERR "md: cannot handle concurrent "
6786 "replacement and reshape.\n");
6789 if (test_bit(In_sync, &rdev->flags)) {
6793 /* This disc is not fully in-sync. However if it
6794 * just stored parity (beyond the recovery_offset),
6795 * when we don't need to be concerned about the
6796 * array being dirty.
6797 * When reshape goes 'backwards', we never have
6798 * partially completed devices, so we only need
6799 * to worry about reshape going forwards.
6801 /* Hack because v0.91 doesn't store recovery_offset properly. */
6802 if (mddev->major_version == 0 &&
6803 mddev->minor_version > 90)
6804 rdev->recovery_offset = reshape_offset;
6806 if (rdev->recovery_offset < reshape_offset) {
6807 /* We need to check old and new layout */
6808 if (!only_parity(rdev->raid_disk,
6811 conf->max_degraded))
6814 if (!only_parity(rdev->raid_disk,
6816 conf->previous_raid_disks,
6817 conf->max_degraded))
6819 dirty_parity_disks++;
6823 * 0 for a fully functional array, 1 or 2 for a degraded array.
6825 mddev->degraded = calc_degraded(conf);
6827 if (has_failed(conf)) {
6828 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6829 " (%d/%d failed)\n",
6830 mdname(mddev), mddev->degraded, conf->raid_disks);
6834 /* device size must be a multiple of chunk size */
6835 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6836 mddev->resync_max_sectors = mddev->dev_sectors;
6838 if (mddev->degraded > dirty_parity_disks &&
6839 mddev->recovery_cp != MaxSector) {
6840 if (mddev->ok_start_degraded)
6842 "md/raid:%s: starting dirty degraded array"
6843 " - data corruption possible.\n",
6847 "md/raid:%s: cannot start dirty degraded array.\n",
6853 if (mddev->degraded == 0)
6854 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6855 " devices, algorithm %d\n", mdname(mddev), conf->level,
6856 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6859 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6860 " out of %d devices, algorithm %d\n",
6861 mdname(mddev), conf->level,
6862 mddev->raid_disks - mddev->degraded,
6863 mddev->raid_disks, mddev->new_layout);
6865 print_raid5_conf(conf);
6867 if (conf->reshape_progress != MaxSector) {
6868 conf->reshape_safe = conf->reshape_progress;
6869 atomic_set(&conf->reshape_stripes, 0);
6870 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6871 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6872 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6873 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6874 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6878 /* Ok, everything is just fine now */
6879 if (mddev->to_remove == &raid5_attrs_group)
6880 mddev->to_remove = NULL;
6881 else if (mddev->kobj.sd &&
6882 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6884 "raid5: failed to create sysfs attributes for %s\n",
6886 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6890 bool discard_supported = true;
6891 /* read-ahead size must cover two whole stripes, which
6892 * is 2 * (datadisks) * chunksize where 'n' is the
6893 * number of raid devices
6895 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6896 int stripe = data_disks *
6897 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6898 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6899 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6901 chunk_size = mddev->chunk_sectors << 9;
6902 blk_queue_io_min(mddev->queue, chunk_size);
6903 blk_queue_io_opt(mddev->queue, chunk_size *
6904 (conf->raid_disks - conf->max_degraded));
6905 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6907 * We can only discard a whole stripe. It doesn't make sense to
6908 * discard data disk but write parity disk
6910 stripe = stripe * PAGE_SIZE;
6911 /* Round up to power of 2, as discard handling
6912 * currently assumes that */
6913 while ((stripe-1) & stripe)
6914 stripe = (stripe | (stripe-1)) + 1;
6915 mddev->queue->limits.discard_alignment = stripe;
6916 mddev->queue->limits.discard_granularity = stripe;
6918 * unaligned part of discard request will be ignored, so can't
6919 * guarantee discard_zeroes_data
6921 mddev->queue->limits.discard_zeroes_data = 0;
6923 blk_queue_max_write_same_sectors(mddev->queue, 0);
6925 rdev_for_each(rdev, mddev) {
6926 disk_stack_limits(mddev->gendisk, rdev->bdev,
6927 rdev->data_offset << 9);
6928 disk_stack_limits(mddev->gendisk, rdev->bdev,
6929 rdev->new_data_offset << 9);
6931 * discard_zeroes_data is required, otherwise data
6932 * could be lost. Consider a scenario: discard a stripe
6933 * (the stripe could be inconsistent if
6934 * discard_zeroes_data is 0); write one disk of the
6935 * stripe (the stripe could be inconsistent again
6936 * depending on which disks are used to calculate
6937 * parity); the disk is broken; The stripe data of this
6940 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6941 !bdev_get_queue(rdev->bdev)->
6942 limits.discard_zeroes_data)
6943 discard_supported = false;
6944 /* Unfortunately, discard_zeroes_data is not currently
6945 * a guarantee - just a hint. So we only allow DISCARD
6946 * if the sysadmin has confirmed that only safe devices
6947 * are in use by setting a module parameter.
6949 if (!devices_handle_discard_safely) {
6950 if (discard_supported) {
6951 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6952 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6954 discard_supported = false;
6958 if (discard_supported &&
6959 mddev->queue->limits.max_discard_sectors >= stripe &&
6960 mddev->queue->limits.discard_granularity >= stripe)
6961 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6964 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6970 md_unregister_thread(&mddev->thread);
6971 print_raid5_conf(conf);
6973 mddev->private = NULL;
6974 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6978 static void raid5_free(struct mddev *mddev, void *priv)
6980 struct r5conf *conf = priv;
6983 mddev->to_remove = &raid5_attrs_group;
6986 static void status(struct seq_file *seq, struct mddev *mddev)
6988 struct r5conf *conf = mddev->private;
6991 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6992 mddev->chunk_sectors / 2, mddev->layout);
6993 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6994 for (i = 0; i < conf->raid_disks; i++)
6995 seq_printf (seq, "%s",
6996 conf->disks[i].rdev &&
6997 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6998 seq_printf (seq, "]");
7001 static void print_raid5_conf (struct r5conf *conf)
7004 struct disk_info *tmp;
7006 printk(KERN_DEBUG "RAID conf printout:\n");
7008 printk("(conf==NULL)\n");
7011 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7013 conf->raid_disks - conf->mddev->degraded);
7015 for (i = 0; i < conf->raid_disks; i++) {
7016 char b[BDEVNAME_SIZE];
7017 tmp = conf->disks + i;
7019 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7020 i, !test_bit(Faulty, &tmp->rdev->flags),
7021 bdevname(tmp->rdev->bdev, b));
7025 static int raid5_spare_active(struct mddev *mddev)
7028 struct r5conf *conf = mddev->private;
7029 struct disk_info *tmp;
7031 unsigned long flags;
7033 for (i = 0; i < conf->raid_disks; i++) {
7034 tmp = conf->disks + i;
7035 if (tmp->replacement
7036 && tmp->replacement->recovery_offset == MaxSector
7037 && !test_bit(Faulty, &tmp->replacement->flags)
7038 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7039 /* Replacement has just become active. */
7041 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7044 /* Replaced device not technically faulty,
7045 * but we need to be sure it gets removed
7046 * and never re-added.
7048 set_bit(Faulty, &tmp->rdev->flags);
7049 sysfs_notify_dirent_safe(
7050 tmp->rdev->sysfs_state);
7052 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7053 } else if (tmp->rdev
7054 && tmp->rdev->recovery_offset == MaxSector
7055 && !test_bit(Faulty, &tmp->rdev->flags)
7056 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7058 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7061 spin_lock_irqsave(&conf->device_lock, flags);
7062 mddev->degraded = calc_degraded(conf);
7063 spin_unlock_irqrestore(&conf->device_lock, flags);
7064 print_raid5_conf(conf);
7068 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7070 struct r5conf *conf = mddev->private;
7072 int number = rdev->raid_disk;
7073 struct md_rdev **rdevp;
7074 struct disk_info *p = conf->disks + number;
7076 print_raid5_conf(conf);
7077 if (rdev == p->rdev)
7079 else if (rdev == p->replacement)
7080 rdevp = &p->replacement;
7084 if (number >= conf->raid_disks &&
7085 conf->reshape_progress == MaxSector)
7086 clear_bit(In_sync, &rdev->flags);
7088 if (test_bit(In_sync, &rdev->flags) ||
7089 atomic_read(&rdev->nr_pending)) {
7093 /* Only remove non-faulty devices if recovery
7096 if (!test_bit(Faulty, &rdev->flags) &&
7097 mddev->recovery_disabled != conf->recovery_disabled &&
7098 !has_failed(conf) &&
7099 (!p->replacement || p->replacement == rdev) &&
7100 number < conf->raid_disks) {
7106 if (atomic_read(&rdev->nr_pending)) {
7107 /* lost the race, try later */
7110 } else if (p->replacement) {
7111 /* We must have just cleared 'rdev' */
7112 p->rdev = p->replacement;
7113 clear_bit(Replacement, &p->replacement->flags);
7114 smp_mb(); /* Make sure other CPUs may see both as identical
7115 * but will never see neither - if they are careful
7117 p->replacement = NULL;
7118 clear_bit(WantReplacement, &rdev->flags);
7120 /* We might have just removed the Replacement as faulty-
7121 * clear the bit just in case
7123 clear_bit(WantReplacement, &rdev->flags);
7126 print_raid5_conf(conf);
7130 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7132 struct r5conf *conf = mddev->private;
7135 struct disk_info *p;
7137 int last = conf->raid_disks - 1;
7139 if (mddev->recovery_disabled == conf->recovery_disabled)
7142 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7143 /* no point adding a device */
7146 if (rdev->raid_disk >= 0)
7147 first = last = rdev->raid_disk;
7150 * find the disk ... but prefer rdev->saved_raid_disk
7153 if (rdev->saved_raid_disk >= 0 &&
7154 rdev->saved_raid_disk >= first &&
7155 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7156 first = rdev->saved_raid_disk;
7158 for (disk = first; disk <= last; disk++) {
7159 p = conf->disks + disk;
7160 if (p->rdev == NULL) {
7161 clear_bit(In_sync, &rdev->flags);
7162 rdev->raid_disk = disk;
7164 if (rdev->saved_raid_disk != disk)
7166 rcu_assign_pointer(p->rdev, rdev);
7170 for (disk = first; disk <= last; disk++) {
7171 p = conf->disks + disk;
7172 if (test_bit(WantReplacement, &p->rdev->flags) &&
7173 p->replacement == NULL) {
7174 clear_bit(In_sync, &rdev->flags);
7175 set_bit(Replacement, &rdev->flags);
7176 rdev->raid_disk = disk;
7179 rcu_assign_pointer(p->replacement, rdev);
7184 print_raid5_conf(conf);
7188 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7190 /* no resync is happening, and there is enough space
7191 * on all devices, so we can resize.
7192 * We need to make sure resync covers any new space.
7193 * If the array is shrinking we should possibly wait until
7194 * any io in the removed space completes, but it hardly seems
7198 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7199 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7200 if (mddev->external_size &&
7201 mddev->array_sectors > newsize)
7203 if (mddev->bitmap) {
7204 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7208 md_set_array_sectors(mddev, newsize);
7209 set_capacity(mddev->gendisk, mddev->array_sectors);
7210 revalidate_disk(mddev->gendisk);
7211 if (sectors > mddev->dev_sectors &&
7212 mddev->recovery_cp > mddev->dev_sectors) {
7213 mddev->recovery_cp = mddev->dev_sectors;
7214 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7216 mddev->dev_sectors = sectors;
7217 mddev->resync_max_sectors = sectors;
7221 static int check_stripe_cache(struct mddev *mddev)
7223 /* Can only proceed if there are plenty of stripe_heads.
7224 * We need a minimum of one full stripe,, and for sensible progress
7225 * it is best to have about 4 times that.
7226 * If we require 4 times, then the default 256 4K stripe_heads will
7227 * allow for chunk sizes up to 256K, which is probably OK.
7228 * If the chunk size is greater, user-space should request more
7229 * stripe_heads first.
7231 struct r5conf *conf = mddev->private;
7232 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7233 > conf->min_nr_stripes ||
7234 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7235 > conf->min_nr_stripes) {
7236 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7238 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7245 static int check_reshape(struct mddev *mddev)
7247 struct r5conf *conf = mddev->private;
7249 if (mddev->delta_disks == 0 &&
7250 mddev->new_layout == mddev->layout &&
7251 mddev->new_chunk_sectors == mddev->chunk_sectors)
7252 return 0; /* nothing to do */
7253 if (has_failed(conf))
7255 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7256 /* We might be able to shrink, but the devices must
7257 * be made bigger first.
7258 * For raid6, 4 is the minimum size.
7259 * Otherwise 2 is the minimum
7262 if (mddev->level == 6)
7264 if (mddev->raid_disks + mddev->delta_disks < min)
7268 if (!check_stripe_cache(mddev))
7271 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7272 mddev->delta_disks > 0)
7273 if (resize_chunks(conf,
7274 conf->previous_raid_disks
7275 + max(0, mddev->delta_disks),
7276 max(mddev->new_chunk_sectors,
7277 mddev->chunk_sectors)
7280 return resize_stripes(conf, (conf->previous_raid_disks
7281 + mddev->delta_disks));
7284 static int raid5_start_reshape(struct mddev *mddev)
7286 struct r5conf *conf = mddev->private;
7287 struct md_rdev *rdev;
7289 unsigned long flags;
7291 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7294 if (!check_stripe_cache(mddev))
7297 if (has_failed(conf))
7300 rdev_for_each(rdev, mddev) {
7301 if (!test_bit(In_sync, &rdev->flags)
7302 && !test_bit(Faulty, &rdev->flags))
7306 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7307 /* Not enough devices even to make a degraded array
7312 /* Refuse to reduce size of the array. Any reductions in
7313 * array size must be through explicit setting of array_size
7316 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7317 < mddev->array_sectors) {
7318 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7319 "before number of disks\n", mdname(mddev));
7323 atomic_set(&conf->reshape_stripes, 0);
7324 spin_lock_irq(&conf->device_lock);
7325 write_seqcount_begin(&conf->gen_lock);
7326 conf->previous_raid_disks = conf->raid_disks;
7327 conf->raid_disks += mddev->delta_disks;
7328 conf->prev_chunk_sectors = conf->chunk_sectors;
7329 conf->chunk_sectors = mddev->new_chunk_sectors;
7330 conf->prev_algo = conf->algorithm;
7331 conf->algorithm = mddev->new_layout;
7333 /* Code that selects data_offset needs to see the generation update
7334 * if reshape_progress has been set - so a memory barrier needed.
7337 if (mddev->reshape_backwards)
7338 conf->reshape_progress = raid5_size(mddev, 0, 0);
7340 conf->reshape_progress = 0;
7341 conf->reshape_safe = conf->reshape_progress;
7342 write_seqcount_end(&conf->gen_lock);
7343 spin_unlock_irq(&conf->device_lock);
7345 /* Now make sure any requests that proceeded on the assumption
7346 * the reshape wasn't running - like Discard or Read - have
7349 mddev_suspend(mddev);
7350 mddev_resume(mddev);
7352 /* Add some new drives, as many as will fit.
7353 * We know there are enough to make the newly sized array work.
7354 * Don't add devices if we are reducing the number of
7355 * devices in the array. This is because it is not possible
7356 * to correctly record the "partially reconstructed" state of
7357 * such devices during the reshape and confusion could result.
7359 if (mddev->delta_disks >= 0) {
7360 rdev_for_each(rdev, mddev)
7361 if (rdev->raid_disk < 0 &&
7362 !test_bit(Faulty, &rdev->flags)) {
7363 if (raid5_add_disk(mddev, rdev) == 0) {
7365 >= conf->previous_raid_disks)
7366 set_bit(In_sync, &rdev->flags);
7368 rdev->recovery_offset = 0;
7370 if (sysfs_link_rdev(mddev, rdev))
7371 /* Failure here is OK */;
7373 } else if (rdev->raid_disk >= conf->previous_raid_disks
7374 && !test_bit(Faulty, &rdev->flags)) {
7375 /* This is a spare that was manually added */
7376 set_bit(In_sync, &rdev->flags);
7379 /* When a reshape changes the number of devices,
7380 * ->degraded is measured against the larger of the
7381 * pre and post number of devices.
7383 spin_lock_irqsave(&conf->device_lock, flags);
7384 mddev->degraded = calc_degraded(conf);
7385 spin_unlock_irqrestore(&conf->device_lock, flags);
7387 mddev->raid_disks = conf->raid_disks;
7388 mddev->reshape_position = conf->reshape_progress;
7389 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7391 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7392 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7393 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7394 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7395 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7396 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7398 if (!mddev->sync_thread) {
7399 mddev->recovery = 0;
7400 spin_lock_irq(&conf->device_lock);
7401 write_seqcount_begin(&conf->gen_lock);
7402 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7403 mddev->new_chunk_sectors =
7404 conf->chunk_sectors = conf->prev_chunk_sectors;
7405 mddev->new_layout = conf->algorithm = conf->prev_algo;
7406 rdev_for_each(rdev, mddev)
7407 rdev->new_data_offset = rdev->data_offset;
7409 conf->generation --;
7410 conf->reshape_progress = MaxSector;
7411 mddev->reshape_position = MaxSector;
7412 write_seqcount_end(&conf->gen_lock);
7413 spin_unlock_irq(&conf->device_lock);
7416 conf->reshape_checkpoint = jiffies;
7417 md_wakeup_thread(mddev->sync_thread);
7418 md_new_event(mddev);
7422 /* This is called from the reshape thread and should make any
7423 * changes needed in 'conf'
7425 static void end_reshape(struct r5conf *conf)
7428 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7429 struct md_rdev *rdev;
7431 spin_lock_irq(&conf->device_lock);
7432 conf->previous_raid_disks = conf->raid_disks;
7433 rdev_for_each(rdev, conf->mddev)
7434 rdev->data_offset = rdev->new_data_offset;
7436 conf->reshape_progress = MaxSector;
7437 spin_unlock_irq(&conf->device_lock);
7438 wake_up(&conf->wait_for_overlap);
7440 /* read-ahead size must cover two whole stripes, which is
7441 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7443 if (conf->mddev->queue) {
7444 int data_disks = conf->raid_disks - conf->max_degraded;
7445 int stripe = data_disks * ((conf->chunk_sectors << 9)
7447 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7448 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7453 /* This is called from the raid5d thread with mddev_lock held.
7454 * It makes config changes to the device.
7456 static void raid5_finish_reshape(struct mddev *mddev)
7458 struct r5conf *conf = mddev->private;
7460 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7462 if (mddev->delta_disks > 0) {
7463 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7464 set_capacity(mddev->gendisk, mddev->array_sectors);
7465 revalidate_disk(mddev->gendisk);
7468 spin_lock_irq(&conf->device_lock);
7469 mddev->degraded = calc_degraded(conf);
7470 spin_unlock_irq(&conf->device_lock);
7471 for (d = conf->raid_disks ;
7472 d < conf->raid_disks - mddev->delta_disks;
7474 struct md_rdev *rdev = conf->disks[d].rdev;
7476 clear_bit(In_sync, &rdev->flags);
7477 rdev = conf->disks[d].replacement;
7479 clear_bit(In_sync, &rdev->flags);
7482 mddev->layout = conf->algorithm;
7483 mddev->chunk_sectors = conf->chunk_sectors;
7484 mddev->reshape_position = MaxSector;
7485 mddev->delta_disks = 0;
7486 mddev->reshape_backwards = 0;
7490 static void raid5_quiesce(struct mddev *mddev, int state)
7492 struct r5conf *conf = mddev->private;
7495 case 2: /* resume for a suspend */
7496 wake_up(&conf->wait_for_overlap);
7499 case 1: /* stop all writes */
7500 lock_all_device_hash_locks_irq(conf);
7501 /* '2' tells resync/reshape to pause so that all
7502 * active stripes can drain
7505 wait_event_cmd(conf->wait_for_quiescent,
7506 atomic_read(&conf->active_stripes) == 0 &&
7507 atomic_read(&conf->active_aligned_reads) == 0,
7508 unlock_all_device_hash_locks_irq(conf),
7509 lock_all_device_hash_locks_irq(conf));
7511 unlock_all_device_hash_locks_irq(conf);
7512 /* allow reshape to continue */
7513 wake_up(&conf->wait_for_overlap);
7516 case 0: /* re-enable writes */
7517 lock_all_device_hash_locks_irq(conf);
7519 wake_up(&conf->wait_for_quiescent);
7520 wake_up(&conf->wait_for_overlap);
7521 unlock_all_device_hash_locks_irq(conf);
7526 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7528 struct r0conf *raid0_conf = mddev->private;
7531 /* for raid0 takeover only one zone is supported */
7532 if (raid0_conf->nr_strip_zones > 1) {
7533 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7535 return ERR_PTR(-EINVAL);
7538 sectors = raid0_conf->strip_zone[0].zone_end;
7539 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7540 mddev->dev_sectors = sectors;
7541 mddev->new_level = level;
7542 mddev->new_layout = ALGORITHM_PARITY_N;
7543 mddev->new_chunk_sectors = mddev->chunk_sectors;
7544 mddev->raid_disks += 1;
7545 mddev->delta_disks = 1;
7546 /* make sure it will be not marked as dirty */
7547 mddev->recovery_cp = MaxSector;
7549 return setup_conf(mddev);
7552 static void *raid5_takeover_raid1(struct mddev *mddev)
7556 if (mddev->raid_disks != 2 ||
7557 mddev->degraded > 1)
7558 return ERR_PTR(-EINVAL);
7560 /* Should check if there are write-behind devices? */
7562 chunksect = 64*2; /* 64K by default */
7564 /* The array must be an exact multiple of chunksize */
7565 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7568 if ((chunksect<<9) < STRIPE_SIZE)
7569 /* array size does not allow a suitable chunk size */
7570 return ERR_PTR(-EINVAL);
7572 mddev->new_level = 5;
7573 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7574 mddev->new_chunk_sectors = chunksect;
7576 return setup_conf(mddev);
7579 static void *raid5_takeover_raid6(struct mddev *mddev)
7583 switch (mddev->layout) {
7584 case ALGORITHM_LEFT_ASYMMETRIC_6:
7585 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7587 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7588 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7590 case ALGORITHM_LEFT_SYMMETRIC_6:
7591 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7593 case ALGORITHM_RIGHT_SYMMETRIC_6:
7594 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7596 case ALGORITHM_PARITY_0_6:
7597 new_layout = ALGORITHM_PARITY_0;
7599 case ALGORITHM_PARITY_N:
7600 new_layout = ALGORITHM_PARITY_N;
7603 return ERR_PTR(-EINVAL);
7605 mddev->new_level = 5;
7606 mddev->new_layout = new_layout;
7607 mddev->delta_disks = -1;
7608 mddev->raid_disks -= 1;
7609 return setup_conf(mddev);
7612 static int raid5_check_reshape(struct mddev *mddev)
7614 /* For a 2-drive array, the layout and chunk size can be changed
7615 * immediately as not restriping is needed.
7616 * For larger arrays we record the new value - after validation
7617 * to be used by a reshape pass.
7619 struct r5conf *conf = mddev->private;
7620 int new_chunk = mddev->new_chunk_sectors;
7622 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7624 if (new_chunk > 0) {
7625 if (!is_power_of_2(new_chunk))
7627 if (new_chunk < (PAGE_SIZE>>9))
7629 if (mddev->array_sectors & (new_chunk-1))
7630 /* not factor of array size */
7634 /* They look valid */
7636 if (mddev->raid_disks == 2) {
7637 /* can make the change immediately */
7638 if (mddev->new_layout >= 0) {
7639 conf->algorithm = mddev->new_layout;
7640 mddev->layout = mddev->new_layout;
7642 if (new_chunk > 0) {
7643 conf->chunk_sectors = new_chunk ;
7644 mddev->chunk_sectors = new_chunk;
7646 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7647 md_wakeup_thread(mddev->thread);
7649 return check_reshape(mddev);
7652 static int raid6_check_reshape(struct mddev *mddev)
7654 int new_chunk = mddev->new_chunk_sectors;
7656 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7658 if (new_chunk > 0) {
7659 if (!is_power_of_2(new_chunk))
7661 if (new_chunk < (PAGE_SIZE >> 9))
7663 if (mddev->array_sectors & (new_chunk-1))
7664 /* not factor of array size */
7668 /* They look valid */
7669 return check_reshape(mddev);
7672 static void *raid5_takeover(struct mddev *mddev)
7674 /* raid5 can take over:
7675 * raid0 - if there is only one strip zone - make it a raid4 layout
7676 * raid1 - if there are two drives. We need to know the chunk size
7677 * raid4 - trivial - just use a raid4 layout.
7678 * raid6 - Providing it is a *_6 layout
7680 if (mddev->level == 0)
7681 return raid45_takeover_raid0(mddev, 5);
7682 if (mddev->level == 1)
7683 return raid5_takeover_raid1(mddev);
7684 if (mddev->level == 4) {
7685 mddev->new_layout = ALGORITHM_PARITY_N;
7686 mddev->new_level = 5;
7687 return setup_conf(mddev);
7689 if (mddev->level == 6)
7690 return raid5_takeover_raid6(mddev);
7692 return ERR_PTR(-EINVAL);
7695 static void *raid4_takeover(struct mddev *mddev)
7697 /* raid4 can take over:
7698 * raid0 - if there is only one strip zone
7699 * raid5 - if layout is right
7701 if (mddev->level == 0)
7702 return raid45_takeover_raid0(mddev, 4);
7703 if (mddev->level == 5 &&
7704 mddev->layout == ALGORITHM_PARITY_N) {
7705 mddev->new_layout = 0;
7706 mddev->new_level = 4;
7707 return setup_conf(mddev);
7709 return ERR_PTR(-EINVAL);
7712 static struct md_personality raid5_personality;
7714 static void *raid6_takeover(struct mddev *mddev)
7716 /* Currently can only take over a raid5. We map the
7717 * personality to an equivalent raid6 personality
7718 * with the Q block at the end.
7722 if (mddev->pers != &raid5_personality)
7723 return ERR_PTR(-EINVAL);
7724 if (mddev->degraded > 1)
7725 return ERR_PTR(-EINVAL);
7726 if (mddev->raid_disks > 253)
7727 return ERR_PTR(-EINVAL);
7728 if (mddev->raid_disks < 3)
7729 return ERR_PTR(-EINVAL);
7731 switch (mddev->layout) {
7732 case ALGORITHM_LEFT_ASYMMETRIC:
7733 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7735 case ALGORITHM_RIGHT_ASYMMETRIC:
7736 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7738 case ALGORITHM_LEFT_SYMMETRIC:
7739 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7741 case ALGORITHM_RIGHT_SYMMETRIC:
7742 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7744 case ALGORITHM_PARITY_0:
7745 new_layout = ALGORITHM_PARITY_0_6;
7747 case ALGORITHM_PARITY_N:
7748 new_layout = ALGORITHM_PARITY_N;
7751 return ERR_PTR(-EINVAL);
7753 mddev->new_level = 6;
7754 mddev->new_layout = new_layout;
7755 mddev->delta_disks = 1;
7756 mddev->raid_disks += 1;
7757 return setup_conf(mddev);
7760 static struct md_personality raid6_personality =
7764 .owner = THIS_MODULE,
7765 .make_request = make_request,
7769 .error_handler = error,
7770 .hot_add_disk = raid5_add_disk,
7771 .hot_remove_disk= raid5_remove_disk,
7772 .spare_active = raid5_spare_active,
7773 .sync_request = sync_request,
7774 .resize = raid5_resize,
7776 .check_reshape = raid6_check_reshape,
7777 .start_reshape = raid5_start_reshape,
7778 .finish_reshape = raid5_finish_reshape,
7779 .quiesce = raid5_quiesce,
7780 .takeover = raid6_takeover,
7781 .congested = raid5_congested,
7782 .mergeable_bvec = raid5_mergeable_bvec,
7784 static struct md_personality raid5_personality =
7788 .owner = THIS_MODULE,
7789 .make_request = make_request,
7793 .error_handler = error,
7794 .hot_add_disk = raid5_add_disk,
7795 .hot_remove_disk= raid5_remove_disk,
7796 .spare_active = raid5_spare_active,
7797 .sync_request = sync_request,
7798 .resize = raid5_resize,
7800 .check_reshape = raid5_check_reshape,
7801 .start_reshape = raid5_start_reshape,
7802 .finish_reshape = raid5_finish_reshape,
7803 .quiesce = raid5_quiesce,
7804 .takeover = raid5_takeover,
7805 .congested = raid5_congested,
7806 .mergeable_bvec = raid5_mergeable_bvec,
7809 static struct md_personality raid4_personality =
7813 .owner = THIS_MODULE,
7814 .make_request = make_request,
7818 .error_handler = error,
7819 .hot_add_disk = raid5_add_disk,
7820 .hot_remove_disk= raid5_remove_disk,
7821 .spare_active = raid5_spare_active,
7822 .sync_request = sync_request,
7823 .resize = raid5_resize,
7825 .check_reshape = raid5_check_reshape,
7826 .start_reshape = raid5_start_reshape,
7827 .finish_reshape = raid5_finish_reshape,
7828 .quiesce = raid5_quiesce,
7829 .takeover = raid4_takeover,
7830 .congested = raid5_congested,
7831 .mergeable_bvec = raid5_mergeable_bvec,
7834 static int __init raid5_init(void)
7836 raid5_wq = alloc_workqueue("raid5wq",
7837 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7840 register_md_personality(&raid6_personality);
7841 register_md_personality(&raid5_personality);
7842 register_md_personality(&raid4_personality);
7846 static void raid5_exit(void)
7848 unregister_md_personality(&raid6_personality);
7849 unregister_md_personality(&raid5_personality);
7850 unregister_md_personality(&raid4_personality);
7851 destroy_workqueue(raid5_wq);
7854 module_init(raid5_init);
7855 module_exit(raid5_exit);
7856 MODULE_LICENSE("GPL");
7857 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7858 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7859 MODULE_ALIAS("md-raid5");
7860 MODULE_ALIAS("md-raid4");
7861 MODULE_ALIAS("md-level-5");
7862 MODULE_ALIAS("md-level-4");
7863 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7864 MODULE_ALIAS("md-raid6");
7865 MODULE_ALIAS("md-level-6");
7867 /* This used to be two separate modules, they were: */
7868 MODULE_ALIAS("raid5");
7869 MODULE_ALIAS("raid6");