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 unsigned long do_wakeup = 0;
347 if (hash == NR_STRIPE_HASH_LOCKS) {
348 size = NR_STRIPE_HASH_LOCKS;
349 hash = NR_STRIPE_HASH_LOCKS - 1;
353 struct list_head *list = &temp_inactive_list[size - 1];
356 * We don't hold any lock here yet, get_active_stripe() might
357 * remove stripes from the list
359 if (!list_empty_careful(list)) {
360 spin_lock_irqsave(conf->hash_locks + hash, flags);
361 if (list_empty(conf->inactive_list + hash) &&
363 atomic_dec(&conf->empty_inactive_list_nr);
364 list_splice_tail_init(list, conf->inactive_list + hash);
365 do_wakeup |= 1 << hash;
366 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
372 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
373 if (do_wakeup & (1 << i))
374 wake_up(&conf->wait_for_stripe[i]);
378 if (atomic_read(&conf->active_stripes) == 0)
379 wake_up(&conf->wait_for_quiescent);
380 if (conf->retry_read_aligned)
381 md_wakeup_thread(conf->mddev->thread);
385 /* should hold conf->device_lock already */
386 static int release_stripe_list(struct r5conf *conf,
387 struct list_head *temp_inactive_list)
389 struct stripe_head *sh;
391 struct llist_node *head;
393 head = llist_del_all(&conf->released_stripes);
394 head = llist_reverse_order(head);
398 sh = llist_entry(head, struct stripe_head, release_list);
399 head = llist_next(head);
400 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
402 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
404 * Don't worry the bit is set here, because if the bit is set
405 * again, the count is always > 1. This is true for
406 * STRIPE_ON_UNPLUG_LIST bit too.
408 hash = sh->hash_lock_index;
409 __release_stripe(conf, sh, &temp_inactive_list[hash]);
416 static void release_stripe(struct stripe_head *sh)
418 struct r5conf *conf = sh->raid_conf;
420 struct list_head list;
424 /* Avoid release_list until the last reference.
426 if (atomic_add_unless(&sh->count, -1, 1))
429 if (unlikely(!conf->mddev->thread) ||
430 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
432 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
434 md_wakeup_thread(conf->mddev->thread);
437 local_irq_save(flags);
438 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
439 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
440 INIT_LIST_HEAD(&list);
441 hash = sh->hash_lock_index;
442 do_release_stripe(conf, sh, &list);
443 spin_unlock(&conf->device_lock);
444 release_inactive_stripe_list(conf, &list, hash);
446 local_irq_restore(flags);
449 static inline void remove_hash(struct stripe_head *sh)
451 pr_debug("remove_hash(), stripe %llu\n",
452 (unsigned long long)sh->sector);
454 hlist_del_init(&sh->hash);
457 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
459 struct hlist_head *hp = stripe_hash(conf, sh->sector);
461 pr_debug("insert_hash(), stripe %llu\n",
462 (unsigned long long)sh->sector);
464 hlist_add_head(&sh->hash, hp);
467 /* find an idle stripe, make sure it is unhashed, and return it. */
468 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
470 struct stripe_head *sh = NULL;
471 struct list_head *first;
473 if (list_empty(conf->inactive_list + hash))
475 first = (conf->inactive_list + hash)->next;
476 sh = list_entry(first, struct stripe_head, lru);
477 list_del_init(first);
479 atomic_inc(&conf->active_stripes);
480 BUG_ON(hash != sh->hash_lock_index);
481 if (list_empty(conf->inactive_list + hash))
482 atomic_inc(&conf->empty_inactive_list_nr);
487 static void shrink_buffers(struct stripe_head *sh)
491 int num = sh->raid_conf->pool_size;
493 for (i = 0; i < num ; i++) {
494 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
498 sh->dev[i].page = NULL;
503 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
506 int num = sh->raid_conf->pool_size;
508 for (i = 0; i < num; i++) {
511 if (!(page = alloc_page(gfp))) {
514 sh->dev[i].page = page;
515 sh->dev[i].orig_page = page;
520 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
521 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
522 struct stripe_head *sh);
524 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
526 struct r5conf *conf = sh->raid_conf;
529 BUG_ON(atomic_read(&sh->count) != 0);
530 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
531 BUG_ON(stripe_operations_active(sh));
532 BUG_ON(sh->batch_head);
534 pr_debug("init_stripe called, stripe %llu\n",
535 (unsigned long long)sector);
537 seq = read_seqcount_begin(&conf->gen_lock);
538 sh->generation = conf->generation - previous;
539 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
541 stripe_set_idx(sector, conf, previous, sh);
544 for (i = sh->disks; i--; ) {
545 struct r5dev *dev = &sh->dev[i];
547 if (dev->toread || dev->read || dev->towrite || dev->written ||
548 test_bit(R5_LOCKED, &dev->flags)) {
549 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
550 (unsigned long long)sh->sector, i, dev->toread,
551 dev->read, dev->towrite, dev->written,
552 test_bit(R5_LOCKED, &dev->flags));
556 raid5_build_block(sh, i, previous);
558 if (read_seqcount_retry(&conf->gen_lock, seq))
560 sh->overwrite_disks = 0;
561 insert_hash(conf, sh);
562 sh->cpu = smp_processor_id();
563 set_bit(STRIPE_BATCH_READY, &sh->state);
566 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
569 struct stripe_head *sh;
571 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
572 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
573 if (sh->sector == sector && sh->generation == generation)
575 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
580 * Need to check if array has failed when deciding whether to:
582 * - remove non-faulty devices
585 * This determination is simple when no reshape is happening.
586 * However if there is a reshape, we need to carefully check
587 * both the before and after sections.
588 * This is because some failed devices may only affect one
589 * of the two sections, and some non-in_sync devices may
590 * be insync in the section most affected by failed devices.
592 static int calc_degraded(struct r5conf *conf)
594 int degraded, degraded2;
599 for (i = 0; i < conf->previous_raid_disks; i++) {
600 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
601 if (rdev && test_bit(Faulty, &rdev->flags))
602 rdev = rcu_dereference(conf->disks[i].replacement);
603 if (!rdev || test_bit(Faulty, &rdev->flags))
605 else if (test_bit(In_sync, &rdev->flags))
608 /* not in-sync or faulty.
609 * If the reshape increases the number of devices,
610 * this is being recovered by the reshape, so
611 * this 'previous' section is not in_sync.
612 * If the number of devices is being reduced however,
613 * the device can only be part of the array if
614 * we are reverting a reshape, so this section will
617 if (conf->raid_disks >= conf->previous_raid_disks)
621 if (conf->raid_disks == conf->previous_raid_disks)
625 for (i = 0; i < conf->raid_disks; i++) {
626 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
627 if (rdev && test_bit(Faulty, &rdev->flags))
628 rdev = rcu_dereference(conf->disks[i].replacement);
629 if (!rdev || test_bit(Faulty, &rdev->flags))
631 else if (test_bit(In_sync, &rdev->flags))
634 /* not in-sync or faulty.
635 * If reshape increases the number of devices, this
636 * section has already been recovered, else it
637 * almost certainly hasn't.
639 if (conf->raid_disks <= conf->previous_raid_disks)
643 if (degraded2 > degraded)
648 static int has_failed(struct r5conf *conf)
652 if (conf->mddev->reshape_position == MaxSector)
653 return conf->mddev->degraded > conf->max_degraded;
655 degraded = calc_degraded(conf);
656 if (degraded > conf->max_degraded)
661 static struct stripe_head *
662 get_active_stripe(struct r5conf *conf, sector_t sector,
663 int previous, int noblock, int noquiesce)
665 struct stripe_head *sh;
666 int hash = stripe_hash_locks_hash(sector);
668 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
670 spin_lock_irq(conf->hash_locks + hash);
673 wait_event_lock_irq(conf->wait_for_quiescent,
674 conf->quiesce == 0 || noquiesce,
675 *(conf->hash_locks + hash));
676 sh = __find_stripe(conf, sector, conf->generation - previous);
678 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
679 sh = get_free_stripe(conf, hash);
680 if (!sh && !test_bit(R5_DID_ALLOC,
682 set_bit(R5_ALLOC_MORE,
685 if (noblock && sh == NULL)
688 set_bit(R5_INACTIVE_BLOCKED,
690 wait_event_exclusive_cmd(
691 conf->wait_for_stripe[hash],
692 !list_empty(conf->inactive_list + hash) &&
693 (atomic_read(&conf->active_stripes)
694 < (conf->max_nr_stripes * 3 / 4)
695 || !test_bit(R5_INACTIVE_BLOCKED,
696 &conf->cache_state)),
697 spin_unlock_irq(conf->hash_locks + hash),
698 spin_lock_irq(conf->hash_locks + hash));
699 clear_bit(R5_INACTIVE_BLOCKED,
702 init_stripe(sh, sector, previous);
703 atomic_inc(&sh->count);
705 } else if (!atomic_inc_not_zero(&sh->count)) {
706 spin_lock(&conf->device_lock);
707 if (!atomic_read(&sh->count)) {
708 if (!test_bit(STRIPE_HANDLE, &sh->state))
709 atomic_inc(&conf->active_stripes);
710 BUG_ON(list_empty(&sh->lru) &&
711 !test_bit(STRIPE_EXPANDING, &sh->state));
712 list_del_init(&sh->lru);
714 sh->group->stripes_cnt--;
718 atomic_inc(&sh->count);
719 spin_unlock(&conf->device_lock);
721 } while (sh == NULL);
723 if (!list_empty(conf->inactive_list + hash))
724 wake_up(&conf->wait_for_stripe[hash]);
726 spin_unlock_irq(conf->hash_locks + hash);
730 static bool is_full_stripe_write(struct stripe_head *sh)
732 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
733 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
736 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
740 spin_lock(&sh2->stripe_lock);
741 spin_lock_nested(&sh1->stripe_lock, 1);
743 spin_lock(&sh1->stripe_lock);
744 spin_lock_nested(&sh2->stripe_lock, 1);
748 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
750 spin_unlock(&sh1->stripe_lock);
751 spin_unlock(&sh2->stripe_lock);
755 /* Only freshly new full stripe normal write stripe can be added to a batch list */
756 static bool stripe_can_batch(struct stripe_head *sh)
758 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
759 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
760 is_full_stripe_write(sh);
763 /* we only do back search */
764 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
766 struct stripe_head *head;
767 sector_t head_sector, tmp_sec;
771 if (!stripe_can_batch(sh))
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 list_del_init(&head->lru);
791 head->group->stripes_cnt--;
795 atomic_inc(&head->count);
796 spin_unlock(&conf->device_lock);
798 spin_unlock_irq(conf->hash_locks + hash);
802 if (!stripe_can_batch(head))
805 lock_two_stripes(head, sh);
806 /* clear_batch_ready clear the flag */
807 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
814 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
816 if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
819 if (head->batch_head) {
820 spin_lock(&head->batch_head->batch_lock);
821 /* This batch list is already running */
822 if (!stripe_can_batch(head)) {
823 spin_unlock(&head->batch_head->batch_lock);
828 * at this point, head's BATCH_READY could be cleared, but we
829 * can still add the stripe to batch list
831 list_add(&sh->batch_list, &head->batch_list);
832 spin_unlock(&head->batch_head->batch_lock);
834 sh->batch_head = head->batch_head;
836 head->batch_head = head;
837 sh->batch_head = head->batch_head;
838 spin_lock(&head->batch_lock);
839 list_add_tail(&sh->batch_list, &head->batch_list);
840 spin_unlock(&head->batch_lock);
843 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
844 if (atomic_dec_return(&conf->preread_active_stripes)
846 md_wakeup_thread(conf->mddev->thread);
848 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
849 int seq = sh->bm_seq;
850 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
851 sh->batch_head->bm_seq > seq)
852 seq = sh->batch_head->bm_seq;
853 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
854 sh->batch_head->bm_seq = seq;
857 atomic_inc(&sh->count);
859 unlock_two_stripes(head, sh);
861 release_stripe(head);
864 /* Determine if 'data_offset' or 'new_data_offset' should be used
865 * in this stripe_head.
867 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
869 sector_t progress = conf->reshape_progress;
870 /* Need a memory barrier to make sure we see the value
871 * of conf->generation, or ->data_offset that was set before
872 * reshape_progress was updated.
875 if (progress == MaxSector)
877 if (sh->generation == conf->generation - 1)
879 /* We are in a reshape, and this is a new-generation stripe,
880 * so use new_data_offset.
886 raid5_end_read_request(struct bio *bi);
888 raid5_end_write_request(struct bio *bi);
890 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
892 struct r5conf *conf = sh->raid_conf;
893 int i, disks = sh->disks;
894 struct stripe_head *head_sh = sh;
898 for (i = disks; i--; ) {
900 int replace_only = 0;
901 struct bio *bi, *rbi;
902 struct md_rdev *rdev, *rrdev = NULL;
905 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
906 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
910 if (test_bit(R5_Discard, &sh->dev[i].flags))
912 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
914 else if (test_and_clear_bit(R5_WantReplace,
915 &sh->dev[i].flags)) {
920 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
924 bi = &sh->dev[i].req;
925 rbi = &sh->dev[i].rreq; /* For writing to replacement */
928 rrdev = rcu_dereference(conf->disks[i].replacement);
929 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
930 rdev = rcu_dereference(conf->disks[i].rdev);
939 /* We raced and saw duplicates */
942 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
947 if (rdev && test_bit(Faulty, &rdev->flags))
950 atomic_inc(&rdev->nr_pending);
951 if (rrdev && test_bit(Faulty, &rrdev->flags))
954 atomic_inc(&rrdev->nr_pending);
957 /* We have already checked bad blocks for reads. Now
958 * need to check for writes. We never accept write errors
959 * on the replacement, so we don't to check rrdev.
961 while ((rw & WRITE) && rdev &&
962 test_bit(WriteErrorSeen, &rdev->flags)) {
965 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
966 &first_bad, &bad_sectors);
971 set_bit(BlockedBadBlocks, &rdev->flags);
972 if (!conf->mddev->external &&
973 conf->mddev->flags) {
974 /* It is very unlikely, but we might
975 * still need to write out the
976 * bad block log - better give it
978 md_check_recovery(conf->mddev);
981 * Because md_wait_for_blocked_rdev
982 * will dec nr_pending, we must
983 * increment it first.
985 atomic_inc(&rdev->nr_pending);
986 md_wait_for_blocked_rdev(rdev, conf->mddev);
988 /* Acknowledged bad block - skip the write */
989 rdev_dec_pending(rdev, conf->mddev);
995 if (s->syncing || s->expanding || s->expanded
997 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
999 set_bit(STRIPE_IO_STARTED, &sh->state);
1002 bi->bi_bdev = rdev->bdev;
1004 bi->bi_end_io = (rw & WRITE)
1005 ? raid5_end_write_request
1006 : raid5_end_read_request;
1007 bi->bi_private = sh;
1009 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1010 __func__, (unsigned long long)sh->sector,
1012 atomic_inc(&sh->count);
1014 atomic_inc(&head_sh->count);
1015 if (use_new_offset(conf, sh))
1016 bi->bi_iter.bi_sector = (sh->sector
1017 + rdev->new_data_offset);
1019 bi->bi_iter.bi_sector = (sh->sector
1020 + rdev->data_offset);
1021 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1022 bi->bi_rw |= REQ_NOMERGE;
1024 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1025 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1026 sh->dev[i].vec.bv_page = sh->dev[i].page;
1028 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1029 bi->bi_io_vec[0].bv_offset = 0;
1030 bi->bi_iter.bi_size = STRIPE_SIZE;
1032 * If this is discard request, set bi_vcnt 0. We don't
1033 * want to confuse SCSI because SCSI will replace payload
1035 if (rw & REQ_DISCARD)
1038 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1040 if (conf->mddev->gendisk)
1041 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1042 bi, disk_devt(conf->mddev->gendisk),
1044 generic_make_request(bi);
1047 if (s->syncing || s->expanding || s->expanded
1049 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1051 set_bit(STRIPE_IO_STARTED, &sh->state);
1054 rbi->bi_bdev = rrdev->bdev;
1056 BUG_ON(!(rw & WRITE));
1057 rbi->bi_end_io = raid5_end_write_request;
1058 rbi->bi_private = sh;
1060 pr_debug("%s: for %llu schedule op %ld on "
1061 "replacement disc %d\n",
1062 __func__, (unsigned long long)sh->sector,
1064 atomic_inc(&sh->count);
1066 atomic_inc(&head_sh->count);
1067 if (use_new_offset(conf, sh))
1068 rbi->bi_iter.bi_sector = (sh->sector
1069 + rrdev->new_data_offset);
1071 rbi->bi_iter.bi_sector = (sh->sector
1072 + rrdev->data_offset);
1073 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1074 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1075 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1077 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1078 rbi->bi_io_vec[0].bv_offset = 0;
1079 rbi->bi_iter.bi_size = STRIPE_SIZE;
1081 * If this is discard request, set bi_vcnt 0. We don't
1082 * want to confuse SCSI because SCSI will replace payload
1084 if (rw & REQ_DISCARD)
1086 if (conf->mddev->gendisk)
1087 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1088 rbi, disk_devt(conf->mddev->gendisk),
1090 generic_make_request(rbi);
1092 if (!rdev && !rrdev) {
1094 set_bit(STRIPE_DEGRADED, &sh->state);
1095 pr_debug("skip op %ld on disc %d for sector %llu\n",
1096 bi->bi_rw, i, (unsigned long long)sh->sector);
1097 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1098 set_bit(STRIPE_HANDLE, &sh->state);
1101 if (!head_sh->batch_head)
1103 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1110 static struct dma_async_tx_descriptor *
1111 async_copy_data(int frombio, struct bio *bio, struct page **page,
1112 sector_t sector, struct dma_async_tx_descriptor *tx,
1113 struct stripe_head *sh)
1116 struct bvec_iter iter;
1117 struct page *bio_page;
1119 struct async_submit_ctl submit;
1120 enum async_tx_flags flags = 0;
1122 if (bio->bi_iter.bi_sector >= sector)
1123 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1125 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1128 flags |= ASYNC_TX_FENCE;
1129 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1131 bio_for_each_segment(bvl, bio, iter) {
1132 int len = bvl.bv_len;
1136 if (page_offset < 0) {
1137 b_offset = -page_offset;
1138 page_offset += b_offset;
1142 if (len > 0 && page_offset + len > STRIPE_SIZE)
1143 clen = STRIPE_SIZE - page_offset;
1148 b_offset += bvl.bv_offset;
1149 bio_page = bvl.bv_page;
1151 if (sh->raid_conf->skip_copy &&
1152 b_offset == 0 && page_offset == 0 &&
1153 clen == STRIPE_SIZE)
1156 tx = async_memcpy(*page, bio_page, page_offset,
1157 b_offset, clen, &submit);
1159 tx = async_memcpy(bio_page, *page, b_offset,
1160 page_offset, clen, &submit);
1162 /* chain the operations */
1163 submit.depend_tx = tx;
1165 if (clen < len) /* hit end of page */
1173 static void ops_complete_biofill(void *stripe_head_ref)
1175 struct stripe_head *sh = stripe_head_ref;
1176 struct bio_list return_bi = BIO_EMPTY_LIST;
1179 pr_debug("%s: stripe %llu\n", __func__,
1180 (unsigned long long)sh->sector);
1182 /* clear completed biofills */
1183 for (i = sh->disks; i--; ) {
1184 struct r5dev *dev = &sh->dev[i];
1186 /* acknowledge completion of a biofill operation */
1187 /* and check if we need to reply to a read request,
1188 * new R5_Wantfill requests are held off until
1189 * !STRIPE_BIOFILL_RUN
1191 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1192 struct bio *rbi, *rbi2;
1197 while (rbi && rbi->bi_iter.bi_sector <
1198 dev->sector + STRIPE_SECTORS) {
1199 rbi2 = r5_next_bio(rbi, dev->sector);
1200 if (!raid5_dec_bi_active_stripes(rbi))
1201 bio_list_add(&return_bi, rbi);
1206 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1208 return_io(&return_bi);
1210 set_bit(STRIPE_HANDLE, &sh->state);
1214 static void ops_run_biofill(struct stripe_head *sh)
1216 struct dma_async_tx_descriptor *tx = NULL;
1217 struct async_submit_ctl submit;
1220 BUG_ON(sh->batch_head);
1221 pr_debug("%s: stripe %llu\n", __func__,
1222 (unsigned long long)sh->sector);
1224 for (i = sh->disks; i--; ) {
1225 struct r5dev *dev = &sh->dev[i];
1226 if (test_bit(R5_Wantfill, &dev->flags)) {
1228 spin_lock_irq(&sh->stripe_lock);
1229 dev->read = rbi = dev->toread;
1231 spin_unlock_irq(&sh->stripe_lock);
1232 while (rbi && rbi->bi_iter.bi_sector <
1233 dev->sector + STRIPE_SECTORS) {
1234 tx = async_copy_data(0, rbi, &dev->page,
1235 dev->sector, tx, sh);
1236 rbi = r5_next_bio(rbi, dev->sector);
1241 atomic_inc(&sh->count);
1242 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1243 async_trigger_callback(&submit);
1246 static void mark_target_uptodate(struct stripe_head *sh, int target)
1253 tgt = &sh->dev[target];
1254 set_bit(R5_UPTODATE, &tgt->flags);
1255 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1256 clear_bit(R5_Wantcompute, &tgt->flags);
1259 static void ops_complete_compute(void *stripe_head_ref)
1261 struct stripe_head *sh = stripe_head_ref;
1263 pr_debug("%s: stripe %llu\n", __func__,
1264 (unsigned long long)sh->sector);
1266 /* mark the computed target(s) as uptodate */
1267 mark_target_uptodate(sh, sh->ops.target);
1268 mark_target_uptodate(sh, sh->ops.target2);
1270 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1271 if (sh->check_state == check_state_compute_run)
1272 sh->check_state = check_state_compute_result;
1273 set_bit(STRIPE_HANDLE, &sh->state);
1277 /* return a pointer to the address conversion region of the scribble buffer */
1278 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1279 struct raid5_percpu *percpu, int i)
1283 addr = flex_array_get(percpu->scribble, i);
1284 return addr + sizeof(struct page *) * (sh->disks + 2);
1287 /* return a pointer to the address conversion region of the scribble buffer */
1288 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1292 addr = flex_array_get(percpu->scribble, i);
1296 static struct dma_async_tx_descriptor *
1297 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1299 int disks = sh->disks;
1300 struct page **xor_srcs = to_addr_page(percpu, 0);
1301 int target = sh->ops.target;
1302 struct r5dev *tgt = &sh->dev[target];
1303 struct page *xor_dest = tgt->page;
1305 struct dma_async_tx_descriptor *tx;
1306 struct async_submit_ctl submit;
1309 BUG_ON(sh->batch_head);
1311 pr_debug("%s: stripe %llu block: %d\n",
1312 __func__, (unsigned long long)sh->sector, target);
1313 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1315 for (i = disks; i--; )
1317 xor_srcs[count++] = sh->dev[i].page;
1319 atomic_inc(&sh->count);
1321 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1322 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1323 if (unlikely(count == 1))
1324 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1326 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1331 /* set_syndrome_sources - populate source buffers for gen_syndrome
1332 * @srcs - (struct page *) array of size sh->disks
1333 * @sh - stripe_head to parse
1335 * Populates srcs in proper layout order for the stripe and returns the
1336 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1337 * destination buffer is recorded in srcs[count] and the Q destination
1338 * is recorded in srcs[count+1]].
1340 static int set_syndrome_sources(struct page **srcs,
1341 struct stripe_head *sh,
1344 int disks = sh->disks;
1345 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1346 int d0_idx = raid6_d0(sh);
1350 for (i = 0; i < disks; i++)
1356 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1357 struct r5dev *dev = &sh->dev[i];
1359 if (i == sh->qd_idx || i == sh->pd_idx ||
1360 (srctype == SYNDROME_SRC_ALL) ||
1361 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1362 test_bit(R5_Wantdrain, &dev->flags)) ||
1363 (srctype == SYNDROME_SRC_WRITTEN &&
1365 srcs[slot] = sh->dev[i].page;
1366 i = raid6_next_disk(i, disks);
1367 } while (i != d0_idx);
1369 return syndrome_disks;
1372 static struct dma_async_tx_descriptor *
1373 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1375 int disks = sh->disks;
1376 struct page **blocks = to_addr_page(percpu, 0);
1378 int qd_idx = sh->qd_idx;
1379 struct dma_async_tx_descriptor *tx;
1380 struct async_submit_ctl submit;
1386 BUG_ON(sh->batch_head);
1387 if (sh->ops.target < 0)
1388 target = sh->ops.target2;
1389 else if (sh->ops.target2 < 0)
1390 target = sh->ops.target;
1392 /* we should only have one valid target */
1395 pr_debug("%s: stripe %llu block: %d\n",
1396 __func__, (unsigned long long)sh->sector, target);
1398 tgt = &sh->dev[target];
1399 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1402 atomic_inc(&sh->count);
1404 if (target == qd_idx) {
1405 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1406 blocks[count] = NULL; /* regenerating p is not necessary */
1407 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1408 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1409 ops_complete_compute, sh,
1410 to_addr_conv(sh, percpu, 0));
1411 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1413 /* Compute any data- or p-drive using XOR */
1415 for (i = disks; i-- ; ) {
1416 if (i == target || i == qd_idx)
1418 blocks[count++] = sh->dev[i].page;
1421 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1422 NULL, ops_complete_compute, sh,
1423 to_addr_conv(sh, percpu, 0));
1424 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1430 static struct dma_async_tx_descriptor *
1431 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1433 int i, count, disks = sh->disks;
1434 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1435 int d0_idx = raid6_d0(sh);
1436 int faila = -1, failb = -1;
1437 int target = sh->ops.target;
1438 int target2 = sh->ops.target2;
1439 struct r5dev *tgt = &sh->dev[target];
1440 struct r5dev *tgt2 = &sh->dev[target2];
1441 struct dma_async_tx_descriptor *tx;
1442 struct page **blocks = to_addr_page(percpu, 0);
1443 struct async_submit_ctl submit;
1445 BUG_ON(sh->batch_head);
1446 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1447 __func__, (unsigned long long)sh->sector, target, target2);
1448 BUG_ON(target < 0 || target2 < 0);
1449 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1450 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1452 /* we need to open-code set_syndrome_sources to handle the
1453 * slot number conversion for 'faila' and 'failb'
1455 for (i = 0; i < disks ; i++)
1460 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1462 blocks[slot] = sh->dev[i].page;
1468 i = raid6_next_disk(i, disks);
1469 } while (i != d0_idx);
1471 BUG_ON(faila == failb);
1474 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1475 __func__, (unsigned long long)sh->sector, faila, failb);
1477 atomic_inc(&sh->count);
1479 if (failb == syndrome_disks+1) {
1480 /* Q disk is one of the missing disks */
1481 if (faila == syndrome_disks) {
1482 /* Missing P+Q, just recompute */
1483 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1484 ops_complete_compute, sh,
1485 to_addr_conv(sh, percpu, 0));
1486 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1487 STRIPE_SIZE, &submit);
1491 int qd_idx = sh->qd_idx;
1493 /* Missing D+Q: recompute D from P, then recompute Q */
1494 if (target == qd_idx)
1495 data_target = target2;
1497 data_target = target;
1500 for (i = disks; i-- ; ) {
1501 if (i == data_target || i == qd_idx)
1503 blocks[count++] = sh->dev[i].page;
1505 dest = sh->dev[data_target].page;
1506 init_async_submit(&submit,
1507 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1509 to_addr_conv(sh, percpu, 0));
1510 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1513 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1514 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1515 ops_complete_compute, sh,
1516 to_addr_conv(sh, percpu, 0));
1517 return async_gen_syndrome(blocks, 0, count+2,
1518 STRIPE_SIZE, &submit);
1521 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1522 ops_complete_compute, sh,
1523 to_addr_conv(sh, percpu, 0));
1524 if (failb == syndrome_disks) {
1525 /* We're missing D+P. */
1526 return async_raid6_datap_recov(syndrome_disks+2,
1530 /* We're missing D+D. */
1531 return async_raid6_2data_recov(syndrome_disks+2,
1532 STRIPE_SIZE, faila, failb,
1538 static void ops_complete_prexor(void *stripe_head_ref)
1540 struct stripe_head *sh = stripe_head_ref;
1542 pr_debug("%s: stripe %llu\n", __func__,
1543 (unsigned long long)sh->sector);
1546 static struct dma_async_tx_descriptor *
1547 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1548 struct dma_async_tx_descriptor *tx)
1550 int disks = sh->disks;
1551 struct page **xor_srcs = to_addr_page(percpu, 0);
1552 int count = 0, pd_idx = sh->pd_idx, i;
1553 struct async_submit_ctl submit;
1555 /* existing parity data subtracted */
1556 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1558 BUG_ON(sh->batch_head);
1559 pr_debug("%s: stripe %llu\n", __func__,
1560 (unsigned long long)sh->sector);
1562 for (i = disks; i--; ) {
1563 struct r5dev *dev = &sh->dev[i];
1564 /* Only process blocks that are known to be uptodate */
1565 if (test_bit(R5_Wantdrain, &dev->flags))
1566 xor_srcs[count++] = dev->page;
1569 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1570 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1571 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1576 static struct dma_async_tx_descriptor *
1577 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1578 struct dma_async_tx_descriptor *tx)
1580 struct page **blocks = to_addr_page(percpu, 0);
1582 struct async_submit_ctl submit;
1584 pr_debug("%s: stripe %llu\n", __func__,
1585 (unsigned long long)sh->sector);
1587 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1589 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1590 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1591 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1596 static struct dma_async_tx_descriptor *
1597 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1599 int disks = sh->disks;
1601 struct stripe_head *head_sh = sh;
1603 pr_debug("%s: stripe %llu\n", __func__,
1604 (unsigned long long)sh->sector);
1606 for (i = disks; i--; ) {
1611 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1616 spin_lock_irq(&sh->stripe_lock);
1617 chosen = dev->towrite;
1618 dev->towrite = NULL;
1619 sh->overwrite_disks = 0;
1620 BUG_ON(dev->written);
1621 wbi = dev->written = chosen;
1622 spin_unlock_irq(&sh->stripe_lock);
1623 WARN_ON(dev->page != dev->orig_page);
1625 while (wbi && wbi->bi_iter.bi_sector <
1626 dev->sector + STRIPE_SECTORS) {
1627 if (wbi->bi_rw & REQ_FUA)
1628 set_bit(R5_WantFUA, &dev->flags);
1629 if (wbi->bi_rw & REQ_SYNC)
1630 set_bit(R5_SyncIO, &dev->flags);
1631 if (wbi->bi_rw & REQ_DISCARD)
1632 set_bit(R5_Discard, &dev->flags);
1634 tx = async_copy_data(1, wbi, &dev->page,
1635 dev->sector, tx, sh);
1636 if (dev->page != dev->orig_page) {
1637 set_bit(R5_SkipCopy, &dev->flags);
1638 clear_bit(R5_UPTODATE, &dev->flags);
1639 clear_bit(R5_OVERWRITE, &dev->flags);
1642 wbi = r5_next_bio(wbi, dev->sector);
1645 if (head_sh->batch_head) {
1646 sh = list_first_entry(&sh->batch_list,
1659 static void ops_complete_reconstruct(void *stripe_head_ref)
1661 struct stripe_head *sh = stripe_head_ref;
1662 int disks = sh->disks;
1663 int pd_idx = sh->pd_idx;
1664 int qd_idx = sh->qd_idx;
1666 bool fua = false, sync = false, discard = false;
1668 pr_debug("%s: stripe %llu\n", __func__,
1669 (unsigned long long)sh->sector);
1671 for (i = disks; i--; ) {
1672 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1673 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1674 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1677 for (i = disks; i--; ) {
1678 struct r5dev *dev = &sh->dev[i];
1680 if (dev->written || i == pd_idx || i == qd_idx) {
1681 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1682 set_bit(R5_UPTODATE, &dev->flags);
1684 set_bit(R5_WantFUA, &dev->flags);
1686 set_bit(R5_SyncIO, &dev->flags);
1690 if (sh->reconstruct_state == reconstruct_state_drain_run)
1691 sh->reconstruct_state = reconstruct_state_drain_result;
1692 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1693 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1695 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1696 sh->reconstruct_state = reconstruct_state_result;
1699 set_bit(STRIPE_HANDLE, &sh->state);
1704 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1705 struct dma_async_tx_descriptor *tx)
1707 int disks = sh->disks;
1708 struct page **xor_srcs;
1709 struct async_submit_ctl submit;
1710 int count, pd_idx = sh->pd_idx, i;
1711 struct page *xor_dest;
1713 unsigned long flags;
1715 struct stripe_head *head_sh = sh;
1718 pr_debug("%s: stripe %llu\n", __func__,
1719 (unsigned long long)sh->sector);
1721 for (i = 0; i < sh->disks; i++) {
1724 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1727 if (i >= sh->disks) {
1728 atomic_inc(&sh->count);
1729 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1730 ops_complete_reconstruct(sh);
1735 xor_srcs = to_addr_page(percpu, j);
1736 /* check if prexor is active which means only process blocks
1737 * that are part of a read-modify-write (written)
1739 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1741 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1742 for (i = disks; i--; ) {
1743 struct r5dev *dev = &sh->dev[i];
1744 if (head_sh->dev[i].written)
1745 xor_srcs[count++] = dev->page;
1748 xor_dest = sh->dev[pd_idx].page;
1749 for (i = disks; i--; ) {
1750 struct r5dev *dev = &sh->dev[i];
1752 xor_srcs[count++] = dev->page;
1756 /* 1/ if we prexor'd then the dest is reused as a source
1757 * 2/ if we did not prexor then we are redoing the parity
1758 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1759 * for the synchronous xor case
1761 last_stripe = !head_sh->batch_head ||
1762 list_first_entry(&sh->batch_list,
1763 struct stripe_head, batch_list) == head_sh;
1765 flags = ASYNC_TX_ACK |
1766 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1768 atomic_inc(&head_sh->count);
1769 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1770 to_addr_conv(sh, percpu, j));
1772 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1773 init_async_submit(&submit, flags, tx, NULL, NULL,
1774 to_addr_conv(sh, percpu, j));
1777 if (unlikely(count == 1))
1778 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1780 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1783 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1790 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1791 struct dma_async_tx_descriptor *tx)
1793 struct async_submit_ctl submit;
1794 struct page **blocks;
1795 int count, i, j = 0;
1796 struct stripe_head *head_sh = sh;
1799 unsigned long txflags;
1801 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1803 for (i = 0; i < sh->disks; i++) {
1804 if (sh->pd_idx == i || sh->qd_idx == i)
1806 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1809 if (i >= sh->disks) {
1810 atomic_inc(&sh->count);
1811 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1812 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1813 ops_complete_reconstruct(sh);
1818 blocks = to_addr_page(percpu, j);
1820 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1821 synflags = SYNDROME_SRC_WRITTEN;
1822 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1824 synflags = SYNDROME_SRC_ALL;
1825 txflags = ASYNC_TX_ACK;
1828 count = set_syndrome_sources(blocks, sh, synflags);
1829 last_stripe = !head_sh->batch_head ||
1830 list_first_entry(&sh->batch_list,
1831 struct stripe_head, batch_list) == head_sh;
1834 atomic_inc(&head_sh->count);
1835 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1836 head_sh, to_addr_conv(sh, percpu, j));
1838 init_async_submit(&submit, 0, tx, NULL, NULL,
1839 to_addr_conv(sh, percpu, j));
1840 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1843 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1849 static void ops_complete_check(void *stripe_head_ref)
1851 struct stripe_head *sh = stripe_head_ref;
1853 pr_debug("%s: stripe %llu\n", __func__,
1854 (unsigned long long)sh->sector);
1856 sh->check_state = check_state_check_result;
1857 set_bit(STRIPE_HANDLE, &sh->state);
1861 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1863 int disks = sh->disks;
1864 int pd_idx = sh->pd_idx;
1865 int qd_idx = sh->qd_idx;
1866 struct page *xor_dest;
1867 struct page **xor_srcs = to_addr_page(percpu, 0);
1868 struct dma_async_tx_descriptor *tx;
1869 struct async_submit_ctl submit;
1873 pr_debug("%s: stripe %llu\n", __func__,
1874 (unsigned long long)sh->sector);
1876 BUG_ON(sh->batch_head);
1878 xor_dest = sh->dev[pd_idx].page;
1879 xor_srcs[count++] = xor_dest;
1880 for (i = disks; i--; ) {
1881 if (i == pd_idx || i == qd_idx)
1883 xor_srcs[count++] = sh->dev[i].page;
1886 init_async_submit(&submit, 0, NULL, NULL, NULL,
1887 to_addr_conv(sh, percpu, 0));
1888 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1889 &sh->ops.zero_sum_result, &submit);
1891 atomic_inc(&sh->count);
1892 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1893 tx = async_trigger_callback(&submit);
1896 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1898 struct page **srcs = to_addr_page(percpu, 0);
1899 struct async_submit_ctl submit;
1902 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1903 (unsigned long long)sh->sector, checkp);
1905 BUG_ON(sh->batch_head);
1906 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1910 atomic_inc(&sh->count);
1911 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1912 sh, to_addr_conv(sh, percpu, 0));
1913 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1914 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1917 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1919 int overlap_clear = 0, i, disks = sh->disks;
1920 struct dma_async_tx_descriptor *tx = NULL;
1921 struct r5conf *conf = sh->raid_conf;
1922 int level = conf->level;
1923 struct raid5_percpu *percpu;
1927 percpu = per_cpu_ptr(conf->percpu, cpu);
1928 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1929 ops_run_biofill(sh);
1933 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1935 tx = ops_run_compute5(sh, percpu);
1937 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1938 tx = ops_run_compute6_1(sh, percpu);
1940 tx = ops_run_compute6_2(sh, percpu);
1942 /* terminate the chain if reconstruct is not set to be run */
1943 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1947 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1949 tx = ops_run_prexor5(sh, percpu, tx);
1951 tx = ops_run_prexor6(sh, percpu, tx);
1954 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1955 tx = ops_run_biodrain(sh, tx);
1959 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1961 ops_run_reconstruct5(sh, percpu, tx);
1963 ops_run_reconstruct6(sh, percpu, tx);
1966 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1967 if (sh->check_state == check_state_run)
1968 ops_run_check_p(sh, percpu);
1969 else if (sh->check_state == check_state_run_q)
1970 ops_run_check_pq(sh, percpu, 0);
1971 else if (sh->check_state == check_state_run_pq)
1972 ops_run_check_pq(sh, percpu, 1);
1977 if (overlap_clear && !sh->batch_head)
1978 for (i = disks; i--; ) {
1979 struct r5dev *dev = &sh->dev[i];
1980 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1981 wake_up(&sh->raid_conf->wait_for_overlap);
1986 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1988 struct stripe_head *sh;
1990 sh = kmem_cache_zalloc(sc, gfp);
1992 spin_lock_init(&sh->stripe_lock);
1993 spin_lock_init(&sh->batch_lock);
1994 INIT_LIST_HEAD(&sh->batch_list);
1995 INIT_LIST_HEAD(&sh->lru);
1996 atomic_set(&sh->count, 1);
2000 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2002 struct stripe_head *sh;
2004 sh = alloc_stripe(conf->slab_cache, gfp);
2008 sh->raid_conf = conf;
2010 if (grow_buffers(sh, gfp)) {
2012 kmem_cache_free(conf->slab_cache, sh);
2015 sh->hash_lock_index =
2016 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2017 /* we just created an active stripe so... */
2018 atomic_inc(&conf->active_stripes);
2021 conf->max_nr_stripes++;
2025 static int grow_stripes(struct r5conf *conf, int num)
2027 struct kmem_cache *sc;
2028 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2030 if (conf->mddev->gendisk)
2031 sprintf(conf->cache_name[0],
2032 "raid%d-%s", conf->level, mdname(conf->mddev));
2034 sprintf(conf->cache_name[0],
2035 "raid%d-%p", conf->level, conf->mddev);
2036 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2038 conf->active_name = 0;
2039 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2040 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2044 conf->slab_cache = sc;
2045 conf->pool_size = devs;
2047 if (!grow_one_stripe(conf, GFP_KERNEL))
2054 * scribble_len - return the required size of the scribble region
2055 * @num - total number of disks in the array
2057 * The size must be enough to contain:
2058 * 1/ a struct page pointer for each device in the array +2
2059 * 2/ room to convert each entry in (1) to its corresponding dma
2060 * (dma_map_page()) or page (page_address()) address.
2062 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2063 * calculate over all devices (not just the data blocks), using zeros in place
2064 * of the P and Q blocks.
2066 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2068 struct flex_array *ret;
2071 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2072 ret = flex_array_alloc(len, cnt, flags);
2075 /* always prealloc all elements, so no locking is required */
2076 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2077 flex_array_free(ret);
2083 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2088 mddev_suspend(conf->mddev);
2090 for_each_present_cpu(cpu) {
2091 struct raid5_percpu *percpu;
2092 struct flex_array *scribble;
2094 percpu = per_cpu_ptr(conf->percpu, cpu);
2095 scribble = scribble_alloc(new_disks,
2096 new_sectors / STRIPE_SECTORS,
2100 flex_array_free(percpu->scribble);
2101 percpu->scribble = scribble;
2108 mddev_resume(conf->mddev);
2112 static int resize_stripes(struct r5conf *conf, int newsize)
2114 /* Make all the stripes able to hold 'newsize' devices.
2115 * New slots in each stripe get 'page' set to a new page.
2117 * This happens in stages:
2118 * 1/ create a new kmem_cache and allocate the required number of
2120 * 2/ gather all the old stripe_heads and transfer the pages across
2121 * to the new stripe_heads. This will have the side effect of
2122 * freezing the array as once all stripe_heads have been collected,
2123 * no IO will be possible. Old stripe heads are freed once their
2124 * pages have been transferred over, and the old kmem_cache is
2125 * freed when all stripes are done.
2126 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2127 * we simple return a failre status - no need to clean anything up.
2128 * 4/ allocate new pages for the new slots in the new stripe_heads.
2129 * If this fails, we don't bother trying the shrink the
2130 * stripe_heads down again, we just leave them as they are.
2131 * As each stripe_head is processed the new one is released into
2134 * Once step2 is started, we cannot afford to wait for a write,
2135 * so we use GFP_NOIO allocations.
2137 struct stripe_head *osh, *nsh;
2138 LIST_HEAD(newstripes);
2139 struct disk_info *ndisks;
2141 struct kmem_cache *sc;
2145 if (newsize <= conf->pool_size)
2146 return 0; /* never bother to shrink */
2148 err = md_allow_write(conf->mddev);
2153 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2154 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2159 /* Need to ensure auto-resizing doesn't interfere */
2160 mutex_lock(&conf->cache_size_mutex);
2162 for (i = conf->max_nr_stripes; i; i--) {
2163 nsh = alloc_stripe(sc, GFP_KERNEL);
2167 nsh->raid_conf = conf;
2168 list_add(&nsh->lru, &newstripes);
2171 /* didn't get enough, give up */
2172 while (!list_empty(&newstripes)) {
2173 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2174 list_del(&nsh->lru);
2175 kmem_cache_free(sc, nsh);
2177 kmem_cache_destroy(sc);
2178 mutex_unlock(&conf->cache_size_mutex);
2181 /* Step 2 - Must use GFP_NOIO now.
2182 * OK, we have enough stripes, start collecting inactive
2183 * stripes and copying them over
2187 list_for_each_entry(nsh, &newstripes, lru) {
2188 lock_device_hash_lock(conf, hash);
2189 wait_event_exclusive_cmd(conf->wait_for_stripe[hash],
2190 !list_empty(conf->inactive_list + hash),
2191 unlock_device_hash_lock(conf, hash),
2192 lock_device_hash_lock(conf, hash));
2193 osh = get_free_stripe(conf, hash);
2194 unlock_device_hash_lock(conf, hash);
2196 for(i=0; i<conf->pool_size; i++) {
2197 nsh->dev[i].page = osh->dev[i].page;
2198 nsh->dev[i].orig_page = osh->dev[i].page;
2200 nsh->hash_lock_index = hash;
2201 kmem_cache_free(conf->slab_cache, osh);
2203 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2204 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2209 kmem_cache_destroy(conf->slab_cache);
2212 * At this point, we are holding all the stripes so the array
2213 * is completely stalled, so now is a good time to resize
2214 * conf->disks and the scribble region
2216 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2218 for (i=0; i<conf->raid_disks; i++)
2219 ndisks[i] = conf->disks[i];
2221 conf->disks = ndisks;
2225 mutex_unlock(&conf->cache_size_mutex);
2226 /* Step 4, return new stripes to service */
2227 while(!list_empty(&newstripes)) {
2228 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2229 list_del_init(&nsh->lru);
2231 for (i=conf->raid_disks; i < newsize; i++)
2232 if (nsh->dev[i].page == NULL) {
2233 struct page *p = alloc_page(GFP_NOIO);
2234 nsh->dev[i].page = p;
2235 nsh->dev[i].orig_page = p;
2239 release_stripe(nsh);
2241 /* critical section pass, GFP_NOIO no longer needed */
2243 conf->slab_cache = sc;
2244 conf->active_name = 1-conf->active_name;
2246 conf->pool_size = newsize;
2250 static int drop_one_stripe(struct r5conf *conf)
2252 struct stripe_head *sh;
2253 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2255 spin_lock_irq(conf->hash_locks + hash);
2256 sh = get_free_stripe(conf, hash);
2257 spin_unlock_irq(conf->hash_locks + hash);
2260 BUG_ON(atomic_read(&sh->count));
2262 kmem_cache_free(conf->slab_cache, sh);
2263 atomic_dec(&conf->active_stripes);
2264 conf->max_nr_stripes--;
2268 static void shrink_stripes(struct r5conf *conf)
2270 while (conf->max_nr_stripes &&
2271 drop_one_stripe(conf))
2274 if (conf->slab_cache)
2275 kmem_cache_destroy(conf->slab_cache);
2276 conf->slab_cache = NULL;
2279 static void raid5_end_read_request(struct bio * bi)
2281 struct stripe_head *sh = bi->bi_private;
2282 struct r5conf *conf = sh->raid_conf;
2283 int disks = sh->disks, i;
2284 char b[BDEVNAME_SIZE];
2285 struct md_rdev *rdev = NULL;
2288 for (i=0 ; i<disks; i++)
2289 if (bi == &sh->dev[i].req)
2292 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2293 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2299 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2300 /* If replacement finished while this request was outstanding,
2301 * 'replacement' might be NULL already.
2302 * In that case it moved down to 'rdev'.
2303 * rdev is not removed until all requests are finished.
2305 rdev = conf->disks[i].replacement;
2307 rdev = conf->disks[i].rdev;
2309 if (use_new_offset(conf, sh))
2310 s = sh->sector + rdev->new_data_offset;
2312 s = sh->sector + rdev->data_offset;
2313 if (!bi->bi_error) {
2314 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2315 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2316 /* Note that this cannot happen on a
2317 * replacement device. We just fail those on
2322 "md/raid:%s: read error corrected"
2323 " (%lu sectors at %llu on %s)\n",
2324 mdname(conf->mddev), STRIPE_SECTORS,
2325 (unsigned long long)s,
2326 bdevname(rdev->bdev, b));
2327 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2328 clear_bit(R5_ReadError, &sh->dev[i].flags);
2329 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2330 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2331 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2333 if (atomic_read(&rdev->read_errors))
2334 atomic_set(&rdev->read_errors, 0);
2336 const char *bdn = bdevname(rdev->bdev, b);
2340 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2341 atomic_inc(&rdev->read_errors);
2342 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2345 "md/raid:%s: read error on replacement device "
2346 "(sector %llu on %s).\n",
2347 mdname(conf->mddev),
2348 (unsigned long long)s,
2350 else if (conf->mddev->degraded >= conf->max_degraded) {
2354 "md/raid:%s: read error not correctable "
2355 "(sector %llu on %s).\n",
2356 mdname(conf->mddev),
2357 (unsigned long long)s,
2359 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2364 "md/raid:%s: read error NOT corrected!! "
2365 "(sector %llu on %s).\n",
2366 mdname(conf->mddev),
2367 (unsigned long long)s,
2369 } else if (atomic_read(&rdev->read_errors)
2370 > conf->max_nr_stripes)
2372 "md/raid:%s: Too many read errors, failing device %s.\n",
2373 mdname(conf->mddev), bdn);
2376 if (set_bad && test_bit(In_sync, &rdev->flags)
2377 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2380 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2381 set_bit(R5_ReadError, &sh->dev[i].flags);
2382 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2384 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2386 clear_bit(R5_ReadError, &sh->dev[i].flags);
2387 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2389 && test_bit(In_sync, &rdev->flags)
2390 && rdev_set_badblocks(
2391 rdev, sh->sector, STRIPE_SECTORS, 0)))
2392 md_error(conf->mddev, rdev);
2395 rdev_dec_pending(rdev, conf->mddev);
2396 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2397 set_bit(STRIPE_HANDLE, &sh->state);
2401 static void raid5_end_write_request(struct bio *bi)
2403 struct stripe_head *sh = bi->bi_private;
2404 struct r5conf *conf = sh->raid_conf;
2405 int disks = sh->disks, i;
2406 struct md_rdev *uninitialized_var(rdev);
2409 int replacement = 0;
2411 for (i = 0 ; i < disks; i++) {
2412 if (bi == &sh->dev[i].req) {
2413 rdev = conf->disks[i].rdev;
2416 if (bi == &sh->dev[i].rreq) {
2417 rdev = conf->disks[i].replacement;
2421 /* rdev was removed and 'replacement'
2422 * replaced it. rdev is not removed
2423 * until all requests are finished.
2425 rdev = conf->disks[i].rdev;
2429 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2430 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2439 md_error(conf->mddev, rdev);
2440 else if (is_badblock(rdev, sh->sector,
2442 &first_bad, &bad_sectors))
2443 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2446 set_bit(STRIPE_DEGRADED, &sh->state);
2447 set_bit(WriteErrorSeen, &rdev->flags);
2448 set_bit(R5_WriteError, &sh->dev[i].flags);
2449 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2450 set_bit(MD_RECOVERY_NEEDED,
2451 &rdev->mddev->recovery);
2452 } else if (is_badblock(rdev, sh->sector,
2454 &first_bad, &bad_sectors)) {
2455 set_bit(R5_MadeGood, &sh->dev[i].flags);
2456 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2457 /* That was a successful write so make
2458 * sure it looks like we already did
2461 set_bit(R5_ReWrite, &sh->dev[i].flags);
2464 rdev_dec_pending(rdev, conf->mddev);
2466 if (sh->batch_head && bi->bi_error && !replacement)
2467 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2469 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2470 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2471 set_bit(STRIPE_HANDLE, &sh->state);
2474 if (sh->batch_head && sh != sh->batch_head)
2475 release_stripe(sh->batch_head);
2478 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2480 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2482 struct r5dev *dev = &sh->dev[i];
2484 bio_init(&dev->req);
2485 dev->req.bi_io_vec = &dev->vec;
2486 dev->req.bi_max_vecs = 1;
2487 dev->req.bi_private = sh;
2489 bio_init(&dev->rreq);
2490 dev->rreq.bi_io_vec = &dev->rvec;
2491 dev->rreq.bi_max_vecs = 1;
2492 dev->rreq.bi_private = sh;
2495 dev->sector = compute_blocknr(sh, i, previous);
2498 static void error(struct mddev *mddev, struct md_rdev *rdev)
2500 char b[BDEVNAME_SIZE];
2501 struct r5conf *conf = mddev->private;
2502 unsigned long flags;
2503 pr_debug("raid456: error called\n");
2505 spin_lock_irqsave(&conf->device_lock, flags);
2506 clear_bit(In_sync, &rdev->flags);
2507 mddev->degraded = calc_degraded(conf);
2508 spin_unlock_irqrestore(&conf->device_lock, flags);
2509 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2511 set_bit(Blocked, &rdev->flags);
2512 set_bit(Faulty, &rdev->flags);
2513 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2514 set_bit(MD_CHANGE_PENDING, &mddev->flags);
2516 "md/raid:%s: Disk failure on %s, disabling device.\n"
2517 "md/raid:%s: Operation continuing on %d devices.\n",
2519 bdevname(rdev->bdev, b),
2521 conf->raid_disks - mddev->degraded);
2525 * Input: a 'big' sector number,
2526 * Output: index of the data and parity disk, and the sector # in them.
2528 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2529 int previous, int *dd_idx,
2530 struct stripe_head *sh)
2532 sector_t stripe, stripe2;
2533 sector_t chunk_number;
2534 unsigned int chunk_offset;
2537 sector_t new_sector;
2538 int algorithm = previous ? conf->prev_algo
2540 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2541 : conf->chunk_sectors;
2542 int raid_disks = previous ? conf->previous_raid_disks
2544 int data_disks = raid_disks - conf->max_degraded;
2546 /* First compute the information on this sector */
2549 * Compute the chunk number and the sector offset inside the chunk
2551 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2552 chunk_number = r_sector;
2555 * Compute the stripe number
2557 stripe = chunk_number;
2558 *dd_idx = sector_div(stripe, data_disks);
2561 * Select the parity disk based on the user selected algorithm.
2563 pd_idx = qd_idx = -1;
2564 switch(conf->level) {
2566 pd_idx = data_disks;
2569 switch (algorithm) {
2570 case ALGORITHM_LEFT_ASYMMETRIC:
2571 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2572 if (*dd_idx >= pd_idx)
2575 case ALGORITHM_RIGHT_ASYMMETRIC:
2576 pd_idx = sector_div(stripe2, raid_disks);
2577 if (*dd_idx >= pd_idx)
2580 case ALGORITHM_LEFT_SYMMETRIC:
2581 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2582 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2584 case ALGORITHM_RIGHT_SYMMETRIC:
2585 pd_idx = sector_div(stripe2, raid_disks);
2586 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2588 case ALGORITHM_PARITY_0:
2592 case ALGORITHM_PARITY_N:
2593 pd_idx = data_disks;
2601 switch (algorithm) {
2602 case ALGORITHM_LEFT_ASYMMETRIC:
2603 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2604 qd_idx = pd_idx + 1;
2605 if (pd_idx == raid_disks-1) {
2606 (*dd_idx)++; /* Q D D D P */
2608 } else if (*dd_idx >= pd_idx)
2609 (*dd_idx) += 2; /* D D P Q D */
2611 case ALGORITHM_RIGHT_ASYMMETRIC:
2612 pd_idx = sector_div(stripe2, raid_disks);
2613 qd_idx = pd_idx + 1;
2614 if (pd_idx == raid_disks-1) {
2615 (*dd_idx)++; /* Q D D D P */
2617 } else if (*dd_idx >= pd_idx)
2618 (*dd_idx) += 2; /* D D P Q D */
2620 case ALGORITHM_LEFT_SYMMETRIC:
2621 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2622 qd_idx = (pd_idx + 1) % raid_disks;
2623 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2625 case ALGORITHM_RIGHT_SYMMETRIC:
2626 pd_idx = sector_div(stripe2, raid_disks);
2627 qd_idx = (pd_idx + 1) % raid_disks;
2628 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2631 case ALGORITHM_PARITY_0:
2636 case ALGORITHM_PARITY_N:
2637 pd_idx = data_disks;
2638 qd_idx = data_disks + 1;
2641 case ALGORITHM_ROTATING_ZERO_RESTART:
2642 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2643 * of blocks for computing Q is different.
2645 pd_idx = sector_div(stripe2, raid_disks);
2646 qd_idx = pd_idx + 1;
2647 if (pd_idx == raid_disks-1) {
2648 (*dd_idx)++; /* Q D D D P */
2650 } else if (*dd_idx >= pd_idx)
2651 (*dd_idx) += 2; /* D D P Q D */
2655 case ALGORITHM_ROTATING_N_RESTART:
2656 /* Same a left_asymmetric, by first stripe is
2657 * D D D P Q rather than
2661 pd_idx = raid_disks - 1 - 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_CONTINUE:
2672 /* Same as left_symmetric but Q is before P */
2673 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2674 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2675 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2679 case ALGORITHM_LEFT_ASYMMETRIC_6:
2680 /* RAID5 left_asymmetric, with Q on last device */
2681 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2682 if (*dd_idx >= pd_idx)
2684 qd_idx = raid_disks - 1;
2687 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2688 pd_idx = sector_div(stripe2, raid_disks-1);
2689 if (*dd_idx >= pd_idx)
2691 qd_idx = raid_disks - 1;
2694 case ALGORITHM_LEFT_SYMMETRIC_6:
2695 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2696 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2697 qd_idx = raid_disks - 1;
2700 case ALGORITHM_RIGHT_SYMMETRIC_6:
2701 pd_idx = sector_div(stripe2, raid_disks-1);
2702 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2703 qd_idx = raid_disks - 1;
2706 case ALGORITHM_PARITY_0_6:
2709 qd_idx = raid_disks - 1;
2719 sh->pd_idx = pd_idx;
2720 sh->qd_idx = qd_idx;
2721 sh->ddf_layout = ddf_layout;
2724 * Finally, compute the new sector number
2726 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2730 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2732 struct r5conf *conf = sh->raid_conf;
2733 int raid_disks = sh->disks;
2734 int data_disks = raid_disks - conf->max_degraded;
2735 sector_t new_sector = sh->sector, check;
2736 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2737 : conf->chunk_sectors;
2738 int algorithm = previous ? conf->prev_algo
2742 sector_t chunk_number;
2743 int dummy1, dd_idx = i;
2745 struct stripe_head sh2;
2747 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2748 stripe = new_sector;
2750 if (i == sh->pd_idx)
2752 switch(conf->level) {
2755 switch (algorithm) {
2756 case ALGORITHM_LEFT_ASYMMETRIC:
2757 case ALGORITHM_RIGHT_ASYMMETRIC:
2761 case ALGORITHM_LEFT_SYMMETRIC:
2762 case ALGORITHM_RIGHT_SYMMETRIC:
2765 i -= (sh->pd_idx + 1);
2767 case ALGORITHM_PARITY_0:
2770 case ALGORITHM_PARITY_N:
2777 if (i == sh->qd_idx)
2778 return 0; /* It is the Q disk */
2779 switch (algorithm) {
2780 case ALGORITHM_LEFT_ASYMMETRIC:
2781 case ALGORITHM_RIGHT_ASYMMETRIC:
2782 case ALGORITHM_ROTATING_ZERO_RESTART:
2783 case ALGORITHM_ROTATING_N_RESTART:
2784 if (sh->pd_idx == raid_disks-1)
2785 i--; /* Q D D D P */
2786 else if (i > sh->pd_idx)
2787 i -= 2; /* D D P Q D */
2789 case ALGORITHM_LEFT_SYMMETRIC:
2790 case ALGORITHM_RIGHT_SYMMETRIC:
2791 if (sh->pd_idx == raid_disks-1)
2792 i--; /* Q D D D P */
2797 i -= (sh->pd_idx + 2);
2800 case ALGORITHM_PARITY_0:
2803 case ALGORITHM_PARITY_N:
2805 case ALGORITHM_ROTATING_N_CONTINUE:
2806 /* Like left_symmetric, but P is before Q */
2807 if (sh->pd_idx == 0)
2808 i--; /* P D D D Q */
2813 i -= (sh->pd_idx + 1);
2816 case ALGORITHM_LEFT_ASYMMETRIC_6:
2817 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2821 case ALGORITHM_LEFT_SYMMETRIC_6:
2822 case ALGORITHM_RIGHT_SYMMETRIC_6:
2824 i += data_disks + 1;
2825 i -= (sh->pd_idx + 1);
2827 case ALGORITHM_PARITY_0_6:
2836 chunk_number = stripe * data_disks + i;
2837 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2839 check = raid5_compute_sector(conf, r_sector,
2840 previous, &dummy1, &sh2);
2841 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2842 || sh2.qd_idx != sh->qd_idx) {
2843 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2844 mdname(conf->mddev));
2851 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2852 int rcw, int expand)
2854 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2855 struct r5conf *conf = sh->raid_conf;
2856 int level = conf->level;
2860 for (i = disks; i--; ) {
2861 struct r5dev *dev = &sh->dev[i];
2864 set_bit(R5_LOCKED, &dev->flags);
2865 set_bit(R5_Wantdrain, &dev->flags);
2867 clear_bit(R5_UPTODATE, &dev->flags);
2871 /* if we are not expanding this is a proper write request, and
2872 * there will be bios with new data to be drained into the
2877 /* False alarm, nothing to do */
2879 sh->reconstruct_state = reconstruct_state_drain_run;
2880 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2882 sh->reconstruct_state = reconstruct_state_run;
2884 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2886 if (s->locked + conf->max_degraded == disks)
2887 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2888 atomic_inc(&conf->pending_full_writes);
2890 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2891 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2892 BUG_ON(level == 6 &&
2893 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2894 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2896 for (i = disks; i--; ) {
2897 struct r5dev *dev = &sh->dev[i];
2898 if (i == pd_idx || i == qd_idx)
2902 (test_bit(R5_UPTODATE, &dev->flags) ||
2903 test_bit(R5_Wantcompute, &dev->flags))) {
2904 set_bit(R5_Wantdrain, &dev->flags);
2905 set_bit(R5_LOCKED, &dev->flags);
2906 clear_bit(R5_UPTODATE, &dev->flags);
2911 /* False alarm - nothing to do */
2913 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2914 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2915 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2916 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2919 /* keep the parity disk(s) locked while asynchronous operations
2922 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2923 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2927 int qd_idx = sh->qd_idx;
2928 struct r5dev *dev = &sh->dev[qd_idx];
2930 set_bit(R5_LOCKED, &dev->flags);
2931 clear_bit(R5_UPTODATE, &dev->flags);
2935 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2936 __func__, (unsigned long long)sh->sector,
2937 s->locked, s->ops_request);
2941 * Each stripe/dev can have one or more bion attached.
2942 * toread/towrite point to the first in a chain.
2943 * The bi_next chain must be in order.
2945 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2946 int forwrite, int previous)
2949 struct r5conf *conf = sh->raid_conf;
2952 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2953 (unsigned long long)bi->bi_iter.bi_sector,
2954 (unsigned long long)sh->sector);
2957 * If several bio share a stripe. The bio bi_phys_segments acts as a
2958 * reference count to avoid race. The reference count should already be
2959 * increased before this function is called (for example, in
2960 * make_request()), so other bio sharing this stripe will not free the
2961 * stripe. If a stripe is owned by one stripe, the stripe lock will
2964 spin_lock_irq(&sh->stripe_lock);
2965 /* Don't allow new IO added to stripes in batch list */
2969 bip = &sh->dev[dd_idx].towrite;
2973 bip = &sh->dev[dd_idx].toread;
2974 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2975 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2977 bip = & (*bip)->bi_next;
2979 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2982 if (!forwrite || previous)
2983 clear_bit(STRIPE_BATCH_READY, &sh->state);
2985 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2989 raid5_inc_bi_active_stripes(bi);
2992 /* check if page is covered */
2993 sector_t sector = sh->dev[dd_idx].sector;
2994 for (bi=sh->dev[dd_idx].towrite;
2995 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2996 bi && bi->bi_iter.bi_sector <= sector;
2997 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2998 if (bio_end_sector(bi) >= sector)
2999 sector = bio_end_sector(bi);
3001 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3002 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3003 sh->overwrite_disks++;
3006 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3007 (unsigned long long)(*bip)->bi_iter.bi_sector,
3008 (unsigned long long)sh->sector, dd_idx);
3010 if (conf->mddev->bitmap && firstwrite) {
3011 /* Cannot hold spinlock over bitmap_startwrite,
3012 * but must ensure this isn't added to a batch until
3013 * we have added to the bitmap and set bm_seq.
3014 * So set STRIPE_BITMAP_PENDING to prevent
3016 * If multiple add_stripe_bio() calls race here they
3017 * much all set STRIPE_BITMAP_PENDING. So only the first one
3018 * to complete "bitmap_startwrite" gets to set
3019 * STRIPE_BIT_DELAY. This is important as once a stripe
3020 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3023 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3024 spin_unlock_irq(&sh->stripe_lock);
3025 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3027 spin_lock_irq(&sh->stripe_lock);
3028 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3029 if (!sh->batch_head) {
3030 sh->bm_seq = conf->seq_flush+1;
3031 set_bit(STRIPE_BIT_DELAY, &sh->state);
3034 spin_unlock_irq(&sh->stripe_lock);
3036 if (stripe_can_batch(sh))
3037 stripe_add_to_batch_list(conf, sh);
3041 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3042 spin_unlock_irq(&sh->stripe_lock);
3046 static void end_reshape(struct r5conf *conf);
3048 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3049 struct stripe_head *sh)
3051 int sectors_per_chunk =
3052 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3054 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3055 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3057 raid5_compute_sector(conf,
3058 stripe * (disks - conf->max_degraded)
3059 *sectors_per_chunk + chunk_offset,
3065 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3066 struct stripe_head_state *s, int disks,
3067 struct bio_list *return_bi)
3070 BUG_ON(sh->batch_head);
3071 for (i = disks; i--; ) {
3075 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3076 struct md_rdev *rdev;
3078 rdev = rcu_dereference(conf->disks[i].rdev);
3079 if (rdev && test_bit(In_sync, &rdev->flags))
3080 atomic_inc(&rdev->nr_pending);
3085 if (!rdev_set_badblocks(
3089 md_error(conf->mddev, rdev);
3090 rdev_dec_pending(rdev, conf->mddev);
3093 spin_lock_irq(&sh->stripe_lock);
3094 /* fail all writes first */
3095 bi = sh->dev[i].towrite;
3096 sh->dev[i].towrite = NULL;
3097 sh->overwrite_disks = 0;
3098 spin_unlock_irq(&sh->stripe_lock);
3102 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3103 wake_up(&conf->wait_for_overlap);
3105 while (bi && bi->bi_iter.bi_sector <
3106 sh->dev[i].sector + STRIPE_SECTORS) {
3107 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3109 bi->bi_error = -EIO;
3110 if (!raid5_dec_bi_active_stripes(bi)) {
3111 md_write_end(conf->mddev);
3112 bio_list_add(return_bi, bi);
3117 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3118 STRIPE_SECTORS, 0, 0);
3120 /* and fail all 'written' */
3121 bi = sh->dev[i].written;
3122 sh->dev[i].written = NULL;
3123 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3124 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3125 sh->dev[i].page = sh->dev[i].orig_page;
3128 if (bi) bitmap_end = 1;
3129 while (bi && bi->bi_iter.bi_sector <
3130 sh->dev[i].sector + STRIPE_SECTORS) {
3131 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3133 bi->bi_error = -EIO;
3134 if (!raid5_dec_bi_active_stripes(bi)) {
3135 md_write_end(conf->mddev);
3136 bio_list_add(return_bi, bi);
3141 /* fail any reads if this device is non-operational and
3142 * the data has not reached the cache yet.
3144 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3145 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3146 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3147 spin_lock_irq(&sh->stripe_lock);
3148 bi = sh->dev[i].toread;
3149 sh->dev[i].toread = NULL;
3150 spin_unlock_irq(&sh->stripe_lock);
3151 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3152 wake_up(&conf->wait_for_overlap);
3155 while (bi && bi->bi_iter.bi_sector <
3156 sh->dev[i].sector + STRIPE_SECTORS) {
3157 struct bio *nextbi =
3158 r5_next_bio(bi, sh->dev[i].sector);
3160 bi->bi_error = -EIO;
3161 if (!raid5_dec_bi_active_stripes(bi))
3162 bio_list_add(return_bi, bi);
3167 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3168 STRIPE_SECTORS, 0, 0);
3169 /* If we were in the middle of a write the parity block might
3170 * still be locked - so just clear all R5_LOCKED flags
3172 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3177 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3178 if (atomic_dec_and_test(&conf->pending_full_writes))
3179 md_wakeup_thread(conf->mddev->thread);
3183 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3184 struct stripe_head_state *s)
3189 BUG_ON(sh->batch_head);
3190 clear_bit(STRIPE_SYNCING, &sh->state);
3191 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3192 wake_up(&conf->wait_for_overlap);
3195 /* There is nothing more to do for sync/check/repair.
3196 * Don't even need to abort as that is handled elsewhere
3197 * if needed, and not always wanted e.g. if there is a known
3199 * For recover/replace we need to record a bad block on all
3200 * non-sync devices, or abort the recovery
3202 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3203 /* During recovery devices cannot be removed, so
3204 * locking and refcounting of rdevs is not needed
3206 for (i = 0; i < conf->raid_disks; i++) {
3207 struct md_rdev *rdev = conf->disks[i].rdev;
3209 && !test_bit(Faulty, &rdev->flags)
3210 && !test_bit(In_sync, &rdev->flags)
3211 && !rdev_set_badblocks(rdev, sh->sector,
3214 rdev = conf->disks[i].replacement;
3216 && !test_bit(Faulty, &rdev->flags)
3217 && !test_bit(In_sync, &rdev->flags)
3218 && !rdev_set_badblocks(rdev, sh->sector,
3223 conf->recovery_disabled =
3224 conf->mddev->recovery_disabled;
3226 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3229 static int want_replace(struct stripe_head *sh, int disk_idx)
3231 struct md_rdev *rdev;
3233 /* Doing recovery so rcu locking not required */
3234 rdev = sh->raid_conf->disks[disk_idx].replacement;
3236 && !test_bit(Faulty, &rdev->flags)
3237 && !test_bit(In_sync, &rdev->flags)
3238 && (rdev->recovery_offset <= sh->sector
3239 || rdev->mddev->recovery_cp <= sh->sector))
3245 /* fetch_block - checks the given member device to see if its data needs
3246 * to be read or computed to satisfy a request.
3248 * Returns 1 when no more member devices need to be checked, otherwise returns
3249 * 0 to tell the loop in handle_stripe_fill to continue
3252 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3253 int disk_idx, int disks)
3255 struct r5dev *dev = &sh->dev[disk_idx];
3256 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3257 &sh->dev[s->failed_num[1]] };
3261 if (test_bit(R5_LOCKED, &dev->flags) ||
3262 test_bit(R5_UPTODATE, &dev->flags))
3263 /* No point reading this as we already have it or have
3264 * decided to get it.
3269 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3270 /* We need this block to directly satisfy a request */
3273 if (s->syncing || s->expanding ||
3274 (s->replacing && want_replace(sh, disk_idx)))
3275 /* When syncing, or expanding we read everything.
3276 * When replacing, we need the replaced block.
3280 if ((s->failed >= 1 && fdev[0]->toread) ||
3281 (s->failed >= 2 && fdev[1]->toread))
3282 /* If we want to read from a failed device, then
3283 * we need to actually read every other device.
3287 /* Sometimes neither read-modify-write nor reconstruct-write
3288 * cycles can work. In those cases we read every block we
3289 * can. Then the parity-update is certain to have enough to
3291 * This can only be a problem when we need to write something,
3292 * and some device has failed. If either of those tests
3293 * fail we need look no further.
3295 if (!s->failed || !s->to_write)
3298 if (test_bit(R5_Insync, &dev->flags) &&
3299 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3300 /* Pre-reads at not permitted until after short delay
3301 * to gather multiple requests. However if this
3302 * device is no Insync, the block could only be be computed
3303 * and there is no need to delay that.
3307 for (i = 0; i < s->failed; i++) {
3308 if (fdev[i]->towrite &&
3309 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3310 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3311 /* If we have a partial write to a failed
3312 * device, then we will need to reconstruct
3313 * the content of that device, so all other
3314 * devices must be read.
3319 /* If we are forced to do a reconstruct-write, either because
3320 * the current RAID6 implementation only supports that, or
3321 * or because parity cannot be trusted and we are currently
3322 * recovering it, there is extra need to be careful.
3323 * If one of the devices that we would need to read, because
3324 * it is not being overwritten (and maybe not written at all)
3325 * is missing/faulty, then we need to read everything we can.
3327 if (sh->raid_conf->level != 6 &&
3328 sh->sector < sh->raid_conf->mddev->recovery_cp)
3329 /* reconstruct-write isn't being forced */
3331 for (i = 0; i < s->failed; i++) {
3332 if (s->failed_num[i] != sh->pd_idx &&
3333 s->failed_num[i] != sh->qd_idx &&
3334 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3335 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3342 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3343 int disk_idx, int disks)
3345 struct r5dev *dev = &sh->dev[disk_idx];
3347 /* is the data in this block needed, and can we get it? */
3348 if (need_this_block(sh, s, disk_idx, disks)) {
3349 /* we would like to get this block, possibly by computing it,
3350 * otherwise read it if the backing disk is insync
3352 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3353 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3354 BUG_ON(sh->batch_head);
3355 if ((s->uptodate == disks - 1) &&
3356 (s->failed && (disk_idx == s->failed_num[0] ||
3357 disk_idx == s->failed_num[1]))) {
3358 /* have disk failed, and we're requested to fetch it;
3361 pr_debug("Computing stripe %llu block %d\n",
3362 (unsigned long long)sh->sector, disk_idx);
3363 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3364 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3365 set_bit(R5_Wantcompute, &dev->flags);
3366 sh->ops.target = disk_idx;
3367 sh->ops.target2 = -1; /* no 2nd target */
3369 /* Careful: from this point on 'uptodate' is in the eye
3370 * of raid_run_ops which services 'compute' operations
3371 * before writes. R5_Wantcompute flags a block that will
3372 * be R5_UPTODATE by the time it is needed for a
3373 * subsequent operation.
3377 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3378 /* Computing 2-failure is *very* expensive; only
3379 * do it if failed >= 2
3382 for (other = disks; other--; ) {
3383 if (other == disk_idx)
3385 if (!test_bit(R5_UPTODATE,
3386 &sh->dev[other].flags))
3390 pr_debug("Computing stripe %llu blocks %d,%d\n",
3391 (unsigned long long)sh->sector,
3393 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3394 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3395 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3396 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3397 sh->ops.target = disk_idx;
3398 sh->ops.target2 = other;
3402 } else if (test_bit(R5_Insync, &dev->flags)) {
3403 set_bit(R5_LOCKED, &dev->flags);
3404 set_bit(R5_Wantread, &dev->flags);
3406 pr_debug("Reading block %d (sync=%d)\n",
3407 disk_idx, s->syncing);
3415 * handle_stripe_fill - read or compute data to satisfy pending requests.
3417 static void handle_stripe_fill(struct stripe_head *sh,
3418 struct stripe_head_state *s,
3423 /* look for blocks to read/compute, skip this if a compute
3424 * is already in flight, or if the stripe contents are in the
3425 * midst of changing due to a write
3427 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3428 !sh->reconstruct_state)
3429 for (i = disks; i--; )
3430 if (fetch_block(sh, s, i, disks))
3432 set_bit(STRIPE_HANDLE, &sh->state);
3435 static void break_stripe_batch_list(struct stripe_head *head_sh,
3436 unsigned long handle_flags);
3437 /* handle_stripe_clean_event
3438 * any written block on an uptodate or failed drive can be returned.
3439 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3440 * never LOCKED, so we don't need to test 'failed' directly.
3442 static void handle_stripe_clean_event(struct r5conf *conf,
3443 struct stripe_head *sh, int disks, struct bio_list *return_bi)
3447 int discard_pending = 0;
3448 struct stripe_head *head_sh = sh;
3449 bool do_endio = false;
3451 for (i = disks; i--; )
3452 if (sh->dev[i].written) {
3454 if (!test_bit(R5_LOCKED, &dev->flags) &&
3455 (test_bit(R5_UPTODATE, &dev->flags) ||
3456 test_bit(R5_Discard, &dev->flags) ||
3457 test_bit(R5_SkipCopy, &dev->flags))) {
3458 /* We can return any write requests */
3459 struct bio *wbi, *wbi2;
3460 pr_debug("Return write for disc %d\n", i);
3461 if (test_and_clear_bit(R5_Discard, &dev->flags))
3462 clear_bit(R5_UPTODATE, &dev->flags);
3463 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3464 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3469 dev->page = dev->orig_page;
3471 dev->written = NULL;
3472 while (wbi && wbi->bi_iter.bi_sector <
3473 dev->sector + STRIPE_SECTORS) {
3474 wbi2 = r5_next_bio(wbi, dev->sector);
3475 if (!raid5_dec_bi_active_stripes(wbi)) {
3476 md_write_end(conf->mddev);
3477 bio_list_add(return_bi, wbi);
3481 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3483 !test_bit(STRIPE_DEGRADED, &sh->state),
3485 if (head_sh->batch_head) {
3486 sh = list_first_entry(&sh->batch_list,
3489 if (sh != head_sh) {
3496 } else if (test_bit(R5_Discard, &dev->flags))
3497 discard_pending = 1;
3498 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3499 WARN_ON(dev->page != dev->orig_page);
3501 if (!discard_pending &&
3502 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3503 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3504 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3505 if (sh->qd_idx >= 0) {
3506 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3507 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3509 /* now that discard is done we can proceed with any sync */
3510 clear_bit(STRIPE_DISCARD, &sh->state);
3512 * SCSI discard will change some bio fields and the stripe has
3513 * no updated data, so remove it from hash list and the stripe
3514 * will be reinitialized
3516 spin_lock_irq(&conf->device_lock);
3519 if (head_sh->batch_head) {
3520 sh = list_first_entry(&sh->batch_list,
3521 struct stripe_head, batch_list);
3525 spin_unlock_irq(&conf->device_lock);
3528 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3529 set_bit(STRIPE_HANDLE, &sh->state);
3533 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3534 if (atomic_dec_and_test(&conf->pending_full_writes))
3535 md_wakeup_thread(conf->mddev->thread);
3537 if (head_sh->batch_head && do_endio)
3538 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3541 static void handle_stripe_dirtying(struct r5conf *conf,
3542 struct stripe_head *sh,
3543 struct stripe_head_state *s,
3546 int rmw = 0, rcw = 0, i;
3547 sector_t recovery_cp = conf->mddev->recovery_cp;
3549 /* Check whether resync is now happening or should start.
3550 * If yes, then the array is dirty (after unclean shutdown or
3551 * initial creation), so parity in some stripes might be inconsistent.
3552 * In this case, we need to always do reconstruct-write, to ensure
3553 * that in case of drive failure or read-error correction, we
3554 * generate correct data from the parity.
3556 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3557 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3559 /* Calculate the real rcw later - for now make it
3560 * look like rcw is cheaper
3563 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3564 conf->rmw_level, (unsigned long long)recovery_cp,
3565 (unsigned long long)sh->sector);
3566 } else for (i = disks; i--; ) {
3567 /* would I have to read this buffer for read_modify_write */
3568 struct r5dev *dev = &sh->dev[i];
3569 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3570 !test_bit(R5_LOCKED, &dev->flags) &&
3571 !(test_bit(R5_UPTODATE, &dev->flags) ||
3572 test_bit(R5_Wantcompute, &dev->flags))) {
3573 if (test_bit(R5_Insync, &dev->flags))
3576 rmw += 2*disks; /* cannot read it */
3578 /* Would I have to read this buffer for reconstruct_write */
3579 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3580 i != sh->pd_idx && i != sh->qd_idx &&
3581 !test_bit(R5_LOCKED, &dev->flags) &&
3582 !(test_bit(R5_UPTODATE, &dev->flags) ||
3583 test_bit(R5_Wantcompute, &dev->flags))) {
3584 if (test_bit(R5_Insync, &dev->flags))
3590 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3591 (unsigned long long)sh->sector, rmw, rcw);
3592 set_bit(STRIPE_HANDLE, &sh->state);
3593 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3594 /* prefer read-modify-write, but need to get some data */
3595 if (conf->mddev->queue)
3596 blk_add_trace_msg(conf->mddev->queue,
3597 "raid5 rmw %llu %d",
3598 (unsigned long long)sh->sector, rmw);
3599 for (i = disks; i--; ) {
3600 struct r5dev *dev = &sh->dev[i];
3601 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3602 !test_bit(R5_LOCKED, &dev->flags) &&
3603 !(test_bit(R5_UPTODATE, &dev->flags) ||
3604 test_bit(R5_Wantcompute, &dev->flags)) &&
3605 test_bit(R5_Insync, &dev->flags)) {
3606 if (test_bit(STRIPE_PREREAD_ACTIVE,
3608 pr_debug("Read_old block %d for r-m-w\n",
3610 set_bit(R5_LOCKED, &dev->flags);
3611 set_bit(R5_Wantread, &dev->flags);
3614 set_bit(STRIPE_DELAYED, &sh->state);
3615 set_bit(STRIPE_HANDLE, &sh->state);
3620 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3621 /* want reconstruct write, but need to get some data */
3624 for (i = disks; i--; ) {
3625 struct r5dev *dev = &sh->dev[i];
3626 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3627 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))) {
3632 if (test_bit(R5_Insync, &dev->flags) &&
3633 test_bit(STRIPE_PREREAD_ACTIVE,
3635 pr_debug("Read_old block "
3636 "%d for Reconstruct\n", i);
3637 set_bit(R5_LOCKED, &dev->flags);
3638 set_bit(R5_Wantread, &dev->flags);
3642 set_bit(STRIPE_DELAYED, &sh->state);
3643 set_bit(STRIPE_HANDLE, &sh->state);
3647 if (rcw && conf->mddev->queue)
3648 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3649 (unsigned long long)sh->sector,
3650 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3653 if (rcw > disks && rmw > disks &&
3654 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3655 set_bit(STRIPE_DELAYED, &sh->state);
3657 /* now if nothing is locked, and if we have enough data,
3658 * we can start a write request
3660 /* since handle_stripe can be called at any time we need to handle the
3661 * case where a compute block operation has been submitted and then a
3662 * subsequent call wants to start a write request. raid_run_ops only
3663 * handles the case where compute block and reconstruct are requested
3664 * simultaneously. If this is not the case then new writes need to be
3665 * held off until the compute completes.
3667 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3668 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3669 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3670 schedule_reconstruction(sh, s, rcw == 0, 0);
3673 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3674 struct stripe_head_state *s, int disks)
3676 struct r5dev *dev = NULL;
3678 BUG_ON(sh->batch_head);
3679 set_bit(STRIPE_HANDLE, &sh->state);
3681 switch (sh->check_state) {
3682 case check_state_idle:
3683 /* start a new check operation if there are no failures */
3684 if (s->failed == 0) {
3685 BUG_ON(s->uptodate != disks);
3686 sh->check_state = check_state_run;
3687 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3688 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3692 dev = &sh->dev[s->failed_num[0]];
3694 case check_state_compute_result:
3695 sh->check_state = check_state_idle;
3697 dev = &sh->dev[sh->pd_idx];
3699 /* check that a write has not made the stripe insync */
3700 if (test_bit(STRIPE_INSYNC, &sh->state))
3703 /* either failed parity check, or recovery is happening */
3704 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3705 BUG_ON(s->uptodate != disks);
3707 set_bit(R5_LOCKED, &dev->flags);
3709 set_bit(R5_Wantwrite, &dev->flags);
3711 clear_bit(STRIPE_DEGRADED, &sh->state);
3712 set_bit(STRIPE_INSYNC, &sh->state);
3714 case check_state_run:
3715 break; /* we will be called again upon completion */
3716 case check_state_check_result:
3717 sh->check_state = check_state_idle;
3719 /* if a failure occurred during the check operation, leave
3720 * STRIPE_INSYNC not set and let the stripe be handled again
3725 /* handle a successful check operation, if parity is correct
3726 * we are done. Otherwise update the mismatch count and repair
3727 * parity if !MD_RECOVERY_CHECK
3729 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3730 /* parity is correct (on disc,
3731 * not in buffer any more)
3733 set_bit(STRIPE_INSYNC, &sh->state);
3735 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3736 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3737 /* don't try to repair!! */
3738 set_bit(STRIPE_INSYNC, &sh->state);
3740 sh->check_state = check_state_compute_run;
3741 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3742 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3743 set_bit(R5_Wantcompute,
3744 &sh->dev[sh->pd_idx].flags);
3745 sh->ops.target = sh->pd_idx;
3746 sh->ops.target2 = -1;
3751 case check_state_compute_run:
3754 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3755 __func__, sh->check_state,
3756 (unsigned long long) sh->sector);
3761 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3762 struct stripe_head_state *s,
3765 int pd_idx = sh->pd_idx;
3766 int qd_idx = sh->qd_idx;
3769 BUG_ON(sh->batch_head);
3770 set_bit(STRIPE_HANDLE, &sh->state);
3772 BUG_ON(s->failed > 2);
3774 /* Want to check and possibly repair P and Q.
3775 * However there could be one 'failed' device, in which
3776 * case we can only check one of them, possibly using the
3777 * other to generate missing data
3780 switch (sh->check_state) {
3781 case check_state_idle:
3782 /* start a new check operation if there are < 2 failures */
3783 if (s->failed == s->q_failed) {
3784 /* The only possible failed device holds Q, so it
3785 * makes sense to check P (If anything else were failed,
3786 * we would have used P to recreate it).
3788 sh->check_state = check_state_run;
3790 if (!s->q_failed && s->failed < 2) {
3791 /* Q is not failed, and we didn't use it to generate
3792 * anything, so it makes sense to check it
3794 if (sh->check_state == check_state_run)
3795 sh->check_state = check_state_run_pq;
3797 sh->check_state = check_state_run_q;
3800 /* discard potentially stale zero_sum_result */
3801 sh->ops.zero_sum_result = 0;
3803 if (sh->check_state == check_state_run) {
3804 /* async_xor_zero_sum destroys the contents of P */
3805 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3808 if (sh->check_state >= check_state_run &&
3809 sh->check_state <= check_state_run_pq) {
3810 /* async_syndrome_zero_sum preserves P and Q, so
3811 * no need to mark them !uptodate here
3813 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3817 /* we have 2-disk failure */
3818 BUG_ON(s->failed != 2);
3820 case check_state_compute_result:
3821 sh->check_state = check_state_idle;
3823 /* check that a write has not made the stripe insync */
3824 if (test_bit(STRIPE_INSYNC, &sh->state))
3827 /* now write out any block on a failed drive,
3828 * or P or Q if they were recomputed
3830 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3831 if (s->failed == 2) {
3832 dev = &sh->dev[s->failed_num[1]];
3834 set_bit(R5_LOCKED, &dev->flags);
3835 set_bit(R5_Wantwrite, &dev->flags);
3837 if (s->failed >= 1) {
3838 dev = &sh->dev[s->failed_num[0]];
3840 set_bit(R5_LOCKED, &dev->flags);
3841 set_bit(R5_Wantwrite, &dev->flags);
3843 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3844 dev = &sh->dev[pd_idx];
3846 set_bit(R5_LOCKED, &dev->flags);
3847 set_bit(R5_Wantwrite, &dev->flags);
3849 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3850 dev = &sh->dev[qd_idx];
3852 set_bit(R5_LOCKED, &dev->flags);
3853 set_bit(R5_Wantwrite, &dev->flags);
3855 clear_bit(STRIPE_DEGRADED, &sh->state);
3857 set_bit(STRIPE_INSYNC, &sh->state);
3859 case check_state_run:
3860 case check_state_run_q:
3861 case check_state_run_pq:
3862 break; /* we will be called again upon completion */
3863 case check_state_check_result:
3864 sh->check_state = check_state_idle;
3866 /* handle a successful check operation, if parity is correct
3867 * we are done. Otherwise update the mismatch count and repair
3868 * parity if !MD_RECOVERY_CHECK
3870 if (sh->ops.zero_sum_result == 0) {
3871 /* both parities are correct */
3873 set_bit(STRIPE_INSYNC, &sh->state);
3875 /* in contrast to the raid5 case we can validate
3876 * parity, but still have a failure to write
3879 sh->check_state = check_state_compute_result;
3880 /* Returning at this point means that we may go
3881 * off and bring p and/or q uptodate again so
3882 * we make sure to check zero_sum_result again
3883 * to verify if p or q need writeback
3887 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3888 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3889 /* don't try to repair!! */
3890 set_bit(STRIPE_INSYNC, &sh->state);
3892 int *target = &sh->ops.target;
3894 sh->ops.target = -1;
3895 sh->ops.target2 = -1;
3896 sh->check_state = check_state_compute_run;
3897 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3898 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3899 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3900 set_bit(R5_Wantcompute,
3901 &sh->dev[pd_idx].flags);
3903 target = &sh->ops.target2;
3906 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3907 set_bit(R5_Wantcompute,
3908 &sh->dev[qd_idx].flags);
3915 case check_state_compute_run:
3918 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3919 __func__, sh->check_state,
3920 (unsigned long long) sh->sector);
3925 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3929 /* We have read all the blocks in this stripe and now we need to
3930 * copy some of them into a target stripe for expand.
3932 struct dma_async_tx_descriptor *tx = NULL;
3933 BUG_ON(sh->batch_head);
3934 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3935 for (i = 0; i < sh->disks; i++)
3936 if (i != sh->pd_idx && i != sh->qd_idx) {
3938 struct stripe_head *sh2;
3939 struct async_submit_ctl submit;
3941 sector_t bn = compute_blocknr(sh, i, 1);
3942 sector_t s = raid5_compute_sector(conf, bn, 0,
3944 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3946 /* so far only the early blocks of this stripe
3947 * have been requested. When later blocks
3948 * get requested, we will try again
3951 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3952 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3953 /* must have already done this block */
3954 release_stripe(sh2);
3958 /* place all the copies on one channel */
3959 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3960 tx = async_memcpy(sh2->dev[dd_idx].page,
3961 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3964 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3965 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3966 for (j = 0; j < conf->raid_disks; j++)
3967 if (j != sh2->pd_idx &&
3969 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3971 if (j == conf->raid_disks) {
3972 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3973 set_bit(STRIPE_HANDLE, &sh2->state);
3975 release_stripe(sh2);
3978 /* done submitting copies, wait for them to complete */
3979 async_tx_quiesce(&tx);
3983 * handle_stripe - do things to a stripe.
3985 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3986 * state of various bits to see what needs to be done.
3988 * return some read requests which now have data
3989 * return some write requests which are safely on storage
3990 * schedule a read on some buffers
3991 * schedule a write of some buffers
3992 * return confirmation of parity correctness
3996 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3998 struct r5conf *conf = sh->raid_conf;
3999 int disks = sh->disks;
4002 int do_recovery = 0;
4004 memset(s, 0, sizeof(*s));
4006 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4007 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4008 s->failed_num[0] = -1;
4009 s->failed_num[1] = -1;
4011 /* Now to look around and see what can be done */
4013 for (i=disks; i--; ) {
4014 struct md_rdev *rdev;
4021 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4023 dev->toread, dev->towrite, dev->written);
4024 /* maybe we can reply to a read
4026 * new wantfill requests are only permitted while
4027 * ops_complete_biofill is guaranteed to be inactive
4029 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4030 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4031 set_bit(R5_Wantfill, &dev->flags);
4033 /* now count some things */
4034 if (test_bit(R5_LOCKED, &dev->flags))
4036 if (test_bit(R5_UPTODATE, &dev->flags))
4038 if (test_bit(R5_Wantcompute, &dev->flags)) {
4040 BUG_ON(s->compute > 2);
4043 if (test_bit(R5_Wantfill, &dev->flags))
4045 else if (dev->toread)
4049 if (!test_bit(R5_OVERWRITE, &dev->flags))
4054 /* Prefer to use the replacement for reads, but only
4055 * if it is recovered enough and has no bad blocks.
4057 rdev = rcu_dereference(conf->disks[i].replacement);
4058 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4059 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4060 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4061 &first_bad, &bad_sectors))
4062 set_bit(R5_ReadRepl, &dev->flags);
4064 if (rdev && !test_bit(Faulty, &rdev->flags))
4065 set_bit(R5_NeedReplace, &dev->flags);
4067 clear_bit(R5_NeedReplace, &dev->flags);
4068 rdev = rcu_dereference(conf->disks[i].rdev);
4069 clear_bit(R5_ReadRepl, &dev->flags);
4071 if (rdev && test_bit(Faulty, &rdev->flags))
4074 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4075 &first_bad, &bad_sectors);
4076 if (s->blocked_rdev == NULL
4077 && (test_bit(Blocked, &rdev->flags)
4080 set_bit(BlockedBadBlocks,
4082 s->blocked_rdev = rdev;
4083 atomic_inc(&rdev->nr_pending);
4086 clear_bit(R5_Insync, &dev->flags);
4090 /* also not in-sync */
4091 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4092 test_bit(R5_UPTODATE, &dev->flags)) {
4093 /* treat as in-sync, but with a read error
4094 * which we can now try to correct
4096 set_bit(R5_Insync, &dev->flags);
4097 set_bit(R5_ReadError, &dev->flags);
4099 } else if (test_bit(In_sync, &rdev->flags))
4100 set_bit(R5_Insync, &dev->flags);
4101 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4102 /* in sync if before recovery_offset */
4103 set_bit(R5_Insync, &dev->flags);
4104 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4105 test_bit(R5_Expanded, &dev->flags))
4106 /* If we've reshaped into here, we assume it is Insync.
4107 * We will shortly update recovery_offset to make
4110 set_bit(R5_Insync, &dev->flags);
4112 if (test_bit(R5_WriteError, &dev->flags)) {
4113 /* This flag does not apply to '.replacement'
4114 * only to .rdev, so make sure to check that*/
4115 struct md_rdev *rdev2 = rcu_dereference(
4116 conf->disks[i].rdev);
4118 clear_bit(R5_Insync, &dev->flags);
4119 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4120 s->handle_bad_blocks = 1;
4121 atomic_inc(&rdev2->nr_pending);
4123 clear_bit(R5_WriteError, &dev->flags);
4125 if (test_bit(R5_MadeGood, &dev->flags)) {
4126 /* This flag does not apply to '.replacement'
4127 * only to .rdev, so make sure to check that*/
4128 struct md_rdev *rdev2 = rcu_dereference(
4129 conf->disks[i].rdev);
4130 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4131 s->handle_bad_blocks = 1;
4132 atomic_inc(&rdev2->nr_pending);
4134 clear_bit(R5_MadeGood, &dev->flags);
4136 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4137 struct md_rdev *rdev2 = rcu_dereference(
4138 conf->disks[i].replacement);
4139 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4140 s->handle_bad_blocks = 1;
4141 atomic_inc(&rdev2->nr_pending);
4143 clear_bit(R5_MadeGoodRepl, &dev->flags);
4145 if (!test_bit(R5_Insync, &dev->flags)) {
4146 /* The ReadError flag will just be confusing now */
4147 clear_bit(R5_ReadError, &dev->flags);
4148 clear_bit(R5_ReWrite, &dev->flags);
4150 if (test_bit(R5_ReadError, &dev->flags))
4151 clear_bit(R5_Insync, &dev->flags);
4152 if (!test_bit(R5_Insync, &dev->flags)) {
4154 s->failed_num[s->failed] = i;
4156 if (rdev && !test_bit(Faulty, &rdev->flags))
4160 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4161 /* If there is a failed device being replaced,
4162 * we must be recovering.
4163 * else if we are after recovery_cp, we must be syncing
4164 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4165 * else we can only be replacing
4166 * sync and recovery both need to read all devices, and so
4167 * use the same flag.
4170 sh->sector >= conf->mddev->recovery_cp ||
4171 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4179 static int clear_batch_ready(struct stripe_head *sh)
4181 /* Return '1' if this is a member of batch, or
4182 * '0' if it is a lone stripe or a head which can now be
4185 struct stripe_head *tmp;
4186 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4187 return (sh->batch_head && sh->batch_head != sh);
4188 spin_lock(&sh->stripe_lock);
4189 if (!sh->batch_head) {
4190 spin_unlock(&sh->stripe_lock);
4195 * this stripe could be added to a batch list before we check
4196 * BATCH_READY, skips it
4198 if (sh->batch_head != sh) {
4199 spin_unlock(&sh->stripe_lock);
4202 spin_lock(&sh->batch_lock);
4203 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4204 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4205 spin_unlock(&sh->batch_lock);
4206 spin_unlock(&sh->stripe_lock);
4209 * BATCH_READY is cleared, no new stripes can be added.
4210 * batch_list can be accessed without lock
4215 static void break_stripe_batch_list(struct stripe_head *head_sh,
4216 unsigned long handle_flags)
4218 struct stripe_head *sh, *next;
4222 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4224 list_del_init(&sh->batch_list);
4226 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4227 (1 << STRIPE_SYNCING) |
4228 (1 << STRIPE_REPLACED) |
4229 (1 << STRIPE_PREREAD_ACTIVE) |
4230 (1 << STRIPE_DELAYED) |
4231 (1 << STRIPE_BIT_DELAY) |
4232 (1 << STRIPE_FULL_WRITE) |
4233 (1 << STRIPE_BIOFILL_RUN) |
4234 (1 << STRIPE_COMPUTE_RUN) |
4235 (1 << STRIPE_OPS_REQ_PENDING) |
4236 (1 << STRIPE_DISCARD) |
4237 (1 << STRIPE_BATCH_READY) |
4238 (1 << STRIPE_BATCH_ERR) |
4239 (1 << STRIPE_BITMAP_PENDING)));
4240 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4241 (1 << STRIPE_REPLACED)));
4243 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4244 (1 << STRIPE_DEGRADED)),
4245 head_sh->state & (1 << STRIPE_INSYNC));
4247 sh->check_state = head_sh->check_state;
4248 sh->reconstruct_state = head_sh->reconstruct_state;
4249 for (i = 0; i < sh->disks; i++) {
4250 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4252 sh->dev[i].flags = head_sh->dev[i].flags &
4253 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4255 spin_lock_irq(&sh->stripe_lock);
4256 sh->batch_head = NULL;
4257 spin_unlock_irq(&sh->stripe_lock);
4258 if (handle_flags == 0 ||
4259 sh->state & handle_flags)
4260 set_bit(STRIPE_HANDLE, &sh->state);
4263 spin_lock_irq(&head_sh->stripe_lock);
4264 head_sh->batch_head = NULL;
4265 spin_unlock_irq(&head_sh->stripe_lock);
4266 for (i = 0; i < head_sh->disks; i++)
4267 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4269 if (head_sh->state & handle_flags)
4270 set_bit(STRIPE_HANDLE, &head_sh->state);
4273 wake_up(&head_sh->raid_conf->wait_for_overlap);
4276 static void handle_stripe(struct stripe_head *sh)
4278 struct stripe_head_state s;
4279 struct r5conf *conf = sh->raid_conf;
4282 int disks = sh->disks;
4283 struct r5dev *pdev, *qdev;
4285 clear_bit(STRIPE_HANDLE, &sh->state);
4286 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4287 /* already being handled, ensure it gets handled
4288 * again when current action finishes */
4289 set_bit(STRIPE_HANDLE, &sh->state);
4293 if (clear_batch_ready(sh) ) {
4294 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4298 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4299 break_stripe_batch_list(sh, 0);
4301 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4302 spin_lock(&sh->stripe_lock);
4303 /* Cannot process 'sync' concurrently with 'discard' */
4304 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4305 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4306 set_bit(STRIPE_SYNCING, &sh->state);
4307 clear_bit(STRIPE_INSYNC, &sh->state);
4308 clear_bit(STRIPE_REPLACED, &sh->state);
4310 spin_unlock(&sh->stripe_lock);
4312 clear_bit(STRIPE_DELAYED, &sh->state);
4314 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4315 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4316 (unsigned long long)sh->sector, sh->state,
4317 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4318 sh->check_state, sh->reconstruct_state);
4320 analyse_stripe(sh, &s);
4322 if (s.handle_bad_blocks) {
4323 set_bit(STRIPE_HANDLE, &sh->state);
4327 if (unlikely(s.blocked_rdev)) {
4328 if (s.syncing || s.expanding || s.expanded ||
4329 s.replacing || s.to_write || s.written) {
4330 set_bit(STRIPE_HANDLE, &sh->state);
4333 /* There is nothing for the blocked_rdev to block */
4334 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4335 s.blocked_rdev = NULL;
4338 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4339 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4340 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4343 pr_debug("locked=%d uptodate=%d to_read=%d"
4344 " to_write=%d failed=%d failed_num=%d,%d\n",
4345 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4346 s.failed_num[0], s.failed_num[1]);
4347 /* check if the array has lost more than max_degraded devices and,
4348 * if so, some requests might need to be failed.
4350 if (s.failed > conf->max_degraded) {
4351 sh->check_state = 0;
4352 sh->reconstruct_state = 0;
4353 break_stripe_batch_list(sh, 0);
4354 if (s.to_read+s.to_write+s.written)
4355 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4356 if (s.syncing + s.replacing)
4357 handle_failed_sync(conf, sh, &s);
4360 /* Now we check to see if any write operations have recently
4364 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4366 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4367 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4368 sh->reconstruct_state = reconstruct_state_idle;
4370 /* All the 'written' buffers and the parity block are ready to
4371 * be written back to disk
4373 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4374 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4375 BUG_ON(sh->qd_idx >= 0 &&
4376 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4377 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4378 for (i = disks; i--; ) {
4379 struct r5dev *dev = &sh->dev[i];
4380 if (test_bit(R5_LOCKED, &dev->flags) &&
4381 (i == sh->pd_idx || i == sh->qd_idx ||
4383 pr_debug("Writing block %d\n", i);
4384 set_bit(R5_Wantwrite, &dev->flags);
4389 if (!test_bit(R5_Insync, &dev->flags) ||
4390 ((i == sh->pd_idx || i == sh->qd_idx) &&
4392 set_bit(STRIPE_INSYNC, &sh->state);
4395 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4396 s.dec_preread_active = 1;
4400 * might be able to return some write requests if the parity blocks
4401 * are safe, or on a failed drive
4403 pdev = &sh->dev[sh->pd_idx];
4404 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4405 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4406 qdev = &sh->dev[sh->qd_idx];
4407 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4408 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4412 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4413 && !test_bit(R5_LOCKED, &pdev->flags)
4414 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4415 test_bit(R5_Discard, &pdev->flags))))) &&
4416 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4417 && !test_bit(R5_LOCKED, &qdev->flags)
4418 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4419 test_bit(R5_Discard, &qdev->flags))))))
4420 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4422 /* Now we might consider reading some blocks, either to check/generate
4423 * parity, or to satisfy requests
4424 * or to load a block that is being partially written.
4426 if (s.to_read || s.non_overwrite
4427 || (conf->level == 6 && s.to_write && s.failed)
4428 || (s.syncing && (s.uptodate + s.compute < disks))
4431 handle_stripe_fill(sh, &s, disks);
4433 /* Now to consider new write requests and what else, if anything
4434 * should be read. We do not handle new writes when:
4435 * 1/ A 'write' operation (copy+xor) is already in flight.
4436 * 2/ A 'check' operation is in flight, as it may clobber the parity
4439 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4440 handle_stripe_dirtying(conf, sh, &s, disks);
4442 /* maybe we need to check and possibly fix the parity for this stripe
4443 * Any reads will already have been scheduled, so we just see if enough
4444 * data is available. The parity check is held off while parity
4445 * dependent operations are in flight.
4447 if (sh->check_state ||
4448 (s.syncing && s.locked == 0 &&
4449 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4450 !test_bit(STRIPE_INSYNC, &sh->state))) {
4451 if (conf->level == 6)
4452 handle_parity_checks6(conf, sh, &s, disks);
4454 handle_parity_checks5(conf, sh, &s, disks);
4457 if ((s.replacing || s.syncing) && s.locked == 0
4458 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4459 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4460 /* Write out to replacement devices where possible */
4461 for (i = 0; i < conf->raid_disks; i++)
4462 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4463 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4464 set_bit(R5_WantReplace, &sh->dev[i].flags);
4465 set_bit(R5_LOCKED, &sh->dev[i].flags);
4469 set_bit(STRIPE_INSYNC, &sh->state);
4470 set_bit(STRIPE_REPLACED, &sh->state);
4472 if ((s.syncing || s.replacing) && s.locked == 0 &&
4473 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4474 test_bit(STRIPE_INSYNC, &sh->state)) {
4475 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4476 clear_bit(STRIPE_SYNCING, &sh->state);
4477 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4478 wake_up(&conf->wait_for_overlap);
4481 /* If the failed drives are just a ReadError, then we might need
4482 * to progress the repair/check process
4484 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4485 for (i = 0; i < s.failed; i++) {
4486 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4487 if (test_bit(R5_ReadError, &dev->flags)
4488 && !test_bit(R5_LOCKED, &dev->flags)
4489 && test_bit(R5_UPTODATE, &dev->flags)
4491 if (!test_bit(R5_ReWrite, &dev->flags)) {
4492 set_bit(R5_Wantwrite, &dev->flags);
4493 set_bit(R5_ReWrite, &dev->flags);
4494 set_bit(R5_LOCKED, &dev->flags);
4497 /* let's read it back */
4498 set_bit(R5_Wantread, &dev->flags);
4499 set_bit(R5_LOCKED, &dev->flags);
4505 /* Finish reconstruct operations initiated by the expansion process */
4506 if (sh->reconstruct_state == reconstruct_state_result) {
4507 struct stripe_head *sh_src
4508 = get_active_stripe(conf, sh->sector, 1, 1, 1);
4509 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4510 /* sh cannot be written until sh_src has been read.
4511 * so arrange for sh to be delayed a little
4513 set_bit(STRIPE_DELAYED, &sh->state);
4514 set_bit(STRIPE_HANDLE, &sh->state);
4515 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4517 atomic_inc(&conf->preread_active_stripes);
4518 release_stripe(sh_src);
4522 release_stripe(sh_src);
4524 sh->reconstruct_state = reconstruct_state_idle;
4525 clear_bit(STRIPE_EXPANDING, &sh->state);
4526 for (i = conf->raid_disks; i--; ) {
4527 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4528 set_bit(R5_LOCKED, &sh->dev[i].flags);
4533 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4534 !sh->reconstruct_state) {
4535 /* Need to write out all blocks after computing parity */
4536 sh->disks = conf->raid_disks;
4537 stripe_set_idx(sh->sector, conf, 0, sh);
4538 schedule_reconstruction(sh, &s, 1, 1);
4539 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4540 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4541 atomic_dec(&conf->reshape_stripes);
4542 wake_up(&conf->wait_for_overlap);
4543 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4546 if (s.expanding && s.locked == 0 &&
4547 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4548 handle_stripe_expansion(conf, sh);
4551 /* wait for this device to become unblocked */
4552 if (unlikely(s.blocked_rdev)) {
4553 if (conf->mddev->external)
4554 md_wait_for_blocked_rdev(s.blocked_rdev,
4557 /* Internal metadata will immediately
4558 * be written by raid5d, so we don't
4559 * need to wait here.
4561 rdev_dec_pending(s.blocked_rdev,
4565 if (s.handle_bad_blocks)
4566 for (i = disks; i--; ) {
4567 struct md_rdev *rdev;
4568 struct r5dev *dev = &sh->dev[i];
4569 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4570 /* We own a safe reference to the rdev */
4571 rdev = conf->disks[i].rdev;
4572 if (!rdev_set_badblocks(rdev, sh->sector,
4574 md_error(conf->mddev, rdev);
4575 rdev_dec_pending(rdev, conf->mddev);
4577 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4578 rdev = conf->disks[i].rdev;
4579 rdev_clear_badblocks(rdev, sh->sector,
4581 rdev_dec_pending(rdev, conf->mddev);
4583 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4584 rdev = conf->disks[i].replacement;
4586 /* rdev have been moved down */
4587 rdev = conf->disks[i].rdev;
4588 rdev_clear_badblocks(rdev, sh->sector,
4590 rdev_dec_pending(rdev, conf->mddev);
4595 raid_run_ops(sh, s.ops_request);
4599 if (s.dec_preread_active) {
4600 /* We delay this until after ops_run_io so that if make_request
4601 * is waiting on a flush, it won't continue until the writes
4602 * have actually been submitted.
4604 atomic_dec(&conf->preread_active_stripes);
4605 if (atomic_read(&conf->preread_active_stripes) <
4607 md_wakeup_thread(conf->mddev->thread);
4610 if (!bio_list_empty(&s.return_bi)) {
4611 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags)) {
4612 spin_lock_irq(&conf->device_lock);
4613 bio_list_merge(&conf->return_bi, &s.return_bi);
4614 spin_unlock_irq(&conf->device_lock);
4615 md_wakeup_thread(conf->mddev->thread);
4617 return_io(&s.return_bi);
4620 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4623 static void raid5_activate_delayed(struct r5conf *conf)
4625 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4626 while (!list_empty(&conf->delayed_list)) {
4627 struct list_head *l = conf->delayed_list.next;
4628 struct stripe_head *sh;
4629 sh = list_entry(l, struct stripe_head, lru);
4631 clear_bit(STRIPE_DELAYED, &sh->state);
4632 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4633 atomic_inc(&conf->preread_active_stripes);
4634 list_add_tail(&sh->lru, &conf->hold_list);
4635 raid5_wakeup_stripe_thread(sh);
4640 static void activate_bit_delay(struct r5conf *conf,
4641 struct list_head *temp_inactive_list)
4643 /* device_lock is held */
4644 struct list_head head;
4645 list_add(&head, &conf->bitmap_list);
4646 list_del_init(&conf->bitmap_list);
4647 while (!list_empty(&head)) {
4648 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4650 list_del_init(&sh->lru);
4651 atomic_inc(&sh->count);
4652 hash = sh->hash_lock_index;
4653 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4657 static int raid5_congested(struct mddev *mddev, int bits)
4659 struct r5conf *conf = mddev->private;
4661 /* No difference between reads and writes. Just check
4662 * how busy the stripe_cache is
4665 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4669 if (atomic_read(&conf->empty_inactive_list_nr))
4675 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4677 struct r5conf *conf = mddev->private;
4678 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4679 unsigned int chunk_sectors;
4680 unsigned int bio_sectors = bio_sectors(bio);
4682 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4683 return chunk_sectors >=
4684 ((sector & (chunk_sectors - 1)) + bio_sectors);
4688 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4689 * later sampled by raid5d.
4691 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4693 unsigned long flags;
4695 spin_lock_irqsave(&conf->device_lock, flags);
4697 bi->bi_next = conf->retry_read_aligned_list;
4698 conf->retry_read_aligned_list = bi;
4700 spin_unlock_irqrestore(&conf->device_lock, flags);
4701 md_wakeup_thread(conf->mddev->thread);
4704 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4708 bi = conf->retry_read_aligned;
4710 conf->retry_read_aligned = NULL;
4713 bi = conf->retry_read_aligned_list;
4715 conf->retry_read_aligned_list = bi->bi_next;
4718 * this sets the active strip count to 1 and the processed
4719 * strip count to zero (upper 8 bits)
4721 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4728 * The "raid5_align_endio" should check if the read succeeded and if it
4729 * did, call bio_endio on the original bio (having bio_put the new bio
4731 * If the read failed..
4733 static void raid5_align_endio(struct bio *bi)
4735 struct bio* raid_bi = bi->bi_private;
4736 struct mddev *mddev;
4737 struct r5conf *conf;
4738 struct md_rdev *rdev;
4739 int error = bi->bi_error;
4743 rdev = (void*)raid_bi->bi_next;
4744 raid_bi->bi_next = NULL;
4745 mddev = rdev->mddev;
4746 conf = mddev->private;
4748 rdev_dec_pending(rdev, conf->mddev);
4751 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4754 if (atomic_dec_and_test(&conf->active_aligned_reads))
4755 wake_up(&conf->wait_for_quiescent);
4759 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4761 add_bio_to_retry(raid_bi, conf);
4764 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4766 struct r5conf *conf = mddev->private;
4768 struct bio* align_bi;
4769 struct md_rdev *rdev;
4770 sector_t end_sector;
4772 if (!in_chunk_boundary(mddev, raid_bio)) {
4773 pr_debug("%s: non aligned\n", __func__);
4777 * use bio_clone_mddev to make a copy of the bio
4779 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4783 * set bi_end_io to a new function, and set bi_private to the
4786 align_bi->bi_end_io = raid5_align_endio;
4787 align_bi->bi_private = raid_bio;
4791 align_bi->bi_iter.bi_sector =
4792 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4795 end_sector = bio_end_sector(align_bi);
4797 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4798 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4799 rdev->recovery_offset < end_sector) {
4800 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4802 (test_bit(Faulty, &rdev->flags) ||
4803 !(test_bit(In_sync, &rdev->flags) ||
4804 rdev->recovery_offset >= end_sector)))
4811 atomic_inc(&rdev->nr_pending);
4813 raid_bio->bi_next = (void*)rdev;
4814 align_bi->bi_bdev = rdev->bdev;
4815 bio_clear_flag(align_bi, BIO_SEG_VALID);
4817 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4818 bio_sectors(align_bi),
4819 &first_bad, &bad_sectors)) {
4821 rdev_dec_pending(rdev, mddev);
4825 /* No reshape active, so we can trust rdev->data_offset */
4826 align_bi->bi_iter.bi_sector += rdev->data_offset;
4828 spin_lock_irq(&conf->device_lock);
4829 wait_event_lock_irq(conf->wait_for_quiescent,
4832 atomic_inc(&conf->active_aligned_reads);
4833 spin_unlock_irq(&conf->device_lock);
4836 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4837 align_bi, disk_devt(mddev->gendisk),
4838 raid_bio->bi_iter.bi_sector);
4839 generic_make_request(align_bi);
4848 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4853 sector_t sector = raid_bio->bi_iter.bi_sector;
4854 unsigned chunk_sects = mddev->chunk_sectors;
4855 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4857 if (sectors < bio_sectors(raid_bio)) {
4858 split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4859 bio_chain(split, raid_bio);
4863 if (!raid5_read_one_chunk(mddev, split)) {
4864 if (split != raid_bio)
4865 generic_make_request(raid_bio);
4868 } while (split != raid_bio);
4873 /* __get_priority_stripe - get the next stripe to process
4875 * Full stripe writes are allowed to pass preread active stripes up until
4876 * the bypass_threshold is exceeded. In general the bypass_count
4877 * increments when the handle_list is handled before the hold_list; however, it
4878 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4879 * stripe with in flight i/o. The bypass_count will be reset when the
4880 * head of the hold_list has changed, i.e. the head was promoted to the
4883 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4885 struct stripe_head *sh = NULL, *tmp;
4886 struct list_head *handle_list = NULL;
4887 struct r5worker_group *wg = NULL;
4889 if (conf->worker_cnt_per_group == 0) {
4890 handle_list = &conf->handle_list;
4891 } else if (group != ANY_GROUP) {
4892 handle_list = &conf->worker_groups[group].handle_list;
4893 wg = &conf->worker_groups[group];
4896 for (i = 0; i < conf->group_cnt; i++) {
4897 handle_list = &conf->worker_groups[i].handle_list;
4898 wg = &conf->worker_groups[i];
4899 if (!list_empty(handle_list))
4904 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4906 list_empty(handle_list) ? "empty" : "busy",
4907 list_empty(&conf->hold_list) ? "empty" : "busy",
4908 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4910 if (!list_empty(handle_list)) {
4911 sh = list_entry(handle_list->next, typeof(*sh), lru);
4913 if (list_empty(&conf->hold_list))
4914 conf->bypass_count = 0;
4915 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4916 if (conf->hold_list.next == conf->last_hold)
4917 conf->bypass_count++;
4919 conf->last_hold = conf->hold_list.next;
4920 conf->bypass_count -= conf->bypass_threshold;
4921 if (conf->bypass_count < 0)
4922 conf->bypass_count = 0;
4925 } else if (!list_empty(&conf->hold_list) &&
4926 ((conf->bypass_threshold &&
4927 conf->bypass_count > conf->bypass_threshold) ||
4928 atomic_read(&conf->pending_full_writes) == 0)) {
4930 list_for_each_entry(tmp, &conf->hold_list, lru) {
4931 if (conf->worker_cnt_per_group == 0 ||
4932 group == ANY_GROUP ||
4933 !cpu_online(tmp->cpu) ||
4934 cpu_to_group(tmp->cpu) == group) {
4941 conf->bypass_count -= conf->bypass_threshold;
4942 if (conf->bypass_count < 0)
4943 conf->bypass_count = 0;
4955 list_del_init(&sh->lru);
4956 BUG_ON(atomic_inc_return(&sh->count) != 1);
4960 struct raid5_plug_cb {
4961 struct blk_plug_cb cb;
4962 struct list_head list;
4963 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4966 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4968 struct raid5_plug_cb *cb = container_of(
4969 blk_cb, struct raid5_plug_cb, cb);
4970 struct stripe_head *sh;
4971 struct mddev *mddev = cb->cb.data;
4972 struct r5conf *conf = mddev->private;
4976 if (cb->list.next && !list_empty(&cb->list)) {
4977 spin_lock_irq(&conf->device_lock);
4978 while (!list_empty(&cb->list)) {
4979 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4980 list_del_init(&sh->lru);
4982 * avoid race release_stripe_plug() sees
4983 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4984 * is still in our list
4986 smp_mb__before_atomic();
4987 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4989 * STRIPE_ON_RELEASE_LIST could be set here. In that
4990 * case, the count is always > 1 here
4992 hash = sh->hash_lock_index;
4993 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
4996 spin_unlock_irq(&conf->device_lock);
4998 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4999 NR_STRIPE_HASH_LOCKS);
5001 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5005 static void release_stripe_plug(struct mddev *mddev,
5006 struct stripe_head *sh)
5008 struct blk_plug_cb *blk_cb = blk_check_plugged(
5009 raid5_unplug, mddev,
5010 sizeof(struct raid5_plug_cb));
5011 struct raid5_plug_cb *cb;
5018 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5020 if (cb->list.next == NULL) {
5022 INIT_LIST_HEAD(&cb->list);
5023 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5024 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5027 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5028 list_add_tail(&sh->lru, &cb->list);
5033 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5035 struct r5conf *conf = mddev->private;
5036 sector_t logical_sector, last_sector;
5037 struct stripe_head *sh;
5041 if (mddev->reshape_position != MaxSector)
5042 /* Skip discard while reshape is happening */
5045 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5046 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5049 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5051 stripe_sectors = conf->chunk_sectors *
5052 (conf->raid_disks - conf->max_degraded);
5053 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5055 sector_div(last_sector, stripe_sectors);
5057 logical_sector *= conf->chunk_sectors;
5058 last_sector *= conf->chunk_sectors;
5060 for (; logical_sector < last_sector;
5061 logical_sector += STRIPE_SECTORS) {
5065 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
5066 prepare_to_wait(&conf->wait_for_overlap, &w,
5067 TASK_UNINTERRUPTIBLE);
5068 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5069 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5074 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5075 spin_lock_irq(&sh->stripe_lock);
5076 for (d = 0; d < conf->raid_disks; d++) {
5077 if (d == sh->pd_idx || d == sh->qd_idx)
5079 if (sh->dev[d].towrite || sh->dev[d].toread) {
5080 set_bit(R5_Overlap, &sh->dev[d].flags);
5081 spin_unlock_irq(&sh->stripe_lock);
5087 set_bit(STRIPE_DISCARD, &sh->state);
5088 finish_wait(&conf->wait_for_overlap, &w);
5089 sh->overwrite_disks = 0;
5090 for (d = 0; d < conf->raid_disks; d++) {
5091 if (d == sh->pd_idx || d == sh->qd_idx)
5093 sh->dev[d].towrite = bi;
5094 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5095 raid5_inc_bi_active_stripes(bi);
5096 sh->overwrite_disks++;
5098 spin_unlock_irq(&sh->stripe_lock);
5099 if (conf->mddev->bitmap) {
5101 d < conf->raid_disks - conf->max_degraded;
5103 bitmap_startwrite(mddev->bitmap,
5107 sh->bm_seq = conf->seq_flush + 1;
5108 set_bit(STRIPE_BIT_DELAY, &sh->state);
5111 set_bit(STRIPE_HANDLE, &sh->state);
5112 clear_bit(STRIPE_DELAYED, &sh->state);
5113 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5114 atomic_inc(&conf->preread_active_stripes);
5115 release_stripe_plug(mddev, sh);
5118 remaining = raid5_dec_bi_active_stripes(bi);
5119 if (remaining == 0) {
5120 md_write_end(mddev);
5125 static void make_request(struct mddev *mddev, struct bio * bi)
5127 struct r5conf *conf = mddev->private;
5129 sector_t new_sector;
5130 sector_t logical_sector, last_sector;
5131 struct stripe_head *sh;
5132 const int rw = bio_data_dir(bi);
5137 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5138 md_flush_request(mddev, bi);
5142 md_write_start(mddev, bi);
5145 * If array is degraded, better not do chunk aligned read because
5146 * later we might have to read it again in order to reconstruct
5147 * data on failed drives.
5149 if (rw == READ && mddev->degraded == 0 &&
5150 mddev->reshape_position == MaxSector) {
5151 bi = chunk_aligned_read(mddev, bi);
5156 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5157 make_discard_request(mddev, bi);
5161 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5162 last_sector = bio_end_sector(bi);
5164 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5166 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5167 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5173 seq = read_seqcount_begin(&conf->gen_lock);
5176 prepare_to_wait(&conf->wait_for_overlap, &w,
5177 TASK_UNINTERRUPTIBLE);
5178 if (unlikely(conf->reshape_progress != MaxSector)) {
5179 /* spinlock is needed as reshape_progress may be
5180 * 64bit on a 32bit platform, and so it might be
5181 * possible to see a half-updated value
5182 * Of course reshape_progress could change after
5183 * the lock is dropped, so once we get a reference
5184 * to the stripe that we think it is, we will have
5187 spin_lock_irq(&conf->device_lock);
5188 if (mddev->reshape_backwards
5189 ? logical_sector < conf->reshape_progress
5190 : logical_sector >= conf->reshape_progress) {
5193 if (mddev->reshape_backwards
5194 ? logical_sector < conf->reshape_safe
5195 : logical_sector >= conf->reshape_safe) {
5196 spin_unlock_irq(&conf->device_lock);
5202 spin_unlock_irq(&conf->device_lock);
5205 new_sector = raid5_compute_sector(conf, logical_sector,
5208 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5209 (unsigned long long)new_sector,
5210 (unsigned long long)logical_sector);
5212 sh = get_active_stripe(conf, new_sector, previous,
5213 (bi->bi_rw&RWA_MASK), 0);
5215 if (unlikely(previous)) {
5216 /* expansion might have moved on while waiting for a
5217 * stripe, so we must do the range check again.
5218 * Expansion could still move past after this
5219 * test, but as we are holding a reference to
5220 * 'sh', we know that if that happens,
5221 * STRIPE_EXPANDING will get set and the expansion
5222 * won't proceed until we finish with the stripe.
5225 spin_lock_irq(&conf->device_lock);
5226 if (mddev->reshape_backwards
5227 ? logical_sector >= conf->reshape_progress
5228 : logical_sector < conf->reshape_progress)
5229 /* mismatch, need to try again */
5231 spin_unlock_irq(&conf->device_lock);
5239 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5240 /* Might have got the wrong stripe_head
5248 logical_sector >= mddev->suspend_lo &&
5249 logical_sector < mddev->suspend_hi) {
5251 /* As the suspend_* range is controlled by
5252 * userspace, we want an interruptible
5255 flush_signals(current);
5256 prepare_to_wait(&conf->wait_for_overlap,
5257 &w, TASK_INTERRUPTIBLE);
5258 if (logical_sector >= mddev->suspend_lo &&
5259 logical_sector < mddev->suspend_hi) {
5266 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5267 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5268 /* Stripe is busy expanding or
5269 * add failed due to overlap. Flush everything
5272 md_wakeup_thread(mddev->thread);
5278 set_bit(STRIPE_HANDLE, &sh->state);
5279 clear_bit(STRIPE_DELAYED, &sh->state);
5280 if ((!sh->batch_head || sh == sh->batch_head) &&
5281 (bi->bi_rw & REQ_SYNC) &&
5282 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5283 atomic_inc(&conf->preread_active_stripes);
5284 release_stripe_plug(mddev, sh);
5286 /* cannot get stripe for read-ahead, just give-up */
5287 bi->bi_error = -EIO;
5291 finish_wait(&conf->wait_for_overlap, &w);
5293 remaining = raid5_dec_bi_active_stripes(bi);
5294 if (remaining == 0) {
5297 md_write_end(mddev);
5299 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5305 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5307 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5309 /* reshaping is quite different to recovery/resync so it is
5310 * handled quite separately ... here.
5312 * On each call to sync_request, we gather one chunk worth of
5313 * destination stripes and flag them as expanding.
5314 * Then we find all the source stripes and request reads.
5315 * As the reads complete, handle_stripe will copy the data
5316 * into the destination stripe and release that stripe.
5318 struct r5conf *conf = mddev->private;
5319 struct stripe_head *sh;
5320 sector_t first_sector, last_sector;
5321 int raid_disks = conf->previous_raid_disks;
5322 int data_disks = raid_disks - conf->max_degraded;
5323 int new_data_disks = conf->raid_disks - conf->max_degraded;
5326 sector_t writepos, readpos, safepos;
5327 sector_t stripe_addr;
5328 int reshape_sectors;
5329 struct list_head stripes;
5332 if (sector_nr == 0) {
5333 /* If restarting in the middle, skip the initial sectors */
5334 if (mddev->reshape_backwards &&
5335 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5336 sector_nr = raid5_size(mddev, 0, 0)
5337 - conf->reshape_progress;
5338 } else if (mddev->reshape_backwards &&
5339 conf->reshape_progress == MaxSector) {
5340 /* shouldn't happen, but just in case, finish up.*/
5341 sector_nr = MaxSector;
5342 } else if (!mddev->reshape_backwards &&
5343 conf->reshape_progress > 0)
5344 sector_nr = conf->reshape_progress;
5345 sector_div(sector_nr, new_data_disks);
5347 mddev->curr_resync_completed = sector_nr;
5348 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5355 /* We need to process a full chunk at a time.
5356 * If old and new chunk sizes differ, we need to process the
5360 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5362 /* We update the metadata at least every 10 seconds, or when
5363 * the data about to be copied would over-write the source of
5364 * the data at the front of the range. i.e. one new_stripe
5365 * along from reshape_progress new_maps to after where
5366 * reshape_safe old_maps to
5368 writepos = conf->reshape_progress;
5369 sector_div(writepos, new_data_disks);
5370 readpos = conf->reshape_progress;
5371 sector_div(readpos, data_disks);
5372 safepos = conf->reshape_safe;
5373 sector_div(safepos, data_disks);
5374 if (mddev->reshape_backwards) {
5375 BUG_ON(writepos < reshape_sectors);
5376 writepos -= reshape_sectors;
5377 readpos += reshape_sectors;
5378 safepos += reshape_sectors;
5380 writepos += reshape_sectors;
5381 /* readpos and safepos are worst-case calculations.
5382 * A negative number is overly pessimistic, and causes
5383 * obvious problems for unsigned storage. So clip to 0.
5385 readpos -= min_t(sector_t, reshape_sectors, readpos);
5386 safepos -= min_t(sector_t, reshape_sectors, safepos);
5389 /* Having calculated the 'writepos' possibly use it
5390 * to set 'stripe_addr' which is where we will write to.
5392 if (mddev->reshape_backwards) {
5393 BUG_ON(conf->reshape_progress == 0);
5394 stripe_addr = writepos;
5395 BUG_ON((mddev->dev_sectors &
5396 ~((sector_t)reshape_sectors - 1))
5397 - reshape_sectors - stripe_addr
5400 BUG_ON(writepos != sector_nr + reshape_sectors);
5401 stripe_addr = sector_nr;
5404 /* 'writepos' is the most advanced device address we might write.
5405 * 'readpos' is the least advanced device address we might read.
5406 * 'safepos' is the least address recorded in the metadata as having
5408 * If there is a min_offset_diff, these are adjusted either by
5409 * increasing the safepos/readpos if diff is negative, or
5410 * increasing writepos if diff is positive.
5411 * If 'readpos' is then behind 'writepos', there is no way that we can
5412 * ensure safety in the face of a crash - that must be done by userspace
5413 * making a backup of the data. So in that case there is no particular
5414 * rush to update metadata.
5415 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5416 * update the metadata to advance 'safepos' to match 'readpos' so that
5417 * we can be safe in the event of a crash.
5418 * So we insist on updating metadata if safepos is behind writepos and
5419 * readpos is beyond writepos.
5420 * In any case, update the metadata every 10 seconds.
5421 * Maybe that number should be configurable, but I'm not sure it is
5422 * worth it.... maybe it could be a multiple of safemode_delay???
5424 if (conf->min_offset_diff < 0) {
5425 safepos += -conf->min_offset_diff;
5426 readpos += -conf->min_offset_diff;
5428 writepos += conf->min_offset_diff;
5430 if ((mddev->reshape_backwards
5431 ? (safepos > writepos && readpos < writepos)
5432 : (safepos < writepos && readpos > writepos)) ||
5433 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5434 /* Cannot proceed until we've updated the superblock... */
5435 wait_event(conf->wait_for_overlap,
5436 atomic_read(&conf->reshape_stripes)==0
5437 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5438 if (atomic_read(&conf->reshape_stripes) != 0)
5440 mddev->reshape_position = conf->reshape_progress;
5441 mddev->curr_resync_completed = sector_nr;
5442 conf->reshape_checkpoint = jiffies;
5443 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5444 md_wakeup_thread(mddev->thread);
5445 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5446 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5447 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5449 spin_lock_irq(&conf->device_lock);
5450 conf->reshape_safe = mddev->reshape_position;
5451 spin_unlock_irq(&conf->device_lock);
5452 wake_up(&conf->wait_for_overlap);
5453 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5456 INIT_LIST_HEAD(&stripes);
5457 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5459 int skipped_disk = 0;
5460 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5461 set_bit(STRIPE_EXPANDING, &sh->state);
5462 atomic_inc(&conf->reshape_stripes);
5463 /* If any of this stripe is beyond the end of the old
5464 * array, then we need to zero those blocks
5466 for (j=sh->disks; j--;) {
5468 if (j == sh->pd_idx)
5470 if (conf->level == 6 &&
5473 s = compute_blocknr(sh, j, 0);
5474 if (s < raid5_size(mddev, 0, 0)) {
5478 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5479 set_bit(R5_Expanded, &sh->dev[j].flags);
5480 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5482 if (!skipped_disk) {
5483 set_bit(STRIPE_EXPAND_READY, &sh->state);
5484 set_bit(STRIPE_HANDLE, &sh->state);
5486 list_add(&sh->lru, &stripes);
5488 spin_lock_irq(&conf->device_lock);
5489 if (mddev->reshape_backwards)
5490 conf->reshape_progress -= reshape_sectors * new_data_disks;
5492 conf->reshape_progress += reshape_sectors * new_data_disks;
5493 spin_unlock_irq(&conf->device_lock);
5494 /* Ok, those stripe are ready. We can start scheduling
5495 * reads on the source stripes.
5496 * The source stripes are determined by mapping the first and last
5497 * block on the destination stripes.
5500 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5503 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5504 * new_data_disks - 1),
5506 if (last_sector >= mddev->dev_sectors)
5507 last_sector = mddev->dev_sectors - 1;
5508 while (first_sector <= last_sector) {
5509 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
5510 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5511 set_bit(STRIPE_HANDLE, &sh->state);
5513 first_sector += STRIPE_SECTORS;
5515 /* Now that the sources are clearly marked, we can release
5516 * the destination stripes
5518 while (!list_empty(&stripes)) {
5519 sh = list_entry(stripes.next, struct stripe_head, lru);
5520 list_del_init(&sh->lru);
5523 /* If this takes us to the resync_max point where we have to pause,
5524 * then we need to write out the superblock.
5526 sector_nr += reshape_sectors;
5527 retn = reshape_sectors;
5529 if (mddev->curr_resync_completed > mddev->resync_max ||
5530 (sector_nr - mddev->curr_resync_completed) * 2
5531 >= mddev->resync_max - mddev->curr_resync_completed) {
5532 /* Cannot proceed until we've updated the superblock... */
5533 wait_event(conf->wait_for_overlap,
5534 atomic_read(&conf->reshape_stripes) == 0
5535 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5536 if (atomic_read(&conf->reshape_stripes) != 0)
5538 mddev->reshape_position = conf->reshape_progress;
5539 mddev->curr_resync_completed = sector_nr;
5540 conf->reshape_checkpoint = jiffies;
5541 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5542 md_wakeup_thread(mddev->thread);
5543 wait_event(mddev->sb_wait,
5544 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5545 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5546 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5548 spin_lock_irq(&conf->device_lock);
5549 conf->reshape_safe = mddev->reshape_position;
5550 spin_unlock_irq(&conf->device_lock);
5551 wake_up(&conf->wait_for_overlap);
5552 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5558 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5560 struct r5conf *conf = mddev->private;
5561 struct stripe_head *sh;
5562 sector_t max_sector = mddev->dev_sectors;
5563 sector_t sync_blocks;
5564 int still_degraded = 0;
5567 if (sector_nr >= max_sector) {
5568 /* just being told to finish up .. nothing much to do */
5570 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5575 if (mddev->curr_resync < max_sector) /* aborted */
5576 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5578 else /* completed sync */
5580 bitmap_close_sync(mddev->bitmap);
5585 /* Allow raid5_quiesce to complete */
5586 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5588 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5589 return reshape_request(mddev, sector_nr, skipped);
5591 /* No need to check resync_max as we never do more than one
5592 * stripe, and as resync_max will always be on a chunk boundary,
5593 * if the check in md_do_sync didn't fire, there is no chance
5594 * of overstepping resync_max here
5597 /* if there is too many failed drives and we are trying
5598 * to resync, then assert that we are finished, because there is
5599 * nothing we can do.
5601 if (mddev->degraded >= conf->max_degraded &&
5602 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5603 sector_t rv = mddev->dev_sectors - sector_nr;
5607 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5609 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5610 sync_blocks >= STRIPE_SECTORS) {
5611 /* we can skip this block, and probably more */
5612 sync_blocks /= STRIPE_SECTORS;
5614 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5617 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5619 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5621 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5622 /* make sure we don't swamp the stripe cache if someone else
5623 * is trying to get access
5625 schedule_timeout_uninterruptible(1);
5627 /* Need to check if array will still be degraded after recovery/resync
5628 * Note in case of > 1 drive failures it's possible we're rebuilding
5629 * one drive while leaving another faulty drive in array.
5632 for (i = 0; i < conf->raid_disks; i++) {
5633 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5635 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5640 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5642 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5643 set_bit(STRIPE_HANDLE, &sh->state);
5647 return STRIPE_SECTORS;
5650 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5652 /* We may not be able to submit a whole bio at once as there
5653 * may not be enough stripe_heads available.
5654 * We cannot pre-allocate enough stripe_heads as we may need
5655 * more than exist in the cache (if we allow ever large chunks).
5656 * So we do one stripe head at a time and record in
5657 * ->bi_hw_segments how many have been done.
5659 * We *know* that this entire raid_bio is in one chunk, so
5660 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5662 struct stripe_head *sh;
5664 sector_t sector, logical_sector, last_sector;
5669 logical_sector = raid_bio->bi_iter.bi_sector &
5670 ~((sector_t)STRIPE_SECTORS-1);
5671 sector = raid5_compute_sector(conf, logical_sector,
5673 last_sector = bio_end_sector(raid_bio);
5675 for (; logical_sector < last_sector;
5676 logical_sector += STRIPE_SECTORS,
5677 sector += STRIPE_SECTORS,
5680 if (scnt < raid5_bi_processed_stripes(raid_bio))
5681 /* already done this stripe */
5684 sh = get_active_stripe(conf, sector, 0, 1, 1);
5687 /* failed to get a stripe - must wait */
5688 raid5_set_bi_processed_stripes(raid_bio, scnt);
5689 conf->retry_read_aligned = raid_bio;
5693 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5695 raid5_set_bi_processed_stripes(raid_bio, scnt);
5696 conf->retry_read_aligned = raid_bio;
5700 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5705 remaining = raid5_dec_bi_active_stripes(raid_bio);
5706 if (remaining == 0) {
5707 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5709 bio_endio(raid_bio);
5711 if (atomic_dec_and_test(&conf->active_aligned_reads))
5712 wake_up(&conf->wait_for_quiescent);
5716 static int handle_active_stripes(struct r5conf *conf, int group,
5717 struct r5worker *worker,
5718 struct list_head *temp_inactive_list)
5720 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5721 int i, batch_size = 0, hash;
5722 bool release_inactive = false;
5724 while (batch_size < MAX_STRIPE_BATCH &&
5725 (sh = __get_priority_stripe(conf, group)) != NULL)
5726 batch[batch_size++] = sh;
5728 if (batch_size == 0) {
5729 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5730 if (!list_empty(temp_inactive_list + i))
5732 if (i == NR_STRIPE_HASH_LOCKS)
5734 release_inactive = true;
5736 spin_unlock_irq(&conf->device_lock);
5738 release_inactive_stripe_list(conf, temp_inactive_list,
5739 NR_STRIPE_HASH_LOCKS);
5741 if (release_inactive) {
5742 spin_lock_irq(&conf->device_lock);
5746 for (i = 0; i < batch_size; i++)
5747 handle_stripe(batch[i]);
5751 spin_lock_irq(&conf->device_lock);
5752 for (i = 0; i < batch_size; i++) {
5753 hash = batch[i]->hash_lock_index;
5754 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5759 static void raid5_do_work(struct work_struct *work)
5761 struct r5worker *worker = container_of(work, struct r5worker, work);
5762 struct r5worker_group *group = worker->group;
5763 struct r5conf *conf = group->conf;
5764 int group_id = group - conf->worker_groups;
5766 struct blk_plug plug;
5768 pr_debug("+++ raid5worker active\n");
5770 blk_start_plug(&plug);
5772 spin_lock_irq(&conf->device_lock);
5774 int batch_size, released;
5776 released = release_stripe_list(conf, worker->temp_inactive_list);
5778 batch_size = handle_active_stripes(conf, group_id, worker,
5779 worker->temp_inactive_list);
5780 worker->working = false;
5781 if (!batch_size && !released)
5783 handled += batch_size;
5785 pr_debug("%d stripes handled\n", handled);
5787 spin_unlock_irq(&conf->device_lock);
5788 blk_finish_plug(&plug);
5790 pr_debug("--- raid5worker inactive\n");
5794 * This is our raid5 kernel thread.
5796 * We scan the hash table for stripes which can be handled now.
5797 * During the scan, completed stripes are saved for us by the interrupt
5798 * handler, so that they will not have to wait for our next wakeup.
5800 static void raid5d(struct md_thread *thread)
5802 struct mddev *mddev = thread->mddev;
5803 struct r5conf *conf = mddev->private;
5805 struct blk_plug plug;
5807 pr_debug("+++ raid5d active\n");
5809 md_check_recovery(mddev);
5811 if (!bio_list_empty(&conf->return_bi) &&
5812 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5813 struct bio_list tmp = BIO_EMPTY_LIST;
5814 spin_lock_irq(&conf->device_lock);
5815 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5816 bio_list_merge(&tmp, &conf->return_bi);
5817 bio_list_init(&conf->return_bi);
5819 spin_unlock_irq(&conf->device_lock);
5823 blk_start_plug(&plug);
5825 spin_lock_irq(&conf->device_lock);
5828 int batch_size, released;
5830 released = release_stripe_list(conf, conf->temp_inactive_list);
5832 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5835 !list_empty(&conf->bitmap_list)) {
5836 /* Now is a good time to flush some bitmap updates */
5838 spin_unlock_irq(&conf->device_lock);
5839 bitmap_unplug(mddev->bitmap);
5840 spin_lock_irq(&conf->device_lock);
5841 conf->seq_write = conf->seq_flush;
5842 activate_bit_delay(conf, conf->temp_inactive_list);
5844 raid5_activate_delayed(conf);
5846 while ((bio = remove_bio_from_retry(conf))) {
5848 spin_unlock_irq(&conf->device_lock);
5849 ok = retry_aligned_read(conf, bio);
5850 spin_lock_irq(&conf->device_lock);
5856 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5857 conf->temp_inactive_list);
5858 if (!batch_size && !released)
5860 handled += batch_size;
5862 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5863 spin_unlock_irq(&conf->device_lock);
5864 md_check_recovery(mddev);
5865 spin_lock_irq(&conf->device_lock);
5868 pr_debug("%d stripes handled\n", handled);
5870 spin_unlock_irq(&conf->device_lock);
5871 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5872 mutex_trylock(&conf->cache_size_mutex)) {
5873 grow_one_stripe(conf, __GFP_NOWARN);
5874 /* Set flag even if allocation failed. This helps
5875 * slow down allocation requests when mem is short
5877 set_bit(R5_DID_ALLOC, &conf->cache_state);
5878 mutex_unlock(&conf->cache_size_mutex);
5881 async_tx_issue_pending_all();
5882 blk_finish_plug(&plug);
5884 pr_debug("--- raid5d inactive\n");
5888 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5890 struct r5conf *conf;
5892 spin_lock(&mddev->lock);
5893 conf = mddev->private;
5895 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5896 spin_unlock(&mddev->lock);
5901 raid5_set_cache_size(struct mddev *mddev, int size)
5903 struct r5conf *conf = mddev->private;
5906 if (size <= 16 || size > 32768)
5909 conf->min_nr_stripes = size;
5910 mutex_lock(&conf->cache_size_mutex);
5911 while (size < conf->max_nr_stripes &&
5912 drop_one_stripe(conf))
5914 mutex_unlock(&conf->cache_size_mutex);
5917 err = md_allow_write(mddev);
5921 mutex_lock(&conf->cache_size_mutex);
5922 while (size > conf->max_nr_stripes)
5923 if (!grow_one_stripe(conf, GFP_KERNEL))
5925 mutex_unlock(&conf->cache_size_mutex);
5929 EXPORT_SYMBOL(raid5_set_cache_size);
5932 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5934 struct r5conf *conf;
5938 if (len >= PAGE_SIZE)
5940 if (kstrtoul(page, 10, &new))
5942 err = mddev_lock(mddev);
5945 conf = mddev->private;
5949 err = raid5_set_cache_size(mddev, new);
5950 mddev_unlock(mddev);
5955 static struct md_sysfs_entry
5956 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5957 raid5_show_stripe_cache_size,
5958 raid5_store_stripe_cache_size);
5961 raid5_show_rmw_level(struct mddev *mddev, char *page)
5963 struct r5conf *conf = mddev->private;
5965 return sprintf(page, "%d\n", conf->rmw_level);
5971 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
5973 struct r5conf *conf = mddev->private;
5979 if (len >= PAGE_SIZE)
5982 if (kstrtoul(page, 10, &new))
5985 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
5988 if (new != PARITY_DISABLE_RMW &&
5989 new != PARITY_ENABLE_RMW &&
5990 new != PARITY_PREFER_RMW)
5993 conf->rmw_level = new;
5997 static struct md_sysfs_entry
5998 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
5999 raid5_show_rmw_level,
6000 raid5_store_rmw_level);
6004 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6006 struct r5conf *conf;
6008 spin_lock(&mddev->lock);
6009 conf = mddev->private;
6011 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6012 spin_unlock(&mddev->lock);
6017 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6019 struct r5conf *conf;
6023 if (len >= PAGE_SIZE)
6025 if (kstrtoul(page, 10, &new))
6028 err = mddev_lock(mddev);
6031 conf = mddev->private;
6034 else if (new > conf->min_nr_stripes)
6037 conf->bypass_threshold = new;
6038 mddev_unlock(mddev);
6042 static struct md_sysfs_entry
6043 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6045 raid5_show_preread_threshold,
6046 raid5_store_preread_threshold);
6049 raid5_show_skip_copy(struct mddev *mddev, char *page)
6051 struct r5conf *conf;
6053 spin_lock(&mddev->lock);
6054 conf = mddev->private;
6056 ret = sprintf(page, "%d\n", conf->skip_copy);
6057 spin_unlock(&mddev->lock);
6062 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6064 struct r5conf *conf;
6068 if (len >= PAGE_SIZE)
6070 if (kstrtoul(page, 10, &new))
6074 err = mddev_lock(mddev);
6077 conf = mddev->private;
6080 else if (new != conf->skip_copy) {
6081 mddev_suspend(mddev);
6082 conf->skip_copy = new;
6084 mddev->queue->backing_dev_info.capabilities |=
6085 BDI_CAP_STABLE_WRITES;
6087 mddev->queue->backing_dev_info.capabilities &=
6088 ~BDI_CAP_STABLE_WRITES;
6089 mddev_resume(mddev);
6091 mddev_unlock(mddev);
6095 static struct md_sysfs_entry
6096 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6097 raid5_show_skip_copy,
6098 raid5_store_skip_copy);
6101 stripe_cache_active_show(struct mddev *mddev, char *page)
6103 struct r5conf *conf = mddev->private;
6105 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6110 static struct md_sysfs_entry
6111 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6114 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6116 struct r5conf *conf;
6118 spin_lock(&mddev->lock);
6119 conf = mddev->private;
6121 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6122 spin_unlock(&mddev->lock);
6126 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6128 int *worker_cnt_per_group,
6129 struct r5worker_group **worker_groups);
6131 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6133 struct r5conf *conf;
6136 struct r5worker_group *new_groups, *old_groups;
6137 int group_cnt, worker_cnt_per_group;
6139 if (len >= PAGE_SIZE)
6141 if (kstrtoul(page, 10, &new))
6144 err = mddev_lock(mddev);
6147 conf = mddev->private;
6150 else if (new != conf->worker_cnt_per_group) {
6151 mddev_suspend(mddev);
6153 old_groups = conf->worker_groups;
6155 flush_workqueue(raid5_wq);
6157 err = alloc_thread_groups(conf, new,
6158 &group_cnt, &worker_cnt_per_group,
6161 spin_lock_irq(&conf->device_lock);
6162 conf->group_cnt = group_cnt;
6163 conf->worker_cnt_per_group = worker_cnt_per_group;
6164 conf->worker_groups = new_groups;
6165 spin_unlock_irq(&conf->device_lock);
6168 kfree(old_groups[0].workers);
6171 mddev_resume(mddev);
6173 mddev_unlock(mddev);
6178 static struct md_sysfs_entry
6179 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6180 raid5_show_group_thread_cnt,
6181 raid5_store_group_thread_cnt);
6183 static struct attribute *raid5_attrs[] = {
6184 &raid5_stripecache_size.attr,
6185 &raid5_stripecache_active.attr,
6186 &raid5_preread_bypass_threshold.attr,
6187 &raid5_group_thread_cnt.attr,
6188 &raid5_skip_copy.attr,
6189 &raid5_rmw_level.attr,
6192 static struct attribute_group raid5_attrs_group = {
6194 .attrs = raid5_attrs,
6197 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6199 int *worker_cnt_per_group,
6200 struct r5worker_group **worker_groups)
6204 struct r5worker *workers;
6206 *worker_cnt_per_group = cnt;
6209 *worker_groups = NULL;
6212 *group_cnt = num_possible_nodes();
6213 size = sizeof(struct r5worker) * cnt;
6214 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6215 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6216 *group_cnt, GFP_NOIO);
6217 if (!*worker_groups || !workers) {
6219 kfree(*worker_groups);
6223 for (i = 0; i < *group_cnt; i++) {
6224 struct r5worker_group *group;
6226 group = &(*worker_groups)[i];
6227 INIT_LIST_HEAD(&group->handle_list);
6229 group->workers = workers + i * cnt;
6231 for (j = 0; j < cnt; j++) {
6232 struct r5worker *worker = group->workers + j;
6233 worker->group = group;
6234 INIT_WORK(&worker->work, raid5_do_work);
6236 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6237 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6244 static void free_thread_groups(struct r5conf *conf)
6246 if (conf->worker_groups)
6247 kfree(conf->worker_groups[0].workers);
6248 kfree(conf->worker_groups);
6249 conf->worker_groups = NULL;
6253 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6255 struct r5conf *conf = mddev->private;
6258 sectors = mddev->dev_sectors;
6260 /* size is defined by the smallest of previous and new size */
6261 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6263 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6264 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6265 return sectors * (raid_disks - conf->max_degraded);
6268 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6270 safe_put_page(percpu->spare_page);
6271 if (percpu->scribble)
6272 flex_array_free(percpu->scribble);
6273 percpu->spare_page = NULL;
6274 percpu->scribble = NULL;
6277 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6279 if (conf->level == 6 && !percpu->spare_page)
6280 percpu->spare_page = alloc_page(GFP_KERNEL);
6281 if (!percpu->scribble)
6282 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6283 conf->previous_raid_disks),
6284 max(conf->chunk_sectors,
6285 conf->prev_chunk_sectors)
6289 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6290 free_scratch_buffer(conf, percpu);
6297 static void raid5_free_percpu(struct r5conf *conf)
6304 #ifdef CONFIG_HOTPLUG_CPU
6305 unregister_cpu_notifier(&conf->cpu_notify);
6309 for_each_possible_cpu(cpu)
6310 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6313 free_percpu(conf->percpu);
6316 static void free_conf(struct r5conf *conf)
6318 if (conf->shrinker.seeks)
6319 unregister_shrinker(&conf->shrinker);
6320 free_thread_groups(conf);
6321 shrink_stripes(conf);
6322 raid5_free_percpu(conf);
6324 kfree(conf->stripe_hashtbl);
6328 #ifdef CONFIG_HOTPLUG_CPU
6329 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6332 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6333 long cpu = (long)hcpu;
6334 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6337 case CPU_UP_PREPARE:
6338 case CPU_UP_PREPARE_FROZEN:
6339 if (alloc_scratch_buffer(conf, percpu)) {
6340 pr_err("%s: failed memory allocation for cpu%ld\n",
6342 return notifier_from_errno(-ENOMEM);
6346 case CPU_DEAD_FROZEN:
6347 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6356 static int raid5_alloc_percpu(struct r5conf *conf)
6361 conf->percpu = alloc_percpu(struct raid5_percpu);
6365 #ifdef CONFIG_HOTPLUG_CPU
6366 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6367 conf->cpu_notify.priority = 0;
6368 err = register_cpu_notifier(&conf->cpu_notify);
6374 for_each_present_cpu(cpu) {
6375 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6377 pr_err("%s: failed memory allocation for cpu%ld\n",
6387 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6388 struct shrink_control *sc)
6390 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6391 unsigned long ret = SHRINK_STOP;
6393 if (mutex_trylock(&conf->cache_size_mutex)) {
6395 while (ret < sc->nr_to_scan &&
6396 conf->max_nr_stripes > conf->min_nr_stripes) {
6397 if (drop_one_stripe(conf) == 0) {
6403 mutex_unlock(&conf->cache_size_mutex);
6408 static unsigned long raid5_cache_count(struct shrinker *shrink,
6409 struct shrink_control *sc)
6411 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6413 if (conf->max_nr_stripes < conf->min_nr_stripes)
6414 /* unlikely, but not impossible */
6416 return conf->max_nr_stripes - conf->min_nr_stripes;
6419 static struct r5conf *setup_conf(struct mddev *mddev)
6421 struct r5conf *conf;
6422 int raid_disk, memory, max_disks;
6423 struct md_rdev *rdev;
6424 struct disk_info *disk;
6427 int group_cnt, worker_cnt_per_group;
6428 struct r5worker_group *new_group;
6430 if (mddev->new_level != 5
6431 && mddev->new_level != 4
6432 && mddev->new_level != 6) {
6433 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6434 mdname(mddev), mddev->new_level);
6435 return ERR_PTR(-EIO);
6437 if ((mddev->new_level == 5
6438 && !algorithm_valid_raid5(mddev->new_layout)) ||
6439 (mddev->new_level == 6
6440 && !algorithm_valid_raid6(mddev->new_layout))) {
6441 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6442 mdname(mddev), mddev->new_layout);
6443 return ERR_PTR(-EIO);
6445 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6446 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6447 mdname(mddev), mddev->raid_disks);
6448 return ERR_PTR(-EINVAL);
6451 if (!mddev->new_chunk_sectors ||
6452 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6453 !is_power_of_2(mddev->new_chunk_sectors)) {
6454 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6455 mdname(mddev), mddev->new_chunk_sectors << 9);
6456 return ERR_PTR(-EINVAL);
6459 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6462 /* Don't enable multi-threading by default*/
6463 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6465 conf->group_cnt = group_cnt;
6466 conf->worker_cnt_per_group = worker_cnt_per_group;
6467 conf->worker_groups = new_group;
6470 spin_lock_init(&conf->device_lock);
6471 seqcount_init(&conf->gen_lock);
6472 mutex_init(&conf->cache_size_mutex);
6473 init_waitqueue_head(&conf->wait_for_quiescent);
6474 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
6475 init_waitqueue_head(&conf->wait_for_stripe[i]);
6477 init_waitqueue_head(&conf->wait_for_overlap);
6478 INIT_LIST_HEAD(&conf->handle_list);
6479 INIT_LIST_HEAD(&conf->hold_list);
6480 INIT_LIST_HEAD(&conf->delayed_list);
6481 INIT_LIST_HEAD(&conf->bitmap_list);
6482 bio_list_init(&conf->return_bi);
6483 init_llist_head(&conf->released_stripes);
6484 atomic_set(&conf->active_stripes, 0);
6485 atomic_set(&conf->preread_active_stripes, 0);
6486 atomic_set(&conf->active_aligned_reads, 0);
6487 conf->bypass_threshold = BYPASS_THRESHOLD;
6488 conf->recovery_disabled = mddev->recovery_disabled - 1;
6490 conf->raid_disks = mddev->raid_disks;
6491 if (mddev->reshape_position == MaxSector)
6492 conf->previous_raid_disks = mddev->raid_disks;
6494 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6495 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6497 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6502 conf->mddev = mddev;
6504 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6507 /* We init hash_locks[0] separately to that it can be used
6508 * as the reference lock in the spin_lock_nest_lock() call
6509 * in lock_all_device_hash_locks_irq in order to convince
6510 * lockdep that we know what we are doing.
6512 spin_lock_init(conf->hash_locks);
6513 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6514 spin_lock_init(conf->hash_locks + i);
6516 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6517 INIT_LIST_HEAD(conf->inactive_list + i);
6519 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6520 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6522 conf->level = mddev->new_level;
6523 conf->chunk_sectors = mddev->new_chunk_sectors;
6524 if (raid5_alloc_percpu(conf) != 0)
6527 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6529 rdev_for_each(rdev, mddev) {
6530 raid_disk = rdev->raid_disk;
6531 if (raid_disk >= max_disks
6534 disk = conf->disks + raid_disk;
6536 if (test_bit(Replacement, &rdev->flags)) {
6537 if (disk->replacement)
6539 disk->replacement = rdev;
6546 if (test_bit(In_sync, &rdev->flags)) {
6547 char b[BDEVNAME_SIZE];
6548 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6550 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6551 } else if (rdev->saved_raid_disk != raid_disk)
6552 /* Cannot rely on bitmap to complete recovery */
6556 conf->level = mddev->new_level;
6557 if (conf->level == 6) {
6558 conf->max_degraded = 2;
6559 if (raid6_call.xor_syndrome)
6560 conf->rmw_level = PARITY_ENABLE_RMW;
6562 conf->rmw_level = PARITY_DISABLE_RMW;
6564 conf->max_degraded = 1;
6565 conf->rmw_level = PARITY_ENABLE_RMW;
6567 conf->algorithm = mddev->new_layout;
6568 conf->reshape_progress = mddev->reshape_position;
6569 if (conf->reshape_progress != MaxSector) {
6570 conf->prev_chunk_sectors = mddev->chunk_sectors;
6571 conf->prev_algo = mddev->layout;
6573 conf->prev_chunk_sectors = conf->chunk_sectors;
6574 conf->prev_algo = conf->algorithm;
6577 conf->min_nr_stripes = NR_STRIPES;
6578 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6579 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6580 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6581 if (grow_stripes(conf, conf->min_nr_stripes)) {
6583 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6584 mdname(mddev), memory);
6587 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6588 mdname(mddev), memory);
6590 * Losing a stripe head costs more than the time to refill it,
6591 * it reduces the queue depth and so can hurt throughput.
6592 * So set it rather large, scaled by number of devices.
6594 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6595 conf->shrinker.scan_objects = raid5_cache_scan;
6596 conf->shrinker.count_objects = raid5_cache_count;
6597 conf->shrinker.batch = 128;
6598 conf->shrinker.flags = 0;
6599 register_shrinker(&conf->shrinker);
6601 sprintf(pers_name, "raid%d", mddev->new_level);
6602 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6603 if (!conf->thread) {
6605 "md/raid:%s: couldn't allocate thread.\n",
6615 return ERR_PTR(-EIO);
6617 return ERR_PTR(-ENOMEM);
6620 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6623 case ALGORITHM_PARITY_0:
6624 if (raid_disk < max_degraded)
6627 case ALGORITHM_PARITY_N:
6628 if (raid_disk >= raid_disks - max_degraded)
6631 case ALGORITHM_PARITY_0_6:
6632 if (raid_disk == 0 ||
6633 raid_disk == raid_disks - 1)
6636 case ALGORITHM_LEFT_ASYMMETRIC_6:
6637 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6638 case ALGORITHM_LEFT_SYMMETRIC_6:
6639 case ALGORITHM_RIGHT_SYMMETRIC_6:
6640 if (raid_disk == raid_disks - 1)
6646 static int run(struct mddev *mddev)
6648 struct r5conf *conf;
6649 int working_disks = 0;
6650 int dirty_parity_disks = 0;
6651 struct md_rdev *rdev;
6652 sector_t reshape_offset = 0;
6654 long long min_offset_diff = 0;
6657 if (mddev->recovery_cp != MaxSector)
6658 printk(KERN_NOTICE "md/raid:%s: not clean"
6659 " -- starting background reconstruction\n",
6662 rdev_for_each(rdev, mddev) {
6664 if (rdev->raid_disk < 0)
6666 diff = (rdev->new_data_offset - rdev->data_offset);
6668 min_offset_diff = diff;
6670 } else if (mddev->reshape_backwards &&
6671 diff < min_offset_diff)
6672 min_offset_diff = diff;
6673 else if (!mddev->reshape_backwards &&
6674 diff > min_offset_diff)
6675 min_offset_diff = diff;
6678 if (mddev->reshape_position != MaxSector) {
6679 /* Check that we can continue the reshape.
6680 * Difficulties arise if the stripe we would write to
6681 * next is at or after the stripe we would read from next.
6682 * For a reshape that changes the number of devices, this
6683 * is only possible for a very short time, and mdadm makes
6684 * sure that time appears to have past before assembling
6685 * the array. So we fail if that time hasn't passed.
6686 * For a reshape that keeps the number of devices the same
6687 * mdadm must be monitoring the reshape can keeping the
6688 * critical areas read-only and backed up. It will start
6689 * the array in read-only mode, so we check for that.
6691 sector_t here_new, here_old;
6693 int max_degraded = (mddev->level == 6 ? 2 : 1);
6697 if (mddev->new_level != mddev->level) {
6698 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6699 "required - aborting.\n",
6703 old_disks = mddev->raid_disks - mddev->delta_disks;
6704 /* reshape_position must be on a new-stripe boundary, and one
6705 * further up in new geometry must map after here in old
6707 * If the chunk sizes are different, then as we perform reshape
6708 * in units of the largest of the two, reshape_position needs
6709 * be a multiple of the largest chunk size times new data disks.
6711 here_new = mddev->reshape_position;
6712 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6713 new_data_disks = mddev->raid_disks - max_degraded;
6714 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6715 printk(KERN_ERR "md/raid:%s: reshape_position not "
6716 "on a stripe boundary\n", mdname(mddev));
6719 reshape_offset = here_new * chunk_sectors;
6720 /* here_new is the stripe we will write to */
6721 here_old = mddev->reshape_position;
6722 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6723 /* here_old is the first stripe that we might need to read
6725 if (mddev->delta_disks == 0) {
6726 /* We cannot be sure it is safe to start an in-place
6727 * reshape. It is only safe if user-space is monitoring
6728 * and taking constant backups.
6729 * mdadm always starts a situation like this in
6730 * readonly mode so it can take control before
6731 * allowing any writes. So just check for that.
6733 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6734 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6735 /* not really in-place - so OK */;
6736 else if (mddev->ro == 0) {
6737 printk(KERN_ERR "md/raid:%s: in-place reshape "
6738 "must be started in read-only mode "
6743 } else if (mddev->reshape_backwards
6744 ? (here_new * chunk_sectors + min_offset_diff <=
6745 here_old * chunk_sectors)
6746 : (here_new * chunk_sectors >=
6747 here_old * chunk_sectors + (-min_offset_diff))) {
6748 /* Reading from the same stripe as writing to - bad */
6749 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6750 "auto-recovery - aborting.\n",
6754 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6756 /* OK, we should be able to continue; */
6758 BUG_ON(mddev->level != mddev->new_level);
6759 BUG_ON(mddev->layout != mddev->new_layout);
6760 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6761 BUG_ON(mddev->delta_disks != 0);
6764 if (mddev->private == NULL)
6765 conf = setup_conf(mddev);
6767 conf = mddev->private;
6770 return PTR_ERR(conf);
6772 conf->min_offset_diff = min_offset_diff;
6773 mddev->thread = conf->thread;
6774 conf->thread = NULL;
6775 mddev->private = conf;
6777 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6779 rdev = conf->disks[i].rdev;
6780 if (!rdev && conf->disks[i].replacement) {
6781 /* The replacement is all we have yet */
6782 rdev = conf->disks[i].replacement;
6783 conf->disks[i].replacement = NULL;
6784 clear_bit(Replacement, &rdev->flags);
6785 conf->disks[i].rdev = rdev;
6789 if (conf->disks[i].replacement &&
6790 conf->reshape_progress != MaxSector) {
6791 /* replacements and reshape simply do not mix. */
6792 printk(KERN_ERR "md: cannot handle concurrent "
6793 "replacement and reshape.\n");
6796 if (test_bit(In_sync, &rdev->flags)) {
6800 /* This disc is not fully in-sync. However if it
6801 * just stored parity (beyond the recovery_offset),
6802 * when we don't need to be concerned about the
6803 * array being dirty.
6804 * When reshape goes 'backwards', we never have
6805 * partially completed devices, so we only need
6806 * to worry about reshape going forwards.
6808 /* Hack because v0.91 doesn't store recovery_offset properly. */
6809 if (mddev->major_version == 0 &&
6810 mddev->minor_version > 90)
6811 rdev->recovery_offset = reshape_offset;
6813 if (rdev->recovery_offset < reshape_offset) {
6814 /* We need to check old and new layout */
6815 if (!only_parity(rdev->raid_disk,
6818 conf->max_degraded))
6821 if (!only_parity(rdev->raid_disk,
6823 conf->previous_raid_disks,
6824 conf->max_degraded))
6826 dirty_parity_disks++;
6830 * 0 for a fully functional array, 1 or 2 for a degraded array.
6832 mddev->degraded = calc_degraded(conf);
6834 if (has_failed(conf)) {
6835 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6836 " (%d/%d failed)\n",
6837 mdname(mddev), mddev->degraded, conf->raid_disks);
6841 /* device size must be a multiple of chunk size */
6842 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6843 mddev->resync_max_sectors = mddev->dev_sectors;
6845 if (mddev->degraded > dirty_parity_disks &&
6846 mddev->recovery_cp != MaxSector) {
6847 if (mddev->ok_start_degraded)
6849 "md/raid:%s: starting dirty degraded array"
6850 " - data corruption possible.\n",
6854 "md/raid:%s: cannot start dirty degraded array.\n",
6860 if (mddev->degraded == 0)
6861 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6862 " devices, algorithm %d\n", mdname(mddev), conf->level,
6863 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6866 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6867 " out of %d devices, algorithm %d\n",
6868 mdname(mddev), conf->level,
6869 mddev->raid_disks - mddev->degraded,
6870 mddev->raid_disks, mddev->new_layout);
6872 print_raid5_conf(conf);
6874 if (conf->reshape_progress != MaxSector) {
6875 conf->reshape_safe = conf->reshape_progress;
6876 atomic_set(&conf->reshape_stripes, 0);
6877 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6878 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6879 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6880 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6881 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6885 /* Ok, everything is just fine now */
6886 if (mddev->to_remove == &raid5_attrs_group)
6887 mddev->to_remove = NULL;
6888 else if (mddev->kobj.sd &&
6889 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6891 "raid5: failed to create sysfs attributes for %s\n",
6893 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6897 bool discard_supported = true;
6898 /* read-ahead size must cover two whole stripes, which
6899 * is 2 * (datadisks) * chunksize where 'n' is the
6900 * number of raid devices
6902 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6903 int stripe = data_disks *
6904 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6905 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6906 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6908 chunk_size = mddev->chunk_sectors << 9;
6909 blk_queue_io_min(mddev->queue, chunk_size);
6910 blk_queue_io_opt(mddev->queue, chunk_size *
6911 (conf->raid_disks - conf->max_degraded));
6912 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6914 * We can only discard a whole stripe. It doesn't make sense to
6915 * discard data disk but write parity disk
6917 stripe = stripe * PAGE_SIZE;
6918 /* Round up to power of 2, as discard handling
6919 * currently assumes that */
6920 while ((stripe-1) & stripe)
6921 stripe = (stripe | (stripe-1)) + 1;
6922 mddev->queue->limits.discard_alignment = stripe;
6923 mddev->queue->limits.discard_granularity = stripe;
6925 * unaligned part of discard request will be ignored, so can't
6926 * guarantee discard_zeroes_data
6928 mddev->queue->limits.discard_zeroes_data = 0;
6930 blk_queue_max_write_same_sectors(mddev->queue, 0);
6932 rdev_for_each(rdev, mddev) {
6933 disk_stack_limits(mddev->gendisk, rdev->bdev,
6934 rdev->data_offset << 9);
6935 disk_stack_limits(mddev->gendisk, rdev->bdev,
6936 rdev->new_data_offset << 9);
6938 * discard_zeroes_data is required, otherwise data
6939 * could be lost. Consider a scenario: discard a stripe
6940 * (the stripe could be inconsistent if
6941 * discard_zeroes_data is 0); write one disk of the
6942 * stripe (the stripe could be inconsistent again
6943 * depending on which disks are used to calculate
6944 * parity); the disk is broken; The stripe data of this
6947 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6948 !bdev_get_queue(rdev->bdev)->
6949 limits.discard_zeroes_data)
6950 discard_supported = false;
6951 /* Unfortunately, discard_zeroes_data is not currently
6952 * a guarantee - just a hint. So we only allow DISCARD
6953 * if the sysadmin has confirmed that only safe devices
6954 * are in use by setting a module parameter.
6956 if (!devices_handle_discard_safely) {
6957 if (discard_supported) {
6958 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6959 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6961 discard_supported = false;
6965 if (discard_supported &&
6966 mddev->queue->limits.max_discard_sectors >= stripe &&
6967 mddev->queue->limits.discard_granularity >= stripe)
6968 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6971 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6977 md_unregister_thread(&mddev->thread);
6978 print_raid5_conf(conf);
6980 mddev->private = NULL;
6981 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6985 static void raid5_free(struct mddev *mddev, void *priv)
6987 struct r5conf *conf = priv;
6990 mddev->to_remove = &raid5_attrs_group;
6993 static void status(struct seq_file *seq, struct mddev *mddev)
6995 struct r5conf *conf = mddev->private;
6998 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6999 conf->chunk_sectors / 2, mddev->layout);
7000 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7001 for (i = 0; i < conf->raid_disks; i++)
7002 seq_printf (seq, "%s",
7003 conf->disks[i].rdev &&
7004 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
7005 seq_printf (seq, "]");
7008 static void print_raid5_conf (struct r5conf *conf)
7011 struct disk_info *tmp;
7013 printk(KERN_DEBUG "RAID conf printout:\n");
7015 printk("(conf==NULL)\n");
7018 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7020 conf->raid_disks - conf->mddev->degraded);
7022 for (i = 0; i < conf->raid_disks; i++) {
7023 char b[BDEVNAME_SIZE];
7024 tmp = conf->disks + i;
7026 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7027 i, !test_bit(Faulty, &tmp->rdev->flags),
7028 bdevname(tmp->rdev->bdev, b));
7032 static int raid5_spare_active(struct mddev *mddev)
7035 struct r5conf *conf = mddev->private;
7036 struct disk_info *tmp;
7038 unsigned long flags;
7040 for (i = 0; i < conf->raid_disks; i++) {
7041 tmp = conf->disks + i;
7042 if (tmp->replacement
7043 && tmp->replacement->recovery_offset == MaxSector
7044 && !test_bit(Faulty, &tmp->replacement->flags)
7045 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7046 /* Replacement has just become active. */
7048 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7051 /* Replaced device not technically faulty,
7052 * but we need to be sure it gets removed
7053 * and never re-added.
7055 set_bit(Faulty, &tmp->rdev->flags);
7056 sysfs_notify_dirent_safe(
7057 tmp->rdev->sysfs_state);
7059 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7060 } else if (tmp->rdev
7061 && tmp->rdev->recovery_offset == MaxSector
7062 && !test_bit(Faulty, &tmp->rdev->flags)
7063 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7065 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7068 spin_lock_irqsave(&conf->device_lock, flags);
7069 mddev->degraded = calc_degraded(conf);
7070 spin_unlock_irqrestore(&conf->device_lock, flags);
7071 print_raid5_conf(conf);
7075 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7077 struct r5conf *conf = mddev->private;
7079 int number = rdev->raid_disk;
7080 struct md_rdev **rdevp;
7081 struct disk_info *p = conf->disks + number;
7083 print_raid5_conf(conf);
7084 if (rdev == p->rdev)
7086 else if (rdev == p->replacement)
7087 rdevp = &p->replacement;
7091 if (number >= conf->raid_disks &&
7092 conf->reshape_progress == MaxSector)
7093 clear_bit(In_sync, &rdev->flags);
7095 if (test_bit(In_sync, &rdev->flags) ||
7096 atomic_read(&rdev->nr_pending)) {
7100 /* Only remove non-faulty devices if recovery
7103 if (!test_bit(Faulty, &rdev->flags) &&
7104 mddev->recovery_disabled != conf->recovery_disabled &&
7105 !has_failed(conf) &&
7106 (!p->replacement || p->replacement == rdev) &&
7107 number < conf->raid_disks) {
7113 if (atomic_read(&rdev->nr_pending)) {
7114 /* lost the race, try later */
7117 } else if (p->replacement) {
7118 /* We must have just cleared 'rdev' */
7119 p->rdev = p->replacement;
7120 clear_bit(Replacement, &p->replacement->flags);
7121 smp_mb(); /* Make sure other CPUs may see both as identical
7122 * but will never see neither - if they are careful
7124 p->replacement = NULL;
7125 clear_bit(WantReplacement, &rdev->flags);
7127 /* We might have just removed the Replacement as faulty-
7128 * clear the bit just in case
7130 clear_bit(WantReplacement, &rdev->flags);
7133 print_raid5_conf(conf);
7137 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7139 struct r5conf *conf = mddev->private;
7142 struct disk_info *p;
7144 int last = conf->raid_disks - 1;
7146 if (mddev->recovery_disabled == conf->recovery_disabled)
7149 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7150 /* no point adding a device */
7153 if (rdev->raid_disk >= 0)
7154 first = last = rdev->raid_disk;
7157 * find the disk ... but prefer rdev->saved_raid_disk
7160 if (rdev->saved_raid_disk >= 0 &&
7161 rdev->saved_raid_disk >= first &&
7162 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7163 first = rdev->saved_raid_disk;
7165 for (disk = first; disk <= last; disk++) {
7166 p = conf->disks + disk;
7167 if (p->rdev == NULL) {
7168 clear_bit(In_sync, &rdev->flags);
7169 rdev->raid_disk = disk;
7171 if (rdev->saved_raid_disk != disk)
7173 rcu_assign_pointer(p->rdev, rdev);
7177 for (disk = first; disk <= last; disk++) {
7178 p = conf->disks + disk;
7179 if (test_bit(WantReplacement, &p->rdev->flags) &&
7180 p->replacement == NULL) {
7181 clear_bit(In_sync, &rdev->flags);
7182 set_bit(Replacement, &rdev->flags);
7183 rdev->raid_disk = disk;
7186 rcu_assign_pointer(p->replacement, rdev);
7191 print_raid5_conf(conf);
7195 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7197 /* no resync is happening, and there is enough space
7198 * on all devices, so we can resize.
7199 * We need to make sure resync covers any new space.
7200 * If the array is shrinking we should possibly wait until
7201 * any io in the removed space completes, but it hardly seems
7205 struct r5conf *conf = mddev->private;
7207 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7208 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7209 if (mddev->external_size &&
7210 mddev->array_sectors > newsize)
7212 if (mddev->bitmap) {
7213 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7217 md_set_array_sectors(mddev, newsize);
7218 set_capacity(mddev->gendisk, mddev->array_sectors);
7219 revalidate_disk(mddev->gendisk);
7220 if (sectors > mddev->dev_sectors &&
7221 mddev->recovery_cp > mddev->dev_sectors) {
7222 mddev->recovery_cp = mddev->dev_sectors;
7223 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7225 mddev->dev_sectors = sectors;
7226 mddev->resync_max_sectors = sectors;
7230 static int check_stripe_cache(struct mddev *mddev)
7232 /* Can only proceed if there are plenty of stripe_heads.
7233 * We need a minimum of one full stripe,, and for sensible progress
7234 * it is best to have about 4 times that.
7235 * If we require 4 times, then the default 256 4K stripe_heads will
7236 * allow for chunk sizes up to 256K, which is probably OK.
7237 * If the chunk size is greater, user-space should request more
7238 * stripe_heads first.
7240 struct r5conf *conf = mddev->private;
7241 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7242 > conf->min_nr_stripes ||
7243 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7244 > conf->min_nr_stripes) {
7245 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7247 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7254 static int check_reshape(struct mddev *mddev)
7256 struct r5conf *conf = mddev->private;
7258 if (mddev->delta_disks == 0 &&
7259 mddev->new_layout == mddev->layout &&
7260 mddev->new_chunk_sectors == mddev->chunk_sectors)
7261 return 0; /* nothing to do */
7262 if (has_failed(conf))
7264 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7265 /* We might be able to shrink, but the devices must
7266 * be made bigger first.
7267 * For raid6, 4 is the minimum size.
7268 * Otherwise 2 is the minimum
7271 if (mddev->level == 6)
7273 if (mddev->raid_disks + mddev->delta_disks < min)
7277 if (!check_stripe_cache(mddev))
7280 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7281 mddev->delta_disks > 0)
7282 if (resize_chunks(conf,
7283 conf->previous_raid_disks
7284 + max(0, mddev->delta_disks),
7285 max(mddev->new_chunk_sectors,
7286 mddev->chunk_sectors)
7289 return resize_stripes(conf, (conf->previous_raid_disks
7290 + mddev->delta_disks));
7293 static int raid5_start_reshape(struct mddev *mddev)
7295 struct r5conf *conf = mddev->private;
7296 struct md_rdev *rdev;
7298 unsigned long flags;
7300 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7303 if (!check_stripe_cache(mddev))
7306 if (has_failed(conf))
7309 rdev_for_each(rdev, mddev) {
7310 if (!test_bit(In_sync, &rdev->flags)
7311 && !test_bit(Faulty, &rdev->flags))
7315 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7316 /* Not enough devices even to make a degraded array
7321 /* Refuse to reduce size of the array. Any reductions in
7322 * array size must be through explicit setting of array_size
7325 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7326 < mddev->array_sectors) {
7327 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7328 "before number of disks\n", mdname(mddev));
7332 atomic_set(&conf->reshape_stripes, 0);
7333 spin_lock_irq(&conf->device_lock);
7334 write_seqcount_begin(&conf->gen_lock);
7335 conf->previous_raid_disks = conf->raid_disks;
7336 conf->raid_disks += mddev->delta_disks;
7337 conf->prev_chunk_sectors = conf->chunk_sectors;
7338 conf->chunk_sectors = mddev->new_chunk_sectors;
7339 conf->prev_algo = conf->algorithm;
7340 conf->algorithm = mddev->new_layout;
7342 /* Code that selects data_offset needs to see the generation update
7343 * if reshape_progress has been set - so a memory barrier needed.
7346 if (mddev->reshape_backwards)
7347 conf->reshape_progress = raid5_size(mddev, 0, 0);
7349 conf->reshape_progress = 0;
7350 conf->reshape_safe = conf->reshape_progress;
7351 write_seqcount_end(&conf->gen_lock);
7352 spin_unlock_irq(&conf->device_lock);
7354 /* Now make sure any requests that proceeded on the assumption
7355 * the reshape wasn't running - like Discard or Read - have
7358 mddev_suspend(mddev);
7359 mddev_resume(mddev);
7361 /* Add some new drives, as many as will fit.
7362 * We know there are enough to make the newly sized array work.
7363 * Don't add devices if we are reducing the number of
7364 * devices in the array. This is because it is not possible
7365 * to correctly record the "partially reconstructed" state of
7366 * such devices during the reshape and confusion could result.
7368 if (mddev->delta_disks >= 0) {
7369 rdev_for_each(rdev, mddev)
7370 if (rdev->raid_disk < 0 &&
7371 !test_bit(Faulty, &rdev->flags)) {
7372 if (raid5_add_disk(mddev, rdev) == 0) {
7374 >= conf->previous_raid_disks)
7375 set_bit(In_sync, &rdev->flags);
7377 rdev->recovery_offset = 0;
7379 if (sysfs_link_rdev(mddev, rdev))
7380 /* Failure here is OK */;
7382 } else if (rdev->raid_disk >= conf->previous_raid_disks
7383 && !test_bit(Faulty, &rdev->flags)) {
7384 /* This is a spare that was manually added */
7385 set_bit(In_sync, &rdev->flags);
7388 /* When a reshape changes the number of devices,
7389 * ->degraded is measured against the larger of the
7390 * pre and post number of devices.
7392 spin_lock_irqsave(&conf->device_lock, flags);
7393 mddev->degraded = calc_degraded(conf);
7394 spin_unlock_irqrestore(&conf->device_lock, flags);
7396 mddev->raid_disks = conf->raid_disks;
7397 mddev->reshape_position = conf->reshape_progress;
7398 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7400 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7401 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7402 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7403 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7404 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7405 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7407 if (!mddev->sync_thread) {
7408 mddev->recovery = 0;
7409 spin_lock_irq(&conf->device_lock);
7410 write_seqcount_begin(&conf->gen_lock);
7411 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7412 mddev->new_chunk_sectors =
7413 conf->chunk_sectors = conf->prev_chunk_sectors;
7414 mddev->new_layout = conf->algorithm = conf->prev_algo;
7415 rdev_for_each(rdev, mddev)
7416 rdev->new_data_offset = rdev->data_offset;
7418 conf->generation --;
7419 conf->reshape_progress = MaxSector;
7420 mddev->reshape_position = MaxSector;
7421 write_seqcount_end(&conf->gen_lock);
7422 spin_unlock_irq(&conf->device_lock);
7425 conf->reshape_checkpoint = jiffies;
7426 md_wakeup_thread(mddev->sync_thread);
7427 md_new_event(mddev);
7431 /* This is called from the reshape thread and should make any
7432 * changes needed in 'conf'
7434 static void end_reshape(struct r5conf *conf)
7437 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7438 struct md_rdev *rdev;
7440 spin_lock_irq(&conf->device_lock);
7441 conf->previous_raid_disks = conf->raid_disks;
7442 rdev_for_each(rdev, conf->mddev)
7443 rdev->data_offset = rdev->new_data_offset;
7445 conf->reshape_progress = MaxSector;
7446 conf->mddev->reshape_position = MaxSector;
7447 spin_unlock_irq(&conf->device_lock);
7448 wake_up(&conf->wait_for_overlap);
7450 /* read-ahead size must cover two whole stripes, which is
7451 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7453 if (conf->mddev->queue) {
7454 int data_disks = conf->raid_disks - conf->max_degraded;
7455 int stripe = data_disks * ((conf->chunk_sectors << 9)
7457 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7458 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7463 /* This is called from the raid5d thread with mddev_lock held.
7464 * It makes config changes to the device.
7466 static void raid5_finish_reshape(struct mddev *mddev)
7468 struct r5conf *conf = mddev->private;
7470 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7472 if (mddev->delta_disks > 0) {
7473 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7474 set_capacity(mddev->gendisk, mddev->array_sectors);
7475 revalidate_disk(mddev->gendisk);
7478 spin_lock_irq(&conf->device_lock);
7479 mddev->degraded = calc_degraded(conf);
7480 spin_unlock_irq(&conf->device_lock);
7481 for (d = conf->raid_disks ;
7482 d < conf->raid_disks - mddev->delta_disks;
7484 struct md_rdev *rdev = conf->disks[d].rdev;
7486 clear_bit(In_sync, &rdev->flags);
7487 rdev = conf->disks[d].replacement;
7489 clear_bit(In_sync, &rdev->flags);
7492 mddev->layout = conf->algorithm;
7493 mddev->chunk_sectors = conf->chunk_sectors;
7494 mddev->reshape_position = MaxSector;
7495 mddev->delta_disks = 0;
7496 mddev->reshape_backwards = 0;
7500 static void raid5_quiesce(struct mddev *mddev, int state)
7502 struct r5conf *conf = mddev->private;
7505 case 2: /* resume for a suspend */
7506 wake_up(&conf->wait_for_overlap);
7509 case 1: /* stop all writes */
7510 lock_all_device_hash_locks_irq(conf);
7511 /* '2' tells resync/reshape to pause so that all
7512 * active stripes can drain
7515 wait_event_cmd(conf->wait_for_quiescent,
7516 atomic_read(&conf->active_stripes) == 0 &&
7517 atomic_read(&conf->active_aligned_reads) == 0,
7518 unlock_all_device_hash_locks_irq(conf),
7519 lock_all_device_hash_locks_irq(conf));
7521 unlock_all_device_hash_locks_irq(conf);
7522 /* allow reshape to continue */
7523 wake_up(&conf->wait_for_overlap);
7526 case 0: /* re-enable writes */
7527 lock_all_device_hash_locks_irq(conf);
7529 wake_up(&conf->wait_for_quiescent);
7530 wake_up(&conf->wait_for_overlap);
7531 unlock_all_device_hash_locks_irq(conf);
7536 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7538 struct r0conf *raid0_conf = mddev->private;
7541 /* for raid0 takeover only one zone is supported */
7542 if (raid0_conf->nr_strip_zones > 1) {
7543 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7545 return ERR_PTR(-EINVAL);
7548 sectors = raid0_conf->strip_zone[0].zone_end;
7549 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7550 mddev->dev_sectors = sectors;
7551 mddev->new_level = level;
7552 mddev->new_layout = ALGORITHM_PARITY_N;
7553 mddev->new_chunk_sectors = mddev->chunk_sectors;
7554 mddev->raid_disks += 1;
7555 mddev->delta_disks = 1;
7556 /* make sure it will be not marked as dirty */
7557 mddev->recovery_cp = MaxSector;
7559 return setup_conf(mddev);
7562 static void *raid5_takeover_raid1(struct mddev *mddev)
7566 if (mddev->raid_disks != 2 ||
7567 mddev->degraded > 1)
7568 return ERR_PTR(-EINVAL);
7570 /* Should check if there are write-behind devices? */
7572 chunksect = 64*2; /* 64K by default */
7574 /* The array must be an exact multiple of chunksize */
7575 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7578 if ((chunksect<<9) < STRIPE_SIZE)
7579 /* array size does not allow a suitable chunk size */
7580 return ERR_PTR(-EINVAL);
7582 mddev->new_level = 5;
7583 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7584 mddev->new_chunk_sectors = chunksect;
7586 return setup_conf(mddev);
7589 static void *raid5_takeover_raid6(struct mddev *mddev)
7593 switch (mddev->layout) {
7594 case ALGORITHM_LEFT_ASYMMETRIC_6:
7595 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7597 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7598 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7600 case ALGORITHM_LEFT_SYMMETRIC_6:
7601 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7603 case ALGORITHM_RIGHT_SYMMETRIC_6:
7604 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7606 case ALGORITHM_PARITY_0_6:
7607 new_layout = ALGORITHM_PARITY_0;
7609 case ALGORITHM_PARITY_N:
7610 new_layout = ALGORITHM_PARITY_N;
7613 return ERR_PTR(-EINVAL);
7615 mddev->new_level = 5;
7616 mddev->new_layout = new_layout;
7617 mddev->delta_disks = -1;
7618 mddev->raid_disks -= 1;
7619 return setup_conf(mddev);
7622 static int raid5_check_reshape(struct mddev *mddev)
7624 /* For a 2-drive array, the layout and chunk size can be changed
7625 * immediately as not restriping is needed.
7626 * For larger arrays we record the new value - after validation
7627 * to be used by a reshape pass.
7629 struct r5conf *conf = mddev->private;
7630 int new_chunk = mddev->new_chunk_sectors;
7632 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7634 if (new_chunk > 0) {
7635 if (!is_power_of_2(new_chunk))
7637 if (new_chunk < (PAGE_SIZE>>9))
7639 if (mddev->array_sectors & (new_chunk-1))
7640 /* not factor of array size */
7644 /* They look valid */
7646 if (mddev->raid_disks == 2) {
7647 /* can make the change immediately */
7648 if (mddev->new_layout >= 0) {
7649 conf->algorithm = mddev->new_layout;
7650 mddev->layout = mddev->new_layout;
7652 if (new_chunk > 0) {
7653 conf->chunk_sectors = new_chunk ;
7654 mddev->chunk_sectors = new_chunk;
7656 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7657 md_wakeup_thread(mddev->thread);
7659 return check_reshape(mddev);
7662 static int raid6_check_reshape(struct mddev *mddev)
7664 int new_chunk = mddev->new_chunk_sectors;
7666 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7668 if (new_chunk > 0) {
7669 if (!is_power_of_2(new_chunk))
7671 if (new_chunk < (PAGE_SIZE >> 9))
7673 if (mddev->array_sectors & (new_chunk-1))
7674 /* not factor of array size */
7678 /* They look valid */
7679 return check_reshape(mddev);
7682 static void *raid5_takeover(struct mddev *mddev)
7684 /* raid5 can take over:
7685 * raid0 - if there is only one strip zone - make it a raid4 layout
7686 * raid1 - if there are two drives. We need to know the chunk size
7687 * raid4 - trivial - just use a raid4 layout.
7688 * raid6 - Providing it is a *_6 layout
7690 if (mddev->level == 0)
7691 return raid45_takeover_raid0(mddev, 5);
7692 if (mddev->level == 1)
7693 return raid5_takeover_raid1(mddev);
7694 if (mddev->level == 4) {
7695 mddev->new_layout = ALGORITHM_PARITY_N;
7696 mddev->new_level = 5;
7697 return setup_conf(mddev);
7699 if (mddev->level == 6)
7700 return raid5_takeover_raid6(mddev);
7702 return ERR_PTR(-EINVAL);
7705 static void *raid4_takeover(struct mddev *mddev)
7707 /* raid4 can take over:
7708 * raid0 - if there is only one strip zone
7709 * raid5 - if layout is right
7711 if (mddev->level == 0)
7712 return raid45_takeover_raid0(mddev, 4);
7713 if (mddev->level == 5 &&
7714 mddev->layout == ALGORITHM_PARITY_N) {
7715 mddev->new_layout = 0;
7716 mddev->new_level = 4;
7717 return setup_conf(mddev);
7719 return ERR_PTR(-EINVAL);
7722 static struct md_personality raid5_personality;
7724 static void *raid6_takeover(struct mddev *mddev)
7726 /* Currently can only take over a raid5. We map the
7727 * personality to an equivalent raid6 personality
7728 * with the Q block at the end.
7732 if (mddev->pers != &raid5_personality)
7733 return ERR_PTR(-EINVAL);
7734 if (mddev->degraded > 1)
7735 return ERR_PTR(-EINVAL);
7736 if (mddev->raid_disks > 253)
7737 return ERR_PTR(-EINVAL);
7738 if (mddev->raid_disks < 3)
7739 return ERR_PTR(-EINVAL);
7741 switch (mddev->layout) {
7742 case ALGORITHM_LEFT_ASYMMETRIC:
7743 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7745 case ALGORITHM_RIGHT_ASYMMETRIC:
7746 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7748 case ALGORITHM_LEFT_SYMMETRIC:
7749 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7751 case ALGORITHM_RIGHT_SYMMETRIC:
7752 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7754 case ALGORITHM_PARITY_0:
7755 new_layout = ALGORITHM_PARITY_0_6;
7757 case ALGORITHM_PARITY_N:
7758 new_layout = ALGORITHM_PARITY_N;
7761 return ERR_PTR(-EINVAL);
7763 mddev->new_level = 6;
7764 mddev->new_layout = new_layout;
7765 mddev->delta_disks = 1;
7766 mddev->raid_disks += 1;
7767 return setup_conf(mddev);
7770 static struct md_personality raid6_personality =
7774 .owner = THIS_MODULE,
7775 .make_request = make_request,
7779 .error_handler = error,
7780 .hot_add_disk = raid5_add_disk,
7781 .hot_remove_disk= raid5_remove_disk,
7782 .spare_active = raid5_spare_active,
7783 .sync_request = sync_request,
7784 .resize = raid5_resize,
7786 .check_reshape = raid6_check_reshape,
7787 .start_reshape = raid5_start_reshape,
7788 .finish_reshape = raid5_finish_reshape,
7789 .quiesce = raid5_quiesce,
7790 .takeover = raid6_takeover,
7791 .congested = raid5_congested,
7793 static struct md_personality raid5_personality =
7797 .owner = THIS_MODULE,
7798 .make_request = make_request,
7802 .error_handler = error,
7803 .hot_add_disk = raid5_add_disk,
7804 .hot_remove_disk= raid5_remove_disk,
7805 .spare_active = raid5_spare_active,
7806 .sync_request = sync_request,
7807 .resize = raid5_resize,
7809 .check_reshape = raid5_check_reshape,
7810 .start_reshape = raid5_start_reshape,
7811 .finish_reshape = raid5_finish_reshape,
7812 .quiesce = raid5_quiesce,
7813 .takeover = raid5_takeover,
7814 .congested = raid5_congested,
7817 static struct md_personality raid4_personality =
7821 .owner = THIS_MODULE,
7822 .make_request = make_request,
7826 .error_handler = error,
7827 .hot_add_disk = raid5_add_disk,
7828 .hot_remove_disk= raid5_remove_disk,
7829 .spare_active = raid5_spare_active,
7830 .sync_request = sync_request,
7831 .resize = raid5_resize,
7833 .check_reshape = raid5_check_reshape,
7834 .start_reshape = raid5_start_reshape,
7835 .finish_reshape = raid5_finish_reshape,
7836 .quiesce = raid5_quiesce,
7837 .takeover = raid4_takeover,
7838 .congested = raid5_congested,
7841 static int __init raid5_init(void)
7843 raid5_wq = alloc_workqueue("raid5wq",
7844 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7847 register_md_personality(&raid6_personality);
7848 register_md_personality(&raid5_personality);
7849 register_md_personality(&raid4_personality);
7853 static void raid5_exit(void)
7855 unregister_md_personality(&raid6_personality);
7856 unregister_md_personality(&raid5_personality);
7857 unregister_md_personality(&raid4_personality);
7858 destroy_workqueue(raid5_wq);
7861 module_init(raid5_init);
7862 module_exit(raid5_exit);
7863 MODULE_LICENSE("GPL");
7864 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7865 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7866 MODULE_ALIAS("md-raid5");
7867 MODULE_ALIAS("md-raid4");
7868 MODULE_ALIAS("md-level-5");
7869 MODULE_ALIAS("md-level-4");
7870 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7871 MODULE_ALIAS("md-raid6");
7872 MODULE_ALIAS("md-level-6");
7874 /* This used to be two separate modules, they were: */
7875 MODULE_ALIAS("raid5");
7876 MODULE_ALIAS("raid6");