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_list *return_bi)
229 while ((bi = bio_list_pop(return_bi)) != NULL) {
230 bi->bi_iter.bi_size = 0;
231 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
237 static void print_raid5_conf (struct r5conf *conf);
239 static int stripe_operations_active(struct stripe_head *sh)
241 return sh->check_state || sh->reconstruct_state ||
242 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
243 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
246 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
248 struct r5conf *conf = sh->raid_conf;
249 struct r5worker_group *group;
251 int i, cpu = sh->cpu;
253 if (!cpu_online(cpu)) {
254 cpu = cpumask_any(cpu_online_mask);
258 if (list_empty(&sh->lru)) {
259 struct r5worker_group *group;
260 group = conf->worker_groups + cpu_to_group(cpu);
261 list_add_tail(&sh->lru, &group->handle_list);
262 group->stripes_cnt++;
266 if (conf->worker_cnt_per_group == 0) {
267 md_wakeup_thread(conf->mddev->thread);
271 group = conf->worker_groups + cpu_to_group(sh->cpu);
273 group->workers[0].working = true;
274 /* at least one worker should run to avoid race */
275 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
277 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
278 /* wakeup more workers */
279 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
280 if (group->workers[i].working == false) {
281 group->workers[i].working = true;
282 queue_work_on(sh->cpu, raid5_wq,
283 &group->workers[i].work);
289 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
290 struct list_head *temp_inactive_list)
292 BUG_ON(!list_empty(&sh->lru));
293 BUG_ON(atomic_read(&conf->active_stripes)==0);
294 if (test_bit(STRIPE_HANDLE, &sh->state)) {
295 if (test_bit(STRIPE_DELAYED, &sh->state) &&
296 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
297 list_add_tail(&sh->lru, &conf->delayed_list);
298 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
299 sh->bm_seq - conf->seq_write > 0)
300 list_add_tail(&sh->lru, &conf->bitmap_list);
302 clear_bit(STRIPE_DELAYED, &sh->state);
303 clear_bit(STRIPE_BIT_DELAY, &sh->state);
304 if (conf->worker_cnt_per_group == 0) {
305 list_add_tail(&sh->lru, &conf->handle_list);
307 raid5_wakeup_stripe_thread(sh);
311 md_wakeup_thread(conf->mddev->thread);
313 BUG_ON(stripe_operations_active(sh));
314 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
315 if (atomic_dec_return(&conf->preread_active_stripes)
317 md_wakeup_thread(conf->mddev->thread);
318 atomic_dec(&conf->active_stripes);
319 if (!test_bit(STRIPE_EXPANDING, &sh->state))
320 list_add_tail(&sh->lru, temp_inactive_list);
324 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
325 struct list_head *temp_inactive_list)
327 if (atomic_dec_and_test(&sh->count))
328 do_release_stripe(conf, sh, temp_inactive_list);
332 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
334 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
335 * given time. Adding stripes only takes device lock, while deleting stripes
336 * only takes hash lock.
338 static void release_inactive_stripe_list(struct r5conf *conf,
339 struct list_head *temp_inactive_list,
343 bool do_wakeup = false;
346 if (hash == NR_STRIPE_HASH_LOCKS) {
347 size = NR_STRIPE_HASH_LOCKS;
348 hash = NR_STRIPE_HASH_LOCKS - 1;
352 struct list_head *list = &temp_inactive_list[size - 1];
355 * We don't hold any lock here yet, raid5_get_active_stripe() might
356 * remove stripes from the list
358 if (!list_empty_careful(list)) {
359 spin_lock_irqsave(conf->hash_locks + hash, flags);
360 if (list_empty(conf->inactive_list + hash) &&
362 atomic_dec(&conf->empty_inactive_list_nr);
363 list_splice_tail_init(list, conf->inactive_list + hash);
365 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
372 wake_up(&conf->wait_for_stripe);
373 if (atomic_read(&conf->active_stripes) == 0)
374 wake_up(&conf->wait_for_quiescent);
375 if (conf->retry_read_aligned)
376 md_wakeup_thread(conf->mddev->thread);
380 /* should hold conf->device_lock already */
381 static int release_stripe_list(struct r5conf *conf,
382 struct list_head *temp_inactive_list)
384 struct stripe_head *sh;
386 struct llist_node *head;
388 head = llist_del_all(&conf->released_stripes);
389 head = llist_reverse_order(head);
393 sh = llist_entry(head, struct stripe_head, release_list);
394 head = llist_next(head);
395 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
397 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
399 * Don't worry the bit is set here, because if the bit is set
400 * again, the count is always > 1. This is true for
401 * STRIPE_ON_UNPLUG_LIST bit too.
403 hash = sh->hash_lock_index;
404 __release_stripe(conf, sh, &temp_inactive_list[hash]);
411 void raid5_release_stripe(struct stripe_head *sh)
413 struct r5conf *conf = sh->raid_conf;
415 struct list_head list;
419 /* Avoid release_list until the last reference.
421 if (atomic_add_unless(&sh->count, -1, 1))
424 if (unlikely(!conf->mddev->thread) ||
425 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
427 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
429 md_wakeup_thread(conf->mddev->thread);
432 local_irq_save(flags);
433 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
434 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
435 INIT_LIST_HEAD(&list);
436 hash = sh->hash_lock_index;
437 do_release_stripe(conf, sh, &list);
438 spin_unlock(&conf->device_lock);
439 release_inactive_stripe_list(conf, &list, hash);
441 local_irq_restore(flags);
444 static inline void remove_hash(struct stripe_head *sh)
446 pr_debug("remove_hash(), stripe %llu\n",
447 (unsigned long long)sh->sector);
449 hlist_del_init(&sh->hash);
452 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
454 struct hlist_head *hp = stripe_hash(conf, sh->sector);
456 pr_debug("insert_hash(), stripe %llu\n",
457 (unsigned long long)sh->sector);
459 hlist_add_head(&sh->hash, hp);
462 /* find an idle stripe, make sure it is unhashed, and return it. */
463 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
465 struct stripe_head *sh = NULL;
466 struct list_head *first;
468 if (list_empty(conf->inactive_list + hash))
470 first = (conf->inactive_list + hash)->next;
471 sh = list_entry(first, struct stripe_head, lru);
472 list_del_init(first);
474 atomic_inc(&conf->active_stripes);
475 BUG_ON(hash != sh->hash_lock_index);
476 if (list_empty(conf->inactive_list + hash))
477 atomic_inc(&conf->empty_inactive_list_nr);
482 static void shrink_buffers(struct stripe_head *sh)
486 int num = sh->raid_conf->pool_size;
488 for (i = 0; i < num ; i++) {
489 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
493 sh->dev[i].page = NULL;
498 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
501 int num = sh->raid_conf->pool_size;
503 for (i = 0; i < num; i++) {
506 if (!(page = alloc_page(gfp))) {
509 sh->dev[i].page = page;
510 sh->dev[i].orig_page = page;
515 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
516 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
517 struct stripe_head *sh);
519 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
521 struct r5conf *conf = sh->raid_conf;
524 BUG_ON(atomic_read(&sh->count) != 0);
525 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
526 BUG_ON(stripe_operations_active(sh));
527 BUG_ON(sh->batch_head);
529 pr_debug("init_stripe called, stripe %llu\n",
530 (unsigned long long)sector);
532 seq = read_seqcount_begin(&conf->gen_lock);
533 sh->generation = conf->generation - previous;
534 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
536 stripe_set_idx(sector, conf, previous, sh);
539 for (i = sh->disks; i--; ) {
540 struct r5dev *dev = &sh->dev[i];
542 if (dev->toread || dev->read || dev->towrite || dev->written ||
543 test_bit(R5_LOCKED, &dev->flags)) {
544 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
545 (unsigned long long)sh->sector, i, dev->toread,
546 dev->read, dev->towrite, dev->written,
547 test_bit(R5_LOCKED, &dev->flags));
551 raid5_build_block(sh, i, previous);
553 if (read_seqcount_retry(&conf->gen_lock, seq))
555 sh->overwrite_disks = 0;
556 insert_hash(conf, sh);
557 sh->cpu = smp_processor_id();
558 set_bit(STRIPE_BATCH_READY, &sh->state);
561 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
564 struct stripe_head *sh;
566 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
567 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
568 if (sh->sector == sector && sh->generation == generation)
570 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
575 * Need to check if array has failed when deciding whether to:
577 * - remove non-faulty devices
580 * This determination is simple when no reshape is happening.
581 * However if there is a reshape, we need to carefully check
582 * both the before and after sections.
583 * This is because some failed devices may only affect one
584 * of the two sections, and some non-in_sync devices may
585 * be insync in the section most affected by failed devices.
587 static int calc_degraded(struct r5conf *conf)
589 int degraded, degraded2;
594 for (i = 0; i < conf->previous_raid_disks; i++) {
595 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
596 if (rdev && test_bit(Faulty, &rdev->flags))
597 rdev = rcu_dereference(conf->disks[i].replacement);
598 if (!rdev || test_bit(Faulty, &rdev->flags))
600 else if (test_bit(In_sync, &rdev->flags))
603 /* not in-sync or faulty.
604 * If the reshape increases the number of devices,
605 * this is being recovered by the reshape, so
606 * this 'previous' section is not in_sync.
607 * If the number of devices is being reduced however,
608 * the device can only be part of the array if
609 * we are reverting a reshape, so this section will
612 if (conf->raid_disks >= conf->previous_raid_disks)
616 if (conf->raid_disks == conf->previous_raid_disks)
620 for (i = 0; i < conf->raid_disks; i++) {
621 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
622 if (rdev && test_bit(Faulty, &rdev->flags))
623 rdev = rcu_dereference(conf->disks[i].replacement);
624 if (!rdev || test_bit(Faulty, &rdev->flags))
626 else if (test_bit(In_sync, &rdev->flags))
629 /* not in-sync or faulty.
630 * If reshape increases the number of devices, this
631 * section has already been recovered, else it
632 * almost certainly hasn't.
634 if (conf->raid_disks <= conf->previous_raid_disks)
638 if (degraded2 > degraded)
643 static int has_failed(struct r5conf *conf)
647 if (conf->mddev->reshape_position == MaxSector)
648 return conf->mddev->degraded > conf->max_degraded;
650 degraded = calc_degraded(conf);
651 if (degraded > conf->max_degraded)
657 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
658 int previous, int noblock, int noquiesce)
660 struct stripe_head *sh;
661 int hash = stripe_hash_locks_hash(sector);
662 int inc_empty_inactive_list_flag;
664 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
666 spin_lock_irq(conf->hash_locks + hash);
669 wait_event_lock_irq(conf->wait_for_quiescent,
670 conf->quiesce == 0 || noquiesce,
671 *(conf->hash_locks + hash));
672 sh = __find_stripe(conf, sector, conf->generation - previous);
674 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
675 sh = get_free_stripe(conf, hash);
676 if (!sh && !test_bit(R5_DID_ALLOC,
678 set_bit(R5_ALLOC_MORE,
681 if (noblock && sh == NULL)
684 set_bit(R5_INACTIVE_BLOCKED,
687 conf->wait_for_stripe,
688 !list_empty(conf->inactive_list + hash) &&
689 (atomic_read(&conf->active_stripes)
690 < (conf->max_nr_stripes * 3 / 4)
691 || !test_bit(R5_INACTIVE_BLOCKED,
692 &conf->cache_state)),
693 *(conf->hash_locks + hash));
694 clear_bit(R5_INACTIVE_BLOCKED,
697 init_stripe(sh, sector, previous);
698 atomic_inc(&sh->count);
700 } else if (!atomic_inc_not_zero(&sh->count)) {
701 spin_lock(&conf->device_lock);
702 if (!atomic_read(&sh->count)) {
703 if (!test_bit(STRIPE_HANDLE, &sh->state))
704 atomic_inc(&conf->active_stripes);
705 BUG_ON(list_empty(&sh->lru) &&
706 !test_bit(STRIPE_EXPANDING, &sh->state));
707 inc_empty_inactive_list_flag = 0;
708 if (!list_empty(conf->inactive_list + hash))
709 inc_empty_inactive_list_flag = 1;
710 list_del_init(&sh->lru);
711 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
712 atomic_inc(&conf->empty_inactive_list_nr);
714 sh->group->stripes_cnt--;
718 atomic_inc(&sh->count);
719 spin_unlock(&conf->device_lock);
721 } while (sh == NULL);
723 spin_unlock_irq(conf->hash_locks + hash);
727 static bool is_full_stripe_write(struct stripe_head *sh)
729 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
730 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
733 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
737 spin_lock(&sh2->stripe_lock);
738 spin_lock_nested(&sh1->stripe_lock, 1);
740 spin_lock(&sh1->stripe_lock);
741 spin_lock_nested(&sh2->stripe_lock, 1);
745 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
747 spin_unlock(&sh1->stripe_lock);
748 spin_unlock(&sh2->stripe_lock);
752 /* Only freshly new full stripe normal write stripe can be added to a batch list */
753 static bool stripe_can_batch(struct stripe_head *sh)
755 struct r5conf *conf = sh->raid_conf;
759 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
760 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
761 is_full_stripe_write(sh);
764 /* we only do back search */
765 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
767 struct stripe_head *head;
768 sector_t head_sector, tmp_sec;
771 int inc_empty_inactive_list_flag;
773 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
774 tmp_sec = sh->sector;
775 if (!sector_div(tmp_sec, conf->chunk_sectors))
777 head_sector = sh->sector - STRIPE_SECTORS;
779 hash = stripe_hash_locks_hash(head_sector);
780 spin_lock_irq(conf->hash_locks + hash);
781 head = __find_stripe(conf, head_sector, conf->generation);
782 if (head && !atomic_inc_not_zero(&head->count)) {
783 spin_lock(&conf->device_lock);
784 if (!atomic_read(&head->count)) {
785 if (!test_bit(STRIPE_HANDLE, &head->state))
786 atomic_inc(&conf->active_stripes);
787 BUG_ON(list_empty(&head->lru) &&
788 !test_bit(STRIPE_EXPANDING, &head->state));
789 inc_empty_inactive_list_flag = 0;
790 if (!list_empty(conf->inactive_list + hash))
791 inc_empty_inactive_list_flag = 1;
792 list_del_init(&head->lru);
793 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
794 atomic_inc(&conf->empty_inactive_list_nr);
796 head->group->stripes_cnt--;
800 atomic_inc(&head->count);
801 spin_unlock(&conf->device_lock);
803 spin_unlock_irq(conf->hash_locks + hash);
807 if (!stripe_can_batch(head))
810 lock_two_stripes(head, sh);
811 /* clear_batch_ready clear the flag */
812 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
819 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
821 if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw ||
822 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
825 if (head->batch_head) {
826 spin_lock(&head->batch_head->batch_lock);
827 /* This batch list is already running */
828 if (!stripe_can_batch(head)) {
829 spin_unlock(&head->batch_head->batch_lock);
834 * at this point, head's BATCH_READY could be cleared, but we
835 * can still add the stripe to batch list
837 list_add(&sh->batch_list, &head->batch_list);
838 spin_unlock(&head->batch_head->batch_lock);
840 sh->batch_head = head->batch_head;
842 head->batch_head = head;
843 sh->batch_head = head->batch_head;
844 spin_lock(&head->batch_lock);
845 list_add_tail(&sh->batch_list, &head->batch_list);
846 spin_unlock(&head->batch_lock);
849 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
850 if (atomic_dec_return(&conf->preread_active_stripes)
852 md_wakeup_thread(conf->mddev->thread);
854 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
855 int seq = sh->bm_seq;
856 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
857 sh->batch_head->bm_seq > seq)
858 seq = sh->batch_head->bm_seq;
859 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
860 sh->batch_head->bm_seq = seq;
863 atomic_inc(&sh->count);
865 unlock_two_stripes(head, sh);
867 raid5_release_stripe(head);
870 /* Determine if 'data_offset' or 'new_data_offset' should be used
871 * in this stripe_head.
873 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
875 sector_t progress = conf->reshape_progress;
876 /* Need a memory barrier to make sure we see the value
877 * of conf->generation, or ->data_offset that was set before
878 * reshape_progress was updated.
881 if (progress == MaxSector)
883 if (sh->generation == conf->generation - 1)
885 /* We are in a reshape, and this is a new-generation stripe,
886 * so use new_data_offset.
892 raid5_end_read_request(struct bio *bi);
894 raid5_end_write_request(struct bio *bi);
896 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
898 struct r5conf *conf = sh->raid_conf;
899 int i, disks = sh->disks;
900 struct stripe_head *head_sh = sh;
904 if (r5l_write_stripe(conf->log, sh) == 0)
906 for (i = disks; i--; ) {
907 int op, op_flags = 0;
908 int replace_only = 0;
909 struct bio *bi, *rbi;
910 struct md_rdev *rdev, *rrdev = NULL;
913 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
915 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
916 op_flags = WRITE_FUA;
917 if (test_bit(R5_Discard, &sh->dev[i].flags))
919 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
921 else if (test_and_clear_bit(R5_WantReplace,
922 &sh->dev[i].flags)) {
927 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
928 op_flags |= REQ_SYNC;
931 bi = &sh->dev[i].req;
932 rbi = &sh->dev[i].rreq; /* For writing to replacement */
935 rrdev = rcu_dereference(conf->disks[i].replacement);
936 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
937 rdev = rcu_dereference(conf->disks[i].rdev);
942 if (op_is_write(op)) {
946 /* We raced and saw duplicates */
949 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
954 if (rdev && test_bit(Faulty, &rdev->flags))
957 atomic_inc(&rdev->nr_pending);
958 if (rrdev && test_bit(Faulty, &rrdev->flags))
961 atomic_inc(&rrdev->nr_pending);
964 /* We have already checked bad blocks for reads. Now
965 * need to check for writes. We never accept write errors
966 * on the replacement, so we don't to check rrdev.
968 while (op_is_write(op) && rdev &&
969 test_bit(WriteErrorSeen, &rdev->flags)) {
972 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
973 &first_bad, &bad_sectors);
978 set_bit(BlockedBadBlocks, &rdev->flags);
979 if (!conf->mddev->external &&
980 conf->mddev->flags) {
981 /* It is very unlikely, but we might
982 * still need to write out the
983 * bad block log - better give it
985 md_check_recovery(conf->mddev);
988 * Because md_wait_for_blocked_rdev
989 * will dec nr_pending, we must
990 * increment it first.
992 atomic_inc(&rdev->nr_pending);
993 md_wait_for_blocked_rdev(rdev, conf->mddev);
995 /* Acknowledged bad block - skip the write */
996 rdev_dec_pending(rdev, conf->mddev);
1002 if (s->syncing || s->expanding || s->expanded
1004 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1006 set_bit(STRIPE_IO_STARTED, &sh->state);
1008 bi->bi_bdev = rdev->bdev;
1009 bio_set_op_attrs(bi, op, op_flags);
1010 bi->bi_end_io = op_is_write(op)
1011 ? raid5_end_write_request
1012 : raid5_end_read_request;
1013 bi->bi_private = sh;
1015 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1016 __func__, (unsigned long long)sh->sector,
1018 atomic_inc(&sh->count);
1020 atomic_inc(&head_sh->count);
1021 if (use_new_offset(conf, sh))
1022 bi->bi_iter.bi_sector = (sh->sector
1023 + rdev->new_data_offset);
1025 bi->bi_iter.bi_sector = (sh->sector
1026 + rdev->data_offset);
1027 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1028 bi->bi_rw |= REQ_NOMERGE;
1030 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1031 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1032 sh->dev[i].vec.bv_page = sh->dev[i].page;
1034 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1035 bi->bi_io_vec[0].bv_offset = 0;
1036 bi->bi_iter.bi_size = STRIPE_SIZE;
1038 * If this is discard request, set bi_vcnt 0. We don't
1039 * want to confuse SCSI because SCSI will replace payload
1041 if (op == REQ_OP_DISCARD)
1044 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1046 if (conf->mddev->gendisk)
1047 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1048 bi, disk_devt(conf->mddev->gendisk),
1050 generic_make_request(bi);
1053 if (s->syncing || s->expanding || s->expanded
1055 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1057 set_bit(STRIPE_IO_STARTED, &sh->state);
1059 rbi->bi_bdev = rrdev->bdev;
1060 bio_set_op_attrs(rbi, op, op_flags);
1061 BUG_ON(!op_is_write(op));
1062 rbi->bi_end_io = raid5_end_write_request;
1063 rbi->bi_private = sh;
1065 pr_debug("%s: for %llu schedule op %d on "
1066 "replacement disc %d\n",
1067 __func__, (unsigned long long)sh->sector,
1069 atomic_inc(&sh->count);
1071 atomic_inc(&head_sh->count);
1072 if (use_new_offset(conf, sh))
1073 rbi->bi_iter.bi_sector = (sh->sector
1074 + rrdev->new_data_offset);
1076 rbi->bi_iter.bi_sector = (sh->sector
1077 + rrdev->data_offset);
1078 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1079 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1080 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1082 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1083 rbi->bi_io_vec[0].bv_offset = 0;
1084 rbi->bi_iter.bi_size = STRIPE_SIZE;
1086 * If this is discard request, set bi_vcnt 0. We don't
1087 * want to confuse SCSI because SCSI will replace payload
1089 if (op == REQ_OP_DISCARD)
1091 if (conf->mddev->gendisk)
1092 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1093 rbi, disk_devt(conf->mddev->gendisk),
1095 generic_make_request(rbi);
1097 if (!rdev && !rrdev) {
1098 if (op_is_write(op))
1099 set_bit(STRIPE_DEGRADED, &sh->state);
1100 pr_debug("skip op %d on disc %d for sector %llu\n",
1101 bi->bi_rw, i, (unsigned long long)sh->sector);
1102 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1103 set_bit(STRIPE_HANDLE, &sh->state);
1106 if (!head_sh->batch_head)
1108 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1115 static struct dma_async_tx_descriptor *
1116 async_copy_data(int frombio, struct bio *bio, struct page **page,
1117 sector_t sector, struct dma_async_tx_descriptor *tx,
1118 struct stripe_head *sh)
1121 struct bvec_iter iter;
1122 struct page *bio_page;
1124 struct async_submit_ctl submit;
1125 enum async_tx_flags flags = 0;
1127 if (bio->bi_iter.bi_sector >= sector)
1128 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1130 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1133 flags |= ASYNC_TX_FENCE;
1134 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1136 bio_for_each_segment(bvl, bio, iter) {
1137 int len = bvl.bv_len;
1141 if (page_offset < 0) {
1142 b_offset = -page_offset;
1143 page_offset += b_offset;
1147 if (len > 0 && page_offset + len > STRIPE_SIZE)
1148 clen = STRIPE_SIZE - page_offset;
1153 b_offset += bvl.bv_offset;
1154 bio_page = bvl.bv_page;
1156 if (sh->raid_conf->skip_copy &&
1157 b_offset == 0 && page_offset == 0 &&
1158 clen == STRIPE_SIZE)
1161 tx = async_memcpy(*page, bio_page, page_offset,
1162 b_offset, clen, &submit);
1164 tx = async_memcpy(bio_page, *page, b_offset,
1165 page_offset, clen, &submit);
1167 /* chain the operations */
1168 submit.depend_tx = tx;
1170 if (clen < len) /* hit end of page */
1178 static void ops_complete_biofill(void *stripe_head_ref)
1180 struct stripe_head *sh = stripe_head_ref;
1181 struct bio_list return_bi = BIO_EMPTY_LIST;
1184 pr_debug("%s: stripe %llu\n", __func__,
1185 (unsigned long long)sh->sector);
1187 /* clear completed biofills */
1188 for (i = sh->disks; i--; ) {
1189 struct r5dev *dev = &sh->dev[i];
1191 /* acknowledge completion of a biofill operation */
1192 /* and check if we need to reply to a read request,
1193 * new R5_Wantfill requests are held off until
1194 * !STRIPE_BIOFILL_RUN
1196 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1197 struct bio *rbi, *rbi2;
1202 while (rbi && rbi->bi_iter.bi_sector <
1203 dev->sector + STRIPE_SECTORS) {
1204 rbi2 = r5_next_bio(rbi, dev->sector);
1205 if (!raid5_dec_bi_active_stripes(rbi))
1206 bio_list_add(&return_bi, rbi);
1211 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1213 return_io(&return_bi);
1215 set_bit(STRIPE_HANDLE, &sh->state);
1216 raid5_release_stripe(sh);
1219 static void ops_run_biofill(struct stripe_head *sh)
1221 struct dma_async_tx_descriptor *tx = NULL;
1222 struct async_submit_ctl submit;
1225 BUG_ON(sh->batch_head);
1226 pr_debug("%s: stripe %llu\n", __func__,
1227 (unsigned long long)sh->sector);
1229 for (i = sh->disks; i--; ) {
1230 struct r5dev *dev = &sh->dev[i];
1231 if (test_bit(R5_Wantfill, &dev->flags)) {
1233 spin_lock_irq(&sh->stripe_lock);
1234 dev->read = rbi = dev->toread;
1236 spin_unlock_irq(&sh->stripe_lock);
1237 while (rbi && rbi->bi_iter.bi_sector <
1238 dev->sector + STRIPE_SECTORS) {
1239 tx = async_copy_data(0, rbi, &dev->page,
1240 dev->sector, tx, sh);
1241 rbi = r5_next_bio(rbi, dev->sector);
1246 atomic_inc(&sh->count);
1247 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1248 async_trigger_callback(&submit);
1251 static void mark_target_uptodate(struct stripe_head *sh, int target)
1258 tgt = &sh->dev[target];
1259 set_bit(R5_UPTODATE, &tgt->flags);
1260 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1261 clear_bit(R5_Wantcompute, &tgt->flags);
1264 static void ops_complete_compute(void *stripe_head_ref)
1266 struct stripe_head *sh = stripe_head_ref;
1268 pr_debug("%s: stripe %llu\n", __func__,
1269 (unsigned long long)sh->sector);
1271 /* mark the computed target(s) as uptodate */
1272 mark_target_uptodate(sh, sh->ops.target);
1273 mark_target_uptodate(sh, sh->ops.target2);
1275 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1276 if (sh->check_state == check_state_compute_run)
1277 sh->check_state = check_state_compute_result;
1278 set_bit(STRIPE_HANDLE, &sh->state);
1279 raid5_release_stripe(sh);
1282 /* return a pointer to the address conversion region of the scribble buffer */
1283 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1284 struct raid5_percpu *percpu, int i)
1288 addr = flex_array_get(percpu->scribble, i);
1289 return addr + sizeof(struct page *) * (sh->disks + 2);
1292 /* return a pointer to the address conversion region of the scribble buffer */
1293 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1297 addr = flex_array_get(percpu->scribble, i);
1301 static struct dma_async_tx_descriptor *
1302 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1304 int disks = sh->disks;
1305 struct page **xor_srcs = to_addr_page(percpu, 0);
1306 int target = sh->ops.target;
1307 struct r5dev *tgt = &sh->dev[target];
1308 struct page *xor_dest = tgt->page;
1310 struct dma_async_tx_descriptor *tx;
1311 struct async_submit_ctl submit;
1314 BUG_ON(sh->batch_head);
1316 pr_debug("%s: stripe %llu block: %d\n",
1317 __func__, (unsigned long long)sh->sector, target);
1318 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1320 for (i = disks; i--; )
1322 xor_srcs[count++] = sh->dev[i].page;
1324 atomic_inc(&sh->count);
1326 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1327 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1328 if (unlikely(count == 1))
1329 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1331 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1336 /* set_syndrome_sources - populate source buffers for gen_syndrome
1337 * @srcs - (struct page *) array of size sh->disks
1338 * @sh - stripe_head to parse
1340 * Populates srcs in proper layout order for the stripe and returns the
1341 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1342 * destination buffer is recorded in srcs[count] and the Q destination
1343 * is recorded in srcs[count+1]].
1345 static int set_syndrome_sources(struct page **srcs,
1346 struct stripe_head *sh,
1349 int disks = sh->disks;
1350 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1351 int d0_idx = raid6_d0(sh);
1355 for (i = 0; i < disks; i++)
1361 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1362 struct r5dev *dev = &sh->dev[i];
1364 if (i == sh->qd_idx || i == sh->pd_idx ||
1365 (srctype == SYNDROME_SRC_ALL) ||
1366 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1367 test_bit(R5_Wantdrain, &dev->flags)) ||
1368 (srctype == SYNDROME_SRC_WRITTEN &&
1370 srcs[slot] = sh->dev[i].page;
1371 i = raid6_next_disk(i, disks);
1372 } while (i != d0_idx);
1374 return syndrome_disks;
1377 static struct dma_async_tx_descriptor *
1378 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1380 int disks = sh->disks;
1381 struct page **blocks = to_addr_page(percpu, 0);
1383 int qd_idx = sh->qd_idx;
1384 struct dma_async_tx_descriptor *tx;
1385 struct async_submit_ctl submit;
1391 BUG_ON(sh->batch_head);
1392 if (sh->ops.target < 0)
1393 target = sh->ops.target2;
1394 else if (sh->ops.target2 < 0)
1395 target = sh->ops.target;
1397 /* we should only have one valid target */
1400 pr_debug("%s: stripe %llu block: %d\n",
1401 __func__, (unsigned long long)sh->sector, target);
1403 tgt = &sh->dev[target];
1404 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1407 atomic_inc(&sh->count);
1409 if (target == qd_idx) {
1410 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1411 blocks[count] = NULL; /* regenerating p is not necessary */
1412 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1413 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1414 ops_complete_compute, sh,
1415 to_addr_conv(sh, percpu, 0));
1416 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1418 /* Compute any data- or p-drive using XOR */
1420 for (i = disks; i-- ; ) {
1421 if (i == target || i == qd_idx)
1423 blocks[count++] = sh->dev[i].page;
1426 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1427 NULL, ops_complete_compute, sh,
1428 to_addr_conv(sh, percpu, 0));
1429 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1435 static struct dma_async_tx_descriptor *
1436 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1438 int i, count, disks = sh->disks;
1439 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1440 int d0_idx = raid6_d0(sh);
1441 int faila = -1, failb = -1;
1442 int target = sh->ops.target;
1443 int target2 = sh->ops.target2;
1444 struct r5dev *tgt = &sh->dev[target];
1445 struct r5dev *tgt2 = &sh->dev[target2];
1446 struct dma_async_tx_descriptor *tx;
1447 struct page **blocks = to_addr_page(percpu, 0);
1448 struct async_submit_ctl submit;
1450 BUG_ON(sh->batch_head);
1451 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1452 __func__, (unsigned long long)sh->sector, target, target2);
1453 BUG_ON(target < 0 || target2 < 0);
1454 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1455 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1457 /* we need to open-code set_syndrome_sources to handle the
1458 * slot number conversion for 'faila' and 'failb'
1460 for (i = 0; i < disks ; i++)
1465 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1467 blocks[slot] = sh->dev[i].page;
1473 i = raid6_next_disk(i, disks);
1474 } while (i != d0_idx);
1476 BUG_ON(faila == failb);
1479 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1480 __func__, (unsigned long long)sh->sector, faila, failb);
1482 atomic_inc(&sh->count);
1484 if (failb == syndrome_disks+1) {
1485 /* Q disk is one of the missing disks */
1486 if (faila == syndrome_disks) {
1487 /* Missing P+Q, just recompute */
1488 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1489 ops_complete_compute, sh,
1490 to_addr_conv(sh, percpu, 0));
1491 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1492 STRIPE_SIZE, &submit);
1496 int qd_idx = sh->qd_idx;
1498 /* Missing D+Q: recompute D from P, then recompute Q */
1499 if (target == qd_idx)
1500 data_target = target2;
1502 data_target = target;
1505 for (i = disks; i-- ; ) {
1506 if (i == data_target || i == qd_idx)
1508 blocks[count++] = sh->dev[i].page;
1510 dest = sh->dev[data_target].page;
1511 init_async_submit(&submit,
1512 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1514 to_addr_conv(sh, percpu, 0));
1515 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1518 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1519 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1520 ops_complete_compute, sh,
1521 to_addr_conv(sh, percpu, 0));
1522 return async_gen_syndrome(blocks, 0, count+2,
1523 STRIPE_SIZE, &submit);
1526 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1527 ops_complete_compute, sh,
1528 to_addr_conv(sh, percpu, 0));
1529 if (failb == syndrome_disks) {
1530 /* We're missing D+P. */
1531 return async_raid6_datap_recov(syndrome_disks+2,
1535 /* We're missing D+D. */
1536 return async_raid6_2data_recov(syndrome_disks+2,
1537 STRIPE_SIZE, faila, failb,
1543 static void ops_complete_prexor(void *stripe_head_ref)
1545 struct stripe_head *sh = stripe_head_ref;
1547 pr_debug("%s: stripe %llu\n", __func__,
1548 (unsigned long long)sh->sector);
1551 static struct dma_async_tx_descriptor *
1552 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1553 struct dma_async_tx_descriptor *tx)
1555 int disks = sh->disks;
1556 struct page **xor_srcs = to_addr_page(percpu, 0);
1557 int count = 0, pd_idx = sh->pd_idx, i;
1558 struct async_submit_ctl submit;
1560 /* existing parity data subtracted */
1561 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1563 BUG_ON(sh->batch_head);
1564 pr_debug("%s: stripe %llu\n", __func__,
1565 (unsigned long long)sh->sector);
1567 for (i = disks; i--; ) {
1568 struct r5dev *dev = &sh->dev[i];
1569 /* Only process blocks that are known to be uptodate */
1570 if (test_bit(R5_Wantdrain, &dev->flags))
1571 xor_srcs[count++] = dev->page;
1574 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1575 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1576 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1581 static struct dma_async_tx_descriptor *
1582 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1583 struct dma_async_tx_descriptor *tx)
1585 struct page **blocks = to_addr_page(percpu, 0);
1587 struct async_submit_ctl submit;
1589 pr_debug("%s: stripe %llu\n", __func__,
1590 (unsigned long long)sh->sector);
1592 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1594 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1595 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1596 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1601 static struct dma_async_tx_descriptor *
1602 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1604 int disks = sh->disks;
1606 struct stripe_head *head_sh = sh;
1608 pr_debug("%s: stripe %llu\n", __func__,
1609 (unsigned long long)sh->sector);
1611 for (i = disks; i--; ) {
1616 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1621 spin_lock_irq(&sh->stripe_lock);
1622 chosen = dev->towrite;
1623 dev->towrite = NULL;
1624 sh->overwrite_disks = 0;
1625 BUG_ON(dev->written);
1626 wbi = dev->written = chosen;
1627 spin_unlock_irq(&sh->stripe_lock);
1628 WARN_ON(dev->page != dev->orig_page);
1630 while (wbi && wbi->bi_iter.bi_sector <
1631 dev->sector + STRIPE_SECTORS) {
1632 if (wbi->bi_rw & REQ_FUA)
1633 set_bit(R5_WantFUA, &dev->flags);
1634 if (wbi->bi_rw & REQ_SYNC)
1635 set_bit(R5_SyncIO, &dev->flags);
1636 if (bio_op(wbi) == REQ_OP_DISCARD)
1637 set_bit(R5_Discard, &dev->flags);
1639 tx = async_copy_data(1, wbi, &dev->page,
1640 dev->sector, tx, sh);
1641 if (dev->page != dev->orig_page) {
1642 set_bit(R5_SkipCopy, &dev->flags);
1643 clear_bit(R5_UPTODATE, &dev->flags);
1644 clear_bit(R5_OVERWRITE, &dev->flags);
1647 wbi = r5_next_bio(wbi, dev->sector);
1650 if (head_sh->batch_head) {
1651 sh = list_first_entry(&sh->batch_list,
1664 static void ops_complete_reconstruct(void *stripe_head_ref)
1666 struct stripe_head *sh = stripe_head_ref;
1667 int disks = sh->disks;
1668 int pd_idx = sh->pd_idx;
1669 int qd_idx = sh->qd_idx;
1671 bool fua = false, sync = false, discard = false;
1673 pr_debug("%s: stripe %llu\n", __func__,
1674 (unsigned long long)sh->sector);
1676 for (i = disks; i--; ) {
1677 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1678 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1679 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1682 for (i = disks; i--; ) {
1683 struct r5dev *dev = &sh->dev[i];
1685 if (dev->written || i == pd_idx || i == qd_idx) {
1686 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1687 set_bit(R5_UPTODATE, &dev->flags);
1689 set_bit(R5_WantFUA, &dev->flags);
1691 set_bit(R5_SyncIO, &dev->flags);
1695 if (sh->reconstruct_state == reconstruct_state_drain_run)
1696 sh->reconstruct_state = reconstruct_state_drain_result;
1697 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1698 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1700 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1701 sh->reconstruct_state = reconstruct_state_result;
1704 set_bit(STRIPE_HANDLE, &sh->state);
1705 raid5_release_stripe(sh);
1709 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1710 struct dma_async_tx_descriptor *tx)
1712 int disks = sh->disks;
1713 struct page **xor_srcs;
1714 struct async_submit_ctl submit;
1715 int count, pd_idx = sh->pd_idx, i;
1716 struct page *xor_dest;
1718 unsigned long flags;
1720 struct stripe_head *head_sh = sh;
1723 pr_debug("%s: stripe %llu\n", __func__,
1724 (unsigned long long)sh->sector);
1726 for (i = 0; i < sh->disks; i++) {
1729 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1732 if (i >= sh->disks) {
1733 atomic_inc(&sh->count);
1734 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1735 ops_complete_reconstruct(sh);
1740 xor_srcs = to_addr_page(percpu, j);
1741 /* check if prexor is active which means only process blocks
1742 * that are part of a read-modify-write (written)
1744 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1746 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1747 for (i = disks; i--; ) {
1748 struct r5dev *dev = &sh->dev[i];
1749 if (head_sh->dev[i].written)
1750 xor_srcs[count++] = dev->page;
1753 xor_dest = sh->dev[pd_idx].page;
1754 for (i = disks; i--; ) {
1755 struct r5dev *dev = &sh->dev[i];
1757 xor_srcs[count++] = dev->page;
1761 /* 1/ if we prexor'd then the dest is reused as a source
1762 * 2/ if we did not prexor then we are redoing the parity
1763 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1764 * for the synchronous xor case
1766 last_stripe = !head_sh->batch_head ||
1767 list_first_entry(&sh->batch_list,
1768 struct stripe_head, batch_list) == head_sh;
1770 flags = ASYNC_TX_ACK |
1771 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1773 atomic_inc(&head_sh->count);
1774 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1775 to_addr_conv(sh, percpu, j));
1777 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1778 init_async_submit(&submit, flags, tx, NULL, NULL,
1779 to_addr_conv(sh, percpu, j));
1782 if (unlikely(count == 1))
1783 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1785 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1788 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1795 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1796 struct dma_async_tx_descriptor *tx)
1798 struct async_submit_ctl submit;
1799 struct page **blocks;
1800 int count, i, j = 0;
1801 struct stripe_head *head_sh = sh;
1804 unsigned long txflags;
1806 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1808 for (i = 0; i < sh->disks; i++) {
1809 if (sh->pd_idx == i || sh->qd_idx == i)
1811 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1814 if (i >= sh->disks) {
1815 atomic_inc(&sh->count);
1816 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1817 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1818 ops_complete_reconstruct(sh);
1823 blocks = to_addr_page(percpu, j);
1825 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1826 synflags = SYNDROME_SRC_WRITTEN;
1827 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1829 synflags = SYNDROME_SRC_ALL;
1830 txflags = ASYNC_TX_ACK;
1833 count = set_syndrome_sources(blocks, sh, synflags);
1834 last_stripe = !head_sh->batch_head ||
1835 list_first_entry(&sh->batch_list,
1836 struct stripe_head, batch_list) == head_sh;
1839 atomic_inc(&head_sh->count);
1840 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1841 head_sh, to_addr_conv(sh, percpu, j));
1843 init_async_submit(&submit, 0, tx, NULL, NULL,
1844 to_addr_conv(sh, percpu, j));
1845 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1848 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1854 static void ops_complete_check(void *stripe_head_ref)
1856 struct stripe_head *sh = stripe_head_ref;
1858 pr_debug("%s: stripe %llu\n", __func__,
1859 (unsigned long long)sh->sector);
1861 sh->check_state = check_state_check_result;
1862 set_bit(STRIPE_HANDLE, &sh->state);
1863 raid5_release_stripe(sh);
1866 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1868 int disks = sh->disks;
1869 int pd_idx = sh->pd_idx;
1870 int qd_idx = sh->qd_idx;
1871 struct page *xor_dest;
1872 struct page **xor_srcs = to_addr_page(percpu, 0);
1873 struct dma_async_tx_descriptor *tx;
1874 struct async_submit_ctl submit;
1878 pr_debug("%s: stripe %llu\n", __func__,
1879 (unsigned long long)sh->sector);
1881 BUG_ON(sh->batch_head);
1883 xor_dest = sh->dev[pd_idx].page;
1884 xor_srcs[count++] = xor_dest;
1885 for (i = disks; i--; ) {
1886 if (i == pd_idx || i == qd_idx)
1888 xor_srcs[count++] = sh->dev[i].page;
1891 init_async_submit(&submit, 0, NULL, NULL, NULL,
1892 to_addr_conv(sh, percpu, 0));
1893 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1894 &sh->ops.zero_sum_result, &submit);
1896 atomic_inc(&sh->count);
1897 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1898 tx = async_trigger_callback(&submit);
1901 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1903 struct page **srcs = to_addr_page(percpu, 0);
1904 struct async_submit_ctl submit;
1907 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1908 (unsigned long long)sh->sector, checkp);
1910 BUG_ON(sh->batch_head);
1911 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1915 atomic_inc(&sh->count);
1916 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1917 sh, to_addr_conv(sh, percpu, 0));
1918 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1919 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1922 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1924 int overlap_clear = 0, i, disks = sh->disks;
1925 struct dma_async_tx_descriptor *tx = NULL;
1926 struct r5conf *conf = sh->raid_conf;
1927 int level = conf->level;
1928 struct raid5_percpu *percpu;
1932 percpu = per_cpu_ptr(conf->percpu, cpu);
1933 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1934 ops_run_biofill(sh);
1938 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1940 tx = ops_run_compute5(sh, percpu);
1942 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1943 tx = ops_run_compute6_1(sh, percpu);
1945 tx = ops_run_compute6_2(sh, percpu);
1947 /* terminate the chain if reconstruct is not set to be run */
1948 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1952 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1954 tx = ops_run_prexor5(sh, percpu, tx);
1956 tx = ops_run_prexor6(sh, percpu, tx);
1959 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1960 tx = ops_run_biodrain(sh, tx);
1964 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1966 ops_run_reconstruct5(sh, percpu, tx);
1968 ops_run_reconstruct6(sh, percpu, tx);
1971 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1972 if (sh->check_state == check_state_run)
1973 ops_run_check_p(sh, percpu);
1974 else if (sh->check_state == check_state_run_q)
1975 ops_run_check_pq(sh, percpu, 0);
1976 else if (sh->check_state == check_state_run_pq)
1977 ops_run_check_pq(sh, percpu, 1);
1982 if (overlap_clear && !sh->batch_head)
1983 for (i = disks; i--; ) {
1984 struct r5dev *dev = &sh->dev[i];
1985 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1986 wake_up(&sh->raid_conf->wait_for_overlap);
1991 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
1994 struct stripe_head *sh;
1997 sh = kmem_cache_zalloc(sc, gfp);
1999 spin_lock_init(&sh->stripe_lock);
2000 spin_lock_init(&sh->batch_lock);
2001 INIT_LIST_HEAD(&sh->batch_list);
2002 INIT_LIST_HEAD(&sh->lru);
2003 atomic_set(&sh->count, 1);
2004 for (i = 0; i < disks; i++) {
2005 struct r5dev *dev = &sh->dev[i];
2007 bio_init(&dev->req);
2008 bio_init(&dev->rreq);
2013 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2015 struct stripe_head *sh;
2017 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size);
2021 sh->raid_conf = conf;
2023 if (grow_buffers(sh, gfp)) {
2025 kmem_cache_free(conf->slab_cache, sh);
2028 sh->hash_lock_index =
2029 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2030 /* we just created an active stripe so... */
2031 atomic_inc(&conf->active_stripes);
2033 raid5_release_stripe(sh);
2034 conf->max_nr_stripes++;
2038 static int grow_stripes(struct r5conf *conf, int num)
2040 struct kmem_cache *sc;
2041 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2043 if (conf->mddev->gendisk)
2044 sprintf(conf->cache_name[0],
2045 "raid%d-%s", conf->level, mdname(conf->mddev));
2047 sprintf(conf->cache_name[0],
2048 "raid%d-%p", conf->level, conf->mddev);
2049 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2051 conf->active_name = 0;
2052 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2053 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2057 conf->slab_cache = sc;
2058 conf->pool_size = devs;
2060 if (!grow_one_stripe(conf, GFP_KERNEL))
2067 * scribble_len - return the required size of the scribble region
2068 * @num - total number of disks in the array
2070 * The size must be enough to contain:
2071 * 1/ a struct page pointer for each device in the array +2
2072 * 2/ room to convert each entry in (1) to its corresponding dma
2073 * (dma_map_page()) or page (page_address()) address.
2075 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2076 * calculate over all devices (not just the data blocks), using zeros in place
2077 * of the P and Q blocks.
2079 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2081 struct flex_array *ret;
2084 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2085 ret = flex_array_alloc(len, cnt, flags);
2088 /* always prealloc all elements, so no locking is required */
2089 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2090 flex_array_free(ret);
2096 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2102 * Never shrink. And mddev_suspend() could deadlock if this is called
2103 * from raid5d. In that case, scribble_disks and scribble_sectors
2104 * should equal to new_disks and new_sectors
2106 if (conf->scribble_disks >= new_disks &&
2107 conf->scribble_sectors >= new_sectors)
2109 mddev_suspend(conf->mddev);
2111 for_each_present_cpu(cpu) {
2112 struct raid5_percpu *percpu;
2113 struct flex_array *scribble;
2115 percpu = per_cpu_ptr(conf->percpu, cpu);
2116 scribble = scribble_alloc(new_disks,
2117 new_sectors / STRIPE_SECTORS,
2121 flex_array_free(percpu->scribble);
2122 percpu->scribble = scribble;
2129 mddev_resume(conf->mddev);
2131 conf->scribble_disks = new_disks;
2132 conf->scribble_sectors = new_sectors;
2137 static int resize_stripes(struct r5conf *conf, int newsize)
2139 /* Make all the stripes able to hold 'newsize' devices.
2140 * New slots in each stripe get 'page' set to a new page.
2142 * This happens in stages:
2143 * 1/ create a new kmem_cache and allocate the required number of
2145 * 2/ gather all the old stripe_heads and transfer the pages across
2146 * to the new stripe_heads. This will have the side effect of
2147 * freezing the array as once all stripe_heads have been collected,
2148 * no IO will be possible. Old stripe heads are freed once their
2149 * pages have been transferred over, and the old kmem_cache is
2150 * freed when all stripes are done.
2151 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2152 * we simple return a failre status - no need to clean anything up.
2153 * 4/ allocate new pages for the new slots in the new stripe_heads.
2154 * If this fails, we don't bother trying the shrink the
2155 * stripe_heads down again, we just leave them as they are.
2156 * As each stripe_head is processed the new one is released into
2159 * Once step2 is started, we cannot afford to wait for a write,
2160 * so we use GFP_NOIO allocations.
2162 struct stripe_head *osh, *nsh;
2163 LIST_HEAD(newstripes);
2164 struct disk_info *ndisks;
2166 struct kmem_cache *sc;
2170 if (newsize <= conf->pool_size)
2171 return 0; /* never bother to shrink */
2173 err = md_allow_write(conf->mddev);
2178 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2179 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2184 /* Need to ensure auto-resizing doesn't interfere */
2185 mutex_lock(&conf->cache_size_mutex);
2187 for (i = conf->max_nr_stripes; i; i--) {
2188 nsh = alloc_stripe(sc, GFP_KERNEL, newsize);
2192 nsh->raid_conf = conf;
2193 list_add(&nsh->lru, &newstripes);
2196 /* didn't get enough, give up */
2197 while (!list_empty(&newstripes)) {
2198 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2199 list_del(&nsh->lru);
2200 kmem_cache_free(sc, nsh);
2202 kmem_cache_destroy(sc);
2203 mutex_unlock(&conf->cache_size_mutex);
2206 /* Step 2 - Must use GFP_NOIO now.
2207 * OK, we have enough stripes, start collecting inactive
2208 * stripes and copying them over
2212 list_for_each_entry(nsh, &newstripes, lru) {
2213 lock_device_hash_lock(conf, hash);
2214 wait_event_cmd(conf->wait_for_stripe,
2215 !list_empty(conf->inactive_list + hash),
2216 unlock_device_hash_lock(conf, hash),
2217 lock_device_hash_lock(conf, hash));
2218 osh = get_free_stripe(conf, hash);
2219 unlock_device_hash_lock(conf, hash);
2221 for(i=0; i<conf->pool_size; i++) {
2222 nsh->dev[i].page = osh->dev[i].page;
2223 nsh->dev[i].orig_page = osh->dev[i].page;
2225 nsh->hash_lock_index = hash;
2226 kmem_cache_free(conf->slab_cache, osh);
2228 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2229 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2234 kmem_cache_destroy(conf->slab_cache);
2237 * At this point, we are holding all the stripes so the array
2238 * is completely stalled, so now is a good time to resize
2239 * conf->disks and the scribble region
2241 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2243 for (i=0; i<conf->raid_disks; i++)
2244 ndisks[i] = conf->disks[i];
2246 conf->disks = ndisks;
2250 mutex_unlock(&conf->cache_size_mutex);
2251 /* Step 4, return new stripes to service */
2252 while(!list_empty(&newstripes)) {
2253 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2254 list_del_init(&nsh->lru);
2256 for (i=conf->raid_disks; i < newsize; i++)
2257 if (nsh->dev[i].page == NULL) {
2258 struct page *p = alloc_page(GFP_NOIO);
2259 nsh->dev[i].page = p;
2260 nsh->dev[i].orig_page = p;
2264 raid5_release_stripe(nsh);
2266 /* critical section pass, GFP_NOIO no longer needed */
2268 conf->slab_cache = sc;
2269 conf->active_name = 1-conf->active_name;
2271 conf->pool_size = newsize;
2275 static int drop_one_stripe(struct r5conf *conf)
2277 struct stripe_head *sh;
2278 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2280 spin_lock_irq(conf->hash_locks + hash);
2281 sh = get_free_stripe(conf, hash);
2282 spin_unlock_irq(conf->hash_locks + hash);
2285 BUG_ON(atomic_read(&sh->count));
2287 kmem_cache_free(conf->slab_cache, sh);
2288 atomic_dec(&conf->active_stripes);
2289 conf->max_nr_stripes--;
2293 static void shrink_stripes(struct r5conf *conf)
2295 while (conf->max_nr_stripes &&
2296 drop_one_stripe(conf))
2299 kmem_cache_destroy(conf->slab_cache);
2300 conf->slab_cache = NULL;
2303 static void raid5_end_read_request(struct bio * bi)
2305 struct stripe_head *sh = bi->bi_private;
2306 struct r5conf *conf = sh->raid_conf;
2307 int disks = sh->disks, i;
2308 char b[BDEVNAME_SIZE];
2309 struct md_rdev *rdev = NULL;
2312 for (i=0 ; i<disks; i++)
2313 if (bi == &sh->dev[i].req)
2316 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2317 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2324 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2325 /* If replacement finished while this request was outstanding,
2326 * 'replacement' might be NULL already.
2327 * In that case it moved down to 'rdev'.
2328 * rdev is not removed until all requests are finished.
2330 rdev = conf->disks[i].replacement;
2332 rdev = conf->disks[i].rdev;
2334 if (use_new_offset(conf, sh))
2335 s = sh->sector + rdev->new_data_offset;
2337 s = sh->sector + rdev->data_offset;
2338 if (!bi->bi_error) {
2339 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2340 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2341 /* Note that this cannot happen on a
2342 * replacement device. We just fail those on
2347 "md/raid:%s: read error corrected"
2348 " (%lu sectors at %llu on %s)\n",
2349 mdname(conf->mddev), STRIPE_SECTORS,
2350 (unsigned long long)s,
2351 bdevname(rdev->bdev, b));
2352 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2353 clear_bit(R5_ReadError, &sh->dev[i].flags);
2354 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2355 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2356 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2358 if (atomic_read(&rdev->read_errors))
2359 atomic_set(&rdev->read_errors, 0);
2361 const char *bdn = bdevname(rdev->bdev, b);
2365 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2366 atomic_inc(&rdev->read_errors);
2367 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2370 "md/raid:%s: read error on replacement device "
2371 "(sector %llu on %s).\n",
2372 mdname(conf->mddev),
2373 (unsigned long long)s,
2375 else if (conf->mddev->degraded >= conf->max_degraded) {
2379 "md/raid:%s: read error not correctable "
2380 "(sector %llu on %s).\n",
2381 mdname(conf->mddev),
2382 (unsigned long long)s,
2384 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2389 "md/raid:%s: read error NOT corrected!! "
2390 "(sector %llu on %s).\n",
2391 mdname(conf->mddev),
2392 (unsigned long long)s,
2394 } else if (atomic_read(&rdev->read_errors)
2395 > conf->max_nr_stripes)
2397 "md/raid:%s: Too many read errors, failing device %s.\n",
2398 mdname(conf->mddev), bdn);
2401 if (set_bad && test_bit(In_sync, &rdev->flags)
2402 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2405 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2406 set_bit(R5_ReadError, &sh->dev[i].flags);
2407 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2409 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2411 clear_bit(R5_ReadError, &sh->dev[i].flags);
2412 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2414 && test_bit(In_sync, &rdev->flags)
2415 && rdev_set_badblocks(
2416 rdev, sh->sector, STRIPE_SECTORS, 0)))
2417 md_error(conf->mddev, rdev);
2420 rdev_dec_pending(rdev, conf->mddev);
2421 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2422 set_bit(STRIPE_HANDLE, &sh->state);
2423 raid5_release_stripe(sh);
2427 static void raid5_end_write_request(struct bio *bi)
2429 struct stripe_head *sh = bi->bi_private;
2430 struct r5conf *conf = sh->raid_conf;
2431 int disks = sh->disks, i;
2432 struct md_rdev *uninitialized_var(rdev);
2435 int replacement = 0;
2437 for (i = 0 ; i < disks; i++) {
2438 if (bi == &sh->dev[i].req) {
2439 rdev = conf->disks[i].rdev;
2442 if (bi == &sh->dev[i].rreq) {
2443 rdev = conf->disks[i].replacement;
2447 /* rdev was removed and 'replacement'
2448 * replaced it. rdev is not removed
2449 * until all requests are finished.
2451 rdev = conf->disks[i].rdev;
2455 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2456 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2466 md_error(conf->mddev, rdev);
2467 else if (is_badblock(rdev, sh->sector,
2469 &first_bad, &bad_sectors))
2470 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2473 set_bit(STRIPE_DEGRADED, &sh->state);
2474 set_bit(WriteErrorSeen, &rdev->flags);
2475 set_bit(R5_WriteError, &sh->dev[i].flags);
2476 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2477 set_bit(MD_RECOVERY_NEEDED,
2478 &rdev->mddev->recovery);
2479 } else if (is_badblock(rdev, sh->sector,
2481 &first_bad, &bad_sectors)) {
2482 set_bit(R5_MadeGood, &sh->dev[i].flags);
2483 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2484 /* That was a successful write so make
2485 * sure it looks like we already did
2488 set_bit(R5_ReWrite, &sh->dev[i].flags);
2491 rdev_dec_pending(rdev, conf->mddev);
2493 if (sh->batch_head && bi->bi_error && !replacement)
2494 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2496 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2497 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2498 set_bit(STRIPE_HANDLE, &sh->state);
2499 raid5_release_stripe(sh);
2501 if (sh->batch_head && sh != sh->batch_head)
2502 raid5_release_stripe(sh->batch_head);
2506 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2508 struct r5dev *dev = &sh->dev[i];
2510 dev->req.bi_io_vec = &dev->vec;
2511 dev->req.bi_max_vecs = 1;
2512 dev->req.bi_private = sh;
2514 dev->rreq.bi_io_vec = &dev->rvec;
2515 dev->rreq.bi_max_vecs = 1;
2516 dev->rreq.bi_private = sh;
2519 dev->sector = raid5_compute_blocknr(sh, i, previous);
2522 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2524 char b[BDEVNAME_SIZE];
2525 struct r5conf *conf = mddev->private;
2526 unsigned long flags;
2527 pr_debug("raid456: error called\n");
2529 spin_lock_irqsave(&conf->device_lock, flags);
2530 clear_bit(In_sync, &rdev->flags);
2531 mddev->degraded = calc_degraded(conf);
2532 spin_unlock_irqrestore(&conf->device_lock, flags);
2533 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2535 set_bit(Blocked, &rdev->flags);
2536 set_bit(Faulty, &rdev->flags);
2537 set_mask_bits(&mddev->flags, 0,
2538 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
2540 "md/raid:%s: Disk failure on %s, disabling device.\n"
2541 "md/raid:%s: Operation continuing on %d devices.\n",
2543 bdevname(rdev->bdev, b),
2545 conf->raid_disks - mddev->degraded);
2549 * Input: a 'big' sector number,
2550 * Output: index of the data and parity disk, and the sector # in them.
2552 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2553 int previous, int *dd_idx,
2554 struct stripe_head *sh)
2556 sector_t stripe, stripe2;
2557 sector_t chunk_number;
2558 unsigned int chunk_offset;
2561 sector_t new_sector;
2562 int algorithm = previous ? conf->prev_algo
2564 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2565 : conf->chunk_sectors;
2566 int raid_disks = previous ? conf->previous_raid_disks
2568 int data_disks = raid_disks - conf->max_degraded;
2570 /* First compute the information on this sector */
2573 * Compute the chunk number and the sector offset inside the chunk
2575 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2576 chunk_number = r_sector;
2579 * Compute the stripe number
2581 stripe = chunk_number;
2582 *dd_idx = sector_div(stripe, data_disks);
2585 * Select the parity disk based on the user selected algorithm.
2587 pd_idx = qd_idx = -1;
2588 switch(conf->level) {
2590 pd_idx = data_disks;
2593 switch (algorithm) {
2594 case ALGORITHM_LEFT_ASYMMETRIC:
2595 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2596 if (*dd_idx >= pd_idx)
2599 case ALGORITHM_RIGHT_ASYMMETRIC:
2600 pd_idx = sector_div(stripe2, raid_disks);
2601 if (*dd_idx >= pd_idx)
2604 case ALGORITHM_LEFT_SYMMETRIC:
2605 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2606 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2608 case ALGORITHM_RIGHT_SYMMETRIC:
2609 pd_idx = sector_div(stripe2, raid_disks);
2610 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2612 case ALGORITHM_PARITY_0:
2616 case ALGORITHM_PARITY_N:
2617 pd_idx = data_disks;
2625 switch (algorithm) {
2626 case ALGORITHM_LEFT_ASYMMETRIC:
2627 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2628 qd_idx = pd_idx + 1;
2629 if (pd_idx == raid_disks-1) {
2630 (*dd_idx)++; /* Q D D D P */
2632 } else if (*dd_idx >= pd_idx)
2633 (*dd_idx) += 2; /* D D P Q D */
2635 case ALGORITHM_RIGHT_ASYMMETRIC:
2636 pd_idx = sector_div(stripe2, raid_disks);
2637 qd_idx = pd_idx + 1;
2638 if (pd_idx == raid_disks-1) {
2639 (*dd_idx)++; /* Q D D D P */
2641 } else if (*dd_idx >= pd_idx)
2642 (*dd_idx) += 2; /* D D P Q D */
2644 case ALGORITHM_LEFT_SYMMETRIC:
2645 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2646 qd_idx = (pd_idx + 1) % raid_disks;
2647 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2649 case ALGORITHM_RIGHT_SYMMETRIC:
2650 pd_idx = sector_div(stripe2, raid_disks);
2651 qd_idx = (pd_idx + 1) % raid_disks;
2652 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2655 case ALGORITHM_PARITY_0:
2660 case ALGORITHM_PARITY_N:
2661 pd_idx = data_disks;
2662 qd_idx = data_disks + 1;
2665 case ALGORITHM_ROTATING_ZERO_RESTART:
2666 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2667 * of blocks for computing Q is different.
2669 pd_idx = sector_div(stripe2, raid_disks);
2670 qd_idx = pd_idx + 1;
2671 if (pd_idx == raid_disks-1) {
2672 (*dd_idx)++; /* Q D D D P */
2674 } else if (*dd_idx >= pd_idx)
2675 (*dd_idx) += 2; /* D D P Q D */
2679 case ALGORITHM_ROTATING_N_RESTART:
2680 /* Same a left_asymmetric, by first stripe is
2681 * D D D P Q rather than
2685 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2686 qd_idx = pd_idx + 1;
2687 if (pd_idx == raid_disks-1) {
2688 (*dd_idx)++; /* Q D D D P */
2690 } else if (*dd_idx >= pd_idx)
2691 (*dd_idx) += 2; /* D D P Q D */
2695 case ALGORITHM_ROTATING_N_CONTINUE:
2696 /* Same as left_symmetric but Q is before P */
2697 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2698 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2699 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2703 case ALGORITHM_LEFT_ASYMMETRIC_6:
2704 /* RAID5 left_asymmetric, with Q on last device */
2705 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2706 if (*dd_idx >= pd_idx)
2708 qd_idx = raid_disks - 1;
2711 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2712 pd_idx = sector_div(stripe2, raid_disks-1);
2713 if (*dd_idx >= pd_idx)
2715 qd_idx = raid_disks - 1;
2718 case ALGORITHM_LEFT_SYMMETRIC_6:
2719 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2720 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2721 qd_idx = raid_disks - 1;
2724 case ALGORITHM_RIGHT_SYMMETRIC_6:
2725 pd_idx = sector_div(stripe2, raid_disks-1);
2726 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2727 qd_idx = raid_disks - 1;
2730 case ALGORITHM_PARITY_0_6:
2733 qd_idx = raid_disks - 1;
2743 sh->pd_idx = pd_idx;
2744 sh->qd_idx = qd_idx;
2745 sh->ddf_layout = ddf_layout;
2748 * Finally, compute the new sector number
2750 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2754 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2756 struct r5conf *conf = sh->raid_conf;
2757 int raid_disks = sh->disks;
2758 int data_disks = raid_disks - conf->max_degraded;
2759 sector_t new_sector = sh->sector, check;
2760 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2761 : conf->chunk_sectors;
2762 int algorithm = previous ? conf->prev_algo
2766 sector_t chunk_number;
2767 int dummy1, dd_idx = i;
2769 struct stripe_head sh2;
2771 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2772 stripe = new_sector;
2774 if (i == sh->pd_idx)
2776 switch(conf->level) {
2779 switch (algorithm) {
2780 case ALGORITHM_LEFT_ASYMMETRIC:
2781 case ALGORITHM_RIGHT_ASYMMETRIC:
2785 case ALGORITHM_LEFT_SYMMETRIC:
2786 case ALGORITHM_RIGHT_SYMMETRIC:
2789 i -= (sh->pd_idx + 1);
2791 case ALGORITHM_PARITY_0:
2794 case ALGORITHM_PARITY_N:
2801 if (i == sh->qd_idx)
2802 return 0; /* It is the Q disk */
2803 switch (algorithm) {
2804 case ALGORITHM_LEFT_ASYMMETRIC:
2805 case ALGORITHM_RIGHT_ASYMMETRIC:
2806 case ALGORITHM_ROTATING_ZERO_RESTART:
2807 case ALGORITHM_ROTATING_N_RESTART:
2808 if (sh->pd_idx == raid_disks-1)
2809 i--; /* Q D D D P */
2810 else if (i > sh->pd_idx)
2811 i -= 2; /* D D P Q D */
2813 case ALGORITHM_LEFT_SYMMETRIC:
2814 case ALGORITHM_RIGHT_SYMMETRIC:
2815 if (sh->pd_idx == raid_disks-1)
2816 i--; /* Q D D D P */
2821 i -= (sh->pd_idx + 2);
2824 case ALGORITHM_PARITY_0:
2827 case ALGORITHM_PARITY_N:
2829 case ALGORITHM_ROTATING_N_CONTINUE:
2830 /* Like left_symmetric, but P is before Q */
2831 if (sh->pd_idx == 0)
2832 i--; /* P D D D Q */
2837 i -= (sh->pd_idx + 1);
2840 case ALGORITHM_LEFT_ASYMMETRIC_6:
2841 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2845 case ALGORITHM_LEFT_SYMMETRIC_6:
2846 case ALGORITHM_RIGHT_SYMMETRIC_6:
2848 i += data_disks + 1;
2849 i -= (sh->pd_idx + 1);
2851 case ALGORITHM_PARITY_0_6:
2860 chunk_number = stripe * data_disks + i;
2861 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2863 check = raid5_compute_sector(conf, r_sector,
2864 previous, &dummy1, &sh2);
2865 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2866 || sh2.qd_idx != sh->qd_idx) {
2867 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2868 mdname(conf->mddev));
2875 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2876 int rcw, int expand)
2878 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2879 struct r5conf *conf = sh->raid_conf;
2880 int level = conf->level;
2884 for (i = disks; i--; ) {
2885 struct r5dev *dev = &sh->dev[i];
2888 set_bit(R5_LOCKED, &dev->flags);
2889 set_bit(R5_Wantdrain, &dev->flags);
2891 clear_bit(R5_UPTODATE, &dev->flags);
2895 /* if we are not expanding this is a proper write request, and
2896 * there will be bios with new data to be drained into the
2901 /* False alarm, nothing to do */
2903 sh->reconstruct_state = reconstruct_state_drain_run;
2904 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2906 sh->reconstruct_state = reconstruct_state_run;
2908 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2910 if (s->locked + conf->max_degraded == disks)
2911 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2912 atomic_inc(&conf->pending_full_writes);
2914 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2915 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2916 BUG_ON(level == 6 &&
2917 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2918 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2920 for (i = disks; i--; ) {
2921 struct r5dev *dev = &sh->dev[i];
2922 if (i == pd_idx || i == qd_idx)
2926 (test_bit(R5_UPTODATE, &dev->flags) ||
2927 test_bit(R5_Wantcompute, &dev->flags))) {
2928 set_bit(R5_Wantdrain, &dev->flags);
2929 set_bit(R5_LOCKED, &dev->flags);
2930 clear_bit(R5_UPTODATE, &dev->flags);
2935 /* False alarm - nothing to do */
2937 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2938 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2939 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2940 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2943 /* keep the parity disk(s) locked while asynchronous operations
2946 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2947 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2951 int qd_idx = sh->qd_idx;
2952 struct r5dev *dev = &sh->dev[qd_idx];
2954 set_bit(R5_LOCKED, &dev->flags);
2955 clear_bit(R5_UPTODATE, &dev->flags);
2959 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2960 __func__, (unsigned long long)sh->sector,
2961 s->locked, s->ops_request);
2965 * Each stripe/dev can have one or more bion attached.
2966 * toread/towrite point to the first in a chain.
2967 * The bi_next chain must be in order.
2969 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2970 int forwrite, int previous)
2973 struct r5conf *conf = sh->raid_conf;
2976 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2977 (unsigned long long)bi->bi_iter.bi_sector,
2978 (unsigned long long)sh->sector);
2981 * If several bio share a stripe. The bio bi_phys_segments acts as a
2982 * reference count to avoid race. The reference count should already be
2983 * increased before this function is called (for example, in
2984 * raid5_make_request()), so other bio sharing this stripe will not free the
2985 * stripe. If a stripe is owned by one stripe, the stripe lock will
2988 spin_lock_irq(&sh->stripe_lock);
2989 /* Don't allow new IO added to stripes in batch list */
2993 bip = &sh->dev[dd_idx].towrite;
2997 bip = &sh->dev[dd_idx].toread;
2998 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2999 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3001 bip = & (*bip)->bi_next;
3003 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3006 if (!forwrite || previous)
3007 clear_bit(STRIPE_BATCH_READY, &sh->state);
3009 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3013 raid5_inc_bi_active_stripes(bi);
3016 /* check if page is covered */
3017 sector_t sector = sh->dev[dd_idx].sector;
3018 for (bi=sh->dev[dd_idx].towrite;
3019 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3020 bi && bi->bi_iter.bi_sector <= sector;
3021 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3022 if (bio_end_sector(bi) >= sector)
3023 sector = bio_end_sector(bi);
3025 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3026 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3027 sh->overwrite_disks++;
3030 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3031 (unsigned long long)(*bip)->bi_iter.bi_sector,
3032 (unsigned long long)sh->sector, dd_idx);
3034 if (conf->mddev->bitmap && firstwrite) {
3035 /* Cannot hold spinlock over bitmap_startwrite,
3036 * but must ensure this isn't added to a batch until
3037 * we have added to the bitmap and set bm_seq.
3038 * So set STRIPE_BITMAP_PENDING to prevent
3040 * If multiple add_stripe_bio() calls race here they
3041 * much all set STRIPE_BITMAP_PENDING. So only the first one
3042 * to complete "bitmap_startwrite" gets to set
3043 * STRIPE_BIT_DELAY. This is important as once a stripe
3044 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3047 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3048 spin_unlock_irq(&sh->stripe_lock);
3049 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3051 spin_lock_irq(&sh->stripe_lock);
3052 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3053 if (!sh->batch_head) {
3054 sh->bm_seq = conf->seq_flush+1;
3055 set_bit(STRIPE_BIT_DELAY, &sh->state);
3058 spin_unlock_irq(&sh->stripe_lock);
3060 if (stripe_can_batch(sh))
3061 stripe_add_to_batch_list(conf, sh);
3065 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3066 spin_unlock_irq(&sh->stripe_lock);
3070 static void end_reshape(struct r5conf *conf);
3072 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3073 struct stripe_head *sh)
3075 int sectors_per_chunk =
3076 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3078 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3079 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3081 raid5_compute_sector(conf,
3082 stripe * (disks - conf->max_degraded)
3083 *sectors_per_chunk + chunk_offset,
3089 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3090 struct stripe_head_state *s, int disks,
3091 struct bio_list *return_bi)
3094 BUG_ON(sh->batch_head);
3095 for (i = disks; i--; ) {
3099 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3100 struct md_rdev *rdev;
3102 rdev = rcu_dereference(conf->disks[i].rdev);
3103 if (rdev && test_bit(In_sync, &rdev->flags) &&
3104 !test_bit(Faulty, &rdev->flags))
3105 atomic_inc(&rdev->nr_pending);
3110 if (!rdev_set_badblocks(
3114 md_error(conf->mddev, rdev);
3115 rdev_dec_pending(rdev, conf->mddev);
3118 spin_lock_irq(&sh->stripe_lock);
3119 /* fail all writes first */
3120 bi = sh->dev[i].towrite;
3121 sh->dev[i].towrite = NULL;
3122 sh->overwrite_disks = 0;
3123 spin_unlock_irq(&sh->stripe_lock);
3127 r5l_stripe_write_finished(sh);
3129 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3130 wake_up(&conf->wait_for_overlap);
3132 while (bi && bi->bi_iter.bi_sector <
3133 sh->dev[i].sector + STRIPE_SECTORS) {
3134 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3136 bi->bi_error = -EIO;
3137 if (!raid5_dec_bi_active_stripes(bi)) {
3138 md_write_end(conf->mddev);
3139 bio_list_add(return_bi, bi);
3144 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3145 STRIPE_SECTORS, 0, 0);
3147 /* and fail all 'written' */
3148 bi = sh->dev[i].written;
3149 sh->dev[i].written = NULL;
3150 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3151 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3152 sh->dev[i].page = sh->dev[i].orig_page;
3155 if (bi) bitmap_end = 1;
3156 while (bi && bi->bi_iter.bi_sector <
3157 sh->dev[i].sector + STRIPE_SECTORS) {
3158 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3160 bi->bi_error = -EIO;
3161 if (!raid5_dec_bi_active_stripes(bi)) {
3162 md_write_end(conf->mddev);
3163 bio_list_add(return_bi, bi);
3168 /* fail any reads if this device is non-operational and
3169 * the data has not reached the cache yet.
3171 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3172 s->failed > conf->max_degraded &&
3173 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3174 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3175 spin_lock_irq(&sh->stripe_lock);
3176 bi = sh->dev[i].toread;
3177 sh->dev[i].toread = NULL;
3178 spin_unlock_irq(&sh->stripe_lock);
3179 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3180 wake_up(&conf->wait_for_overlap);
3183 while (bi && bi->bi_iter.bi_sector <
3184 sh->dev[i].sector + STRIPE_SECTORS) {
3185 struct bio *nextbi =
3186 r5_next_bio(bi, sh->dev[i].sector);
3188 bi->bi_error = -EIO;
3189 if (!raid5_dec_bi_active_stripes(bi))
3190 bio_list_add(return_bi, bi);
3195 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3196 STRIPE_SECTORS, 0, 0);
3197 /* If we were in the middle of a write the parity block might
3198 * still be locked - so just clear all R5_LOCKED flags
3200 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3205 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3206 if (atomic_dec_and_test(&conf->pending_full_writes))
3207 md_wakeup_thread(conf->mddev->thread);
3211 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3212 struct stripe_head_state *s)
3217 BUG_ON(sh->batch_head);
3218 clear_bit(STRIPE_SYNCING, &sh->state);
3219 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3220 wake_up(&conf->wait_for_overlap);
3223 /* There is nothing more to do for sync/check/repair.
3224 * Don't even need to abort as that is handled elsewhere
3225 * if needed, and not always wanted e.g. if there is a known
3227 * For recover/replace we need to record a bad block on all
3228 * non-sync devices, or abort the recovery
3230 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3231 /* During recovery devices cannot be removed, so
3232 * locking and refcounting of rdevs is not needed
3235 for (i = 0; i < conf->raid_disks; i++) {
3236 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3238 && !test_bit(Faulty, &rdev->flags)
3239 && !test_bit(In_sync, &rdev->flags)
3240 && !rdev_set_badblocks(rdev, sh->sector,
3243 rdev = rcu_dereference(conf->disks[i].replacement);
3245 && !test_bit(Faulty, &rdev->flags)
3246 && !test_bit(In_sync, &rdev->flags)
3247 && !rdev_set_badblocks(rdev, sh->sector,
3253 conf->recovery_disabled =
3254 conf->mddev->recovery_disabled;
3256 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3259 static int want_replace(struct stripe_head *sh, int disk_idx)
3261 struct md_rdev *rdev;
3265 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3267 && !test_bit(Faulty, &rdev->flags)
3268 && !test_bit(In_sync, &rdev->flags)
3269 && (rdev->recovery_offset <= sh->sector
3270 || rdev->mddev->recovery_cp <= sh->sector))
3276 /* fetch_block - checks the given member device to see if its data needs
3277 * to be read or computed to satisfy a request.
3279 * Returns 1 when no more member devices need to be checked, otherwise returns
3280 * 0 to tell the loop in handle_stripe_fill to continue
3283 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3284 int disk_idx, int disks)
3286 struct r5dev *dev = &sh->dev[disk_idx];
3287 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3288 &sh->dev[s->failed_num[1]] };
3292 if (test_bit(R5_LOCKED, &dev->flags) ||
3293 test_bit(R5_UPTODATE, &dev->flags))
3294 /* No point reading this as we already have it or have
3295 * decided to get it.
3300 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3301 /* We need this block to directly satisfy a request */
3304 if (s->syncing || s->expanding ||
3305 (s->replacing && want_replace(sh, disk_idx)))
3306 /* When syncing, or expanding we read everything.
3307 * When replacing, we need the replaced block.
3311 if ((s->failed >= 1 && fdev[0]->toread) ||
3312 (s->failed >= 2 && fdev[1]->toread))
3313 /* If we want to read from a failed device, then
3314 * we need to actually read every other device.
3318 /* Sometimes neither read-modify-write nor reconstruct-write
3319 * cycles can work. In those cases we read every block we
3320 * can. Then the parity-update is certain to have enough to
3322 * This can only be a problem when we need to write something,
3323 * and some device has failed. If either of those tests
3324 * fail we need look no further.
3326 if (!s->failed || !s->to_write)
3329 if (test_bit(R5_Insync, &dev->flags) &&
3330 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3331 /* Pre-reads at not permitted until after short delay
3332 * to gather multiple requests. However if this
3333 * device is no Insync, the block could only be be computed
3334 * and there is no need to delay that.
3338 for (i = 0; i < s->failed && i < 2; i++) {
3339 if (fdev[i]->towrite &&
3340 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3341 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3342 /* If we have a partial write to a failed
3343 * device, then we will need to reconstruct
3344 * the content of that device, so all other
3345 * devices must be read.
3350 /* If we are forced to do a reconstruct-write, either because
3351 * the current RAID6 implementation only supports that, or
3352 * or because parity cannot be trusted and we are currently
3353 * recovering it, there is extra need to be careful.
3354 * If one of the devices that we would need to read, because
3355 * it is not being overwritten (and maybe not written at all)
3356 * is missing/faulty, then we need to read everything we can.
3358 if (sh->raid_conf->level != 6 &&
3359 sh->sector < sh->raid_conf->mddev->recovery_cp)
3360 /* reconstruct-write isn't being forced */
3362 for (i = 0; i < s->failed && i < 2; i++) {
3363 if (s->failed_num[i] != sh->pd_idx &&
3364 s->failed_num[i] != sh->qd_idx &&
3365 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3366 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3373 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3374 int disk_idx, int disks)
3376 struct r5dev *dev = &sh->dev[disk_idx];
3378 /* is the data in this block needed, and can we get it? */
3379 if (need_this_block(sh, s, disk_idx, disks)) {
3380 /* we would like to get this block, possibly by computing it,
3381 * otherwise read it if the backing disk is insync
3383 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3384 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3385 BUG_ON(sh->batch_head);
3386 if ((s->uptodate == disks - 1) &&
3387 (s->failed && (disk_idx == s->failed_num[0] ||
3388 disk_idx == s->failed_num[1]))) {
3389 /* have disk failed, and we're requested to fetch it;
3392 pr_debug("Computing stripe %llu block %d\n",
3393 (unsigned long long)sh->sector, disk_idx);
3394 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3395 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3396 set_bit(R5_Wantcompute, &dev->flags);
3397 sh->ops.target = disk_idx;
3398 sh->ops.target2 = -1; /* no 2nd target */
3400 /* Careful: from this point on 'uptodate' is in the eye
3401 * of raid_run_ops which services 'compute' operations
3402 * before writes. R5_Wantcompute flags a block that will
3403 * be R5_UPTODATE by the time it is needed for a
3404 * subsequent operation.
3408 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3409 /* Computing 2-failure is *very* expensive; only
3410 * do it if failed >= 2
3413 for (other = disks; other--; ) {
3414 if (other == disk_idx)
3416 if (!test_bit(R5_UPTODATE,
3417 &sh->dev[other].flags))
3421 pr_debug("Computing stripe %llu blocks %d,%d\n",
3422 (unsigned long long)sh->sector,
3424 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3425 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3426 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3427 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3428 sh->ops.target = disk_idx;
3429 sh->ops.target2 = other;
3433 } else if (test_bit(R5_Insync, &dev->flags)) {
3434 set_bit(R5_LOCKED, &dev->flags);
3435 set_bit(R5_Wantread, &dev->flags);
3437 pr_debug("Reading block %d (sync=%d)\n",
3438 disk_idx, s->syncing);
3446 * handle_stripe_fill - read or compute data to satisfy pending requests.
3448 static void handle_stripe_fill(struct stripe_head *sh,
3449 struct stripe_head_state *s,
3454 /* look for blocks to read/compute, skip this if a compute
3455 * is already in flight, or if the stripe contents are in the
3456 * midst of changing due to a write
3458 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3459 !sh->reconstruct_state)
3460 for (i = disks; i--; )
3461 if (fetch_block(sh, s, i, disks))
3463 set_bit(STRIPE_HANDLE, &sh->state);
3466 static void break_stripe_batch_list(struct stripe_head *head_sh,
3467 unsigned long handle_flags);
3468 /* handle_stripe_clean_event
3469 * any written block on an uptodate or failed drive can be returned.
3470 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3471 * never LOCKED, so we don't need to test 'failed' directly.
3473 static void handle_stripe_clean_event(struct r5conf *conf,
3474 struct stripe_head *sh, int disks, struct bio_list *return_bi)
3478 int discard_pending = 0;
3479 struct stripe_head *head_sh = sh;
3480 bool do_endio = false;
3482 for (i = disks; i--; )
3483 if (sh->dev[i].written) {
3485 if (!test_bit(R5_LOCKED, &dev->flags) &&
3486 (test_bit(R5_UPTODATE, &dev->flags) ||
3487 test_bit(R5_Discard, &dev->flags) ||
3488 test_bit(R5_SkipCopy, &dev->flags))) {
3489 /* We can return any write requests */
3490 struct bio *wbi, *wbi2;
3491 pr_debug("Return write for disc %d\n", i);
3492 if (test_and_clear_bit(R5_Discard, &dev->flags))
3493 clear_bit(R5_UPTODATE, &dev->flags);
3494 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3495 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3500 dev->page = dev->orig_page;
3502 dev->written = NULL;
3503 while (wbi && wbi->bi_iter.bi_sector <
3504 dev->sector + STRIPE_SECTORS) {
3505 wbi2 = r5_next_bio(wbi, dev->sector);
3506 if (!raid5_dec_bi_active_stripes(wbi)) {
3507 md_write_end(conf->mddev);
3508 bio_list_add(return_bi, wbi);
3512 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3514 !test_bit(STRIPE_DEGRADED, &sh->state),
3516 if (head_sh->batch_head) {
3517 sh = list_first_entry(&sh->batch_list,
3520 if (sh != head_sh) {
3527 } else if (test_bit(R5_Discard, &dev->flags))
3528 discard_pending = 1;
3531 r5l_stripe_write_finished(sh);
3533 if (!discard_pending &&
3534 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3536 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3537 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3538 if (sh->qd_idx >= 0) {
3539 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3540 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3542 /* now that discard is done we can proceed with any sync */
3543 clear_bit(STRIPE_DISCARD, &sh->state);
3545 * SCSI discard will change some bio fields and the stripe has
3546 * no updated data, so remove it from hash list and the stripe
3547 * will be reinitialized
3550 hash = sh->hash_lock_index;
3551 spin_lock_irq(conf->hash_locks + hash);
3553 spin_unlock_irq(conf->hash_locks + hash);
3554 if (head_sh->batch_head) {
3555 sh = list_first_entry(&sh->batch_list,
3556 struct stripe_head, batch_list);
3562 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3563 set_bit(STRIPE_HANDLE, &sh->state);
3567 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3568 if (atomic_dec_and_test(&conf->pending_full_writes))
3569 md_wakeup_thread(conf->mddev->thread);
3571 if (head_sh->batch_head && do_endio)
3572 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3575 static void handle_stripe_dirtying(struct r5conf *conf,
3576 struct stripe_head *sh,
3577 struct stripe_head_state *s,
3580 int rmw = 0, rcw = 0, i;
3581 sector_t recovery_cp = conf->mddev->recovery_cp;
3583 /* Check whether resync is now happening or should start.
3584 * If yes, then the array is dirty (after unclean shutdown or
3585 * initial creation), so parity in some stripes might be inconsistent.
3586 * In this case, we need to always do reconstruct-write, to ensure
3587 * that in case of drive failure or read-error correction, we
3588 * generate correct data from the parity.
3590 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3591 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3593 /* Calculate the real rcw later - for now make it
3594 * look like rcw is cheaper
3597 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3598 conf->rmw_level, (unsigned long long)recovery_cp,
3599 (unsigned long long)sh->sector);
3600 } else for (i = disks; i--; ) {
3601 /* would I have to read this buffer for read_modify_write */
3602 struct r5dev *dev = &sh->dev[i];
3603 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3604 !test_bit(R5_LOCKED, &dev->flags) &&
3605 !(test_bit(R5_UPTODATE, &dev->flags) ||
3606 test_bit(R5_Wantcompute, &dev->flags))) {
3607 if (test_bit(R5_Insync, &dev->flags))
3610 rmw += 2*disks; /* cannot read it */
3612 /* Would I have to read this buffer for reconstruct_write */
3613 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3614 i != sh->pd_idx && i != sh->qd_idx &&
3615 !test_bit(R5_LOCKED, &dev->flags) &&
3616 !(test_bit(R5_UPTODATE, &dev->flags) ||
3617 test_bit(R5_Wantcompute, &dev->flags))) {
3618 if (test_bit(R5_Insync, &dev->flags))
3624 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3625 (unsigned long long)sh->sector, rmw, rcw);
3626 set_bit(STRIPE_HANDLE, &sh->state);
3627 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3628 /* prefer read-modify-write, but need to get some data */
3629 if (conf->mddev->queue)
3630 blk_add_trace_msg(conf->mddev->queue,
3631 "raid5 rmw %llu %d",
3632 (unsigned long long)sh->sector, rmw);
3633 for (i = disks; i--; ) {
3634 struct r5dev *dev = &sh->dev[i];
3635 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3636 !test_bit(R5_LOCKED, &dev->flags) &&
3637 !(test_bit(R5_UPTODATE, &dev->flags) ||
3638 test_bit(R5_Wantcompute, &dev->flags)) &&
3639 test_bit(R5_Insync, &dev->flags)) {
3640 if (test_bit(STRIPE_PREREAD_ACTIVE,
3642 pr_debug("Read_old block %d for r-m-w\n",
3644 set_bit(R5_LOCKED, &dev->flags);
3645 set_bit(R5_Wantread, &dev->flags);
3648 set_bit(STRIPE_DELAYED, &sh->state);
3649 set_bit(STRIPE_HANDLE, &sh->state);
3654 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3655 /* want reconstruct write, but need to get some data */
3658 for (i = disks; i--; ) {
3659 struct r5dev *dev = &sh->dev[i];
3660 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3661 i != sh->pd_idx && i != sh->qd_idx &&
3662 !test_bit(R5_LOCKED, &dev->flags) &&
3663 !(test_bit(R5_UPTODATE, &dev->flags) ||
3664 test_bit(R5_Wantcompute, &dev->flags))) {
3666 if (test_bit(R5_Insync, &dev->flags) &&
3667 test_bit(STRIPE_PREREAD_ACTIVE,
3669 pr_debug("Read_old block "
3670 "%d for Reconstruct\n", i);
3671 set_bit(R5_LOCKED, &dev->flags);
3672 set_bit(R5_Wantread, &dev->flags);
3676 set_bit(STRIPE_DELAYED, &sh->state);
3677 set_bit(STRIPE_HANDLE, &sh->state);
3681 if (rcw && conf->mddev->queue)
3682 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3683 (unsigned long long)sh->sector,
3684 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3687 if (rcw > disks && rmw > disks &&
3688 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3689 set_bit(STRIPE_DELAYED, &sh->state);
3691 /* now if nothing is locked, and if we have enough data,
3692 * we can start a write request
3694 /* since handle_stripe can be called at any time we need to handle the
3695 * case where a compute block operation has been submitted and then a
3696 * subsequent call wants to start a write request. raid_run_ops only
3697 * handles the case where compute block and reconstruct are requested
3698 * simultaneously. If this is not the case then new writes need to be
3699 * held off until the compute completes.
3701 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3702 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3703 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3704 schedule_reconstruction(sh, s, rcw == 0, 0);
3707 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3708 struct stripe_head_state *s, int disks)
3710 struct r5dev *dev = NULL;
3712 BUG_ON(sh->batch_head);
3713 set_bit(STRIPE_HANDLE, &sh->state);
3715 switch (sh->check_state) {
3716 case check_state_idle:
3717 /* start a new check operation if there are no failures */
3718 if (s->failed == 0) {
3719 BUG_ON(s->uptodate != disks);
3720 sh->check_state = check_state_run;
3721 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3722 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3726 dev = &sh->dev[s->failed_num[0]];
3728 case check_state_compute_result:
3729 sh->check_state = check_state_idle;
3731 dev = &sh->dev[sh->pd_idx];
3733 /* check that a write has not made the stripe insync */
3734 if (test_bit(STRIPE_INSYNC, &sh->state))
3737 /* either failed parity check, or recovery is happening */
3738 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3739 BUG_ON(s->uptodate != disks);
3741 set_bit(R5_LOCKED, &dev->flags);
3743 set_bit(R5_Wantwrite, &dev->flags);
3745 clear_bit(STRIPE_DEGRADED, &sh->state);
3746 set_bit(STRIPE_INSYNC, &sh->state);
3748 case check_state_run:
3749 break; /* we will be called again upon completion */
3750 case check_state_check_result:
3751 sh->check_state = check_state_idle;
3753 /* if a failure occurred during the check operation, leave
3754 * STRIPE_INSYNC not set and let the stripe be handled again
3759 /* handle a successful check operation, if parity is correct
3760 * we are done. Otherwise update the mismatch count and repair
3761 * parity if !MD_RECOVERY_CHECK
3763 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3764 /* parity is correct (on disc,
3765 * not in buffer any more)
3767 set_bit(STRIPE_INSYNC, &sh->state);
3769 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3770 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3771 /* don't try to repair!! */
3772 set_bit(STRIPE_INSYNC, &sh->state);
3774 sh->check_state = check_state_compute_run;
3775 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3776 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3777 set_bit(R5_Wantcompute,
3778 &sh->dev[sh->pd_idx].flags);
3779 sh->ops.target = sh->pd_idx;
3780 sh->ops.target2 = -1;
3785 case check_state_compute_run:
3788 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3789 __func__, sh->check_state,
3790 (unsigned long long) sh->sector);
3795 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3796 struct stripe_head_state *s,
3799 int pd_idx = sh->pd_idx;
3800 int qd_idx = sh->qd_idx;
3803 BUG_ON(sh->batch_head);
3804 set_bit(STRIPE_HANDLE, &sh->state);
3806 BUG_ON(s->failed > 2);
3808 /* Want to check and possibly repair P and Q.
3809 * However there could be one 'failed' device, in which
3810 * case we can only check one of them, possibly using the
3811 * other to generate missing data
3814 switch (sh->check_state) {
3815 case check_state_idle:
3816 /* start a new check operation if there are < 2 failures */
3817 if (s->failed == s->q_failed) {
3818 /* The only possible failed device holds Q, so it
3819 * makes sense to check P (If anything else were failed,
3820 * we would have used P to recreate it).
3822 sh->check_state = check_state_run;
3824 if (!s->q_failed && s->failed < 2) {
3825 /* Q is not failed, and we didn't use it to generate
3826 * anything, so it makes sense to check it
3828 if (sh->check_state == check_state_run)
3829 sh->check_state = check_state_run_pq;
3831 sh->check_state = check_state_run_q;
3834 /* discard potentially stale zero_sum_result */
3835 sh->ops.zero_sum_result = 0;
3837 if (sh->check_state == check_state_run) {
3838 /* async_xor_zero_sum destroys the contents of P */
3839 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3842 if (sh->check_state >= check_state_run &&
3843 sh->check_state <= check_state_run_pq) {
3844 /* async_syndrome_zero_sum preserves P and Q, so
3845 * no need to mark them !uptodate here
3847 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3851 /* we have 2-disk failure */
3852 BUG_ON(s->failed != 2);
3854 case check_state_compute_result:
3855 sh->check_state = check_state_idle;
3857 /* check that a write has not made the stripe insync */
3858 if (test_bit(STRIPE_INSYNC, &sh->state))
3861 /* now write out any block on a failed drive,
3862 * or P or Q if they were recomputed
3864 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3865 if (s->failed == 2) {
3866 dev = &sh->dev[s->failed_num[1]];
3868 set_bit(R5_LOCKED, &dev->flags);
3869 set_bit(R5_Wantwrite, &dev->flags);
3871 if (s->failed >= 1) {
3872 dev = &sh->dev[s->failed_num[0]];
3874 set_bit(R5_LOCKED, &dev->flags);
3875 set_bit(R5_Wantwrite, &dev->flags);
3877 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3878 dev = &sh->dev[pd_idx];
3880 set_bit(R5_LOCKED, &dev->flags);
3881 set_bit(R5_Wantwrite, &dev->flags);
3883 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3884 dev = &sh->dev[qd_idx];
3886 set_bit(R5_LOCKED, &dev->flags);
3887 set_bit(R5_Wantwrite, &dev->flags);
3889 clear_bit(STRIPE_DEGRADED, &sh->state);
3891 set_bit(STRIPE_INSYNC, &sh->state);
3893 case check_state_run:
3894 case check_state_run_q:
3895 case check_state_run_pq:
3896 break; /* we will be called again upon completion */
3897 case check_state_check_result:
3898 sh->check_state = check_state_idle;
3900 /* handle a successful check operation, if parity is correct
3901 * we are done. Otherwise update the mismatch count and repair
3902 * parity if !MD_RECOVERY_CHECK
3904 if (sh->ops.zero_sum_result == 0) {
3905 /* both parities are correct */
3907 set_bit(STRIPE_INSYNC, &sh->state);
3909 /* in contrast to the raid5 case we can validate
3910 * parity, but still have a failure to write
3913 sh->check_state = check_state_compute_result;
3914 /* Returning at this point means that we may go
3915 * off and bring p and/or q uptodate again so
3916 * we make sure to check zero_sum_result again
3917 * to verify if p or q need writeback
3921 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3922 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3923 /* don't try to repair!! */
3924 set_bit(STRIPE_INSYNC, &sh->state);
3926 int *target = &sh->ops.target;
3928 sh->ops.target = -1;
3929 sh->ops.target2 = -1;
3930 sh->check_state = check_state_compute_run;
3931 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3932 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3933 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3934 set_bit(R5_Wantcompute,
3935 &sh->dev[pd_idx].flags);
3937 target = &sh->ops.target2;
3940 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3941 set_bit(R5_Wantcompute,
3942 &sh->dev[qd_idx].flags);
3949 case check_state_compute_run:
3952 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3953 __func__, sh->check_state,
3954 (unsigned long long) sh->sector);
3959 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3963 /* We have read all the blocks in this stripe and now we need to
3964 * copy some of them into a target stripe for expand.
3966 struct dma_async_tx_descriptor *tx = NULL;
3967 BUG_ON(sh->batch_head);
3968 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3969 for (i = 0; i < sh->disks; i++)
3970 if (i != sh->pd_idx && i != sh->qd_idx) {
3972 struct stripe_head *sh2;
3973 struct async_submit_ctl submit;
3975 sector_t bn = raid5_compute_blocknr(sh, i, 1);
3976 sector_t s = raid5_compute_sector(conf, bn, 0,
3978 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
3980 /* so far only the early blocks of this stripe
3981 * have been requested. When later blocks
3982 * get requested, we will try again
3985 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3986 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3987 /* must have already done this block */
3988 raid5_release_stripe(sh2);
3992 /* place all the copies on one channel */
3993 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3994 tx = async_memcpy(sh2->dev[dd_idx].page,
3995 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3998 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3999 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4000 for (j = 0; j < conf->raid_disks; j++)
4001 if (j != sh2->pd_idx &&
4003 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4005 if (j == conf->raid_disks) {
4006 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4007 set_bit(STRIPE_HANDLE, &sh2->state);
4009 raid5_release_stripe(sh2);
4012 /* done submitting copies, wait for them to complete */
4013 async_tx_quiesce(&tx);
4017 * handle_stripe - do things to a stripe.
4019 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4020 * state of various bits to see what needs to be done.
4022 * return some read requests which now have data
4023 * return some write requests which are safely on storage
4024 * schedule a read on some buffers
4025 * schedule a write of some buffers
4026 * return confirmation of parity correctness
4030 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4032 struct r5conf *conf = sh->raid_conf;
4033 int disks = sh->disks;
4036 int do_recovery = 0;
4038 memset(s, 0, sizeof(*s));
4040 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4041 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4042 s->failed_num[0] = -1;
4043 s->failed_num[1] = -1;
4044 s->log_failed = r5l_log_disk_error(conf);
4046 /* Now to look around and see what can be done */
4048 for (i=disks; i--; ) {
4049 struct md_rdev *rdev;
4056 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4058 dev->toread, dev->towrite, dev->written);
4059 /* maybe we can reply to a read
4061 * new wantfill requests are only permitted while
4062 * ops_complete_biofill is guaranteed to be inactive
4064 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4065 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4066 set_bit(R5_Wantfill, &dev->flags);
4068 /* now count some things */
4069 if (test_bit(R5_LOCKED, &dev->flags))
4071 if (test_bit(R5_UPTODATE, &dev->flags))
4073 if (test_bit(R5_Wantcompute, &dev->flags)) {
4075 BUG_ON(s->compute > 2);
4078 if (test_bit(R5_Wantfill, &dev->flags))
4080 else if (dev->toread)
4084 if (!test_bit(R5_OVERWRITE, &dev->flags))
4089 /* Prefer to use the replacement for reads, but only
4090 * if it is recovered enough and has no bad blocks.
4092 rdev = rcu_dereference(conf->disks[i].replacement);
4093 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4094 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4095 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4096 &first_bad, &bad_sectors))
4097 set_bit(R5_ReadRepl, &dev->flags);
4099 if (rdev && !test_bit(Faulty, &rdev->flags))
4100 set_bit(R5_NeedReplace, &dev->flags);
4102 clear_bit(R5_NeedReplace, &dev->flags);
4103 rdev = rcu_dereference(conf->disks[i].rdev);
4104 clear_bit(R5_ReadRepl, &dev->flags);
4106 if (rdev && test_bit(Faulty, &rdev->flags))
4109 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4110 &first_bad, &bad_sectors);
4111 if (s->blocked_rdev == NULL
4112 && (test_bit(Blocked, &rdev->flags)
4115 set_bit(BlockedBadBlocks,
4117 s->blocked_rdev = rdev;
4118 atomic_inc(&rdev->nr_pending);
4121 clear_bit(R5_Insync, &dev->flags);
4125 /* also not in-sync */
4126 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4127 test_bit(R5_UPTODATE, &dev->flags)) {
4128 /* treat as in-sync, but with a read error
4129 * which we can now try to correct
4131 set_bit(R5_Insync, &dev->flags);
4132 set_bit(R5_ReadError, &dev->flags);
4134 } else if (test_bit(In_sync, &rdev->flags))
4135 set_bit(R5_Insync, &dev->flags);
4136 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4137 /* in sync if before recovery_offset */
4138 set_bit(R5_Insync, &dev->flags);
4139 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4140 test_bit(R5_Expanded, &dev->flags))
4141 /* If we've reshaped into here, we assume it is Insync.
4142 * We will shortly update recovery_offset to make
4145 set_bit(R5_Insync, &dev->flags);
4147 if (test_bit(R5_WriteError, &dev->flags)) {
4148 /* This flag does not apply to '.replacement'
4149 * only to .rdev, so make sure to check that*/
4150 struct md_rdev *rdev2 = rcu_dereference(
4151 conf->disks[i].rdev);
4153 clear_bit(R5_Insync, &dev->flags);
4154 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4155 s->handle_bad_blocks = 1;
4156 atomic_inc(&rdev2->nr_pending);
4158 clear_bit(R5_WriteError, &dev->flags);
4160 if (test_bit(R5_MadeGood, &dev->flags)) {
4161 /* This flag does not apply to '.replacement'
4162 * only to .rdev, so make sure to check that*/
4163 struct md_rdev *rdev2 = rcu_dereference(
4164 conf->disks[i].rdev);
4165 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4166 s->handle_bad_blocks = 1;
4167 atomic_inc(&rdev2->nr_pending);
4169 clear_bit(R5_MadeGood, &dev->flags);
4171 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4172 struct md_rdev *rdev2 = rcu_dereference(
4173 conf->disks[i].replacement);
4174 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4175 s->handle_bad_blocks = 1;
4176 atomic_inc(&rdev2->nr_pending);
4178 clear_bit(R5_MadeGoodRepl, &dev->flags);
4180 if (!test_bit(R5_Insync, &dev->flags)) {
4181 /* The ReadError flag will just be confusing now */
4182 clear_bit(R5_ReadError, &dev->flags);
4183 clear_bit(R5_ReWrite, &dev->flags);
4185 if (test_bit(R5_ReadError, &dev->flags))
4186 clear_bit(R5_Insync, &dev->flags);
4187 if (!test_bit(R5_Insync, &dev->flags)) {
4189 s->failed_num[s->failed] = i;
4191 if (rdev && !test_bit(Faulty, &rdev->flags))
4195 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4196 /* If there is a failed device being replaced,
4197 * we must be recovering.
4198 * else if we are after recovery_cp, we must be syncing
4199 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4200 * else we can only be replacing
4201 * sync and recovery both need to read all devices, and so
4202 * use the same flag.
4205 sh->sector >= conf->mddev->recovery_cp ||
4206 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4214 static int clear_batch_ready(struct stripe_head *sh)
4216 /* Return '1' if this is a member of batch, or
4217 * '0' if it is a lone stripe or a head which can now be
4220 struct stripe_head *tmp;
4221 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4222 return (sh->batch_head && sh->batch_head != sh);
4223 spin_lock(&sh->stripe_lock);
4224 if (!sh->batch_head) {
4225 spin_unlock(&sh->stripe_lock);
4230 * this stripe could be added to a batch list before we check
4231 * BATCH_READY, skips it
4233 if (sh->batch_head != sh) {
4234 spin_unlock(&sh->stripe_lock);
4237 spin_lock(&sh->batch_lock);
4238 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4239 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4240 spin_unlock(&sh->batch_lock);
4241 spin_unlock(&sh->stripe_lock);
4244 * BATCH_READY is cleared, no new stripes can be added.
4245 * batch_list can be accessed without lock
4250 static void break_stripe_batch_list(struct stripe_head *head_sh,
4251 unsigned long handle_flags)
4253 struct stripe_head *sh, *next;
4257 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4259 list_del_init(&sh->batch_list);
4261 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4262 (1 << STRIPE_SYNCING) |
4263 (1 << STRIPE_REPLACED) |
4264 (1 << STRIPE_DELAYED) |
4265 (1 << STRIPE_BIT_DELAY) |
4266 (1 << STRIPE_FULL_WRITE) |
4267 (1 << STRIPE_BIOFILL_RUN) |
4268 (1 << STRIPE_COMPUTE_RUN) |
4269 (1 << STRIPE_OPS_REQ_PENDING) |
4270 (1 << STRIPE_DISCARD) |
4271 (1 << STRIPE_BATCH_READY) |
4272 (1 << STRIPE_BATCH_ERR) |
4273 (1 << STRIPE_BITMAP_PENDING)),
4274 "stripe state: %lx\n", sh->state);
4275 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4276 (1 << STRIPE_REPLACED)),
4277 "head stripe state: %lx\n", head_sh->state);
4279 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4280 (1 << STRIPE_PREREAD_ACTIVE) |
4281 (1 << STRIPE_DEGRADED)),
4282 head_sh->state & (1 << STRIPE_INSYNC));
4284 sh->check_state = head_sh->check_state;
4285 sh->reconstruct_state = head_sh->reconstruct_state;
4286 for (i = 0; i < sh->disks; i++) {
4287 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4289 sh->dev[i].flags = head_sh->dev[i].flags &
4290 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4292 spin_lock_irq(&sh->stripe_lock);
4293 sh->batch_head = NULL;
4294 spin_unlock_irq(&sh->stripe_lock);
4295 if (handle_flags == 0 ||
4296 sh->state & handle_flags)
4297 set_bit(STRIPE_HANDLE, &sh->state);
4298 raid5_release_stripe(sh);
4300 spin_lock_irq(&head_sh->stripe_lock);
4301 head_sh->batch_head = NULL;
4302 spin_unlock_irq(&head_sh->stripe_lock);
4303 for (i = 0; i < head_sh->disks; i++)
4304 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4306 if (head_sh->state & handle_flags)
4307 set_bit(STRIPE_HANDLE, &head_sh->state);
4310 wake_up(&head_sh->raid_conf->wait_for_overlap);
4313 static void handle_stripe(struct stripe_head *sh)
4315 struct stripe_head_state s;
4316 struct r5conf *conf = sh->raid_conf;
4319 int disks = sh->disks;
4320 struct r5dev *pdev, *qdev;
4322 clear_bit(STRIPE_HANDLE, &sh->state);
4323 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4324 /* already being handled, ensure it gets handled
4325 * again when current action finishes */
4326 set_bit(STRIPE_HANDLE, &sh->state);
4330 if (clear_batch_ready(sh) ) {
4331 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4335 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4336 break_stripe_batch_list(sh, 0);
4338 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4339 spin_lock(&sh->stripe_lock);
4340 /* Cannot process 'sync' concurrently with 'discard' */
4341 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4342 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4343 set_bit(STRIPE_SYNCING, &sh->state);
4344 clear_bit(STRIPE_INSYNC, &sh->state);
4345 clear_bit(STRIPE_REPLACED, &sh->state);
4347 spin_unlock(&sh->stripe_lock);
4349 clear_bit(STRIPE_DELAYED, &sh->state);
4351 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4352 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4353 (unsigned long long)sh->sector, sh->state,
4354 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4355 sh->check_state, sh->reconstruct_state);
4357 analyse_stripe(sh, &s);
4359 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4362 if (s.handle_bad_blocks) {
4363 set_bit(STRIPE_HANDLE, &sh->state);
4367 if (unlikely(s.blocked_rdev)) {
4368 if (s.syncing || s.expanding || s.expanded ||
4369 s.replacing || s.to_write || s.written) {
4370 set_bit(STRIPE_HANDLE, &sh->state);
4373 /* There is nothing for the blocked_rdev to block */
4374 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4375 s.blocked_rdev = NULL;
4378 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4379 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4380 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4383 pr_debug("locked=%d uptodate=%d to_read=%d"
4384 " to_write=%d failed=%d failed_num=%d,%d\n",
4385 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4386 s.failed_num[0], s.failed_num[1]);
4387 /* check if the array has lost more than max_degraded devices and,
4388 * if so, some requests might need to be failed.
4390 if (s.failed > conf->max_degraded || s.log_failed) {
4391 sh->check_state = 0;
4392 sh->reconstruct_state = 0;
4393 break_stripe_batch_list(sh, 0);
4394 if (s.to_read+s.to_write+s.written)
4395 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4396 if (s.syncing + s.replacing)
4397 handle_failed_sync(conf, sh, &s);
4400 /* Now we check to see if any write operations have recently
4404 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4406 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4407 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4408 sh->reconstruct_state = reconstruct_state_idle;
4410 /* All the 'written' buffers and the parity block are ready to
4411 * be written back to disk
4413 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4414 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4415 BUG_ON(sh->qd_idx >= 0 &&
4416 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4417 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4418 for (i = disks; i--; ) {
4419 struct r5dev *dev = &sh->dev[i];
4420 if (test_bit(R5_LOCKED, &dev->flags) &&
4421 (i == sh->pd_idx || i == sh->qd_idx ||
4423 pr_debug("Writing block %d\n", i);
4424 set_bit(R5_Wantwrite, &dev->flags);
4429 if (!test_bit(R5_Insync, &dev->flags) ||
4430 ((i == sh->pd_idx || i == sh->qd_idx) &&
4432 set_bit(STRIPE_INSYNC, &sh->state);
4435 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4436 s.dec_preread_active = 1;
4440 * might be able to return some write requests if the parity blocks
4441 * are safe, or on a failed drive
4443 pdev = &sh->dev[sh->pd_idx];
4444 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4445 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4446 qdev = &sh->dev[sh->qd_idx];
4447 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4448 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4452 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4453 && !test_bit(R5_LOCKED, &pdev->flags)
4454 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4455 test_bit(R5_Discard, &pdev->flags))))) &&
4456 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4457 && !test_bit(R5_LOCKED, &qdev->flags)
4458 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4459 test_bit(R5_Discard, &qdev->flags))))))
4460 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4462 /* Now we might consider reading some blocks, either to check/generate
4463 * parity, or to satisfy requests
4464 * or to load a block that is being partially written.
4466 if (s.to_read || s.non_overwrite
4467 || (conf->level == 6 && s.to_write && s.failed)
4468 || (s.syncing && (s.uptodate + s.compute < disks))
4471 handle_stripe_fill(sh, &s, disks);
4473 /* Now to consider new write requests and what else, if anything
4474 * should be read. We do not handle new writes when:
4475 * 1/ A 'write' operation (copy+xor) is already in flight.
4476 * 2/ A 'check' operation is in flight, as it may clobber the parity
4479 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4480 handle_stripe_dirtying(conf, sh, &s, disks);
4482 /* maybe we need to check and possibly fix the parity for this stripe
4483 * Any reads will already have been scheduled, so we just see if enough
4484 * data is available. The parity check is held off while parity
4485 * dependent operations are in flight.
4487 if (sh->check_state ||
4488 (s.syncing && s.locked == 0 &&
4489 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4490 !test_bit(STRIPE_INSYNC, &sh->state))) {
4491 if (conf->level == 6)
4492 handle_parity_checks6(conf, sh, &s, disks);
4494 handle_parity_checks5(conf, sh, &s, disks);
4497 if ((s.replacing || s.syncing) && s.locked == 0
4498 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4499 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4500 /* Write out to replacement devices where possible */
4501 for (i = 0; i < conf->raid_disks; i++)
4502 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4503 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4504 set_bit(R5_WantReplace, &sh->dev[i].flags);
4505 set_bit(R5_LOCKED, &sh->dev[i].flags);
4509 set_bit(STRIPE_INSYNC, &sh->state);
4510 set_bit(STRIPE_REPLACED, &sh->state);
4512 if ((s.syncing || s.replacing) && s.locked == 0 &&
4513 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4514 test_bit(STRIPE_INSYNC, &sh->state)) {
4515 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4516 clear_bit(STRIPE_SYNCING, &sh->state);
4517 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4518 wake_up(&conf->wait_for_overlap);
4521 /* If the failed drives are just a ReadError, then we might need
4522 * to progress the repair/check process
4524 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4525 for (i = 0; i < s.failed; i++) {
4526 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4527 if (test_bit(R5_ReadError, &dev->flags)
4528 && !test_bit(R5_LOCKED, &dev->flags)
4529 && test_bit(R5_UPTODATE, &dev->flags)
4531 if (!test_bit(R5_ReWrite, &dev->flags)) {
4532 set_bit(R5_Wantwrite, &dev->flags);
4533 set_bit(R5_ReWrite, &dev->flags);
4534 set_bit(R5_LOCKED, &dev->flags);
4537 /* let's read it back */
4538 set_bit(R5_Wantread, &dev->flags);
4539 set_bit(R5_LOCKED, &dev->flags);
4545 /* Finish reconstruct operations initiated by the expansion process */
4546 if (sh->reconstruct_state == reconstruct_state_result) {
4547 struct stripe_head *sh_src
4548 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4549 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4550 /* sh cannot be written until sh_src has been read.
4551 * so arrange for sh to be delayed a little
4553 set_bit(STRIPE_DELAYED, &sh->state);
4554 set_bit(STRIPE_HANDLE, &sh->state);
4555 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4557 atomic_inc(&conf->preread_active_stripes);
4558 raid5_release_stripe(sh_src);
4562 raid5_release_stripe(sh_src);
4564 sh->reconstruct_state = reconstruct_state_idle;
4565 clear_bit(STRIPE_EXPANDING, &sh->state);
4566 for (i = conf->raid_disks; i--; ) {
4567 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4568 set_bit(R5_LOCKED, &sh->dev[i].flags);
4573 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4574 !sh->reconstruct_state) {
4575 /* Need to write out all blocks after computing parity */
4576 sh->disks = conf->raid_disks;
4577 stripe_set_idx(sh->sector, conf, 0, sh);
4578 schedule_reconstruction(sh, &s, 1, 1);
4579 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4580 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4581 atomic_dec(&conf->reshape_stripes);
4582 wake_up(&conf->wait_for_overlap);
4583 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4586 if (s.expanding && s.locked == 0 &&
4587 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4588 handle_stripe_expansion(conf, sh);
4591 /* wait for this device to become unblocked */
4592 if (unlikely(s.blocked_rdev)) {
4593 if (conf->mddev->external)
4594 md_wait_for_blocked_rdev(s.blocked_rdev,
4597 /* Internal metadata will immediately
4598 * be written by raid5d, so we don't
4599 * need to wait here.
4601 rdev_dec_pending(s.blocked_rdev,
4605 if (s.handle_bad_blocks)
4606 for (i = disks; i--; ) {
4607 struct md_rdev *rdev;
4608 struct r5dev *dev = &sh->dev[i];
4609 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4610 /* We own a safe reference to the rdev */
4611 rdev = conf->disks[i].rdev;
4612 if (!rdev_set_badblocks(rdev, sh->sector,
4614 md_error(conf->mddev, rdev);
4615 rdev_dec_pending(rdev, conf->mddev);
4617 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4618 rdev = conf->disks[i].rdev;
4619 rdev_clear_badblocks(rdev, sh->sector,
4621 rdev_dec_pending(rdev, conf->mddev);
4623 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4624 rdev = conf->disks[i].replacement;
4626 /* rdev have been moved down */
4627 rdev = conf->disks[i].rdev;
4628 rdev_clear_badblocks(rdev, sh->sector,
4630 rdev_dec_pending(rdev, conf->mddev);
4635 raid_run_ops(sh, s.ops_request);
4639 if (s.dec_preread_active) {
4640 /* We delay this until after ops_run_io so that if make_request
4641 * is waiting on a flush, it won't continue until the writes
4642 * have actually been submitted.
4644 atomic_dec(&conf->preread_active_stripes);
4645 if (atomic_read(&conf->preread_active_stripes) <
4647 md_wakeup_thread(conf->mddev->thread);
4650 if (!bio_list_empty(&s.return_bi)) {
4651 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags) &&
4652 (s.failed <= conf->max_degraded ||
4653 conf->mddev->external == 0)) {
4654 spin_lock_irq(&conf->device_lock);
4655 bio_list_merge(&conf->return_bi, &s.return_bi);
4656 spin_unlock_irq(&conf->device_lock);
4657 md_wakeup_thread(conf->mddev->thread);
4659 return_io(&s.return_bi);
4662 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4665 static void raid5_activate_delayed(struct r5conf *conf)
4667 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4668 while (!list_empty(&conf->delayed_list)) {
4669 struct list_head *l = conf->delayed_list.next;
4670 struct stripe_head *sh;
4671 sh = list_entry(l, struct stripe_head, lru);
4673 clear_bit(STRIPE_DELAYED, &sh->state);
4674 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4675 atomic_inc(&conf->preread_active_stripes);
4676 list_add_tail(&sh->lru, &conf->hold_list);
4677 raid5_wakeup_stripe_thread(sh);
4682 static void activate_bit_delay(struct r5conf *conf,
4683 struct list_head *temp_inactive_list)
4685 /* device_lock is held */
4686 struct list_head head;
4687 list_add(&head, &conf->bitmap_list);
4688 list_del_init(&conf->bitmap_list);
4689 while (!list_empty(&head)) {
4690 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4692 list_del_init(&sh->lru);
4693 atomic_inc(&sh->count);
4694 hash = sh->hash_lock_index;
4695 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4699 static int raid5_congested(struct mddev *mddev, int bits)
4701 struct r5conf *conf = mddev->private;
4703 /* No difference between reads and writes. Just check
4704 * how busy the stripe_cache is
4707 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4711 if (atomic_read(&conf->empty_inactive_list_nr))
4717 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4719 struct r5conf *conf = mddev->private;
4720 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4721 unsigned int chunk_sectors;
4722 unsigned int bio_sectors = bio_sectors(bio);
4724 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4725 return chunk_sectors >=
4726 ((sector & (chunk_sectors - 1)) + bio_sectors);
4730 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4731 * later sampled by raid5d.
4733 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4735 unsigned long flags;
4737 spin_lock_irqsave(&conf->device_lock, flags);
4739 bi->bi_next = conf->retry_read_aligned_list;
4740 conf->retry_read_aligned_list = bi;
4742 spin_unlock_irqrestore(&conf->device_lock, flags);
4743 md_wakeup_thread(conf->mddev->thread);
4746 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4750 bi = conf->retry_read_aligned;
4752 conf->retry_read_aligned = NULL;
4755 bi = conf->retry_read_aligned_list;
4757 conf->retry_read_aligned_list = bi->bi_next;
4760 * this sets the active strip count to 1 and the processed
4761 * strip count to zero (upper 8 bits)
4763 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4770 * The "raid5_align_endio" should check if the read succeeded and if it
4771 * did, call bio_endio on the original bio (having bio_put the new bio
4773 * If the read failed..
4775 static void raid5_align_endio(struct bio *bi)
4777 struct bio* raid_bi = bi->bi_private;
4778 struct mddev *mddev;
4779 struct r5conf *conf;
4780 struct md_rdev *rdev;
4781 int error = bi->bi_error;
4785 rdev = (void*)raid_bi->bi_next;
4786 raid_bi->bi_next = NULL;
4787 mddev = rdev->mddev;
4788 conf = mddev->private;
4790 rdev_dec_pending(rdev, conf->mddev);
4793 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4796 if (atomic_dec_and_test(&conf->active_aligned_reads))
4797 wake_up(&conf->wait_for_quiescent);
4801 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4803 add_bio_to_retry(raid_bi, conf);
4806 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4808 struct r5conf *conf = mddev->private;
4810 struct bio* align_bi;
4811 struct md_rdev *rdev;
4812 sector_t end_sector;
4814 if (!in_chunk_boundary(mddev, raid_bio)) {
4815 pr_debug("%s: non aligned\n", __func__);
4819 * use bio_clone_mddev to make a copy of the bio
4821 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4825 * set bi_end_io to a new function, and set bi_private to the
4828 align_bi->bi_end_io = raid5_align_endio;
4829 align_bi->bi_private = raid_bio;
4833 align_bi->bi_iter.bi_sector =
4834 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4837 end_sector = bio_end_sector(align_bi);
4839 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4840 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4841 rdev->recovery_offset < end_sector) {
4842 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4844 (test_bit(Faulty, &rdev->flags) ||
4845 !(test_bit(In_sync, &rdev->flags) ||
4846 rdev->recovery_offset >= end_sector)))
4853 atomic_inc(&rdev->nr_pending);
4855 raid_bio->bi_next = (void*)rdev;
4856 align_bi->bi_bdev = rdev->bdev;
4857 bio_clear_flag(align_bi, BIO_SEG_VALID);
4859 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4860 bio_sectors(align_bi),
4861 &first_bad, &bad_sectors)) {
4863 rdev_dec_pending(rdev, mddev);
4867 /* No reshape active, so we can trust rdev->data_offset */
4868 align_bi->bi_iter.bi_sector += rdev->data_offset;
4870 spin_lock_irq(&conf->device_lock);
4871 wait_event_lock_irq(conf->wait_for_quiescent,
4874 atomic_inc(&conf->active_aligned_reads);
4875 spin_unlock_irq(&conf->device_lock);
4878 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4879 align_bi, disk_devt(mddev->gendisk),
4880 raid_bio->bi_iter.bi_sector);
4881 generic_make_request(align_bi);
4890 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4895 sector_t sector = raid_bio->bi_iter.bi_sector;
4896 unsigned chunk_sects = mddev->chunk_sectors;
4897 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4899 if (sectors < bio_sectors(raid_bio)) {
4900 split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4901 bio_chain(split, raid_bio);
4905 if (!raid5_read_one_chunk(mddev, split)) {
4906 if (split != raid_bio)
4907 generic_make_request(raid_bio);
4910 } while (split != raid_bio);
4915 /* __get_priority_stripe - get the next stripe to process
4917 * Full stripe writes are allowed to pass preread active stripes up until
4918 * the bypass_threshold is exceeded. In general the bypass_count
4919 * increments when the handle_list is handled before the hold_list; however, it
4920 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4921 * stripe with in flight i/o. The bypass_count will be reset when the
4922 * head of the hold_list has changed, i.e. the head was promoted to the
4925 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4927 struct stripe_head *sh = NULL, *tmp;
4928 struct list_head *handle_list = NULL;
4929 struct r5worker_group *wg = NULL;
4931 if (conf->worker_cnt_per_group == 0) {
4932 handle_list = &conf->handle_list;
4933 } else if (group != ANY_GROUP) {
4934 handle_list = &conf->worker_groups[group].handle_list;
4935 wg = &conf->worker_groups[group];
4938 for (i = 0; i < conf->group_cnt; i++) {
4939 handle_list = &conf->worker_groups[i].handle_list;
4940 wg = &conf->worker_groups[i];
4941 if (!list_empty(handle_list))
4946 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4948 list_empty(handle_list) ? "empty" : "busy",
4949 list_empty(&conf->hold_list) ? "empty" : "busy",
4950 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4952 if (!list_empty(handle_list)) {
4953 sh = list_entry(handle_list->next, typeof(*sh), lru);
4955 if (list_empty(&conf->hold_list))
4956 conf->bypass_count = 0;
4957 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4958 if (conf->hold_list.next == conf->last_hold)
4959 conf->bypass_count++;
4961 conf->last_hold = conf->hold_list.next;
4962 conf->bypass_count -= conf->bypass_threshold;
4963 if (conf->bypass_count < 0)
4964 conf->bypass_count = 0;
4967 } else if (!list_empty(&conf->hold_list) &&
4968 ((conf->bypass_threshold &&
4969 conf->bypass_count > conf->bypass_threshold) ||
4970 atomic_read(&conf->pending_full_writes) == 0)) {
4972 list_for_each_entry(tmp, &conf->hold_list, lru) {
4973 if (conf->worker_cnt_per_group == 0 ||
4974 group == ANY_GROUP ||
4975 !cpu_online(tmp->cpu) ||
4976 cpu_to_group(tmp->cpu) == group) {
4983 conf->bypass_count -= conf->bypass_threshold;
4984 if (conf->bypass_count < 0)
4985 conf->bypass_count = 0;
4997 list_del_init(&sh->lru);
4998 BUG_ON(atomic_inc_return(&sh->count) != 1);
5002 struct raid5_plug_cb {
5003 struct blk_plug_cb cb;
5004 struct list_head list;
5005 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5008 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5010 struct raid5_plug_cb *cb = container_of(
5011 blk_cb, struct raid5_plug_cb, cb);
5012 struct stripe_head *sh;
5013 struct mddev *mddev = cb->cb.data;
5014 struct r5conf *conf = mddev->private;
5018 if (cb->list.next && !list_empty(&cb->list)) {
5019 spin_lock_irq(&conf->device_lock);
5020 while (!list_empty(&cb->list)) {
5021 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5022 list_del_init(&sh->lru);
5024 * avoid race release_stripe_plug() sees
5025 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5026 * is still in our list
5028 smp_mb__before_atomic();
5029 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5031 * STRIPE_ON_RELEASE_LIST could be set here. In that
5032 * case, the count is always > 1 here
5034 hash = sh->hash_lock_index;
5035 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5038 spin_unlock_irq(&conf->device_lock);
5040 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5041 NR_STRIPE_HASH_LOCKS);
5043 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5047 static void release_stripe_plug(struct mddev *mddev,
5048 struct stripe_head *sh)
5050 struct blk_plug_cb *blk_cb = blk_check_plugged(
5051 raid5_unplug, mddev,
5052 sizeof(struct raid5_plug_cb));
5053 struct raid5_plug_cb *cb;
5056 raid5_release_stripe(sh);
5060 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5062 if (cb->list.next == NULL) {
5064 INIT_LIST_HEAD(&cb->list);
5065 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5066 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5069 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5070 list_add_tail(&sh->lru, &cb->list);
5072 raid5_release_stripe(sh);
5075 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5077 struct r5conf *conf = mddev->private;
5078 sector_t logical_sector, last_sector;
5079 struct stripe_head *sh;
5083 if (mddev->reshape_position != MaxSector)
5084 /* Skip discard while reshape is happening */
5087 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5088 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5091 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5093 stripe_sectors = conf->chunk_sectors *
5094 (conf->raid_disks - conf->max_degraded);
5095 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5097 sector_div(last_sector, stripe_sectors);
5099 logical_sector *= conf->chunk_sectors;
5100 last_sector *= conf->chunk_sectors;
5102 for (; logical_sector < last_sector;
5103 logical_sector += STRIPE_SECTORS) {
5107 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5108 prepare_to_wait(&conf->wait_for_overlap, &w,
5109 TASK_UNINTERRUPTIBLE);
5110 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5111 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5112 raid5_release_stripe(sh);
5116 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5117 spin_lock_irq(&sh->stripe_lock);
5118 for (d = 0; d < conf->raid_disks; d++) {
5119 if (d == sh->pd_idx || d == sh->qd_idx)
5121 if (sh->dev[d].towrite || sh->dev[d].toread) {
5122 set_bit(R5_Overlap, &sh->dev[d].flags);
5123 spin_unlock_irq(&sh->stripe_lock);
5124 raid5_release_stripe(sh);
5129 set_bit(STRIPE_DISCARD, &sh->state);
5130 finish_wait(&conf->wait_for_overlap, &w);
5131 sh->overwrite_disks = 0;
5132 for (d = 0; d < conf->raid_disks; d++) {
5133 if (d == sh->pd_idx || d == sh->qd_idx)
5135 sh->dev[d].towrite = bi;
5136 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5137 raid5_inc_bi_active_stripes(bi);
5138 sh->overwrite_disks++;
5140 spin_unlock_irq(&sh->stripe_lock);
5141 if (conf->mddev->bitmap) {
5143 d < conf->raid_disks - conf->max_degraded;
5145 bitmap_startwrite(mddev->bitmap,
5149 sh->bm_seq = conf->seq_flush + 1;
5150 set_bit(STRIPE_BIT_DELAY, &sh->state);
5153 set_bit(STRIPE_HANDLE, &sh->state);
5154 clear_bit(STRIPE_DELAYED, &sh->state);
5155 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5156 atomic_inc(&conf->preread_active_stripes);
5157 release_stripe_plug(mddev, sh);
5160 remaining = raid5_dec_bi_active_stripes(bi);
5161 if (remaining == 0) {
5162 md_write_end(mddev);
5167 static void raid5_make_request(struct mddev *mddev, struct bio * bi)
5169 struct r5conf *conf = mddev->private;
5171 sector_t new_sector;
5172 sector_t logical_sector, last_sector;
5173 struct stripe_head *sh;
5174 const int rw = bio_data_dir(bi);
5179 if (unlikely(bi->bi_rw & REQ_PREFLUSH)) {
5180 int ret = r5l_handle_flush_request(conf->log, bi);
5184 if (ret == -ENODEV) {
5185 md_flush_request(mddev, bi);
5188 /* ret == -EAGAIN, fallback */
5191 md_write_start(mddev, bi);
5194 * If array is degraded, better not do chunk aligned read because
5195 * later we might have to read it again in order to reconstruct
5196 * data on failed drives.
5198 if (rw == READ && mddev->degraded == 0 &&
5199 mddev->reshape_position == MaxSector) {
5200 bi = chunk_aligned_read(mddev, bi);
5205 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5206 make_discard_request(mddev, bi);
5210 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5211 last_sector = bio_end_sector(bi);
5213 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5215 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5216 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5222 seq = read_seqcount_begin(&conf->gen_lock);
5225 prepare_to_wait(&conf->wait_for_overlap, &w,
5226 TASK_UNINTERRUPTIBLE);
5227 if (unlikely(conf->reshape_progress != MaxSector)) {
5228 /* spinlock is needed as reshape_progress may be
5229 * 64bit on a 32bit platform, and so it might be
5230 * possible to see a half-updated value
5231 * Of course reshape_progress could change after
5232 * the lock is dropped, so once we get a reference
5233 * to the stripe that we think it is, we will have
5236 spin_lock_irq(&conf->device_lock);
5237 if (mddev->reshape_backwards
5238 ? logical_sector < conf->reshape_progress
5239 : logical_sector >= conf->reshape_progress) {
5242 if (mddev->reshape_backwards
5243 ? logical_sector < conf->reshape_safe
5244 : logical_sector >= conf->reshape_safe) {
5245 spin_unlock_irq(&conf->device_lock);
5251 spin_unlock_irq(&conf->device_lock);
5254 new_sector = raid5_compute_sector(conf, logical_sector,
5257 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5258 (unsigned long long)new_sector,
5259 (unsigned long long)logical_sector);
5261 sh = raid5_get_active_stripe(conf, new_sector, previous,
5262 (bi->bi_rw & REQ_RAHEAD), 0);
5264 if (unlikely(previous)) {
5265 /* expansion might have moved on while waiting for a
5266 * stripe, so we must do the range check again.
5267 * Expansion could still move past after this
5268 * test, but as we are holding a reference to
5269 * 'sh', we know that if that happens,
5270 * STRIPE_EXPANDING will get set and the expansion
5271 * won't proceed until we finish with the stripe.
5274 spin_lock_irq(&conf->device_lock);
5275 if (mddev->reshape_backwards
5276 ? logical_sector >= conf->reshape_progress
5277 : logical_sector < conf->reshape_progress)
5278 /* mismatch, need to try again */
5280 spin_unlock_irq(&conf->device_lock);
5282 raid5_release_stripe(sh);
5288 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5289 /* Might have got the wrong stripe_head
5292 raid5_release_stripe(sh);
5297 logical_sector >= mddev->suspend_lo &&
5298 logical_sector < mddev->suspend_hi) {
5299 raid5_release_stripe(sh);
5300 /* As the suspend_* range is controlled by
5301 * userspace, we want an interruptible
5304 flush_signals(current);
5305 prepare_to_wait(&conf->wait_for_overlap,
5306 &w, TASK_INTERRUPTIBLE);
5307 if (logical_sector >= mddev->suspend_lo &&
5308 logical_sector < mddev->suspend_hi) {
5315 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5316 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5317 /* Stripe is busy expanding or
5318 * add failed due to overlap. Flush everything
5321 md_wakeup_thread(mddev->thread);
5322 raid5_release_stripe(sh);
5327 set_bit(STRIPE_HANDLE, &sh->state);
5328 clear_bit(STRIPE_DELAYED, &sh->state);
5329 if ((!sh->batch_head || sh == sh->batch_head) &&
5330 (bi->bi_rw & REQ_SYNC) &&
5331 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5332 atomic_inc(&conf->preread_active_stripes);
5333 release_stripe_plug(mddev, sh);
5335 /* cannot get stripe for read-ahead, just give-up */
5336 bi->bi_error = -EIO;
5340 finish_wait(&conf->wait_for_overlap, &w);
5342 remaining = raid5_dec_bi_active_stripes(bi);
5343 if (remaining == 0) {
5346 md_write_end(mddev);
5348 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5354 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5356 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5358 /* reshaping is quite different to recovery/resync so it is
5359 * handled quite separately ... here.
5361 * On each call to sync_request, we gather one chunk worth of
5362 * destination stripes and flag them as expanding.
5363 * Then we find all the source stripes and request reads.
5364 * As the reads complete, handle_stripe will copy the data
5365 * into the destination stripe and release that stripe.
5367 struct r5conf *conf = mddev->private;
5368 struct stripe_head *sh;
5369 sector_t first_sector, last_sector;
5370 int raid_disks = conf->previous_raid_disks;
5371 int data_disks = raid_disks - conf->max_degraded;
5372 int new_data_disks = conf->raid_disks - conf->max_degraded;
5375 sector_t writepos, readpos, safepos;
5376 sector_t stripe_addr;
5377 int reshape_sectors;
5378 struct list_head stripes;
5381 if (sector_nr == 0) {
5382 /* If restarting in the middle, skip the initial sectors */
5383 if (mddev->reshape_backwards &&
5384 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5385 sector_nr = raid5_size(mddev, 0, 0)
5386 - conf->reshape_progress;
5387 } else if (mddev->reshape_backwards &&
5388 conf->reshape_progress == MaxSector) {
5389 /* shouldn't happen, but just in case, finish up.*/
5390 sector_nr = MaxSector;
5391 } else if (!mddev->reshape_backwards &&
5392 conf->reshape_progress > 0)
5393 sector_nr = conf->reshape_progress;
5394 sector_div(sector_nr, new_data_disks);
5396 mddev->curr_resync_completed = sector_nr;
5397 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5404 /* We need to process a full chunk at a time.
5405 * If old and new chunk sizes differ, we need to process the
5409 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5411 /* We update the metadata at least every 10 seconds, or when
5412 * the data about to be copied would over-write the source of
5413 * the data at the front of the range. i.e. one new_stripe
5414 * along from reshape_progress new_maps to after where
5415 * reshape_safe old_maps to
5417 writepos = conf->reshape_progress;
5418 sector_div(writepos, new_data_disks);
5419 readpos = conf->reshape_progress;
5420 sector_div(readpos, data_disks);
5421 safepos = conf->reshape_safe;
5422 sector_div(safepos, data_disks);
5423 if (mddev->reshape_backwards) {
5424 BUG_ON(writepos < reshape_sectors);
5425 writepos -= reshape_sectors;
5426 readpos += reshape_sectors;
5427 safepos += reshape_sectors;
5429 writepos += reshape_sectors;
5430 /* readpos and safepos are worst-case calculations.
5431 * A negative number is overly pessimistic, and causes
5432 * obvious problems for unsigned storage. So clip to 0.
5434 readpos -= min_t(sector_t, reshape_sectors, readpos);
5435 safepos -= min_t(sector_t, reshape_sectors, safepos);
5438 /* Having calculated the 'writepos' possibly use it
5439 * to set 'stripe_addr' which is where we will write to.
5441 if (mddev->reshape_backwards) {
5442 BUG_ON(conf->reshape_progress == 0);
5443 stripe_addr = writepos;
5444 BUG_ON((mddev->dev_sectors &
5445 ~((sector_t)reshape_sectors - 1))
5446 - reshape_sectors - stripe_addr
5449 BUG_ON(writepos != sector_nr + reshape_sectors);
5450 stripe_addr = sector_nr;
5453 /* 'writepos' is the most advanced device address we might write.
5454 * 'readpos' is the least advanced device address we might read.
5455 * 'safepos' is the least address recorded in the metadata as having
5457 * If there is a min_offset_diff, these are adjusted either by
5458 * increasing the safepos/readpos if diff is negative, or
5459 * increasing writepos if diff is positive.
5460 * If 'readpos' is then behind 'writepos', there is no way that we can
5461 * ensure safety in the face of a crash - that must be done by userspace
5462 * making a backup of the data. So in that case there is no particular
5463 * rush to update metadata.
5464 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5465 * update the metadata to advance 'safepos' to match 'readpos' so that
5466 * we can be safe in the event of a crash.
5467 * So we insist on updating metadata if safepos is behind writepos and
5468 * readpos is beyond writepos.
5469 * In any case, update the metadata every 10 seconds.
5470 * Maybe that number should be configurable, but I'm not sure it is
5471 * worth it.... maybe it could be a multiple of safemode_delay???
5473 if (conf->min_offset_diff < 0) {
5474 safepos += -conf->min_offset_diff;
5475 readpos += -conf->min_offset_diff;
5477 writepos += conf->min_offset_diff;
5479 if ((mddev->reshape_backwards
5480 ? (safepos > writepos && readpos < writepos)
5481 : (safepos < writepos && readpos > writepos)) ||
5482 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5483 /* Cannot proceed until we've updated the superblock... */
5484 wait_event(conf->wait_for_overlap,
5485 atomic_read(&conf->reshape_stripes)==0
5486 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5487 if (atomic_read(&conf->reshape_stripes) != 0)
5489 mddev->reshape_position = conf->reshape_progress;
5490 mddev->curr_resync_completed = sector_nr;
5491 conf->reshape_checkpoint = jiffies;
5492 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5493 md_wakeup_thread(mddev->thread);
5494 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5495 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5496 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5498 spin_lock_irq(&conf->device_lock);
5499 conf->reshape_safe = mddev->reshape_position;
5500 spin_unlock_irq(&conf->device_lock);
5501 wake_up(&conf->wait_for_overlap);
5502 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5505 INIT_LIST_HEAD(&stripes);
5506 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5508 int skipped_disk = 0;
5509 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5510 set_bit(STRIPE_EXPANDING, &sh->state);
5511 atomic_inc(&conf->reshape_stripes);
5512 /* If any of this stripe is beyond the end of the old
5513 * array, then we need to zero those blocks
5515 for (j=sh->disks; j--;) {
5517 if (j == sh->pd_idx)
5519 if (conf->level == 6 &&
5522 s = raid5_compute_blocknr(sh, j, 0);
5523 if (s < raid5_size(mddev, 0, 0)) {
5527 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5528 set_bit(R5_Expanded, &sh->dev[j].flags);
5529 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5531 if (!skipped_disk) {
5532 set_bit(STRIPE_EXPAND_READY, &sh->state);
5533 set_bit(STRIPE_HANDLE, &sh->state);
5535 list_add(&sh->lru, &stripes);
5537 spin_lock_irq(&conf->device_lock);
5538 if (mddev->reshape_backwards)
5539 conf->reshape_progress -= reshape_sectors * new_data_disks;
5541 conf->reshape_progress += reshape_sectors * new_data_disks;
5542 spin_unlock_irq(&conf->device_lock);
5543 /* Ok, those stripe are ready. We can start scheduling
5544 * reads on the source stripes.
5545 * The source stripes are determined by mapping the first and last
5546 * block on the destination stripes.
5549 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5552 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5553 * new_data_disks - 1),
5555 if (last_sector >= mddev->dev_sectors)
5556 last_sector = mddev->dev_sectors - 1;
5557 while (first_sector <= last_sector) {
5558 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5559 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5560 set_bit(STRIPE_HANDLE, &sh->state);
5561 raid5_release_stripe(sh);
5562 first_sector += STRIPE_SECTORS;
5564 /* Now that the sources are clearly marked, we can release
5565 * the destination stripes
5567 while (!list_empty(&stripes)) {
5568 sh = list_entry(stripes.next, struct stripe_head, lru);
5569 list_del_init(&sh->lru);
5570 raid5_release_stripe(sh);
5572 /* If this takes us to the resync_max point where we have to pause,
5573 * then we need to write out the superblock.
5575 sector_nr += reshape_sectors;
5576 retn = reshape_sectors;
5578 if (mddev->curr_resync_completed > mddev->resync_max ||
5579 (sector_nr - mddev->curr_resync_completed) * 2
5580 >= mddev->resync_max - mddev->curr_resync_completed) {
5581 /* Cannot proceed until we've updated the superblock... */
5582 wait_event(conf->wait_for_overlap,
5583 atomic_read(&conf->reshape_stripes) == 0
5584 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5585 if (atomic_read(&conf->reshape_stripes) != 0)
5587 mddev->reshape_position = conf->reshape_progress;
5588 mddev->curr_resync_completed = sector_nr;
5589 conf->reshape_checkpoint = jiffies;
5590 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5591 md_wakeup_thread(mddev->thread);
5592 wait_event(mddev->sb_wait,
5593 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5594 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5595 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5597 spin_lock_irq(&conf->device_lock);
5598 conf->reshape_safe = mddev->reshape_position;
5599 spin_unlock_irq(&conf->device_lock);
5600 wake_up(&conf->wait_for_overlap);
5601 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5607 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
5610 struct r5conf *conf = mddev->private;
5611 struct stripe_head *sh;
5612 sector_t max_sector = mddev->dev_sectors;
5613 sector_t sync_blocks;
5614 int still_degraded = 0;
5617 if (sector_nr >= max_sector) {
5618 /* just being told to finish up .. nothing much to do */
5620 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5625 if (mddev->curr_resync < max_sector) /* aborted */
5626 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5628 else /* completed sync */
5630 bitmap_close_sync(mddev->bitmap);
5635 /* Allow raid5_quiesce to complete */
5636 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5638 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5639 return reshape_request(mddev, sector_nr, skipped);
5641 /* No need to check resync_max as we never do more than one
5642 * stripe, and as resync_max will always be on a chunk boundary,
5643 * if the check in md_do_sync didn't fire, there is no chance
5644 * of overstepping resync_max here
5647 /* if there is too many failed drives and we are trying
5648 * to resync, then assert that we are finished, because there is
5649 * nothing we can do.
5651 if (mddev->degraded >= conf->max_degraded &&
5652 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5653 sector_t rv = mddev->dev_sectors - sector_nr;
5657 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5659 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5660 sync_blocks >= STRIPE_SECTORS) {
5661 /* we can skip this block, and probably more */
5662 sync_blocks /= STRIPE_SECTORS;
5664 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5667 bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5669 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5671 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5672 /* make sure we don't swamp the stripe cache if someone else
5673 * is trying to get access
5675 schedule_timeout_uninterruptible(1);
5677 /* Need to check if array will still be degraded after recovery/resync
5678 * Note in case of > 1 drive failures it's possible we're rebuilding
5679 * one drive while leaving another faulty drive in array.
5682 for (i = 0; i < conf->raid_disks; i++) {
5683 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5685 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5690 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5692 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5693 set_bit(STRIPE_HANDLE, &sh->state);
5695 raid5_release_stripe(sh);
5697 return STRIPE_SECTORS;
5700 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5702 /* We may not be able to submit a whole bio at once as there
5703 * may not be enough stripe_heads available.
5704 * We cannot pre-allocate enough stripe_heads as we may need
5705 * more than exist in the cache (if we allow ever large chunks).
5706 * So we do one stripe head at a time and record in
5707 * ->bi_hw_segments how many have been done.
5709 * We *know* that this entire raid_bio is in one chunk, so
5710 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5712 struct stripe_head *sh;
5714 sector_t sector, logical_sector, last_sector;
5719 logical_sector = raid_bio->bi_iter.bi_sector &
5720 ~((sector_t)STRIPE_SECTORS-1);
5721 sector = raid5_compute_sector(conf, logical_sector,
5723 last_sector = bio_end_sector(raid_bio);
5725 for (; logical_sector < last_sector;
5726 logical_sector += STRIPE_SECTORS,
5727 sector += STRIPE_SECTORS,
5730 if (scnt < raid5_bi_processed_stripes(raid_bio))
5731 /* already done this stripe */
5734 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5737 /* failed to get a stripe - must wait */
5738 raid5_set_bi_processed_stripes(raid_bio, scnt);
5739 conf->retry_read_aligned = raid_bio;
5743 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5744 raid5_release_stripe(sh);
5745 raid5_set_bi_processed_stripes(raid_bio, scnt);
5746 conf->retry_read_aligned = raid_bio;
5750 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5752 raid5_release_stripe(sh);
5755 remaining = raid5_dec_bi_active_stripes(raid_bio);
5756 if (remaining == 0) {
5757 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5759 bio_endio(raid_bio);
5761 if (atomic_dec_and_test(&conf->active_aligned_reads))
5762 wake_up(&conf->wait_for_quiescent);
5766 static int handle_active_stripes(struct r5conf *conf, int group,
5767 struct r5worker *worker,
5768 struct list_head *temp_inactive_list)
5770 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5771 int i, batch_size = 0, hash;
5772 bool release_inactive = false;
5774 while (batch_size < MAX_STRIPE_BATCH &&
5775 (sh = __get_priority_stripe(conf, group)) != NULL)
5776 batch[batch_size++] = sh;
5778 if (batch_size == 0) {
5779 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5780 if (!list_empty(temp_inactive_list + i))
5782 if (i == NR_STRIPE_HASH_LOCKS) {
5783 spin_unlock_irq(&conf->device_lock);
5784 r5l_flush_stripe_to_raid(conf->log);
5785 spin_lock_irq(&conf->device_lock);
5788 release_inactive = true;
5790 spin_unlock_irq(&conf->device_lock);
5792 release_inactive_stripe_list(conf, temp_inactive_list,
5793 NR_STRIPE_HASH_LOCKS);
5795 r5l_flush_stripe_to_raid(conf->log);
5796 if (release_inactive) {
5797 spin_lock_irq(&conf->device_lock);
5801 for (i = 0; i < batch_size; i++)
5802 handle_stripe(batch[i]);
5803 r5l_write_stripe_run(conf->log);
5807 spin_lock_irq(&conf->device_lock);
5808 for (i = 0; i < batch_size; i++) {
5809 hash = batch[i]->hash_lock_index;
5810 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5815 static void raid5_do_work(struct work_struct *work)
5817 struct r5worker *worker = container_of(work, struct r5worker, work);
5818 struct r5worker_group *group = worker->group;
5819 struct r5conf *conf = group->conf;
5820 int group_id = group - conf->worker_groups;
5822 struct blk_plug plug;
5824 pr_debug("+++ raid5worker active\n");
5826 blk_start_plug(&plug);
5828 spin_lock_irq(&conf->device_lock);
5830 int batch_size, released;
5832 released = release_stripe_list(conf, worker->temp_inactive_list);
5834 batch_size = handle_active_stripes(conf, group_id, worker,
5835 worker->temp_inactive_list);
5836 worker->working = false;
5837 if (!batch_size && !released)
5839 handled += batch_size;
5841 pr_debug("%d stripes handled\n", handled);
5843 spin_unlock_irq(&conf->device_lock);
5844 blk_finish_plug(&plug);
5846 pr_debug("--- raid5worker inactive\n");
5850 * This is our raid5 kernel thread.
5852 * We scan the hash table for stripes which can be handled now.
5853 * During the scan, completed stripes are saved for us by the interrupt
5854 * handler, so that they will not have to wait for our next wakeup.
5856 static void raid5d(struct md_thread *thread)
5858 struct mddev *mddev = thread->mddev;
5859 struct r5conf *conf = mddev->private;
5861 struct blk_plug plug;
5863 pr_debug("+++ raid5d active\n");
5865 md_check_recovery(mddev);
5867 if (!bio_list_empty(&conf->return_bi) &&
5868 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5869 struct bio_list tmp = BIO_EMPTY_LIST;
5870 spin_lock_irq(&conf->device_lock);
5871 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5872 bio_list_merge(&tmp, &conf->return_bi);
5873 bio_list_init(&conf->return_bi);
5875 spin_unlock_irq(&conf->device_lock);
5879 blk_start_plug(&plug);
5881 spin_lock_irq(&conf->device_lock);
5884 int batch_size, released;
5886 released = release_stripe_list(conf, conf->temp_inactive_list);
5888 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5891 !list_empty(&conf->bitmap_list)) {
5892 /* Now is a good time to flush some bitmap updates */
5894 spin_unlock_irq(&conf->device_lock);
5895 bitmap_unplug(mddev->bitmap);
5896 spin_lock_irq(&conf->device_lock);
5897 conf->seq_write = conf->seq_flush;
5898 activate_bit_delay(conf, conf->temp_inactive_list);
5900 raid5_activate_delayed(conf);
5902 while ((bio = remove_bio_from_retry(conf))) {
5904 spin_unlock_irq(&conf->device_lock);
5905 ok = retry_aligned_read(conf, bio);
5906 spin_lock_irq(&conf->device_lock);
5912 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5913 conf->temp_inactive_list);
5914 if (!batch_size && !released)
5916 handled += batch_size;
5918 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5919 spin_unlock_irq(&conf->device_lock);
5920 md_check_recovery(mddev);
5921 spin_lock_irq(&conf->device_lock);
5924 pr_debug("%d stripes handled\n", handled);
5926 spin_unlock_irq(&conf->device_lock);
5927 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5928 mutex_trylock(&conf->cache_size_mutex)) {
5929 grow_one_stripe(conf, __GFP_NOWARN);
5930 /* Set flag even if allocation failed. This helps
5931 * slow down allocation requests when mem is short
5933 set_bit(R5_DID_ALLOC, &conf->cache_state);
5934 mutex_unlock(&conf->cache_size_mutex);
5937 r5l_flush_stripe_to_raid(conf->log);
5939 async_tx_issue_pending_all();
5940 blk_finish_plug(&plug);
5942 pr_debug("--- raid5d inactive\n");
5946 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5948 struct r5conf *conf;
5950 spin_lock(&mddev->lock);
5951 conf = mddev->private;
5953 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5954 spin_unlock(&mddev->lock);
5959 raid5_set_cache_size(struct mddev *mddev, int size)
5961 struct r5conf *conf = mddev->private;
5964 if (size <= 16 || size > 32768)
5967 conf->min_nr_stripes = size;
5968 mutex_lock(&conf->cache_size_mutex);
5969 while (size < conf->max_nr_stripes &&
5970 drop_one_stripe(conf))
5972 mutex_unlock(&conf->cache_size_mutex);
5975 err = md_allow_write(mddev);
5979 mutex_lock(&conf->cache_size_mutex);
5980 while (size > conf->max_nr_stripes)
5981 if (!grow_one_stripe(conf, GFP_KERNEL))
5983 mutex_unlock(&conf->cache_size_mutex);
5987 EXPORT_SYMBOL(raid5_set_cache_size);
5990 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5992 struct r5conf *conf;
5996 if (len >= PAGE_SIZE)
5998 if (kstrtoul(page, 10, &new))
6000 err = mddev_lock(mddev);
6003 conf = mddev->private;
6007 err = raid5_set_cache_size(mddev, new);
6008 mddev_unlock(mddev);
6013 static struct md_sysfs_entry
6014 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6015 raid5_show_stripe_cache_size,
6016 raid5_store_stripe_cache_size);
6019 raid5_show_rmw_level(struct mddev *mddev, char *page)
6021 struct r5conf *conf = mddev->private;
6023 return sprintf(page, "%d\n", conf->rmw_level);
6029 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6031 struct r5conf *conf = mddev->private;
6037 if (len >= PAGE_SIZE)
6040 if (kstrtoul(page, 10, &new))
6043 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6046 if (new != PARITY_DISABLE_RMW &&
6047 new != PARITY_ENABLE_RMW &&
6048 new != PARITY_PREFER_RMW)
6051 conf->rmw_level = new;
6055 static struct md_sysfs_entry
6056 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6057 raid5_show_rmw_level,
6058 raid5_store_rmw_level);
6062 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6064 struct r5conf *conf;
6066 spin_lock(&mddev->lock);
6067 conf = mddev->private;
6069 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6070 spin_unlock(&mddev->lock);
6075 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6077 struct r5conf *conf;
6081 if (len >= PAGE_SIZE)
6083 if (kstrtoul(page, 10, &new))
6086 err = mddev_lock(mddev);
6089 conf = mddev->private;
6092 else if (new > conf->min_nr_stripes)
6095 conf->bypass_threshold = new;
6096 mddev_unlock(mddev);
6100 static struct md_sysfs_entry
6101 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6103 raid5_show_preread_threshold,
6104 raid5_store_preread_threshold);
6107 raid5_show_skip_copy(struct mddev *mddev, char *page)
6109 struct r5conf *conf;
6111 spin_lock(&mddev->lock);
6112 conf = mddev->private;
6114 ret = sprintf(page, "%d\n", conf->skip_copy);
6115 spin_unlock(&mddev->lock);
6120 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6122 struct r5conf *conf;
6126 if (len >= PAGE_SIZE)
6128 if (kstrtoul(page, 10, &new))
6132 err = mddev_lock(mddev);
6135 conf = mddev->private;
6138 else if (new != conf->skip_copy) {
6139 mddev_suspend(mddev);
6140 conf->skip_copy = new;
6142 mddev->queue->backing_dev_info.capabilities |=
6143 BDI_CAP_STABLE_WRITES;
6145 mddev->queue->backing_dev_info.capabilities &=
6146 ~BDI_CAP_STABLE_WRITES;
6147 mddev_resume(mddev);
6149 mddev_unlock(mddev);
6153 static struct md_sysfs_entry
6154 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6155 raid5_show_skip_copy,
6156 raid5_store_skip_copy);
6159 stripe_cache_active_show(struct mddev *mddev, char *page)
6161 struct r5conf *conf = mddev->private;
6163 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6168 static struct md_sysfs_entry
6169 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6172 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6174 struct r5conf *conf;
6176 spin_lock(&mddev->lock);
6177 conf = mddev->private;
6179 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6180 spin_unlock(&mddev->lock);
6184 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6186 int *worker_cnt_per_group,
6187 struct r5worker_group **worker_groups);
6189 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6191 struct r5conf *conf;
6194 struct r5worker_group *new_groups, *old_groups;
6195 int group_cnt, worker_cnt_per_group;
6197 if (len >= PAGE_SIZE)
6199 if (kstrtoul(page, 10, &new))
6202 err = mddev_lock(mddev);
6205 conf = mddev->private;
6208 else if (new != conf->worker_cnt_per_group) {
6209 mddev_suspend(mddev);
6211 old_groups = conf->worker_groups;
6213 flush_workqueue(raid5_wq);
6215 err = alloc_thread_groups(conf, new,
6216 &group_cnt, &worker_cnt_per_group,
6219 spin_lock_irq(&conf->device_lock);
6220 conf->group_cnt = group_cnt;
6221 conf->worker_cnt_per_group = worker_cnt_per_group;
6222 conf->worker_groups = new_groups;
6223 spin_unlock_irq(&conf->device_lock);
6226 kfree(old_groups[0].workers);
6229 mddev_resume(mddev);
6231 mddev_unlock(mddev);
6236 static struct md_sysfs_entry
6237 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6238 raid5_show_group_thread_cnt,
6239 raid5_store_group_thread_cnt);
6241 static struct attribute *raid5_attrs[] = {
6242 &raid5_stripecache_size.attr,
6243 &raid5_stripecache_active.attr,
6244 &raid5_preread_bypass_threshold.attr,
6245 &raid5_group_thread_cnt.attr,
6246 &raid5_skip_copy.attr,
6247 &raid5_rmw_level.attr,
6250 static struct attribute_group raid5_attrs_group = {
6252 .attrs = raid5_attrs,
6255 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6257 int *worker_cnt_per_group,
6258 struct r5worker_group **worker_groups)
6262 struct r5worker *workers;
6264 *worker_cnt_per_group = cnt;
6267 *worker_groups = NULL;
6270 *group_cnt = num_possible_nodes();
6271 size = sizeof(struct r5worker) * cnt;
6272 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6273 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6274 *group_cnt, GFP_NOIO);
6275 if (!*worker_groups || !workers) {
6277 kfree(*worker_groups);
6281 for (i = 0; i < *group_cnt; i++) {
6282 struct r5worker_group *group;
6284 group = &(*worker_groups)[i];
6285 INIT_LIST_HEAD(&group->handle_list);
6287 group->workers = workers + i * cnt;
6289 for (j = 0; j < cnt; j++) {
6290 struct r5worker *worker = group->workers + j;
6291 worker->group = group;
6292 INIT_WORK(&worker->work, raid5_do_work);
6294 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6295 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6302 static void free_thread_groups(struct r5conf *conf)
6304 if (conf->worker_groups)
6305 kfree(conf->worker_groups[0].workers);
6306 kfree(conf->worker_groups);
6307 conf->worker_groups = NULL;
6311 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6313 struct r5conf *conf = mddev->private;
6316 sectors = mddev->dev_sectors;
6318 /* size is defined by the smallest of previous and new size */
6319 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6321 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6322 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6323 return sectors * (raid_disks - conf->max_degraded);
6326 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6328 safe_put_page(percpu->spare_page);
6329 if (percpu->scribble)
6330 flex_array_free(percpu->scribble);
6331 percpu->spare_page = NULL;
6332 percpu->scribble = NULL;
6335 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6337 if (conf->level == 6 && !percpu->spare_page)
6338 percpu->spare_page = alloc_page(GFP_KERNEL);
6339 if (!percpu->scribble)
6340 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6341 conf->previous_raid_disks),
6342 max(conf->chunk_sectors,
6343 conf->prev_chunk_sectors)
6347 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6348 free_scratch_buffer(conf, percpu);
6355 static void raid5_free_percpu(struct r5conf *conf)
6362 #ifdef CONFIG_HOTPLUG_CPU
6363 unregister_cpu_notifier(&conf->cpu_notify);
6367 for_each_possible_cpu(cpu)
6368 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6371 free_percpu(conf->percpu);
6374 static void free_conf(struct r5conf *conf)
6377 r5l_exit_log(conf->log);
6378 if (conf->shrinker.seeks)
6379 unregister_shrinker(&conf->shrinker);
6381 free_thread_groups(conf);
6382 shrink_stripes(conf);
6383 raid5_free_percpu(conf);
6385 kfree(conf->stripe_hashtbl);
6389 #ifdef CONFIG_HOTPLUG_CPU
6390 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6393 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6394 long cpu = (long)hcpu;
6395 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6398 case CPU_UP_PREPARE:
6399 case CPU_UP_PREPARE_FROZEN:
6400 if (alloc_scratch_buffer(conf, percpu)) {
6401 pr_err("%s: failed memory allocation for cpu%ld\n",
6403 return notifier_from_errno(-ENOMEM);
6407 case CPU_DEAD_FROZEN:
6408 case CPU_UP_CANCELED:
6409 case CPU_UP_CANCELED_FROZEN:
6410 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6419 static int raid5_alloc_percpu(struct r5conf *conf)
6424 conf->percpu = alloc_percpu(struct raid5_percpu);
6428 #ifdef CONFIG_HOTPLUG_CPU
6429 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6430 conf->cpu_notify.priority = 0;
6431 err = register_cpu_notifier(&conf->cpu_notify);
6437 for_each_present_cpu(cpu) {
6438 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6440 pr_err("%s: failed memory allocation for cpu%ld\n",
6448 conf->scribble_disks = max(conf->raid_disks,
6449 conf->previous_raid_disks);
6450 conf->scribble_sectors = max(conf->chunk_sectors,
6451 conf->prev_chunk_sectors);
6456 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6457 struct shrink_control *sc)
6459 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6460 unsigned long ret = SHRINK_STOP;
6462 if (mutex_trylock(&conf->cache_size_mutex)) {
6464 while (ret < sc->nr_to_scan &&
6465 conf->max_nr_stripes > conf->min_nr_stripes) {
6466 if (drop_one_stripe(conf) == 0) {
6472 mutex_unlock(&conf->cache_size_mutex);
6477 static unsigned long raid5_cache_count(struct shrinker *shrink,
6478 struct shrink_control *sc)
6480 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6482 if (conf->max_nr_stripes < conf->min_nr_stripes)
6483 /* unlikely, but not impossible */
6485 return conf->max_nr_stripes - conf->min_nr_stripes;
6488 static struct r5conf *setup_conf(struct mddev *mddev)
6490 struct r5conf *conf;
6491 int raid_disk, memory, max_disks;
6492 struct md_rdev *rdev;
6493 struct disk_info *disk;
6496 int group_cnt, worker_cnt_per_group;
6497 struct r5worker_group *new_group;
6499 if (mddev->new_level != 5
6500 && mddev->new_level != 4
6501 && mddev->new_level != 6) {
6502 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6503 mdname(mddev), mddev->new_level);
6504 return ERR_PTR(-EIO);
6506 if ((mddev->new_level == 5
6507 && !algorithm_valid_raid5(mddev->new_layout)) ||
6508 (mddev->new_level == 6
6509 && !algorithm_valid_raid6(mddev->new_layout))) {
6510 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6511 mdname(mddev), mddev->new_layout);
6512 return ERR_PTR(-EIO);
6514 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6515 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6516 mdname(mddev), mddev->raid_disks);
6517 return ERR_PTR(-EINVAL);
6520 if (!mddev->new_chunk_sectors ||
6521 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6522 !is_power_of_2(mddev->new_chunk_sectors)) {
6523 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6524 mdname(mddev), mddev->new_chunk_sectors << 9);
6525 return ERR_PTR(-EINVAL);
6528 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6531 /* Don't enable multi-threading by default*/
6532 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6534 conf->group_cnt = group_cnt;
6535 conf->worker_cnt_per_group = worker_cnt_per_group;
6536 conf->worker_groups = new_group;
6539 spin_lock_init(&conf->device_lock);
6540 seqcount_init(&conf->gen_lock);
6541 mutex_init(&conf->cache_size_mutex);
6542 init_waitqueue_head(&conf->wait_for_quiescent);
6543 init_waitqueue_head(&conf->wait_for_stripe);
6544 init_waitqueue_head(&conf->wait_for_overlap);
6545 INIT_LIST_HEAD(&conf->handle_list);
6546 INIT_LIST_HEAD(&conf->hold_list);
6547 INIT_LIST_HEAD(&conf->delayed_list);
6548 INIT_LIST_HEAD(&conf->bitmap_list);
6549 bio_list_init(&conf->return_bi);
6550 init_llist_head(&conf->released_stripes);
6551 atomic_set(&conf->active_stripes, 0);
6552 atomic_set(&conf->preread_active_stripes, 0);
6553 atomic_set(&conf->active_aligned_reads, 0);
6554 conf->bypass_threshold = BYPASS_THRESHOLD;
6555 conf->recovery_disabled = mddev->recovery_disabled - 1;
6557 conf->raid_disks = mddev->raid_disks;
6558 if (mddev->reshape_position == MaxSector)
6559 conf->previous_raid_disks = mddev->raid_disks;
6561 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6562 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6564 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6569 conf->mddev = mddev;
6571 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6574 /* We init hash_locks[0] separately to that it can be used
6575 * as the reference lock in the spin_lock_nest_lock() call
6576 * in lock_all_device_hash_locks_irq in order to convince
6577 * lockdep that we know what we are doing.
6579 spin_lock_init(conf->hash_locks);
6580 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6581 spin_lock_init(conf->hash_locks + i);
6583 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6584 INIT_LIST_HEAD(conf->inactive_list + i);
6586 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6587 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6589 conf->level = mddev->new_level;
6590 conf->chunk_sectors = mddev->new_chunk_sectors;
6591 if (raid5_alloc_percpu(conf) != 0)
6594 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6596 rdev_for_each(rdev, mddev) {
6597 raid_disk = rdev->raid_disk;
6598 if (raid_disk >= max_disks
6599 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6601 disk = conf->disks + raid_disk;
6603 if (test_bit(Replacement, &rdev->flags)) {
6604 if (disk->replacement)
6606 disk->replacement = rdev;
6613 if (test_bit(In_sync, &rdev->flags)) {
6614 char b[BDEVNAME_SIZE];
6615 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6617 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6618 } else if (rdev->saved_raid_disk != raid_disk)
6619 /* Cannot rely on bitmap to complete recovery */
6623 conf->level = mddev->new_level;
6624 if (conf->level == 6) {
6625 conf->max_degraded = 2;
6626 if (raid6_call.xor_syndrome)
6627 conf->rmw_level = PARITY_ENABLE_RMW;
6629 conf->rmw_level = PARITY_DISABLE_RMW;
6631 conf->max_degraded = 1;
6632 conf->rmw_level = PARITY_ENABLE_RMW;
6634 conf->algorithm = mddev->new_layout;
6635 conf->reshape_progress = mddev->reshape_position;
6636 if (conf->reshape_progress != MaxSector) {
6637 conf->prev_chunk_sectors = mddev->chunk_sectors;
6638 conf->prev_algo = mddev->layout;
6640 conf->prev_chunk_sectors = conf->chunk_sectors;
6641 conf->prev_algo = conf->algorithm;
6644 conf->min_nr_stripes = NR_STRIPES;
6645 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6646 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6647 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6648 if (grow_stripes(conf, conf->min_nr_stripes)) {
6650 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6651 mdname(mddev), memory);
6654 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6655 mdname(mddev), memory);
6657 * Losing a stripe head costs more than the time to refill it,
6658 * it reduces the queue depth and so can hurt throughput.
6659 * So set it rather large, scaled by number of devices.
6661 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6662 conf->shrinker.scan_objects = raid5_cache_scan;
6663 conf->shrinker.count_objects = raid5_cache_count;
6664 conf->shrinker.batch = 128;
6665 conf->shrinker.flags = 0;
6666 register_shrinker(&conf->shrinker);
6668 sprintf(pers_name, "raid%d", mddev->new_level);
6669 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6670 if (!conf->thread) {
6672 "md/raid:%s: couldn't allocate thread.\n",
6682 return ERR_PTR(-EIO);
6684 return ERR_PTR(-ENOMEM);
6687 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6690 case ALGORITHM_PARITY_0:
6691 if (raid_disk < max_degraded)
6694 case ALGORITHM_PARITY_N:
6695 if (raid_disk >= raid_disks - max_degraded)
6698 case ALGORITHM_PARITY_0_6:
6699 if (raid_disk == 0 ||
6700 raid_disk == raid_disks - 1)
6703 case ALGORITHM_LEFT_ASYMMETRIC_6:
6704 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6705 case ALGORITHM_LEFT_SYMMETRIC_6:
6706 case ALGORITHM_RIGHT_SYMMETRIC_6:
6707 if (raid_disk == raid_disks - 1)
6713 static int raid5_run(struct mddev *mddev)
6715 struct r5conf *conf;
6716 int working_disks = 0;
6717 int dirty_parity_disks = 0;
6718 struct md_rdev *rdev;
6719 struct md_rdev *journal_dev = NULL;
6720 sector_t reshape_offset = 0;
6722 long long min_offset_diff = 0;
6725 if (mddev->recovery_cp != MaxSector)
6726 printk(KERN_NOTICE "md/raid:%s: not clean"
6727 " -- starting background reconstruction\n",
6730 rdev_for_each(rdev, mddev) {
6733 if (test_bit(Journal, &rdev->flags)) {
6737 if (rdev->raid_disk < 0)
6739 diff = (rdev->new_data_offset - rdev->data_offset);
6741 min_offset_diff = diff;
6743 } else if (mddev->reshape_backwards &&
6744 diff < min_offset_diff)
6745 min_offset_diff = diff;
6746 else if (!mddev->reshape_backwards &&
6747 diff > min_offset_diff)
6748 min_offset_diff = diff;
6751 if (mddev->reshape_position != MaxSector) {
6752 /* Check that we can continue the reshape.
6753 * Difficulties arise if the stripe we would write to
6754 * next is at or after the stripe we would read from next.
6755 * For a reshape that changes the number of devices, this
6756 * is only possible for a very short time, and mdadm makes
6757 * sure that time appears to have past before assembling
6758 * the array. So we fail if that time hasn't passed.
6759 * For a reshape that keeps the number of devices the same
6760 * mdadm must be monitoring the reshape can keeping the
6761 * critical areas read-only and backed up. It will start
6762 * the array in read-only mode, so we check for that.
6764 sector_t here_new, here_old;
6766 int max_degraded = (mddev->level == 6 ? 2 : 1);
6771 printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6776 if (mddev->new_level != mddev->level) {
6777 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6778 "required - aborting.\n",
6782 old_disks = mddev->raid_disks - mddev->delta_disks;
6783 /* reshape_position must be on a new-stripe boundary, and one
6784 * further up in new geometry must map after here in old
6786 * If the chunk sizes are different, then as we perform reshape
6787 * in units of the largest of the two, reshape_position needs
6788 * be a multiple of the largest chunk size times new data disks.
6790 here_new = mddev->reshape_position;
6791 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6792 new_data_disks = mddev->raid_disks - max_degraded;
6793 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6794 printk(KERN_ERR "md/raid:%s: reshape_position not "
6795 "on a stripe boundary\n", mdname(mddev));
6798 reshape_offset = here_new * chunk_sectors;
6799 /* here_new is the stripe we will write to */
6800 here_old = mddev->reshape_position;
6801 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6802 /* here_old is the first stripe that we might need to read
6804 if (mddev->delta_disks == 0) {
6805 /* We cannot be sure it is safe to start an in-place
6806 * reshape. It is only safe if user-space is monitoring
6807 * and taking constant backups.
6808 * mdadm always starts a situation like this in
6809 * readonly mode so it can take control before
6810 * allowing any writes. So just check for that.
6812 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6813 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6814 /* not really in-place - so OK */;
6815 else if (mddev->ro == 0) {
6816 printk(KERN_ERR "md/raid:%s: in-place reshape "
6817 "must be started in read-only mode "
6822 } else if (mddev->reshape_backwards
6823 ? (here_new * chunk_sectors + min_offset_diff <=
6824 here_old * chunk_sectors)
6825 : (here_new * chunk_sectors >=
6826 here_old * chunk_sectors + (-min_offset_diff))) {
6827 /* Reading from the same stripe as writing to - bad */
6828 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6829 "auto-recovery - aborting.\n",
6833 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6835 /* OK, we should be able to continue; */
6837 BUG_ON(mddev->level != mddev->new_level);
6838 BUG_ON(mddev->layout != mddev->new_layout);
6839 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6840 BUG_ON(mddev->delta_disks != 0);
6843 if (mddev->private == NULL)
6844 conf = setup_conf(mddev);
6846 conf = mddev->private;
6849 return PTR_ERR(conf);
6851 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
6853 pr_err("md/raid:%s: journal disk is missing, force array readonly\n",
6856 set_disk_ro(mddev->gendisk, 1);
6857 } else if (mddev->recovery_cp == MaxSector)
6858 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
6861 conf->min_offset_diff = min_offset_diff;
6862 mddev->thread = conf->thread;
6863 conf->thread = NULL;
6864 mddev->private = conf;
6866 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6868 rdev = conf->disks[i].rdev;
6869 if (!rdev && conf->disks[i].replacement) {
6870 /* The replacement is all we have yet */
6871 rdev = conf->disks[i].replacement;
6872 conf->disks[i].replacement = NULL;
6873 clear_bit(Replacement, &rdev->flags);
6874 conf->disks[i].rdev = rdev;
6878 if (conf->disks[i].replacement &&
6879 conf->reshape_progress != MaxSector) {
6880 /* replacements and reshape simply do not mix. */
6881 printk(KERN_ERR "md: cannot handle concurrent "
6882 "replacement and reshape.\n");
6885 if (test_bit(In_sync, &rdev->flags)) {
6889 /* This disc is not fully in-sync. However if it
6890 * just stored parity (beyond the recovery_offset),
6891 * when we don't need to be concerned about the
6892 * array being dirty.
6893 * When reshape goes 'backwards', we never have
6894 * partially completed devices, so we only need
6895 * to worry about reshape going forwards.
6897 /* Hack because v0.91 doesn't store recovery_offset properly. */
6898 if (mddev->major_version == 0 &&
6899 mddev->minor_version > 90)
6900 rdev->recovery_offset = reshape_offset;
6902 if (rdev->recovery_offset < reshape_offset) {
6903 /* We need to check old and new layout */
6904 if (!only_parity(rdev->raid_disk,
6907 conf->max_degraded))
6910 if (!only_parity(rdev->raid_disk,
6912 conf->previous_raid_disks,
6913 conf->max_degraded))
6915 dirty_parity_disks++;
6919 * 0 for a fully functional array, 1 or 2 for a degraded array.
6921 mddev->degraded = calc_degraded(conf);
6923 if (has_failed(conf)) {
6924 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6925 " (%d/%d failed)\n",
6926 mdname(mddev), mddev->degraded, conf->raid_disks);
6930 /* device size must be a multiple of chunk size */
6931 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6932 mddev->resync_max_sectors = mddev->dev_sectors;
6934 if (mddev->degraded > dirty_parity_disks &&
6935 mddev->recovery_cp != MaxSector) {
6936 if (mddev->ok_start_degraded)
6938 "md/raid:%s: starting dirty degraded array"
6939 " - data corruption possible.\n",
6943 "md/raid:%s: cannot start dirty degraded array.\n",
6949 if (mddev->degraded == 0)
6950 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6951 " devices, algorithm %d\n", mdname(mddev), conf->level,
6952 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6955 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6956 " out of %d devices, algorithm %d\n",
6957 mdname(mddev), conf->level,
6958 mddev->raid_disks - mddev->degraded,
6959 mddev->raid_disks, mddev->new_layout);
6961 print_raid5_conf(conf);
6963 if (conf->reshape_progress != MaxSector) {
6964 conf->reshape_safe = conf->reshape_progress;
6965 atomic_set(&conf->reshape_stripes, 0);
6966 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6967 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6968 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6969 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6970 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6974 /* Ok, everything is just fine now */
6975 if (mddev->to_remove == &raid5_attrs_group)
6976 mddev->to_remove = NULL;
6977 else if (mddev->kobj.sd &&
6978 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6980 "raid5: failed to create sysfs attributes for %s\n",
6982 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6986 bool discard_supported = true;
6987 /* read-ahead size must cover two whole stripes, which
6988 * is 2 * (datadisks) * chunksize where 'n' is the
6989 * number of raid devices
6991 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6992 int stripe = data_disks *
6993 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6994 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6995 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6997 chunk_size = mddev->chunk_sectors << 9;
6998 blk_queue_io_min(mddev->queue, chunk_size);
6999 blk_queue_io_opt(mddev->queue, chunk_size *
7000 (conf->raid_disks - conf->max_degraded));
7001 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7003 * We can only discard a whole stripe. It doesn't make sense to
7004 * discard data disk but write parity disk
7006 stripe = stripe * PAGE_SIZE;
7007 /* Round up to power of 2, as discard handling
7008 * currently assumes that */
7009 while ((stripe-1) & stripe)
7010 stripe = (stripe | (stripe-1)) + 1;
7011 mddev->queue->limits.discard_alignment = stripe;
7012 mddev->queue->limits.discard_granularity = stripe;
7014 * unaligned part of discard request will be ignored, so can't
7015 * guarantee discard_zeroes_data
7017 mddev->queue->limits.discard_zeroes_data = 0;
7019 blk_queue_max_write_same_sectors(mddev->queue, 0);
7021 rdev_for_each(rdev, mddev) {
7022 disk_stack_limits(mddev->gendisk, rdev->bdev,
7023 rdev->data_offset << 9);
7024 disk_stack_limits(mddev->gendisk, rdev->bdev,
7025 rdev->new_data_offset << 9);
7027 * discard_zeroes_data is required, otherwise data
7028 * could be lost. Consider a scenario: discard a stripe
7029 * (the stripe could be inconsistent if
7030 * discard_zeroes_data is 0); write one disk of the
7031 * stripe (the stripe could be inconsistent again
7032 * depending on which disks are used to calculate
7033 * parity); the disk is broken; The stripe data of this
7036 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7037 !bdev_get_queue(rdev->bdev)->
7038 limits.discard_zeroes_data)
7039 discard_supported = false;
7040 /* Unfortunately, discard_zeroes_data is not currently
7041 * a guarantee - just a hint. So we only allow DISCARD
7042 * if the sysadmin has confirmed that only safe devices
7043 * are in use by setting a module parameter.
7045 if (!devices_handle_discard_safely) {
7046 if (discard_supported) {
7047 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7048 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7050 discard_supported = false;
7054 if (discard_supported &&
7055 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7056 mddev->queue->limits.discard_granularity >= stripe)
7057 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7060 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7065 char b[BDEVNAME_SIZE];
7067 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7068 mdname(mddev), bdevname(journal_dev->bdev, b));
7069 r5l_init_log(conf, journal_dev);
7074 md_unregister_thread(&mddev->thread);
7075 print_raid5_conf(conf);
7077 mddev->private = NULL;
7078 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7082 static void raid5_free(struct mddev *mddev, void *priv)
7084 struct r5conf *conf = priv;
7087 mddev->to_remove = &raid5_attrs_group;
7090 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7092 struct r5conf *conf = mddev->private;
7095 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7096 conf->chunk_sectors / 2, mddev->layout);
7097 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7099 for (i = 0; i < conf->raid_disks; i++) {
7100 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7101 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7104 seq_printf (seq, "]");
7107 static void print_raid5_conf (struct r5conf *conf)
7110 struct disk_info *tmp;
7112 printk(KERN_DEBUG "RAID conf printout:\n");
7114 printk("(conf==NULL)\n");
7117 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7119 conf->raid_disks - conf->mddev->degraded);
7121 for (i = 0; i < conf->raid_disks; i++) {
7122 char b[BDEVNAME_SIZE];
7123 tmp = conf->disks + i;
7125 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7126 i, !test_bit(Faulty, &tmp->rdev->flags),
7127 bdevname(tmp->rdev->bdev, b));
7131 static int raid5_spare_active(struct mddev *mddev)
7134 struct r5conf *conf = mddev->private;
7135 struct disk_info *tmp;
7137 unsigned long flags;
7139 for (i = 0; i < conf->raid_disks; i++) {
7140 tmp = conf->disks + i;
7141 if (tmp->replacement
7142 && tmp->replacement->recovery_offset == MaxSector
7143 && !test_bit(Faulty, &tmp->replacement->flags)
7144 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7145 /* Replacement has just become active. */
7147 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7150 /* Replaced device not technically faulty,
7151 * but we need to be sure it gets removed
7152 * and never re-added.
7154 set_bit(Faulty, &tmp->rdev->flags);
7155 sysfs_notify_dirent_safe(
7156 tmp->rdev->sysfs_state);
7158 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7159 } else if (tmp->rdev
7160 && tmp->rdev->recovery_offset == MaxSector
7161 && !test_bit(Faulty, &tmp->rdev->flags)
7162 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7164 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7167 spin_lock_irqsave(&conf->device_lock, flags);
7168 mddev->degraded = calc_degraded(conf);
7169 spin_unlock_irqrestore(&conf->device_lock, flags);
7170 print_raid5_conf(conf);
7174 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7176 struct r5conf *conf = mddev->private;
7178 int number = rdev->raid_disk;
7179 struct md_rdev **rdevp;
7180 struct disk_info *p = conf->disks + number;
7182 print_raid5_conf(conf);
7183 if (test_bit(Journal, &rdev->flags) && conf->log) {
7184 struct r5l_log *log;
7186 * we can't wait pending write here, as this is called in
7187 * raid5d, wait will deadlock.
7189 if (atomic_read(&mddev->writes_pending))
7197 if (rdev == p->rdev)
7199 else if (rdev == p->replacement)
7200 rdevp = &p->replacement;
7204 if (number >= conf->raid_disks &&
7205 conf->reshape_progress == MaxSector)
7206 clear_bit(In_sync, &rdev->flags);
7208 if (test_bit(In_sync, &rdev->flags) ||
7209 atomic_read(&rdev->nr_pending)) {
7213 /* Only remove non-faulty devices if recovery
7216 if (!test_bit(Faulty, &rdev->flags) &&
7217 mddev->recovery_disabled != conf->recovery_disabled &&
7218 !has_failed(conf) &&
7219 (!p->replacement || p->replacement == rdev) &&
7220 number < conf->raid_disks) {
7225 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7227 if (atomic_read(&rdev->nr_pending)) {
7228 /* lost the race, try later */
7233 if (p->replacement) {
7234 /* We must have just cleared 'rdev' */
7235 p->rdev = p->replacement;
7236 clear_bit(Replacement, &p->replacement->flags);
7237 smp_mb(); /* Make sure other CPUs may see both as identical
7238 * but will never see neither - if they are careful
7240 p->replacement = NULL;
7241 clear_bit(WantReplacement, &rdev->flags);
7243 /* We might have just removed the Replacement as faulty-
7244 * clear the bit just in case
7246 clear_bit(WantReplacement, &rdev->flags);
7249 print_raid5_conf(conf);
7253 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7255 struct r5conf *conf = mddev->private;
7258 struct disk_info *p;
7260 int last = conf->raid_disks - 1;
7262 if (test_bit(Journal, &rdev->flags)) {
7263 char b[BDEVNAME_SIZE];
7267 rdev->raid_disk = 0;
7269 * The array is in readonly mode if journal is missing, so no
7270 * write requests running. We should be safe
7272 r5l_init_log(conf, rdev);
7273 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7274 mdname(mddev), bdevname(rdev->bdev, b));
7277 if (mddev->recovery_disabled == conf->recovery_disabled)
7280 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7281 /* no point adding a device */
7284 if (rdev->raid_disk >= 0)
7285 first = last = rdev->raid_disk;
7288 * find the disk ... but prefer rdev->saved_raid_disk
7291 if (rdev->saved_raid_disk >= 0 &&
7292 rdev->saved_raid_disk >= first &&
7293 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7294 first = rdev->saved_raid_disk;
7296 for (disk = first; disk <= last; disk++) {
7297 p = conf->disks + disk;
7298 if (p->rdev == NULL) {
7299 clear_bit(In_sync, &rdev->flags);
7300 rdev->raid_disk = disk;
7302 if (rdev->saved_raid_disk != disk)
7304 rcu_assign_pointer(p->rdev, rdev);
7308 for (disk = first; disk <= last; disk++) {
7309 p = conf->disks + disk;
7310 if (test_bit(WantReplacement, &p->rdev->flags) &&
7311 p->replacement == NULL) {
7312 clear_bit(In_sync, &rdev->flags);
7313 set_bit(Replacement, &rdev->flags);
7314 rdev->raid_disk = disk;
7317 rcu_assign_pointer(p->replacement, rdev);
7322 print_raid5_conf(conf);
7326 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7328 /* no resync is happening, and there is enough space
7329 * on all devices, so we can resize.
7330 * We need to make sure resync covers any new space.
7331 * If the array is shrinking we should possibly wait until
7332 * any io in the removed space completes, but it hardly seems
7336 struct r5conf *conf = mddev->private;
7340 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7341 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7342 if (mddev->external_size &&
7343 mddev->array_sectors > newsize)
7345 if (mddev->bitmap) {
7346 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7350 md_set_array_sectors(mddev, newsize);
7351 set_capacity(mddev->gendisk, mddev->array_sectors);
7352 revalidate_disk(mddev->gendisk);
7353 if (sectors > mddev->dev_sectors &&
7354 mddev->recovery_cp > mddev->dev_sectors) {
7355 mddev->recovery_cp = mddev->dev_sectors;
7356 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7358 mddev->dev_sectors = sectors;
7359 mddev->resync_max_sectors = sectors;
7363 static int check_stripe_cache(struct mddev *mddev)
7365 /* Can only proceed if there are plenty of stripe_heads.
7366 * We need a minimum of one full stripe,, and for sensible progress
7367 * it is best to have about 4 times that.
7368 * If we require 4 times, then the default 256 4K stripe_heads will
7369 * allow for chunk sizes up to 256K, which is probably OK.
7370 * If the chunk size is greater, user-space should request more
7371 * stripe_heads first.
7373 struct r5conf *conf = mddev->private;
7374 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7375 > conf->min_nr_stripes ||
7376 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7377 > conf->min_nr_stripes) {
7378 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7380 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7387 static int check_reshape(struct mddev *mddev)
7389 struct r5conf *conf = mddev->private;
7393 if (mddev->delta_disks == 0 &&
7394 mddev->new_layout == mddev->layout &&
7395 mddev->new_chunk_sectors == mddev->chunk_sectors)
7396 return 0; /* nothing to do */
7397 if (has_failed(conf))
7399 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7400 /* We might be able to shrink, but the devices must
7401 * be made bigger first.
7402 * For raid6, 4 is the minimum size.
7403 * Otherwise 2 is the minimum
7406 if (mddev->level == 6)
7408 if (mddev->raid_disks + mddev->delta_disks < min)
7412 if (!check_stripe_cache(mddev))
7415 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7416 mddev->delta_disks > 0)
7417 if (resize_chunks(conf,
7418 conf->previous_raid_disks
7419 + max(0, mddev->delta_disks),
7420 max(mddev->new_chunk_sectors,
7421 mddev->chunk_sectors)
7424 return resize_stripes(conf, (conf->previous_raid_disks
7425 + mddev->delta_disks));
7428 static int raid5_start_reshape(struct mddev *mddev)
7430 struct r5conf *conf = mddev->private;
7431 struct md_rdev *rdev;
7433 unsigned long flags;
7435 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7438 if (!check_stripe_cache(mddev))
7441 if (has_failed(conf))
7444 rdev_for_each(rdev, mddev) {
7445 if (!test_bit(In_sync, &rdev->flags)
7446 && !test_bit(Faulty, &rdev->flags))
7450 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7451 /* Not enough devices even to make a degraded array
7456 /* Refuse to reduce size of the array. Any reductions in
7457 * array size must be through explicit setting of array_size
7460 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7461 < mddev->array_sectors) {
7462 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7463 "before number of disks\n", mdname(mddev));
7467 atomic_set(&conf->reshape_stripes, 0);
7468 spin_lock_irq(&conf->device_lock);
7469 write_seqcount_begin(&conf->gen_lock);
7470 conf->previous_raid_disks = conf->raid_disks;
7471 conf->raid_disks += mddev->delta_disks;
7472 conf->prev_chunk_sectors = conf->chunk_sectors;
7473 conf->chunk_sectors = mddev->new_chunk_sectors;
7474 conf->prev_algo = conf->algorithm;
7475 conf->algorithm = mddev->new_layout;
7477 /* Code that selects data_offset needs to see the generation update
7478 * if reshape_progress has been set - so a memory barrier needed.
7481 if (mddev->reshape_backwards)
7482 conf->reshape_progress = raid5_size(mddev, 0, 0);
7484 conf->reshape_progress = 0;
7485 conf->reshape_safe = conf->reshape_progress;
7486 write_seqcount_end(&conf->gen_lock);
7487 spin_unlock_irq(&conf->device_lock);
7489 /* Now make sure any requests that proceeded on the assumption
7490 * the reshape wasn't running - like Discard or Read - have
7493 mddev_suspend(mddev);
7494 mddev_resume(mddev);
7496 /* Add some new drives, as many as will fit.
7497 * We know there are enough to make the newly sized array work.
7498 * Don't add devices if we are reducing the number of
7499 * devices in the array. This is because it is not possible
7500 * to correctly record the "partially reconstructed" state of
7501 * such devices during the reshape and confusion could result.
7503 if (mddev->delta_disks >= 0) {
7504 rdev_for_each(rdev, mddev)
7505 if (rdev->raid_disk < 0 &&
7506 !test_bit(Faulty, &rdev->flags)) {
7507 if (raid5_add_disk(mddev, rdev) == 0) {
7509 >= conf->previous_raid_disks)
7510 set_bit(In_sync, &rdev->flags);
7512 rdev->recovery_offset = 0;
7514 if (sysfs_link_rdev(mddev, rdev))
7515 /* Failure here is OK */;
7517 } else if (rdev->raid_disk >= conf->previous_raid_disks
7518 && !test_bit(Faulty, &rdev->flags)) {
7519 /* This is a spare that was manually added */
7520 set_bit(In_sync, &rdev->flags);
7523 /* When a reshape changes the number of devices,
7524 * ->degraded is measured against the larger of the
7525 * pre and post number of devices.
7527 spin_lock_irqsave(&conf->device_lock, flags);
7528 mddev->degraded = calc_degraded(conf);
7529 spin_unlock_irqrestore(&conf->device_lock, flags);
7531 mddev->raid_disks = conf->raid_disks;
7532 mddev->reshape_position = conf->reshape_progress;
7533 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7535 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7536 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7537 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7538 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7539 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7540 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7542 if (!mddev->sync_thread) {
7543 mddev->recovery = 0;
7544 spin_lock_irq(&conf->device_lock);
7545 write_seqcount_begin(&conf->gen_lock);
7546 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7547 mddev->new_chunk_sectors =
7548 conf->chunk_sectors = conf->prev_chunk_sectors;
7549 mddev->new_layout = conf->algorithm = conf->prev_algo;
7550 rdev_for_each(rdev, mddev)
7551 rdev->new_data_offset = rdev->data_offset;
7553 conf->generation --;
7554 conf->reshape_progress = MaxSector;
7555 mddev->reshape_position = MaxSector;
7556 write_seqcount_end(&conf->gen_lock);
7557 spin_unlock_irq(&conf->device_lock);
7560 conf->reshape_checkpoint = jiffies;
7561 md_wakeup_thread(mddev->sync_thread);
7562 md_new_event(mddev);
7566 /* This is called from the reshape thread and should make any
7567 * changes needed in 'conf'
7569 static void end_reshape(struct r5conf *conf)
7572 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7573 struct md_rdev *rdev;
7575 spin_lock_irq(&conf->device_lock);
7576 conf->previous_raid_disks = conf->raid_disks;
7577 rdev_for_each(rdev, conf->mddev)
7578 rdev->data_offset = rdev->new_data_offset;
7580 conf->reshape_progress = MaxSector;
7581 conf->mddev->reshape_position = MaxSector;
7582 spin_unlock_irq(&conf->device_lock);
7583 wake_up(&conf->wait_for_overlap);
7585 /* read-ahead size must cover two whole stripes, which is
7586 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7588 if (conf->mddev->queue) {
7589 int data_disks = conf->raid_disks - conf->max_degraded;
7590 int stripe = data_disks * ((conf->chunk_sectors << 9)
7592 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7593 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7598 /* This is called from the raid5d thread with mddev_lock held.
7599 * It makes config changes to the device.
7601 static void raid5_finish_reshape(struct mddev *mddev)
7603 struct r5conf *conf = mddev->private;
7605 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7607 if (mddev->delta_disks > 0) {
7608 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7610 set_capacity(mddev->gendisk, mddev->array_sectors);
7611 revalidate_disk(mddev->gendisk);
7615 spin_lock_irq(&conf->device_lock);
7616 mddev->degraded = calc_degraded(conf);
7617 spin_unlock_irq(&conf->device_lock);
7618 for (d = conf->raid_disks ;
7619 d < conf->raid_disks - mddev->delta_disks;
7621 struct md_rdev *rdev = conf->disks[d].rdev;
7623 clear_bit(In_sync, &rdev->flags);
7624 rdev = conf->disks[d].replacement;
7626 clear_bit(In_sync, &rdev->flags);
7629 mddev->layout = conf->algorithm;
7630 mddev->chunk_sectors = conf->chunk_sectors;
7631 mddev->reshape_position = MaxSector;
7632 mddev->delta_disks = 0;
7633 mddev->reshape_backwards = 0;
7637 static void raid5_quiesce(struct mddev *mddev, int state)
7639 struct r5conf *conf = mddev->private;
7642 case 2: /* resume for a suspend */
7643 wake_up(&conf->wait_for_overlap);
7646 case 1: /* stop all writes */
7647 lock_all_device_hash_locks_irq(conf);
7648 /* '2' tells resync/reshape to pause so that all
7649 * active stripes can drain
7652 wait_event_cmd(conf->wait_for_quiescent,
7653 atomic_read(&conf->active_stripes) == 0 &&
7654 atomic_read(&conf->active_aligned_reads) == 0,
7655 unlock_all_device_hash_locks_irq(conf),
7656 lock_all_device_hash_locks_irq(conf));
7658 unlock_all_device_hash_locks_irq(conf);
7659 /* allow reshape to continue */
7660 wake_up(&conf->wait_for_overlap);
7663 case 0: /* re-enable writes */
7664 lock_all_device_hash_locks_irq(conf);
7666 wake_up(&conf->wait_for_quiescent);
7667 wake_up(&conf->wait_for_overlap);
7668 unlock_all_device_hash_locks_irq(conf);
7671 r5l_quiesce(conf->log, state);
7674 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7676 struct r0conf *raid0_conf = mddev->private;
7679 /* for raid0 takeover only one zone is supported */
7680 if (raid0_conf->nr_strip_zones > 1) {
7681 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7683 return ERR_PTR(-EINVAL);
7686 sectors = raid0_conf->strip_zone[0].zone_end;
7687 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7688 mddev->dev_sectors = sectors;
7689 mddev->new_level = level;
7690 mddev->new_layout = ALGORITHM_PARITY_N;
7691 mddev->new_chunk_sectors = mddev->chunk_sectors;
7692 mddev->raid_disks += 1;
7693 mddev->delta_disks = 1;
7694 /* make sure it will be not marked as dirty */
7695 mddev->recovery_cp = MaxSector;
7697 return setup_conf(mddev);
7700 static void *raid5_takeover_raid1(struct mddev *mddev)
7704 if (mddev->raid_disks != 2 ||
7705 mddev->degraded > 1)
7706 return ERR_PTR(-EINVAL);
7708 /* Should check if there are write-behind devices? */
7710 chunksect = 64*2; /* 64K by default */
7712 /* The array must be an exact multiple of chunksize */
7713 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7716 if ((chunksect<<9) < STRIPE_SIZE)
7717 /* array size does not allow a suitable chunk size */
7718 return ERR_PTR(-EINVAL);
7720 mddev->new_level = 5;
7721 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7722 mddev->new_chunk_sectors = chunksect;
7724 return setup_conf(mddev);
7727 static void *raid5_takeover_raid6(struct mddev *mddev)
7731 switch (mddev->layout) {
7732 case ALGORITHM_LEFT_ASYMMETRIC_6:
7733 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7735 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7736 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7738 case ALGORITHM_LEFT_SYMMETRIC_6:
7739 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7741 case ALGORITHM_RIGHT_SYMMETRIC_6:
7742 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7744 case ALGORITHM_PARITY_0_6:
7745 new_layout = ALGORITHM_PARITY_0;
7747 case ALGORITHM_PARITY_N:
7748 new_layout = ALGORITHM_PARITY_N;
7751 return ERR_PTR(-EINVAL);
7753 mddev->new_level = 5;
7754 mddev->new_layout = new_layout;
7755 mddev->delta_disks = -1;
7756 mddev->raid_disks -= 1;
7757 return setup_conf(mddev);
7760 static int raid5_check_reshape(struct mddev *mddev)
7762 /* For a 2-drive array, the layout and chunk size can be changed
7763 * immediately as not restriping is needed.
7764 * For larger arrays we record the new value - after validation
7765 * to be used by a reshape pass.
7767 struct r5conf *conf = mddev->private;
7768 int new_chunk = mddev->new_chunk_sectors;
7770 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7772 if (new_chunk > 0) {
7773 if (!is_power_of_2(new_chunk))
7775 if (new_chunk < (PAGE_SIZE>>9))
7777 if (mddev->array_sectors & (new_chunk-1))
7778 /* not factor of array size */
7782 /* They look valid */
7784 if (mddev->raid_disks == 2) {
7785 /* can make the change immediately */
7786 if (mddev->new_layout >= 0) {
7787 conf->algorithm = mddev->new_layout;
7788 mddev->layout = mddev->new_layout;
7790 if (new_chunk > 0) {
7791 conf->chunk_sectors = new_chunk ;
7792 mddev->chunk_sectors = new_chunk;
7794 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7795 md_wakeup_thread(mddev->thread);
7797 return check_reshape(mddev);
7800 static int raid6_check_reshape(struct mddev *mddev)
7802 int new_chunk = mddev->new_chunk_sectors;
7804 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7806 if (new_chunk > 0) {
7807 if (!is_power_of_2(new_chunk))
7809 if (new_chunk < (PAGE_SIZE >> 9))
7811 if (mddev->array_sectors & (new_chunk-1))
7812 /* not factor of array size */
7816 /* They look valid */
7817 return check_reshape(mddev);
7820 static void *raid5_takeover(struct mddev *mddev)
7822 /* raid5 can take over:
7823 * raid0 - if there is only one strip zone - make it a raid4 layout
7824 * raid1 - if there are two drives. We need to know the chunk size
7825 * raid4 - trivial - just use a raid4 layout.
7826 * raid6 - Providing it is a *_6 layout
7828 if (mddev->level == 0)
7829 return raid45_takeover_raid0(mddev, 5);
7830 if (mddev->level == 1)
7831 return raid5_takeover_raid1(mddev);
7832 if (mddev->level == 4) {
7833 mddev->new_layout = ALGORITHM_PARITY_N;
7834 mddev->new_level = 5;
7835 return setup_conf(mddev);
7837 if (mddev->level == 6)
7838 return raid5_takeover_raid6(mddev);
7840 return ERR_PTR(-EINVAL);
7843 static void *raid4_takeover(struct mddev *mddev)
7845 /* raid4 can take over:
7846 * raid0 - if there is only one strip zone
7847 * raid5 - if layout is right
7849 if (mddev->level == 0)
7850 return raid45_takeover_raid0(mddev, 4);
7851 if (mddev->level == 5 &&
7852 mddev->layout == ALGORITHM_PARITY_N) {
7853 mddev->new_layout = 0;
7854 mddev->new_level = 4;
7855 return setup_conf(mddev);
7857 return ERR_PTR(-EINVAL);
7860 static struct md_personality raid5_personality;
7862 static void *raid6_takeover(struct mddev *mddev)
7864 /* Currently can only take over a raid5. We map the
7865 * personality to an equivalent raid6 personality
7866 * with the Q block at the end.
7870 if (mddev->pers != &raid5_personality)
7871 return ERR_PTR(-EINVAL);
7872 if (mddev->degraded > 1)
7873 return ERR_PTR(-EINVAL);
7874 if (mddev->raid_disks > 253)
7875 return ERR_PTR(-EINVAL);
7876 if (mddev->raid_disks < 3)
7877 return ERR_PTR(-EINVAL);
7879 switch (mddev->layout) {
7880 case ALGORITHM_LEFT_ASYMMETRIC:
7881 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7883 case ALGORITHM_RIGHT_ASYMMETRIC:
7884 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7886 case ALGORITHM_LEFT_SYMMETRIC:
7887 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7889 case ALGORITHM_RIGHT_SYMMETRIC:
7890 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7892 case ALGORITHM_PARITY_0:
7893 new_layout = ALGORITHM_PARITY_0_6;
7895 case ALGORITHM_PARITY_N:
7896 new_layout = ALGORITHM_PARITY_N;
7899 return ERR_PTR(-EINVAL);
7901 mddev->new_level = 6;
7902 mddev->new_layout = new_layout;
7903 mddev->delta_disks = 1;
7904 mddev->raid_disks += 1;
7905 return setup_conf(mddev);
7908 static struct md_personality raid6_personality =
7912 .owner = THIS_MODULE,
7913 .make_request = raid5_make_request,
7916 .status = raid5_status,
7917 .error_handler = raid5_error,
7918 .hot_add_disk = raid5_add_disk,
7919 .hot_remove_disk= raid5_remove_disk,
7920 .spare_active = raid5_spare_active,
7921 .sync_request = raid5_sync_request,
7922 .resize = raid5_resize,
7924 .check_reshape = raid6_check_reshape,
7925 .start_reshape = raid5_start_reshape,
7926 .finish_reshape = raid5_finish_reshape,
7927 .quiesce = raid5_quiesce,
7928 .takeover = raid6_takeover,
7929 .congested = raid5_congested,
7931 static struct md_personality raid5_personality =
7935 .owner = THIS_MODULE,
7936 .make_request = raid5_make_request,
7939 .status = raid5_status,
7940 .error_handler = raid5_error,
7941 .hot_add_disk = raid5_add_disk,
7942 .hot_remove_disk= raid5_remove_disk,
7943 .spare_active = raid5_spare_active,
7944 .sync_request = raid5_sync_request,
7945 .resize = raid5_resize,
7947 .check_reshape = raid5_check_reshape,
7948 .start_reshape = raid5_start_reshape,
7949 .finish_reshape = raid5_finish_reshape,
7950 .quiesce = raid5_quiesce,
7951 .takeover = raid5_takeover,
7952 .congested = raid5_congested,
7955 static struct md_personality raid4_personality =
7959 .owner = THIS_MODULE,
7960 .make_request = raid5_make_request,
7963 .status = raid5_status,
7964 .error_handler = raid5_error,
7965 .hot_add_disk = raid5_add_disk,
7966 .hot_remove_disk= raid5_remove_disk,
7967 .spare_active = raid5_spare_active,
7968 .sync_request = raid5_sync_request,
7969 .resize = raid5_resize,
7971 .check_reshape = raid5_check_reshape,
7972 .start_reshape = raid5_start_reshape,
7973 .finish_reshape = raid5_finish_reshape,
7974 .quiesce = raid5_quiesce,
7975 .takeover = raid4_takeover,
7976 .congested = raid5_congested,
7979 static int __init raid5_init(void)
7981 raid5_wq = alloc_workqueue("raid5wq",
7982 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7985 register_md_personality(&raid6_personality);
7986 register_md_personality(&raid5_personality);
7987 register_md_personality(&raid4_personality);
7991 static void raid5_exit(void)
7993 unregister_md_personality(&raid6_personality);
7994 unregister_md_personality(&raid5_personality);
7995 unregister_md_personality(&raid4_personality);
7996 destroy_workqueue(raid5_wq);
7999 module_init(raid5_init);
8000 module_exit(raid5_exit);
8001 MODULE_LICENSE("GPL");
8002 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8003 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8004 MODULE_ALIAS("md-raid5");
8005 MODULE_ALIAS("md-raid4");
8006 MODULE_ALIAS("md-level-5");
8007 MODULE_ALIAS("md-level-4");
8008 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8009 MODULE_ALIAS("md-raid6");
8010 MODULE_ALIAS("md-level-6");
8012 /* This used to be two separate modules, they were: */
8013 MODULE_ALIAS("raid5");
8014 MODULE_ALIAS("raid6");