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 kmem_cache_destroy(conf->slab_cache);
2275 conf->slab_cache = NULL;
2278 static void raid5_end_read_request(struct bio * bi)
2280 struct stripe_head *sh = bi->bi_private;
2281 struct r5conf *conf = sh->raid_conf;
2282 int disks = sh->disks, i;
2283 char b[BDEVNAME_SIZE];
2284 struct md_rdev *rdev = NULL;
2287 for (i=0 ; i<disks; i++)
2288 if (bi == &sh->dev[i].req)
2291 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2292 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2298 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2299 /* If replacement finished while this request was outstanding,
2300 * 'replacement' might be NULL already.
2301 * In that case it moved down to 'rdev'.
2302 * rdev is not removed until all requests are finished.
2304 rdev = conf->disks[i].replacement;
2306 rdev = conf->disks[i].rdev;
2308 if (use_new_offset(conf, sh))
2309 s = sh->sector + rdev->new_data_offset;
2311 s = sh->sector + rdev->data_offset;
2312 if (!bi->bi_error) {
2313 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2314 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2315 /* Note that this cannot happen on a
2316 * replacement device. We just fail those on
2321 "md/raid:%s: read error corrected"
2322 " (%lu sectors at %llu on %s)\n",
2323 mdname(conf->mddev), STRIPE_SECTORS,
2324 (unsigned long long)s,
2325 bdevname(rdev->bdev, b));
2326 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2327 clear_bit(R5_ReadError, &sh->dev[i].flags);
2328 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2329 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2330 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2332 if (atomic_read(&rdev->read_errors))
2333 atomic_set(&rdev->read_errors, 0);
2335 const char *bdn = bdevname(rdev->bdev, b);
2339 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2340 atomic_inc(&rdev->read_errors);
2341 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2344 "md/raid:%s: read error on replacement device "
2345 "(sector %llu on %s).\n",
2346 mdname(conf->mddev),
2347 (unsigned long long)s,
2349 else if (conf->mddev->degraded >= conf->max_degraded) {
2353 "md/raid:%s: read error not correctable "
2354 "(sector %llu on %s).\n",
2355 mdname(conf->mddev),
2356 (unsigned long long)s,
2358 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2363 "md/raid:%s: read error NOT corrected!! "
2364 "(sector %llu on %s).\n",
2365 mdname(conf->mddev),
2366 (unsigned long long)s,
2368 } else if (atomic_read(&rdev->read_errors)
2369 > conf->max_nr_stripes)
2371 "md/raid:%s: Too many read errors, failing device %s.\n",
2372 mdname(conf->mddev), bdn);
2375 if (set_bad && test_bit(In_sync, &rdev->flags)
2376 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2379 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2380 set_bit(R5_ReadError, &sh->dev[i].flags);
2381 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2383 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2385 clear_bit(R5_ReadError, &sh->dev[i].flags);
2386 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2388 && test_bit(In_sync, &rdev->flags)
2389 && rdev_set_badblocks(
2390 rdev, sh->sector, STRIPE_SECTORS, 0)))
2391 md_error(conf->mddev, rdev);
2394 rdev_dec_pending(rdev, conf->mddev);
2395 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2396 set_bit(STRIPE_HANDLE, &sh->state);
2400 static void raid5_end_write_request(struct bio *bi)
2402 struct stripe_head *sh = bi->bi_private;
2403 struct r5conf *conf = sh->raid_conf;
2404 int disks = sh->disks, i;
2405 struct md_rdev *uninitialized_var(rdev);
2408 int replacement = 0;
2410 for (i = 0 ; i < disks; i++) {
2411 if (bi == &sh->dev[i].req) {
2412 rdev = conf->disks[i].rdev;
2415 if (bi == &sh->dev[i].rreq) {
2416 rdev = conf->disks[i].replacement;
2420 /* rdev was removed and 'replacement'
2421 * replaced it. rdev is not removed
2422 * until all requests are finished.
2424 rdev = conf->disks[i].rdev;
2428 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2429 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2438 md_error(conf->mddev, rdev);
2439 else if (is_badblock(rdev, sh->sector,
2441 &first_bad, &bad_sectors))
2442 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2445 set_bit(STRIPE_DEGRADED, &sh->state);
2446 set_bit(WriteErrorSeen, &rdev->flags);
2447 set_bit(R5_WriteError, &sh->dev[i].flags);
2448 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2449 set_bit(MD_RECOVERY_NEEDED,
2450 &rdev->mddev->recovery);
2451 } else if (is_badblock(rdev, sh->sector,
2453 &first_bad, &bad_sectors)) {
2454 set_bit(R5_MadeGood, &sh->dev[i].flags);
2455 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2456 /* That was a successful write so make
2457 * sure it looks like we already did
2460 set_bit(R5_ReWrite, &sh->dev[i].flags);
2463 rdev_dec_pending(rdev, conf->mddev);
2465 if (sh->batch_head && bi->bi_error && !replacement)
2466 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2468 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2469 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2470 set_bit(STRIPE_HANDLE, &sh->state);
2473 if (sh->batch_head && sh != sh->batch_head)
2474 release_stripe(sh->batch_head);
2477 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2479 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2481 struct r5dev *dev = &sh->dev[i];
2483 bio_init(&dev->req);
2484 dev->req.bi_io_vec = &dev->vec;
2485 dev->req.bi_max_vecs = 1;
2486 dev->req.bi_private = sh;
2488 bio_init(&dev->rreq);
2489 dev->rreq.bi_io_vec = &dev->rvec;
2490 dev->rreq.bi_max_vecs = 1;
2491 dev->rreq.bi_private = sh;
2494 dev->sector = compute_blocknr(sh, i, previous);
2497 static void error(struct mddev *mddev, struct md_rdev *rdev)
2499 char b[BDEVNAME_SIZE];
2500 struct r5conf *conf = mddev->private;
2501 unsigned long flags;
2502 pr_debug("raid456: error called\n");
2504 spin_lock_irqsave(&conf->device_lock, flags);
2505 clear_bit(In_sync, &rdev->flags);
2506 mddev->degraded = calc_degraded(conf);
2507 spin_unlock_irqrestore(&conf->device_lock, flags);
2508 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2510 set_bit(Blocked, &rdev->flags);
2511 set_bit(Faulty, &rdev->flags);
2512 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2513 set_bit(MD_CHANGE_PENDING, &mddev->flags);
2515 "md/raid:%s: Disk failure on %s, disabling device.\n"
2516 "md/raid:%s: Operation continuing on %d devices.\n",
2518 bdevname(rdev->bdev, b),
2520 conf->raid_disks - mddev->degraded);
2524 * Input: a 'big' sector number,
2525 * Output: index of the data and parity disk, and the sector # in them.
2527 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2528 int previous, int *dd_idx,
2529 struct stripe_head *sh)
2531 sector_t stripe, stripe2;
2532 sector_t chunk_number;
2533 unsigned int chunk_offset;
2536 sector_t new_sector;
2537 int algorithm = previous ? conf->prev_algo
2539 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2540 : conf->chunk_sectors;
2541 int raid_disks = previous ? conf->previous_raid_disks
2543 int data_disks = raid_disks - conf->max_degraded;
2545 /* First compute the information on this sector */
2548 * Compute the chunk number and the sector offset inside the chunk
2550 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2551 chunk_number = r_sector;
2554 * Compute the stripe number
2556 stripe = chunk_number;
2557 *dd_idx = sector_div(stripe, data_disks);
2560 * Select the parity disk based on the user selected algorithm.
2562 pd_idx = qd_idx = -1;
2563 switch(conf->level) {
2565 pd_idx = data_disks;
2568 switch (algorithm) {
2569 case ALGORITHM_LEFT_ASYMMETRIC:
2570 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2571 if (*dd_idx >= pd_idx)
2574 case ALGORITHM_RIGHT_ASYMMETRIC:
2575 pd_idx = sector_div(stripe2, raid_disks);
2576 if (*dd_idx >= pd_idx)
2579 case ALGORITHM_LEFT_SYMMETRIC:
2580 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2581 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2583 case ALGORITHM_RIGHT_SYMMETRIC:
2584 pd_idx = sector_div(stripe2, raid_disks);
2585 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2587 case ALGORITHM_PARITY_0:
2591 case ALGORITHM_PARITY_N:
2592 pd_idx = data_disks;
2600 switch (algorithm) {
2601 case ALGORITHM_LEFT_ASYMMETRIC:
2602 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2603 qd_idx = pd_idx + 1;
2604 if (pd_idx == raid_disks-1) {
2605 (*dd_idx)++; /* Q D D D P */
2607 } else if (*dd_idx >= pd_idx)
2608 (*dd_idx) += 2; /* D D P Q D */
2610 case ALGORITHM_RIGHT_ASYMMETRIC:
2611 pd_idx = sector_div(stripe2, raid_disks);
2612 qd_idx = pd_idx + 1;
2613 if (pd_idx == raid_disks-1) {
2614 (*dd_idx)++; /* Q D D D P */
2616 } else if (*dd_idx >= pd_idx)
2617 (*dd_idx) += 2; /* D D P Q D */
2619 case ALGORITHM_LEFT_SYMMETRIC:
2620 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2621 qd_idx = (pd_idx + 1) % raid_disks;
2622 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2624 case ALGORITHM_RIGHT_SYMMETRIC:
2625 pd_idx = sector_div(stripe2, raid_disks);
2626 qd_idx = (pd_idx + 1) % raid_disks;
2627 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2630 case ALGORITHM_PARITY_0:
2635 case ALGORITHM_PARITY_N:
2636 pd_idx = data_disks;
2637 qd_idx = data_disks + 1;
2640 case ALGORITHM_ROTATING_ZERO_RESTART:
2641 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2642 * of blocks for computing Q is different.
2644 pd_idx = sector_div(stripe2, raid_disks);
2645 qd_idx = pd_idx + 1;
2646 if (pd_idx == raid_disks-1) {
2647 (*dd_idx)++; /* Q D D D P */
2649 } else if (*dd_idx >= pd_idx)
2650 (*dd_idx) += 2; /* D D P Q D */
2654 case ALGORITHM_ROTATING_N_RESTART:
2655 /* Same a left_asymmetric, by first stripe is
2656 * D D D P Q rather than
2660 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2661 qd_idx = pd_idx + 1;
2662 if (pd_idx == raid_disks-1) {
2663 (*dd_idx)++; /* Q D D D P */
2665 } else if (*dd_idx >= pd_idx)
2666 (*dd_idx) += 2; /* D D P Q D */
2670 case ALGORITHM_ROTATING_N_CONTINUE:
2671 /* Same as left_symmetric but Q is before P */
2672 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2673 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2674 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2678 case ALGORITHM_LEFT_ASYMMETRIC_6:
2679 /* RAID5 left_asymmetric, with Q on last device */
2680 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2681 if (*dd_idx >= pd_idx)
2683 qd_idx = raid_disks - 1;
2686 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2687 pd_idx = sector_div(stripe2, raid_disks-1);
2688 if (*dd_idx >= pd_idx)
2690 qd_idx = raid_disks - 1;
2693 case ALGORITHM_LEFT_SYMMETRIC_6:
2694 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2695 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2696 qd_idx = raid_disks - 1;
2699 case ALGORITHM_RIGHT_SYMMETRIC_6:
2700 pd_idx = sector_div(stripe2, raid_disks-1);
2701 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2702 qd_idx = raid_disks - 1;
2705 case ALGORITHM_PARITY_0_6:
2708 qd_idx = raid_disks - 1;
2718 sh->pd_idx = pd_idx;
2719 sh->qd_idx = qd_idx;
2720 sh->ddf_layout = ddf_layout;
2723 * Finally, compute the new sector number
2725 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2729 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2731 struct r5conf *conf = sh->raid_conf;
2732 int raid_disks = sh->disks;
2733 int data_disks = raid_disks - conf->max_degraded;
2734 sector_t new_sector = sh->sector, check;
2735 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2736 : conf->chunk_sectors;
2737 int algorithm = previous ? conf->prev_algo
2741 sector_t chunk_number;
2742 int dummy1, dd_idx = i;
2744 struct stripe_head sh2;
2746 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2747 stripe = new_sector;
2749 if (i == sh->pd_idx)
2751 switch(conf->level) {
2754 switch (algorithm) {
2755 case ALGORITHM_LEFT_ASYMMETRIC:
2756 case ALGORITHM_RIGHT_ASYMMETRIC:
2760 case ALGORITHM_LEFT_SYMMETRIC:
2761 case ALGORITHM_RIGHT_SYMMETRIC:
2764 i -= (sh->pd_idx + 1);
2766 case ALGORITHM_PARITY_0:
2769 case ALGORITHM_PARITY_N:
2776 if (i == sh->qd_idx)
2777 return 0; /* It is the Q disk */
2778 switch (algorithm) {
2779 case ALGORITHM_LEFT_ASYMMETRIC:
2780 case ALGORITHM_RIGHT_ASYMMETRIC:
2781 case ALGORITHM_ROTATING_ZERO_RESTART:
2782 case ALGORITHM_ROTATING_N_RESTART:
2783 if (sh->pd_idx == raid_disks-1)
2784 i--; /* Q D D D P */
2785 else if (i > sh->pd_idx)
2786 i -= 2; /* D D P Q D */
2788 case ALGORITHM_LEFT_SYMMETRIC:
2789 case ALGORITHM_RIGHT_SYMMETRIC:
2790 if (sh->pd_idx == raid_disks-1)
2791 i--; /* Q D D D P */
2796 i -= (sh->pd_idx + 2);
2799 case ALGORITHM_PARITY_0:
2802 case ALGORITHM_PARITY_N:
2804 case ALGORITHM_ROTATING_N_CONTINUE:
2805 /* Like left_symmetric, but P is before Q */
2806 if (sh->pd_idx == 0)
2807 i--; /* P D D D Q */
2812 i -= (sh->pd_idx + 1);
2815 case ALGORITHM_LEFT_ASYMMETRIC_6:
2816 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2820 case ALGORITHM_LEFT_SYMMETRIC_6:
2821 case ALGORITHM_RIGHT_SYMMETRIC_6:
2823 i += data_disks + 1;
2824 i -= (sh->pd_idx + 1);
2826 case ALGORITHM_PARITY_0_6:
2835 chunk_number = stripe * data_disks + i;
2836 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2838 check = raid5_compute_sector(conf, r_sector,
2839 previous, &dummy1, &sh2);
2840 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2841 || sh2.qd_idx != sh->qd_idx) {
2842 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2843 mdname(conf->mddev));
2850 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2851 int rcw, int expand)
2853 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2854 struct r5conf *conf = sh->raid_conf;
2855 int level = conf->level;
2859 for (i = disks; i--; ) {
2860 struct r5dev *dev = &sh->dev[i];
2863 set_bit(R5_LOCKED, &dev->flags);
2864 set_bit(R5_Wantdrain, &dev->flags);
2866 clear_bit(R5_UPTODATE, &dev->flags);
2870 /* if we are not expanding this is a proper write request, and
2871 * there will be bios with new data to be drained into the
2876 /* False alarm, nothing to do */
2878 sh->reconstruct_state = reconstruct_state_drain_run;
2879 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2881 sh->reconstruct_state = reconstruct_state_run;
2883 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2885 if (s->locked + conf->max_degraded == disks)
2886 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2887 atomic_inc(&conf->pending_full_writes);
2889 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2890 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2891 BUG_ON(level == 6 &&
2892 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2893 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2895 for (i = disks; i--; ) {
2896 struct r5dev *dev = &sh->dev[i];
2897 if (i == pd_idx || i == qd_idx)
2901 (test_bit(R5_UPTODATE, &dev->flags) ||
2902 test_bit(R5_Wantcompute, &dev->flags))) {
2903 set_bit(R5_Wantdrain, &dev->flags);
2904 set_bit(R5_LOCKED, &dev->flags);
2905 clear_bit(R5_UPTODATE, &dev->flags);
2910 /* False alarm - nothing to do */
2912 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2913 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2914 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2915 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2918 /* keep the parity disk(s) locked while asynchronous operations
2921 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2922 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2926 int qd_idx = sh->qd_idx;
2927 struct r5dev *dev = &sh->dev[qd_idx];
2929 set_bit(R5_LOCKED, &dev->flags);
2930 clear_bit(R5_UPTODATE, &dev->flags);
2934 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2935 __func__, (unsigned long long)sh->sector,
2936 s->locked, s->ops_request);
2940 * Each stripe/dev can have one or more bion attached.
2941 * toread/towrite point to the first in a chain.
2942 * The bi_next chain must be in order.
2944 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2945 int forwrite, int previous)
2948 struct r5conf *conf = sh->raid_conf;
2951 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2952 (unsigned long long)bi->bi_iter.bi_sector,
2953 (unsigned long long)sh->sector);
2956 * If several bio share a stripe. The bio bi_phys_segments acts as a
2957 * reference count to avoid race. The reference count should already be
2958 * increased before this function is called (for example, in
2959 * make_request()), so other bio sharing this stripe will not free the
2960 * stripe. If a stripe is owned by one stripe, the stripe lock will
2963 spin_lock_irq(&sh->stripe_lock);
2964 /* Don't allow new IO added to stripes in batch list */
2968 bip = &sh->dev[dd_idx].towrite;
2972 bip = &sh->dev[dd_idx].toread;
2973 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2974 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2976 bip = & (*bip)->bi_next;
2978 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2981 if (!forwrite || previous)
2982 clear_bit(STRIPE_BATCH_READY, &sh->state);
2984 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2988 raid5_inc_bi_active_stripes(bi);
2991 /* check if page is covered */
2992 sector_t sector = sh->dev[dd_idx].sector;
2993 for (bi=sh->dev[dd_idx].towrite;
2994 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2995 bi && bi->bi_iter.bi_sector <= sector;
2996 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2997 if (bio_end_sector(bi) >= sector)
2998 sector = bio_end_sector(bi);
3000 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3001 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3002 sh->overwrite_disks++;
3005 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3006 (unsigned long long)(*bip)->bi_iter.bi_sector,
3007 (unsigned long long)sh->sector, dd_idx);
3009 if (conf->mddev->bitmap && firstwrite) {
3010 /* Cannot hold spinlock over bitmap_startwrite,
3011 * but must ensure this isn't added to a batch until
3012 * we have added to the bitmap and set bm_seq.
3013 * So set STRIPE_BITMAP_PENDING to prevent
3015 * If multiple add_stripe_bio() calls race here they
3016 * much all set STRIPE_BITMAP_PENDING. So only the first one
3017 * to complete "bitmap_startwrite" gets to set
3018 * STRIPE_BIT_DELAY. This is important as once a stripe
3019 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3022 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3023 spin_unlock_irq(&sh->stripe_lock);
3024 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3026 spin_lock_irq(&sh->stripe_lock);
3027 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3028 if (!sh->batch_head) {
3029 sh->bm_seq = conf->seq_flush+1;
3030 set_bit(STRIPE_BIT_DELAY, &sh->state);
3033 spin_unlock_irq(&sh->stripe_lock);
3035 if (stripe_can_batch(sh))
3036 stripe_add_to_batch_list(conf, sh);
3040 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3041 spin_unlock_irq(&sh->stripe_lock);
3045 static void end_reshape(struct r5conf *conf);
3047 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3048 struct stripe_head *sh)
3050 int sectors_per_chunk =
3051 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3053 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3054 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3056 raid5_compute_sector(conf,
3057 stripe * (disks - conf->max_degraded)
3058 *sectors_per_chunk + chunk_offset,
3064 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3065 struct stripe_head_state *s, int disks,
3066 struct bio_list *return_bi)
3069 BUG_ON(sh->batch_head);
3070 for (i = disks; i--; ) {
3074 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3075 struct md_rdev *rdev;
3077 rdev = rcu_dereference(conf->disks[i].rdev);
3078 if (rdev && test_bit(In_sync, &rdev->flags))
3079 atomic_inc(&rdev->nr_pending);
3084 if (!rdev_set_badblocks(
3088 md_error(conf->mddev, rdev);
3089 rdev_dec_pending(rdev, conf->mddev);
3092 spin_lock_irq(&sh->stripe_lock);
3093 /* fail all writes first */
3094 bi = sh->dev[i].towrite;
3095 sh->dev[i].towrite = NULL;
3096 sh->overwrite_disks = 0;
3097 spin_unlock_irq(&sh->stripe_lock);
3101 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3102 wake_up(&conf->wait_for_overlap);
3104 while (bi && bi->bi_iter.bi_sector <
3105 sh->dev[i].sector + STRIPE_SECTORS) {
3106 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3108 bi->bi_error = -EIO;
3109 if (!raid5_dec_bi_active_stripes(bi)) {
3110 md_write_end(conf->mddev);
3111 bio_list_add(return_bi, bi);
3116 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3117 STRIPE_SECTORS, 0, 0);
3119 /* and fail all 'written' */
3120 bi = sh->dev[i].written;
3121 sh->dev[i].written = NULL;
3122 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3123 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3124 sh->dev[i].page = sh->dev[i].orig_page;
3127 if (bi) bitmap_end = 1;
3128 while (bi && bi->bi_iter.bi_sector <
3129 sh->dev[i].sector + STRIPE_SECTORS) {
3130 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3132 bi->bi_error = -EIO;
3133 if (!raid5_dec_bi_active_stripes(bi)) {
3134 md_write_end(conf->mddev);
3135 bio_list_add(return_bi, bi);
3140 /* fail any reads if this device is non-operational and
3141 * the data has not reached the cache yet.
3143 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3144 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3145 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3146 spin_lock_irq(&sh->stripe_lock);
3147 bi = sh->dev[i].toread;
3148 sh->dev[i].toread = NULL;
3149 spin_unlock_irq(&sh->stripe_lock);
3150 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3151 wake_up(&conf->wait_for_overlap);
3154 while (bi && bi->bi_iter.bi_sector <
3155 sh->dev[i].sector + STRIPE_SECTORS) {
3156 struct bio *nextbi =
3157 r5_next_bio(bi, sh->dev[i].sector);
3159 bi->bi_error = -EIO;
3160 if (!raid5_dec_bi_active_stripes(bi))
3161 bio_list_add(return_bi, bi);
3166 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3167 STRIPE_SECTORS, 0, 0);
3168 /* If we were in the middle of a write the parity block might
3169 * still be locked - so just clear all R5_LOCKED flags
3171 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3176 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3177 if (atomic_dec_and_test(&conf->pending_full_writes))
3178 md_wakeup_thread(conf->mddev->thread);
3182 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3183 struct stripe_head_state *s)
3188 BUG_ON(sh->batch_head);
3189 clear_bit(STRIPE_SYNCING, &sh->state);
3190 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3191 wake_up(&conf->wait_for_overlap);
3194 /* There is nothing more to do for sync/check/repair.
3195 * Don't even need to abort as that is handled elsewhere
3196 * if needed, and not always wanted e.g. if there is a known
3198 * For recover/replace we need to record a bad block on all
3199 * non-sync devices, or abort the recovery
3201 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3202 /* During recovery devices cannot be removed, so
3203 * locking and refcounting of rdevs is not needed
3205 for (i = 0; i < conf->raid_disks; i++) {
3206 struct md_rdev *rdev = conf->disks[i].rdev;
3208 && !test_bit(Faulty, &rdev->flags)
3209 && !test_bit(In_sync, &rdev->flags)
3210 && !rdev_set_badblocks(rdev, sh->sector,
3213 rdev = conf->disks[i].replacement;
3215 && !test_bit(Faulty, &rdev->flags)
3216 && !test_bit(In_sync, &rdev->flags)
3217 && !rdev_set_badblocks(rdev, sh->sector,
3222 conf->recovery_disabled =
3223 conf->mddev->recovery_disabled;
3225 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3228 static int want_replace(struct stripe_head *sh, int disk_idx)
3230 struct md_rdev *rdev;
3232 /* Doing recovery so rcu locking not required */
3233 rdev = sh->raid_conf->disks[disk_idx].replacement;
3235 && !test_bit(Faulty, &rdev->flags)
3236 && !test_bit(In_sync, &rdev->flags)
3237 && (rdev->recovery_offset <= sh->sector
3238 || rdev->mddev->recovery_cp <= sh->sector))
3244 /* fetch_block - checks the given member device to see if its data needs
3245 * to be read or computed to satisfy a request.
3247 * Returns 1 when no more member devices need to be checked, otherwise returns
3248 * 0 to tell the loop in handle_stripe_fill to continue
3251 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3252 int disk_idx, int disks)
3254 struct r5dev *dev = &sh->dev[disk_idx];
3255 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3256 &sh->dev[s->failed_num[1]] };
3260 if (test_bit(R5_LOCKED, &dev->flags) ||
3261 test_bit(R5_UPTODATE, &dev->flags))
3262 /* No point reading this as we already have it or have
3263 * decided to get it.
3268 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3269 /* We need this block to directly satisfy a request */
3272 if (s->syncing || s->expanding ||
3273 (s->replacing && want_replace(sh, disk_idx)))
3274 /* When syncing, or expanding we read everything.
3275 * When replacing, we need the replaced block.
3279 if ((s->failed >= 1 && fdev[0]->toread) ||
3280 (s->failed >= 2 && fdev[1]->toread))
3281 /* If we want to read from a failed device, then
3282 * we need to actually read every other device.
3286 /* Sometimes neither read-modify-write nor reconstruct-write
3287 * cycles can work. In those cases we read every block we
3288 * can. Then the parity-update is certain to have enough to
3290 * This can only be a problem when we need to write something,
3291 * and some device has failed. If either of those tests
3292 * fail we need look no further.
3294 if (!s->failed || !s->to_write)
3297 if (test_bit(R5_Insync, &dev->flags) &&
3298 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3299 /* Pre-reads at not permitted until after short delay
3300 * to gather multiple requests. However if this
3301 * device is no Insync, the block could only be be computed
3302 * and there is no need to delay that.
3306 for (i = 0; i < s->failed && i < 2; i++) {
3307 if (fdev[i]->towrite &&
3308 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3309 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3310 /* If we have a partial write to a failed
3311 * device, then we will need to reconstruct
3312 * the content of that device, so all other
3313 * devices must be read.
3318 /* If we are forced to do a reconstruct-write, either because
3319 * the current RAID6 implementation only supports that, or
3320 * or because parity cannot be trusted and we are currently
3321 * recovering it, there is extra need to be careful.
3322 * If one of the devices that we would need to read, because
3323 * it is not being overwritten (and maybe not written at all)
3324 * is missing/faulty, then we need to read everything we can.
3326 if (sh->raid_conf->level != 6 &&
3327 sh->sector < sh->raid_conf->mddev->recovery_cp)
3328 /* reconstruct-write isn't being forced */
3330 for (i = 0; i < s->failed && i < 2; i++) {
3331 if (s->failed_num[i] != sh->pd_idx &&
3332 s->failed_num[i] != sh->qd_idx &&
3333 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3334 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3341 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3342 int disk_idx, int disks)
3344 struct r5dev *dev = &sh->dev[disk_idx];
3346 /* is the data in this block needed, and can we get it? */
3347 if (need_this_block(sh, s, disk_idx, disks)) {
3348 /* we would like to get this block, possibly by computing it,
3349 * otherwise read it if the backing disk is insync
3351 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3352 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3353 BUG_ON(sh->batch_head);
3354 if ((s->uptodate == disks - 1) &&
3355 (s->failed && (disk_idx == s->failed_num[0] ||
3356 disk_idx == s->failed_num[1]))) {
3357 /* have disk failed, and we're requested to fetch it;
3360 pr_debug("Computing stripe %llu block %d\n",
3361 (unsigned long long)sh->sector, disk_idx);
3362 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3363 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3364 set_bit(R5_Wantcompute, &dev->flags);
3365 sh->ops.target = disk_idx;
3366 sh->ops.target2 = -1; /* no 2nd target */
3368 /* Careful: from this point on 'uptodate' is in the eye
3369 * of raid_run_ops which services 'compute' operations
3370 * before writes. R5_Wantcompute flags a block that will
3371 * be R5_UPTODATE by the time it is needed for a
3372 * subsequent operation.
3376 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3377 /* Computing 2-failure is *very* expensive; only
3378 * do it if failed >= 2
3381 for (other = disks; other--; ) {
3382 if (other == disk_idx)
3384 if (!test_bit(R5_UPTODATE,
3385 &sh->dev[other].flags))
3389 pr_debug("Computing stripe %llu blocks %d,%d\n",
3390 (unsigned long long)sh->sector,
3392 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3393 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3394 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3395 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3396 sh->ops.target = disk_idx;
3397 sh->ops.target2 = other;
3401 } else if (test_bit(R5_Insync, &dev->flags)) {
3402 set_bit(R5_LOCKED, &dev->flags);
3403 set_bit(R5_Wantread, &dev->flags);
3405 pr_debug("Reading block %d (sync=%d)\n",
3406 disk_idx, s->syncing);
3414 * handle_stripe_fill - read or compute data to satisfy pending requests.
3416 static void handle_stripe_fill(struct stripe_head *sh,
3417 struct stripe_head_state *s,
3422 /* look for blocks to read/compute, skip this if a compute
3423 * is already in flight, or if the stripe contents are in the
3424 * midst of changing due to a write
3426 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3427 !sh->reconstruct_state)
3428 for (i = disks; i--; )
3429 if (fetch_block(sh, s, i, disks))
3431 set_bit(STRIPE_HANDLE, &sh->state);
3434 static void break_stripe_batch_list(struct stripe_head *head_sh,
3435 unsigned long handle_flags);
3436 /* handle_stripe_clean_event
3437 * any written block on an uptodate or failed drive can be returned.
3438 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3439 * never LOCKED, so we don't need to test 'failed' directly.
3441 static void handle_stripe_clean_event(struct r5conf *conf,
3442 struct stripe_head *sh, int disks, struct bio_list *return_bi)
3446 int discard_pending = 0;
3447 struct stripe_head *head_sh = sh;
3448 bool do_endio = false;
3450 for (i = disks; i--; )
3451 if (sh->dev[i].written) {
3453 if (!test_bit(R5_LOCKED, &dev->flags) &&
3454 (test_bit(R5_UPTODATE, &dev->flags) ||
3455 test_bit(R5_Discard, &dev->flags) ||
3456 test_bit(R5_SkipCopy, &dev->flags))) {
3457 /* We can return any write requests */
3458 struct bio *wbi, *wbi2;
3459 pr_debug("Return write for disc %d\n", i);
3460 if (test_and_clear_bit(R5_Discard, &dev->flags))
3461 clear_bit(R5_UPTODATE, &dev->flags);
3462 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3463 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3468 dev->page = dev->orig_page;
3470 dev->written = NULL;
3471 while (wbi && wbi->bi_iter.bi_sector <
3472 dev->sector + STRIPE_SECTORS) {
3473 wbi2 = r5_next_bio(wbi, dev->sector);
3474 if (!raid5_dec_bi_active_stripes(wbi)) {
3475 md_write_end(conf->mddev);
3476 bio_list_add(return_bi, wbi);
3480 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3482 !test_bit(STRIPE_DEGRADED, &sh->state),
3484 if (head_sh->batch_head) {
3485 sh = list_first_entry(&sh->batch_list,
3488 if (sh != head_sh) {
3495 } else if (test_bit(R5_Discard, &dev->flags))
3496 discard_pending = 1;
3497 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3498 WARN_ON(dev->page != dev->orig_page);
3500 if (!discard_pending &&
3501 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3502 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3503 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3504 if (sh->qd_idx >= 0) {
3505 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3506 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3508 /* now that discard is done we can proceed with any sync */
3509 clear_bit(STRIPE_DISCARD, &sh->state);
3511 * SCSI discard will change some bio fields and the stripe has
3512 * no updated data, so remove it from hash list and the stripe
3513 * will be reinitialized
3515 spin_lock_irq(&conf->device_lock);
3518 if (head_sh->batch_head) {
3519 sh = list_first_entry(&sh->batch_list,
3520 struct stripe_head, batch_list);
3524 spin_unlock_irq(&conf->device_lock);
3527 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3528 set_bit(STRIPE_HANDLE, &sh->state);
3532 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3533 if (atomic_dec_and_test(&conf->pending_full_writes))
3534 md_wakeup_thread(conf->mddev->thread);
3536 if (head_sh->batch_head && do_endio)
3537 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3540 static void handle_stripe_dirtying(struct r5conf *conf,
3541 struct stripe_head *sh,
3542 struct stripe_head_state *s,
3545 int rmw = 0, rcw = 0, i;
3546 sector_t recovery_cp = conf->mddev->recovery_cp;
3548 /* Check whether resync is now happening or should start.
3549 * If yes, then the array is dirty (after unclean shutdown or
3550 * initial creation), so parity in some stripes might be inconsistent.
3551 * In this case, we need to always do reconstruct-write, to ensure
3552 * that in case of drive failure or read-error correction, we
3553 * generate correct data from the parity.
3555 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3556 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3558 /* Calculate the real rcw later - for now make it
3559 * look like rcw is cheaper
3562 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3563 conf->rmw_level, (unsigned long long)recovery_cp,
3564 (unsigned long long)sh->sector);
3565 } else for (i = disks; i--; ) {
3566 /* would I have to read this buffer for read_modify_write */
3567 struct r5dev *dev = &sh->dev[i];
3568 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3569 !test_bit(R5_LOCKED, &dev->flags) &&
3570 !(test_bit(R5_UPTODATE, &dev->flags) ||
3571 test_bit(R5_Wantcompute, &dev->flags))) {
3572 if (test_bit(R5_Insync, &dev->flags))
3575 rmw += 2*disks; /* cannot read it */
3577 /* Would I have to read this buffer for reconstruct_write */
3578 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3579 i != sh->pd_idx && i != sh->qd_idx &&
3580 !test_bit(R5_LOCKED, &dev->flags) &&
3581 !(test_bit(R5_UPTODATE, &dev->flags) ||
3582 test_bit(R5_Wantcompute, &dev->flags))) {
3583 if (test_bit(R5_Insync, &dev->flags))
3589 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3590 (unsigned long long)sh->sector, rmw, rcw);
3591 set_bit(STRIPE_HANDLE, &sh->state);
3592 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3593 /* prefer read-modify-write, but need to get some data */
3594 if (conf->mddev->queue)
3595 blk_add_trace_msg(conf->mddev->queue,
3596 "raid5 rmw %llu %d",
3597 (unsigned long long)sh->sector, rmw);
3598 for (i = disks; i--; ) {
3599 struct r5dev *dev = &sh->dev[i];
3600 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3601 !test_bit(R5_LOCKED, &dev->flags) &&
3602 !(test_bit(R5_UPTODATE, &dev->flags) ||
3603 test_bit(R5_Wantcompute, &dev->flags)) &&
3604 test_bit(R5_Insync, &dev->flags)) {
3605 if (test_bit(STRIPE_PREREAD_ACTIVE,
3607 pr_debug("Read_old block %d for r-m-w\n",
3609 set_bit(R5_LOCKED, &dev->flags);
3610 set_bit(R5_Wantread, &dev->flags);
3613 set_bit(STRIPE_DELAYED, &sh->state);
3614 set_bit(STRIPE_HANDLE, &sh->state);
3619 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3620 /* want reconstruct write, but need to get some data */
3623 for (i = disks; i--; ) {
3624 struct r5dev *dev = &sh->dev[i];
3625 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3626 i != sh->pd_idx && i != sh->qd_idx &&
3627 !test_bit(R5_LOCKED, &dev->flags) &&
3628 !(test_bit(R5_UPTODATE, &dev->flags) ||
3629 test_bit(R5_Wantcompute, &dev->flags))) {
3631 if (test_bit(R5_Insync, &dev->flags) &&
3632 test_bit(STRIPE_PREREAD_ACTIVE,
3634 pr_debug("Read_old block "
3635 "%d for Reconstruct\n", i);
3636 set_bit(R5_LOCKED, &dev->flags);
3637 set_bit(R5_Wantread, &dev->flags);
3641 set_bit(STRIPE_DELAYED, &sh->state);
3642 set_bit(STRIPE_HANDLE, &sh->state);
3646 if (rcw && conf->mddev->queue)
3647 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3648 (unsigned long long)sh->sector,
3649 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3652 if (rcw > disks && rmw > disks &&
3653 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3654 set_bit(STRIPE_DELAYED, &sh->state);
3656 /* now if nothing is locked, and if we have enough data,
3657 * we can start a write request
3659 /* since handle_stripe can be called at any time we need to handle the
3660 * case where a compute block operation has been submitted and then a
3661 * subsequent call wants to start a write request. raid_run_ops only
3662 * handles the case where compute block and reconstruct are requested
3663 * simultaneously. If this is not the case then new writes need to be
3664 * held off until the compute completes.
3666 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3667 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3668 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3669 schedule_reconstruction(sh, s, rcw == 0, 0);
3672 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3673 struct stripe_head_state *s, int disks)
3675 struct r5dev *dev = NULL;
3677 BUG_ON(sh->batch_head);
3678 set_bit(STRIPE_HANDLE, &sh->state);
3680 switch (sh->check_state) {
3681 case check_state_idle:
3682 /* start a new check operation if there are no failures */
3683 if (s->failed == 0) {
3684 BUG_ON(s->uptodate != disks);
3685 sh->check_state = check_state_run;
3686 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3687 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3691 dev = &sh->dev[s->failed_num[0]];
3693 case check_state_compute_result:
3694 sh->check_state = check_state_idle;
3696 dev = &sh->dev[sh->pd_idx];
3698 /* check that a write has not made the stripe insync */
3699 if (test_bit(STRIPE_INSYNC, &sh->state))
3702 /* either failed parity check, or recovery is happening */
3703 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3704 BUG_ON(s->uptodate != disks);
3706 set_bit(R5_LOCKED, &dev->flags);
3708 set_bit(R5_Wantwrite, &dev->flags);
3710 clear_bit(STRIPE_DEGRADED, &sh->state);
3711 set_bit(STRIPE_INSYNC, &sh->state);
3713 case check_state_run:
3714 break; /* we will be called again upon completion */
3715 case check_state_check_result:
3716 sh->check_state = check_state_idle;
3718 /* if a failure occurred during the check operation, leave
3719 * STRIPE_INSYNC not set and let the stripe be handled again
3724 /* handle a successful check operation, if parity is correct
3725 * we are done. Otherwise update the mismatch count and repair
3726 * parity if !MD_RECOVERY_CHECK
3728 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3729 /* parity is correct (on disc,
3730 * not in buffer any more)
3732 set_bit(STRIPE_INSYNC, &sh->state);
3734 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3735 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3736 /* don't try to repair!! */
3737 set_bit(STRIPE_INSYNC, &sh->state);
3739 sh->check_state = check_state_compute_run;
3740 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3741 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3742 set_bit(R5_Wantcompute,
3743 &sh->dev[sh->pd_idx].flags);
3744 sh->ops.target = sh->pd_idx;
3745 sh->ops.target2 = -1;
3750 case check_state_compute_run:
3753 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3754 __func__, sh->check_state,
3755 (unsigned long long) sh->sector);
3760 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3761 struct stripe_head_state *s,
3764 int pd_idx = sh->pd_idx;
3765 int qd_idx = sh->qd_idx;
3768 BUG_ON(sh->batch_head);
3769 set_bit(STRIPE_HANDLE, &sh->state);
3771 BUG_ON(s->failed > 2);
3773 /* Want to check and possibly repair P and Q.
3774 * However there could be one 'failed' device, in which
3775 * case we can only check one of them, possibly using the
3776 * other to generate missing data
3779 switch (sh->check_state) {
3780 case check_state_idle:
3781 /* start a new check operation if there are < 2 failures */
3782 if (s->failed == s->q_failed) {
3783 /* The only possible failed device holds Q, so it
3784 * makes sense to check P (If anything else were failed,
3785 * we would have used P to recreate it).
3787 sh->check_state = check_state_run;
3789 if (!s->q_failed && s->failed < 2) {
3790 /* Q is not failed, and we didn't use it to generate
3791 * anything, so it makes sense to check it
3793 if (sh->check_state == check_state_run)
3794 sh->check_state = check_state_run_pq;
3796 sh->check_state = check_state_run_q;
3799 /* discard potentially stale zero_sum_result */
3800 sh->ops.zero_sum_result = 0;
3802 if (sh->check_state == check_state_run) {
3803 /* async_xor_zero_sum destroys the contents of P */
3804 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3807 if (sh->check_state >= check_state_run &&
3808 sh->check_state <= check_state_run_pq) {
3809 /* async_syndrome_zero_sum preserves P and Q, so
3810 * no need to mark them !uptodate here
3812 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3816 /* we have 2-disk failure */
3817 BUG_ON(s->failed != 2);
3819 case check_state_compute_result:
3820 sh->check_state = check_state_idle;
3822 /* check that a write has not made the stripe insync */
3823 if (test_bit(STRIPE_INSYNC, &sh->state))
3826 /* now write out any block on a failed drive,
3827 * or P or Q if they were recomputed
3829 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3830 if (s->failed == 2) {
3831 dev = &sh->dev[s->failed_num[1]];
3833 set_bit(R5_LOCKED, &dev->flags);
3834 set_bit(R5_Wantwrite, &dev->flags);
3836 if (s->failed >= 1) {
3837 dev = &sh->dev[s->failed_num[0]];
3839 set_bit(R5_LOCKED, &dev->flags);
3840 set_bit(R5_Wantwrite, &dev->flags);
3842 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3843 dev = &sh->dev[pd_idx];
3845 set_bit(R5_LOCKED, &dev->flags);
3846 set_bit(R5_Wantwrite, &dev->flags);
3848 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3849 dev = &sh->dev[qd_idx];
3851 set_bit(R5_LOCKED, &dev->flags);
3852 set_bit(R5_Wantwrite, &dev->flags);
3854 clear_bit(STRIPE_DEGRADED, &sh->state);
3856 set_bit(STRIPE_INSYNC, &sh->state);
3858 case check_state_run:
3859 case check_state_run_q:
3860 case check_state_run_pq:
3861 break; /* we will be called again upon completion */
3862 case check_state_check_result:
3863 sh->check_state = check_state_idle;
3865 /* handle a successful check operation, if parity is correct
3866 * we are done. Otherwise update the mismatch count and repair
3867 * parity if !MD_RECOVERY_CHECK
3869 if (sh->ops.zero_sum_result == 0) {
3870 /* both parities are correct */
3872 set_bit(STRIPE_INSYNC, &sh->state);
3874 /* in contrast to the raid5 case we can validate
3875 * parity, but still have a failure to write
3878 sh->check_state = check_state_compute_result;
3879 /* Returning at this point means that we may go
3880 * off and bring p and/or q uptodate again so
3881 * we make sure to check zero_sum_result again
3882 * to verify if p or q need writeback
3886 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3887 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3888 /* don't try to repair!! */
3889 set_bit(STRIPE_INSYNC, &sh->state);
3891 int *target = &sh->ops.target;
3893 sh->ops.target = -1;
3894 sh->ops.target2 = -1;
3895 sh->check_state = check_state_compute_run;
3896 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3897 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3898 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3899 set_bit(R5_Wantcompute,
3900 &sh->dev[pd_idx].flags);
3902 target = &sh->ops.target2;
3905 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3906 set_bit(R5_Wantcompute,
3907 &sh->dev[qd_idx].flags);
3914 case check_state_compute_run:
3917 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3918 __func__, sh->check_state,
3919 (unsigned long long) sh->sector);
3924 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3928 /* We have read all the blocks in this stripe and now we need to
3929 * copy some of them into a target stripe for expand.
3931 struct dma_async_tx_descriptor *tx = NULL;
3932 BUG_ON(sh->batch_head);
3933 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3934 for (i = 0; i < sh->disks; i++)
3935 if (i != sh->pd_idx && i != sh->qd_idx) {
3937 struct stripe_head *sh2;
3938 struct async_submit_ctl submit;
3940 sector_t bn = compute_blocknr(sh, i, 1);
3941 sector_t s = raid5_compute_sector(conf, bn, 0,
3943 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3945 /* so far only the early blocks of this stripe
3946 * have been requested. When later blocks
3947 * get requested, we will try again
3950 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3951 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3952 /* must have already done this block */
3953 release_stripe(sh2);
3957 /* place all the copies on one channel */
3958 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3959 tx = async_memcpy(sh2->dev[dd_idx].page,
3960 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3963 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3964 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3965 for (j = 0; j < conf->raid_disks; j++)
3966 if (j != sh2->pd_idx &&
3968 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3970 if (j == conf->raid_disks) {
3971 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3972 set_bit(STRIPE_HANDLE, &sh2->state);
3974 release_stripe(sh2);
3977 /* done submitting copies, wait for them to complete */
3978 async_tx_quiesce(&tx);
3982 * handle_stripe - do things to a stripe.
3984 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3985 * state of various bits to see what needs to be done.
3987 * return some read requests which now have data
3988 * return some write requests which are safely on storage
3989 * schedule a read on some buffers
3990 * schedule a write of some buffers
3991 * return confirmation of parity correctness
3995 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3997 struct r5conf *conf = sh->raid_conf;
3998 int disks = sh->disks;
4001 int do_recovery = 0;
4003 memset(s, 0, sizeof(*s));
4005 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4006 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4007 s->failed_num[0] = -1;
4008 s->failed_num[1] = -1;
4010 /* Now to look around and see what can be done */
4012 for (i=disks; i--; ) {
4013 struct md_rdev *rdev;
4020 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4022 dev->toread, dev->towrite, dev->written);
4023 /* maybe we can reply to a read
4025 * new wantfill requests are only permitted while
4026 * ops_complete_biofill is guaranteed to be inactive
4028 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4029 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4030 set_bit(R5_Wantfill, &dev->flags);
4032 /* now count some things */
4033 if (test_bit(R5_LOCKED, &dev->flags))
4035 if (test_bit(R5_UPTODATE, &dev->flags))
4037 if (test_bit(R5_Wantcompute, &dev->flags)) {
4039 BUG_ON(s->compute > 2);
4042 if (test_bit(R5_Wantfill, &dev->flags))
4044 else if (dev->toread)
4048 if (!test_bit(R5_OVERWRITE, &dev->flags))
4053 /* Prefer to use the replacement for reads, but only
4054 * if it is recovered enough and has no bad blocks.
4056 rdev = rcu_dereference(conf->disks[i].replacement);
4057 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4058 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4059 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4060 &first_bad, &bad_sectors))
4061 set_bit(R5_ReadRepl, &dev->flags);
4063 if (rdev && !test_bit(Faulty, &rdev->flags))
4064 set_bit(R5_NeedReplace, &dev->flags);
4066 clear_bit(R5_NeedReplace, &dev->flags);
4067 rdev = rcu_dereference(conf->disks[i].rdev);
4068 clear_bit(R5_ReadRepl, &dev->flags);
4070 if (rdev && test_bit(Faulty, &rdev->flags))
4073 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4074 &first_bad, &bad_sectors);
4075 if (s->blocked_rdev == NULL
4076 && (test_bit(Blocked, &rdev->flags)
4079 set_bit(BlockedBadBlocks,
4081 s->blocked_rdev = rdev;
4082 atomic_inc(&rdev->nr_pending);
4085 clear_bit(R5_Insync, &dev->flags);
4089 /* also not in-sync */
4090 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4091 test_bit(R5_UPTODATE, &dev->flags)) {
4092 /* treat as in-sync, but with a read error
4093 * which we can now try to correct
4095 set_bit(R5_Insync, &dev->flags);
4096 set_bit(R5_ReadError, &dev->flags);
4098 } else if (test_bit(In_sync, &rdev->flags))
4099 set_bit(R5_Insync, &dev->flags);
4100 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4101 /* in sync if before recovery_offset */
4102 set_bit(R5_Insync, &dev->flags);
4103 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4104 test_bit(R5_Expanded, &dev->flags))
4105 /* If we've reshaped into here, we assume it is Insync.
4106 * We will shortly update recovery_offset to make
4109 set_bit(R5_Insync, &dev->flags);
4111 if (test_bit(R5_WriteError, &dev->flags)) {
4112 /* This flag does not apply to '.replacement'
4113 * only to .rdev, so make sure to check that*/
4114 struct md_rdev *rdev2 = rcu_dereference(
4115 conf->disks[i].rdev);
4117 clear_bit(R5_Insync, &dev->flags);
4118 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4119 s->handle_bad_blocks = 1;
4120 atomic_inc(&rdev2->nr_pending);
4122 clear_bit(R5_WriteError, &dev->flags);
4124 if (test_bit(R5_MadeGood, &dev->flags)) {
4125 /* This flag does not apply to '.replacement'
4126 * only to .rdev, so make sure to check that*/
4127 struct md_rdev *rdev2 = rcu_dereference(
4128 conf->disks[i].rdev);
4129 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4130 s->handle_bad_blocks = 1;
4131 atomic_inc(&rdev2->nr_pending);
4133 clear_bit(R5_MadeGood, &dev->flags);
4135 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4136 struct md_rdev *rdev2 = rcu_dereference(
4137 conf->disks[i].replacement);
4138 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4139 s->handle_bad_blocks = 1;
4140 atomic_inc(&rdev2->nr_pending);
4142 clear_bit(R5_MadeGoodRepl, &dev->flags);
4144 if (!test_bit(R5_Insync, &dev->flags)) {
4145 /* The ReadError flag will just be confusing now */
4146 clear_bit(R5_ReadError, &dev->flags);
4147 clear_bit(R5_ReWrite, &dev->flags);
4149 if (test_bit(R5_ReadError, &dev->flags))
4150 clear_bit(R5_Insync, &dev->flags);
4151 if (!test_bit(R5_Insync, &dev->flags)) {
4153 s->failed_num[s->failed] = i;
4155 if (rdev && !test_bit(Faulty, &rdev->flags))
4159 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4160 /* If there is a failed device being replaced,
4161 * we must be recovering.
4162 * else if we are after recovery_cp, we must be syncing
4163 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4164 * else we can only be replacing
4165 * sync and recovery both need to read all devices, and so
4166 * use the same flag.
4169 sh->sector >= conf->mddev->recovery_cp ||
4170 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4178 static int clear_batch_ready(struct stripe_head *sh)
4180 /* Return '1' if this is a member of batch, or
4181 * '0' if it is a lone stripe or a head which can now be
4184 struct stripe_head *tmp;
4185 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4186 return (sh->batch_head && sh->batch_head != sh);
4187 spin_lock(&sh->stripe_lock);
4188 if (!sh->batch_head) {
4189 spin_unlock(&sh->stripe_lock);
4194 * this stripe could be added to a batch list before we check
4195 * BATCH_READY, skips it
4197 if (sh->batch_head != sh) {
4198 spin_unlock(&sh->stripe_lock);
4201 spin_lock(&sh->batch_lock);
4202 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4203 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4204 spin_unlock(&sh->batch_lock);
4205 spin_unlock(&sh->stripe_lock);
4208 * BATCH_READY is cleared, no new stripes can be added.
4209 * batch_list can be accessed without lock
4214 static void break_stripe_batch_list(struct stripe_head *head_sh,
4215 unsigned long handle_flags)
4217 struct stripe_head *sh, *next;
4221 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4223 list_del_init(&sh->batch_list);
4225 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4226 (1 << STRIPE_SYNCING) |
4227 (1 << STRIPE_REPLACED) |
4228 (1 << STRIPE_PREREAD_ACTIVE) |
4229 (1 << STRIPE_DELAYED) |
4230 (1 << STRIPE_BIT_DELAY) |
4231 (1 << STRIPE_FULL_WRITE) |
4232 (1 << STRIPE_BIOFILL_RUN) |
4233 (1 << STRIPE_COMPUTE_RUN) |
4234 (1 << STRIPE_OPS_REQ_PENDING) |
4235 (1 << STRIPE_DISCARD) |
4236 (1 << STRIPE_BATCH_READY) |
4237 (1 << STRIPE_BATCH_ERR) |
4238 (1 << STRIPE_BITMAP_PENDING)));
4239 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4240 (1 << STRIPE_REPLACED)));
4242 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4243 (1 << STRIPE_DEGRADED)),
4244 head_sh->state & (1 << STRIPE_INSYNC));
4246 sh->check_state = head_sh->check_state;
4247 sh->reconstruct_state = head_sh->reconstruct_state;
4248 for (i = 0; i < sh->disks; i++) {
4249 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4251 sh->dev[i].flags = head_sh->dev[i].flags &
4252 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4254 spin_lock_irq(&sh->stripe_lock);
4255 sh->batch_head = NULL;
4256 spin_unlock_irq(&sh->stripe_lock);
4257 if (handle_flags == 0 ||
4258 sh->state & handle_flags)
4259 set_bit(STRIPE_HANDLE, &sh->state);
4262 spin_lock_irq(&head_sh->stripe_lock);
4263 head_sh->batch_head = NULL;
4264 spin_unlock_irq(&head_sh->stripe_lock);
4265 for (i = 0; i < head_sh->disks; i++)
4266 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4268 if (head_sh->state & handle_flags)
4269 set_bit(STRIPE_HANDLE, &head_sh->state);
4272 wake_up(&head_sh->raid_conf->wait_for_overlap);
4275 static void handle_stripe(struct stripe_head *sh)
4277 struct stripe_head_state s;
4278 struct r5conf *conf = sh->raid_conf;
4281 int disks = sh->disks;
4282 struct r5dev *pdev, *qdev;
4284 clear_bit(STRIPE_HANDLE, &sh->state);
4285 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4286 /* already being handled, ensure it gets handled
4287 * again when current action finishes */
4288 set_bit(STRIPE_HANDLE, &sh->state);
4292 if (clear_batch_ready(sh) ) {
4293 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4297 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4298 break_stripe_batch_list(sh, 0);
4300 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4301 spin_lock(&sh->stripe_lock);
4302 /* Cannot process 'sync' concurrently with 'discard' */
4303 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4304 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4305 set_bit(STRIPE_SYNCING, &sh->state);
4306 clear_bit(STRIPE_INSYNC, &sh->state);
4307 clear_bit(STRIPE_REPLACED, &sh->state);
4309 spin_unlock(&sh->stripe_lock);
4311 clear_bit(STRIPE_DELAYED, &sh->state);
4313 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4314 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4315 (unsigned long long)sh->sector, sh->state,
4316 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4317 sh->check_state, sh->reconstruct_state);
4319 analyse_stripe(sh, &s);
4321 if (s.handle_bad_blocks) {
4322 set_bit(STRIPE_HANDLE, &sh->state);
4326 if (unlikely(s.blocked_rdev)) {
4327 if (s.syncing || s.expanding || s.expanded ||
4328 s.replacing || s.to_write || s.written) {
4329 set_bit(STRIPE_HANDLE, &sh->state);
4332 /* There is nothing for the blocked_rdev to block */
4333 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4334 s.blocked_rdev = NULL;
4337 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4338 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4339 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4342 pr_debug("locked=%d uptodate=%d to_read=%d"
4343 " to_write=%d failed=%d failed_num=%d,%d\n",
4344 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4345 s.failed_num[0], s.failed_num[1]);
4346 /* check if the array has lost more than max_degraded devices and,
4347 * if so, some requests might need to be failed.
4349 if (s.failed > conf->max_degraded) {
4350 sh->check_state = 0;
4351 sh->reconstruct_state = 0;
4352 break_stripe_batch_list(sh, 0);
4353 if (s.to_read+s.to_write+s.written)
4354 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4355 if (s.syncing + s.replacing)
4356 handle_failed_sync(conf, sh, &s);
4359 /* Now we check to see if any write operations have recently
4363 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4365 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4366 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4367 sh->reconstruct_state = reconstruct_state_idle;
4369 /* All the 'written' buffers and the parity block are ready to
4370 * be written back to disk
4372 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4373 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4374 BUG_ON(sh->qd_idx >= 0 &&
4375 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4376 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4377 for (i = disks; i--; ) {
4378 struct r5dev *dev = &sh->dev[i];
4379 if (test_bit(R5_LOCKED, &dev->flags) &&
4380 (i == sh->pd_idx || i == sh->qd_idx ||
4382 pr_debug("Writing block %d\n", i);
4383 set_bit(R5_Wantwrite, &dev->flags);
4388 if (!test_bit(R5_Insync, &dev->flags) ||
4389 ((i == sh->pd_idx || i == sh->qd_idx) &&
4391 set_bit(STRIPE_INSYNC, &sh->state);
4394 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4395 s.dec_preread_active = 1;
4399 * might be able to return some write requests if the parity blocks
4400 * are safe, or on a failed drive
4402 pdev = &sh->dev[sh->pd_idx];
4403 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4404 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4405 qdev = &sh->dev[sh->qd_idx];
4406 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4407 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4411 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4412 && !test_bit(R5_LOCKED, &pdev->flags)
4413 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4414 test_bit(R5_Discard, &pdev->flags))))) &&
4415 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4416 && !test_bit(R5_LOCKED, &qdev->flags)
4417 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4418 test_bit(R5_Discard, &qdev->flags))))))
4419 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4421 /* Now we might consider reading some blocks, either to check/generate
4422 * parity, or to satisfy requests
4423 * or to load a block that is being partially written.
4425 if (s.to_read || s.non_overwrite
4426 || (conf->level == 6 && s.to_write && s.failed)
4427 || (s.syncing && (s.uptodate + s.compute < disks))
4430 handle_stripe_fill(sh, &s, disks);
4432 /* Now to consider new write requests and what else, if anything
4433 * should be read. We do not handle new writes when:
4434 * 1/ A 'write' operation (copy+xor) is already in flight.
4435 * 2/ A 'check' operation is in flight, as it may clobber the parity
4438 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4439 handle_stripe_dirtying(conf, sh, &s, disks);
4441 /* maybe we need to check and possibly fix the parity for this stripe
4442 * Any reads will already have been scheduled, so we just see if enough
4443 * data is available. The parity check is held off while parity
4444 * dependent operations are in flight.
4446 if (sh->check_state ||
4447 (s.syncing && s.locked == 0 &&
4448 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4449 !test_bit(STRIPE_INSYNC, &sh->state))) {
4450 if (conf->level == 6)
4451 handle_parity_checks6(conf, sh, &s, disks);
4453 handle_parity_checks5(conf, sh, &s, disks);
4456 if ((s.replacing || s.syncing) && s.locked == 0
4457 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4458 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4459 /* Write out to replacement devices where possible */
4460 for (i = 0; i < conf->raid_disks; i++)
4461 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4462 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4463 set_bit(R5_WantReplace, &sh->dev[i].flags);
4464 set_bit(R5_LOCKED, &sh->dev[i].flags);
4468 set_bit(STRIPE_INSYNC, &sh->state);
4469 set_bit(STRIPE_REPLACED, &sh->state);
4471 if ((s.syncing || s.replacing) && s.locked == 0 &&
4472 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4473 test_bit(STRIPE_INSYNC, &sh->state)) {
4474 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4475 clear_bit(STRIPE_SYNCING, &sh->state);
4476 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4477 wake_up(&conf->wait_for_overlap);
4480 /* If the failed drives are just a ReadError, then we might need
4481 * to progress the repair/check process
4483 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4484 for (i = 0; i < s.failed; i++) {
4485 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4486 if (test_bit(R5_ReadError, &dev->flags)
4487 && !test_bit(R5_LOCKED, &dev->flags)
4488 && test_bit(R5_UPTODATE, &dev->flags)
4490 if (!test_bit(R5_ReWrite, &dev->flags)) {
4491 set_bit(R5_Wantwrite, &dev->flags);
4492 set_bit(R5_ReWrite, &dev->flags);
4493 set_bit(R5_LOCKED, &dev->flags);
4496 /* let's read it back */
4497 set_bit(R5_Wantread, &dev->flags);
4498 set_bit(R5_LOCKED, &dev->flags);
4504 /* Finish reconstruct operations initiated by the expansion process */
4505 if (sh->reconstruct_state == reconstruct_state_result) {
4506 struct stripe_head *sh_src
4507 = get_active_stripe(conf, sh->sector, 1, 1, 1);
4508 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4509 /* sh cannot be written until sh_src has been read.
4510 * so arrange for sh to be delayed a little
4512 set_bit(STRIPE_DELAYED, &sh->state);
4513 set_bit(STRIPE_HANDLE, &sh->state);
4514 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4516 atomic_inc(&conf->preread_active_stripes);
4517 release_stripe(sh_src);
4521 release_stripe(sh_src);
4523 sh->reconstruct_state = reconstruct_state_idle;
4524 clear_bit(STRIPE_EXPANDING, &sh->state);
4525 for (i = conf->raid_disks; i--; ) {
4526 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4527 set_bit(R5_LOCKED, &sh->dev[i].flags);
4532 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4533 !sh->reconstruct_state) {
4534 /* Need to write out all blocks after computing parity */
4535 sh->disks = conf->raid_disks;
4536 stripe_set_idx(sh->sector, conf, 0, sh);
4537 schedule_reconstruction(sh, &s, 1, 1);
4538 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4539 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4540 atomic_dec(&conf->reshape_stripes);
4541 wake_up(&conf->wait_for_overlap);
4542 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4545 if (s.expanding && s.locked == 0 &&
4546 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4547 handle_stripe_expansion(conf, sh);
4550 /* wait for this device to become unblocked */
4551 if (unlikely(s.blocked_rdev)) {
4552 if (conf->mddev->external)
4553 md_wait_for_blocked_rdev(s.blocked_rdev,
4556 /* Internal metadata will immediately
4557 * be written by raid5d, so we don't
4558 * need to wait here.
4560 rdev_dec_pending(s.blocked_rdev,
4564 if (s.handle_bad_blocks)
4565 for (i = disks; i--; ) {
4566 struct md_rdev *rdev;
4567 struct r5dev *dev = &sh->dev[i];
4568 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4569 /* We own a safe reference to the rdev */
4570 rdev = conf->disks[i].rdev;
4571 if (!rdev_set_badblocks(rdev, sh->sector,
4573 md_error(conf->mddev, rdev);
4574 rdev_dec_pending(rdev, conf->mddev);
4576 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4577 rdev = conf->disks[i].rdev;
4578 rdev_clear_badblocks(rdev, sh->sector,
4580 rdev_dec_pending(rdev, conf->mddev);
4582 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4583 rdev = conf->disks[i].replacement;
4585 /* rdev have been moved down */
4586 rdev = conf->disks[i].rdev;
4587 rdev_clear_badblocks(rdev, sh->sector,
4589 rdev_dec_pending(rdev, conf->mddev);
4594 raid_run_ops(sh, s.ops_request);
4598 if (s.dec_preread_active) {
4599 /* We delay this until after ops_run_io so that if make_request
4600 * is waiting on a flush, it won't continue until the writes
4601 * have actually been submitted.
4603 atomic_dec(&conf->preread_active_stripes);
4604 if (atomic_read(&conf->preread_active_stripes) <
4606 md_wakeup_thread(conf->mddev->thread);
4609 if (!bio_list_empty(&s.return_bi)) {
4610 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags)) {
4611 spin_lock_irq(&conf->device_lock);
4612 bio_list_merge(&conf->return_bi, &s.return_bi);
4613 spin_unlock_irq(&conf->device_lock);
4614 md_wakeup_thread(conf->mddev->thread);
4616 return_io(&s.return_bi);
4619 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4622 static void raid5_activate_delayed(struct r5conf *conf)
4624 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4625 while (!list_empty(&conf->delayed_list)) {
4626 struct list_head *l = conf->delayed_list.next;
4627 struct stripe_head *sh;
4628 sh = list_entry(l, struct stripe_head, lru);
4630 clear_bit(STRIPE_DELAYED, &sh->state);
4631 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4632 atomic_inc(&conf->preread_active_stripes);
4633 list_add_tail(&sh->lru, &conf->hold_list);
4634 raid5_wakeup_stripe_thread(sh);
4639 static void activate_bit_delay(struct r5conf *conf,
4640 struct list_head *temp_inactive_list)
4642 /* device_lock is held */
4643 struct list_head head;
4644 list_add(&head, &conf->bitmap_list);
4645 list_del_init(&conf->bitmap_list);
4646 while (!list_empty(&head)) {
4647 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4649 list_del_init(&sh->lru);
4650 atomic_inc(&sh->count);
4651 hash = sh->hash_lock_index;
4652 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4656 static int raid5_congested(struct mddev *mddev, int bits)
4658 struct r5conf *conf = mddev->private;
4660 /* No difference between reads and writes. Just check
4661 * how busy the stripe_cache is
4664 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4668 if (atomic_read(&conf->empty_inactive_list_nr))
4674 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4676 struct r5conf *conf = mddev->private;
4677 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4678 unsigned int chunk_sectors;
4679 unsigned int bio_sectors = bio_sectors(bio);
4681 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4682 return chunk_sectors >=
4683 ((sector & (chunk_sectors - 1)) + bio_sectors);
4687 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4688 * later sampled by raid5d.
4690 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4692 unsigned long flags;
4694 spin_lock_irqsave(&conf->device_lock, flags);
4696 bi->bi_next = conf->retry_read_aligned_list;
4697 conf->retry_read_aligned_list = bi;
4699 spin_unlock_irqrestore(&conf->device_lock, flags);
4700 md_wakeup_thread(conf->mddev->thread);
4703 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4707 bi = conf->retry_read_aligned;
4709 conf->retry_read_aligned = NULL;
4712 bi = conf->retry_read_aligned_list;
4714 conf->retry_read_aligned_list = bi->bi_next;
4717 * this sets the active strip count to 1 and the processed
4718 * strip count to zero (upper 8 bits)
4720 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4727 * The "raid5_align_endio" should check if the read succeeded and if it
4728 * did, call bio_endio on the original bio (having bio_put the new bio
4730 * If the read failed..
4732 static void raid5_align_endio(struct bio *bi)
4734 struct bio* raid_bi = bi->bi_private;
4735 struct mddev *mddev;
4736 struct r5conf *conf;
4737 struct md_rdev *rdev;
4738 int error = bi->bi_error;
4742 rdev = (void*)raid_bi->bi_next;
4743 raid_bi->bi_next = NULL;
4744 mddev = rdev->mddev;
4745 conf = mddev->private;
4747 rdev_dec_pending(rdev, conf->mddev);
4750 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4753 if (atomic_dec_and_test(&conf->active_aligned_reads))
4754 wake_up(&conf->wait_for_quiescent);
4758 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4760 add_bio_to_retry(raid_bi, conf);
4763 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4765 struct r5conf *conf = mddev->private;
4767 struct bio* align_bi;
4768 struct md_rdev *rdev;
4769 sector_t end_sector;
4771 if (!in_chunk_boundary(mddev, raid_bio)) {
4772 pr_debug("%s: non aligned\n", __func__);
4776 * use bio_clone_mddev to make a copy of the bio
4778 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4782 * set bi_end_io to a new function, and set bi_private to the
4785 align_bi->bi_end_io = raid5_align_endio;
4786 align_bi->bi_private = raid_bio;
4790 align_bi->bi_iter.bi_sector =
4791 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4794 end_sector = bio_end_sector(align_bi);
4796 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4797 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4798 rdev->recovery_offset < end_sector) {
4799 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4801 (test_bit(Faulty, &rdev->flags) ||
4802 !(test_bit(In_sync, &rdev->flags) ||
4803 rdev->recovery_offset >= end_sector)))
4810 atomic_inc(&rdev->nr_pending);
4812 raid_bio->bi_next = (void*)rdev;
4813 align_bi->bi_bdev = rdev->bdev;
4814 bio_clear_flag(align_bi, BIO_SEG_VALID);
4816 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4817 bio_sectors(align_bi),
4818 &first_bad, &bad_sectors)) {
4820 rdev_dec_pending(rdev, mddev);
4824 /* No reshape active, so we can trust rdev->data_offset */
4825 align_bi->bi_iter.bi_sector += rdev->data_offset;
4827 spin_lock_irq(&conf->device_lock);
4828 wait_event_lock_irq(conf->wait_for_quiescent,
4831 atomic_inc(&conf->active_aligned_reads);
4832 spin_unlock_irq(&conf->device_lock);
4835 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4836 align_bi, disk_devt(mddev->gendisk),
4837 raid_bio->bi_iter.bi_sector);
4838 generic_make_request(align_bi);
4847 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4852 sector_t sector = raid_bio->bi_iter.bi_sector;
4853 unsigned chunk_sects = mddev->chunk_sectors;
4854 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4856 if (sectors < bio_sectors(raid_bio)) {
4857 split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4858 bio_chain(split, raid_bio);
4862 if (!raid5_read_one_chunk(mddev, split)) {
4863 if (split != raid_bio)
4864 generic_make_request(raid_bio);
4867 } while (split != raid_bio);
4872 /* __get_priority_stripe - get the next stripe to process
4874 * Full stripe writes are allowed to pass preread active stripes up until
4875 * the bypass_threshold is exceeded. In general the bypass_count
4876 * increments when the handle_list is handled before the hold_list; however, it
4877 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4878 * stripe with in flight i/o. The bypass_count will be reset when the
4879 * head of the hold_list has changed, i.e. the head was promoted to the
4882 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4884 struct stripe_head *sh = NULL, *tmp;
4885 struct list_head *handle_list = NULL;
4886 struct r5worker_group *wg = NULL;
4888 if (conf->worker_cnt_per_group == 0) {
4889 handle_list = &conf->handle_list;
4890 } else if (group != ANY_GROUP) {
4891 handle_list = &conf->worker_groups[group].handle_list;
4892 wg = &conf->worker_groups[group];
4895 for (i = 0; i < conf->group_cnt; i++) {
4896 handle_list = &conf->worker_groups[i].handle_list;
4897 wg = &conf->worker_groups[i];
4898 if (!list_empty(handle_list))
4903 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4905 list_empty(handle_list) ? "empty" : "busy",
4906 list_empty(&conf->hold_list) ? "empty" : "busy",
4907 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4909 if (!list_empty(handle_list)) {
4910 sh = list_entry(handle_list->next, typeof(*sh), lru);
4912 if (list_empty(&conf->hold_list))
4913 conf->bypass_count = 0;
4914 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4915 if (conf->hold_list.next == conf->last_hold)
4916 conf->bypass_count++;
4918 conf->last_hold = conf->hold_list.next;
4919 conf->bypass_count -= conf->bypass_threshold;
4920 if (conf->bypass_count < 0)
4921 conf->bypass_count = 0;
4924 } else if (!list_empty(&conf->hold_list) &&
4925 ((conf->bypass_threshold &&
4926 conf->bypass_count > conf->bypass_threshold) ||
4927 atomic_read(&conf->pending_full_writes) == 0)) {
4929 list_for_each_entry(tmp, &conf->hold_list, lru) {
4930 if (conf->worker_cnt_per_group == 0 ||
4931 group == ANY_GROUP ||
4932 !cpu_online(tmp->cpu) ||
4933 cpu_to_group(tmp->cpu) == group) {
4940 conf->bypass_count -= conf->bypass_threshold;
4941 if (conf->bypass_count < 0)
4942 conf->bypass_count = 0;
4954 list_del_init(&sh->lru);
4955 BUG_ON(atomic_inc_return(&sh->count) != 1);
4959 struct raid5_plug_cb {
4960 struct blk_plug_cb cb;
4961 struct list_head list;
4962 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4965 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4967 struct raid5_plug_cb *cb = container_of(
4968 blk_cb, struct raid5_plug_cb, cb);
4969 struct stripe_head *sh;
4970 struct mddev *mddev = cb->cb.data;
4971 struct r5conf *conf = mddev->private;
4975 if (cb->list.next && !list_empty(&cb->list)) {
4976 spin_lock_irq(&conf->device_lock);
4977 while (!list_empty(&cb->list)) {
4978 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4979 list_del_init(&sh->lru);
4981 * avoid race release_stripe_plug() sees
4982 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4983 * is still in our list
4985 smp_mb__before_atomic();
4986 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4988 * STRIPE_ON_RELEASE_LIST could be set here. In that
4989 * case, the count is always > 1 here
4991 hash = sh->hash_lock_index;
4992 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
4995 spin_unlock_irq(&conf->device_lock);
4997 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4998 NR_STRIPE_HASH_LOCKS);
5000 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5004 static void release_stripe_plug(struct mddev *mddev,
5005 struct stripe_head *sh)
5007 struct blk_plug_cb *blk_cb = blk_check_plugged(
5008 raid5_unplug, mddev,
5009 sizeof(struct raid5_plug_cb));
5010 struct raid5_plug_cb *cb;
5017 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5019 if (cb->list.next == NULL) {
5021 INIT_LIST_HEAD(&cb->list);
5022 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5023 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5026 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5027 list_add_tail(&sh->lru, &cb->list);
5032 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5034 struct r5conf *conf = mddev->private;
5035 sector_t logical_sector, last_sector;
5036 struct stripe_head *sh;
5040 if (mddev->reshape_position != MaxSector)
5041 /* Skip discard while reshape is happening */
5044 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5045 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5048 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5050 stripe_sectors = conf->chunk_sectors *
5051 (conf->raid_disks - conf->max_degraded);
5052 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5054 sector_div(last_sector, stripe_sectors);
5056 logical_sector *= conf->chunk_sectors;
5057 last_sector *= conf->chunk_sectors;
5059 for (; logical_sector < last_sector;
5060 logical_sector += STRIPE_SECTORS) {
5064 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
5065 prepare_to_wait(&conf->wait_for_overlap, &w,
5066 TASK_UNINTERRUPTIBLE);
5067 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5068 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5073 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5074 spin_lock_irq(&sh->stripe_lock);
5075 for (d = 0; d < conf->raid_disks; d++) {
5076 if (d == sh->pd_idx || d == sh->qd_idx)
5078 if (sh->dev[d].towrite || sh->dev[d].toread) {
5079 set_bit(R5_Overlap, &sh->dev[d].flags);
5080 spin_unlock_irq(&sh->stripe_lock);
5086 set_bit(STRIPE_DISCARD, &sh->state);
5087 finish_wait(&conf->wait_for_overlap, &w);
5088 sh->overwrite_disks = 0;
5089 for (d = 0; d < conf->raid_disks; d++) {
5090 if (d == sh->pd_idx || d == sh->qd_idx)
5092 sh->dev[d].towrite = bi;
5093 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5094 raid5_inc_bi_active_stripes(bi);
5095 sh->overwrite_disks++;
5097 spin_unlock_irq(&sh->stripe_lock);
5098 if (conf->mddev->bitmap) {
5100 d < conf->raid_disks - conf->max_degraded;
5102 bitmap_startwrite(mddev->bitmap,
5106 sh->bm_seq = conf->seq_flush + 1;
5107 set_bit(STRIPE_BIT_DELAY, &sh->state);
5110 set_bit(STRIPE_HANDLE, &sh->state);
5111 clear_bit(STRIPE_DELAYED, &sh->state);
5112 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5113 atomic_inc(&conf->preread_active_stripes);
5114 release_stripe_plug(mddev, sh);
5117 remaining = raid5_dec_bi_active_stripes(bi);
5118 if (remaining == 0) {
5119 md_write_end(mddev);
5124 static void make_request(struct mddev *mddev, struct bio * bi)
5126 struct r5conf *conf = mddev->private;
5128 sector_t new_sector;
5129 sector_t logical_sector, last_sector;
5130 struct stripe_head *sh;
5131 const int rw = bio_data_dir(bi);
5136 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5137 md_flush_request(mddev, bi);
5141 md_write_start(mddev, bi);
5144 * If array is degraded, better not do chunk aligned read because
5145 * later we might have to read it again in order to reconstruct
5146 * data on failed drives.
5148 if (rw == READ && mddev->degraded == 0 &&
5149 mddev->reshape_position == MaxSector) {
5150 bi = chunk_aligned_read(mddev, bi);
5155 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5156 make_discard_request(mddev, bi);
5160 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5161 last_sector = bio_end_sector(bi);
5163 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5165 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5166 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5172 seq = read_seqcount_begin(&conf->gen_lock);
5175 prepare_to_wait(&conf->wait_for_overlap, &w,
5176 TASK_UNINTERRUPTIBLE);
5177 if (unlikely(conf->reshape_progress != MaxSector)) {
5178 /* spinlock is needed as reshape_progress may be
5179 * 64bit on a 32bit platform, and so it might be
5180 * possible to see a half-updated value
5181 * Of course reshape_progress could change after
5182 * the lock is dropped, so once we get a reference
5183 * to the stripe that we think it is, we will have
5186 spin_lock_irq(&conf->device_lock);
5187 if (mddev->reshape_backwards
5188 ? logical_sector < conf->reshape_progress
5189 : logical_sector >= conf->reshape_progress) {
5192 if (mddev->reshape_backwards
5193 ? logical_sector < conf->reshape_safe
5194 : logical_sector >= conf->reshape_safe) {
5195 spin_unlock_irq(&conf->device_lock);
5201 spin_unlock_irq(&conf->device_lock);
5204 new_sector = raid5_compute_sector(conf, logical_sector,
5207 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5208 (unsigned long long)new_sector,
5209 (unsigned long long)logical_sector);
5211 sh = get_active_stripe(conf, new_sector, previous,
5212 (bi->bi_rw&RWA_MASK), 0);
5214 if (unlikely(previous)) {
5215 /* expansion might have moved on while waiting for a
5216 * stripe, so we must do the range check again.
5217 * Expansion could still move past after this
5218 * test, but as we are holding a reference to
5219 * 'sh', we know that if that happens,
5220 * STRIPE_EXPANDING will get set and the expansion
5221 * won't proceed until we finish with the stripe.
5224 spin_lock_irq(&conf->device_lock);
5225 if (mddev->reshape_backwards
5226 ? logical_sector >= conf->reshape_progress
5227 : logical_sector < conf->reshape_progress)
5228 /* mismatch, need to try again */
5230 spin_unlock_irq(&conf->device_lock);
5238 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5239 /* Might have got the wrong stripe_head
5247 logical_sector >= mddev->suspend_lo &&
5248 logical_sector < mddev->suspend_hi) {
5250 /* As the suspend_* range is controlled by
5251 * userspace, we want an interruptible
5254 flush_signals(current);
5255 prepare_to_wait(&conf->wait_for_overlap,
5256 &w, TASK_INTERRUPTIBLE);
5257 if (logical_sector >= mddev->suspend_lo &&
5258 logical_sector < mddev->suspend_hi) {
5265 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5266 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5267 /* Stripe is busy expanding or
5268 * add failed due to overlap. Flush everything
5271 md_wakeup_thread(mddev->thread);
5277 set_bit(STRIPE_HANDLE, &sh->state);
5278 clear_bit(STRIPE_DELAYED, &sh->state);
5279 if ((!sh->batch_head || sh == sh->batch_head) &&
5280 (bi->bi_rw & REQ_SYNC) &&
5281 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5282 atomic_inc(&conf->preread_active_stripes);
5283 release_stripe_plug(mddev, sh);
5285 /* cannot get stripe for read-ahead, just give-up */
5286 bi->bi_error = -EIO;
5290 finish_wait(&conf->wait_for_overlap, &w);
5292 remaining = raid5_dec_bi_active_stripes(bi);
5293 if (remaining == 0) {
5296 md_write_end(mddev);
5298 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5304 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5306 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5308 /* reshaping is quite different to recovery/resync so it is
5309 * handled quite separately ... here.
5311 * On each call to sync_request, we gather one chunk worth of
5312 * destination stripes and flag them as expanding.
5313 * Then we find all the source stripes and request reads.
5314 * As the reads complete, handle_stripe will copy the data
5315 * into the destination stripe and release that stripe.
5317 struct r5conf *conf = mddev->private;
5318 struct stripe_head *sh;
5319 sector_t first_sector, last_sector;
5320 int raid_disks = conf->previous_raid_disks;
5321 int data_disks = raid_disks - conf->max_degraded;
5322 int new_data_disks = conf->raid_disks - conf->max_degraded;
5325 sector_t writepos, readpos, safepos;
5326 sector_t stripe_addr;
5327 int reshape_sectors;
5328 struct list_head stripes;
5331 if (sector_nr == 0) {
5332 /* If restarting in the middle, skip the initial sectors */
5333 if (mddev->reshape_backwards &&
5334 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5335 sector_nr = raid5_size(mddev, 0, 0)
5336 - conf->reshape_progress;
5337 } else if (mddev->reshape_backwards &&
5338 conf->reshape_progress == MaxSector) {
5339 /* shouldn't happen, but just in case, finish up.*/
5340 sector_nr = MaxSector;
5341 } else if (!mddev->reshape_backwards &&
5342 conf->reshape_progress > 0)
5343 sector_nr = conf->reshape_progress;
5344 sector_div(sector_nr, new_data_disks);
5346 mddev->curr_resync_completed = sector_nr;
5347 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5354 /* We need to process a full chunk at a time.
5355 * If old and new chunk sizes differ, we need to process the
5359 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5361 /* We update the metadata at least every 10 seconds, or when
5362 * the data about to be copied would over-write the source of
5363 * the data at the front of the range. i.e. one new_stripe
5364 * along from reshape_progress new_maps to after where
5365 * reshape_safe old_maps to
5367 writepos = conf->reshape_progress;
5368 sector_div(writepos, new_data_disks);
5369 readpos = conf->reshape_progress;
5370 sector_div(readpos, data_disks);
5371 safepos = conf->reshape_safe;
5372 sector_div(safepos, data_disks);
5373 if (mddev->reshape_backwards) {
5374 BUG_ON(writepos < reshape_sectors);
5375 writepos -= reshape_sectors;
5376 readpos += reshape_sectors;
5377 safepos += reshape_sectors;
5379 writepos += reshape_sectors;
5380 /* readpos and safepos are worst-case calculations.
5381 * A negative number is overly pessimistic, and causes
5382 * obvious problems for unsigned storage. So clip to 0.
5384 readpos -= min_t(sector_t, reshape_sectors, readpos);
5385 safepos -= min_t(sector_t, reshape_sectors, safepos);
5388 /* Having calculated the 'writepos' possibly use it
5389 * to set 'stripe_addr' which is where we will write to.
5391 if (mddev->reshape_backwards) {
5392 BUG_ON(conf->reshape_progress == 0);
5393 stripe_addr = writepos;
5394 BUG_ON((mddev->dev_sectors &
5395 ~((sector_t)reshape_sectors - 1))
5396 - reshape_sectors - stripe_addr
5399 BUG_ON(writepos != sector_nr + reshape_sectors);
5400 stripe_addr = sector_nr;
5403 /* 'writepos' is the most advanced device address we might write.
5404 * 'readpos' is the least advanced device address we might read.
5405 * 'safepos' is the least address recorded in the metadata as having
5407 * If there is a min_offset_diff, these are adjusted either by
5408 * increasing the safepos/readpos if diff is negative, or
5409 * increasing writepos if diff is positive.
5410 * If 'readpos' is then behind 'writepos', there is no way that we can
5411 * ensure safety in the face of a crash - that must be done by userspace
5412 * making a backup of the data. So in that case there is no particular
5413 * rush to update metadata.
5414 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5415 * update the metadata to advance 'safepos' to match 'readpos' so that
5416 * we can be safe in the event of a crash.
5417 * So we insist on updating metadata if safepos is behind writepos and
5418 * readpos is beyond writepos.
5419 * In any case, update the metadata every 10 seconds.
5420 * Maybe that number should be configurable, but I'm not sure it is
5421 * worth it.... maybe it could be a multiple of safemode_delay???
5423 if (conf->min_offset_diff < 0) {
5424 safepos += -conf->min_offset_diff;
5425 readpos += -conf->min_offset_diff;
5427 writepos += conf->min_offset_diff;
5429 if ((mddev->reshape_backwards
5430 ? (safepos > writepos && readpos < writepos)
5431 : (safepos < writepos && readpos > writepos)) ||
5432 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5433 /* Cannot proceed until we've updated the superblock... */
5434 wait_event(conf->wait_for_overlap,
5435 atomic_read(&conf->reshape_stripes)==0
5436 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5437 if (atomic_read(&conf->reshape_stripes) != 0)
5439 mddev->reshape_position = conf->reshape_progress;
5440 mddev->curr_resync_completed = sector_nr;
5441 conf->reshape_checkpoint = jiffies;
5442 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5443 md_wakeup_thread(mddev->thread);
5444 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5445 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5446 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5448 spin_lock_irq(&conf->device_lock);
5449 conf->reshape_safe = mddev->reshape_position;
5450 spin_unlock_irq(&conf->device_lock);
5451 wake_up(&conf->wait_for_overlap);
5452 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5455 INIT_LIST_HEAD(&stripes);
5456 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5458 int skipped_disk = 0;
5459 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5460 set_bit(STRIPE_EXPANDING, &sh->state);
5461 atomic_inc(&conf->reshape_stripes);
5462 /* If any of this stripe is beyond the end of the old
5463 * array, then we need to zero those blocks
5465 for (j=sh->disks; j--;) {
5467 if (j == sh->pd_idx)
5469 if (conf->level == 6 &&
5472 s = compute_blocknr(sh, j, 0);
5473 if (s < raid5_size(mddev, 0, 0)) {
5477 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5478 set_bit(R5_Expanded, &sh->dev[j].flags);
5479 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5481 if (!skipped_disk) {
5482 set_bit(STRIPE_EXPAND_READY, &sh->state);
5483 set_bit(STRIPE_HANDLE, &sh->state);
5485 list_add(&sh->lru, &stripes);
5487 spin_lock_irq(&conf->device_lock);
5488 if (mddev->reshape_backwards)
5489 conf->reshape_progress -= reshape_sectors * new_data_disks;
5491 conf->reshape_progress += reshape_sectors * new_data_disks;
5492 spin_unlock_irq(&conf->device_lock);
5493 /* Ok, those stripe are ready. We can start scheduling
5494 * reads on the source stripes.
5495 * The source stripes are determined by mapping the first and last
5496 * block on the destination stripes.
5499 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5502 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5503 * new_data_disks - 1),
5505 if (last_sector >= mddev->dev_sectors)
5506 last_sector = mddev->dev_sectors - 1;
5507 while (first_sector <= last_sector) {
5508 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
5509 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5510 set_bit(STRIPE_HANDLE, &sh->state);
5512 first_sector += STRIPE_SECTORS;
5514 /* Now that the sources are clearly marked, we can release
5515 * the destination stripes
5517 while (!list_empty(&stripes)) {
5518 sh = list_entry(stripes.next, struct stripe_head, lru);
5519 list_del_init(&sh->lru);
5522 /* If this takes us to the resync_max point where we have to pause,
5523 * then we need to write out the superblock.
5525 sector_nr += reshape_sectors;
5526 retn = reshape_sectors;
5528 if (mddev->curr_resync_completed > mddev->resync_max ||
5529 (sector_nr - mddev->curr_resync_completed) * 2
5530 >= mddev->resync_max - mddev->curr_resync_completed) {
5531 /* Cannot proceed until we've updated the superblock... */
5532 wait_event(conf->wait_for_overlap,
5533 atomic_read(&conf->reshape_stripes) == 0
5534 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5535 if (atomic_read(&conf->reshape_stripes) != 0)
5537 mddev->reshape_position = conf->reshape_progress;
5538 mddev->curr_resync_completed = sector_nr;
5539 conf->reshape_checkpoint = jiffies;
5540 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5541 md_wakeup_thread(mddev->thread);
5542 wait_event(mddev->sb_wait,
5543 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5544 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5545 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5547 spin_lock_irq(&conf->device_lock);
5548 conf->reshape_safe = mddev->reshape_position;
5549 spin_unlock_irq(&conf->device_lock);
5550 wake_up(&conf->wait_for_overlap);
5551 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5557 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5559 struct r5conf *conf = mddev->private;
5560 struct stripe_head *sh;
5561 sector_t max_sector = mddev->dev_sectors;
5562 sector_t sync_blocks;
5563 int still_degraded = 0;
5566 if (sector_nr >= max_sector) {
5567 /* just being told to finish up .. nothing much to do */
5569 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5574 if (mddev->curr_resync < max_sector) /* aborted */
5575 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5577 else /* completed sync */
5579 bitmap_close_sync(mddev->bitmap);
5584 /* Allow raid5_quiesce to complete */
5585 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5587 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5588 return reshape_request(mddev, sector_nr, skipped);
5590 /* No need to check resync_max as we never do more than one
5591 * stripe, and as resync_max will always be on a chunk boundary,
5592 * if the check in md_do_sync didn't fire, there is no chance
5593 * of overstepping resync_max here
5596 /* if there is too many failed drives and we are trying
5597 * to resync, then assert that we are finished, because there is
5598 * nothing we can do.
5600 if (mddev->degraded >= conf->max_degraded &&
5601 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5602 sector_t rv = mddev->dev_sectors - sector_nr;
5606 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5608 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5609 sync_blocks >= STRIPE_SECTORS) {
5610 /* we can skip this block, and probably more */
5611 sync_blocks /= STRIPE_SECTORS;
5613 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5616 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5618 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5620 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5621 /* make sure we don't swamp the stripe cache if someone else
5622 * is trying to get access
5624 schedule_timeout_uninterruptible(1);
5626 /* Need to check if array will still be degraded after recovery/resync
5627 * Note in case of > 1 drive failures it's possible we're rebuilding
5628 * one drive while leaving another faulty drive in array.
5631 for (i = 0; i < conf->raid_disks; i++) {
5632 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5634 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5639 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5641 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5642 set_bit(STRIPE_HANDLE, &sh->state);
5646 return STRIPE_SECTORS;
5649 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5651 /* We may not be able to submit a whole bio at once as there
5652 * may not be enough stripe_heads available.
5653 * We cannot pre-allocate enough stripe_heads as we may need
5654 * more than exist in the cache (if we allow ever large chunks).
5655 * So we do one stripe head at a time and record in
5656 * ->bi_hw_segments how many have been done.
5658 * We *know* that this entire raid_bio is in one chunk, so
5659 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5661 struct stripe_head *sh;
5663 sector_t sector, logical_sector, last_sector;
5668 logical_sector = raid_bio->bi_iter.bi_sector &
5669 ~((sector_t)STRIPE_SECTORS-1);
5670 sector = raid5_compute_sector(conf, logical_sector,
5672 last_sector = bio_end_sector(raid_bio);
5674 for (; logical_sector < last_sector;
5675 logical_sector += STRIPE_SECTORS,
5676 sector += STRIPE_SECTORS,
5679 if (scnt < raid5_bi_processed_stripes(raid_bio))
5680 /* already done this stripe */
5683 sh = get_active_stripe(conf, sector, 0, 1, 1);
5686 /* failed to get a stripe - must wait */
5687 raid5_set_bi_processed_stripes(raid_bio, scnt);
5688 conf->retry_read_aligned = raid_bio;
5692 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5694 raid5_set_bi_processed_stripes(raid_bio, scnt);
5695 conf->retry_read_aligned = raid_bio;
5699 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5704 remaining = raid5_dec_bi_active_stripes(raid_bio);
5705 if (remaining == 0) {
5706 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5708 bio_endio(raid_bio);
5710 if (atomic_dec_and_test(&conf->active_aligned_reads))
5711 wake_up(&conf->wait_for_quiescent);
5715 static int handle_active_stripes(struct r5conf *conf, int group,
5716 struct r5worker *worker,
5717 struct list_head *temp_inactive_list)
5719 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5720 int i, batch_size = 0, hash;
5721 bool release_inactive = false;
5723 while (batch_size < MAX_STRIPE_BATCH &&
5724 (sh = __get_priority_stripe(conf, group)) != NULL)
5725 batch[batch_size++] = sh;
5727 if (batch_size == 0) {
5728 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5729 if (!list_empty(temp_inactive_list + i))
5731 if (i == NR_STRIPE_HASH_LOCKS)
5733 release_inactive = true;
5735 spin_unlock_irq(&conf->device_lock);
5737 release_inactive_stripe_list(conf, temp_inactive_list,
5738 NR_STRIPE_HASH_LOCKS);
5740 if (release_inactive) {
5741 spin_lock_irq(&conf->device_lock);
5745 for (i = 0; i < batch_size; i++)
5746 handle_stripe(batch[i]);
5750 spin_lock_irq(&conf->device_lock);
5751 for (i = 0; i < batch_size; i++) {
5752 hash = batch[i]->hash_lock_index;
5753 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5758 static void raid5_do_work(struct work_struct *work)
5760 struct r5worker *worker = container_of(work, struct r5worker, work);
5761 struct r5worker_group *group = worker->group;
5762 struct r5conf *conf = group->conf;
5763 int group_id = group - conf->worker_groups;
5765 struct blk_plug plug;
5767 pr_debug("+++ raid5worker active\n");
5769 blk_start_plug(&plug);
5771 spin_lock_irq(&conf->device_lock);
5773 int batch_size, released;
5775 released = release_stripe_list(conf, worker->temp_inactive_list);
5777 batch_size = handle_active_stripes(conf, group_id, worker,
5778 worker->temp_inactive_list);
5779 worker->working = false;
5780 if (!batch_size && !released)
5782 handled += batch_size;
5784 pr_debug("%d stripes handled\n", handled);
5786 spin_unlock_irq(&conf->device_lock);
5787 blk_finish_plug(&plug);
5789 pr_debug("--- raid5worker inactive\n");
5793 * This is our raid5 kernel thread.
5795 * We scan the hash table for stripes which can be handled now.
5796 * During the scan, completed stripes are saved for us by the interrupt
5797 * handler, so that they will not have to wait for our next wakeup.
5799 static void raid5d(struct md_thread *thread)
5801 struct mddev *mddev = thread->mddev;
5802 struct r5conf *conf = mddev->private;
5804 struct blk_plug plug;
5806 pr_debug("+++ raid5d active\n");
5808 md_check_recovery(mddev);
5810 if (!bio_list_empty(&conf->return_bi) &&
5811 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5812 struct bio_list tmp = BIO_EMPTY_LIST;
5813 spin_lock_irq(&conf->device_lock);
5814 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5815 bio_list_merge(&tmp, &conf->return_bi);
5816 bio_list_init(&conf->return_bi);
5818 spin_unlock_irq(&conf->device_lock);
5822 blk_start_plug(&plug);
5824 spin_lock_irq(&conf->device_lock);
5827 int batch_size, released;
5829 released = release_stripe_list(conf, conf->temp_inactive_list);
5831 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5834 !list_empty(&conf->bitmap_list)) {
5835 /* Now is a good time to flush some bitmap updates */
5837 spin_unlock_irq(&conf->device_lock);
5838 bitmap_unplug(mddev->bitmap);
5839 spin_lock_irq(&conf->device_lock);
5840 conf->seq_write = conf->seq_flush;
5841 activate_bit_delay(conf, conf->temp_inactive_list);
5843 raid5_activate_delayed(conf);
5845 while ((bio = remove_bio_from_retry(conf))) {
5847 spin_unlock_irq(&conf->device_lock);
5848 ok = retry_aligned_read(conf, bio);
5849 spin_lock_irq(&conf->device_lock);
5855 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5856 conf->temp_inactive_list);
5857 if (!batch_size && !released)
5859 handled += batch_size;
5861 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5862 spin_unlock_irq(&conf->device_lock);
5863 md_check_recovery(mddev);
5864 spin_lock_irq(&conf->device_lock);
5867 pr_debug("%d stripes handled\n", handled);
5869 spin_unlock_irq(&conf->device_lock);
5870 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5871 mutex_trylock(&conf->cache_size_mutex)) {
5872 grow_one_stripe(conf, __GFP_NOWARN);
5873 /* Set flag even if allocation failed. This helps
5874 * slow down allocation requests when mem is short
5876 set_bit(R5_DID_ALLOC, &conf->cache_state);
5877 mutex_unlock(&conf->cache_size_mutex);
5880 async_tx_issue_pending_all();
5881 blk_finish_plug(&plug);
5883 pr_debug("--- raid5d inactive\n");
5887 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5889 struct r5conf *conf;
5891 spin_lock(&mddev->lock);
5892 conf = mddev->private;
5894 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5895 spin_unlock(&mddev->lock);
5900 raid5_set_cache_size(struct mddev *mddev, int size)
5902 struct r5conf *conf = mddev->private;
5905 if (size <= 16 || size > 32768)
5908 conf->min_nr_stripes = size;
5909 mutex_lock(&conf->cache_size_mutex);
5910 while (size < conf->max_nr_stripes &&
5911 drop_one_stripe(conf))
5913 mutex_unlock(&conf->cache_size_mutex);
5916 err = md_allow_write(mddev);
5920 mutex_lock(&conf->cache_size_mutex);
5921 while (size > conf->max_nr_stripes)
5922 if (!grow_one_stripe(conf, GFP_KERNEL))
5924 mutex_unlock(&conf->cache_size_mutex);
5928 EXPORT_SYMBOL(raid5_set_cache_size);
5931 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5933 struct r5conf *conf;
5937 if (len >= PAGE_SIZE)
5939 if (kstrtoul(page, 10, &new))
5941 err = mddev_lock(mddev);
5944 conf = mddev->private;
5948 err = raid5_set_cache_size(mddev, new);
5949 mddev_unlock(mddev);
5954 static struct md_sysfs_entry
5955 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5956 raid5_show_stripe_cache_size,
5957 raid5_store_stripe_cache_size);
5960 raid5_show_rmw_level(struct mddev *mddev, char *page)
5962 struct r5conf *conf = mddev->private;
5964 return sprintf(page, "%d\n", conf->rmw_level);
5970 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
5972 struct r5conf *conf = mddev->private;
5978 if (len >= PAGE_SIZE)
5981 if (kstrtoul(page, 10, &new))
5984 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
5987 if (new != PARITY_DISABLE_RMW &&
5988 new != PARITY_ENABLE_RMW &&
5989 new != PARITY_PREFER_RMW)
5992 conf->rmw_level = new;
5996 static struct md_sysfs_entry
5997 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
5998 raid5_show_rmw_level,
5999 raid5_store_rmw_level);
6003 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6005 struct r5conf *conf;
6007 spin_lock(&mddev->lock);
6008 conf = mddev->private;
6010 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6011 spin_unlock(&mddev->lock);
6016 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6018 struct r5conf *conf;
6022 if (len >= PAGE_SIZE)
6024 if (kstrtoul(page, 10, &new))
6027 err = mddev_lock(mddev);
6030 conf = mddev->private;
6033 else if (new > conf->min_nr_stripes)
6036 conf->bypass_threshold = new;
6037 mddev_unlock(mddev);
6041 static struct md_sysfs_entry
6042 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6044 raid5_show_preread_threshold,
6045 raid5_store_preread_threshold);
6048 raid5_show_skip_copy(struct mddev *mddev, char *page)
6050 struct r5conf *conf;
6052 spin_lock(&mddev->lock);
6053 conf = mddev->private;
6055 ret = sprintf(page, "%d\n", conf->skip_copy);
6056 spin_unlock(&mddev->lock);
6061 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6063 struct r5conf *conf;
6067 if (len >= PAGE_SIZE)
6069 if (kstrtoul(page, 10, &new))
6073 err = mddev_lock(mddev);
6076 conf = mddev->private;
6079 else if (new != conf->skip_copy) {
6080 mddev_suspend(mddev);
6081 conf->skip_copy = new;
6083 mddev->queue->backing_dev_info.capabilities |=
6084 BDI_CAP_STABLE_WRITES;
6086 mddev->queue->backing_dev_info.capabilities &=
6087 ~BDI_CAP_STABLE_WRITES;
6088 mddev_resume(mddev);
6090 mddev_unlock(mddev);
6094 static struct md_sysfs_entry
6095 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6096 raid5_show_skip_copy,
6097 raid5_store_skip_copy);
6100 stripe_cache_active_show(struct mddev *mddev, char *page)
6102 struct r5conf *conf = mddev->private;
6104 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6109 static struct md_sysfs_entry
6110 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6113 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6115 struct r5conf *conf;
6117 spin_lock(&mddev->lock);
6118 conf = mddev->private;
6120 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6121 spin_unlock(&mddev->lock);
6125 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6127 int *worker_cnt_per_group,
6128 struct r5worker_group **worker_groups);
6130 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6132 struct r5conf *conf;
6135 struct r5worker_group *new_groups, *old_groups;
6136 int group_cnt, worker_cnt_per_group;
6138 if (len >= PAGE_SIZE)
6140 if (kstrtoul(page, 10, &new))
6143 err = mddev_lock(mddev);
6146 conf = mddev->private;
6149 else if (new != conf->worker_cnt_per_group) {
6150 mddev_suspend(mddev);
6152 old_groups = conf->worker_groups;
6154 flush_workqueue(raid5_wq);
6156 err = alloc_thread_groups(conf, new,
6157 &group_cnt, &worker_cnt_per_group,
6160 spin_lock_irq(&conf->device_lock);
6161 conf->group_cnt = group_cnt;
6162 conf->worker_cnt_per_group = worker_cnt_per_group;
6163 conf->worker_groups = new_groups;
6164 spin_unlock_irq(&conf->device_lock);
6167 kfree(old_groups[0].workers);
6170 mddev_resume(mddev);
6172 mddev_unlock(mddev);
6177 static struct md_sysfs_entry
6178 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6179 raid5_show_group_thread_cnt,
6180 raid5_store_group_thread_cnt);
6182 static struct attribute *raid5_attrs[] = {
6183 &raid5_stripecache_size.attr,
6184 &raid5_stripecache_active.attr,
6185 &raid5_preread_bypass_threshold.attr,
6186 &raid5_group_thread_cnt.attr,
6187 &raid5_skip_copy.attr,
6188 &raid5_rmw_level.attr,
6191 static struct attribute_group raid5_attrs_group = {
6193 .attrs = raid5_attrs,
6196 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6198 int *worker_cnt_per_group,
6199 struct r5worker_group **worker_groups)
6203 struct r5worker *workers;
6205 *worker_cnt_per_group = cnt;
6208 *worker_groups = NULL;
6211 *group_cnt = num_possible_nodes();
6212 size = sizeof(struct r5worker) * cnt;
6213 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6214 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6215 *group_cnt, GFP_NOIO);
6216 if (!*worker_groups || !workers) {
6218 kfree(*worker_groups);
6222 for (i = 0; i < *group_cnt; i++) {
6223 struct r5worker_group *group;
6225 group = &(*worker_groups)[i];
6226 INIT_LIST_HEAD(&group->handle_list);
6228 group->workers = workers + i * cnt;
6230 for (j = 0; j < cnt; j++) {
6231 struct r5worker *worker = group->workers + j;
6232 worker->group = group;
6233 INIT_WORK(&worker->work, raid5_do_work);
6235 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6236 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6243 static void free_thread_groups(struct r5conf *conf)
6245 if (conf->worker_groups)
6246 kfree(conf->worker_groups[0].workers);
6247 kfree(conf->worker_groups);
6248 conf->worker_groups = NULL;
6252 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6254 struct r5conf *conf = mddev->private;
6257 sectors = mddev->dev_sectors;
6259 /* size is defined by the smallest of previous and new size */
6260 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6262 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6263 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6264 return sectors * (raid_disks - conf->max_degraded);
6267 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6269 safe_put_page(percpu->spare_page);
6270 if (percpu->scribble)
6271 flex_array_free(percpu->scribble);
6272 percpu->spare_page = NULL;
6273 percpu->scribble = NULL;
6276 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6278 if (conf->level == 6 && !percpu->spare_page)
6279 percpu->spare_page = alloc_page(GFP_KERNEL);
6280 if (!percpu->scribble)
6281 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6282 conf->previous_raid_disks),
6283 max(conf->chunk_sectors,
6284 conf->prev_chunk_sectors)
6288 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6289 free_scratch_buffer(conf, percpu);
6296 static void raid5_free_percpu(struct r5conf *conf)
6303 #ifdef CONFIG_HOTPLUG_CPU
6304 unregister_cpu_notifier(&conf->cpu_notify);
6308 for_each_possible_cpu(cpu)
6309 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6312 free_percpu(conf->percpu);
6315 static void free_conf(struct r5conf *conf)
6317 if (conf->shrinker.seeks)
6318 unregister_shrinker(&conf->shrinker);
6319 free_thread_groups(conf);
6320 shrink_stripes(conf);
6321 raid5_free_percpu(conf);
6323 kfree(conf->stripe_hashtbl);
6327 #ifdef CONFIG_HOTPLUG_CPU
6328 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6331 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6332 long cpu = (long)hcpu;
6333 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6336 case CPU_UP_PREPARE:
6337 case CPU_UP_PREPARE_FROZEN:
6338 if (alloc_scratch_buffer(conf, percpu)) {
6339 pr_err("%s: failed memory allocation for cpu%ld\n",
6341 return notifier_from_errno(-ENOMEM);
6345 case CPU_DEAD_FROZEN:
6346 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6355 static int raid5_alloc_percpu(struct r5conf *conf)
6360 conf->percpu = alloc_percpu(struct raid5_percpu);
6364 #ifdef CONFIG_HOTPLUG_CPU
6365 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6366 conf->cpu_notify.priority = 0;
6367 err = register_cpu_notifier(&conf->cpu_notify);
6373 for_each_present_cpu(cpu) {
6374 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6376 pr_err("%s: failed memory allocation for cpu%ld\n",
6386 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6387 struct shrink_control *sc)
6389 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6390 unsigned long ret = SHRINK_STOP;
6392 if (mutex_trylock(&conf->cache_size_mutex)) {
6394 while (ret < sc->nr_to_scan &&
6395 conf->max_nr_stripes > conf->min_nr_stripes) {
6396 if (drop_one_stripe(conf) == 0) {
6402 mutex_unlock(&conf->cache_size_mutex);
6407 static unsigned long raid5_cache_count(struct shrinker *shrink,
6408 struct shrink_control *sc)
6410 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6412 if (conf->max_nr_stripes < conf->min_nr_stripes)
6413 /* unlikely, but not impossible */
6415 return conf->max_nr_stripes - conf->min_nr_stripes;
6418 static struct r5conf *setup_conf(struct mddev *mddev)
6420 struct r5conf *conf;
6421 int raid_disk, memory, max_disks;
6422 struct md_rdev *rdev;
6423 struct disk_info *disk;
6426 int group_cnt, worker_cnt_per_group;
6427 struct r5worker_group *new_group;
6429 if (mddev->new_level != 5
6430 && mddev->new_level != 4
6431 && mddev->new_level != 6) {
6432 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6433 mdname(mddev), mddev->new_level);
6434 return ERR_PTR(-EIO);
6436 if ((mddev->new_level == 5
6437 && !algorithm_valid_raid5(mddev->new_layout)) ||
6438 (mddev->new_level == 6
6439 && !algorithm_valid_raid6(mddev->new_layout))) {
6440 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6441 mdname(mddev), mddev->new_layout);
6442 return ERR_PTR(-EIO);
6444 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6445 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6446 mdname(mddev), mddev->raid_disks);
6447 return ERR_PTR(-EINVAL);
6450 if (!mddev->new_chunk_sectors ||
6451 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6452 !is_power_of_2(mddev->new_chunk_sectors)) {
6453 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6454 mdname(mddev), mddev->new_chunk_sectors << 9);
6455 return ERR_PTR(-EINVAL);
6458 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6461 /* Don't enable multi-threading by default*/
6462 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6464 conf->group_cnt = group_cnt;
6465 conf->worker_cnt_per_group = worker_cnt_per_group;
6466 conf->worker_groups = new_group;
6469 spin_lock_init(&conf->device_lock);
6470 seqcount_init(&conf->gen_lock);
6471 mutex_init(&conf->cache_size_mutex);
6472 init_waitqueue_head(&conf->wait_for_quiescent);
6473 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
6474 init_waitqueue_head(&conf->wait_for_stripe[i]);
6476 init_waitqueue_head(&conf->wait_for_overlap);
6477 INIT_LIST_HEAD(&conf->handle_list);
6478 INIT_LIST_HEAD(&conf->hold_list);
6479 INIT_LIST_HEAD(&conf->delayed_list);
6480 INIT_LIST_HEAD(&conf->bitmap_list);
6481 bio_list_init(&conf->return_bi);
6482 init_llist_head(&conf->released_stripes);
6483 atomic_set(&conf->active_stripes, 0);
6484 atomic_set(&conf->preread_active_stripes, 0);
6485 atomic_set(&conf->active_aligned_reads, 0);
6486 conf->bypass_threshold = BYPASS_THRESHOLD;
6487 conf->recovery_disabled = mddev->recovery_disabled - 1;
6489 conf->raid_disks = mddev->raid_disks;
6490 if (mddev->reshape_position == MaxSector)
6491 conf->previous_raid_disks = mddev->raid_disks;
6493 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6494 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6496 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6501 conf->mddev = mddev;
6503 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6506 /* We init hash_locks[0] separately to that it can be used
6507 * as the reference lock in the spin_lock_nest_lock() call
6508 * in lock_all_device_hash_locks_irq in order to convince
6509 * lockdep that we know what we are doing.
6511 spin_lock_init(conf->hash_locks);
6512 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6513 spin_lock_init(conf->hash_locks + i);
6515 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6516 INIT_LIST_HEAD(conf->inactive_list + i);
6518 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6519 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6521 conf->level = mddev->new_level;
6522 conf->chunk_sectors = mddev->new_chunk_sectors;
6523 if (raid5_alloc_percpu(conf) != 0)
6526 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6528 rdev_for_each(rdev, mddev) {
6529 raid_disk = rdev->raid_disk;
6530 if (raid_disk >= max_disks
6533 disk = conf->disks + raid_disk;
6535 if (test_bit(Replacement, &rdev->flags)) {
6536 if (disk->replacement)
6538 disk->replacement = rdev;
6545 if (test_bit(In_sync, &rdev->flags)) {
6546 char b[BDEVNAME_SIZE];
6547 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6549 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6550 } else if (rdev->saved_raid_disk != raid_disk)
6551 /* Cannot rely on bitmap to complete recovery */
6555 conf->level = mddev->new_level;
6556 if (conf->level == 6) {
6557 conf->max_degraded = 2;
6558 if (raid6_call.xor_syndrome)
6559 conf->rmw_level = PARITY_ENABLE_RMW;
6561 conf->rmw_level = PARITY_DISABLE_RMW;
6563 conf->max_degraded = 1;
6564 conf->rmw_level = PARITY_ENABLE_RMW;
6566 conf->algorithm = mddev->new_layout;
6567 conf->reshape_progress = mddev->reshape_position;
6568 if (conf->reshape_progress != MaxSector) {
6569 conf->prev_chunk_sectors = mddev->chunk_sectors;
6570 conf->prev_algo = mddev->layout;
6572 conf->prev_chunk_sectors = conf->chunk_sectors;
6573 conf->prev_algo = conf->algorithm;
6576 conf->min_nr_stripes = NR_STRIPES;
6577 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6578 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6579 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6580 if (grow_stripes(conf, conf->min_nr_stripes)) {
6582 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6583 mdname(mddev), memory);
6586 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6587 mdname(mddev), memory);
6589 * Losing a stripe head costs more than the time to refill it,
6590 * it reduces the queue depth and so can hurt throughput.
6591 * So set it rather large, scaled by number of devices.
6593 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6594 conf->shrinker.scan_objects = raid5_cache_scan;
6595 conf->shrinker.count_objects = raid5_cache_count;
6596 conf->shrinker.batch = 128;
6597 conf->shrinker.flags = 0;
6598 register_shrinker(&conf->shrinker);
6600 sprintf(pers_name, "raid%d", mddev->new_level);
6601 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6602 if (!conf->thread) {
6604 "md/raid:%s: couldn't allocate thread.\n",
6614 return ERR_PTR(-EIO);
6616 return ERR_PTR(-ENOMEM);
6619 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6622 case ALGORITHM_PARITY_0:
6623 if (raid_disk < max_degraded)
6626 case ALGORITHM_PARITY_N:
6627 if (raid_disk >= raid_disks - max_degraded)
6630 case ALGORITHM_PARITY_0_6:
6631 if (raid_disk == 0 ||
6632 raid_disk == raid_disks - 1)
6635 case ALGORITHM_LEFT_ASYMMETRIC_6:
6636 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6637 case ALGORITHM_LEFT_SYMMETRIC_6:
6638 case ALGORITHM_RIGHT_SYMMETRIC_6:
6639 if (raid_disk == raid_disks - 1)
6645 static int run(struct mddev *mddev)
6647 struct r5conf *conf;
6648 int working_disks = 0;
6649 int dirty_parity_disks = 0;
6650 struct md_rdev *rdev;
6651 sector_t reshape_offset = 0;
6653 long long min_offset_diff = 0;
6656 if (mddev->recovery_cp != MaxSector)
6657 printk(KERN_NOTICE "md/raid:%s: not clean"
6658 " -- starting background reconstruction\n",
6661 rdev_for_each(rdev, mddev) {
6663 if (rdev->raid_disk < 0)
6665 diff = (rdev->new_data_offset - rdev->data_offset);
6667 min_offset_diff = diff;
6669 } else if (mddev->reshape_backwards &&
6670 diff < min_offset_diff)
6671 min_offset_diff = diff;
6672 else if (!mddev->reshape_backwards &&
6673 diff > min_offset_diff)
6674 min_offset_diff = diff;
6677 if (mddev->reshape_position != MaxSector) {
6678 /* Check that we can continue the reshape.
6679 * Difficulties arise if the stripe we would write to
6680 * next is at or after the stripe we would read from next.
6681 * For a reshape that changes the number of devices, this
6682 * is only possible for a very short time, and mdadm makes
6683 * sure that time appears to have past before assembling
6684 * the array. So we fail if that time hasn't passed.
6685 * For a reshape that keeps the number of devices the same
6686 * mdadm must be monitoring the reshape can keeping the
6687 * critical areas read-only and backed up. It will start
6688 * the array in read-only mode, so we check for that.
6690 sector_t here_new, here_old;
6692 int max_degraded = (mddev->level == 6 ? 2 : 1);
6696 if (mddev->new_level != mddev->level) {
6697 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6698 "required - aborting.\n",
6702 old_disks = mddev->raid_disks - mddev->delta_disks;
6703 /* reshape_position must be on a new-stripe boundary, and one
6704 * further up in new geometry must map after here in old
6706 * If the chunk sizes are different, then as we perform reshape
6707 * in units of the largest of the two, reshape_position needs
6708 * be a multiple of the largest chunk size times new data disks.
6710 here_new = mddev->reshape_position;
6711 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6712 new_data_disks = mddev->raid_disks - max_degraded;
6713 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6714 printk(KERN_ERR "md/raid:%s: reshape_position not "
6715 "on a stripe boundary\n", mdname(mddev));
6718 reshape_offset = here_new * chunk_sectors;
6719 /* here_new is the stripe we will write to */
6720 here_old = mddev->reshape_position;
6721 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6722 /* here_old is the first stripe that we might need to read
6724 if (mddev->delta_disks == 0) {
6725 /* We cannot be sure it is safe to start an in-place
6726 * reshape. It is only safe if user-space is monitoring
6727 * and taking constant backups.
6728 * mdadm always starts a situation like this in
6729 * readonly mode so it can take control before
6730 * allowing any writes. So just check for that.
6732 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6733 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6734 /* not really in-place - so OK */;
6735 else if (mddev->ro == 0) {
6736 printk(KERN_ERR "md/raid:%s: in-place reshape "
6737 "must be started in read-only mode "
6742 } else if (mddev->reshape_backwards
6743 ? (here_new * chunk_sectors + min_offset_diff <=
6744 here_old * chunk_sectors)
6745 : (here_new * chunk_sectors >=
6746 here_old * chunk_sectors + (-min_offset_diff))) {
6747 /* Reading from the same stripe as writing to - bad */
6748 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6749 "auto-recovery - aborting.\n",
6753 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6755 /* OK, we should be able to continue; */
6757 BUG_ON(mddev->level != mddev->new_level);
6758 BUG_ON(mddev->layout != mddev->new_layout);
6759 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6760 BUG_ON(mddev->delta_disks != 0);
6763 if (mddev->private == NULL)
6764 conf = setup_conf(mddev);
6766 conf = mddev->private;
6769 return PTR_ERR(conf);
6771 conf->min_offset_diff = min_offset_diff;
6772 mddev->thread = conf->thread;
6773 conf->thread = NULL;
6774 mddev->private = conf;
6776 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6778 rdev = conf->disks[i].rdev;
6779 if (!rdev && conf->disks[i].replacement) {
6780 /* The replacement is all we have yet */
6781 rdev = conf->disks[i].replacement;
6782 conf->disks[i].replacement = NULL;
6783 clear_bit(Replacement, &rdev->flags);
6784 conf->disks[i].rdev = rdev;
6788 if (conf->disks[i].replacement &&
6789 conf->reshape_progress != MaxSector) {
6790 /* replacements and reshape simply do not mix. */
6791 printk(KERN_ERR "md: cannot handle concurrent "
6792 "replacement and reshape.\n");
6795 if (test_bit(In_sync, &rdev->flags)) {
6799 /* This disc is not fully in-sync. However if it
6800 * just stored parity (beyond the recovery_offset),
6801 * when we don't need to be concerned about the
6802 * array being dirty.
6803 * When reshape goes 'backwards', we never have
6804 * partially completed devices, so we only need
6805 * to worry about reshape going forwards.
6807 /* Hack because v0.91 doesn't store recovery_offset properly. */
6808 if (mddev->major_version == 0 &&
6809 mddev->minor_version > 90)
6810 rdev->recovery_offset = reshape_offset;
6812 if (rdev->recovery_offset < reshape_offset) {
6813 /* We need to check old and new layout */
6814 if (!only_parity(rdev->raid_disk,
6817 conf->max_degraded))
6820 if (!only_parity(rdev->raid_disk,
6822 conf->previous_raid_disks,
6823 conf->max_degraded))
6825 dirty_parity_disks++;
6829 * 0 for a fully functional array, 1 or 2 for a degraded array.
6831 mddev->degraded = calc_degraded(conf);
6833 if (has_failed(conf)) {
6834 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6835 " (%d/%d failed)\n",
6836 mdname(mddev), mddev->degraded, conf->raid_disks);
6840 /* device size must be a multiple of chunk size */
6841 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6842 mddev->resync_max_sectors = mddev->dev_sectors;
6844 if (mddev->degraded > dirty_parity_disks &&
6845 mddev->recovery_cp != MaxSector) {
6846 if (mddev->ok_start_degraded)
6848 "md/raid:%s: starting dirty degraded array"
6849 " - data corruption possible.\n",
6853 "md/raid:%s: cannot start dirty degraded array.\n",
6859 if (mddev->degraded == 0)
6860 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6861 " devices, algorithm %d\n", mdname(mddev), conf->level,
6862 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6865 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6866 " out of %d devices, algorithm %d\n",
6867 mdname(mddev), conf->level,
6868 mddev->raid_disks - mddev->degraded,
6869 mddev->raid_disks, mddev->new_layout);
6871 print_raid5_conf(conf);
6873 if (conf->reshape_progress != MaxSector) {
6874 conf->reshape_safe = conf->reshape_progress;
6875 atomic_set(&conf->reshape_stripes, 0);
6876 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6877 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6878 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6879 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6880 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6884 /* Ok, everything is just fine now */
6885 if (mddev->to_remove == &raid5_attrs_group)
6886 mddev->to_remove = NULL;
6887 else if (mddev->kobj.sd &&
6888 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6890 "raid5: failed to create sysfs attributes for %s\n",
6892 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6896 bool discard_supported = true;
6897 /* read-ahead size must cover two whole stripes, which
6898 * is 2 * (datadisks) * chunksize where 'n' is the
6899 * number of raid devices
6901 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6902 int stripe = data_disks *
6903 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6904 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6905 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6907 chunk_size = mddev->chunk_sectors << 9;
6908 blk_queue_io_min(mddev->queue, chunk_size);
6909 blk_queue_io_opt(mddev->queue, chunk_size *
6910 (conf->raid_disks - conf->max_degraded));
6911 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6913 * We can only discard a whole stripe. It doesn't make sense to
6914 * discard data disk but write parity disk
6916 stripe = stripe * PAGE_SIZE;
6917 /* Round up to power of 2, as discard handling
6918 * currently assumes that */
6919 while ((stripe-1) & stripe)
6920 stripe = (stripe | (stripe-1)) + 1;
6921 mddev->queue->limits.discard_alignment = stripe;
6922 mddev->queue->limits.discard_granularity = stripe;
6924 * unaligned part of discard request will be ignored, so can't
6925 * guarantee discard_zeroes_data
6927 mddev->queue->limits.discard_zeroes_data = 0;
6929 blk_queue_max_write_same_sectors(mddev->queue, 0);
6931 rdev_for_each(rdev, mddev) {
6932 disk_stack_limits(mddev->gendisk, rdev->bdev,
6933 rdev->data_offset << 9);
6934 disk_stack_limits(mddev->gendisk, rdev->bdev,
6935 rdev->new_data_offset << 9);
6937 * discard_zeroes_data is required, otherwise data
6938 * could be lost. Consider a scenario: discard a stripe
6939 * (the stripe could be inconsistent if
6940 * discard_zeroes_data is 0); write one disk of the
6941 * stripe (the stripe could be inconsistent again
6942 * depending on which disks are used to calculate
6943 * parity); the disk is broken; The stripe data of this
6946 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6947 !bdev_get_queue(rdev->bdev)->
6948 limits.discard_zeroes_data)
6949 discard_supported = false;
6950 /* Unfortunately, discard_zeroes_data is not currently
6951 * a guarantee - just a hint. So we only allow DISCARD
6952 * if the sysadmin has confirmed that only safe devices
6953 * are in use by setting a module parameter.
6955 if (!devices_handle_discard_safely) {
6956 if (discard_supported) {
6957 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6958 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6960 discard_supported = false;
6964 if (discard_supported &&
6965 mddev->queue->limits.max_discard_sectors >= stripe &&
6966 mddev->queue->limits.discard_granularity >= stripe)
6967 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6970 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6976 md_unregister_thread(&mddev->thread);
6977 print_raid5_conf(conf);
6979 mddev->private = NULL;
6980 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6984 static void raid5_free(struct mddev *mddev, void *priv)
6986 struct r5conf *conf = priv;
6989 mddev->to_remove = &raid5_attrs_group;
6992 static void status(struct seq_file *seq, struct mddev *mddev)
6994 struct r5conf *conf = mddev->private;
6997 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6998 conf->chunk_sectors / 2, mddev->layout);
6999 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7000 for (i = 0; i < conf->raid_disks; i++)
7001 seq_printf (seq, "%s",
7002 conf->disks[i].rdev &&
7003 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
7004 seq_printf (seq, "]");
7007 static void print_raid5_conf (struct r5conf *conf)
7010 struct disk_info *tmp;
7012 printk(KERN_DEBUG "RAID conf printout:\n");
7014 printk("(conf==NULL)\n");
7017 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7019 conf->raid_disks - conf->mddev->degraded);
7021 for (i = 0; i < conf->raid_disks; i++) {
7022 char b[BDEVNAME_SIZE];
7023 tmp = conf->disks + i;
7025 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7026 i, !test_bit(Faulty, &tmp->rdev->flags),
7027 bdevname(tmp->rdev->bdev, b));
7031 static int raid5_spare_active(struct mddev *mddev)
7034 struct r5conf *conf = mddev->private;
7035 struct disk_info *tmp;
7037 unsigned long flags;
7039 for (i = 0; i < conf->raid_disks; i++) {
7040 tmp = conf->disks + i;
7041 if (tmp->replacement
7042 && tmp->replacement->recovery_offset == MaxSector
7043 && !test_bit(Faulty, &tmp->replacement->flags)
7044 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7045 /* Replacement has just become active. */
7047 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7050 /* Replaced device not technically faulty,
7051 * but we need to be sure it gets removed
7052 * and never re-added.
7054 set_bit(Faulty, &tmp->rdev->flags);
7055 sysfs_notify_dirent_safe(
7056 tmp->rdev->sysfs_state);
7058 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7059 } else if (tmp->rdev
7060 && tmp->rdev->recovery_offset == MaxSector
7061 && !test_bit(Faulty, &tmp->rdev->flags)
7062 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7064 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7067 spin_lock_irqsave(&conf->device_lock, flags);
7068 mddev->degraded = calc_degraded(conf);
7069 spin_unlock_irqrestore(&conf->device_lock, flags);
7070 print_raid5_conf(conf);
7074 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7076 struct r5conf *conf = mddev->private;
7078 int number = rdev->raid_disk;
7079 struct md_rdev **rdevp;
7080 struct disk_info *p = conf->disks + number;
7082 print_raid5_conf(conf);
7083 if (rdev == p->rdev)
7085 else if (rdev == p->replacement)
7086 rdevp = &p->replacement;
7090 if (number >= conf->raid_disks &&
7091 conf->reshape_progress == MaxSector)
7092 clear_bit(In_sync, &rdev->flags);
7094 if (test_bit(In_sync, &rdev->flags) ||
7095 atomic_read(&rdev->nr_pending)) {
7099 /* Only remove non-faulty devices if recovery
7102 if (!test_bit(Faulty, &rdev->flags) &&
7103 mddev->recovery_disabled != conf->recovery_disabled &&
7104 !has_failed(conf) &&
7105 (!p->replacement || p->replacement == rdev) &&
7106 number < conf->raid_disks) {
7112 if (atomic_read(&rdev->nr_pending)) {
7113 /* lost the race, try later */
7116 } else if (p->replacement) {
7117 /* We must have just cleared 'rdev' */
7118 p->rdev = p->replacement;
7119 clear_bit(Replacement, &p->replacement->flags);
7120 smp_mb(); /* Make sure other CPUs may see both as identical
7121 * but will never see neither - if they are careful
7123 p->replacement = NULL;
7124 clear_bit(WantReplacement, &rdev->flags);
7126 /* We might have just removed the Replacement as faulty-
7127 * clear the bit just in case
7129 clear_bit(WantReplacement, &rdev->flags);
7132 print_raid5_conf(conf);
7136 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7138 struct r5conf *conf = mddev->private;
7141 struct disk_info *p;
7143 int last = conf->raid_disks - 1;
7145 if (mddev->recovery_disabled == conf->recovery_disabled)
7148 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7149 /* no point adding a device */
7152 if (rdev->raid_disk >= 0)
7153 first = last = rdev->raid_disk;
7156 * find the disk ... but prefer rdev->saved_raid_disk
7159 if (rdev->saved_raid_disk >= 0 &&
7160 rdev->saved_raid_disk >= first &&
7161 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7162 first = rdev->saved_raid_disk;
7164 for (disk = first; disk <= last; disk++) {
7165 p = conf->disks + disk;
7166 if (p->rdev == NULL) {
7167 clear_bit(In_sync, &rdev->flags);
7168 rdev->raid_disk = disk;
7170 if (rdev->saved_raid_disk != disk)
7172 rcu_assign_pointer(p->rdev, rdev);
7176 for (disk = first; disk <= last; disk++) {
7177 p = conf->disks + disk;
7178 if (test_bit(WantReplacement, &p->rdev->flags) &&
7179 p->replacement == NULL) {
7180 clear_bit(In_sync, &rdev->flags);
7181 set_bit(Replacement, &rdev->flags);
7182 rdev->raid_disk = disk;
7185 rcu_assign_pointer(p->replacement, rdev);
7190 print_raid5_conf(conf);
7194 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7196 /* no resync is happening, and there is enough space
7197 * on all devices, so we can resize.
7198 * We need to make sure resync covers any new space.
7199 * If the array is shrinking we should possibly wait until
7200 * any io in the removed space completes, but it hardly seems
7204 struct r5conf *conf = mddev->private;
7206 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7207 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7208 if (mddev->external_size &&
7209 mddev->array_sectors > newsize)
7211 if (mddev->bitmap) {
7212 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7216 md_set_array_sectors(mddev, newsize);
7217 set_capacity(mddev->gendisk, mddev->array_sectors);
7218 revalidate_disk(mddev->gendisk);
7219 if (sectors > mddev->dev_sectors &&
7220 mddev->recovery_cp > mddev->dev_sectors) {
7221 mddev->recovery_cp = mddev->dev_sectors;
7222 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7224 mddev->dev_sectors = sectors;
7225 mddev->resync_max_sectors = sectors;
7229 static int check_stripe_cache(struct mddev *mddev)
7231 /* Can only proceed if there are plenty of stripe_heads.
7232 * We need a minimum of one full stripe,, and for sensible progress
7233 * it is best to have about 4 times that.
7234 * If we require 4 times, then the default 256 4K stripe_heads will
7235 * allow for chunk sizes up to 256K, which is probably OK.
7236 * If the chunk size is greater, user-space should request more
7237 * stripe_heads first.
7239 struct r5conf *conf = mddev->private;
7240 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7241 > conf->min_nr_stripes ||
7242 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7243 > conf->min_nr_stripes) {
7244 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7246 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7253 static int check_reshape(struct mddev *mddev)
7255 struct r5conf *conf = mddev->private;
7257 if (mddev->delta_disks == 0 &&
7258 mddev->new_layout == mddev->layout &&
7259 mddev->new_chunk_sectors == mddev->chunk_sectors)
7260 return 0; /* nothing to do */
7261 if (has_failed(conf))
7263 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7264 /* We might be able to shrink, but the devices must
7265 * be made bigger first.
7266 * For raid6, 4 is the minimum size.
7267 * Otherwise 2 is the minimum
7270 if (mddev->level == 6)
7272 if (mddev->raid_disks + mddev->delta_disks < min)
7276 if (!check_stripe_cache(mddev))
7279 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7280 mddev->delta_disks > 0)
7281 if (resize_chunks(conf,
7282 conf->previous_raid_disks
7283 + max(0, mddev->delta_disks),
7284 max(mddev->new_chunk_sectors,
7285 mddev->chunk_sectors)
7288 return resize_stripes(conf, (conf->previous_raid_disks
7289 + mddev->delta_disks));
7292 static int raid5_start_reshape(struct mddev *mddev)
7294 struct r5conf *conf = mddev->private;
7295 struct md_rdev *rdev;
7297 unsigned long flags;
7299 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7302 if (!check_stripe_cache(mddev))
7305 if (has_failed(conf))
7308 rdev_for_each(rdev, mddev) {
7309 if (!test_bit(In_sync, &rdev->flags)
7310 && !test_bit(Faulty, &rdev->flags))
7314 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7315 /* Not enough devices even to make a degraded array
7320 /* Refuse to reduce size of the array. Any reductions in
7321 * array size must be through explicit setting of array_size
7324 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7325 < mddev->array_sectors) {
7326 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7327 "before number of disks\n", mdname(mddev));
7331 atomic_set(&conf->reshape_stripes, 0);
7332 spin_lock_irq(&conf->device_lock);
7333 write_seqcount_begin(&conf->gen_lock);
7334 conf->previous_raid_disks = conf->raid_disks;
7335 conf->raid_disks += mddev->delta_disks;
7336 conf->prev_chunk_sectors = conf->chunk_sectors;
7337 conf->chunk_sectors = mddev->new_chunk_sectors;
7338 conf->prev_algo = conf->algorithm;
7339 conf->algorithm = mddev->new_layout;
7341 /* Code that selects data_offset needs to see the generation update
7342 * if reshape_progress has been set - so a memory barrier needed.
7345 if (mddev->reshape_backwards)
7346 conf->reshape_progress = raid5_size(mddev, 0, 0);
7348 conf->reshape_progress = 0;
7349 conf->reshape_safe = conf->reshape_progress;
7350 write_seqcount_end(&conf->gen_lock);
7351 spin_unlock_irq(&conf->device_lock);
7353 /* Now make sure any requests that proceeded on the assumption
7354 * the reshape wasn't running - like Discard or Read - have
7357 mddev_suspend(mddev);
7358 mddev_resume(mddev);
7360 /* Add some new drives, as many as will fit.
7361 * We know there are enough to make the newly sized array work.
7362 * Don't add devices if we are reducing the number of
7363 * devices in the array. This is because it is not possible
7364 * to correctly record the "partially reconstructed" state of
7365 * such devices during the reshape and confusion could result.
7367 if (mddev->delta_disks >= 0) {
7368 rdev_for_each(rdev, mddev)
7369 if (rdev->raid_disk < 0 &&
7370 !test_bit(Faulty, &rdev->flags)) {
7371 if (raid5_add_disk(mddev, rdev) == 0) {
7373 >= conf->previous_raid_disks)
7374 set_bit(In_sync, &rdev->flags);
7376 rdev->recovery_offset = 0;
7378 if (sysfs_link_rdev(mddev, rdev))
7379 /* Failure here is OK */;
7381 } else if (rdev->raid_disk >= conf->previous_raid_disks
7382 && !test_bit(Faulty, &rdev->flags)) {
7383 /* This is a spare that was manually added */
7384 set_bit(In_sync, &rdev->flags);
7387 /* When a reshape changes the number of devices,
7388 * ->degraded is measured against the larger of the
7389 * pre and post number of devices.
7391 spin_lock_irqsave(&conf->device_lock, flags);
7392 mddev->degraded = calc_degraded(conf);
7393 spin_unlock_irqrestore(&conf->device_lock, flags);
7395 mddev->raid_disks = conf->raid_disks;
7396 mddev->reshape_position = conf->reshape_progress;
7397 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7399 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7400 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7401 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7402 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7403 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7404 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7406 if (!mddev->sync_thread) {
7407 mddev->recovery = 0;
7408 spin_lock_irq(&conf->device_lock);
7409 write_seqcount_begin(&conf->gen_lock);
7410 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7411 mddev->new_chunk_sectors =
7412 conf->chunk_sectors = conf->prev_chunk_sectors;
7413 mddev->new_layout = conf->algorithm = conf->prev_algo;
7414 rdev_for_each(rdev, mddev)
7415 rdev->new_data_offset = rdev->data_offset;
7417 conf->generation --;
7418 conf->reshape_progress = MaxSector;
7419 mddev->reshape_position = MaxSector;
7420 write_seqcount_end(&conf->gen_lock);
7421 spin_unlock_irq(&conf->device_lock);
7424 conf->reshape_checkpoint = jiffies;
7425 md_wakeup_thread(mddev->sync_thread);
7426 md_new_event(mddev);
7430 /* This is called from the reshape thread and should make any
7431 * changes needed in 'conf'
7433 static void end_reshape(struct r5conf *conf)
7436 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7437 struct md_rdev *rdev;
7439 spin_lock_irq(&conf->device_lock);
7440 conf->previous_raid_disks = conf->raid_disks;
7441 rdev_for_each(rdev, conf->mddev)
7442 rdev->data_offset = rdev->new_data_offset;
7444 conf->reshape_progress = MaxSector;
7445 conf->mddev->reshape_position = MaxSector;
7446 spin_unlock_irq(&conf->device_lock);
7447 wake_up(&conf->wait_for_overlap);
7449 /* read-ahead size must cover two whole stripes, which is
7450 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7452 if (conf->mddev->queue) {
7453 int data_disks = conf->raid_disks - conf->max_degraded;
7454 int stripe = data_disks * ((conf->chunk_sectors << 9)
7456 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7457 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7462 /* This is called from the raid5d thread with mddev_lock held.
7463 * It makes config changes to the device.
7465 static void raid5_finish_reshape(struct mddev *mddev)
7467 struct r5conf *conf = mddev->private;
7469 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7471 if (mddev->delta_disks > 0) {
7472 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7473 set_capacity(mddev->gendisk, mddev->array_sectors);
7474 revalidate_disk(mddev->gendisk);
7477 spin_lock_irq(&conf->device_lock);
7478 mddev->degraded = calc_degraded(conf);
7479 spin_unlock_irq(&conf->device_lock);
7480 for (d = conf->raid_disks ;
7481 d < conf->raid_disks - mddev->delta_disks;
7483 struct md_rdev *rdev = conf->disks[d].rdev;
7485 clear_bit(In_sync, &rdev->flags);
7486 rdev = conf->disks[d].replacement;
7488 clear_bit(In_sync, &rdev->flags);
7491 mddev->layout = conf->algorithm;
7492 mddev->chunk_sectors = conf->chunk_sectors;
7493 mddev->reshape_position = MaxSector;
7494 mddev->delta_disks = 0;
7495 mddev->reshape_backwards = 0;
7499 static void raid5_quiesce(struct mddev *mddev, int state)
7501 struct r5conf *conf = mddev->private;
7504 case 2: /* resume for a suspend */
7505 wake_up(&conf->wait_for_overlap);
7508 case 1: /* stop all writes */
7509 lock_all_device_hash_locks_irq(conf);
7510 /* '2' tells resync/reshape to pause so that all
7511 * active stripes can drain
7514 wait_event_cmd(conf->wait_for_quiescent,
7515 atomic_read(&conf->active_stripes) == 0 &&
7516 atomic_read(&conf->active_aligned_reads) == 0,
7517 unlock_all_device_hash_locks_irq(conf),
7518 lock_all_device_hash_locks_irq(conf));
7520 unlock_all_device_hash_locks_irq(conf);
7521 /* allow reshape to continue */
7522 wake_up(&conf->wait_for_overlap);
7525 case 0: /* re-enable writes */
7526 lock_all_device_hash_locks_irq(conf);
7528 wake_up(&conf->wait_for_quiescent);
7529 wake_up(&conf->wait_for_overlap);
7530 unlock_all_device_hash_locks_irq(conf);
7535 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7537 struct r0conf *raid0_conf = mddev->private;
7540 /* for raid0 takeover only one zone is supported */
7541 if (raid0_conf->nr_strip_zones > 1) {
7542 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7544 return ERR_PTR(-EINVAL);
7547 sectors = raid0_conf->strip_zone[0].zone_end;
7548 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7549 mddev->dev_sectors = sectors;
7550 mddev->new_level = level;
7551 mddev->new_layout = ALGORITHM_PARITY_N;
7552 mddev->new_chunk_sectors = mddev->chunk_sectors;
7553 mddev->raid_disks += 1;
7554 mddev->delta_disks = 1;
7555 /* make sure it will be not marked as dirty */
7556 mddev->recovery_cp = MaxSector;
7558 return setup_conf(mddev);
7561 static void *raid5_takeover_raid1(struct mddev *mddev)
7565 if (mddev->raid_disks != 2 ||
7566 mddev->degraded > 1)
7567 return ERR_PTR(-EINVAL);
7569 /* Should check if there are write-behind devices? */
7571 chunksect = 64*2; /* 64K by default */
7573 /* The array must be an exact multiple of chunksize */
7574 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7577 if ((chunksect<<9) < STRIPE_SIZE)
7578 /* array size does not allow a suitable chunk size */
7579 return ERR_PTR(-EINVAL);
7581 mddev->new_level = 5;
7582 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7583 mddev->new_chunk_sectors = chunksect;
7585 return setup_conf(mddev);
7588 static void *raid5_takeover_raid6(struct mddev *mddev)
7592 switch (mddev->layout) {
7593 case ALGORITHM_LEFT_ASYMMETRIC_6:
7594 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7596 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7597 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7599 case ALGORITHM_LEFT_SYMMETRIC_6:
7600 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7602 case ALGORITHM_RIGHT_SYMMETRIC_6:
7603 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7605 case ALGORITHM_PARITY_0_6:
7606 new_layout = ALGORITHM_PARITY_0;
7608 case ALGORITHM_PARITY_N:
7609 new_layout = ALGORITHM_PARITY_N;
7612 return ERR_PTR(-EINVAL);
7614 mddev->new_level = 5;
7615 mddev->new_layout = new_layout;
7616 mddev->delta_disks = -1;
7617 mddev->raid_disks -= 1;
7618 return setup_conf(mddev);
7621 static int raid5_check_reshape(struct mddev *mddev)
7623 /* For a 2-drive array, the layout and chunk size can be changed
7624 * immediately as not restriping is needed.
7625 * For larger arrays we record the new value - after validation
7626 * to be used by a reshape pass.
7628 struct r5conf *conf = mddev->private;
7629 int new_chunk = mddev->new_chunk_sectors;
7631 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7633 if (new_chunk > 0) {
7634 if (!is_power_of_2(new_chunk))
7636 if (new_chunk < (PAGE_SIZE>>9))
7638 if (mddev->array_sectors & (new_chunk-1))
7639 /* not factor of array size */
7643 /* They look valid */
7645 if (mddev->raid_disks == 2) {
7646 /* can make the change immediately */
7647 if (mddev->new_layout >= 0) {
7648 conf->algorithm = mddev->new_layout;
7649 mddev->layout = mddev->new_layout;
7651 if (new_chunk > 0) {
7652 conf->chunk_sectors = new_chunk ;
7653 mddev->chunk_sectors = new_chunk;
7655 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7656 md_wakeup_thread(mddev->thread);
7658 return check_reshape(mddev);
7661 static int raid6_check_reshape(struct mddev *mddev)
7663 int new_chunk = mddev->new_chunk_sectors;
7665 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7667 if (new_chunk > 0) {
7668 if (!is_power_of_2(new_chunk))
7670 if (new_chunk < (PAGE_SIZE >> 9))
7672 if (mddev->array_sectors & (new_chunk-1))
7673 /* not factor of array size */
7677 /* They look valid */
7678 return check_reshape(mddev);
7681 static void *raid5_takeover(struct mddev *mddev)
7683 /* raid5 can take over:
7684 * raid0 - if there is only one strip zone - make it a raid4 layout
7685 * raid1 - if there are two drives. We need to know the chunk size
7686 * raid4 - trivial - just use a raid4 layout.
7687 * raid6 - Providing it is a *_6 layout
7689 if (mddev->level == 0)
7690 return raid45_takeover_raid0(mddev, 5);
7691 if (mddev->level == 1)
7692 return raid5_takeover_raid1(mddev);
7693 if (mddev->level == 4) {
7694 mddev->new_layout = ALGORITHM_PARITY_N;
7695 mddev->new_level = 5;
7696 return setup_conf(mddev);
7698 if (mddev->level == 6)
7699 return raid5_takeover_raid6(mddev);
7701 return ERR_PTR(-EINVAL);
7704 static void *raid4_takeover(struct mddev *mddev)
7706 /* raid4 can take over:
7707 * raid0 - if there is only one strip zone
7708 * raid5 - if layout is right
7710 if (mddev->level == 0)
7711 return raid45_takeover_raid0(mddev, 4);
7712 if (mddev->level == 5 &&
7713 mddev->layout == ALGORITHM_PARITY_N) {
7714 mddev->new_layout = 0;
7715 mddev->new_level = 4;
7716 return setup_conf(mddev);
7718 return ERR_PTR(-EINVAL);
7721 static struct md_personality raid5_personality;
7723 static void *raid6_takeover(struct mddev *mddev)
7725 /* Currently can only take over a raid5. We map the
7726 * personality to an equivalent raid6 personality
7727 * with the Q block at the end.
7731 if (mddev->pers != &raid5_personality)
7732 return ERR_PTR(-EINVAL);
7733 if (mddev->degraded > 1)
7734 return ERR_PTR(-EINVAL);
7735 if (mddev->raid_disks > 253)
7736 return ERR_PTR(-EINVAL);
7737 if (mddev->raid_disks < 3)
7738 return ERR_PTR(-EINVAL);
7740 switch (mddev->layout) {
7741 case ALGORITHM_LEFT_ASYMMETRIC:
7742 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7744 case ALGORITHM_RIGHT_ASYMMETRIC:
7745 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7747 case ALGORITHM_LEFT_SYMMETRIC:
7748 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7750 case ALGORITHM_RIGHT_SYMMETRIC:
7751 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7753 case ALGORITHM_PARITY_0:
7754 new_layout = ALGORITHM_PARITY_0_6;
7756 case ALGORITHM_PARITY_N:
7757 new_layout = ALGORITHM_PARITY_N;
7760 return ERR_PTR(-EINVAL);
7762 mddev->new_level = 6;
7763 mddev->new_layout = new_layout;
7764 mddev->delta_disks = 1;
7765 mddev->raid_disks += 1;
7766 return setup_conf(mddev);
7769 static struct md_personality raid6_personality =
7773 .owner = THIS_MODULE,
7774 .make_request = make_request,
7778 .error_handler = error,
7779 .hot_add_disk = raid5_add_disk,
7780 .hot_remove_disk= raid5_remove_disk,
7781 .spare_active = raid5_spare_active,
7782 .sync_request = sync_request,
7783 .resize = raid5_resize,
7785 .check_reshape = raid6_check_reshape,
7786 .start_reshape = raid5_start_reshape,
7787 .finish_reshape = raid5_finish_reshape,
7788 .quiesce = raid5_quiesce,
7789 .takeover = raid6_takeover,
7790 .congested = raid5_congested,
7792 static struct md_personality raid5_personality =
7796 .owner = THIS_MODULE,
7797 .make_request = make_request,
7801 .error_handler = error,
7802 .hot_add_disk = raid5_add_disk,
7803 .hot_remove_disk= raid5_remove_disk,
7804 .spare_active = raid5_spare_active,
7805 .sync_request = sync_request,
7806 .resize = raid5_resize,
7808 .check_reshape = raid5_check_reshape,
7809 .start_reshape = raid5_start_reshape,
7810 .finish_reshape = raid5_finish_reshape,
7811 .quiesce = raid5_quiesce,
7812 .takeover = raid5_takeover,
7813 .congested = raid5_congested,
7816 static struct md_personality raid4_personality =
7820 .owner = THIS_MODULE,
7821 .make_request = make_request,
7825 .error_handler = error,
7826 .hot_add_disk = raid5_add_disk,
7827 .hot_remove_disk= raid5_remove_disk,
7828 .spare_active = raid5_spare_active,
7829 .sync_request = sync_request,
7830 .resize = raid5_resize,
7832 .check_reshape = raid5_check_reshape,
7833 .start_reshape = raid5_start_reshape,
7834 .finish_reshape = raid5_finish_reshape,
7835 .quiesce = raid5_quiesce,
7836 .takeover = raid4_takeover,
7837 .congested = raid5_congested,
7840 static int __init raid5_init(void)
7842 raid5_wq = alloc_workqueue("raid5wq",
7843 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7846 register_md_personality(&raid6_personality);
7847 register_md_personality(&raid5_personality);
7848 register_md_personality(&raid4_personality);
7852 static void raid5_exit(void)
7854 unregister_md_personality(&raid6_personality);
7855 unregister_md_personality(&raid5_personality);
7856 unregister_md_personality(&raid4_personality);
7857 destroy_workqueue(raid5_wq);
7860 module_init(raid5_init);
7861 module_exit(raid5_exit);
7862 MODULE_LICENSE("GPL");
7863 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7864 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7865 MODULE_ALIAS("md-raid5");
7866 MODULE_ALIAS("md-raid4");
7867 MODULE_ALIAS("md-level-5");
7868 MODULE_ALIAS("md-level-4");
7869 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7870 MODULE_ALIAS("md-raid6");
7871 MODULE_ALIAS("md-level-6");
7873 /* This used to be two separate modules, they were: */
7874 MODULE_ALIAS("raid5");
7875 MODULE_ALIAS("raid6");