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);
1009 bi->bi_bdev = rdev->bdev;
1010 bio_set_op_attrs(bi, op, op_flags);
1011 bi->bi_end_io = op_is_write(op)
1012 ? raid5_end_write_request
1013 : raid5_end_read_request;
1014 bi->bi_private = sh;
1016 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1017 __func__, (unsigned long long)sh->sector,
1019 atomic_inc(&sh->count);
1021 atomic_inc(&head_sh->count);
1022 if (use_new_offset(conf, sh))
1023 bi->bi_iter.bi_sector = (sh->sector
1024 + rdev->new_data_offset);
1026 bi->bi_iter.bi_sector = (sh->sector
1027 + rdev->data_offset);
1028 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1029 bi->bi_rw |= REQ_NOMERGE;
1031 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1032 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1033 sh->dev[i].vec.bv_page = sh->dev[i].page;
1035 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1036 bi->bi_io_vec[0].bv_offset = 0;
1037 bi->bi_iter.bi_size = STRIPE_SIZE;
1039 * If this is discard request, set bi_vcnt 0. We don't
1040 * want to confuse SCSI because SCSI will replace payload
1042 if (op == REQ_OP_DISCARD)
1045 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1047 if (conf->mddev->gendisk)
1048 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1049 bi, disk_devt(conf->mddev->gendisk),
1051 generic_make_request(bi);
1054 if (s->syncing || s->expanding || s->expanded
1056 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1058 set_bit(STRIPE_IO_STARTED, &sh->state);
1061 rbi->bi_bdev = rrdev->bdev;
1062 bio_set_op_attrs(rbi, op, op_flags);
1063 BUG_ON(!op_is_write(op));
1064 rbi->bi_end_io = raid5_end_write_request;
1065 rbi->bi_private = sh;
1067 pr_debug("%s: for %llu schedule op %d on "
1068 "replacement disc %d\n",
1069 __func__, (unsigned long long)sh->sector,
1071 atomic_inc(&sh->count);
1073 atomic_inc(&head_sh->count);
1074 if (use_new_offset(conf, sh))
1075 rbi->bi_iter.bi_sector = (sh->sector
1076 + rrdev->new_data_offset);
1078 rbi->bi_iter.bi_sector = (sh->sector
1079 + rrdev->data_offset);
1080 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1081 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1082 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1084 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1085 rbi->bi_io_vec[0].bv_offset = 0;
1086 rbi->bi_iter.bi_size = STRIPE_SIZE;
1088 * If this is discard request, set bi_vcnt 0. We don't
1089 * want to confuse SCSI because SCSI will replace payload
1091 if (op == REQ_OP_DISCARD)
1093 if (conf->mddev->gendisk)
1094 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1095 rbi, disk_devt(conf->mddev->gendisk),
1097 generic_make_request(rbi);
1099 if (!rdev && !rrdev) {
1100 if (op_is_write(op))
1101 set_bit(STRIPE_DEGRADED, &sh->state);
1102 pr_debug("skip op %d on disc %d for sector %llu\n",
1103 bi->bi_rw, i, (unsigned long long)sh->sector);
1104 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1105 set_bit(STRIPE_HANDLE, &sh->state);
1108 if (!head_sh->batch_head)
1110 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1117 static struct dma_async_tx_descriptor *
1118 async_copy_data(int frombio, struct bio *bio, struct page **page,
1119 sector_t sector, struct dma_async_tx_descriptor *tx,
1120 struct stripe_head *sh)
1123 struct bvec_iter iter;
1124 struct page *bio_page;
1126 struct async_submit_ctl submit;
1127 enum async_tx_flags flags = 0;
1129 if (bio->bi_iter.bi_sector >= sector)
1130 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1132 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1135 flags |= ASYNC_TX_FENCE;
1136 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1138 bio_for_each_segment(bvl, bio, iter) {
1139 int len = bvl.bv_len;
1143 if (page_offset < 0) {
1144 b_offset = -page_offset;
1145 page_offset += b_offset;
1149 if (len > 0 && page_offset + len > STRIPE_SIZE)
1150 clen = STRIPE_SIZE - page_offset;
1155 b_offset += bvl.bv_offset;
1156 bio_page = bvl.bv_page;
1158 if (sh->raid_conf->skip_copy &&
1159 b_offset == 0 && page_offset == 0 &&
1160 clen == STRIPE_SIZE)
1163 tx = async_memcpy(*page, bio_page, page_offset,
1164 b_offset, clen, &submit);
1166 tx = async_memcpy(bio_page, *page, b_offset,
1167 page_offset, clen, &submit);
1169 /* chain the operations */
1170 submit.depend_tx = tx;
1172 if (clen < len) /* hit end of page */
1180 static void ops_complete_biofill(void *stripe_head_ref)
1182 struct stripe_head *sh = stripe_head_ref;
1183 struct bio_list return_bi = BIO_EMPTY_LIST;
1186 pr_debug("%s: stripe %llu\n", __func__,
1187 (unsigned long long)sh->sector);
1189 /* clear completed biofills */
1190 for (i = sh->disks; i--; ) {
1191 struct r5dev *dev = &sh->dev[i];
1193 /* acknowledge completion of a biofill operation */
1194 /* and check if we need to reply to a read request,
1195 * new R5_Wantfill requests are held off until
1196 * !STRIPE_BIOFILL_RUN
1198 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1199 struct bio *rbi, *rbi2;
1204 while (rbi && rbi->bi_iter.bi_sector <
1205 dev->sector + STRIPE_SECTORS) {
1206 rbi2 = r5_next_bio(rbi, dev->sector);
1207 if (!raid5_dec_bi_active_stripes(rbi))
1208 bio_list_add(&return_bi, rbi);
1213 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1215 return_io(&return_bi);
1217 set_bit(STRIPE_HANDLE, &sh->state);
1218 raid5_release_stripe(sh);
1221 static void ops_run_biofill(struct stripe_head *sh)
1223 struct dma_async_tx_descriptor *tx = NULL;
1224 struct async_submit_ctl submit;
1227 BUG_ON(sh->batch_head);
1228 pr_debug("%s: stripe %llu\n", __func__,
1229 (unsigned long long)sh->sector);
1231 for (i = sh->disks; i--; ) {
1232 struct r5dev *dev = &sh->dev[i];
1233 if (test_bit(R5_Wantfill, &dev->flags)) {
1235 spin_lock_irq(&sh->stripe_lock);
1236 dev->read = rbi = dev->toread;
1238 spin_unlock_irq(&sh->stripe_lock);
1239 while (rbi && rbi->bi_iter.bi_sector <
1240 dev->sector + STRIPE_SECTORS) {
1241 tx = async_copy_data(0, rbi, &dev->page,
1242 dev->sector, tx, sh);
1243 rbi = r5_next_bio(rbi, dev->sector);
1248 atomic_inc(&sh->count);
1249 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1250 async_trigger_callback(&submit);
1253 static void mark_target_uptodate(struct stripe_head *sh, int target)
1260 tgt = &sh->dev[target];
1261 set_bit(R5_UPTODATE, &tgt->flags);
1262 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1263 clear_bit(R5_Wantcompute, &tgt->flags);
1266 static void ops_complete_compute(void *stripe_head_ref)
1268 struct stripe_head *sh = stripe_head_ref;
1270 pr_debug("%s: stripe %llu\n", __func__,
1271 (unsigned long long)sh->sector);
1273 /* mark the computed target(s) as uptodate */
1274 mark_target_uptodate(sh, sh->ops.target);
1275 mark_target_uptodate(sh, sh->ops.target2);
1277 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1278 if (sh->check_state == check_state_compute_run)
1279 sh->check_state = check_state_compute_result;
1280 set_bit(STRIPE_HANDLE, &sh->state);
1281 raid5_release_stripe(sh);
1284 /* return a pointer to the address conversion region of the scribble buffer */
1285 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1286 struct raid5_percpu *percpu, int i)
1290 addr = flex_array_get(percpu->scribble, i);
1291 return addr + sizeof(struct page *) * (sh->disks + 2);
1294 /* return a pointer to the address conversion region of the scribble buffer */
1295 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1299 addr = flex_array_get(percpu->scribble, i);
1303 static struct dma_async_tx_descriptor *
1304 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1306 int disks = sh->disks;
1307 struct page **xor_srcs = to_addr_page(percpu, 0);
1308 int target = sh->ops.target;
1309 struct r5dev *tgt = &sh->dev[target];
1310 struct page *xor_dest = tgt->page;
1312 struct dma_async_tx_descriptor *tx;
1313 struct async_submit_ctl submit;
1316 BUG_ON(sh->batch_head);
1318 pr_debug("%s: stripe %llu block: %d\n",
1319 __func__, (unsigned long long)sh->sector, target);
1320 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1322 for (i = disks; i--; )
1324 xor_srcs[count++] = sh->dev[i].page;
1326 atomic_inc(&sh->count);
1328 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1329 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1330 if (unlikely(count == 1))
1331 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1333 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1338 /* set_syndrome_sources - populate source buffers for gen_syndrome
1339 * @srcs - (struct page *) array of size sh->disks
1340 * @sh - stripe_head to parse
1342 * Populates srcs in proper layout order for the stripe and returns the
1343 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1344 * destination buffer is recorded in srcs[count] and the Q destination
1345 * is recorded in srcs[count+1]].
1347 static int set_syndrome_sources(struct page **srcs,
1348 struct stripe_head *sh,
1351 int disks = sh->disks;
1352 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1353 int d0_idx = raid6_d0(sh);
1357 for (i = 0; i < disks; i++)
1363 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1364 struct r5dev *dev = &sh->dev[i];
1366 if (i == sh->qd_idx || i == sh->pd_idx ||
1367 (srctype == SYNDROME_SRC_ALL) ||
1368 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1369 test_bit(R5_Wantdrain, &dev->flags)) ||
1370 (srctype == SYNDROME_SRC_WRITTEN &&
1372 srcs[slot] = sh->dev[i].page;
1373 i = raid6_next_disk(i, disks);
1374 } while (i != d0_idx);
1376 return syndrome_disks;
1379 static struct dma_async_tx_descriptor *
1380 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1382 int disks = sh->disks;
1383 struct page **blocks = to_addr_page(percpu, 0);
1385 int qd_idx = sh->qd_idx;
1386 struct dma_async_tx_descriptor *tx;
1387 struct async_submit_ctl submit;
1393 BUG_ON(sh->batch_head);
1394 if (sh->ops.target < 0)
1395 target = sh->ops.target2;
1396 else if (sh->ops.target2 < 0)
1397 target = sh->ops.target;
1399 /* we should only have one valid target */
1402 pr_debug("%s: stripe %llu block: %d\n",
1403 __func__, (unsigned long long)sh->sector, target);
1405 tgt = &sh->dev[target];
1406 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1409 atomic_inc(&sh->count);
1411 if (target == qd_idx) {
1412 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1413 blocks[count] = NULL; /* regenerating p is not necessary */
1414 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1415 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1416 ops_complete_compute, sh,
1417 to_addr_conv(sh, percpu, 0));
1418 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1420 /* Compute any data- or p-drive using XOR */
1422 for (i = disks; i-- ; ) {
1423 if (i == target || i == qd_idx)
1425 blocks[count++] = sh->dev[i].page;
1428 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1429 NULL, ops_complete_compute, sh,
1430 to_addr_conv(sh, percpu, 0));
1431 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1437 static struct dma_async_tx_descriptor *
1438 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1440 int i, count, disks = sh->disks;
1441 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1442 int d0_idx = raid6_d0(sh);
1443 int faila = -1, failb = -1;
1444 int target = sh->ops.target;
1445 int target2 = sh->ops.target2;
1446 struct r5dev *tgt = &sh->dev[target];
1447 struct r5dev *tgt2 = &sh->dev[target2];
1448 struct dma_async_tx_descriptor *tx;
1449 struct page **blocks = to_addr_page(percpu, 0);
1450 struct async_submit_ctl submit;
1452 BUG_ON(sh->batch_head);
1453 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1454 __func__, (unsigned long long)sh->sector, target, target2);
1455 BUG_ON(target < 0 || target2 < 0);
1456 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1457 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1459 /* we need to open-code set_syndrome_sources to handle the
1460 * slot number conversion for 'faila' and 'failb'
1462 for (i = 0; i < disks ; i++)
1467 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1469 blocks[slot] = sh->dev[i].page;
1475 i = raid6_next_disk(i, disks);
1476 } while (i != d0_idx);
1478 BUG_ON(faila == failb);
1481 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1482 __func__, (unsigned long long)sh->sector, faila, failb);
1484 atomic_inc(&sh->count);
1486 if (failb == syndrome_disks+1) {
1487 /* Q disk is one of the missing disks */
1488 if (faila == syndrome_disks) {
1489 /* Missing P+Q, just recompute */
1490 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1491 ops_complete_compute, sh,
1492 to_addr_conv(sh, percpu, 0));
1493 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1494 STRIPE_SIZE, &submit);
1498 int qd_idx = sh->qd_idx;
1500 /* Missing D+Q: recompute D from P, then recompute Q */
1501 if (target == qd_idx)
1502 data_target = target2;
1504 data_target = target;
1507 for (i = disks; i-- ; ) {
1508 if (i == data_target || i == qd_idx)
1510 blocks[count++] = sh->dev[i].page;
1512 dest = sh->dev[data_target].page;
1513 init_async_submit(&submit,
1514 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1516 to_addr_conv(sh, percpu, 0));
1517 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1520 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1521 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1522 ops_complete_compute, sh,
1523 to_addr_conv(sh, percpu, 0));
1524 return async_gen_syndrome(blocks, 0, count+2,
1525 STRIPE_SIZE, &submit);
1528 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1529 ops_complete_compute, sh,
1530 to_addr_conv(sh, percpu, 0));
1531 if (failb == syndrome_disks) {
1532 /* We're missing D+P. */
1533 return async_raid6_datap_recov(syndrome_disks+2,
1537 /* We're missing D+D. */
1538 return async_raid6_2data_recov(syndrome_disks+2,
1539 STRIPE_SIZE, faila, failb,
1545 static void ops_complete_prexor(void *stripe_head_ref)
1547 struct stripe_head *sh = stripe_head_ref;
1549 pr_debug("%s: stripe %llu\n", __func__,
1550 (unsigned long long)sh->sector);
1553 static struct dma_async_tx_descriptor *
1554 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1555 struct dma_async_tx_descriptor *tx)
1557 int disks = sh->disks;
1558 struct page **xor_srcs = to_addr_page(percpu, 0);
1559 int count = 0, pd_idx = sh->pd_idx, i;
1560 struct async_submit_ctl submit;
1562 /* existing parity data subtracted */
1563 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1565 BUG_ON(sh->batch_head);
1566 pr_debug("%s: stripe %llu\n", __func__,
1567 (unsigned long long)sh->sector);
1569 for (i = disks; i--; ) {
1570 struct r5dev *dev = &sh->dev[i];
1571 /* Only process blocks that are known to be uptodate */
1572 if (test_bit(R5_Wantdrain, &dev->flags))
1573 xor_srcs[count++] = dev->page;
1576 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1577 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1578 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1583 static struct dma_async_tx_descriptor *
1584 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1585 struct dma_async_tx_descriptor *tx)
1587 struct page **blocks = to_addr_page(percpu, 0);
1589 struct async_submit_ctl submit;
1591 pr_debug("%s: stripe %llu\n", __func__,
1592 (unsigned long long)sh->sector);
1594 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1596 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1597 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1598 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1603 static struct dma_async_tx_descriptor *
1604 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1606 int disks = sh->disks;
1608 struct stripe_head *head_sh = sh;
1610 pr_debug("%s: stripe %llu\n", __func__,
1611 (unsigned long long)sh->sector);
1613 for (i = disks; i--; ) {
1618 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1623 spin_lock_irq(&sh->stripe_lock);
1624 chosen = dev->towrite;
1625 dev->towrite = NULL;
1626 sh->overwrite_disks = 0;
1627 BUG_ON(dev->written);
1628 wbi = dev->written = chosen;
1629 spin_unlock_irq(&sh->stripe_lock);
1630 WARN_ON(dev->page != dev->orig_page);
1632 while (wbi && wbi->bi_iter.bi_sector <
1633 dev->sector + STRIPE_SECTORS) {
1634 if (wbi->bi_rw & REQ_FUA)
1635 set_bit(R5_WantFUA, &dev->flags);
1636 if (wbi->bi_rw & REQ_SYNC)
1637 set_bit(R5_SyncIO, &dev->flags);
1638 if (bio_op(wbi) == REQ_OP_DISCARD)
1639 set_bit(R5_Discard, &dev->flags);
1641 tx = async_copy_data(1, wbi, &dev->page,
1642 dev->sector, tx, sh);
1643 if (dev->page != dev->orig_page) {
1644 set_bit(R5_SkipCopy, &dev->flags);
1645 clear_bit(R5_UPTODATE, &dev->flags);
1646 clear_bit(R5_OVERWRITE, &dev->flags);
1649 wbi = r5_next_bio(wbi, dev->sector);
1652 if (head_sh->batch_head) {
1653 sh = list_first_entry(&sh->batch_list,
1666 static void ops_complete_reconstruct(void *stripe_head_ref)
1668 struct stripe_head *sh = stripe_head_ref;
1669 int disks = sh->disks;
1670 int pd_idx = sh->pd_idx;
1671 int qd_idx = sh->qd_idx;
1673 bool fua = false, sync = false, discard = false;
1675 pr_debug("%s: stripe %llu\n", __func__,
1676 (unsigned long long)sh->sector);
1678 for (i = disks; i--; ) {
1679 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1680 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1681 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1684 for (i = disks; i--; ) {
1685 struct r5dev *dev = &sh->dev[i];
1687 if (dev->written || i == pd_idx || i == qd_idx) {
1688 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1689 set_bit(R5_UPTODATE, &dev->flags);
1691 set_bit(R5_WantFUA, &dev->flags);
1693 set_bit(R5_SyncIO, &dev->flags);
1697 if (sh->reconstruct_state == reconstruct_state_drain_run)
1698 sh->reconstruct_state = reconstruct_state_drain_result;
1699 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1700 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1702 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1703 sh->reconstruct_state = reconstruct_state_result;
1706 set_bit(STRIPE_HANDLE, &sh->state);
1707 raid5_release_stripe(sh);
1711 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1712 struct dma_async_tx_descriptor *tx)
1714 int disks = sh->disks;
1715 struct page **xor_srcs;
1716 struct async_submit_ctl submit;
1717 int count, pd_idx = sh->pd_idx, i;
1718 struct page *xor_dest;
1720 unsigned long flags;
1722 struct stripe_head *head_sh = sh;
1725 pr_debug("%s: stripe %llu\n", __func__,
1726 (unsigned long long)sh->sector);
1728 for (i = 0; i < sh->disks; i++) {
1731 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1734 if (i >= sh->disks) {
1735 atomic_inc(&sh->count);
1736 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1737 ops_complete_reconstruct(sh);
1742 xor_srcs = to_addr_page(percpu, j);
1743 /* check if prexor is active which means only process blocks
1744 * that are part of a read-modify-write (written)
1746 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1748 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1749 for (i = disks; i--; ) {
1750 struct r5dev *dev = &sh->dev[i];
1751 if (head_sh->dev[i].written)
1752 xor_srcs[count++] = dev->page;
1755 xor_dest = sh->dev[pd_idx].page;
1756 for (i = disks; i--; ) {
1757 struct r5dev *dev = &sh->dev[i];
1759 xor_srcs[count++] = dev->page;
1763 /* 1/ if we prexor'd then the dest is reused as a source
1764 * 2/ if we did not prexor then we are redoing the parity
1765 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1766 * for the synchronous xor case
1768 last_stripe = !head_sh->batch_head ||
1769 list_first_entry(&sh->batch_list,
1770 struct stripe_head, batch_list) == head_sh;
1772 flags = ASYNC_TX_ACK |
1773 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1775 atomic_inc(&head_sh->count);
1776 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1777 to_addr_conv(sh, percpu, j));
1779 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1780 init_async_submit(&submit, flags, tx, NULL, NULL,
1781 to_addr_conv(sh, percpu, j));
1784 if (unlikely(count == 1))
1785 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1787 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1790 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1797 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1798 struct dma_async_tx_descriptor *tx)
1800 struct async_submit_ctl submit;
1801 struct page **blocks;
1802 int count, i, j = 0;
1803 struct stripe_head *head_sh = sh;
1806 unsigned long txflags;
1808 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1810 for (i = 0; i < sh->disks; i++) {
1811 if (sh->pd_idx == i || sh->qd_idx == i)
1813 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1816 if (i >= sh->disks) {
1817 atomic_inc(&sh->count);
1818 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1819 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1820 ops_complete_reconstruct(sh);
1825 blocks = to_addr_page(percpu, j);
1827 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1828 synflags = SYNDROME_SRC_WRITTEN;
1829 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1831 synflags = SYNDROME_SRC_ALL;
1832 txflags = ASYNC_TX_ACK;
1835 count = set_syndrome_sources(blocks, sh, synflags);
1836 last_stripe = !head_sh->batch_head ||
1837 list_first_entry(&sh->batch_list,
1838 struct stripe_head, batch_list) == head_sh;
1841 atomic_inc(&head_sh->count);
1842 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1843 head_sh, to_addr_conv(sh, percpu, j));
1845 init_async_submit(&submit, 0, tx, NULL, NULL,
1846 to_addr_conv(sh, percpu, j));
1847 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1850 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1856 static void ops_complete_check(void *stripe_head_ref)
1858 struct stripe_head *sh = stripe_head_ref;
1860 pr_debug("%s: stripe %llu\n", __func__,
1861 (unsigned long long)sh->sector);
1863 sh->check_state = check_state_check_result;
1864 set_bit(STRIPE_HANDLE, &sh->state);
1865 raid5_release_stripe(sh);
1868 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1870 int disks = sh->disks;
1871 int pd_idx = sh->pd_idx;
1872 int qd_idx = sh->qd_idx;
1873 struct page *xor_dest;
1874 struct page **xor_srcs = to_addr_page(percpu, 0);
1875 struct dma_async_tx_descriptor *tx;
1876 struct async_submit_ctl submit;
1880 pr_debug("%s: stripe %llu\n", __func__,
1881 (unsigned long long)sh->sector);
1883 BUG_ON(sh->batch_head);
1885 xor_dest = sh->dev[pd_idx].page;
1886 xor_srcs[count++] = xor_dest;
1887 for (i = disks; i--; ) {
1888 if (i == pd_idx || i == qd_idx)
1890 xor_srcs[count++] = sh->dev[i].page;
1893 init_async_submit(&submit, 0, NULL, NULL, NULL,
1894 to_addr_conv(sh, percpu, 0));
1895 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1896 &sh->ops.zero_sum_result, &submit);
1898 atomic_inc(&sh->count);
1899 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1900 tx = async_trigger_callback(&submit);
1903 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1905 struct page **srcs = to_addr_page(percpu, 0);
1906 struct async_submit_ctl submit;
1909 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1910 (unsigned long long)sh->sector, checkp);
1912 BUG_ON(sh->batch_head);
1913 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1917 atomic_inc(&sh->count);
1918 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1919 sh, to_addr_conv(sh, percpu, 0));
1920 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1921 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1924 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1926 int overlap_clear = 0, i, disks = sh->disks;
1927 struct dma_async_tx_descriptor *tx = NULL;
1928 struct r5conf *conf = sh->raid_conf;
1929 int level = conf->level;
1930 struct raid5_percpu *percpu;
1934 percpu = per_cpu_ptr(conf->percpu, cpu);
1935 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1936 ops_run_biofill(sh);
1940 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1942 tx = ops_run_compute5(sh, percpu);
1944 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1945 tx = ops_run_compute6_1(sh, percpu);
1947 tx = ops_run_compute6_2(sh, percpu);
1949 /* terminate the chain if reconstruct is not set to be run */
1950 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1954 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1956 tx = ops_run_prexor5(sh, percpu, tx);
1958 tx = ops_run_prexor6(sh, percpu, tx);
1961 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1962 tx = ops_run_biodrain(sh, tx);
1966 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1968 ops_run_reconstruct5(sh, percpu, tx);
1970 ops_run_reconstruct6(sh, percpu, tx);
1973 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1974 if (sh->check_state == check_state_run)
1975 ops_run_check_p(sh, percpu);
1976 else if (sh->check_state == check_state_run_q)
1977 ops_run_check_pq(sh, percpu, 0);
1978 else if (sh->check_state == check_state_run_pq)
1979 ops_run_check_pq(sh, percpu, 1);
1984 if (overlap_clear && !sh->batch_head)
1985 for (i = disks; i--; ) {
1986 struct r5dev *dev = &sh->dev[i];
1987 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1988 wake_up(&sh->raid_conf->wait_for_overlap);
1993 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1995 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);
2007 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2009 struct stripe_head *sh;
2011 sh = alloc_stripe(conf->slab_cache, gfp);
2015 sh->raid_conf = conf;
2017 if (grow_buffers(sh, gfp)) {
2019 kmem_cache_free(conf->slab_cache, sh);
2022 sh->hash_lock_index =
2023 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2024 /* we just created an active stripe so... */
2025 atomic_inc(&conf->active_stripes);
2027 raid5_release_stripe(sh);
2028 conf->max_nr_stripes++;
2032 static int grow_stripes(struct r5conf *conf, int num)
2034 struct kmem_cache *sc;
2035 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2037 if (conf->mddev->gendisk)
2038 sprintf(conf->cache_name[0],
2039 "raid%d-%s", conf->level, mdname(conf->mddev));
2041 sprintf(conf->cache_name[0],
2042 "raid%d-%p", conf->level, conf->mddev);
2043 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2045 conf->active_name = 0;
2046 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2047 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2051 conf->slab_cache = sc;
2052 conf->pool_size = devs;
2054 if (!grow_one_stripe(conf, GFP_KERNEL))
2061 * scribble_len - return the required size of the scribble region
2062 * @num - total number of disks in the array
2064 * The size must be enough to contain:
2065 * 1/ a struct page pointer for each device in the array +2
2066 * 2/ room to convert each entry in (1) to its corresponding dma
2067 * (dma_map_page()) or page (page_address()) address.
2069 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2070 * calculate over all devices (not just the data blocks), using zeros in place
2071 * of the P and Q blocks.
2073 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2075 struct flex_array *ret;
2078 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2079 ret = flex_array_alloc(len, cnt, flags);
2082 /* always prealloc all elements, so no locking is required */
2083 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2084 flex_array_free(ret);
2090 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2096 * Never shrink. And mddev_suspend() could deadlock if this is called
2097 * from raid5d. In that case, scribble_disks and scribble_sectors
2098 * should equal to new_disks and new_sectors
2100 if (conf->scribble_disks >= new_disks &&
2101 conf->scribble_sectors >= new_sectors)
2103 mddev_suspend(conf->mddev);
2105 for_each_present_cpu(cpu) {
2106 struct raid5_percpu *percpu;
2107 struct flex_array *scribble;
2109 percpu = per_cpu_ptr(conf->percpu, cpu);
2110 scribble = scribble_alloc(new_disks,
2111 new_sectors / STRIPE_SECTORS,
2115 flex_array_free(percpu->scribble);
2116 percpu->scribble = scribble;
2123 mddev_resume(conf->mddev);
2125 conf->scribble_disks = new_disks;
2126 conf->scribble_sectors = new_sectors;
2131 static int resize_stripes(struct r5conf *conf, int newsize)
2133 /* Make all the stripes able to hold 'newsize' devices.
2134 * New slots in each stripe get 'page' set to a new page.
2136 * This happens in stages:
2137 * 1/ create a new kmem_cache and allocate the required number of
2139 * 2/ gather all the old stripe_heads and transfer the pages across
2140 * to the new stripe_heads. This will have the side effect of
2141 * freezing the array as once all stripe_heads have been collected,
2142 * no IO will be possible. Old stripe heads are freed once their
2143 * pages have been transferred over, and the old kmem_cache is
2144 * freed when all stripes are done.
2145 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2146 * we simple return a failre status - no need to clean anything up.
2147 * 4/ allocate new pages for the new slots in the new stripe_heads.
2148 * If this fails, we don't bother trying the shrink the
2149 * stripe_heads down again, we just leave them as they are.
2150 * As each stripe_head is processed the new one is released into
2153 * Once step2 is started, we cannot afford to wait for a write,
2154 * so we use GFP_NOIO allocations.
2156 struct stripe_head *osh, *nsh;
2157 LIST_HEAD(newstripes);
2158 struct disk_info *ndisks;
2160 struct kmem_cache *sc;
2164 if (newsize <= conf->pool_size)
2165 return 0; /* never bother to shrink */
2167 err = md_allow_write(conf->mddev);
2172 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2173 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2178 /* Need to ensure auto-resizing doesn't interfere */
2179 mutex_lock(&conf->cache_size_mutex);
2181 for (i = conf->max_nr_stripes; i; i--) {
2182 nsh = alloc_stripe(sc, GFP_KERNEL);
2186 nsh->raid_conf = conf;
2187 list_add(&nsh->lru, &newstripes);
2190 /* didn't get enough, give up */
2191 while (!list_empty(&newstripes)) {
2192 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2193 list_del(&nsh->lru);
2194 kmem_cache_free(sc, nsh);
2196 kmem_cache_destroy(sc);
2197 mutex_unlock(&conf->cache_size_mutex);
2200 /* Step 2 - Must use GFP_NOIO now.
2201 * OK, we have enough stripes, start collecting inactive
2202 * stripes and copying them over
2206 list_for_each_entry(nsh, &newstripes, lru) {
2207 lock_device_hash_lock(conf, hash);
2208 wait_event_cmd(conf->wait_for_stripe,
2209 !list_empty(conf->inactive_list + hash),
2210 unlock_device_hash_lock(conf, hash),
2211 lock_device_hash_lock(conf, hash));
2212 osh = get_free_stripe(conf, hash);
2213 unlock_device_hash_lock(conf, hash);
2215 for(i=0; i<conf->pool_size; i++) {
2216 nsh->dev[i].page = osh->dev[i].page;
2217 nsh->dev[i].orig_page = osh->dev[i].page;
2219 nsh->hash_lock_index = hash;
2220 kmem_cache_free(conf->slab_cache, osh);
2222 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2223 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2228 kmem_cache_destroy(conf->slab_cache);
2231 * At this point, we are holding all the stripes so the array
2232 * is completely stalled, so now is a good time to resize
2233 * conf->disks and the scribble region
2235 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2237 for (i=0; i<conf->raid_disks; i++)
2238 ndisks[i] = conf->disks[i];
2240 conf->disks = ndisks;
2244 mutex_unlock(&conf->cache_size_mutex);
2245 /* Step 4, return new stripes to service */
2246 while(!list_empty(&newstripes)) {
2247 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2248 list_del_init(&nsh->lru);
2250 for (i=conf->raid_disks; i < newsize; i++)
2251 if (nsh->dev[i].page == NULL) {
2252 struct page *p = alloc_page(GFP_NOIO);
2253 nsh->dev[i].page = p;
2254 nsh->dev[i].orig_page = p;
2258 raid5_release_stripe(nsh);
2260 /* critical section pass, GFP_NOIO no longer needed */
2262 conf->slab_cache = sc;
2263 conf->active_name = 1-conf->active_name;
2265 conf->pool_size = newsize;
2269 static int drop_one_stripe(struct r5conf *conf)
2271 struct stripe_head *sh;
2272 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2274 spin_lock_irq(conf->hash_locks + hash);
2275 sh = get_free_stripe(conf, hash);
2276 spin_unlock_irq(conf->hash_locks + hash);
2279 BUG_ON(atomic_read(&sh->count));
2281 kmem_cache_free(conf->slab_cache, sh);
2282 atomic_dec(&conf->active_stripes);
2283 conf->max_nr_stripes--;
2287 static void shrink_stripes(struct r5conf *conf)
2289 while (conf->max_nr_stripes &&
2290 drop_one_stripe(conf))
2293 kmem_cache_destroy(conf->slab_cache);
2294 conf->slab_cache = NULL;
2297 static void raid5_end_read_request(struct bio * bi)
2299 struct stripe_head *sh = bi->bi_private;
2300 struct r5conf *conf = sh->raid_conf;
2301 int disks = sh->disks, i;
2302 char b[BDEVNAME_SIZE];
2303 struct md_rdev *rdev = NULL;
2306 for (i=0 ; i<disks; i++)
2307 if (bi == &sh->dev[i].req)
2310 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2311 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2317 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2318 /* If replacement finished while this request was outstanding,
2319 * 'replacement' might be NULL already.
2320 * In that case it moved down to 'rdev'.
2321 * rdev is not removed until all requests are finished.
2323 rdev = conf->disks[i].replacement;
2325 rdev = conf->disks[i].rdev;
2327 if (use_new_offset(conf, sh))
2328 s = sh->sector + rdev->new_data_offset;
2330 s = sh->sector + rdev->data_offset;
2331 if (!bi->bi_error) {
2332 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2333 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2334 /* Note that this cannot happen on a
2335 * replacement device. We just fail those on
2340 "md/raid:%s: read error corrected"
2341 " (%lu sectors at %llu on %s)\n",
2342 mdname(conf->mddev), STRIPE_SECTORS,
2343 (unsigned long long)s,
2344 bdevname(rdev->bdev, b));
2345 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2346 clear_bit(R5_ReadError, &sh->dev[i].flags);
2347 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2348 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2349 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2351 if (atomic_read(&rdev->read_errors))
2352 atomic_set(&rdev->read_errors, 0);
2354 const char *bdn = bdevname(rdev->bdev, b);
2358 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2359 atomic_inc(&rdev->read_errors);
2360 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2363 "md/raid:%s: read error on replacement device "
2364 "(sector %llu on %s).\n",
2365 mdname(conf->mddev),
2366 (unsigned long long)s,
2368 else if (conf->mddev->degraded >= conf->max_degraded) {
2372 "md/raid:%s: read error not correctable "
2373 "(sector %llu on %s).\n",
2374 mdname(conf->mddev),
2375 (unsigned long long)s,
2377 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2382 "md/raid:%s: read error NOT corrected!! "
2383 "(sector %llu on %s).\n",
2384 mdname(conf->mddev),
2385 (unsigned long long)s,
2387 } else if (atomic_read(&rdev->read_errors)
2388 > conf->max_nr_stripes)
2390 "md/raid:%s: Too many read errors, failing device %s.\n",
2391 mdname(conf->mddev), bdn);
2394 if (set_bad && test_bit(In_sync, &rdev->flags)
2395 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2398 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2399 set_bit(R5_ReadError, &sh->dev[i].flags);
2400 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2402 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2404 clear_bit(R5_ReadError, &sh->dev[i].flags);
2405 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2407 && test_bit(In_sync, &rdev->flags)
2408 && rdev_set_badblocks(
2409 rdev, sh->sector, STRIPE_SECTORS, 0)))
2410 md_error(conf->mddev, rdev);
2413 rdev_dec_pending(rdev, conf->mddev);
2414 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2415 set_bit(STRIPE_HANDLE, &sh->state);
2416 raid5_release_stripe(sh);
2419 static void raid5_end_write_request(struct bio *bi)
2421 struct stripe_head *sh = bi->bi_private;
2422 struct r5conf *conf = sh->raid_conf;
2423 int disks = sh->disks, i;
2424 struct md_rdev *uninitialized_var(rdev);
2427 int replacement = 0;
2429 for (i = 0 ; i < disks; i++) {
2430 if (bi == &sh->dev[i].req) {
2431 rdev = conf->disks[i].rdev;
2434 if (bi == &sh->dev[i].rreq) {
2435 rdev = conf->disks[i].replacement;
2439 /* rdev was removed and 'replacement'
2440 * replaced it. rdev is not removed
2441 * until all requests are finished.
2443 rdev = conf->disks[i].rdev;
2447 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2448 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2457 md_error(conf->mddev, rdev);
2458 else if (is_badblock(rdev, sh->sector,
2460 &first_bad, &bad_sectors))
2461 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2464 set_bit(STRIPE_DEGRADED, &sh->state);
2465 set_bit(WriteErrorSeen, &rdev->flags);
2466 set_bit(R5_WriteError, &sh->dev[i].flags);
2467 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2468 set_bit(MD_RECOVERY_NEEDED,
2469 &rdev->mddev->recovery);
2470 } else if (is_badblock(rdev, sh->sector,
2472 &first_bad, &bad_sectors)) {
2473 set_bit(R5_MadeGood, &sh->dev[i].flags);
2474 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2475 /* That was a successful write so make
2476 * sure it looks like we already did
2479 set_bit(R5_ReWrite, &sh->dev[i].flags);
2482 rdev_dec_pending(rdev, conf->mddev);
2484 if (sh->batch_head && bi->bi_error && !replacement)
2485 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2487 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2488 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2489 set_bit(STRIPE_HANDLE, &sh->state);
2490 raid5_release_stripe(sh);
2492 if (sh->batch_head && sh != sh->batch_head)
2493 raid5_release_stripe(sh->batch_head);
2496 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2498 struct r5dev *dev = &sh->dev[i];
2500 bio_init(&dev->req);
2501 dev->req.bi_io_vec = &dev->vec;
2502 dev->req.bi_max_vecs = 1;
2503 dev->req.bi_private = sh;
2505 bio_init(&dev->rreq);
2506 dev->rreq.bi_io_vec = &dev->rvec;
2507 dev->rreq.bi_max_vecs = 1;
2508 dev->rreq.bi_private = sh;
2511 dev->sector = raid5_compute_blocknr(sh, i, previous);
2514 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2516 char b[BDEVNAME_SIZE];
2517 struct r5conf *conf = mddev->private;
2518 unsigned long flags;
2519 pr_debug("raid456: error called\n");
2521 spin_lock_irqsave(&conf->device_lock, flags);
2522 clear_bit(In_sync, &rdev->flags);
2523 mddev->degraded = calc_degraded(conf);
2524 spin_unlock_irqrestore(&conf->device_lock, flags);
2525 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2527 set_bit(Blocked, &rdev->flags);
2528 set_bit(Faulty, &rdev->flags);
2529 set_mask_bits(&mddev->flags, 0,
2530 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
2532 "md/raid:%s: Disk failure on %s, disabling device.\n"
2533 "md/raid:%s: Operation continuing on %d devices.\n",
2535 bdevname(rdev->bdev, b),
2537 conf->raid_disks - mddev->degraded);
2541 * Input: a 'big' sector number,
2542 * Output: index of the data and parity disk, and the sector # in them.
2544 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2545 int previous, int *dd_idx,
2546 struct stripe_head *sh)
2548 sector_t stripe, stripe2;
2549 sector_t chunk_number;
2550 unsigned int chunk_offset;
2553 sector_t new_sector;
2554 int algorithm = previous ? conf->prev_algo
2556 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2557 : conf->chunk_sectors;
2558 int raid_disks = previous ? conf->previous_raid_disks
2560 int data_disks = raid_disks - conf->max_degraded;
2562 /* First compute the information on this sector */
2565 * Compute the chunk number and the sector offset inside the chunk
2567 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2568 chunk_number = r_sector;
2571 * Compute the stripe number
2573 stripe = chunk_number;
2574 *dd_idx = sector_div(stripe, data_disks);
2577 * Select the parity disk based on the user selected algorithm.
2579 pd_idx = qd_idx = -1;
2580 switch(conf->level) {
2582 pd_idx = data_disks;
2585 switch (algorithm) {
2586 case ALGORITHM_LEFT_ASYMMETRIC:
2587 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2588 if (*dd_idx >= pd_idx)
2591 case ALGORITHM_RIGHT_ASYMMETRIC:
2592 pd_idx = sector_div(stripe2, raid_disks);
2593 if (*dd_idx >= pd_idx)
2596 case ALGORITHM_LEFT_SYMMETRIC:
2597 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2598 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2600 case ALGORITHM_RIGHT_SYMMETRIC:
2601 pd_idx = sector_div(stripe2, raid_disks);
2602 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2604 case ALGORITHM_PARITY_0:
2608 case ALGORITHM_PARITY_N:
2609 pd_idx = data_disks;
2617 switch (algorithm) {
2618 case ALGORITHM_LEFT_ASYMMETRIC:
2619 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2620 qd_idx = pd_idx + 1;
2621 if (pd_idx == raid_disks-1) {
2622 (*dd_idx)++; /* Q D D D P */
2624 } else if (*dd_idx >= pd_idx)
2625 (*dd_idx) += 2; /* D D P Q D */
2627 case ALGORITHM_RIGHT_ASYMMETRIC:
2628 pd_idx = sector_div(stripe2, raid_disks);
2629 qd_idx = pd_idx + 1;
2630 if (pd_idx == raid_disks-1) {
2631 (*dd_idx)++; /* Q D D D P */
2633 } else if (*dd_idx >= pd_idx)
2634 (*dd_idx) += 2; /* D D P Q D */
2636 case ALGORITHM_LEFT_SYMMETRIC:
2637 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2638 qd_idx = (pd_idx + 1) % raid_disks;
2639 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2641 case ALGORITHM_RIGHT_SYMMETRIC:
2642 pd_idx = sector_div(stripe2, raid_disks);
2643 qd_idx = (pd_idx + 1) % raid_disks;
2644 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2647 case ALGORITHM_PARITY_0:
2652 case ALGORITHM_PARITY_N:
2653 pd_idx = data_disks;
2654 qd_idx = data_disks + 1;
2657 case ALGORITHM_ROTATING_ZERO_RESTART:
2658 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2659 * of blocks for computing Q is different.
2661 pd_idx = sector_div(stripe2, raid_disks);
2662 qd_idx = pd_idx + 1;
2663 if (pd_idx == raid_disks-1) {
2664 (*dd_idx)++; /* Q D D D P */
2666 } else if (*dd_idx >= pd_idx)
2667 (*dd_idx) += 2; /* D D P Q D */
2671 case ALGORITHM_ROTATING_N_RESTART:
2672 /* Same a left_asymmetric, by first stripe is
2673 * D D D P Q rather than
2677 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2678 qd_idx = pd_idx + 1;
2679 if (pd_idx == raid_disks-1) {
2680 (*dd_idx)++; /* Q D D D P */
2682 } else if (*dd_idx >= pd_idx)
2683 (*dd_idx) += 2; /* D D P Q D */
2687 case ALGORITHM_ROTATING_N_CONTINUE:
2688 /* Same as left_symmetric but Q is before P */
2689 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2690 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2691 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2695 case ALGORITHM_LEFT_ASYMMETRIC_6:
2696 /* RAID5 left_asymmetric, with Q on last device */
2697 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2698 if (*dd_idx >= pd_idx)
2700 qd_idx = raid_disks - 1;
2703 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2704 pd_idx = sector_div(stripe2, raid_disks-1);
2705 if (*dd_idx >= pd_idx)
2707 qd_idx = raid_disks - 1;
2710 case ALGORITHM_LEFT_SYMMETRIC_6:
2711 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2712 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2713 qd_idx = raid_disks - 1;
2716 case ALGORITHM_RIGHT_SYMMETRIC_6:
2717 pd_idx = sector_div(stripe2, raid_disks-1);
2718 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2719 qd_idx = raid_disks - 1;
2722 case ALGORITHM_PARITY_0_6:
2725 qd_idx = raid_disks - 1;
2735 sh->pd_idx = pd_idx;
2736 sh->qd_idx = qd_idx;
2737 sh->ddf_layout = ddf_layout;
2740 * Finally, compute the new sector number
2742 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2746 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2748 struct r5conf *conf = sh->raid_conf;
2749 int raid_disks = sh->disks;
2750 int data_disks = raid_disks - conf->max_degraded;
2751 sector_t new_sector = sh->sector, check;
2752 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2753 : conf->chunk_sectors;
2754 int algorithm = previous ? conf->prev_algo
2758 sector_t chunk_number;
2759 int dummy1, dd_idx = i;
2761 struct stripe_head sh2;
2763 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2764 stripe = new_sector;
2766 if (i == sh->pd_idx)
2768 switch(conf->level) {
2771 switch (algorithm) {
2772 case ALGORITHM_LEFT_ASYMMETRIC:
2773 case ALGORITHM_RIGHT_ASYMMETRIC:
2777 case ALGORITHM_LEFT_SYMMETRIC:
2778 case ALGORITHM_RIGHT_SYMMETRIC:
2781 i -= (sh->pd_idx + 1);
2783 case ALGORITHM_PARITY_0:
2786 case ALGORITHM_PARITY_N:
2793 if (i == sh->qd_idx)
2794 return 0; /* It is the Q disk */
2795 switch (algorithm) {
2796 case ALGORITHM_LEFT_ASYMMETRIC:
2797 case ALGORITHM_RIGHT_ASYMMETRIC:
2798 case ALGORITHM_ROTATING_ZERO_RESTART:
2799 case ALGORITHM_ROTATING_N_RESTART:
2800 if (sh->pd_idx == raid_disks-1)
2801 i--; /* Q D D D P */
2802 else if (i > sh->pd_idx)
2803 i -= 2; /* D D P Q D */
2805 case ALGORITHM_LEFT_SYMMETRIC:
2806 case ALGORITHM_RIGHT_SYMMETRIC:
2807 if (sh->pd_idx == raid_disks-1)
2808 i--; /* Q D D D P */
2813 i -= (sh->pd_idx + 2);
2816 case ALGORITHM_PARITY_0:
2819 case ALGORITHM_PARITY_N:
2821 case ALGORITHM_ROTATING_N_CONTINUE:
2822 /* Like left_symmetric, but P is before Q */
2823 if (sh->pd_idx == 0)
2824 i--; /* P D D D Q */
2829 i -= (sh->pd_idx + 1);
2832 case ALGORITHM_LEFT_ASYMMETRIC_6:
2833 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2837 case ALGORITHM_LEFT_SYMMETRIC_6:
2838 case ALGORITHM_RIGHT_SYMMETRIC_6:
2840 i += data_disks + 1;
2841 i -= (sh->pd_idx + 1);
2843 case ALGORITHM_PARITY_0_6:
2852 chunk_number = stripe * data_disks + i;
2853 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2855 check = raid5_compute_sector(conf, r_sector,
2856 previous, &dummy1, &sh2);
2857 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2858 || sh2.qd_idx != sh->qd_idx) {
2859 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2860 mdname(conf->mddev));
2867 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2868 int rcw, int expand)
2870 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2871 struct r5conf *conf = sh->raid_conf;
2872 int level = conf->level;
2876 for (i = disks; i--; ) {
2877 struct r5dev *dev = &sh->dev[i];
2880 set_bit(R5_LOCKED, &dev->flags);
2881 set_bit(R5_Wantdrain, &dev->flags);
2883 clear_bit(R5_UPTODATE, &dev->flags);
2887 /* if we are not expanding this is a proper write request, and
2888 * there will be bios with new data to be drained into the
2893 /* False alarm, nothing to do */
2895 sh->reconstruct_state = reconstruct_state_drain_run;
2896 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2898 sh->reconstruct_state = reconstruct_state_run;
2900 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2902 if (s->locked + conf->max_degraded == disks)
2903 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2904 atomic_inc(&conf->pending_full_writes);
2906 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2907 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2908 BUG_ON(level == 6 &&
2909 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2910 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2912 for (i = disks; i--; ) {
2913 struct r5dev *dev = &sh->dev[i];
2914 if (i == pd_idx || i == qd_idx)
2918 (test_bit(R5_UPTODATE, &dev->flags) ||
2919 test_bit(R5_Wantcompute, &dev->flags))) {
2920 set_bit(R5_Wantdrain, &dev->flags);
2921 set_bit(R5_LOCKED, &dev->flags);
2922 clear_bit(R5_UPTODATE, &dev->flags);
2927 /* False alarm - nothing to do */
2929 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2930 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2931 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2932 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2935 /* keep the parity disk(s) locked while asynchronous operations
2938 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2939 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2943 int qd_idx = sh->qd_idx;
2944 struct r5dev *dev = &sh->dev[qd_idx];
2946 set_bit(R5_LOCKED, &dev->flags);
2947 clear_bit(R5_UPTODATE, &dev->flags);
2951 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2952 __func__, (unsigned long long)sh->sector,
2953 s->locked, s->ops_request);
2957 * Each stripe/dev can have one or more bion attached.
2958 * toread/towrite point to the first in a chain.
2959 * The bi_next chain must be in order.
2961 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2962 int forwrite, int previous)
2965 struct r5conf *conf = sh->raid_conf;
2968 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2969 (unsigned long long)bi->bi_iter.bi_sector,
2970 (unsigned long long)sh->sector);
2973 * If several bio share a stripe. The bio bi_phys_segments acts as a
2974 * reference count to avoid race. The reference count should already be
2975 * increased before this function is called (for example, in
2976 * raid5_make_request()), so other bio sharing this stripe will not free the
2977 * stripe. If a stripe is owned by one stripe, the stripe lock will
2980 spin_lock_irq(&sh->stripe_lock);
2981 /* Don't allow new IO added to stripes in batch list */
2985 bip = &sh->dev[dd_idx].towrite;
2989 bip = &sh->dev[dd_idx].toread;
2990 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2991 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2993 bip = & (*bip)->bi_next;
2995 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2998 if (!forwrite || previous)
2999 clear_bit(STRIPE_BATCH_READY, &sh->state);
3001 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3005 raid5_inc_bi_active_stripes(bi);
3008 /* check if page is covered */
3009 sector_t sector = sh->dev[dd_idx].sector;
3010 for (bi=sh->dev[dd_idx].towrite;
3011 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3012 bi && bi->bi_iter.bi_sector <= sector;
3013 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3014 if (bio_end_sector(bi) >= sector)
3015 sector = bio_end_sector(bi);
3017 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3018 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3019 sh->overwrite_disks++;
3022 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3023 (unsigned long long)(*bip)->bi_iter.bi_sector,
3024 (unsigned long long)sh->sector, dd_idx);
3026 if (conf->mddev->bitmap && firstwrite) {
3027 /* Cannot hold spinlock over bitmap_startwrite,
3028 * but must ensure this isn't added to a batch until
3029 * we have added to the bitmap and set bm_seq.
3030 * So set STRIPE_BITMAP_PENDING to prevent
3032 * If multiple add_stripe_bio() calls race here they
3033 * much all set STRIPE_BITMAP_PENDING. So only the first one
3034 * to complete "bitmap_startwrite" gets to set
3035 * STRIPE_BIT_DELAY. This is important as once a stripe
3036 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3039 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3040 spin_unlock_irq(&sh->stripe_lock);
3041 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3043 spin_lock_irq(&sh->stripe_lock);
3044 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3045 if (!sh->batch_head) {
3046 sh->bm_seq = conf->seq_flush+1;
3047 set_bit(STRIPE_BIT_DELAY, &sh->state);
3050 spin_unlock_irq(&sh->stripe_lock);
3052 if (stripe_can_batch(sh))
3053 stripe_add_to_batch_list(conf, sh);
3057 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3058 spin_unlock_irq(&sh->stripe_lock);
3062 static void end_reshape(struct r5conf *conf);
3064 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3065 struct stripe_head *sh)
3067 int sectors_per_chunk =
3068 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3070 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3071 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3073 raid5_compute_sector(conf,
3074 stripe * (disks - conf->max_degraded)
3075 *sectors_per_chunk + chunk_offset,
3081 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3082 struct stripe_head_state *s, int disks,
3083 struct bio_list *return_bi)
3086 BUG_ON(sh->batch_head);
3087 for (i = disks; i--; ) {
3091 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3092 struct md_rdev *rdev;
3094 rdev = rcu_dereference(conf->disks[i].rdev);
3095 if (rdev && test_bit(In_sync, &rdev->flags) &&
3096 !test_bit(Faulty, &rdev->flags))
3097 atomic_inc(&rdev->nr_pending);
3102 if (!rdev_set_badblocks(
3106 md_error(conf->mddev, rdev);
3107 rdev_dec_pending(rdev, conf->mddev);
3110 spin_lock_irq(&sh->stripe_lock);
3111 /* fail all writes first */
3112 bi = sh->dev[i].towrite;
3113 sh->dev[i].towrite = NULL;
3114 sh->overwrite_disks = 0;
3115 spin_unlock_irq(&sh->stripe_lock);
3119 r5l_stripe_write_finished(sh);
3121 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3122 wake_up(&conf->wait_for_overlap);
3124 while (bi && bi->bi_iter.bi_sector <
3125 sh->dev[i].sector + STRIPE_SECTORS) {
3126 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3128 bi->bi_error = -EIO;
3129 if (!raid5_dec_bi_active_stripes(bi)) {
3130 md_write_end(conf->mddev);
3131 bio_list_add(return_bi, bi);
3136 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3137 STRIPE_SECTORS, 0, 0);
3139 /* and fail all 'written' */
3140 bi = sh->dev[i].written;
3141 sh->dev[i].written = NULL;
3142 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3143 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3144 sh->dev[i].page = sh->dev[i].orig_page;
3147 if (bi) bitmap_end = 1;
3148 while (bi && bi->bi_iter.bi_sector <
3149 sh->dev[i].sector + STRIPE_SECTORS) {
3150 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3152 bi->bi_error = -EIO;
3153 if (!raid5_dec_bi_active_stripes(bi)) {
3154 md_write_end(conf->mddev);
3155 bio_list_add(return_bi, bi);
3160 /* fail any reads if this device is non-operational and
3161 * the data has not reached the cache yet.
3163 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3164 s->failed > conf->max_degraded &&
3165 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3166 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3167 spin_lock_irq(&sh->stripe_lock);
3168 bi = sh->dev[i].toread;
3169 sh->dev[i].toread = NULL;
3170 spin_unlock_irq(&sh->stripe_lock);
3171 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3172 wake_up(&conf->wait_for_overlap);
3175 while (bi && bi->bi_iter.bi_sector <
3176 sh->dev[i].sector + STRIPE_SECTORS) {
3177 struct bio *nextbi =
3178 r5_next_bio(bi, sh->dev[i].sector);
3180 bi->bi_error = -EIO;
3181 if (!raid5_dec_bi_active_stripes(bi))
3182 bio_list_add(return_bi, bi);
3187 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3188 STRIPE_SECTORS, 0, 0);
3189 /* If we were in the middle of a write the parity block might
3190 * still be locked - so just clear all R5_LOCKED flags
3192 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3197 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3198 if (atomic_dec_and_test(&conf->pending_full_writes))
3199 md_wakeup_thread(conf->mddev->thread);
3203 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3204 struct stripe_head_state *s)
3209 BUG_ON(sh->batch_head);
3210 clear_bit(STRIPE_SYNCING, &sh->state);
3211 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3212 wake_up(&conf->wait_for_overlap);
3215 /* There is nothing more to do for sync/check/repair.
3216 * Don't even need to abort as that is handled elsewhere
3217 * if needed, and not always wanted e.g. if there is a known
3219 * For recover/replace we need to record a bad block on all
3220 * non-sync devices, or abort the recovery
3222 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3223 /* During recovery devices cannot be removed, so
3224 * locking and refcounting of rdevs is not needed
3227 for (i = 0; i < conf->raid_disks; i++) {
3228 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3230 && !test_bit(Faulty, &rdev->flags)
3231 && !test_bit(In_sync, &rdev->flags)
3232 && !rdev_set_badblocks(rdev, sh->sector,
3235 rdev = rcu_dereference(conf->disks[i].replacement);
3237 && !test_bit(Faulty, &rdev->flags)
3238 && !test_bit(In_sync, &rdev->flags)
3239 && !rdev_set_badblocks(rdev, sh->sector,
3245 conf->recovery_disabled =
3246 conf->mddev->recovery_disabled;
3248 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3251 static int want_replace(struct stripe_head *sh, int disk_idx)
3253 struct md_rdev *rdev;
3257 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3259 && !test_bit(Faulty, &rdev->flags)
3260 && !test_bit(In_sync, &rdev->flags)
3261 && (rdev->recovery_offset <= sh->sector
3262 || rdev->mddev->recovery_cp <= sh->sector))
3268 /* fetch_block - checks the given member device to see if its data needs
3269 * to be read or computed to satisfy a request.
3271 * Returns 1 when no more member devices need to be checked, otherwise returns
3272 * 0 to tell the loop in handle_stripe_fill to continue
3275 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3276 int disk_idx, int disks)
3278 struct r5dev *dev = &sh->dev[disk_idx];
3279 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3280 &sh->dev[s->failed_num[1]] };
3284 if (test_bit(R5_LOCKED, &dev->flags) ||
3285 test_bit(R5_UPTODATE, &dev->flags))
3286 /* No point reading this as we already have it or have
3287 * decided to get it.
3292 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3293 /* We need this block to directly satisfy a request */
3296 if (s->syncing || s->expanding ||
3297 (s->replacing && want_replace(sh, disk_idx)))
3298 /* When syncing, or expanding we read everything.
3299 * When replacing, we need the replaced block.
3303 if ((s->failed >= 1 && fdev[0]->toread) ||
3304 (s->failed >= 2 && fdev[1]->toread))
3305 /* If we want to read from a failed device, then
3306 * we need to actually read every other device.
3310 /* Sometimes neither read-modify-write nor reconstruct-write
3311 * cycles can work. In those cases we read every block we
3312 * can. Then the parity-update is certain to have enough to
3314 * This can only be a problem when we need to write something,
3315 * and some device has failed. If either of those tests
3316 * fail we need look no further.
3318 if (!s->failed || !s->to_write)
3321 if (test_bit(R5_Insync, &dev->flags) &&
3322 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3323 /* Pre-reads at not permitted until after short delay
3324 * to gather multiple requests. However if this
3325 * device is no Insync, the block could only be be computed
3326 * and there is no need to delay that.
3330 for (i = 0; i < s->failed && i < 2; i++) {
3331 if (fdev[i]->towrite &&
3332 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3333 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3334 /* If we have a partial write to a failed
3335 * device, then we will need to reconstruct
3336 * the content of that device, so all other
3337 * devices must be read.
3342 /* If we are forced to do a reconstruct-write, either because
3343 * the current RAID6 implementation only supports that, or
3344 * or because parity cannot be trusted and we are currently
3345 * recovering it, there is extra need to be careful.
3346 * If one of the devices that we would need to read, because
3347 * it is not being overwritten (and maybe not written at all)
3348 * is missing/faulty, then we need to read everything we can.
3350 if (sh->raid_conf->level != 6 &&
3351 sh->sector < sh->raid_conf->mddev->recovery_cp)
3352 /* reconstruct-write isn't being forced */
3354 for (i = 0; i < s->failed && i < 2; i++) {
3355 if (s->failed_num[i] != sh->pd_idx &&
3356 s->failed_num[i] != sh->qd_idx &&
3357 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3358 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3365 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3366 int disk_idx, int disks)
3368 struct r5dev *dev = &sh->dev[disk_idx];
3370 /* is the data in this block needed, and can we get it? */
3371 if (need_this_block(sh, s, disk_idx, disks)) {
3372 /* we would like to get this block, possibly by computing it,
3373 * otherwise read it if the backing disk is insync
3375 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3376 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3377 BUG_ON(sh->batch_head);
3378 if ((s->uptodate == disks - 1) &&
3379 (s->failed && (disk_idx == s->failed_num[0] ||
3380 disk_idx == s->failed_num[1]))) {
3381 /* have disk failed, and we're requested to fetch it;
3384 pr_debug("Computing stripe %llu block %d\n",
3385 (unsigned long long)sh->sector, disk_idx);
3386 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3387 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3388 set_bit(R5_Wantcompute, &dev->flags);
3389 sh->ops.target = disk_idx;
3390 sh->ops.target2 = -1; /* no 2nd target */
3392 /* Careful: from this point on 'uptodate' is in the eye
3393 * of raid_run_ops which services 'compute' operations
3394 * before writes. R5_Wantcompute flags a block that will
3395 * be R5_UPTODATE by the time it is needed for a
3396 * subsequent operation.
3400 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3401 /* Computing 2-failure is *very* expensive; only
3402 * do it if failed >= 2
3405 for (other = disks; other--; ) {
3406 if (other == disk_idx)
3408 if (!test_bit(R5_UPTODATE,
3409 &sh->dev[other].flags))
3413 pr_debug("Computing stripe %llu blocks %d,%d\n",
3414 (unsigned long long)sh->sector,
3416 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3417 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3418 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3419 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3420 sh->ops.target = disk_idx;
3421 sh->ops.target2 = other;
3425 } else if (test_bit(R5_Insync, &dev->flags)) {
3426 set_bit(R5_LOCKED, &dev->flags);
3427 set_bit(R5_Wantread, &dev->flags);
3429 pr_debug("Reading block %d (sync=%d)\n",
3430 disk_idx, s->syncing);
3438 * handle_stripe_fill - read or compute data to satisfy pending requests.
3440 static void handle_stripe_fill(struct stripe_head *sh,
3441 struct stripe_head_state *s,
3446 /* look for blocks to read/compute, skip this if a compute
3447 * is already in flight, or if the stripe contents are in the
3448 * midst of changing due to a write
3450 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3451 !sh->reconstruct_state)
3452 for (i = disks; i--; )
3453 if (fetch_block(sh, s, i, disks))
3455 set_bit(STRIPE_HANDLE, &sh->state);
3458 static void break_stripe_batch_list(struct stripe_head *head_sh,
3459 unsigned long handle_flags);
3460 /* handle_stripe_clean_event
3461 * any written block on an uptodate or failed drive can be returned.
3462 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3463 * never LOCKED, so we don't need to test 'failed' directly.
3465 static void handle_stripe_clean_event(struct r5conf *conf,
3466 struct stripe_head *sh, int disks, struct bio_list *return_bi)
3470 int discard_pending = 0;
3471 struct stripe_head *head_sh = sh;
3472 bool do_endio = false;
3474 for (i = disks; i--; )
3475 if (sh->dev[i].written) {
3477 if (!test_bit(R5_LOCKED, &dev->flags) &&
3478 (test_bit(R5_UPTODATE, &dev->flags) ||
3479 test_bit(R5_Discard, &dev->flags) ||
3480 test_bit(R5_SkipCopy, &dev->flags))) {
3481 /* We can return any write requests */
3482 struct bio *wbi, *wbi2;
3483 pr_debug("Return write for disc %d\n", i);
3484 if (test_and_clear_bit(R5_Discard, &dev->flags))
3485 clear_bit(R5_UPTODATE, &dev->flags);
3486 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3487 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3492 dev->page = dev->orig_page;
3494 dev->written = NULL;
3495 while (wbi && wbi->bi_iter.bi_sector <
3496 dev->sector + STRIPE_SECTORS) {
3497 wbi2 = r5_next_bio(wbi, dev->sector);
3498 if (!raid5_dec_bi_active_stripes(wbi)) {
3499 md_write_end(conf->mddev);
3500 bio_list_add(return_bi, wbi);
3504 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3506 !test_bit(STRIPE_DEGRADED, &sh->state),
3508 if (head_sh->batch_head) {
3509 sh = list_first_entry(&sh->batch_list,
3512 if (sh != head_sh) {
3519 } else if (test_bit(R5_Discard, &dev->flags))
3520 discard_pending = 1;
3523 r5l_stripe_write_finished(sh);
3525 if (!discard_pending &&
3526 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3528 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3529 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3530 if (sh->qd_idx >= 0) {
3531 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3532 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3534 /* now that discard is done we can proceed with any sync */
3535 clear_bit(STRIPE_DISCARD, &sh->state);
3537 * SCSI discard will change some bio fields and the stripe has
3538 * no updated data, so remove it from hash list and the stripe
3539 * will be reinitialized
3542 hash = sh->hash_lock_index;
3543 spin_lock_irq(conf->hash_locks + hash);
3545 spin_unlock_irq(conf->hash_locks + hash);
3546 if (head_sh->batch_head) {
3547 sh = list_first_entry(&sh->batch_list,
3548 struct stripe_head, batch_list);
3554 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3555 set_bit(STRIPE_HANDLE, &sh->state);
3559 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3560 if (atomic_dec_and_test(&conf->pending_full_writes))
3561 md_wakeup_thread(conf->mddev->thread);
3563 if (head_sh->batch_head && do_endio)
3564 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3567 static void handle_stripe_dirtying(struct r5conf *conf,
3568 struct stripe_head *sh,
3569 struct stripe_head_state *s,
3572 int rmw = 0, rcw = 0, i;
3573 sector_t recovery_cp = conf->mddev->recovery_cp;
3575 /* Check whether resync is now happening or should start.
3576 * If yes, then the array is dirty (after unclean shutdown or
3577 * initial creation), so parity in some stripes might be inconsistent.
3578 * In this case, we need to always do reconstruct-write, to ensure
3579 * that in case of drive failure or read-error correction, we
3580 * generate correct data from the parity.
3582 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3583 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3585 /* Calculate the real rcw later - for now make it
3586 * look like rcw is cheaper
3589 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3590 conf->rmw_level, (unsigned long long)recovery_cp,
3591 (unsigned long long)sh->sector);
3592 } else for (i = disks; i--; ) {
3593 /* would I have to read this buffer for read_modify_write */
3594 struct r5dev *dev = &sh->dev[i];
3595 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3596 !test_bit(R5_LOCKED, &dev->flags) &&
3597 !(test_bit(R5_UPTODATE, &dev->flags) ||
3598 test_bit(R5_Wantcompute, &dev->flags))) {
3599 if (test_bit(R5_Insync, &dev->flags))
3602 rmw += 2*disks; /* cannot read it */
3604 /* Would I have to read this buffer for reconstruct_write */
3605 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3606 i != sh->pd_idx && i != sh->qd_idx &&
3607 !test_bit(R5_LOCKED, &dev->flags) &&
3608 !(test_bit(R5_UPTODATE, &dev->flags) ||
3609 test_bit(R5_Wantcompute, &dev->flags))) {
3610 if (test_bit(R5_Insync, &dev->flags))
3616 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3617 (unsigned long long)sh->sector, rmw, rcw);
3618 set_bit(STRIPE_HANDLE, &sh->state);
3619 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3620 /* prefer read-modify-write, but need to get some data */
3621 if (conf->mddev->queue)
3622 blk_add_trace_msg(conf->mddev->queue,
3623 "raid5 rmw %llu %d",
3624 (unsigned long long)sh->sector, rmw);
3625 for (i = disks; i--; ) {
3626 struct r5dev *dev = &sh->dev[i];
3627 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3628 !test_bit(R5_LOCKED, &dev->flags) &&
3629 !(test_bit(R5_UPTODATE, &dev->flags) ||
3630 test_bit(R5_Wantcompute, &dev->flags)) &&
3631 test_bit(R5_Insync, &dev->flags)) {
3632 if (test_bit(STRIPE_PREREAD_ACTIVE,
3634 pr_debug("Read_old block %d for r-m-w\n",
3636 set_bit(R5_LOCKED, &dev->flags);
3637 set_bit(R5_Wantread, &dev->flags);
3640 set_bit(STRIPE_DELAYED, &sh->state);
3641 set_bit(STRIPE_HANDLE, &sh->state);
3646 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3647 /* want reconstruct write, but need to get some data */
3650 for (i = disks; i--; ) {
3651 struct r5dev *dev = &sh->dev[i];
3652 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3653 i != sh->pd_idx && i != sh->qd_idx &&
3654 !test_bit(R5_LOCKED, &dev->flags) &&
3655 !(test_bit(R5_UPTODATE, &dev->flags) ||
3656 test_bit(R5_Wantcompute, &dev->flags))) {
3658 if (test_bit(R5_Insync, &dev->flags) &&
3659 test_bit(STRIPE_PREREAD_ACTIVE,
3661 pr_debug("Read_old block "
3662 "%d for Reconstruct\n", i);
3663 set_bit(R5_LOCKED, &dev->flags);
3664 set_bit(R5_Wantread, &dev->flags);
3668 set_bit(STRIPE_DELAYED, &sh->state);
3669 set_bit(STRIPE_HANDLE, &sh->state);
3673 if (rcw && conf->mddev->queue)
3674 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3675 (unsigned long long)sh->sector,
3676 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3679 if (rcw > disks && rmw > disks &&
3680 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3681 set_bit(STRIPE_DELAYED, &sh->state);
3683 /* now if nothing is locked, and if we have enough data,
3684 * we can start a write request
3686 /* since handle_stripe can be called at any time we need to handle the
3687 * case where a compute block operation has been submitted and then a
3688 * subsequent call wants to start a write request. raid_run_ops only
3689 * handles the case where compute block and reconstruct are requested
3690 * simultaneously. If this is not the case then new writes need to be
3691 * held off until the compute completes.
3693 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3694 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3695 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3696 schedule_reconstruction(sh, s, rcw == 0, 0);
3699 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3700 struct stripe_head_state *s, int disks)
3702 struct r5dev *dev = NULL;
3704 BUG_ON(sh->batch_head);
3705 set_bit(STRIPE_HANDLE, &sh->state);
3707 switch (sh->check_state) {
3708 case check_state_idle:
3709 /* start a new check operation if there are no failures */
3710 if (s->failed == 0) {
3711 BUG_ON(s->uptodate != disks);
3712 sh->check_state = check_state_run;
3713 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3714 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3718 dev = &sh->dev[s->failed_num[0]];
3720 case check_state_compute_result:
3721 sh->check_state = check_state_idle;
3723 dev = &sh->dev[sh->pd_idx];
3725 /* check that a write has not made the stripe insync */
3726 if (test_bit(STRIPE_INSYNC, &sh->state))
3729 /* either failed parity check, or recovery is happening */
3730 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3731 BUG_ON(s->uptodate != disks);
3733 set_bit(R5_LOCKED, &dev->flags);
3735 set_bit(R5_Wantwrite, &dev->flags);
3737 clear_bit(STRIPE_DEGRADED, &sh->state);
3738 set_bit(STRIPE_INSYNC, &sh->state);
3740 case check_state_run:
3741 break; /* we will be called again upon completion */
3742 case check_state_check_result:
3743 sh->check_state = check_state_idle;
3745 /* if a failure occurred during the check operation, leave
3746 * STRIPE_INSYNC not set and let the stripe be handled again
3751 /* handle a successful check operation, if parity is correct
3752 * we are done. Otherwise update the mismatch count and repair
3753 * parity if !MD_RECOVERY_CHECK
3755 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3756 /* parity is correct (on disc,
3757 * not in buffer any more)
3759 set_bit(STRIPE_INSYNC, &sh->state);
3761 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3762 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3763 /* don't try to repair!! */
3764 set_bit(STRIPE_INSYNC, &sh->state);
3766 sh->check_state = check_state_compute_run;
3767 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3768 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3769 set_bit(R5_Wantcompute,
3770 &sh->dev[sh->pd_idx].flags);
3771 sh->ops.target = sh->pd_idx;
3772 sh->ops.target2 = -1;
3777 case check_state_compute_run:
3780 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3781 __func__, sh->check_state,
3782 (unsigned long long) sh->sector);
3787 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3788 struct stripe_head_state *s,
3791 int pd_idx = sh->pd_idx;
3792 int qd_idx = sh->qd_idx;
3795 BUG_ON(sh->batch_head);
3796 set_bit(STRIPE_HANDLE, &sh->state);
3798 BUG_ON(s->failed > 2);
3800 /* Want to check and possibly repair P and Q.
3801 * However there could be one 'failed' device, in which
3802 * case we can only check one of them, possibly using the
3803 * other to generate missing data
3806 switch (sh->check_state) {
3807 case check_state_idle:
3808 /* start a new check operation if there are < 2 failures */
3809 if (s->failed == s->q_failed) {
3810 /* The only possible failed device holds Q, so it
3811 * makes sense to check P (If anything else were failed,
3812 * we would have used P to recreate it).
3814 sh->check_state = check_state_run;
3816 if (!s->q_failed && s->failed < 2) {
3817 /* Q is not failed, and we didn't use it to generate
3818 * anything, so it makes sense to check it
3820 if (sh->check_state == check_state_run)
3821 sh->check_state = check_state_run_pq;
3823 sh->check_state = check_state_run_q;
3826 /* discard potentially stale zero_sum_result */
3827 sh->ops.zero_sum_result = 0;
3829 if (sh->check_state == check_state_run) {
3830 /* async_xor_zero_sum destroys the contents of P */
3831 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3834 if (sh->check_state >= check_state_run &&
3835 sh->check_state <= check_state_run_pq) {
3836 /* async_syndrome_zero_sum preserves P and Q, so
3837 * no need to mark them !uptodate here
3839 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3843 /* we have 2-disk failure */
3844 BUG_ON(s->failed != 2);
3846 case check_state_compute_result:
3847 sh->check_state = check_state_idle;
3849 /* check that a write has not made the stripe insync */
3850 if (test_bit(STRIPE_INSYNC, &sh->state))
3853 /* now write out any block on a failed drive,
3854 * or P or Q if they were recomputed
3856 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3857 if (s->failed == 2) {
3858 dev = &sh->dev[s->failed_num[1]];
3860 set_bit(R5_LOCKED, &dev->flags);
3861 set_bit(R5_Wantwrite, &dev->flags);
3863 if (s->failed >= 1) {
3864 dev = &sh->dev[s->failed_num[0]];
3866 set_bit(R5_LOCKED, &dev->flags);
3867 set_bit(R5_Wantwrite, &dev->flags);
3869 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3870 dev = &sh->dev[pd_idx];
3872 set_bit(R5_LOCKED, &dev->flags);
3873 set_bit(R5_Wantwrite, &dev->flags);
3875 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3876 dev = &sh->dev[qd_idx];
3878 set_bit(R5_LOCKED, &dev->flags);
3879 set_bit(R5_Wantwrite, &dev->flags);
3881 clear_bit(STRIPE_DEGRADED, &sh->state);
3883 set_bit(STRIPE_INSYNC, &sh->state);
3885 case check_state_run:
3886 case check_state_run_q:
3887 case check_state_run_pq:
3888 break; /* we will be called again upon completion */
3889 case check_state_check_result:
3890 sh->check_state = check_state_idle;
3892 /* handle a successful check operation, if parity is correct
3893 * we are done. Otherwise update the mismatch count and repair
3894 * parity if !MD_RECOVERY_CHECK
3896 if (sh->ops.zero_sum_result == 0) {
3897 /* both parities are correct */
3899 set_bit(STRIPE_INSYNC, &sh->state);
3901 /* in contrast to the raid5 case we can validate
3902 * parity, but still have a failure to write
3905 sh->check_state = check_state_compute_result;
3906 /* Returning at this point means that we may go
3907 * off and bring p and/or q uptodate again so
3908 * we make sure to check zero_sum_result again
3909 * to verify if p or q need writeback
3913 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3914 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3915 /* don't try to repair!! */
3916 set_bit(STRIPE_INSYNC, &sh->state);
3918 int *target = &sh->ops.target;
3920 sh->ops.target = -1;
3921 sh->ops.target2 = -1;
3922 sh->check_state = check_state_compute_run;
3923 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3924 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3925 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3926 set_bit(R5_Wantcompute,
3927 &sh->dev[pd_idx].flags);
3929 target = &sh->ops.target2;
3932 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3933 set_bit(R5_Wantcompute,
3934 &sh->dev[qd_idx].flags);
3941 case check_state_compute_run:
3944 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3945 __func__, sh->check_state,
3946 (unsigned long long) sh->sector);
3951 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3955 /* We have read all the blocks in this stripe and now we need to
3956 * copy some of them into a target stripe for expand.
3958 struct dma_async_tx_descriptor *tx = NULL;
3959 BUG_ON(sh->batch_head);
3960 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3961 for (i = 0; i < sh->disks; i++)
3962 if (i != sh->pd_idx && i != sh->qd_idx) {
3964 struct stripe_head *sh2;
3965 struct async_submit_ctl submit;
3967 sector_t bn = raid5_compute_blocknr(sh, i, 1);
3968 sector_t s = raid5_compute_sector(conf, bn, 0,
3970 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
3972 /* so far only the early blocks of this stripe
3973 * have been requested. When later blocks
3974 * get requested, we will try again
3977 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3978 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3979 /* must have already done this block */
3980 raid5_release_stripe(sh2);
3984 /* place all the copies on one channel */
3985 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3986 tx = async_memcpy(sh2->dev[dd_idx].page,
3987 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3990 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3991 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3992 for (j = 0; j < conf->raid_disks; j++)
3993 if (j != sh2->pd_idx &&
3995 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3997 if (j == conf->raid_disks) {
3998 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3999 set_bit(STRIPE_HANDLE, &sh2->state);
4001 raid5_release_stripe(sh2);
4004 /* done submitting copies, wait for them to complete */
4005 async_tx_quiesce(&tx);
4009 * handle_stripe - do things to a stripe.
4011 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4012 * state of various bits to see what needs to be done.
4014 * return some read requests which now have data
4015 * return some write requests which are safely on storage
4016 * schedule a read on some buffers
4017 * schedule a write of some buffers
4018 * return confirmation of parity correctness
4022 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4024 struct r5conf *conf = sh->raid_conf;
4025 int disks = sh->disks;
4028 int do_recovery = 0;
4030 memset(s, 0, sizeof(*s));
4032 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4033 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4034 s->failed_num[0] = -1;
4035 s->failed_num[1] = -1;
4036 s->log_failed = r5l_log_disk_error(conf);
4038 /* Now to look around and see what can be done */
4040 for (i=disks; i--; ) {
4041 struct md_rdev *rdev;
4048 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4050 dev->toread, dev->towrite, dev->written);
4051 /* maybe we can reply to a read
4053 * new wantfill requests are only permitted while
4054 * ops_complete_biofill is guaranteed to be inactive
4056 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4057 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4058 set_bit(R5_Wantfill, &dev->flags);
4060 /* now count some things */
4061 if (test_bit(R5_LOCKED, &dev->flags))
4063 if (test_bit(R5_UPTODATE, &dev->flags))
4065 if (test_bit(R5_Wantcompute, &dev->flags)) {
4067 BUG_ON(s->compute > 2);
4070 if (test_bit(R5_Wantfill, &dev->flags))
4072 else if (dev->toread)
4076 if (!test_bit(R5_OVERWRITE, &dev->flags))
4081 /* Prefer to use the replacement for reads, but only
4082 * if it is recovered enough and has no bad blocks.
4084 rdev = rcu_dereference(conf->disks[i].replacement);
4085 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4086 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4087 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4088 &first_bad, &bad_sectors))
4089 set_bit(R5_ReadRepl, &dev->flags);
4091 if (rdev && !test_bit(Faulty, &rdev->flags))
4092 set_bit(R5_NeedReplace, &dev->flags);
4094 clear_bit(R5_NeedReplace, &dev->flags);
4095 rdev = rcu_dereference(conf->disks[i].rdev);
4096 clear_bit(R5_ReadRepl, &dev->flags);
4098 if (rdev && test_bit(Faulty, &rdev->flags))
4101 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4102 &first_bad, &bad_sectors);
4103 if (s->blocked_rdev == NULL
4104 && (test_bit(Blocked, &rdev->flags)
4107 set_bit(BlockedBadBlocks,
4109 s->blocked_rdev = rdev;
4110 atomic_inc(&rdev->nr_pending);
4113 clear_bit(R5_Insync, &dev->flags);
4117 /* also not in-sync */
4118 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4119 test_bit(R5_UPTODATE, &dev->flags)) {
4120 /* treat as in-sync, but with a read error
4121 * which we can now try to correct
4123 set_bit(R5_Insync, &dev->flags);
4124 set_bit(R5_ReadError, &dev->flags);
4126 } else if (test_bit(In_sync, &rdev->flags))
4127 set_bit(R5_Insync, &dev->flags);
4128 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4129 /* in sync if before recovery_offset */
4130 set_bit(R5_Insync, &dev->flags);
4131 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4132 test_bit(R5_Expanded, &dev->flags))
4133 /* If we've reshaped into here, we assume it is Insync.
4134 * We will shortly update recovery_offset to make
4137 set_bit(R5_Insync, &dev->flags);
4139 if (test_bit(R5_WriteError, &dev->flags)) {
4140 /* This flag does not apply to '.replacement'
4141 * only to .rdev, so make sure to check that*/
4142 struct md_rdev *rdev2 = rcu_dereference(
4143 conf->disks[i].rdev);
4145 clear_bit(R5_Insync, &dev->flags);
4146 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4147 s->handle_bad_blocks = 1;
4148 atomic_inc(&rdev2->nr_pending);
4150 clear_bit(R5_WriteError, &dev->flags);
4152 if (test_bit(R5_MadeGood, &dev->flags)) {
4153 /* This flag does not apply to '.replacement'
4154 * only to .rdev, so make sure to check that*/
4155 struct md_rdev *rdev2 = rcu_dereference(
4156 conf->disks[i].rdev);
4157 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4158 s->handle_bad_blocks = 1;
4159 atomic_inc(&rdev2->nr_pending);
4161 clear_bit(R5_MadeGood, &dev->flags);
4163 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4164 struct md_rdev *rdev2 = rcu_dereference(
4165 conf->disks[i].replacement);
4166 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4167 s->handle_bad_blocks = 1;
4168 atomic_inc(&rdev2->nr_pending);
4170 clear_bit(R5_MadeGoodRepl, &dev->flags);
4172 if (!test_bit(R5_Insync, &dev->flags)) {
4173 /* The ReadError flag will just be confusing now */
4174 clear_bit(R5_ReadError, &dev->flags);
4175 clear_bit(R5_ReWrite, &dev->flags);
4177 if (test_bit(R5_ReadError, &dev->flags))
4178 clear_bit(R5_Insync, &dev->flags);
4179 if (!test_bit(R5_Insync, &dev->flags)) {
4181 s->failed_num[s->failed] = i;
4183 if (rdev && !test_bit(Faulty, &rdev->flags))
4187 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4188 /* If there is a failed device being replaced,
4189 * we must be recovering.
4190 * else if we are after recovery_cp, we must be syncing
4191 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4192 * else we can only be replacing
4193 * sync and recovery both need to read all devices, and so
4194 * use the same flag.
4197 sh->sector >= conf->mddev->recovery_cp ||
4198 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4206 static int clear_batch_ready(struct stripe_head *sh)
4208 /* Return '1' if this is a member of batch, or
4209 * '0' if it is a lone stripe or a head which can now be
4212 struct stripe_head *tmp;
4213 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4214 return (sh->batch_head && sh->batch_head != sh);
4215 spin_lock(&sh->stripe_lock);
4216 if (!sh->batch_head) {
4217 spin_unlock(&sh->stripe_lock);
4222 * this stripe could be added to a batch list before we check
4223 * BATCH_READY, skips it
4225 if (sh->batch_head != sh) {
4226 spin_unlock(&sh->stripe_lock);
4229 spin_lock(&sh->batch_lock);
4230 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4231 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4232 spin_unlock(&sh->batch_lock);
4233 spin_unlock(&sh->stripe_lock);
4236 * BATCH_READY is cleared, no new stripes can be added.
4237 * batch_list can be accessed without lock
4242 static void break_stripe_batch_list(struct stripe_head *head_sh,
4243 unsigned long handle_flags)
4245 struct stripe_head *sh, *next;
4249 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4251 list_del_init(&sh->batch_list);
4253 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4254 (1 << STRIPE_SYNCING) |
4255 (1 << STRIPE_REPLACED) |
4256 (1 << STRIPE_DELAYED) |
4257 (1 << STRIPE_BIT_DELAY) |
4258 (1 << STRIPE_FULL_WRITE) |
4259 (1 << STRIPE_BIOFILL_RUN) |
4260 (1 << STRIPE_COMPUTE_RUN) |
4261 (1 << STRIPE_OPS_REQ_PENDING) |
4262 (1 << STRIPE_DISCARD) |
4263 (1 << STRIPE_BATCH_READY) |
4264 (1 << STRIPE_BATCH_ERR) |
4265 (1 << STRIPE_BITMAP_PENDING)),
4266 "stripe state: %lx\n", sh->state);
4267 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4268 (1 << STRIPE_REPLACED)),
4269 "head stripe state: %lx\n", head_sh->state);
4271 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4272 (1 << STRIPE_PREREAD_ACTIVE) |
4273 (1 << STRIPE_DEGRADED)),
4274 head_sh->state & (1 << STRIPE_INSYNC));
4276 sh->check_state = head_sh->check_state;
4277 sh->reconstruct_state = head_sh->reconstruct_state;
4278 for (i = 0; i < sh->disks; i++) {
4279 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4281 sh->dev[i].flags = head_sh->dev[i].flags &
4282 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4284 spin_lock_irq(&sh->stripe_lock);
4285 sh->batch_head = NULL;
4286 spin_unlock_irq(&sh->stripe_lock);
4287 if (handle_flags == 0 ||
4288 sh->state & handle_flags)
4289 set_bit(STRIPE_HANDLE, &sh->state);
4290 raid5_release_stripe(sh);
4292 spin_lock_irq(&head_sh->stripe_lock);
4293 head_sh->batch_head = NULL;
4294 spin_unlock_irq(&head_sh->stripe_lock);
4295 for (i = 0; i < head_sh->disks; i++)
4296 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4298 if (head_sh->state & handle_flags)
4299 set_bit(STRIPE_HANDLE, &head_sh->state);
4302 wake_up(&head_sh->raid_conf->wait_for_overlap);
4305 static void handle_stripe(struct stripe_head *sh)
4307 struct stripe_head_state s;
4308 struct r5conf *conf = sh->raid_conf;
4311 int disks = sh->disks;
4312 struct r5dev *pdev, *qdev;
4314 clear_bit(STRIPE_HANDLE, &sh->state);
4315 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4316 /* already being handled, ensure it gets handled
4317 * again when current action finishes */
4318 set_bit(STRIPE_HANDLE, &sh->state);
4322 if (clear_batch_ready(sh) ) {
4323 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4327 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4328 break_stripe_batch_list(sh, 0);
4330 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4331 spin_lock(&sh->stripe_lock);
4332 /* Cannot process 'sync' concurrently with 'discard' */
4333 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4334 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4335 set_bit(STRIPE_SYNCING, &sh->state);
4336 clear_bit(STRIPE_INSYNC, &sh->state);
4337 clear_bit(STRIPE_REPLACED, &sh->state);
4339 spin_unlock(&sh->stripe_lock);
4341 clear_bit(STRIPE_DELAYED, &sh->state);
4343 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4344 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4345 (unsigned long long)sh->sector, sh->state,
4346 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4347 sh->check_state, sh->reconstruct_state);
4349 analyse_stripe(sh, &s);
4351 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4354 if (s.handle_bad_blocks) {
4355 set_bit(STRIPE_HANDLE, &sh->state);
4359 if (unlikely(s.blocked_rdev)) {
4360 if (s.syncing || s.expanding || s.expanded ||
4361 s.replacing || s.to_write || s.written) {
4362 set_bit(STRIPE_HANDLE, &sh->state);
4365 /* There is nothing for the blocked_rdev to block */
4366 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4367 s.blocked_rdev = NULL;
4370 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4371 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4372 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4375 pr_debug("locked=%d uptodate=%d to_read=%d"
4376 " to_write=%d failed=%d failed_num=%d,%d\n",
4377 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4378 s.failed_num[0], s.failed_num[1]);
4379 /* check if the array has lost more than max_degraded devices and,
4380 * if so, some requests might need to be failed.
4382 if (s.failed > conf->max_degraded || s.log_failed) {
4383 sh->check_state = 0;
4384 sh->reconstruct_state = 0;
4385 break_stripe_batch_list(sh, 0);
4386 if (s.to_read+s.to_write+s.written)
4387 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4388 if (s.syncing + s.replacing)
4389 handle_failed_sync(conf, sh, &s);
4392 /* Now we check to see if any write operations have recently
4396 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4398 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4399 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4400 sh->reconstruct_state = reconstruct_state_idle;
4402 /* All the 'written' buffers and the parity block are ready to
4403 * be written back to disk
4405 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4406 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4407 BUG_ON(sh->qd_idx >= 0 &&
4408 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4409 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4410 for (i = disks; i--; ) {
4411 struct r5dev *dev = &sh->dev[i];
4412 if (test_bit(R5_LOCKED, &dev->flags) &&
4413 (i == sh->pd_idx || i == sh->qd_idx ||
4415 pr_debug("Writing block %d\n", i);
4416 set_bit(R5_Wantwrite, &dev->flags);
4421 if (!test_bit(R5_Insync, &dev->flags) ||
4422 ((i == sh->pd_idx || i == sh->qd_idx) &&
4424 set_bit(STRIPE_INSYNC, &sh->state);
4427 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4428 s.dec_preread_active = 1;
4432 * might be able to return some write requests if the parity blocks
4433 * are safe, or on a failed drive
4435 pdev = &sh->dev[sh->pd_idx];
4436 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4437 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4438 qdev = &sh->dev[sh->qd_idx];
4439 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4440 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4444 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4445 && !test_bit(R5_LOCKED, &pdev->flags)
4446 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4447 test_bit(R5_Discard, &pdev->flags))))) &&
4448 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4449 && !test_bit(R5_LOCKED, &qdev->flags)
4450 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4451 test_bit(R5_Discard, &qdev->flags))))))
4452 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4454 /* Now we might consider reading some blocks, either to check/generate
4455 * parity, or to satisfy requests
4456 * or to load a block that is being partially written.
4458 if (s.to_read || s.non_overwrite
4459 || (conf->level == 6 && s.to_write && s.failed)
4460 || (s.syncing && (s.uptodate + s.compute < disks))
4463 handle_stripe_fill(sh, &s, disks);
4465 /* Now to consider new write requests and what else, if anything
4466 * should be read. We do not handle new writes when:
4467 * 1/ A 'write' operation (copy+xor) is already in flight.
4468 * 2/ A 'check' operation is in flight, as it may clobber the parity
4471 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4472 handle_stripe_dirtying(conf, sh, &s, disks);
4474 /* maybe we need to check and possibly fix the parity for this stripe
4475 * Any reads will already have been scheduled, so we just see if enough
4476 * data is available. The parity check is held off while parity
4477 * dependent operations are in flight.
4479 if (sh->check_state ||
4480 (s.syncing && s.locked == 0 &&
4481 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4482 !test_bit(STRIPE_INSYNC, &sh->state))) {
4483 if (conf->level == 6)
4484 handle_parity_checks6(conf, sh, &s, disks);
4486 handle_parity_checks5(conf, sh, &s, disks);
4489 if ((s.replacing || s.syncing) && s.locked == 0
4490 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4491 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4492 /* Write out to replacement devices where possible */
4493 for (i = 0; i < conf->raid_disks; i++)
4494 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4495 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4496 set_bit(R5_WantReplace, &sh->dev[i].flags);
4497 set_bit(R5_LOCKED, &sh->dev[i].flags);
4501 set_bit(STRIPE_INSYNC, &sh->state);
4502 set_bit(STRIPE_REPLACED, &sh->state);
4504 if ((s.syncing || s.replacing) && s.locked == 0 &&
4505 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4506 test_bit(STRIPE_INSYNC, &sh->state)) {
4507 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4508 clear_bit(STRIPE_SYNCING, &sh->state);
4509 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4510 wake_up(&conf->wait_for_overlap);
4513 /* If the failed drives are just a ReadError, then we might need
4514 * to progress the repair/check process
4516 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4517 for (i = 0; i < s.failed; i++) {
4518 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4519 if (test_bit(R5_ReadError, &dev->flags)
4520 && !test_bit(R5_LOCKED, &dev->flags)
4521 && test_bit(R5_UPTODATE, &dev->flags)
4523 if (!test_bit(R5_ReWrite, &dev->flags)) {
4524 set_bit(R5_Wantwrite, &dev->flags);
4525 set_bit(R5_ReWrite, &dev->flags);
4526 set_bit(R5_LOCKED, &dev->flags);
4529 /* let's read it back */
4530 set_bit(R5_Wantread, &dev->flags);
4531 set_bit(R5_LOCKED, &dev->flags);
4537 /* Finish reconstruct operations initiated by the expansion process */
4538 if (sh->reconstruct_state == reconstruct_state_result) {
4539 struct stripe_head *sh_src
4540 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4541 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4542 /* sh cannot be written until sh_src has been read.
4543 * so arrange for sh to be delayed a little
4545 set_bit(STRIPE_DELAYED, &sh->state);
4546 set_bit(STRIPE_HANDLE, &sh->state);
4547 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4549 atomic_inc(&conf->preread_active_stripes);
4550 raid5_release_stripe(sh_src);
4554 raid5_release_stripe(sh_src);
4556 sh->reconstruct_state = reconstruct_state_idle;
4557 clear_bit(STRIPE_EXPANDING, &sh->state);
4558 for (i = conf->raid_disks; i--; ) {
4559 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4560 set_bit(R5_LOCKED, &sh->dev[i].flags);
4565 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4566 !sh->reconstruct_state) {
4567 /* Need to write out all blocks after computing parity */
4568 sh->disks = conf->raid_disks;
4569 stripe_set_idx(sh->sector, conf, 0, sh);
4570 schedule_reconstruction(sh, &s, 1, 1);
4571 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4572 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4573 atomic_dec(&conf->reshape_stripes);
4574 wake_up(&conf->wait_for_overlap);
4575 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4578 if (s.expanding && s.locked == 0 &&
4579 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4580 handle_stripe_expansion(conf, sh);
4583 /* wait for this device to become unblocked */
4584 if (unlikely(s.blocked_rdev)) {
4585 if (conf->mddev->external)
4586 md_wait_for_blocked_rdev(s.blocked_rdev,
4589 /* Internal metadata will immediately
4590 * be written by raid5d, so we don't
4591 * need to wait here.
4593 rdev_dec_pending(s.blocked_rdev,
4597 if (s.handle_bad_blocks)
4598 for (i = disks; i--; ) {
4599 struct md_rdev *rdev;
4600 struct r5dev *dev = &sh->dev[i];
4601 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4602 /* We own a safe reference to the rdev */
4603 rdev = conf->disks[i].rdev;
4604 if (!rdev_set_badblocks(rdev, sh->sector,
4606 md_error(conf->mddev, rdev);
4607 rdev_dec_pending(rdev, conf->mddev);
4609 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4610 rdev = conf->disks[i].rdev;
4611 rdev_clear_badblocks(rdev, sh->sector,
4613 rdev_dec_pending(rdev, conf->mddev);
4615 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4616 rdev = conf->disks[i].replacement;
4618 /* rdev have been moved down */
4619 rdev = conf->disks[i].rdev;
4620 rdev_clear_badblocks(rdev, sh->sector,
4622 rdev_dec_pending(rdev, conf->mddev);
4627 raid_run_ops(sh, s.ops_request);
4631 if (s.dec_preread_active) {
4632 /* We delay this until after ops_run_io so that if make_request
4633 * is waiting on a flush, it won't continue until the writes
4634 * have actually been submitted.
4636 atomic_dec(&conf->preread_active_stripes);
4637 if (atomic_read(&conf->preread_active_stripes) <
4639 md_wakeup_thread(conf->mddev->thread);
4642 if (!bio_list_empty(&s.return_bi)) {
4643 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags) &&
4644 (s.failed <= conf->max_degraded ||
4645 conf->mddev->external == 0)) {
4646 spin_lock_irq(&conf->device_lock);
4647 bio_list_merge(&conf->return_bi, &s.return_bi);
4648 spin_unlock_irq(&conf->device_lock);
4649 md_wakeup_thread(conf->mddev->thread);
4651 return_io(&s.return_bi);
4654 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4657 static void raid5_activate_delayed(struct r5conf *conf)
4659 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4660 while (!list_empty(&conf->delayed_list)) {
4661 struct list_head *l = conf->delayed_list.next;
4662 struct stripe_head *sh;
4663 sh = list_entry(l, struct stripe_head, lru);
4665 clear_bit(STRIPE_DELAYED, &sh->state);
4666 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4667 atomic_inc(&conf->preread_active_stripes);
4668 list_add_tail(&sh->lru, &conf->hold_list);
4669 raid5_wakeup_stripe_thread(sh);
4674 static void activate_bit_delay(struct r5conf *conf,
4675 struct list_head *temp_inactive_list)
4677 /* device_lock is held */
4678 struct list_head head;
4679 list_add(&head, &conf->bitmap_list);
4680 list_del_init(&conf->bitmap_list);
4681 while (!list_empty(&head)) {
4682 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4684 list_del_init(&sh->lru);
4685 atomic_inc(&sh->count);
4686 hash = sh->hash_lock_index;
4687 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4691 static int raid5_congested(struct mddev *mddev, int bits)
4693 struct r5conf *conf = mddev->private;
4695 /* No difference between reads and writes. Just check
4696 * how busy the stripe_cache is
4699 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4703 if (atomic_read(&conf->empty_inactive_list_nr))
4709 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4711 struct r5conf *conf = mddev->private;
4712 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4713 unsigned int chunk_sectors;
4714 unsigned int bio_sectors = bio_sectors(bio);
4716 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4717 return chunk_sectors >=
4718 ((sector & (chunk_sectors - 1)) + bio_sectors);
4722 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4723 * later sampled by raid5d.
4725 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4727 unsigned long flags;
4729 spin_lock_irqsave(&conf->device_lock, flags);
4731 bi->bi_next = conf->retry_read_aligned_list;
4732 conf->retry_read_aligned_list = bi;
4734 spin_unlock_irqrestore(&conf->device_lock, flags);
4735 md_wakeup_thread(conf->mddev->thread);
4738 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4742 bi = conf->retry_read_aligned;
4744 conf->retry_read_aligned = NULL;
4747 bi = conf->retry_read_aligned_list;
4749 conf->retry_read_aligned_list = bi->bi_next;
4752 * this sets the active strip count to 1 and the processed
4753 * strip count to zero (upper 8 bits)
4755 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4762 * The "raid5_align_endio" should check if the read succeeded and if it
4763 * did, call bio_endio on the original bio (having bio_put the new bio
4765 * If the read failed..
4767 static void raid5_align_endio(struct bio *bi)
4769 struct bio* raid_bi = bi->bi_private;
4770 struct mddev *mddev;
4771 struct r5conf *conf;
4772 struct md_rdev *rdev;
4773 int error = bi->bi_error;
4777 rdev = (void*)raid_bi->bi_next;
4778 raid_bi->bi_next = NULL;
4779 mddev = rdev->mddev;
4780 conf = mddev->private;
4782 rdev_dec_pending(rdev, conf->mddev);
4785 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4788 if (atomic_dec_and_test(&conf->active_aligned_reads))
4789 wake_up(&conf->wait_for_quiescent);
4793 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4795 add_bio_to_retry(raid_bi, conf);
4798 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4800 struct r5conf *conf = mddev->private;
4802 struct bio* align_bi;
4803 struct md_rdev *rdev;
4804 sector_t end_sector;
4806 if (!in_chunk_boundary(mddev, raid_bio)) {
4807 pr_debug("%s: non aligned\n", __func__);
4811 * use bio_clone_mddev to make a copy of the bio
4813 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4817 * set bi_end_io to a new function, and set bi_private to the
4820 align_bi->bi_end_io = raid5_align_endio;
4821 align_bi->bi_private = raid_bio;
4825 align_bi->bi_iter.bi_sector =
4826 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4829 end_sector = bio_end_sector(align_bi);
4831 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4832 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4833 rdev->recovery_offset < end_sector) {
4834 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4836 (test_bit(Faulty, &rdev->flags) ||
4837 !(test_bit(In_sync, &rdev->flags) ||
4838 rdev->recovery_offset >= end_sector)))
4845 atomic_inc(&rdev->nr_pending);
4847 raid_bio->bi_next = (void*)rdev;
4848 align_bi->bi_bdev = rdev->bdev;
4849 bio_clear_flag(align_bi, BIO_SEG_VALID);
4851 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4852 bio_sectors(align_bi),
4853 &first_bad, &bad_sectors)) {
4855 rdev_dec_pending(rdev, mddev);
4859 /* No reshape active, so we can trust rdev->data_offset */
4860 align_bi->bi_iter.bi_sector += rdev->data_offset;
4862 spin_lock_irq(&conf->device_lock);
4863 wait_event_lock_irq(conf->wait_for_quiescent,
4866 atomic_inc(&conf->active_aligned_reads);
4867 spin_unlock_irq(&conf->device_lock);
4870 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4871 align_bi, disk_devt(mddev->gendisk),
4872 raid_bio->bi_iter.bi_sector);
4873 generic_make_request(align_bi);
4882 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4887 sector_t sector = raid_bio->bi_iter.bi_sector;
4888 unsigned chunk_sects = mddev->chunk_sectors;
4889 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4891 if (sectors < bio_sectors(raid_bio)) {
4892 split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4893 bio_chain(split, raid_bio);
4897 if (!raid5_read_one_chunk(mddev, split)) {
4898 if (split != raid_bio)
4899 generic_make_request(raid_bio);
4902 } while (split != raid_bio);
4907 /* __get_priority_stripe - get the next stripe to process
4909 * Full stripe writes are allowed to pass preread active stripes up until
4910 * the bypass_threshold is exceeded. In general the bypass_count
4911 * increments when the handle_list is handled before the hold_list; however, it
4912 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4913 * stripe with in flight i/o. The bypass_count will be reset when the
4914 * head of the hold_list has changed, i.e. the head was promoted to the
4917 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4919 struct stripe_head *sh = NULL, *tmp;
4920 struct list_head *handle_list = NULL;
4921 struct r5worker_group *wg = NULL;
4923 if (conf->worker_cnt_per_group == 0) {
4924 handle_list = &conf->handle_list;
4925 } else if (group != ANY_GROUP) {
4926 handle_list = &conf->worker_groups[group].handle_list;
4927 wg = &conf->worker_groups[group];
4930 for (i = 0; i < conf->group_cnt; i++) {
4931 handle_list = &conf->worker_groups[i].handle_list;
4932 wg = &conf->worker_groups[i];
4933 if (!list_empty(handle_list))
4938 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4940 list_empty(handle_list) ? "empty" : "busy",
4941 list_empty(&conf->hold_list) ? "empty" : "busy",
4942 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4944 if (!list_empty(handle_list)) {
4945 sh = list_entry(handle_list->next, typeof(*sh), lru);
4947 if (list_empty(&conf->hold_list))
4948 conf->bypass_count = 0;
4949 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4950 if (conf->hold_list.next == conf->last_hold)
4951 conf->bypass_count++;
4953 conf->last_hold = conf->hold_list.next;
4954 conf->bypass_count -= conf->bypass_threshold;
4955 if (conf->bypass_count < 0)
4956 conf->bypass_count = 0;
4959 } else if (!list_empty(&conf->hold_list) &&
4960 ((conf->bypass_threshold &&
4961 conf->bypass_count > conf->bypass_threshold) ||
4962 atomic_read(&conf->pending_full_writes) == 0)) {
4964 list_for_each_entry(tmp, &conf->hold_list, lru) {
4965 if (conf->worker_cnt_per_group == 0 ||
4966 group == ANY_GROUP ||
4967 !cpu_online(tmp->cpu) ||
4968 cpu_to_group(tmp->cpu) == group) {
4975 conf->bypass_count -= conf->bypass_threshold;
4976 if (conf->bypass_count < 0)
4977 conf->bypass_count = 0;
4989 list_del_init(&sh->lru);
4990 BUG_ON(atomic_inc_return(&sh->count) != 1);
4994 struct raid5_plug_cb {
4995 struct blk_plug_cb cb;
4996 struct list_head list;
4997 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5000 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5002 struct raid5_plug_cb *cb = container_of(
5003 blk_cb, struct raid5_plug_cb, cb);
5004 struct stripe_head *sh;
5005 struct mddev *mddev = cb->cb.data;
5006 struct r5conf *conf = mddev->private;
5010 if (cb->list.next && !list_empty(&cb->list)) {
5011 spin_lock_irq(&conf->device_lock);
5012 while (!list_empty(&cb->list)) {
5013 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5014 list_del_init(&sh->lru);
5016 * avoid race release_stripe_plug() sees
5017 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5018 * is still in our list
5020 smp_mb__before_atomic();
5021 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5023 * STRIPE_ON_RELEASE_LIST could be set here. In that
5024 * case, the count is always > 1 here
5026 hash = sh->hash_lock_index;
5027 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5030 spin_unlock_irq(&conf->device_lock);
5032 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5033 NR_STRIPE_HASH_LOCKS);
5035 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5039 static void release_stripe_plug(struct mddev *mddev,
5040 struct stripe_head *sh)
5042 struct blk_plug_cb *blk_cb = blk_check_plugged(
5043 raid5_unplug, mddev,
5044 sizeof(struct raid5_plug_cb));
5045 struct raid5_plug_cb *cb;
5048 raid5_release_stripe(sh);
5052 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5054 if (cb->list.next == NULL) {
5056 INIT_LIST_HEAD(&cb->list);
5057 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5058 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5061 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5062 list_add_tail(&sh->lru, &cb->list);
5064 raid5_release_stripe(sh);
5067 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5069 struct r5conf *conf = mddev->private;
5070 sector_t logical_sector, last_sector;
5071 struct stripe_head *sh;
5075 if (mddev->reshape_position != MaxSector)
5076 /* Skip discard while reshape is happening */
5079 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5080 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5083 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5085 stripe_sectors = conf->chunk_sectors *
5086 (conf->raid_disks - conf->max_degraded);
5087 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5089 sector_div(last_sector, stripe_sectors);
5091 logical_sector *= conf->chunk_sectors;
5092 last_sector *= conf->chunk_sectors;
5094 for (; logical_sector < last_sector;
5095 logical_sector += STRIPE_SECTORS) {
5099 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5100 prepare_to_wait(&conf->wait_for_overlap, &w,
5101 TASK_UNINTERRUPTIBLE);
5102 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5103 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5104 raid5_release_stripe(sh);
5108 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5109 spin_lock_irq(&sh->stripe_lock);
5110 for (d = 0; d < conf->raid_disks; d++) {
5111 if (d == sh->pd_idx || d == sh->qd_idx)
5113 if (sh->dev[d].towrite || sh->dev[d].toread) {
5114 set_bit(R5_Overlap, &sh->dev[d].flags);
5115 spin_unlock_irq(&sh->stripe_lock);
5116 raid5_release_stripe(sh);
5121 set_bit(STRIPE_DISCARD, &sh->state);
5122 finish_wait(&conf->wait_for_overlap, &w);
5123 sh->overwrite_disks = 0;
5124 for (d = 0; d < conf->raid_disks; d++) {
5125 if (d == sh->pd_idx || d == sh->qd_idx)
5127 sh->dev[d].towrite = bi;
5128 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5129 raid5_inc_bi_active_stripes(bi);
5130 sh->overwrite_disks++;
5132 spin_unlock_irq(&sh->stripe_lock);
5133 if (conf->mddev->bitmap) {
5135 d < conf->raid_disks - conf->max_degraded;
5137 bitmap_startwrite(mddev->bitmap,
5141 sh->bm_seq = conf->seq_flush + 1;
5142 set_bit(STRIPE_BIT_DELAY, &sh->state);
5145 set_bit(STRIPE_HANDLE, &sh->state);
5146 clear_bit(STRIPE_DELAYED, &sh->state);
5147 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5148 atomic_inc(&conf->preread_active_stripes);
5149 release_stripe_plug(mddev, sh);
5152 remaining = raid5_dec_bi_active_stripes(bi);
5153 if (remaining == 0) {
5154 md_write_end(mddev);
5159 static void raid5_make_request(struct mddev *mddev, struct bio * bi)
5161 struct r5conf *conf = mddev->private;
5163 sector_t new_sector;
5164 sector_t logical_sector, last_sector;
5165 struct stripe_head *sh;
5166 const int rw = bio_data_dir(bi);
5171 if (unlikely(bi->bi_rw & REQ_PREFLUSH)) {
5172 int ret = r5l_handle_flush_request(conf->log, bi);
5176 if (ret == -ENODEV) {
5177 md_flush_request(mddev, bi);
5180 /* ret == -EAGAIN, fallback */
5183 md_write_start(mddev, bi);
5186 * If array is degraded, better not do chunk aligned read because
5187 * later we might have to read it again in order to reconstruct
5188 * data on failed drives.
5190 if (rw == READ && mddev->degraded == 0 &&
5191 mddev->reshape_position == MaxSector) {
5192 bi = chunk_aligned_read(mddev, bi);
5197 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5198 make_discard_request(mddev, bi);
5202 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5203 last_sector = bio_end_sector(bi);
5205 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5207 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5208 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5214 seq = read_seqcount_begin(&conf->gen_lock);
5217 prepare_to_wait(&conf->wait_for_overlap, &w,
5218 TASK_UNINTERRUPTIBLE);
5219 if (unlikely(conf->reshape_progress != MaxSector)) {
5220 /* spinlock is needed as reshape_progress may be
5221 * 64bit on a 32bit platform, and so it might be
5222 * possible to see a half-updated value
5223 * Of course reshape_progress could change after
5224 * the lock is dropped, so once we get a reference
5225 * to the stripe that we think it is, we will have
5228 spin_lock_irq(&conf->device_lock);
5229 if (mddev->reshape_backwards
5230 ? logical_sector < conf->reshape_progress
5231 : logical_sector >= conf->reshape_progress) {
5234 if (mddev->reshape_backwards
5235 ? logical_sector < conf->reshape_safe
5236 : logical_sector >= conf->reshape_safe) {
5237 spin_unlock_irq(&conf->device_lock);
5243 spin_unlock_irq(&conf->device_lock);
5246 new_sector = raid5_compute_sector(conf, logical_sector,
5249 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5250 (unsigned long long)new_sector,
5251 (unsigned long long)logical_sector);
5253 sh = raid5_get_active_stripe(conf, new_sector, previous,
5254 (bi->bi_rw & REQ_RAHEAD), 0);
5256 if (unlikely(previous)) {
5257 /* expansion might have moved on while waiting for a
5258 * stripe, so we must do the range check again.
5259 * Expansion could still move past after this
5260 * test, but as we are holding a reference to
5261 * 'sh', we know that if that happens,
5262 * STRIPE_EXPANDING will get set and the expansion
5263 * won't proceed until we finish with the stripe.
5266 spin_lock_irq(&conf->device_lock);
5267 if (mddev->reshape_backwards
5268 ? logical_sector >= conf->reshape_progress
5269 : logical_sector < conf->reshape_progress)
5270 /* mismatch, need to try again */
5272 spin_unlock_irq(&conf->device_lock);
5274 raid5_release_stripe(sh);
5280 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5281 /* Might have got the wrong stripe_head
5284 raid5_release_stripe(sh);
5289 logical_sector >= mddev->suspend_lo &&
5290 logical_sector < mddev->suspend_hi) {
5291 raid5_release_stripe(sh);
5292 /* As the suspend_* range is controlled by
5293 * userspace, we want an interruptible
5296 flush_signals(current);
5297 prepare_to_wait(&conf->wait_for_overlap,
5298 &w, TASK_INTERRUPTIBLE);
5299 if (logical_sector >= mddev->suspend_lo &&
5300 logical_sector < mddev->suspend_hi) {
5307 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5308 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5309 /* Stripe is busy expanding or
5310 * add failed due to overlap. Flush everything
5313 md_wakeup_thread(mddev->thread);
5314 raid5_release_stripe(sh);
5319 set_bit(STRIPE_HANDLE, &sh->state);
5320 clear_bit(STRIPE_DELAYED, &sh->state);
5321 if ((!sh->batch_head || sh == sh->batch_head) &&
5322 (bi->bi_rw & REQ_SYNC) &&
5323 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5324 atomic_inc(&conf->preread_active_stripes);
5325 release_stripe_plug(mddev, sh);
5327 /* cannot get stripe for read-ahead, just give-up */
5328 bi->bi_error = -EIO;
5332 finish_wait(&conf->wait_for_overlap, &w);
5334 remaining = raid5_dec_bi_active_stripes(bi);
5335 if (remaining == 0) {
5338 md_write_end(mddev);
5340 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5346 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5348 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5350 /* reshaping is quite different to recovery/resync so it is
5351 * handled quite separately ... here.
5353 * On each call to sync_request, we gather one chunk worth of
5354 * destination stripes and flag them as expanding.
5355 * Then we find all the source stripes and request reads.
5356 * As the reads complete, handle_stripe will copy the data
5357 * into the destination stripe and release that stripe.
5359 struct r5conf *conf = mddev->private;
5360 struct stripe_head *sh;
5361 sector_t first_sector, last_sector;
5362 int raid_disks = conf->previous_raid_disks;
5363 int data_disks = raid_disks - conf->max_degraded;
5364 int new_data_disks = conf->raid_disks - conf->max_degraded;
5367 sector_t writepos, readpos, safepos;
5368 sector_t stripe_addr;
5369 int reshape_sectors;
5370 struct list_head stripes;
5373 if (sector_nr == 0) {
5374 /* If restarting in the middle, skip the initial sectors */
5375 if (mddev->reshape_backwards &&
5376 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5377 sector_nr = raid5_size(mddev, 0, 0)
5378 - conf->reshape_progress;
5379 } else if (mddev->reshape_backwards &&
5380 conf->reshape_progress == MaxSector) {
5381 /* shouldn't happen, but just in case, finish up.*/
5382 sector_nr = MaxSector;
5383 } else if (!mddev->reshape_backwards &&
5384 conf->reshape_progress > 0)
5385 sector_nr = conf->reshape_progress;
5386 sector_div(sector_nr, new_data_disks);
5388 mddev->curr_resync_completed = sector_nr;
5389 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5396 /* We need to process a full chunk at a time.
5397 * If old and new chunk sizes differ, we need to process the
5401 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5403 /* We update the metadata at least every 10 seconds, or when
5404 * the data about to be copied would over-write the source of
5405 * the data at the front of the range. i.e. one new_stripe
5406 * along from reshape_progress new_maps to after where
5407 * reshape_safe old_maps to
5409 writepos = conf->reshape_progress;
5410 sector_div(writepos, new_data_disks);
5411 readpos = conf->reshape_progress;
5412 sector_div(readpos, data_disks);
5413 safepos = conf->reshape_safe;
5414 sector_div(safepos, data_disks);
5415 if (mddev->reshape_backwards) {
5416 BUG_ON(writepos < reshape_sectors);
5417 writepos -= reshape_sectors;
5418 readpos += reshape_sectors;
5419 safepos += reshape_sectors;
5421 writepos += reshape_sectors;
5422 /* readpos and safepos are worst-case calculations.
5423 * A negative number is overly pessimistic, and causes
5424 * obvious problems for unsigned storage. So clip to 0.
5426 readpos -= min_t(sector_t, reshape_sectors, readpos);
5427 safepos -= min_t(sector_t, reshape_sectors, safepos);
5430 /* Having calculated the 'writepos' possibly use it
5431 * to set 'stripe_addr' which is where we will write to.
5433 if (mddev->reshape_backwards) {
5434 BUG_ON(conf->reshape_progress == 0);
5435 stripe_addr = writepos;
5436 BUG_ON((mddev->dev_sectors &
5437 ~((sector_t)reshape_sectors - 1))
5438 - reshape_sectors - stripe_addr
5441 BUG_ON(writepos != sector_nr + reshape_sectors);
5442 stripe_addr = sector_nr;
5445 /* 'writepos' is the most advanced device address we might write.
5446 * 'readpos' is the least advanced device address we might read.
5447 * 'safepos' is the least address recorded in the metadata as having
5449 * If there is a min_offset_diff, these are adjusted either by
5450 * increasing the safepos/readpos if diff is negative, or
5451 * increasing writepos if diff is positive.
5452 * If 'readpos' is then behind 'writepos', there is no way that we can
5453 * ensure safety in the face of a crash - that must be done by userspace
5454 * making a backup of the data. So in that case there is no particular
5455 * rush to update metadata.
5456 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5457 * update the metadata to advance 'safepos' to match 'readpos' so that
5458 * we can be safe in the event of a crash.
5459 * So we insist on updating metadata if safepos is behind writepos and
5460 * readpos is beyond writepos.
5461 * In any case, update the metadata every 10 seconds.
5462 * Maybe that number should be configurable, but I'm not sure it is
5463 * worth it.... maybe it could be a multiple of safemode_delay???
5465 if (conf->min_offset_diff < 0) {
5466 safepos += -conf->min_offset_diff;
5467 readpos += -conf->min_offset_diff;
5469 writepos += conf->min_offset_diff;
5471 if ((mddev->reshape_backwards
5472 ? (safepos > writepos && readpos < writepos)
5473 : (safepos < writepos && readpos > writepos)) ||
5474 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5475 /* Cannot proceed until we've updated the superblock... */
5476 wait_event(conf->wait_for_overlap,
5477 atomic_read(&conf->reshape_stripes)==0
5478 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5479 if (atomic_read(&conf->reshape_stripes) != 0)
5481 mddev->reshape_position = conf->reshape_progress;
5482 mddev->curr_resync_completed = sector_nr;
5483 conf->reshape_checkpoint = jiffies;
5484 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5485 md_wakeup_thread(mddev->thread);
5486 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5487 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5488 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5490 spin_lock_irq(&conf->device_lock);
5491 conf->reshape_safe = mddev->reshape_position;
5492 spin_unlock_irq(&conf->device_lock);
5493 wake_up(&conf->wait_for_overlap);
5494 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5497 INIT_LIST_HEAD(&stripes);
5498 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5500 int skipped_disk = 0;
5501 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5502 set_bit(STRIPE_EXPANDING, &sh->state);
5503 atomic_inc(&conf->reshape_stripes);
5504 /* If any of this stripe is beyond the end of the old
5505 * array, then we need to zero those blocks
5507 for (j=sh->disks; j--;) {
5509 if (j == sh->pd_idx)
5511 if (conf->level == 6 &&
5514 s = raid5_compute_blocknr(sh, j, 0);
5515 if (s < raid5_size(mddev, 0, 0)) {
5519 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5520 set_bit(R5_Expanded, &sh->dev[j].flags);
5521 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5523 if (!skipped_disk) {
5524 set_bit(STRIPE_EXPAND_READY, &sh->state);
5525 set_bit(STRIPE_HANDLE, &sh->state);
5527 list_add(&sh->lru, &stripes);
5529 spin_lock_irq(&conf->device_lock);
5530 if (mddev->reshape_backwards)
5531 conf->reshape_progress -= reshape_sectors * new_data_disks;
5533 conf->reshape_progress += reshape_sectors * new_data_disks;
5534 spin_unlock_irq(&conf->device_lock);
5535 /* Ok, those stripe are ready. We can start scheduling
5536 * reads on the source stripes.
5537 * The source stripes are determined by mapping the first and last
5538 * block on the destination stripes.
5541 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5544 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5545 * new_data_disks - 1),
5547 if (last_sector >= mddev->dev_sectors)
5548 last_sector = mddev->dev_sectors - 1;
5549 while (first_sector <= last_sector) {
5550 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5551 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5552 set_bit(STRIPE_HANDLE, &sh->state);
5553 raid5_release_stripe(sh);
5554 first_sector += STRIPE_SECTORS;
5556 /* Now that the sources are clearly marked, we can release
5557 * the destination stripes
5559 while (!list_empty(&stripes)) {
5560 sh = list_entry(stripes.next, struct stripe_head, lru);
5561 list_del_init(&sh->lru);
5562 raid5_release_stripe(sh);
5564 /* If this takes us to the resync_max point where we have to pause,
5565 * then we need to write out the superblock.
5567 sector_nr += reshape_sectors;
5568 retn = reshape_sectors;
5570 if (mddev->curr_resync_completed > mddev->resync_max ||
5571 (sector_nr - mddev->curr_resync_completed) * 2
5572 >= mddev->resync_max - mddev->curr_resync_completed) {
5573 /* Cannot proceed until we've updated the superblock... */
5574 wait_event(conf->wait_for_overlap,
5575 atomic_read(&conf->reshape_stripes) == 0
5576 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5577 if (atomic_read(&conf->reshape_stripes) != 0)
5579 mddev->reshape_position = conf->reshape_progress;
5580 mddev->curr_resync_completed = sector_nr;
5581 conf->reshape_checkpoint = jiffies;
5582 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5583 md_wakeup_thread(mddev->thread);
5584 wait_event(mddev->sb_wait,
5585 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5586 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5587 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5589 spin_lock_irq(&conf->device_lock);
5590 conf->reshape_safe = mddev->reshape_position;
5591 spin_unlock_irq(&conf->device_lock);
5592 wake_up(&conf->wait_for_overlap);
5593 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5599 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
5602 struct r5conf *conf = mddev->private;
5603 struct stripe_head *sh;
5604 sector_t max_sector = mddev->dev_sectors;
5605 sector_t sync_blocks;
5606 int still_degraded = 0;
5609 if (sector_nr >= max_sector) {
5610 /* just being told to finish up .. nothing much to do */
5612 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5617 if (mddev->curr_resync < max_sector) /* aborted */
5618 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5620 else /* completed sync */
5622 bitmap_close_sync(mddev->bitmap);
5627 /* Allow raid5_quiesce to complete */
5628 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5630 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5631 return reshape_request(mddev, sector_nr, skipped);
5633 /* No need to check resync_max as we never do more than one
5634 * stripe, and as resync_max will always be on a chunk boundary,
5635 * if the check in md_do_sync didn't fire, there is no chance
5636 * of overstepping resync_max here
5639 /* if there is too many failed drives and we are trying
5640 * to resync, then assert that we are finished, because there is
5641 * nothing we can do.
5643 if (mddev->degraded >= conf->max_degraded &&
5644 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5645 sector_t rv = mddev->dev_sectors - sector_nr;
5649 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5651 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5652 sync_blocks >= STRIPE_SECTORS) {
5653 /* we can skip this block, and probably more */
5654 sync_blocks /= STRIPE_SECTORS;
5656 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5659 bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5661 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5663 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5664 /* make sure we don't swamp the stripe cache if someone else
5665 * is trying to get access
5667 schedule_timeout_uninterruptible(1);
5669 /* Need to check if array will still be degraded after recovery/resync
5670 * Note in case of > 1 drive failures it's possible we're rebuilding
5671 * one drive while leaving another faulty drive in array.
5674 for (i = 0; i < conf->raid_disks; i++) {
5675 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5677 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5682 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5684 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5685 set_bit(STRIPE_HANDLE, &sh->state);
5687 raid5_release_stripe(sh);
5689 return STRIPE_SECTORS;
5692 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5694 /* We may not be able to submit a whole bio at once as there
5695 * may not be enough stripe_heads available.
5696 * We cannot pre-allocate enough stripe_heads as we may need
5697 * more than exist in the cache (if we allow ever large chunks).
5698 * So we do one stripe head at a time and record in
5699 * ->bi_hw_segments how many have been done.
5701 * We *know* that this entire raid_bio is in one chunk, so
5702 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5704 struct stripe_head *sh;
5706 sector_t sector, logical_sector, last_sector;
5711 logical_sector = raid_bio->bi_iter.bi_sector &
5712 ~((sector_t)STRIPE_SECTORS-1);
5713 sector = raid5_compute_sector(conf, logical_sector,
5715 last_sector = bio_end_sector(raid_bio);
5717 for (; logical_sector < last_sector;
5718 logical_sector += STRIPE_SECTORS,
5719 sector += STRIPE_SECTORS,
5722 if (scnt < raid5_bi_processed_stripes(raid_bio))
5723 /* already done this stripe */
5726 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5729 /* failed to get a stripe - must wait */
5730 raid5_set_bi_processed_stripes(raid_bio, scnt);
5731 conf->retry_read_aligned = raid_bio;
5735 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5736 raid5_release_stripe(sh);
5737 raid5_set_bi_processed_stripes(raid_bio, scnt);
5738 conf->retry_read_aligned = raid_bio;
5742 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5744 raid5_release_stripe(sh);
5747 remaining = raid5_dec_bi_active_stripes(raid_bio);
5748 if (remaining == 0) {
5749 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5751 bio_endio(raid_bio);
5753 if (atomic_dec_and_test(&conf->active_aligned_reads))
5754 wake_up(&conf->wait_for_quiescent);
5758 static int handle_active_stripes(struct r5conf *conf, int group,
5759 struct r5worker *worker,
5760 struct list_head *temp_inactive_list)
5762 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5763 int i, batch_size = 0, hash;
5764 bool release_inactive = false;
5766 while (batch_size < MAX_STRIPE_BATCH &&
5767 (sh = __get_priority_stripe(conf, group)) != NULL)
5768 batch[batch_size++] = sh;
5770 if (batch_size == 0) {
5771 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5772 if (!list_empty(temp_inactive_list + i))
5774 if (i == NR_STRIPE_HASH_LOCKS) {
5775 spin_unlock_irq(&conf->device_lock);
5776 r5l_flush_stripe_to_raid(conf->log);
5777 spin_lock_irq(&conf->device_lock);
5780 release_inactive = true;
5782 spin_unlock_irq(&conf->device_lock);
5784 release_inactive_stripe_list(conf, temp_inactive_list,
5785 NR_STRIPE_HASH_LOCKS);
5787 r5l_flush_stripe_to_raid(conf->log);
5788 if (release_inactive) {
5789 spin_lock_irq(&conf->device_lock);
5793 for (i = 0; i < batch_size; i++)
5794 handle_stripe(batch[i]);
5795 r5l_write_stripe_run(conf->log);
5799 spin_lock_irq(&conf->device_lock);
5800 for (i = 0; i < batch_size; i++) {
5801 hash = batch[i]->hash_lock_index;
5802 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5807 static void raid5_do_work(struct work_struct *work)
5809 struct r5worker *worker = container_of(work, struct r5worker, work);
5810 struct r5worker_group *group = worker->group;
5811 struct r5conf *conf = group->conf;
5812 int group_id = group - conf->worker_groups;
5814 struct blk_plug plug;
5816 pr_debug("+++ raid5worker active\n");
5818 blk_start_plug(&plug);
5820 spin_lock_irq(&conf->device_lock);
5822 int batch_size, released;
5824 released = release_stripe_list(conf, worker->temp_inactive_list);
5826 batch_size = handle_active_stripes(conf, group_id, worker,
5827 worker->temp_inactive_list);
5828 worker->working = false;
5829 if (!batch_size && !released)
5831 handled += batch_size;
5833 pr_debug("%d stripes handled\n", handled);
5835 spin_unlock_irq(&conf->device_lock);
5836 blk_finish_plug(&plug);
5838 pr_debug("--- raid5worker inactive\n");
5842 * This is our raid5 kernel thread.
5844 * We scan the hash table for stripes which can be handled now.
5845 * During the scan, completed stripes are saved for us by the interrupt
5846 * handler, so that they will not have to wait for our next wakeup.
5848 static void raid5d(struct md_thread *thread)
5850 struct mddev *mddev = thread->mddev;
5851 struct r5conf *conf = mddev->private;
5853 struct blk_plug plug;
5855 pr_debug("+++ raid5d active\n");
5857 md_check_recovery(mddev);
5859 if (!bio_list_empty(&conf->return_bi) &&
5860 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5861 struct bio_list tmp = BIO_EMPTY_LIST;
5862 spin_lock_irq(&conf->device_lock);
5863 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5864 bio_list_merge(&tmp, &conf->return_bi);
5865 bio_list_init(&conf->return_bi);
5867 spin_unlock_irq(&conf->device_lock);
5871 blk_start_plug(&plug);
5873 spin_lock_irq(&conf->device_lock);
5876 int batch_size, released;
5878 released = release_stripe_list(conf, conf->temp_inactive_list);
5880 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5883 !list_empty(&conf->bitmap_list)) {
5884 /* Now is a good time to flush some bitmap updates */
5886 spin_unlock_irq(&conf->device_lock);
5887 bitmap_unplug(mddev->bitmap);
5888 spin_lock_irq(&conf->device_lock);
5889 conf->seq_write = conf->seq_flush;
5890 activate_bit_delay(conf, conf->temp_inactive_list);
5892 raid5_activate_delayed(conf);
5894 while ((bio = remove_bio_from_retry(conf))) {
5896 spin_unlock_irq(&conf->device_lock);
5897 ok = retry_aligned_read(conf, bio);
5898 spin_lock_irq(&conf->device_lock);
5904 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5905 conf->temp_inactive_list);
5906 if (!batch_size && !released)
5908 handled += batch_size;
5910 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5911 spin_unlock_irq(&conf->device_lock);
5912 md_check_recovery(mddev);
5913 spin_lock_irq(&conf->device_lock);
5916 pr_debug("%d stripes handled\n", handled);
5918 spin_unlock_irq(&conf->device_lock);
5919 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5920 mutex_trylock(&conf->cache_size_mutex)) {
5921 grow_one_stripe(conf, __GFP_NOWARN);
5922 /* Set flag even if allocation failed. This helps
5923 * slow down allocation requests when mem is short
5925 set_bit(R5_DID_ALLOC, &conf->cache_state);
5926 mutex_unlock(&conf->cache_size_mutex);
5929 r5l_flush_stripe_to_raid(conf->log);
5931 async_tx_issue_pending_all();
5932 blk_finish_plug(&plug);
5934 pr_debug("--- raid5d inactive\n");
5938 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5940 struct r5conf *conf;
5942 spin_lock(&mddev->lock);
5943 conf = mddev->private;
5945 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5946 spin_unlock(&mddev->lock);
5951 raid5_set_cache_size(struct mddev *mddev, int size)
5953 struct r5conf *conf = mddev->private;
5956 if (size <= 16 || size > 32768)
5959 conf->min_nr_stripes = size;
5960 mutex_lock(&conf->cache_size_mutex);
5961 while (size < conf->max_nr_stripes &&
5962 drop_one_stripe(conf))
5964 mutex_unlock(&conf->cache_size_mutex);
5967 err = md_allow_write(mddev);
5971 mutex_lock(&conf->cache_size_mutex);
5972 while (size > conf->max_nr_stripes)
5973 if (!grow_one_stripe(conf, GFP_KERNEL))
5975 mutex_unlock(&conf->cache_size_mutex);
5979 EXPORT_SYMBOL(raid5_set_cache_size);
5982 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5984 struct r5conf *conf;
5988 if (len >= PAGE_SIZE)
5990 if (kstrtoul(page, 10, &new))
5992 err = mddev_lock(mddev);
5995 conf = mddev->private;
5999 err = raid5_set_cache_size(mddev, new);
6000 mddev_unlock(mddev);
6005 static struct md_sysfs_entry
6006 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6007 raid5_show_stripe_cache_size,
6008 raid5_store_stripe_cache_size);
6011 raid5_show_rmw_level(struct mddev *mddev, char *page)
6013 struct r5conf *conf = mddev->private;
6015 return sprintf(page, "%d\n", conf->rmw_level);
6021 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6023 struct r5conf *conf = mddev->private;
6029 if (len >= PAGE_SIZE)
6032 if (kstrtoul(page, 10, &new))
6035 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6038 if (new != PARITY_DISABLE_RMW &&
6039 new != PARITY_ENABLE_RMW &&
6040 new != PARITY_PREFER_RMW)
6043 conf->rmw_level = new;
6047 static struct md_sysfs_entry
6048 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6049 raid5_show_rmw_level,
6050 raid5_store_rmw_level);
6054 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6056 struct r5conf *conf;
6058 spin_lock(&mddev->lock);
6059 conf = mddev->private;
6061 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6062 spin_unlock(&mddev->lock);
6067 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6069 struct r5conf *conf;
6073 if (len >= PAGE_SIZE)
6075 if (kstrtoul(page, 10, &new))
6078 err = mddev_lock(mddev);
6081 conf = mddev->private;
6084 else if (new > conf->min_nr_stripes)
6087 conf->bypass_threshold = new;
6088 mddev_unlock(mddev);
6092 static struct md_sysfs_entry
6093 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6095 raid5_show_preread_threshold,
6096 raid5_store_preread_threshold);
6099 raid5_show_skip_copy(struct mddev *mddev, char *page)
6101 struct r5conf *conf;
6103 spin_lock(&mddev->lock);
6104 conf = mddev->private;
6106 ret = sprintf(page, "%d\n", conf->skip_copy);
6107 spin_unlock(&mddev->lock);
6112 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6114 struct r5conf *conf;
6118 if (len >= PAGE_SIZE)
6120 if (kstrtoul(page, 10, &new))
6124 err = mddev_lock(mddev);
6127 conf = mddev->private;
6130 else if (new != conf->skip_copy) {
6131 mddev_suspend(mddev);
6132 conf->skip_copy = new;
6134 mddev->queue->backing_dev_info.capabilities |=
6135 BDI_CAP_STABLE_WRITES;
6137 mddev->queue->backing_dev_info.capabilities &=
6138 ~BDI_CAP_STABLE_WRITES;
6139 mddev_resume(mddev);
6141 mddev_unlock(mddev);
6145 static struct md_sysfs_entry
6146 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6147 raid5_show_skip_copy,
6148 raid5_store_skip_copy);
6151 stripe_cache_active_show(struct mddev *mddev, char *page)
6153 struct r5conf *conf = mddev->private;
6155 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6160 static struct md_sysfs_entry
6161 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6164 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6166 struct r5conf *conf;
6168 spin_lock(&mddev->lock);
6169 conf = mddev->private;
6171 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6172 spin_unlock(&mddev->lock);
6176 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6178 int *worker_cnt_per_group,
6179 struct r5worker_group **worker_groups);
6181 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6183 struct r5conf *conf;
6186 struct r5worker_group *new_groups, *old_groups;
6187 int group_cnt, worker_cnt_per_group;
6189 if (len >= PAGE_SIZE)
6191 if (kstrtoul(page, 10, &new))
6194 err = mddev_lock(mddev);
6197 conf = mddev->private;
6200 else if (new != conf->worker_cnt_per_group) {
6201 mddev_suspend(mddev);
6203 old_groups = conf->worker_groups;
6205 flush_workqueue(raid5_wq);
6207 err = alloc_thread_groups(conf, new,
6208 &group_cnt, &worker_cnt_per_group,
6211 spin_lock_irq(&conf->device_lock);
6212 conf->group_cnt = group_cnt;
6213 conf->worker_cnt_per_group = worker_cnt_per_group;
6214 conf->worker_groups = new_groups;
6215 spin_unlock_irq(&conf->device_lock);
6218 kfree(old_groups[0].workers);
6221 mddev_resume(mddev);
6223 mddev_unlock(mddev);
6228 static struct md_sysfs_entry
6229 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6230 raid5_show_group_thread_cnt,
6231 raid5_store_group_thread_cnt);
6233 static struct attribute *raid5_attrs[] = {
6234 &raid5_stripecache_size.attr,
6235 &raid5_stripecache_active.attr,
6236 &raid5_preread_bypass_threshold.attr,
6237 &raid5_group_thread_cnt.attr,
6238 &raid5_skip_copy.attr,
6239 &raid5_rmw_level.attr,
6242 static struct attribute_group raid5_attrs_group = {
6244 .attrs = raid5_attrs,
6247 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6249 int *worker_cnt_per_group,
6250 struct r5worker_group **worker_groups)
6254 struct r5worker *workers;
6256 *worker_cnt_per_group = cnt;
6259 *worker_groups = NULL;
6262 *group_cnt = num_possible_nodes();
6263 size = sizeof(struct r5worker) * cnt;
6264 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6265 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6266 *group_cnt, GFP_NOIO);
6267 if (!*worker_groups || !workers) {
6269 kfree(*worker_groups);
6273 for (i = 0; i < *group_cnt; i++) {
6274 struct r5worker_group *group;
6276 group = &(*worker_groups)[i];
6277 INIT_LIST_HEAD(&group->handle_list);
6279 group->workers = workers + i * cnt;
6281 for (j = 0; j < cnt; j++) {
6282 struct r5worker *worker = group->workers + j;
6283 worker->group = group;
6284 INIT_WORK(&worker->work, raid5_do_work);
6286 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6287 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6294 static void free_thread_groups(struct r5conf *conf)
6296 if (conf->worker_groups)
6297 kfree(conf->worker_groups[0].workers);
6298 kfree(conf->worker_groups);
6299 conf->worker_groups = NULL;
6303 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6305 struct r5conf *conf = mddev->private;
6308 sectors = mddev->dev_sectors;
6310 /* size is defined by the smallest of previous and new size */
6311 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6313 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6314 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6315 return sectors * (raid_disks - conf->max_degraded);
6318 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6320 safe_put_page(percpu->spare_page);
6321 if (percpu->scribble)
6322 flex_array_free(percpu->scribble);
6323 percpu->spare_page = NULL;
6324 percpu->scribble = NULL;
6327 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6329 if (conf->level == 6 && !percpu->spare_page)
6330 percpu->spare_page = alloc_page(GFP_KERNEL);
6331 if (!percpu->scribble)
6332 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6333 conf->previous_raid_disks),
6334 max(conf->chunk_sectors,
6335 conf->prev_chunk_sectors)
6339 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6340 free_scratch_buffer(conf, percpu);
6347 static void raid5_free_percpu(struct r5conf *conf)
6354 #ifdef CONFIG_HOTPLUG_CPU
6355 unregister_cpu_notifier(&conf->cpu_notify);
6359 for_each_possible_cpu(cpu)
6360 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6363 free_percpu(conf->percpu);
6366 static void free_conf(struct r5conf *conf)
6369 r5l_exit_log(conf->log);
6370 if (conf->shrinker.seeks)
6371 unregister_shrinker(&conf->shrinker);
6373 free_thread_groups(conf);
6374 shrink_stripes(conf);
6375 raid5_free_percpu(conf);
6377 kfree(conf->stripe_hashtbl);
6381 #ifdef CONFIG_HOTPLUG_CPU
6382 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6385 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6386 long cpu = (long)hcpu;
6387 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6390 case CPU_UP_PREPARE:
6391 case CPU_UP_PREPARE_FROZEN:
6392 if (alloc_scratch_buffer(conf, percpu)) {
6393 pr_err("%s: failed memory allocation for cpu%ld\n",
6395 return notifier_from_errno(-ENOMEM);
6399 case CPU_DEAD_FROZEN:
6400 case CPU_UP_CANCELED:
6401 case CPU_UP_CANCELED_FROZEN:
6402 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6411 static int raid5_alloc_percpu(struct r5conf *conf)
6416 conf->percpu = alloc_percpu(struct raid5_percpu);
6420 #ifdef CONFIG_HOTPLUG_CPU
6421 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6422 conf->cpu_notify.priority = 0;
6423 err = register_cpu_notifier(&conf->cpu_notify);
6429 for_each_present_cpu(cpu) {
6430 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6432 pr_err("%s: failed memory allocation for cpu%ld\n",
6440 conf->scribble_disks = max(conf->raid_disks,
6441 conf->previous_raid_disks);
6442 conf->scribble_sectors = max(conf->chunk_sectors,
6443 conf->prev_chunk_sectors);
6448 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6449 struct shrink_control *sc)
6451 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6452 unsigned long ret = SHRINK_STOP;
6454 if (mutex_trylock(&conf->cache_size_mutex)) {
6456 while (ret < sc->nr_to_scan &&
6457 conf->max_nr_stripes > conf->min_nr_stripes) {
6458 if (drop_one_stripe(conf) == 0) {
6464 mutex_unlock(&conf->cache_size_mutex);
6469 static unsigned long raid5_cache_count(struct shrinker *shrink,
6470 struct shrink_control *sc)
6472 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6474 if (conf->max_nr_stripes < conf->min_nr_stripes)
6475 /* unlikely, but not impossible */
6477 return conf->max_nr_stripes - conf->min_nr_stripes;
6480 static struct r5conf *setup_conf(struct mddev *mddev)
6482 struct r5conf *conf;
6483 int raid_disk, memory, max_disks;
6484 struct md_rdev *rdev;
6485 struct disk_info *disk;
6488 int group_cnt, worker_cnt_per_group;
6489 struct r5worker_group *new_group;
6491 if (mddev->new_level != 5
6492 && mddev->new_level != 4
6493 && mddev->new_level != 6) {
6494 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6495 mdname(mddev), mddev->new_level);
6496 return ERR_PTR(-EIO);
6498 if ((mddev->new_level == 5
6499 && !algorithm_valid_raid5(mddev->new_layout)) ||
6500 (mddev->new_level == 6
6501 && !algorithm_valid_raid6(mddev->new_layout))) {
6502 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6503 mdname(mddev), mddev->new_layout);
6504 return ERR_PTR(-EIO);
6506 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6507 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6508 mdname(mddev), mddev->raid_disks);
6509 return ERR_PTR(-EINVAL);
6512 if (!mddev->new_chunk_sectors ||
6513 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6514 !is_power_of_2(mddev->new_chunk_sectors)) {
6515 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6516 mdname(mddev), mddev->new_chunk_sectors << 9);
6517 return ERR_PTR(-EINVAL);
6520 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6523 /* Don't enable multi-threading by default*/
6524 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6526 conf->group_cnt = group_cnt;
6527 conf->worker_cnt_per_group = worker_cnt_per_group;
6528 conf->worker_groups = new_group;
6531 spin_lock_init(&conf->device_lock);
6532 seqcount_init(&conf->gen_lock);
6533 mutex_init(&conf->cache_size_mutex);
6534 init_waitqueue_head(&conf->wait_for_quiescent);
6535 init_waitqueue_head(&conf->wait_for_stripe);
6536 init_waitqueue_head(&conf->wait_for_overlap);
6537 INIT_LIST_HEAD(&conf->handle_list);
6538 INIT_LIST_HEAD(&conf->hold_list);
6539 INIT_LIST_HEAD(&conf->delayed_list);
6540 INIT_LIST_HEAD(&conf->bitmap_list);
6541 bio_list_init(&conf->return_bi);
6542 init_llist_head(&conf->released_stripes);
6543 atomic_set(&conf->active_stripes, 0);
6544 atomic_set(&conf->preread_active_stripes, 0);
6545 atomic_set(&conf->active_aligned_reads, 0);
6546 conf->bypass_threshold = BYPASS_THRESHOLD;
6547 conf->recovery_disabled = mddev->recovery_disabled - 1;
6549 conf->raid_disks = mddev->raid_disks;
6550 if (mddev->reshape_position == MaxSector)
6551 conf->previous_raid_disks = mddev->raid_disks;
6553 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6554 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6556 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6561 conf->mddev = mddev;
6563 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6566 /* We init hash_locks[0] separately to that it can be used
6567 * as the reference lock in the spin_lock_nest_lock() call
6568 * in lock_all_device_hash_locks_irq in order to convince
6569 * lockdep that we know what we are doing.
6571 spin_lock_init(conf->hash_locks);
6572 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6573 spin_lock_init(conf->hash_locks + i);
6575 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6576 INIT_LIST_HEAD(conf->inactive_list + i);
6578 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6579 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6581 conf->level = mddev->new_level;
6582 conf->chunk_sectors = mddev->new_chunk_sectors;
6583 if (raid5_alloc_percpu(conf) != 0)
6586 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6588 rdev_for_each(rdev, mddev) {
6589 raid_disk = rdev->raid_disk;
6590 if (raid_disk >= max_disks
6591 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6593 disk = conf->disks + raid_disk;
6595 if (test_bit(Replacement, &rdev->flags)) {
6596 if (disk->replacement)
6598 disk->replacement = rdev;
6605 if (test_bit(In_sync, &rdev->flags)) {
6606 char b[BDEVNAME_SIZE];
6607 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6609 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6610 } else if (rdev->saved_raid_disk != raid_disk)
6611 /* Cannot rely on bitmap to complete recovery */
6615 conf->level = mddev->new_level;
6616 if (conf->level == 6) {
6617 conf->max_degraded = 2;
6618 if (raid6_call.xor_syndrome)
6619 conf->rmw_level = PARITY_ENABLE_RMW;
6621 conf->rmw_level = PARITY_DISABLE_RMW;
6623 conf->max_degraded = 1;
6624 conf->rmw_level = PARITY_ENABLE_RMW;
6626 conf->algorithm = mddev->new_layout;
6627 conf->reshape_progress = mddev->reshape_position;
6628 if (conf->reshape_progress != MaxSector) {
6629 conf->prev_chunk_sectors = mddev->chunk_sectors;
6630 conf->prev_algo = mddev->layout;
6632 conf->prev_chunk_sectors = conf->chunk_sectors;
6633 conf->prev_algo = conf->algorithm;
6636 conf->min_nr_stripes = NR_STRIPES;
6637 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6638 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6639 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6640 if (grow_stripes(conf, conf->min_nr_stripes)) {
6642 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6643 mdname(mddev), memory);
6646 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6647 mdname(mddev), memory);
6649 * Losing a stripe head costs more than the time to refill it,
6650 * it reduces the queue depth and so can hurt throughput.
6651 * So set it rather large, scaled by number of devices.
6653 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6654 conf->shrinker.scan_objects = raid5_cache_scan;
6655 conf->shrinker.count_objects = raid5_cache_count;
6656 conf->shrinker.batch = 128;
6657 conf->shrinker.flags = 0;
6658 register_shrinker(&conf->shrinker);
6660 sprintf(pers_name, "raid%d", mddev->new_level);
6661 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6662 if (!conf->thread) {
6664 "md/raid:%s: couldn't allocate thread.\n",
6674 return ERR_PTR(-EIO);
6676 return ERR_PTR(-ENOMEM);
6679 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6682 case ALGORITHM_PARITY_0:
6683 if (raid_disk < max_degraded)
6686 case ALGORITHM_PARITY_N:
6687 if (raid_disk >= raid_disks - max_degraded)
6690 case ALGORITHM_PARITY_0_6:
6691 if (raid_disk == 0 ||
6692 raid_disk == raid_disks - 1)
6695 case ALGORITHM_LEFT_ASYMMETRIC_6:
6696 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6697 case ALGORITHM_LEFT_SYMMETRIC_6:
6698 case ALGORITHM_RIGHT_SYMMETRIC_6:
6699 if (raid_disk == raid_disks - 1)
6705 static int raid5_run(struct mddev *mddev)
6707 struct r5conf *conf;
6708 int working_disks = 0;
6709 int dirty_parity_disks = 0;
6710 struct md_rdev *rdev;
6711 struct md_rdev *journal_dev = NULL;
6712 sector_t reshape_offset = 0;
6714 long long min_offset_diff = 0;
6717 if (mddev->recovery_cp != MaxSector)
6718 printk(KERN_NOTICE "md/raid:%s: not clean"
6719 " -- starting background reconstruction\n",
6722 rdev_for_each(rdev, mddev) {
6725 if (test_bit(Journal, &rdev->flags)) {
6729 if (rdev->raid_disk < 0)
6731 diff = (rdev->new_data_offset - rdev->data_offset);
6733 min_offset_diff = diff;
6735 } else if (mddev->reshape_backwards &&
6736 diff < min_offset_diff)
6737 min_offset_diff = diff;
6738 else if (!mddev->reshape_backwards &&
6739 diff > min_offset_diff)
6740 min_offset_diff = diff;
6743 if (mddev->reshape_position != MaxSector) {
6744 /* Check that we can continue the reshape.
6745 * Difficulties arise if the stripe we would write to
6746 * next is at or after the stripe we would read from next.
6747 * For a reshape that changes the number of devices, this
6748 * is only possible for a very short time, and mdadm makes
6749 * sure that time appears to have past before assembling
6750 * the array. So we fail if that time hasn't passed.
6751 * For a reshape that keeps the number of devices the same
6752 * mdadm must be monitoring the reshape can keeping the
6753 * critical areas read-only and backed up. It will start
6754 * the array in read-only mode, so we check for that.
6756 sector_t here_new, here_old;
6758 int max_degraded = (mddev->level == 6 ? 2 : 1);
6763 printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6768 if (mddev->new_level != mddev->level) {
6769 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6770 "required - aborting.\n",
6774 old_disks = mddev->raid_disks - mddev->delta_disks;
6775 /* reshape_position must be on a new-stripe boundary, and one
6776 * further up in new geometry must map after here in old
6778 * If the chunk sizes are different, then as we perform reshape
6779 * in units of the largest of the two, reshape_position needs
6780 * be a multiple of the largest chunk size times new data disks.
6782 here_new = mddev->reshape_position;
6783 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6784 new_data_disks = mddev->raid_disks - max_degraded;
6785 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6786 printk(KERN_ERR "md/raid:%s: reshape_position not "
6787 "on a stripe boundary\n", mdname(mddev));
6790 reshape_offset = here_new * chunk_sectors;
6791 /* here_new is the stripe we will write to */
6792 here_old = mddev->reshape_position;
6793 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6794 /* here_old is the first stripe that we might need to read
6796 if (mddev->delta_disks == 0) {
6797 /* We cannot be sure it is safe to start an in-place
6798 * reshape. It is only safe if user-space is monitoring
6799 * and taking constant backups.
6800 * mdadm always starts a situation like this in
6801 * readonly mode so it can take control before
6802 * allowing any writes. So just check for that.
6804 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6805 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6806 /* not really in-place - so OK */;
6807 else if (mddev->ro == 0) {
6808 printk(KERN_ERR "md/raid:%s: in-place reshape "
6809 "must be started in read-only mode "
6814 } else if (mddev->reshape_backwards
6815 ? (here_new * chunk_sectors + min_offset_diff <=
6816 here_old * chunk_sectors)
6817 : (here_new * chunk_sectors >=
6818 here_old * chunk_sectors + (-min_offset_diff))) {
6819 /* Reading from the same stripe as writing to - bad */
6820 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6821 "auto-recovery - aborting.\n",
6825 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6827 /* OK, we should be able to continue; */
6829 BUG_ON(mddev->level != mddev->new_level);
6830 BUG_ON(mddev->layout != mddev->new_layout);
6831 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6832 BUG_ON(mddev->delta_disks != 0);
6835 if (mddev->private == NULL)
6836 conf = setup_conf(mddev);
6838 conf = mddev->private;
6841 return PTR_ERR(conf);
6843 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) && !journal_dev) {
6844 printk(KERN_ERR "md/raid:%s: journal disk is missing, force array readonly\n",
6847 set_disk_ro(mddev->gendisk, 1);
6850 conf->min_offset_diff = min_offset_diff;
6851 mddev->thread = conf->thread;
6852 conf->thread = NULL;
6853 mddev->private = conf;
6855 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6857 rdev = conf->disks[i].rdev;
6858 if (!rdev && conf->disks[i].replacement) {
6859 /* The replacement is all we have yet */
6860 rdev = conf->disks[i].replacement;
6861 conf->disks[i].replacement = NULL;
6862 clear_bit(Replacement, &rdev->flags);
6863 conf->disks[i].rdev = rdev;
6867 if (conf->disks[i].replacement &&
6868 conf->reshape_progress != MaxSector) {
6869 /* replacements and reshape simply do not mix. */
6870 printk(KERN_ERR "md: cannot handle concurrent "
6871 "replacement and reshape.\n");
6874 if (test_bit(In_sync, &rdev->flags)) {
6878 /* This disc is not fully in-sync. However if it
6879 * just stored parity (beyond the recovery_offset),
6880 * when we don't need to be concerned about the
6881 * array being dirty.
6882 * When reshape goes 'backwards', we never have
6883 * partially completed devices, so we only need
6884 * to worry about reshape going forwards.
6886 /* Hack because v0.91 doesn't store recovery_offset properly. */
6887 if (mddev->major_version == 0 &&
6888 mddev->minor_version > 90)
6889 rdev->recovery_offset = reshape_offset;
6891 if (rdev->recovery_offset < reshape_offset) {
6892 /* We need to check old and new layout */
6893 if (!only_parity(rdev->raid_disk,
6896 conf->max_degraded))
6899 if (!only_parity(rdev->raid_disk,
6901 conf->previous_raid_disks,
6902 conf->max_degraded))
6904 dirty_parity_disks++;
6908 * 0 for a fully functional array, 1 or 2 for a degraded array.
6910 mddev->degraded = calc_degraded(conf);
6912 if (has_failed(conf)) {
6913 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6914 " (%d/%d failed)\n",
6915 mdname(mddev), mddev->degraded, conf->raid_disks);
6919 /* device size must be a multiple of chunk size */
6920 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6921 mddev->resync_max_sectors = mddev->dev_sectors;
6923 if (mddev->degraded > dirty_parity_disks &&
6924 mddev->recovery_cp != MaxSector) {
6925 if (mddev->ok_start_degraded)
6927 "md/raid:%s: starting dirty degraded array"
6928 " - data corruption possible.\n",
6932 "md/raid:%s: cannot start dirty degraded array.\n",
6938 if (mddev->degraded == 0)
6939 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6940 " devices, algorithm %d\n", mdname(mddev), conf->level,
6941 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6944 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6945 " out of %d devices, algorithm %d\n",
6946 mdname(mddev), conf->level,
6947 mddev->raid_disks - mddev->degraded,
6948 mddev->raid_disks, mddev->new_layout);
6950 print_raid5_conf(conf);
6952 if (conf->reshape_progress != MaxSector) {
6953 conf->reshape_safe = conf->reshape_progress;
6954 atomic_set(&conf->reshape_stripes, 0);
6955 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6956 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6957 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6958 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6959 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6963 /* Ok, everything is just fine now */
6964 if (mddev->to_remove == &raid5_attrs_group)
6965 mddev->to_remove = NULL;
6966 else if (mddev->kobj.sd &&
6967 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6969 "raid5: failed to create sysfs attributes for %s\n",
6971 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6975 bool discard_supported = true;
6976 /* read-ahead size must cover two whole stripes, which
6977 * is 2 * (datadisks) * chunksize where 'n' is the
6978 * number of raid devices
6980 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6981 int stripe = data_disks *
6982 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6983 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6984 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6986 chunk_size = mddev->chunk_sectors << 9;
6987 blk_queue_io_min(mddev->queue, chunk_size);
6988 blk_queue_io_opt(mddev->queue, chunk_size *
6989 (conf->raid_disks - conf->max_degraded));
6990 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6992 * We can only discard a whole stripe. It doesn't make sense to
6993 * discard data disk but write parity disk
6995 stripe = stripe * PAGE_SIZE;
6996 /* Round up to power of 2, as discard handling
6997 * currently assumes that */
6998 while ((stripe-1) & stripe)
6999 stripe = (stripe | (stripe-1)) + 1;
7000 mddev->queue->limits.discard_alignment = stripe;
7001 mddev->queue->limits.discard_granularity = stripe;
7003 * unaligned part of discard request will be ignored, so can't
7004 * guarantee discard_zeroes_data
7006 mddev->queue->limits.discard_zeroes_data = 0;
7008 blk_queue_max_write_same_sectors(mddev->queue, 0);
7010 rdev_for_each(rdev, mddev) {
7011 disk_stack_limits(mddev->gendisk, rdev->bdev,
7012 rdev->data_offset << 9);
7013 disk_stack_limits(mddev->gendisk, rdev->bdev,
7014 rdev->new_data_offset << 9);
7016 * discard_zeroes_data is required, otherwise data
7017 * could be lost. Consider a scenario: discard a stripe
7018 * (the stripe could be inconsistent if
7019 * discard_zeroes_data is 0); write one disk of the
7020 * stripe (the stripe could be inconsistent again
7021 * depending on which disks are used to calculate
7022 * parity); the disk is broken; The stripe data of this
7025 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7026 !bdev_get_queue(rdev->bdev)->
7027 limits.discard_zeroes_data)
7028 discard_supported = false;
7029 /* Unfortunately, discard_zeroes_data is not currently
7030 * a guarantee - just a hint. So we only allow DISCARD
7031 * if the sysadmin has confirmed that only safe devices
7032 * are in use by setting a module parameter.
7034 if (!devices_handle_discard_safely) {
7035 if (discard_supported) {
7036 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7037 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7039 discard_supported = false;
7043 if (discard_supported &&
7044 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7045 mddev->queue->limits.discard_granularity >= stripe)
7046 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7049 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7054 char b[BDEVNAME_SIZE];
7056 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7057 mdname(mddev), bdevname(journal_dev->bdev, b));
7058 r5l_init_log(conf, journal_dev);
7063 md_unregister_thread(&mddev->thread);
7064 print_raid5_conf(conf);
7066 mddev->private = NULL;
7067 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7071 static void raid5_free(struct mddev *mddev, void *priv)
7073 struct r5conf *conf = priv;
7076 mddev->to_remove = &raid5_attrs_group;
7079 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7081 struct r5conf *conf = mddev->private;
7084 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7085 conf->chunk_sectors / 2, mddev->layout);
7086 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7088 for (i = 0; i < conf->raid_disks; i++) {
7089 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7090 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7093 seq_printf (seq, "]");
7096 static void print_raid5_conf (struct r5conf *conf)
7099 struct disk_info *tmp;
7101 printk(KERN_DEBUG "RAID conf printout:\n");
7103 printk("(conf==NULL)\n");
7106 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7108 conf->raid_disks - conf->mddev->degraded);
7110 for (i = 0; i < conf->raid_disks; i++) {
7111 char b[BDEVNAME_SIZE];
7112 tmp = conf->disks + i;
7114 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7115 i, !test_bit(Faulty, &tmp->rdev->flags),
7116 bdevname(tmp->rdev->bdev, b));
7120 static int raid5_spare_active(struct mddev *mddev)
7123 struct r5conf *conf = mddev->private;
7124 struct disk_info *tmp;
7126 unsigned long flags;
7128 for (i = 0; i < conf->raid_disks; i++) {
7129 tmp = conf->disks + i;
7130 if (tmp->replacement
7131 && tmp->replacement->recovery_offset == MaxSector
7132 && !test_bit(Faulty, &tmp->replacement->flags)
7133 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7134 /* Replacement has just become active. */
7136 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7139 /* Replaced device not technically faulty,
7140 * but we need to be sure it gets removed
7141 * and never re-added.
7143 set_bit(Faulty, &tmp->rdev->flags);
7144 sysfs_notify_dirent_safe(
7145 tmp->rdev->sysfs_state);
7147 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7148 } else if (tmp->rdev
7149 && tmp->rdev->recovery_offset == MaxSector
7150 && !test_bit(Faulty, &tmp->rdev->flags)
7151 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7153 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7156 spin_lock_irqsave(&conf->device_lock, flags);
7157 mddev->degraded = calc_degraded(conf);
7158 spin_unlock_irqrestore(&conf->device_lock, flags);
7159 print_raid5_conf(conf);
7163 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7165 struct r5conf *conf = mddev->private;
7167 int number = rdev->raid_disk;
7168 struct md_rdev **rdevp;
7169 struct disk_info *p = conf->disks + number;
7171 print_raid5_conf(conf);
7172 if (test_bit(Journal, &rdev->flags) && conf->log) {
7173 struct r5l_log *log;
7175 * we can't wait pending write here, as this is called in
7176 * raid5d, wait will deadlock.
7178 if (atomic_read(&mddev->writes_pending))
7186 if (rdev == p->rdev)
7188 else if (rdev == p->replacement)
7189 rdevp = &p->replacement;
7193 if (number >= conf->raid_disks &&
7194 conf->reshape_progress == MaxSector)
7195 clear_bit(In_sync, &rdev->flags);
7197 if (test_bit(In_sync, &rdev->flags) ||
7198 atomic_read(&rdev->nr_pending)) {
7202 /* Only remove non-faulty devices if recovery
7205 if (!test_bit(Faulty, &rdev->flags) &&
7206 mddev->recovery_disabled != conf->recovery_disabled &&
7207 !has_failed(conf) &&
7208 (!p->replacement || p->replacement == rdev) &&
7209 number < conf->raid_disks) {
7214 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7216 if (atomic_read(&rdev->nr_pending)) {
7217 /* lost the race, try later */
7222 if (p->replacement) {
7223 /* We must have just cleared 'rdev' */
7224 p->rdev = p->replacement;
7225 clear_bit(Replacement, &p->replacement->flags);
7226 smp_mb(); /* Make sure other CPUs may see both as identical
7227 * but will never see neither - if they are careful
7229 p->replacement = NULL;
7230 clear_bit(WantReplacement, &rdev->flags);
7232 /* We might have just removed the Replacement as faulty-
7233 * clear the bit just in case
7235 clear_bit(WantReplacement, &rdev->flags);
7238 print_raid5_conf(conf);
7242 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7244 struct r5conf *conf = mddev->private;
7247 struct disk_info *p;
7249 int last = conf->raid_disks - 1;
7251 if (test_bit(Journal, &rdev->flags)) {
7252 char b[BDEVNAME_SIZE];
7256 rdev->raid_disk = 0;
7258 * The array is in readonly mode if journal is missing, so no
7259 * write requests running. We should be safe
7261 r5l_init_log(conf, rdev);
7262 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7263 mdname(mddev), bdevname(rdev->bdev, b));
7266 if (mddev->recovery_disabled == conf->recovery_disabled)
7269 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7270 /* no point adding a device */
7273 if (rdev->raid_disk >= 0)
7274 first = last = rdev->raid_disk;
7277 * find the disk ... but prefer rdev->saved_raid_disk
7280 if (rdev->saved_raid_disk >= 0 &&
7281 rdev->saved_raid_disk >= first &&
7282 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7283 first = rdev->saved_raid_disk;
7285 for (disk = first; disk <= last; disk++) {
7286 p = conf->disks + disk;
7287 if (p->rdev == NULL) {
7288 clear_bit(In_sync, &rdev->flags);
7289 rdev->raid_disk = disk;
7291 if (rdev->saved_raid_disk != disk)
7293 rcu_assign_pointer(p->rdev, rdev);
7297 for (disk = first; disk <= last; disk++) {
7298 p = conf->disks + disk;
7299 if (test_bit(WantReplacement, &p->rdev->flags) &&
7300 p->replacement == NULL) {
7301 clear_bit(In_sync, &rdev->flags);
7302 set_bit(Replacement, &rdev->flags);
7303 rdev->raid_disk = disk;
7306 rcu_assign_pointer(p->replacement, rdev);
7311 print_raid5_conf(conf);
7315 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7317 /* no resync is happening, and there is enough space
7318 * on all devices, so we can resize.
7319 * We need to make sure resync covers any new space.
7320 * If the array is shrinking we should possibly wait until
7321 * any io in the removed space completes, but it hardly seems
7325 struct r5conf *conf = mddev->private;
7329 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7330 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7331 if (mddev->external_size &&
7332 mddev->array_sectors > newsize)
7334 if (mddev->bitmap) {
7335 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7339 md_set_array_sectors(mddev, newsize);
7340 set_capacity(mddev->gendisk, mddev->array_sectors);
7341 revalidate_disk(mddev->gendisk);
7342 if (sectors > mddev->dev_sectors &&
7343 mddev->recovery_cp > mddev->dev_sectors) {
7344 mddev->recovery_cp = mddev->dev_sectors;
7345 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7347 mddev->dev_sectors = sectors;
7348 mddev->resync_max_sectors = sectors;
7352 static int check_stripe_cache(struct mddev *mddev)
7354 /* Can only proceed if there are plenty of stripe_heads.
7355 * We need a minimum of one full stripe,, and for sensible progress
7356 * it is best to have about 4 times that.
7357 * If we require 4 times, then the default 256 4K stripe_heads will
7358 * allow for chunk sizes up to 256K, which is probably OK.
7359 * If the chunk size is greater, user-space should request more
7360 * stripe_heads first.
7362 struct r5conf *conf = mddev->private;
7363 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7364 > conf->min_nr_stripes ||
7365 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7366 > conf->min_nr_stripes) {
7367 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7369 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7376 static int check_reshape(struct mddev *mddev)
7378 struct r5conf *conf = mddev->private;
7382 if (mddev->delta_disks == 0 &&
7383 mddev->new_layout == mddev->layout &&
7384 mddev->new_chunk_sectors == mddev->chunk_sectors)
7385 return 0; /* nothing to do */
7386 if (has_failed(conf))
7388 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7389 /* We might be able to shrink, but the devices must
7390 * be made bigger first.
7391 * For raid6, 4 is the minimum size.
7392 * Otherwise 2 is the minimum
7395 if (mddev->level == 6)
7397 if (mddev->raid_disks + mddev->delta_disks < min)
7401 if (!check_stripe_cache(mddev))
7404 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7405 mddev->delta_disks > 0)
7406 if (resize_chunks(conf,
7407 conf->previous_raid_disks
7408 + max(0, mddev->delta_disks),
7409 max(mddev->new_chunk_sectors,
7410 mddev->chunk_sectors)
7413 return resize_stripes(conf, (conf->previous_raid_disks
7414 + mddev->delta_disks));
7417 static int raid5_start_reshape(struct mddev *mddev)
7419 struct r5conf *conf = mddev->private;
7420 struct md_rdev *rdev;
7422 unsigned long flags;
7424 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7427 if (!check_stripe_cache(mddev))
7430 if (has_failed(conf))
7433 rdev_for_each(rdev, mddev) {
7434 if (!test_bit(In_sync, &rdev->flags)
7435 && !test_bit(Faulty, &rdev->flags))
7439 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7440 /* Not enough devices even to make a degraded array
7445 /* Refuse to reduce size of the array. Any reductions in
7446 * array size must be through explicit setting of array_size
7449 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7450 < mddev->array_sectors) {
7451 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7452 "before number of disks\n", mdname(mddev));
7456 atomic_set(&conf->reshape_stripes, 0);
7457 spin_lock_irq(&conf->device_lock);
7458 write_seqcount_begin(&conf->gen_lock);
7459 conf->previous_raid_disks = conf->raid_disks;
7460 conf->raid_disks += mddev->delta_disks;
7461 conf->prev_chunk_sectors = conf->chunk_sectors;
7462 conf->chunk_sectors = mddev->new_chunk_sectors;
7463 conf->prev_algo = conf->algorithm;
7464 conf->algorithm = mddev->new_layout;
7466 /* Code that selects data_offset needs to see the generation update
7467 * if reshape_progress has been set - so a memory barrier needed.
7470 if (mddev->reshape_backwards)
7471 conf->reshape_progress = raid5_size(mddev, 0, 0);
7473 conf->reshape_progress = 0;
7474 conf->reshape_safe = conf->reshape_progress;
7475 write_seqcount_end(&conf->gen_lock);
7476 spin_unlock_irq(&conf->device_lock);
7478 /* Now make sure any requests that proceeded on the assumption
7479 * the reshape wasn't running - like Discard or Read - have
7482 mddev_suspend(mddev);
7483 mddev_resume(mddev);
7485 /* Add some new drives, as many as will fit.
7486 * We know there are enough to make the newly sized array work.
7487 * Don't add devices if we are reducing the number of
7488 * devices in the array. This is because it is not possible
7489 * to correctly record the "partially reconstructed" state of
7490 * such devices during the reshape and confusion could result.
7492 if (mddev->delta_disks >= 0) {
7493 rdev_for_each(rdev, mddev)
7494 if (rdev->raid_disk < 0 &&
7495 !test_bit(Faulty, &rdev->flags)) {
7496 if (raid5_add_disk(mddev, rdev) == 0) {
7498 >= conf->previous_raid_disks)
7499 set_bit(In_sync, &rdev->flags);
7501 rdev->recovery_offset = 0;
7503 if (sysfs_link_rdev(mddev, rdev))
7504 /* Failure here is OK */;
7506 } else if (rdev->raid_disk >= conf->previous_raid_disks
7507 && !test_bit(Faulty, &rdev->flags)) {
7508 /* This is a spare that was manually added */
7509 set_bit(In_sync, &rdev->flags);
7512 /* When a reshape changes the number of devices,
7513 * ->degraded is measured against the larger of the
7514 * pre and post number of devices.
7516 spin_lock_irqsave(&conf->device_lock, flags);
7517 mddev->degraded = calc_degraded(conf);
7518 spin_unlock_irqrestore(&conf->device_lock, flags);
7520 mddev->raid_disks = conf->raid_disks;
7521 mddev->reshape_position = conf->reshape_progress;
7522 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7524 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7525 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7526 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7527 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7528 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7529 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7531 if (!mddev->sync_thread) {
7532 mddev->recovery = 0;
7533 spin_lock_irq(&conf->device_lock);
7534 write_seqcount_begin(&conf->gen_lock);
7535 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7536 mddev->new_chunk_sectors =
7537 conf->chunk_sectors = conf->prev_chunk_sectors;
7538 mddev->new_layout = conf->algorithm = conf->prev_algo;
7539 rdev_for_each(rdev, mddev)
7540 rdev->new_data_offset = rdev->data_offset;
7542 conf->generation --;
7543 conf->reshape_progress = MaxSector;
7544 mddev->reshape_position = MaxSector;
7545 write_seqcount_end(&conf->gen_lock);
7546 spin_unlock_irq(&conf->device_lock);
7549 conf->reshape_checkpoint = jiffies;
7550 md_wakeup_thread(mddev->sync_thread);
7551 md_new_event(mddev);
7555 /* This is called from the reshape thread and should make any
7556 * changes needed in 'conf'
7558 static void end_reshape(struct r5conf *conf)
7561 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7562 struct md_rdev *rdev;
7564 spin_lock_irq(&conf->device_lock);
7565 conf->previous_raid_disks = conf->raid_disks;
7566 rdev_for_each(rdev, conf->mddev)
7567 rdev->data_offset = rdev->new_data_offset;
7569 conf->reshape_progress = MaxSector;
7570 conf->mddev->reshape_position = MaxSector;
7571 spin_unlock_irq(&conf->device_lock);
7572 wake_up(&conf->wait_for_overlap);
7574 /* read-ahead size must cover two whole stripes, which is
7575 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7577 if (conf->mddev->queue) {
7578 int data_disks = conf->raid_disks - conf->max_degraded;
7579 int stripe = data_disks * ((conf->chunk_sectors << 9)
7581 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7582 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7587 /* This is called from the raid5d thread with mddev_lock held.
7588 * It makes config changes to the device.
7590 static void raid5_finish_reshape(struct mddev *mddev)
7592 struct r5conf *conf = mddev->private;
7594 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7596 if (mddev->delta_disks > 0) {
7597 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7599 set_capacity(mddev->gendisk, mddev->array_sectors);
7600 revalidate_disk(mddev->gendisk);
7604 spin_lock_irq(&conf->device_lock);
7605 mddev->degraded = calc_degraded(conf);
7606 spin_unlock_irq(&conf->device_lock);
7607 for (d = conf->raid_disks ;
7608 d < conf->raid_disks - mddev->delta_disks;
7610 struct md_rdev *rdev = conf->disks[d].rdev;
7612 clear_bit(In_sync, &rdev->flags);
7613 rdev = conf->disks[d].replacement;
7615 clear_bit(In_sync, &rdev->flags);
7618 mddev->layout = conf->algorithm;
7619 mddev->chunk_sectors = conf->chunk_sectors;
7620 mddev->reshape_position = MaxSector;
7621 mddev->delta_disks = 0;
7622 mddev->reshape_backwards = 0;
7626 static void raid5_quiesce(struct mddev *mddev, int state)
7628 struct r5conf *conf = mddev->private;
7631 case 2: /* resume for a suspend */
7632 wake_up(&conf->wait_for_overlap);
7635 case 1: /* stop all writes */
7636 lock_all_device_hash_locks_irq(conf);
7637 /* '2' tells resync/reshape to pause so that all
7638 * active stripes can drain
7641 wait_event_cmd(conf->wait_for_quiescent,
7642 atomic_read(&conf->active_stripes) == 0 &&
7643 atomic_read(&conf->active_aligned_reads) == 0,
7644 unlock_all_device_hash_locks_irq(conf),
7645 lock_all_device_hash_locks_irq(conf));
7647 unlock_all_device_hash_locks_irq(conf);
7648 /* allow reshape to continue */
7649 wake_up(&conf->wait_for_overlap);
7652 case 0: /* re-enable writes */
7653 lock_all_device_hash_locks_irq(conf);
7655 wake_up(&conf->wait_for_quiescent);
7656 wake_up(&conf->wait_for_overlap);
7657 unlock_all_device_hash_locks_irq(conf);
7660 r5l_quiesce(conf->log, state);
7663 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7665 struct r0conf *raid0_conf = mddev->private;
7668 /* for raid0 takeover only one zone is supported */
7669 if (raid0_conf->nr_strip_zones > 1) {
7670 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7672 return ERR_PTR(-EINVAL);
7675 sectors = raid0_conf->strip_zone[0].zone_end;
7676 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7677 mddev->dev_sectors = sectors;
7678 mddev->new_level = level;
7679 mddev->new_layout = ALGORITHM_PARITY_N;
7680 mddev->new_chunk_sectors = mddev->chunk_sectors;
7681 mddev->raid_disks += 1;
7682 mddev->delta_disks = 1;
7683 /* make sure it will be not marked as dirty */
7684 mddev->recovery_cp = MaxSector;
7686 return setup_conf(mddev);
7689 static void *raid5_takeover_raid1(struct mddev *mddev)
7693 if (mddev->raid_disks != 2 ||
7694 mddev->degraded > 1)
7695 return ERR_PTR(-EINVAL);
7697 /* Should check if there are write-behind devices? */
7699 chunksect = 64*2; /* 64K by default */
7701 /* The array must be an exact multiple of chunksize */
7702 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7705 if ((chunksect<<9) < STRIPE_SIZE)
7706 /* array size does not allow a suitable chunk size */
7707 return ERR_PTR(-EINVAL);
7709 mddev->new_level = 5;
7710 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7711 mddev->new_chunk_sectors = chunksect;
7713 return setup_conf(mddev);
7716 static void *raid5_takeover_raid6(struct mddev *mddev)
7720 switch (mddev->layout) {
7721 case ALGORITHM_LEFT_ASYMMETRIC_6:
7722 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7724 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7725 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7727 case ALGORITHM_LEFT_SYMMETRIC_6:
7728 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7730 case ALGORITHM_RIGHT_SYMMETRIC_6:
7731 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7733 case ALGORITHM_PARITY_0_6:
7734 new_layout = ALGORITHM_PARITY_0;
7736 case ALGORITHM_PARITY_N:
7737 new_layout = ALGORITHM_PARITY_N;
7740 return ERR_PTR(-EINVAL);
7742 mddev->new_level = 5;
7743 mddev->new_layout = new_layout;
7744 mddev->delta_disks = -1;
7745 mddev->raid_disks -= 1;
7746 return setup_conf(mddev);
7749 static int raid5_check_reshape(struct mddev *mddev)
7751 /* For a 2-drive array, the layout and chunk size can be changed
7752 * immediately as not restriping is needed.
7753 * For larger arrays we record the new value - after validation
7754 * to be used by a reshape pass.
7756 struct r5conf *conf = mddev->private;
7757 int new_chunk = mddev->new_chunk_sectors;
7759 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7761 if (new_chunk > 0) {
7762 if (!is_power_of_2(new_chunk))
7764 if (new_chunk < (PAGE_SIZE>>9))
7766 if (mddev->array_sectors & (new_chunk-1))
7767 /* not factor of array size */
7771 /* They look valid */
7773 if (mddev->raid_disks == 2) {
7774 /* can make the change immediately */
7775 if (mddev->new_layout >= 0) {
7776 conf->algorithm = mddev->new_layout;
7777 mddev->layout = mddev->new_layout;
7779 if (new_chunk > 0) {
7780 conf->chunk_sectors = new_chunk ;
7781 mddev->chunk_sectors = new_chunk;
7783 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7784 md_wakeup_thread(mddev->thread);
7786 return check_reshape(mddev);
7789 static int raid6_check_reshape(struct mddev *mddev)
7791 int new_chunk = mddev->new_chunk_sectors;
7793 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7795 if (new_chunk > 0) {
7796 if (!is_power_of_2(new_chunk))
7798 if (new_chunk < (PAGE_SIZE >> 9))
7800 if (mddev->array_sectors & (new_chunk-1))
7801 /* not factor of array size */
7805 /* They look valid */
7806 return check_reshape(mddev);
7809 static void *raid5_takeover(struct mddev *mddev)
7811 /* raid5 can take over:
7812 * raid0 - if there is only one strip zone - make it a raid4 layout
7813 * raid1 - if there are two drives. We need to know the chunk size
7814 * raid4 - trivial - just use a raid4 layout.
7815 * raid6 - Providing it is a *_6 layout
7817 if (mddev->level == 0)
7818 return raid45_takeover_raid0(mddev, 5);
7819 if (mddev->level == 1)
7820 return raid5_takeover_raid1(mddev);
7821 if (mddev->level == 4) {
7822 mddev->new_layout = ALGORITHM_PARITY_N;
7823 mddev->new_level = 5;
7824 return setup_conf(mddev);
7826 if (mddev->level == 6)
7827 return raid5_takeover_raid6(mddev);
7829 return ERR_PTR(-EINVAL);
7832 static void *raid4_takeover(struct mddev *mddev)
7834 /* raid4 can take over:
7835 * raid0 - if there is only one strip zone
7836 * raid5 - if layout is right
7838 if (mddev->level == 0)
7839 return raid45_takeover_raid0(mddev, 4);
7840 if (mddev->level == 5 &&
7841 mddev->layout == ALGORITHM_PARITY_N) {
7842 mddev->new_layout = 0;
7843 mddev->new_level = 4;
7844 return setup_conf(mddev);
7846 return ERR_PTR(-EINVAL);
7849 static struct md_personality raid5_personality;
7851 static void *raid6_takeover(struct mddev *mddev)
7853 /* Currently can only take over a raid5. We map the
7854 * personality to an equivalent raid6 personality
7855 * with the Q block at the end.
7859 if (mddev->pers != &raid5_personality)
7860 return ERR_PTR(-EINVAL);
7861 if (mddev->degraded > 1)
7862 return ERR_PTR(-EINVAL);
7863 if (mddev->raid_disks > 253)
7864 return ERR_PTR(-EINVAL);
7865 if (mddev->raid_disks < 3)
7866 return ERR_PTR(-EINVAL);
7868 switch (mddev->layout) {
7869 case ALGORITHM_LEFT_ASYMMETRIC:
7870 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7872 case ALGORITHM_RIGHT_ASYMMETRIC:
7873 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7875 case ALGORITHM_LEFT_SYMMETRIC:
7876 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7878 case ALGORITHM_RIGHT_SYMMETRIC:
7879 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7881 case ALGORITHM_PARITY_0:
7882 new_layout = ALGORITHM_PARITY_0_6;
7884 case ALGORITHM_PARITY_N:
7885 new_layout = ALGORITHM_PARITY_N;
7888 return ERR_PTR(-EINVAL);
7890 mddev->new_level = 6;
7891 mddev->new_layout = new_layout;
7892 mddev->delta_disks = 1;
7893 mddev->raid_disks += 1;
7894 return setup_conf(mddev);
7897 static struct md_personality raid6_personality =
7901 .owner = THIS_MODULE,
7902 .make_request = raid5_make_request,
7905 .status = raid5_status,
7906 .error_handler = raid5_error,
7907 .hot_add_disk = raid5_add_disk,
7908 .hot_remove_disk= raid5_remove_disk,
7909 .spare_active = raid5_spare_active,
7910 .sync_request = raid5_sync_request,
7911 .resize = raid5_resize,
7913 .check_reshape = raid6_check_reshape,
7914 .start_reshape = raid5_start_reshape,
7915 .finish_reshape = raid5_finish_reshape,
7916 .quiesce = raid5_quiesce,
7917 .takeover = raid6_takeover,
7918 .congested = raid5_congested,
7920 static struct md_personality raid5_personality =
7924 .owner = THIS_MODULE,
7925 .make_request = raid5_make_request,
7928 .status = raid5_status,
7929 .error_handler = raid5_error,
7930 .hot_add_disk = raid5_add_disk,
7931 .hot_remove_disk= raid5_remove_disk,
7932 .spare_active = raid5_spare_active,
7933 .sync_request = raid5_sync_request,
7934 .resize = raid5_resize,
7936 .check_reshape = raid5_check_reshape,
7937 .start_reshape = raid5_start_reshape,
7938 .finish_reshape = raid5_finish_reshape,
7939 .quiesce = raid5_quiesce,
7940 .takeover = raid5_takeover,
7941 .congested = raid5_congested,
7944 static struct md_personality raid4_personality =
7948 .owner = THIS_MODULE,
7949 .make_request = raid5_make_request,
7952 .status = raid5_status,
7953 .error_handler = raid5_error,
7954 .hot_add_disk = raid5_add_disk,
7955 .hot_remove_disk= raid5_remove_disk,
7956 .spare_active = raid5_spare_active,
7957 .sync_request = raid5_sync_request,
7958 .resize = raid5_resize,
7960 .check_reshape = raid5_check_reshape,
7961 .start_reshape = raid5_start_reshape,
7962 .finish_reshape = raid5_finish_reshape,
7963 .quiesce = raid5_quiesce,
7964 .takeover = raid4_takeover,
7965 .congested = raid5_congested,
7968 static int __init raid5_init(void)
7970 raid5_wq = alloc_workqueue("raid5wq",
7971 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7974 register_md_personality(&raid6_personality);
7975 register_md_personality(&raid5_personality);
7976 register_md_personality(&raid4_personality);
7980 static void raid5_exit(void)
7982 unregister_md_personality(&raid6_personality);
7983 unregister_md_personality(&raid5_personality);
7984 unregister_md_personality(&raid4_personality);
7985 destroy_workqueue(raid5_wq);
7988 module_init(raid5_init);
7989 module_exit(raid5_exit);
7990 MODULE_LICENSE("GPL");
7991 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7992 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7993 MODULE_ALIAS("md-raid5");
7994 MODULE_ALIAS("md-raid4");
7995 MODULE_ALIAS("md-level-5");
7996 MODULE_ALIAS("md-level-4");
7997 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7998 MODULE_ALIAS("md-raid6");
7999 MODULE_ALIAS("md-level-6");
8001 /* This used to be two separate modules, they were: */
8002 MODULE_ALIAS("raid5");
8003 MODULE_ALIAS("raid6");