2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
58 #include <trace/events/block.h>
65 #define cpu_to_group(cpu) cpu_to_node(cpu)
66 #define ANY_GROUP NUMA_NO_NODE
68 static bool devices_handle_discard_safely = false;
69 module_param(devices_handle_discard_safely, bool, 0644);
70 MODULE_PARM_DESC(devices_handle_discard_safely,
71 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
72 static struct workqueue_struct *raid5_wq;
77 #define NR_STRIPES 256
78 #define STRIPE_SIZE PAGE_SIZE
79 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
80 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
81 #define IO_THRESHOLD 1
82 #define BYPASS_THRESHOLD 1
83 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
84 #define HASH_MASK (NR_HASH - 1)
85 #define MAX_STRIPE_BATCH 8
87 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
89 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
90 return &conf->stripe_hashtbl[hash];
93 static inline int stripe_hash_locks_hash(sector_t sect)
95 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
98 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
100 spin_lock_irq(conf->hash_locks + hash);
101 spin_lock(&conf->device_lock);
104 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
106 spin_unlock(&conf->device_lock);
107 spin_unlock_irq(conf->hash_locks + hash);
110 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
114 spin_lock(conf->hash_locks);
115 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
116 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
117 spin_lock(&conf->device_lock);
120 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
123 spin_unlock(&conf->device_lock);
124 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
125 spin_unlock(conf->hash_locks + i - 1);
129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
130 * order without overlap. There may be several bio's per stripe+device, and
131 * a bio could span several devices.
132 * When walking this list for a particular stripe+device, we must never proceed
133 * beyond a bio that extends past this device, as the next bio might no longer
135 * This function is used to determine the 'next' bio in the list, given the sector
136 * of the current stripe+device
138 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
140 int sectors = bio_sectors(bio);
141 if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
148 * We maintain a biased count of active stripes in the bottom 16 bits of
149 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
151 static inline int raid5_bi_processed_stripes(struct bio *bio)
153 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
154 return (atomic_read(segments) >> 16) & 0xffff;
157 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
159 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
160 return atomic_sub_return(1, segments) & 0xffff;
163 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
165 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
166 atomic_inc(segments);
169 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
172 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
176 old = atomic_read(segments);
177 new = (old & 0xffff) | (cnt << 16);
178 } while (atomic_cmpxchg(segments, old, new) != old);
181 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
183 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
184 atomic_set(segments, cnt);
187 /* Find first data disk in a raid6 stripe */
188 static inline int raid6_d0(struct stripe_head *sh)
191 /* ddf always start from first device */
193 /* md starts just after Q block */
194 if (sh->qd_idx == sh->disks - 1)
197 return sh->qd_idx + 1;
199 static inline int raid6_next_disk(int disk, int raid_disks)
202 return (disk < raid_disks) ? disk : 0;
205 /* When walking through the disks in a raid5, starting at raid6_d0,
206 * We need to map each disk to a 'slot', where the data disks are slot
207 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
208 * is raid_disks-1. This help does that mapping.
210 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
211 int *count, int syndrome_disks)
217 if (idx == sh->pd_idx)
218 return syndrome_disks;
219 if (idx == sh->qd_idx)
220 return syndrome_disks + 1;
226 static void return_io(struct bio_list *return_bi)
229 while ((bi = bio_list_pop(return_bi)) != NULL) {
230 bi->bi_iter.bi_size = 0;
231 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
237 static void print_raid5_conf (struct r5conf *conf);
239 static int stripe_operations_active(struct stripe_head *sh)
241 return sh->check_state || sh->reconstruct_state ||
242 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
243 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
246 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
248 struct r5conf *conf = sh->raid_conf;
249 struct r5worker_group *group;
251 int i, cpu = sh->cpu;
253 if (!cpu_online(cpu)) {
254 cpu = cpumask_any(cpu_online_mask);
258 if (list_empty(&sh->lru)) {
259 struct r5worker_group *group;
260 group = conf->worker_groups + cpu_to_group(cpu);
261 list_add_tail(&sh->lru, &group->handle_list);
262 group->stripes_cnt++;
266 if (conf->worker_cnt_per_group == 0) {
267 md_wakeup_thread(conf->mddev->thread);
271 group = conf->worker_groups + cpu_to_group(sh->cpu);
273 group->workers[0].working = true;
274 /* at least one worker should run to avoid race */
275 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
277 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
278 /* wakeup more workers */
279 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
280 if (group->workers[i].working == false) {
281 group->workers[i].working = true;
282 queue_work_on(sh->cpu, raid5_wq,
283 &group->workers[i].work);
289 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
290 struct list_head *temp_inactive_list)
292 BUG_ON(!list_empty(&sh->lru));
293 BUG_ON(atomic_read(&conf->active_stripes)==0);
294 if (test_bit(STRIPE_HANDLE, &sh->state)) {
295 if (test_bit(STRIPE_DELAYED, &sh->state) &&
296 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
297 list_add_tail(&sh->lru, &conf->delayed_list);
298 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
299 sh->bm_seq - conf->seq_write > 0)
300 list_add_tail(&sh->lru, &conf->bitmap_list);
302 clear_bit(STRIPE_DELAYED, &sh->state);
303 clear_bit(STRIPE_BIT_DELAY, &sh->state);
304 if (conf->worker_cnt_per_group == 0) {
305 list_add_tail(&sh->lru, &conf->handle_list);
307 raid5_wakeup_stripe_thread(sh);
311 md_wakeup_thread(conf->mddev->thread);
313 BUG_ON(stripe_operations_active(sh));
314 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
315 if (atomic_dec_return(&conf->preread_active_stripes)
317 md_wakeup_thread(conf->mddev->thread);
318 atomic_dec(&conf->active_stripes);
319 if (!test_bit(STRIPE_EXPANDING, &sh->state))
320 list_add_tail(&sh->lru, temp_inactive_list);
324 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
325 struct list_head *temp_inactive_list)
327 if (atomic_dec_and_test(&sh->count))
328 do_release_stripe(conf, sh, temp_inactive_list);
332 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
334 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
335 * given time. Adding stripes only takes device lock, while deleting stripes
336 * only takes hash lock.
338 static void release_inactive_stripe_list(struct r5conf *conf,
339 struct list_head *temp_inactive_list,
343 bool do_wakeup = false;
346 if (hash == NR_STRIPE_HASH_LOCKS) {
347 size = NR_STRIPE_HASH_LOCKS;
348 hash = NR_STRIPE_HASH_LOCKS - 1;
352 struct list_head *list = &temp_inactive_list[size - 1];
355 * We don't hold any lock here yet, raid5_get_active_stripe() might
356 * remove stripes from the list
358 if (!list_empty_careful(list)) {
359 spin_lock_irqsave(conf->hash_locks + hash, flags);
360 if (list_empty(conf->inactive_list + hash) &&
362 atomic_dec(&conf->empty_inactive_list_nr);
363 list_splice_tail_init(list, conf->inactive_list + hash);
365 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
372 wake_up(&conf->wait_for_stripe);
373 if (atomic_read(&conf->active_stripes) == 0)
374 wake_up(&conf->wait_for_quiescent);
375 if (conf->retry_read_aligned)
376 md_wakeup_thread(conf->mddev->thread);
380 /* should hold conf->device_lock already */
381 static int release_stripe_list(struct r5conf *conf,
382 struct list_head *temp_inactive_list)
384 struct stripe_head *sh;
386 struct llist_node *head;
388 head = llist_del_all(&conf->released_stripes);
389 head = llist_reverse_order(head);
393 sh = llist_entry(head, struct stripe_head, release_list);
394 head = llist_next(head);
395 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
397 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
399 * Don't worry the bit is set here, because if the bit is set
400 * again, the count is always > 1. This is true for
401 * STRIPE_ON_UNPLUG_LIST bit too.
403 hash = sh->hash_lock_index;
404 __release_stripe(conf, sh, &temp_inactive_list[hash]);
411 void raid5_release_stripe(struct stripe_head *sh)
413 struct r5conf *conf = sh->raid_conf;
415 struct list_head list;
419 /* Avoid release_list until the last reference.
421 if (atomic_add_unless(&sh->count, -1, 1))
424 if (unlikely(!conf->mddev->thread) ||
425 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
427 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
429 md_wakeup_thread(conf->mddev->thread);
432 local_irq_save(flags);
433 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
434 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
435 INIT_LIST_HEAD(&list);
436 hash = sh->hash_lock_index;
437 do_release_stripe(conf, sh, &list);
438 spin_unlock(&conf->device_lock);
439 release_inactive_stripe_list(conf, &list, hash);
441 local_irq_restore(flags);
444 static inline void remove_hash(struct stripe_head *sh)
446 pr_debug("remove_hash(), stripe %llu\n",
447 (unsigned long long)sh->sector);
449 hlist_del_init(&sh->hash);
452 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
454 struct hlist_head *hp = stripe_hash(conf, sh->sector);
456 pr_debug("insert_hash(), stripe %llu\n",
457 (unsigned long long)sh->sector);
459 hlist_add_head(&sh->hash, hp);
462 /* find an idle stripe, make sure it is unhashed, and return it. */
463 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
465 struct stripe_head *sh = NULL;
466 struct list_head *first;
468 if (list_empty(conf->inactive_list + hash))
470 first = (conf->inactive_list + hash)->next;
471 sh = list_entry(first, struct stripe_head, lru);
472 list_del_init(first);
474 atomic_inc(&conf->active_stripes);
475 BUG_ON(hash != sh->hash_lock_index);
476 if (list_empty(conf->inactive_list + hash))
477 atomic_inc(&conf->empty_inactive_list_nr);
482 static void shrink_buffers(struct stripe_head *sh)
486 int num = sh->raid_conf->pool_size;
488 for (i = 0; i < num ; i++) {
489 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
493 sh->dev[i].page = NULL;
498 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
501 int num = sh->raid_conf->pool_size;
503 for (i = 0; i < num; i++) {
506 if (!(page = alloc_page(gfp))) {
509 sh->dev[i].page = page;
510 sh->dev[i].orig_page = page;
515 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
516 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
517 struct stripe_head *sh);
519 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
521 struct r5conf *conf = sh->raid_conf;
524 BUG_ON(atomic_read(&sh->count) != 0);
525 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
526 BUG_ON(stripe_operations_active(sh));
527 BUG_ON(sh->batch_head);
529 pr_debug("init_stripe called, stripe %llu\n",
530 (unsigned long long)sector);
532 seq = read_seqcount_begin(&conf->gen_lock);
533 sh->generation = conf->generation - previous;
534 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
536 stripe_set_idx(sector, conf, previous, sh);
539 for (i = sh->disks; i--; ) {
540 struct r5dev *dev = &sh->dev[i];
542 if (dev->toread || dev->read || dev->towrite || dev->written ||
543 test_bit(R5_LOCKED, &dev->flags)) {
544 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
545 (unsigned long long)sh->sector, i, dev->toread,
546 dev->read, dev->towrite, dev->written,
547 test_bit(R5_LOCKED, &dev->flags));
551 raid5_build_block(sh, i, previous);
553 if (read_seqcount_retry(&conf->gen_lock, seq))
555 sh->overwrite_disks = 0;
556 insert_hash(conf, sh);
557 sh->cpu = smp_processor_id();
558 set_bit(STRIPE_BATCH_READY, &sh->state);
561 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
564 struct stripe_head *sh;
566 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
567 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
568 if (sh->sector == sector && sh->generation == generation)
570 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
575 * Need to check if array has failed when deciding whether to:
577 * - remove non-faulty devices
580 * This determination is simple when no reshape is happening.
581 * However if there is a reshape, we need to carefully check
582 * both the before and after sections.
583 * This is because some failed devices may only affect one
584 * of the two sections, and some non-in_sync devices may
585 * be insync in the section most affected by failed devices.
587 static int calc_degraded(struct r5conf *conf)
589 int degraded, degraded2;
594 for (i = 0; i < conf->previous_raid_disks; i++) {
595 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
596 if (rdev && test_bit(Faulty, &rdev->flags))
597 rdev = rcu_dereference(conf->disks[i].replacement);
598 if (!rdev || test_bit(Faulty, &rdev->flags))
600 else if (test_bit(In_sync, &rdev->flags))
603 /* not in-sync or faulty.
604 * If the reshape increases the number of devices,
605 * this is being recovered by the reshape, so
606 * this 'previous' section is not in_sync.
607 * If the number of devices is being reduced however,
608 * the device can only be part of the array if
609 * we are reverting a reshape, so this section will
612 if (conf->raid_disks >= conf->previous_raid_disks)
616 if (conf->raid_disks == conf->previous_raid_disks)
620 for (i = 0; i < conf->raid_disks; i++) {
621 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
622 if (rdev && test_bit(Faulty, &rdev->flags))
623 rdev = rcu_dereference(conf->disks[i].replacement);
624 if (!rdev || test_bit(Faulty, &rdev->flags))
626 else if (test_bit(In_sync, &rdev->flags))
629 /* not in-sync or faulty.
630 * If reshape increases the number of devices, this
631 * section has already been recovered, else it
632 * almost certainly hasn't.
634 if (conf->raid_disks <= conf->previous_raid_disks)
638 if (degraded2 > degraded)
643 static int has_failed(struct r5conf *conf)
647 if (conf->mddev->reshape_position == MaxSector)
648 return conf->mddev->degraded > conf->max_degraded;
650 degraded = calc_degraded(conf);
651 if (degraded > conf->max_degraded)
657 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
658 int previous, int noblock, int noquiesce)
660 struct stripe_head *sh;
661 int hash = stripe_hash_locks_hash(sector);
663 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
665 spin_lock_irq(conf->hash_locks + hash);
668 wait_event_lock_irq(conf->wait_for_quiescent,
669 conf->quiesce == 0 || noquiesce,
670 *(conf->hash_locks + hash));
671 sh = __find_stripe(conf, sector, conf->generation - previous);
673 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
674 sh = get_free_stripe(conf, hash);
675 if (!sh && !test_bit(R5_DID_ALLOC,
677 set_bit(R5_ALLOC_MORE,
680 if (noblock && sh == NULL)
683 set_bit(R5_INACTIVE_BLOCKED,
686 conf->wait_for_stripe,
687 !list_empty(conf->inactive_list + hash) &&
688 (atomic_read(&conf->active_stripes)
689 < (conf->max_nr_stripes * 3 / 4)
690 || !test_bit(R5_INACTIVE_BLOCKED,
691 &conf->cache_state)),
692 *(conf->hash_locks + hash));
693 clear_bit(R5_INACTIVE_BLOCKED,
696 init_stripe(sh, sector, previous);
697 atomic_inc(&sh->count);
699 } else if (!atomic_inc_not_zero(&sh->count)) {
700 spin_lock(&conf->device_lock);
701 if (!atomic_read(&sh->count)) {
702 if (!test_bit(STRIPE_HANDLE, &sh->state))
703 atomic_inc(&conf->active_stripes);
704 BUG_ON(list_empty(&sh->lru) &&
705 !test_bit(STRIPE_EXPANDING, &sh->state));
706 list_del_init(&sh->lru);
708 sh->group->stripes_cnt--;
712 atomic_inc(&sh->count);
713 spin_unlock(&conf->device_lock);
715 } while (sh == NULL);
717 spin_unlock_irq(conf->hash_locks + hash);
721 static bool is_full_stripe_write(struct stripe_head *sh)
723 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
724 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
727 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
731 spin_lock(&sh2->stripe_lock);
732 spin_lock_nested(&sh1->stripe_lock, 1);
734 spin_lock(&sh1->stripe_lock);
735 spin_lock_nested(&sh2->stripe_lock, 1);
739 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
741 spin_unlock(&sh1->stripe_lock);
742 spin_unlock(&sh2->stripe_lock);
746 /* Only freshly new full stripe normal write stripe can be added to a batch list */
747 static bool stripe_can_batch(struct stripe_head *sh)
749 struct r5conf *conf = sh->raid_conf;
753 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
754 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
755 is_full_stripe_write(sh);
758 /* we only do back search */
759 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
761 struct stripe_head *head;
762 sector_t head_sector, tmp_sec;
766 if (!stripe_can_batch(sh))
768 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
769 tmp_sec = sh->sector;
770 if (!sector_div(tmp_sec, conf->chunk_sectors))
772 head_sector = sh->sector - STRIPE_SECTORS;
774 hash = stripe_hash_locks_hash(head_sector);
775 spin_lock_irq(conf->hash_locks + hash);
776 head = __find_stripe(conf, head_sector, conf->generation);
777 if (head && !atomic_inc_not_zero(&head->count)) {
778 spin_lock(&conf->device_lock);
779 if (!atomic_read(&head->count)) {
780 if (!test_bit(STRIPE_HANDLE, &head->state))
781 atomic_inc(&conf->active_stripes);
782 BUG_ON(list_empty(&head->lru) &&
783 !test_bit(STRIPE_EXPANDING, &head->state));
784 list_del_init(&head->lru);
786 head->group->stripes_cnt--;
790 atomic_inc(&head->count);
791 spin_unlock(&conf->device_lock);
793 spin_unlock_irq(conf->hash_locks + hash);
797 if (!stripe_can_batch(head))
800 lock_two_stripes(head, sh);
801 /* clear_batch_ready clear the flag */
802 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
809 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
811 if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
814 if (head->batch_head) {
815 spin_lock(&head->batch_head->batch_lock);
816 /* This batch list is already running */
817 if (!stripe_can_batch(head)) {
818 spin_unlock(&head->batch_head->batch_lock);
823 * at this point, head's BATCH_READY could be cleared, but we
824 * can still add the stripe to batch list
826 list_add(&sh->batch_list, &head->batch_list);
827 spin_unlock(&head->batch_head->batch_lock);
829 sh->batch_head = head->batch_head;
831 head->batch_head = head;
832 sh->batch_head = head->batch_head;
833 spin_lock(&head->batch_lock);
834 list_add_tail(&sh->batch_list, &head->batch_list);
835 spin_unlock(&head->batch_lock);
838 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
839 if (atomic_dec_return(&conf->preread_active_stripes)
841 md_wakeup_thread(conf->mddev->thread);
843 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
844 int seq = sh->bm_seq;
845 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
846 sh->batch_head->bm_seq > seq)
847 seq = sh->batch_head->bm_seq;
848 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
849 sh->batch_head->bm_seq = seq;
852 atomic_inc(&sh->count);
854 unlock_two_stripes(head, sh);
856 raid5_release_stripe(head);
859 /* Determine if 'data_offset' or 'new_data_offset' should be used
860 * in this stripe_head.
862 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
864 sector_t progress = conf->reshape_progress;
865 /* Need a memory barrier to make sure we see the value
866 * of conf->generation, or ->data_offset that was set before
867 * reshape_progress was updated.
870 if (progress == MaxSector)
872 if (sh->generation == conf->generation - 1)
874 /* We are in a reshape, and this is a new-generation stripe,
875 * so use new_data_offset.
881 raid5_end_read_request(struct bio *bi);
883 raid5_end_write_request(struct bio *bi);
885 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
887 struct r5conf *conf = sh->raid_conf;
888 int i, disks = sh->disks;
889 struct stripe_head *head_sh = sh;
893 if (r5l_write_stripe(conf->log, sh) == 0)
895 for (i = disks; i--; ) {
897 int replace_only = 0;
898 struct bio *bi, *rbi;
899 struct md_rdev *rdev, *rrdev = NULL;
902 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
903 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
907 if (test_bit(R5_Discard, &sh->dev[i].flags))
909 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
911 else if (test_and_clear_bit(R5_WantReplace,
912 &sh->dev[i].flags)) {
917 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
921 bi = &sh->dev[i].req;
922 rbi = &sh->dev[i].rreq; /* For writing to replacement */
925 rrdev = rcu_dereference(conf->disks[i].replacement);
926 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
927 rdev = rcu_dereference(conf->disks[i].rdev);
936 /* We raced and saw duplicates */
939 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
944 if (rdev && test_bit(Faulty, &rdev->flags))
947 atomic_inc(&rdev->nr_pending);
948 if (rrdev && test_bit(Faulty, &rrdev->flags))
951 atomic_inc(&rrdev->nr_pending);
954 /* We have already checked bad blocks for reads. Now
955 * need to check for writes. We never accept write errors
956 * on the replacement, so we don't to check rrdev.
958 while ((rw & WRITE) && rdev &&
959 test_bit(WriteErrorSeen, &rdev->flags)) {
962 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
963 &first_bad, &bad_sectors);
968 set_bit(BlockedBadBlocks, &rdev->flags);
969 if (!conf->mddev->external &&
970 conf->mddev->flags) {
971 /* It is very unlikely, but we might
972 * still need to write out the
973 * bad block log - better give it
975 md_check_recovery(conf->mddev);
978 * Because md_wait_for_blocked_rdev
979 * will dec nr_pending, we must
980 * increment it first.
982 atomic_inc(&rdev->nr_pending);
983 md_wait_for_blocked_rdev(rdev, conf->mddev);
985 /* Acknowledged bad block - skip the write */
986 rdev_dec_pending(rdev, conf->mddev);
992 if (s->syncing || s->expanding || s->expanded
994 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
996 set_bit(STRIPE_IO_STARTED, &sh->state);
999 bi->bi_bdev = rdev->bdev;
1001 bi->bi_end_io = (rw & WRITE)
1002 ? raid5_end_write_request
1003 : raid5_end_read_request;
1004 bi->bi_private = sh;
1006 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1007 __func__, (unsigned long long)sh->sector,
1009 atomic_inc(&sh->count);
1011 atomic_inc(&head_sh->count);
1012 if (use_new_offset(conf, sh))
1013 bi->bi_iter.bi_sector = (sh->sector
1014 + rdev->new_data_offset);
1016 bi->bi_iter.bi_sector = (sh->sector
1017 + rdev->data_offset);
1018 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1019 bi->bi_rw |= REQ_NOMERGE;
1021 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1022 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1023 sh->dev[i].vec.bv_page = sh->dev[i].page;
1025 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1026 bi->bi_io_vec[0].bv_offset = 0;
1027 bi->bi_iter.bi_size = STRIPE_SIZE;
1029 * If this is discard request, set bi_vcnt 0. We don't
1030 * want to confuse SCSI because SCSI will replace payload
1032 if (rw & REQ_DISCARD)
1035 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1037 if (conf->mddev->gendisk)
1038 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1039 bi, disk_devt(conf->mddev->gendisk),
1041 generic_make_request(bi);
1044 if (s->syncing || s->expanding || s->expanded
1046 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1048 set_bit(STRIPE_IO_STARTED, &sh->state);
1051 rbi->bi_bdev = rrdev->bdev;
1053 BUG_ON(!(rw & WRITE));
1054 rbi->bi_end_io = raid5_end_write_request;
1055 rbi->bi_private = sh;
1057 pr_debug("%s: for %llu schedule op %ld on "
1058 "replacement disc %d\n",
1059 __func__, (unsigned long long)sh->sector,
1061 atomic_inc(&sh->count);
1063 atomic_inc(&head_sh->count);
1064 if (use_new_offset(conf, sh))
1065 rbi->bi_iter.bi_sector = (sh->sector
1066 + rrdev->new_data_offset);
1068 rbi->bi_iter.bi_sector = (sh->sector
1069 + rrdev->data_offset);
1070 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1071 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1072 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1074 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1075 rbi->bi_io_vec[0].bv_offset = 0;
1076 rbi->bi_iter.bi_size = STRIPE_SIZE;
1078 * If this is discard request, set bi_vcnt 0. We don't
1079 * want to confuse SCSI because SCSI will replace payload
1081 if (rw & REQ_DISCARD)
1083 if (conf->mddev->gendisk)
1084 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1085 rbi, disk_devt(conf->mddev->gendisk),
1087 generic_make_request(rbi);
1089 if (!rdev && !rrdev) {
1091 set_bit(STRIPE_DEGRADED, &sh->state);
1092 pr_debug("skip op %ld on disc %d for sector %llu\n",
1093 bi->bi_rw, i, (unsigned long long)sh->sector);
1094 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1095 set_bit(STRIPE_HANDLE, &sh->state);
1098 if (!head_sh->batch_head)
1100 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1107 static struct dma_async_tx_descriptor *
1108 async_copy_data(int frombio, struct bio *bio, struct page **page,
1109 sector_t sector, struct dma_async_tx_descriptor *tx,
1110 struct stripe_head *sh)
1113 struct bvec_iter iter;
1114 struct page *bio_page;
1116 struct async_submit_ctl submit;
1117 enum async_tx_flags flags = 0;
1119 if (bio->bi_iter.bi_sector >= sector)
1120 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1122 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1125 flags |= ASYNC_TX_FENCE;
1126 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1128 bio_for_each_segment(bvl, bio, iter) {
1129 int len = bvl.bv_len;
1133 if (page_offset < 0) {
1134 b_offset = -page_offset;
1135 page_offset += b_offset;
1139 if (len > 0 && page_offset + len > STRIPE_SIZE)
1140 clen = STRIPE_SIZE - page_offset;
1145 b_offset += bvl.bv_offset;
1146 bio_page = bvl.bv_page;
1148 if (sh->raid_conf->skip_copy &&
1149 b_offset == 0 && page_offset == 0 &&
1150 clen == STRIPE_SIZE)
1153 tx = async_memcpy(*page, bio_page, page_offset,
1154 b_offset, clen, &submit);
1156 tx = async_memcpy(bio_page, *page, b_offset,
1157 page_offset, clen, &submit);
1159 /* chain the operations */
1160 submit.depend_tx = tx;
1162 if (clen < len) /* hit end of page */
1170 static void ops_complete_biofill(void *stripe_head_ref)
1172 struct stripe_head *sh = stripe_head_ref;
1173 struct bio_list return_bi = BIO_EMPTY_LIST;
1176 pr_debug("%s: stripe %llu\n", __func__,
1177 (unsigned long long)sh->sector);
1179 /* clear completed biofills */
1180 for (i = sh->disks; i--; ) {
1181 struct r5dev *dev = &sh->dev[i];
1183 /* acknowledge completion of a biofill operation */
1184 /* and check if we need to reply to a read request,
1185 * new R5_Wantfill requests are held off until
1186 * !STRIPE_BIOFILL_RUN
1188 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1189 struct bio *rbi, *rbi2;
1194 while (rbi && rbi->bi_iter.bi_sector <
1195 dev->sector + STRIPE_SECTORS) {
1196 rbi2 = r5_next_bio(rbi, dev->sector);
1197 if (!raid5_dec_bi_active_stripes(rbi))
1198 bio_list_add(&return_bi, rbi);
1203 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1205 return_io(&return_bi);
1207 set_bit(STRIPE_HANDLE, &sh->state);
1208 raid5_release_stripe(sh);
1211 static void ops_run_biofill(struct stripe_head *sh)
1213 struct dma_async_tx_descriptor *tx = NULL;
1214 struct async_submit_ctl submit;
1217 BUG_ON(sh->batch_head);
1218 pr_debug("%s: stripe %llu\n", __func__,
1219 (unsigned long long)sh->sector);
1221 for (i = sh->disks; i--; ) {
1222 struct r5dev *dev = &sh->dev[i];
1223 if (test_bit(R5_Wantfill, &dev->flags)) {
1225 spin_lock_irq(&sh->stripe_lock);
1226 dev->read = rbi = dev->toread;
1228 spin_unlock_irq(&sh->stripe_lock);
1229 while (rbi && rbi->bi_iter.bi_sector <
1230 dev->sector + STRIPE_SECTORS) {
1231 tx = async_copy_data(0, rbi, &dev->page,
1232 dev->sector, tx, sh);
1233 rbi = r5_next_bio(rbi, dev->sector);
1238 atomic_inc(&sh->count);
1239 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1240 async_trigger_callback(&submit);
1243 static void mark_target_uptodate(struct stripe_head *sh, int target)
1250 tgt = &sh->dev[target];
1251 set_bit(R5_UPTODATE, &tgt->flags);
1252 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1253 clear_bit(R5_Wantcompute, &tgt->flags);
1256 static void ops_complete_compute(void *stripe_head_ref)
1258 struct stripe_head *sh = stripe_head_ref;
1260 pr_debug("%s: stripe %llu\n", __func__,
1261 (unsigned long long)sh->sector);
1263 /* mark the computed target(s) as uptodate */
1264 mark_target_uptodate(sh, sh->ops.target);
1265 mark_target_uptodate(sh, sh->ops.target2);
1267 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1268 if (sh->check_state == check_state_compute_run)
1269 sh->check_state = check_state_compute_result;
1270 set_bit(STRIPE_HANDLE, &sh->state);
1271 raid5_release_stripe(sh);
1274 /* return a pointer to the address conversion region of the scribble buffer */
1275 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1276 struct raid5_percpu *percpu, int i)
1280 addr = flex_array_get(percpu->scribble, i);
1281 return addr + sizeof(struct page *) * (sh->disks + 2);
1284 /* return a pointer to the address conversion region of the scribble buffer */
1285 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1289 addr = flex_array_get(percpu->scribble, i);
1293 static struct dma_async_tx_descriptor *
1294 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1296 int disks = sh->disks;
1297 struct page **xor_srcs = to_addr_page(percpu, 0);
1298 int target = sh->ops.target;
1299 struct r5dev *tgt = &sh->dev[target];
1300 struct page *xor_dest = tgt->page;
1302 struct dma_async_tx_descriptor *tx;
1303 struct async_submit_ctl submit;
1306 BUG_ON(sh->batch_head);
1308 pr_debug("%s: stripe %llu block: %d\n",
1309 __func__, (unsigned long long)sh->sector, target);
1310 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1312 for (i = disks; i--; )
1314 xor_srcs[count++] = sh->dev[i].page;
1316 atomic_inc(&sh->count);
1318 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1319 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1320 if (unlikely(count == 1))
1321 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1323 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1328 /* set_syndrome_sources - populate source buffers for gen_syndrome
1329 * @srcs - (struct page *) array of size sh->disks
1330 * @sh - stripe_head to parse
1332 * Populates srcs in proper layout order for the stripe and returns the
1333 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1334 * destination buffer is recorded in srcs[count] and the Q destination
1335 * is recorded in srcs[count+1]].
1337 static int set_syndrome_sources(struct page **srcs,
1338 struct stripe_head *sh,
1341 int disks = sh->disks;
1342 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1343 int d0_idx = raid6_d0(sh);
1347 for (i = 0; i < disks; i++)
1353 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1354 struct r5dev *dev = &sh->dev[i];
1356 if (i == sh->qd_idx || i == sh->pd_idx ||
1357 (srctype == SYNDROME_SRC_ALL) ||
1358 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1359 test_bit(R5_Wantdrain, &dev->flags)) ||
1360 (srctype == SYNDROME_SRC_WRITTEN &&
1362 srcs[slot] = sh->dev[i].page;
1363 i = raid6_next_disk(i, disks);
1364 } while (i != d0_idx);
1366 return syndrome_disks;
1369 static struct dma_async_tx_descriptor *
1370 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1372 int disks = sh->disks;
1373 struct page **blocks = to_addr_page(percpu, 0);
1375 int qd_idx = sh->qd_idx;
1376 struct dma_async_tx_descriptor *tx;
1377 struct async_submit_ctl submit;
1383 BUG_ON(sh->batch_head);
1384 if (sh->ops.target < 0)
1385 target = sh->ops.target2;
1386 else if (sh->ops.target2 < 0)
1387 target = sh->ops.target;
1389 /* we should only have one valid target */
1392 pr_debug("%s: stripe %llu block: %d\n",
1393 __func__, (unsigned long long)sh->sector, target);
1395 tgt = &sh->dev[target];
1396 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1399 atomic_inc(&sh->count);
1401 if (target == qd_idx) {
1402 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1403 blocks[count] = NULL; /* regenerating p is not necessary */
1404 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1405 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1406 ops_complete_compute, sh,
1407 to_addr_conv(sh, percpu, 0));
1408 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1410 /* Compute any data- or p-drive using XOR */
1412 for (i = disks; i-- ; ) {
1413 if (i == target || i == qd_idx)
1415 blocks[count++] = sh->dev[i].page;
1418 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1419 NULL, ops_complete_compute, sh,
1420 to_addr_conv(sh, percpu, 0));
1421 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1427 static struct dma_async_tx_descriptor *
1428 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1430 int i, count, disks = sh->disks;
1431 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1432 int d0_idx = raid6_d0(sh);
1433 int faila = -1, failb = -1;
1434 int target = sh->ops.target;
1435 int target2 = sh->ops.target2;
1436 struct r5dev *tgt = &sh->dev[target];
1437 struct r5dev *tgt2 = &sh->dev[target2];
1438 struct dma_async_tx_descriptor *tx;
1439 struct page **blocks = to_addr_page(percpu, 0);
1440 struct async_submit_ctl submit;
1442 BUG_ON(sh->batch_head);
1443 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1444 __func__, (unsigned long long)sh->sector, target, target2);
1445 BUG_ON(target < 0 || target2 < 0);
1446 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1447 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1449 /* we need to open-code set_syndrome_sources to handle the
1450 * slot number conversion for 'faila' and 'failb'
1452 for (i = 0; i < disks ; i++)
1457 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1459 blocks[slot] = sh->dev[i].page;
1465 i = raid6_next_disk(i, disks);
1466 } while (i != d0_idx);
1468 BUG_ON(faila == failb);
1471 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1472 __func__, (unsigned long long)sh->sector, faila, failb);
1474 atomic_inc(&sh->count);
1476 if (failb == syndrome_disks+1) {
1477 /* Q disk is one of the missing disks */
1478 if (faila == syndrome_disks) {
1479 /* Missing P+Q, just recompute */
1480 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1481 ops_complete_compute, sh,
1482 to_addr_conv(sh, percpu, 0));
1483 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1484 STRIPE_SIZE, &submit);
1488 int qd_idx = sh->qd_idx;
1490 /* Missing D+Q: recompute D from P, then recompute Q */
1491 if (target == qd_idx)
1492 data_target = target2;
1494 data_target = target;
1497 for (i = disks; i-- ; ) {
1498 if (i == data_target || i == qd_idx)
1500 blocks[count++] = sh->dev[i].page;
1502 dest = sh->dev[data_target].page;
1503 init_async_submit(&submit,
1504 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1506 to_addr_conv(sh, percpu, 0));
1507 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1510 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1511 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1512 ops_complete_compute, sh,
1513 to_addr_conv(sh, percpu, 0));
1514 return async_gen_syndrome(blocks, 0, count+2,
1515 STRIPE_SIZE, &submit);
1518 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1519 ops_complete_compute, sh,
1520 to_addr_conv(sh, percpu, 0));
1521 if (failb == syndrome_disks) {
1522 /* We're missing D+P. */
1523 return async_raid6_datap_recov(syndrome_disks+2,
1527 /* We're missing D+D. */
1528 return async_raid6_2data_recov(syndrome_disks+2,
1529 STRIPE_SIZE, faila, failb,
1535 static void ops_complete_prexor(void *stripe_head_ref)
1537 struct stripe_head *sh = stripe_head_ref;
1539 pr_debug("%s: stripe %llu\n", __func__,
1540 (unsigned long long)sh->sector);
1543 static struct dma_async_tx_descriptor *
1544 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1545 struct dma_async_tx_descriptor *tx)
1547 int disks = sh->disks;
1548 struct page **xor_srcs = to_addr_page(percpu, 0);
1549 int count = 0, pd_idx = sh->pd_idx, i;
1550 struct async_submit_ctl submit;
1552 /* existing parity data subtracted */
1553 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1555 BUG_ON(sh->batch_head);
1556 pr_debug("%s: stripe %llu\n", __func__,
1557 (unsigned long long)sh->sector);
1559 for (i = disks; i--; ) {
1560 struct r5dev *dev = &sh->dev[i];
1561 /* Only process blocks that are known to be uptodate */
1562 if (test_bit(R5_Wantdrain, &dev->flags))
1563 xor_srcs[count++] = dev->page;
1566 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1567 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1568 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1573 static struct dma_async_tx_descriptor *
1574 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1575 struct dma_async_tx_descriptor *tx)
1577 struct page **blocks = to_addr_page(percpu, 0);
1579 struct async_submit_ctl submit;
1581 pr_debug("%s: stripe %llu\n", __func__,
1582 (unsigned long long)sh->sector);
1584 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1586 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1587 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1588 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1593 static struct dma_async_tx_descriptor *
1594 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1596 int disks = sh->disks;
1598 struct stripe_head *head_sh = sh;
1600 pr_debug("%s: stripe %llu\n", __func__,
1601 (unsigned long long)sh->sector);
1603 for (i = disks; i--; ) {
1608 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1613 spin_lock_irq(&sh->stripe_lock);
1614 chosen = dev->towrite;
1615 dev->towrite = NULL;
1616 sh->overwrite_disks = 0;
1617 BUG_ON(dev->written);
1618 wbi = dev->written = chosen;
1619 spin_unlock_irq(&sh->stripe_lock);
1620 WARN_ON(dev->page != dev->orig_page);
1622 while (wbi && wbi->bi_iter.bi_sector <
1623 dev->sector + STRIPE_SECTORS) {
1624 if (wbi->bi_rw & REQ_FUA)
1625 set_bit(R5_WantFUA, &dev->flags);
1626 if (wbi->bi_rw & REQ_SYNC)
1627 set_bit(R5_SyncIO, &dev->flags);
1628 if (wbi->bi_rw & REQ_DISCARD)
1629 set_bit(R5_Discard, &dev->flags);
1631 tx = async_copy_data(1, wbi, &dev->page,
1632 dev->sector, tx, sh);
1633 if (dev->page != dev->orig_page) {
1634 set_bit(R5_SkipCopy, &dev->flags);
1635 clear_bit(R5_UPTODATE, &dev->flags);
1636 clear_bit(R5_OVERWRITE, &dev->flags);
1639 wbi = r5_next_bio(wbi, dev->sector);
1642 if (head_sh->batch_head) {
1643 sh = list_first_entry(&sh->batch_list,
1656 static void ops_complete_reconstruct(void *stripe_head_ref)
1658 struct stripe_head *sh = stripe_head_ref;
1659 int disks = sh->disks;
1660 int pd_idx = sh->pd_idx;
1661 int qd_idx = sh->qd_idx;
1663 bool fua = false, sync = false, discard = false;
1665 pr_debug("%s: stripe %llu\n", __func__,
1666 (unsigned long long)sh->sector);
1668 for (i = disks; i--; ) {
1669 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1670 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1671 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1674 for (i = disks; i--; ) {
1675 struct r5dev *dev = &sh->dev[i];
1677 if (dev->written || i == pd_idx || i == qd_idx) {
1678 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1679 set_bit(R5_UPTODATE, &dev->flags);
1681 set_bit(R5_WantFUA, &dev->flags);
1683 set_bit(R5_SyncIO, &dev->flags);
1687 if (sh->reconstruct_state == reconstruct_state_drain_run)
1688 sh->reconstruct_state = reconstruct_state_drain_result;
1689 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1690 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1692 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1693 sh->reconstruct_state = reconstruct_state_result;
1696 set_bit(STRIPE_HANDLE, &sh->state);
1697 raid5_release_stripe(sh);
1701 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1702 struct dma_async_tx_descriptor *tx)
1704 int disks = sh->disks;
1705 struct page **xor_srcs;
1706 struct async_submit_ctl submit;
1707 int count, pd_idx = sh->pd_idx, i;
1708 struct page *xor_dest;
1710 unsigned long flags;
1712 struct stripe_head *head_sh = sh;
1715 pr_debug("%s: stripe %llu\n", __func__,
1716 (unsigned long long)sh->sector);
1718 for (i = 0; i < sh->disks; i++) {
1721 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1724 if (i >= sh->disks) {
1725 atomic_inc(&sh->count);
1726 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1727 ops_complete_reconstruct(sh);
1732 xor_srcs = to_addr_page(percpu, j);
1733 /* check if prexor is active which means only process blocks
1734 * that are part of a read-modify-write (written)
1736 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1738 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1739 for (i = disks; i--; ) {
1740 struct r5dev *dev = &sh->dev[i];
1741 if (head_sh->dev[i].written)
1742 xor_srcs[count++] = dev->page;
1745 xor_dest = sh->dev[pd_idx].page;
1746 for (i = disks; i--; ) {
1747 struct r5dev *dev = &sh->dev[i];
1749 xor_srcs[count++] = dev->page;
1753 /* 1/ if we prexor'd then the dest is reused as a source
1754 * 2/ if we did not prexor then we are redoing the parity
1755 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1756 * for the synchronous xor case
1758 last_stripe = !head_sh->batch_head ||
1759 list_first_entry(&sh->batch_list,
1760 struct stripe_head, batch_list) == head_sh;
1762 flags = ASYNC_TX_ACK |
1763 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1765 atomic_inc(&head_sh->count);
1766 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1767 to_addr_conv(sh, percpu, j));
1769 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1770 init_async_submit(&submit, flags, tx, NULL, NULL,
1771 to_addr_conv(sh, percpu, j));
1774 if (unlikely(count == 1))
1775 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1777 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1780 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1787 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1788 struct dma_async_tx_descriptor *tx)
1790 struct async_submit_ctl submit;
1791 struct page **blocks;
1792 int count, i, j = 0;
1793 struct stripe_head *head_sh = sh;
1796 unsigned long txflags;
1798 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1800 for (i = 0; i < sh->disks; i++) {
1801 if (sh->pd_idx == i || sh->qd_idx == i)
1803 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1806 if (i >= sh->disks) {
1807 atomic_inc(&sh->count);
1808 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1809 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1810 ops_complete_reconstruct(sh);
1815 blocks = to_addr_page(percpu, j);
1817 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1818 synflags = SYNDROME_SRC_WRITTEN;
1819 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1821 synflags = SYNDROME_SRC_ALL;
1822 txflags = ASYNC_TX_ACK;
1825 count = set_syndrome_sources(blocks, sh, synflags);
1826 last_stripe = !head_sh->batch_head ||
1827 list_first_entry(&sh->batch_list,
1828 struct stripe_head, batch_list) == head_sh;
1831 atomic_inc(&head_sh->count);
1832 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1833 head_sh, to_addr_conv(sh, percpu, j));
1835 init_async_submit(&submit, 0, tx, NULL, NULL,
1836 to_addr_conv(sh, percpu, j));
1837 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1840 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1846 static void ops_complete_check(void *stripe_head_ref)
1848 struct stripe_head *sh = stripe_head_ref;
1850 pr_debug("%s: stripe %llu\n", __func__,
1851 (unsigned long long)sh->sector);
1853 sh->check_state = check_state_check_result;
1854 set_bit(STRIPE_HANDLE, &sh->state);
1855 raid5_release_stripe(sh);
1858 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1860 int disks = sh->disks;
1861 int pd_idx = sh->pd_idx;
1862 int qd_idx = sh->qd_idx;
1863 struct page *xor_dest;
1864 struct page **xor_srcs = to_addr_page(percpu, 0);
1865 struct dma_async_tx_descriptor *tx;
1866 struct async_submit_ctl submit;
1870 pr_debug("%s: stripe %llu\n", __func__,
1871 (unsigned long long)sh->sector);
1873 BUG_ON(sh->batch_head);
1875 xor_dest = sh->dev[pd_idx].page;
1876 xor_srcs[count++] = xor_dest;
1877 for (i = disks; i--; ) {
1878 if (i == pd_idx || i == qd_idx)
1880 xor_srcs[count++] = sh->dev[i].page;
1883 init_async_submit(&submit, 0, NULL, NULL, NULL,
1884 to_addr_conv(sh, percpu, 0));
1885 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1886 &sh->ops.zero_sum_result, &submit);
1888 atomic_inc(&sh->count);
1889 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1890 tx = async_trigger_callback(&submit);
1893 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1895 struct page **srcs = to_addr_page(percpu, 0);
1896 struct async_submit_ctl submit;
1899 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1900 (unsigned long long)sh->sector, checkp);
1902 BUG_ON(sh->batch_head);
1903 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1907 atomic_inc(&sh->count);
1908 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1909 sh, to_addr_conv(sh, percpu, 0));
1910 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1911 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1914 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1916 int overlap_clear = 0, i, disks = sh->disks;
1917 struct dma_async_tx_descriptor *tx = NULL;
1918 struct r5conf *conf = sh->raid_conf;
1919 int level = conf->level;
1920 struct raid5_percpu *percpu;
1924 percpu = per_cpu_ptr(conf->percpu, cpu);
1925 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1926 ops_run_biofill(sh);
1930 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1932 tx = ops_run_compute5(sh, percpu);
1934 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1935 tx = ops_run_compute6_1(sh, percpu);
1937 tx = ops_run_compute6_2(sh, percpu);
1939 /* terminate the chain if reconstruct is not set to be run */
1940 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1944 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1946 tx = ops_run_prexor5(sh, percpu, tx);
1948 tx = ops_run_prexor6(sh, percpu, tx);
1951 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1952 tx = ops_run_biodrain(sh, tx);
1956 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1958 ops_run_reconstruct5(sh, percpu, tx);
1960 ops_run_reconstruct6(sh, percpu, tx);
1963 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1964 if (sh->check_state == check_state_run)
1965 ops_run_check_p(sh, percpu);
1966 else if (sh->check_state == check_state_run_q)
1967 ops_run_check_pq(sh, percpu, 0);
1968 else if (sh->check_state == check_state_run_pq)
1969 ops_run_check_pq(sh, percpu, 1);
1974 if (overlap_clear && !sh->batch_head)
1975 for (i = disks; i--; ) {
1976 struct r5dev *dev = &sh->dev[i];
1977 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1978 wake_up(&sh->raid_conf->wait_for_overlap);
1983 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1985 struct stripe_head *sh;
1987 sh = kmem_cache_zalloc(sc, gfp);
1989 spin_lock_init(&sh->stripe_lock);
1990 spin_lock_init(&sh->batch_lock);
1991 INIT_LIST_HEAD(&sh->batch_list);
1992 INIT_LIST_HEAD(&sh->lru);
1993 atomic_set(&sh->count, 1);
1997 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
1999 struct stripe_head *sh;
2001 sh = alloc_stripe(conf->slab_cache, gfp);
2005 sh->raid_conf = conf;
2007 if (grow_buffers(sh, gfp)) {
2009 kmem_cache_free(conf->slab_cache, sh);
2012 sh->hash_lock_index =
2013 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2014 /* we just created an active stripe so... */
2015 atomic_inc(&conf->active_stripes);
2017 raid5_release_stripe(sh);
2018 conf->max_nr_stripes++;
2022 static int grow_stripes(struct r5conf *conf, int num)
2024 struct kmem_cache *sc;
2025 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2027 if (conf->mddev->gendisk)
2028 sprintf(conf->cache_name[0],
2029 "raid%d-%s", conf->level, mdname(conf->mddev));
2031 sprintf(conf->cache_name[0],
2032 "raid%d-%p", conf->level, conf->mddev);
2033 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2035 conf->active_name = 0;
2036 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2037 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2041 conf->slab_cache = sc;
2042 conf->pool_size = devs;
2044 if (!grow_one_stripe(conf, GFP_KERNEL))
2051 * scribble_len - return the required size of the scribble region
2052 * @num - total number of disks in the array
2054 * The size must be enough to contain:
2055 * 1/ a struct page pointer for each device in the array +2
2056 * 2/ room to convert each entry in (1) to its corresponding dma
2057 * (dma_map_page()) or page (page_address()) address.
2059 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2060 * calculate over all devices (not just the data blocks), using zeros in place
2061 * of the P and Q blocks.
2063 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2065 struct flex_array *ret;
2068 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2069 ret = flex_array_alloc(len, cnt, flags);
2072 /* always prealloc all elements, so no locking is required */
2073 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2074 flex_array_free(ret);
2080 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2086 * Never shrink. And mddev_suspend() could deadlock if this is called
2087 * from raid5d. In that case, scribble_disks and scribble_sectors
2088 * should equal to new_disks and new_sectors
2090 if (conf->scribble_disks >= new_disks &&
2091 conf->scribble_sectors >= new_sectors)
2093 mddev_suspend(conf->mddev);
2095 for_each_present_cpu(cpu) {
2096 struct raid5_percpu *percpu;
2097 struct flex_array *scribble;
2099 percpu = per_cpu_ptr(conf->percpu, cpu);
2100 scribble = scribble_alloc(new_disks,
2101 new_sectors / STRIPE_SECTORS,
2105 flex_array_free(percpu->scribble);
2106 percpu->scribble = scribble;
2113 mddev_resume(conf->mddev);
2115 conf->scribble_disks = new_disks;
2116 conf->scribble_sectors = new_sectors;
2121 static int resize_stripes(struct r5conf *conf, int newsize)
2123 /* Make all the stripes able to hold 'newsize' devices.
2124 * New slots in each stripe get 'page' set to a new page.
2126 * This happens in stages:
2127 * 1/ create a new kmem_cache and allocate the required number of
2129 * 2/ gather all the old stripe_heads and transfer the pages across
2130 * to the new stripe_heads. This will have the side effect of
2131 * freezing the array as once all stripe_heads have been collected,
2132 * no IO will be possible. Old stripe heads are freed once their
2133 * pages have been transferred over, and the old kmem_cache is
2134 * freed when all stripes are done.
2135 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2136 * we simple return a failre status - no need to clean anything up.
2137 * 4/ allocate new pages for the new slots in the new stripe_heads.
2138 * If this fails, we don't bother trying the shrink the
2139 * stripe_heads down again, we just leave them as they are.
2140 * As each stripe_head is processed the new one is released into
2143 * Once step2 is started, we cannot afford to wait for a write,
2144 * so we use GFP_NOIO allocations.
2146 struct stripe_head *osh, *nsh;
2147 LIST_HEAD(newstripes);
2148 struct disk_info *ndisks;
2150 struct kmem_cache *sc;
2154 if (newsize <= conf->pool_size)
2155 return 0; /* never bother to shrink */
2157 err = md_allow_write(conf->mddev);
2162 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2163 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2168 /* Need to ensure auto-resizing doesn't interfere */
2169 mutex_lock(&conf->cache_size_mutex);
2171 for (i = conf->max_nr_stripes; i; i--) {
2172 nsh = alloc_stripe(sc, GFP_KERNEL);
2176 nsh->raid_conf = conf;
2177 list_add(&nsh->lru, &newstripes);
2180 /* didn't get enough, give up */
2181 while (!list_empty(&newstripes)) {
2182 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2183 list_del(&nsh->lru);
2184 kmem_cache_free(sc, nsh);
2186 kmem_cache_destroy(sc);
2187 mutex_unlock(&conf->cache_size_mutex);
2190 /* Step 2 - Must use GFP_NOIO now.
2191 * OK, we have enough stripes, start collecting inactive
2192 * stripes and copying them over
2196 list_for_each_entry(nsh, &newstripes, lru) {
2197 lock_device_hash_lock(conf, hash);
2198 wait_event_cmd(conf->wait_for_stripe,
2199 !list_empty(conf->inactive_list + hash),
2200 unlock_device_hash_lock(conf, hash),
2201 lock_device_hash_lock(conf, hash));
2202 osh = get_free_stripe(conf, hash);
2203 unlock_device_hash_lock(conf, hash);
2205 for(i=0; i<conf->pool_size; i++) {
2206 nsh->dev[i].page = osh->dev[i].page;
2207 nsh->dev[i].orig_page = osh->dev[i].page;
2209 nsh->hash_lock_index = hash;
2210 kmem_cache_free(conf->slab_cache, osh);
2212 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2213 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2218 kmem_cache_destroy(conf->slab_cache);
2221 * At this point, we are holding all the stripes so the array
2222 * is completely stalled, so now is a good time to resize
2223 * conf->disks and the scribble region
2225 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2227 for (i=0; i<conf->raid_disks; i++)
2228 ndisks[i] = conf->disks[i];
2230 conf->disks = ndisks;
2234 mutex_unlock(&conf->cache_size_mutex);
2235 /* Step 4, return new stripes to service */
2236 while(!list_empty(&newstripes)) {
2237 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2238 list_del_init(&nsh->lru);
2240 for (i=conf->raid_disks; i < newsize; i++)
2241 if (nsh->dev[i].page == NULL) {
2242 struct page *p = alloc_page(GFP_NOIO);
2243 nsh->dev[i].page = p;
2244 nsh->dev[i].orig_page = p;
2248 raid5_release_stripe(nsh);
2250 /* critical section pass, GFP_NOIO no longer needed */
2252 conf->slab_cache = sc;
2253 conf->active_name = 1-conf->active_name;
2255 conf->pool_size = newsize;
2259 static int drop_one_stripe(struct r5conf *conf)
2261 struct stripe_head *sh;
2262 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2264 spin_lock_irq(conf->hash_locks + hash);
2265 sh = get_free_stripe(conf, hash);
2266 spin_unlock_irq(conf->hash_locks + hash);
2269 BUG_ON(atomic_read(&sh->count));
2271 kmem_cache_free(conf->slab_cache, sh);
2272 atomic_dec(&conf->active_stripes);
2273 conf->max_nr_stripes--;
2277 static void shrink_stripes(struct r5conf *conf)
2279 while (conf->max_nr_stripes &&
2280 drop_one_stripe(conf))
2283 kmem_cache_destroy(conf->slab_cache);
2284 conf->slab_cache = NULL;
2287 static void raid5_end_read_request(struct bio * bi)
2289 struct stripe_head *sh = bi->bi_private;
2290 struct r5conf *conf = sh->raid_conf;
2291 int disks = sh->disks, i;
2292 char b[BDEVNAME_SIZE];
2293 struct md_rdev *rdev = NULL;
2296 for (i=0 ; i<disks; i++)
2297 if (bi == &sh->dev[i].req)
2300 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2301 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2307 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2308 /* If replacement finished while this request was outstanding,
2309 * 'replacement' might be NULL already.
2310 * In that case it moved down to 'rdev'.
2311 * rdev is not removed until all requests are finished.
2313 rdev = conf->disks[i].replacement;
2315 rdev = conf->disks[i].rdev;
2317 if (use_new_offset(conf, sh))
2318 s = sh->sector + rdev->new_data_offset;
2320 s = sh->sector + rdev->data_offset;
2321 if (!bi->bi_error) {
2322 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2323 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2324 /* Note that this cannot happen on a
2325 * replacement device. We just fail those on
2330 "md/raid:%s: read error corrected"
2331 " (%lu sectors at %llu on %s)\n",
2332 mdname(conf->mddev), STRIPE_SECTORS,
2333 (unsigned long long)s,
2334 bdevname(rdev->bdev, b));
2335 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2336 clear_bit(R5_ReadError, &sh->dev[i].flags);
2337 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2338 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2339 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2341 if (atomic_read(&rdev->read_errors))
2342 atomic_set(&rdev->read_errors, 0);
2344 const char *bdn = bdevname(rdev->bdev, b);
2348 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2349 atomic_inc(&rdev->read_errors);
2350 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2353 "md/raid:%s: read error on replacement device "
2354 "(sector %llu on %s).\n",
2355 mdname(conf->mddev),
2356 (unsigned long long)s,
2358 else if (conf->mddev->degraded >= conf->max_degraded) {
2362 "md/raid:%s: read error not correctable "
2363 "(sector %llu on %s).\n",
2364 mdname(conf->mddev),
2365 (unsigned long long)s,
2367 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2372 "md/raid:%s: read error NOT corrected!! "
2373 "(sector %llu on %s).\n",
2374 mdname(conf->mddev),
2375 (unsigned long long)s,
2377 } else if (atomic_read(&rdev->read_errors)
2378 > conf->max_nr_stripes)
2380 "md/raid:%s: Too many read errors, failing device %s.\n",
2381 mdname(conf->mddev), bdn);
2384 if (set_bad && test_bit(In_sync, &rdev->flags)
2385 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2388 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2389 set_bit(R5_ReadError, &sh->dev[i].flags);
2390 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2392 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2394 clear_bit(R5_ReadError, &sh->dev[i].flags);
2395 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2397 && test_bit(In_sync, &rdev->flags)
2398 && rdev_set_badblocks(
2399 rdev, sh->sector, STRIPE_SECTORS, 0)))
2400 md_error(conf->mddev, rdev);
2403 rdev_dec_pending(rdev, conf->mddev);
2404 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2405 set_bit(STRIPE_HANDLE, &sh->state);
2406 raid5_release_stripe(sh);
2409 static void raid5_end_write_request(struct bio *bi)
2411 struct stripe_head *sh = bi->bi_private;
2412 struct r5conf *conf = sh->raid_conf;
2413 int disks = sh->disks, i;
2414 struct md_rdev *uninitialized_var(rdev);
2417 int replacement = 0;
2419 for (i = 0 ; i < disks; i++) {
2420 if (bi == &sh->dev[i].req) {
2421 rdev = conf->disks[i].rdev;
2424 if (bi == &sh->dev[i].rreq) {
2425 rdev = conf->disks[i].replacement;
2429 /* rdev was removed and 'replacement'
2430 * replaced it. rdev is not removed
2431 * until all requests are finished.
2433 rdev = conf->disks[i].rdev;
2437 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2438 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2447 md_error(conf->mddev, rdev);
2448 else if (is_badblock(rdev, sh->sector,
2450 &first_bad, &bad_sectors))
2451 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2454 set_bit(STRIPE_DEGRADED, &sh->state);
2455 set_bit(WriteErrorSeen, &rdev->flags);
2456 set_bit(R5_WriteError, &sh->dev[i].flags);
2457 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2458 set_bit(MD_RECOVERY_NEEDED,
2459 &rdev->mddev->recovery);
2460 } else if (is_badblock(rdev, sh->sector,
2462 &first_bad, &bad_sectors)) {
2463 set_bit(R5_MadeGood, &sh->dev[i].flags);
2464 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2465 /* That was a successful write so make
2466 * sure it looks like we already did
2469 set_bit(R5_ReWrite, &sh->dev[i].flags);
2472 rdev_dec_pending(rdev, conf->mddev);
2474 if (sh->batch_head && bi->bi_error && !replacement)
2475 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2477 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2478 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2479 set_bit(STRIPE_HANDLE, &sh->state);
2480 raid5_release_stripe(sh);
2482 if (sh->batch_head && sh != sh->batch_head)
2483 raid5_release_stripe(sh->batch_head);
2486 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2488 struct r5dev *dev = &sh->dev[i];
2490 bio_init(&dev->req);
2491 dev->req.bi_io_vec = &dev->vec;
2492 dev->req.bi_max_vecs = 1;
2493 dev->req.bi_private = sh;
2495 bio_init(&dev->rreq);
2496 dev->rreq.bi_io_vec = &dev->rvec;
2497 dev->rreq.bi_max_vecs = 1;
2498 dev->rreq.bi_private = sh;
2501 dev->sector = raid5_compute_blocknr(sh, i, previous);
2504 static void error(struct mddev *mddev, struct md_rdev *rdev)
2506 char b[BDEVNAME_SIZE];
2507 struct r5conf *conf = mddev->private;
2508 unsigned long flags;
2509 pr_debug("raid456: error called\n");
2511 spin_lock_irqsave(&conf->device_lock, flags);
2512 clear_bit(In_sync, &rdev->flags);
2513 mddev->degraded = calc_degraded(conf);
2514 spin_unlock_irqrestore(&conf->device_lock, flags);
2515 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2517 set_bit(Blocked, &rdev->flags);
2518 set_bit(Faulty, &rdev->flags);
2519 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2520 set_bit(MD_CHANGE_PENDING, &mddev->flags);
2522 "md/raid:%s: Disk failure on %s, disabling device.\n"
2523 "md/raid:%s: Operation continuing on %d devices.\n",
2525 bdevname(rdev->bdev, b),
2527 conf->raid_disks - mddev->degraded);
2531 * Input: a 'big' sector number,
2532 * Output: index of the data and parity disk, and the sector # in them.
2534 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2535 int previous, int *dd_idx,
2536 struct stripe_head *sh)
2538 sector_t stripe, stripe2;
2539 sector_t chunk_number;
2540 unsigned int chunk_offset;
2543 sector_t new_sector;
2544 int algorithm = previous ? conf->prev_algo
2546 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2547 : conf->chunk_sectors;
2548 int raid_disks = previous ? conf->previous_raid_disks
2550 int data_disks = raid_disks - conf->max_degraded;
2552 /* First compute the information on this sector */
2555 * Compute the chunk number and the sector offset inside the chunk
2557 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2558 chunk_number = r_sector;
2561 * Compute the stripe number
2563 stripe = chunk_number;
2564 *dd_idx = sector_div(stripe, data_disks);
2567 * Select the parity disk based on the user selected algorithm.
2569 pd_idx = qd_idx = -1;
2570 switch(conf->level) {
2572 pd_idx = data_disks;
2575 switch (algorithm) {
2576 case ALGORITHM_LEFT_ASYMMETRIC:
2577 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2578 if (*dd_idx >= pd_idx)
2581 case ALGORITHM_RIGHT_ASYMMETRIC:
2582 pd_idx = sector_div(stripe2, raid_disks);
2583 if (*dd_idx >= pd_idx)
2586 case ALGORITHM_LEFT_SYMMETRIC:
2587 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2588 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2590 case ALGORITHM_RIGHT_SYMMETRIC:
2591 pd_idx = sector_div(stripe2, raid_disks);
2592 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2594 case ALGORITHM_PARITY_0:
2598 case ALGORITHM_PARITY_N:
2599 pd_idx = data_disks;
2607 switch (algorithm) {
2608 case ALGORITHM_LEFT_ASYMMETRIC:
2609 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2610 qd_idx = pd_idx + 1;
2611 if (pd_idx == raid_disks-1) {
2612 (*dd_idx)++; /* Q D D D P */
2614 } else if (*dd_idx >= pd_idx)
2615 (*dd_idx) += 2; /* D D P Q D */
2617 case ALGORITHM_RIGHT_ASYMMETRIC:
2618 pd_idx = sector_div(stripe2, raid_disks);
2619 qd_idx = pd_idx + 1;
2620 if (pd_idx == raid_disks-1) {
2621 (*dd_idx)++; /* Q D D D P */
2623 } else if (*dd_idx >= pd_idx)
2624 (*dd_idx) += 2; /* D D P Q D */
2626 case ALGORITHM_LEFT_SYMMETRIC:
2627 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2628 qd_idx = (pd_idx + 1) % raid_disks;
2629 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2631 case ALGORITHM_RIGHT_SYMMETRIC:
2632 pd_idx = sector_div(stripe2, raid_disks);
2633 qd_idx = (pd_idx + 1) % raid_disks;
2634 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2637 case ALGORITHM_PARITY_0:
2642 case ALGORITHM_PARITY_N:
2643 pd_idx = data_disks;
2644 qd_idx = data_disks + 1;
2647 case ALGORITHM_ROTATING_ZERO_RESTART:
2648 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2649 * of blocks for computing Q is different.
2651 pd_idx = sector_div(stripe2, raid_disks);
2652 qd_idx = pd_idx + 1;
2653 if (pd_idx == raid_disks-1) {
2654 (*dd_idx)++; /* Q D D D P */
2656 } else if (*dd_idx >= pd_idx)
2657 (*dd_idx) += 2; /* D D P Q D */
2661 case ALGORITHM_ROTATING_N_RESTART:
2662 /* Same a left_asymmetric, by first stripe is
2663 * D D D P Q rather than
2667 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2668 qd_idx = pd_idx + 1;
2669 if (pd_idx == raid_disks-1) {
2670 (*dd_idx)++; /* Q D D D P */
2672 } else if (*dd_idx >= pd_idx)
2673 (*dd_idx) += 2; /* D D P Q D */
2677 case ALGORITHM_ROTATING_N_CONTINUE:
2678 /* Same as left_symmetric but Q is before P */
2679 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2680 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2681 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2685 case ALGORITHM_LEFT_ASYMMETRIC_6:
2686 /* RAID5 left_asymmetric, with Q on last device */
2687 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2688 if (*dd_idx >= pd_idx)
2690 qd_idx = raid_disks - 1;
2693 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2694 pd_idx = sector_div(stripe2, raid_disks-1);
2695 if (*dd_idx >= pd_idx)
2697 qd_idx = raid_disks - 1;
2700 case ALGORITHM_LEFT_SYMMETRIC_6:
2701 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2702 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2703 qd_idx = raid_disks - 1;
2706 case ALGORITHM_RIGHT_SYMMETRIC_6:
2707 pd_idx = sector_div(stripe2, raid_disks-1);
2708 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2709 qd_idx = raid_disks - 1;
2712 case ALGORITHM_PARITY_0_6:
2715 qd_idx = raid_disks - 1;
2725 sh->pd_idx = pd_idx;
2726 sh->qd_idx = qd_idx;
2727 sh->ddf_layout = ddf_layout;
2730 * Finally, compute the new sector number
2732 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2736 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2738 struct r5conf *conf = sh->raid_conf;
2739 int raid_disks = sh->disks;
2740 int data_disks = raid_disks - conf->max_degraded;
2741 sector_t new_sector = sh->sector, check;
2742 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2743 : conf->chunk_sectors;
2744 int algorithm = previous ? conf->prev_algo
2748 sector_t chunk_number;
2749 int dummy1, dd_idx = i;
2751 struct stripe_head sh2;
2753 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2754 stripe = new_sector;
2756 if (i == sh->pd_idx)
2758 switch(conf->level) {
2761 switch (algorithm) {
2762 case ALGORITHM_LEFT_ASYMMETRIC:
2763 case ALGORITHM_RIGHT_ASYMMETRIC:
2767 case ALGORITHM_LEFT_SYMMETRIC:
2768 case ALGORITHM_RIGHT_SYMMETRIC:
2771 i -= (sh->pd_idx + 1);
2773 case ALGORITHM_PARITY_0:
2776 case ALGORITHM_PARITY_N:
2783 if (i == sh->qd_idx)
2784 return 0; /* It is the Q disk */
2785 switch (algorithm) {
2786 case ALGORITHM_LEFT_ASYMMETRIC:
2787 case ALGORITHM_RIGHT_ASYMMETRIC:
2788 case ALGORITHM_ROTATING_ZERO_RESTART:
2789 case ALGORITHM_ROTATING_N_RESTART:
2790 if (sh->pd_idx == raid_disks-1)
2791 i--; /* Q D D D P */
2792 else if (i > sh->pd_idx)
2793 i -= 2; /* D D P Q D */
2795 case ALGORITHM_LEFT_SYMMETRIC:
2796 case ALGORITHM_RIGHT_SYMMETRIC:
2797 if (sh->pd_idx == raid_disks-1)
2798 i--; /* Q D D D P */
2803 i -= (sh->pd_idx + 2);
2806 case ALGORITHM_PARITY_0:
2809 case ALGORITHM_PARITY_N:
2811 case ALGORITHM_ROTATING_N_CONTINUE:
2812 /* Like left_symmetric, but P is before Q */
2813 if (sh->pd_idx == 0)
2814 i--; /* P D D D Q */
2819 i -= (sh->pd_idx + 1);
2822 case ALGORITHM_LEFT_ASYMMETRIC_6:
2823 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2827 case ALGORITHM_LEFT_SYMMETRIC_6:
2828 case ALGORITHM_RIGHT_SYMMETRIC_6:
2830 i += data_disks + 1;
2831 i -= (sh->pd_idx + 1);
2833 case ALGORITHM_PARITY_0_6:
2842 chunk_number = stripe * data_disks + i;
2843 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2845 check = raid5_compute_sector(conf, r_sector,
2846 previous, &dummy1, &sh2);
2847 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2848 || sh2.qd_idx != sh->qd_idx) {
2849 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2850 mdname(conf->mddev));
2857 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2858 int rcw, int expand)
2860 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2861 struct r5conf *conf = sh->raid_conf;
2862 int level = conf->level;
2866 for (i = disks; i--; ) {
2867 struct r5dev *dev = &sh->dev[i];
2870 set_bit(R5_LOCKED, &dev->flags);
2871 set_bit(R5_Wantdrain, &dev->flags);
2873 clear_bit(R5_UPTODATE, &dev->flags);
2877 /* if we are not expanding this is a proper write request, and
2878 * there will be bios with new data to be drained into the
2883 /* False alarm, nothing to do */
2885 sh->reconstruct_state = reconstruct_state_drain_run;
2886 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2888 sh->reconstruct_state = reconstruct_state_run;
2890 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2892 if (s->locked + conf->max_degraded == disks)
2893 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2894 atomic_inc(&conf->pending_full_writes);
2896 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2897 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2898 BUG_ON(level == 6 &&
2899 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2900 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2902 for (i = disks; i--; ) {
2903 struct r5dev *dev = &sh->dev[i];
2904 if (i == pd_idx || i == qd_idx)
2908 (test_bit(R5_UPTODATE, &dev->flags) ||
2909 test_bit(R5_Wantcompute, &dev->flags))) {
2910 set_bit(R5_Wantdrain, &dev->flags);
2911 set_bit(R5_LOCKED, &dev->flags);
2912 clear_bit(R5_UPTODATE, &dev->flags);
2917 /* False alarm - nothing to do */
2919 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2920 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2921 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2922 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2925 /* keep the parity disk(s) locked while asynchronous operations
2928 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2929 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2933 int qd_idx = sh->qd_idx;
2934 struct r5dev *dev = &sh->dev[qd_idx];
2936 set_bit(R5_LOCKED, &dev->flags);
2937 clear_bit(R5_UPTODATE, &dev->flags);
2941 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2942 __func__, (unsigned long long)sh->sector,
2943 s->locked, s->ops_request);
2947 * Each stripe/dev can have one or more bion attached.
2948 * toread/towrite point to the first in a chain.
2949 * The bi_next chain must be in order.
2951 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2952 int forwrite, int previous)
2955 struct r5conf *conf = sh->raid_conf;
2958 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2959 (unsigned long long)bi->bi_iter.bi_sector,
2960 (unsigned long long)sh->sector);
2963 * If several bio share a stripe. The bio bi_phys_segments acts as a
2964 * reference count to avoid race. The reference count should already be
2965 * increased before this function is called (for example, in
2966 * make_request()), so other bio sharing this stripe will not free the
2967 * stripe. If a stripe is owned by one stripe, the stripe lock will
2970 spin_lock_irq(&sh->stripe_lock);
2971 /* Don't allow new IO added to stripes in batch list */
2975 bip = &sh->dev[dd_idx].towrite;
2979 bip = &sh->dev[dd_idx].toread;
2980 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2981 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2983 bip = & (*bip)->bi_next;
2985 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2988 if (!forwrite || previous)
2989 clear_bit(STRIPE_BATCH_READY, &sh->state);
2991 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2995 raid5_inc_bi_active_stripes(bi);
2998 /* check if page is covered */
2999 sector_t sector = sh->dev[dd_idx].sector;
3000 for (bi=sh->dev[dd_idx].towrite;
3001 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3002 bi && bi->bi_iter.bi_sector <= sector;
3003 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3004 if (bio_end_sector(bi) >= sector)
3005 sector = bio_end_sector(bi);
3007 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3008 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3009 sh->overwrite_disks++;
3012 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3013 (unsigned long long)(*bip)->bi_iter.bi_sector,
3014 (unsigned long long)sh->sector, dd_idx);
3016 if (conf->mddev->bitmap && firstwrite) {
3017 /* Cannot hold spinlock over bitmap_startwrite,
3018 * but must ensure this isn't added to a batch until
3019 * we have added to the bitmap and set bm_seq.
3020 * So set STRIPE_BITMAP_PENDING to prevent
3022 * If multiple add_stripe_bio() calls race here they
3023 * much all set STRIPE_BITMAP_PENDING. So only the first one
3024 * to complete "bitmap_startwrite" gets to set
3025 * STRIPE_BIT_DELAY. This is important as once a stripe
3026 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3029 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3030 spin_unlock_irq(&sh->stripe_lock);
3031 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3033 spin_lock_irq(&sh->stripe_lock);
3034 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3035 if (!sh->batch_head) {
3036 sh->bm_seq = conf->seq_flush+1;
3037 set_bit(STRIPE_BIT_DELAY, &sh->state);
3040 spin_unlock_irq(&sh->stripe_lock);
3042 if (stripe_can_batch(sh))
3043 stripe_add_to_batch_list(conf, sh);
3047 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3048 spin_unlock_irq(&sh->stripe_lock);
3052 static void end_reshape(struct r5conf *conf);
3054 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3055 struct stripe_head *sh)
3057 int sectors_per_chunk =
3058 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3060 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3061 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3063 raid5_compute_sector(conf,
3064 stripe * (disks - conf->max_degraded)
3065 *sectors_per_chunk + chunk_offset,
3071 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3072 struct stripe_head_state *s, int disks,
3073 struct bio_list *return_bi)
3076 BUG_ON(sh->batch_head);
3077 for (i = disks; i--; ) {
3081 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3082 struct md_rdev *rdev;
3084 rdev = rcu_dereference(conf->disks[i].rdev);
3085 if (rdev && test_bit(In_sync, &rdev->flags))
3086 atomic_inc(&rdev->nr_pending);
3091 if (!rdev_set_badblocks(
3095 md_error(conf->mddev, rdev);
3096 rdev_dec_pending(rdev, conf->mddev);
3099 spin_lock_irq(&sh->stripe_lock);
3100 /* fail all writes first */
3101 bi = sh->dev[i].towrite;
3102 sh->dev[i].towrite = NULL;
3103 sh->overwrite_disks = 0;
3104 spin_unlock_irq(&sh->stripe_lock);
3108 r5l_stripe_write_finished(sh);
3110 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3111 wake_up(&conf->wait_for_overlap);
3113 while (bi && bi->bi_iter.bi_sector <
3114 sh->dev[i].sector + STRIPE_SECTORS) {
3115 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3117 bi->bi_error = -EIO;
3118 if (!raid5_dec_bi_active_stripes(bi)) {
3119 md_write_end(conf->mddev);
3120 bio_list_add(return_bi, bi);
3125 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3126 STRIPE_SECTORS, 0, 0);
3128 /* and fail all 'written' */
3129 bi = sh->dev[i].written;
3130 sh->dev[i].written = NULL;
3131 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3132 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3133 sh->dev[i].page = sh->dev[i].orig_page;
3136 if (bi) bitmap_end = 1;
3137 while (bi && bi->bi_iter.bi_sector <
3138 sh->dev[i].sector + STRIPE_SECTORS) {
3139 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3141 bi->bi_error = -EIO;
3142 if (!raid5_dec_bi_active_stripes(bi)) {
3143 md_write_end(conf->mddev);
3144 bio_list_add(return_bi, bi);
3149 /* fail any reads if this device is non-operational and
3150 * the data has not reached the cache yet.
3152 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3153 s->failed > conf->max_degraded &&
3154 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3155 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3156 spin_lock_irq(&sh->stripe_lock);
3157 bi = sh->dev[i].toread;
3158 sh->dev[i].toread = NULL;
3159 spin_unlock_irq(&sh->stripe_lock);
3160 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3161 wake_up(&conf->wait_for_overlap);
3164 while (bi && bi->bi_iter.bi_sector <
3165 sh->dev[i].sector + STRIPE_SECTORS) {
3166 struct bio *nextbi =
3167 r5_next_bio(bi, sh->dev[i].sector);
3169 bi->bi_error = -EIO;
3170 if (!raid5_dec_bi_active_stripes(bi))
3171 bio_list_add(return_bi, bi);
3176 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3177 STRIPE_SECTORS, 0, 0);
3178 /* If we were in the middle of a write the parity block might
3179 * still be locked - so just clear all R5_LOCKED flags
3181 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3186 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3187 if (atomic_dec_and_test(&conf->pending_full_writes))
3188 md_wakeup_thread(conf->mddev->thread);
3192 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3193 struct stripe_head_state *s)
3198 BUG_ON(sh->batch_head);
3199 clear_bit(STRIPE_SYNCING, &sh->state);
3200 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3201 wake_up(&conf->wait_for_overlap);
3204 /* There is nothing more to do for sync/check/repair.
3205 * Don't even need to abort as that is handled elsewhere
3206 * if needed, and not always wanted e.g. if there is a known
3208 * For recover/replace we need to record a bad block on all
3209 * non-sync devices, or abort the recovery
3211 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3212 /* During recovery devices cannot be removed, so
3213 * locking and refcounting of rdevs is not needed
3215 for (i = 0; i < conf->raid_disks; i++) {
3216 struct md_rdev *rdev = conf->disks[i].rdev;
3218 && !test_bit(Faulty, &rdev->flags)
3219 && !test_bit(In_sync, &rdev->flags)
3220 && !rdev_set_badblocks(rdev, sh->sector,
3223 rdev = conf->disks[i].replacement;
3225 && !test_bit(Faulty, &rdev->flags)
3226 && !test_bit(In_sync, &rdev->flags)
3227 && !rdev_set_badblocks(rdev, sh->sector,
3232 conf->recovery_disabled =
3233 conf->mddev->recovery_disabled;
3235 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3238 static int want_replace(struct stripe_head *sh, int disk_idx)
3240 struct md_rdev *rdev;
3242 /* Doing recovery so rcu locking not required */
3243 rdev = sh->raid_conf->disks[disk_idx].replacement;
3245 && !test_bit(Faulty, &rdev->flags)
3246 && !test_bit(In_sync, &rdev->flags)
3247 && (rdev->recovery_offset <= sh->sector
3248 || rdev->mddev->recovery_cp <= sh->sector))
3254 /* fetch_block - checks the given member device to see if its data needs
3255 * to be read or computed to satisfy a request.
3257 * Returns 1 when no more member devices need to be checked, otherwise returns
3258 * 0 to tell the loop in handle_stripe_fill to continue
3261 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3262 int disk_idx, int disks)
3264 struct r5dev *dev = &sh->dev[disk_idx];
3265 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3266 &sh->dev[s->failed_num[1]] };
3270 if (test_bit(R5_LOCKED, &dev->flags) ||
3271 test_bit(R5_UPTODATE, &dev->flags))
3272 /* No point reading this as we already have it or have
3273 * decided to get it.
3278 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3279 /* We need this block to directly satisfy a request */
3282 if (s->syncing || s->expanding ||
3283 (s->replacing && want_replace(sh, disk_idx)))
3284 /* When syncing, or expanding we read everything.
3285 * When replacing, we need the replaced block.
3289 if ((s->failed >= 1 && fdev[0]->toread) ||
3290 (s->failed >= 2 && fdev[1]->toread))
3291 /* If we want to read from a failed device, then
3292 * we need to actually read every other device.
3296 /* Sometimes neither read-modify-write nor reconstruct-write
3297 * cycles can work. In those cases we read every block we
3298 * can. Then the parity-update is certain to have enough to
3300 * This can only be a problem when we need to write something,
3301 * and some device has failed. If either of those tests
3302 * fail we need look no further.
3304 if (!s->failed || !s->to_write)
3307 if (test_bit(R5_Insync, &dev->flags) &&
3308 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3309 /* Pre-reads at not permitted until after short delay
3310 * to gather multiple requests. However if this
3311 * device is no Insync, the block could only be be computed
3312 * and there is no need to delay that.
3316 for (i = 0; i < s->failed && i < 2; i++) {
3317 if (fdev[i]->towrite &&
3318 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3319 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3320 /* If we have a partial write to a failed
3321 * device, then we will need to reconstruct
3322 * the content of that device, so all other
3323 * devices must be read.
3328 /* If we are forced to do a reconstruct-write, either because
3329 * the current RAID6 implementation only supports that, or
3330 * or because parity cannot be trusted and we are currently
3331 * recovering it, there is extra need to be careful.
3332 * If one of the devices that we would need to read, because
3333 * it is not being overwritten (and maybe not written at all)
3334 * is missing/faulty, then we need to read everything we can.
3336 if (sh->raid_conf->level != 6 &&
3337 sh->sector < sh->raid_conf->mddev->recovery_cp)
3338 /* reconstruct-write isn't being forced */
3340 for (i = 0; i < s->failed && i < 2; i++) {
3341 if (s->failed_num[i] != sh->pd_idx &&
3342 s->failed_num[i] != sh->qd_idx &&
3343 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3344 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3351 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3352 int disk_idx, int disks)
3354 struct r5dev *dev = &sh->dev[disk_idx];
3356 /* is the data in this block needed, and can we get it? */
3357 if (need_this_block(sh, s, disk_idx, disks)) {
3358 /* we would like to get this block, possibly by computing it,
3359 * otherwise read it if the backing disk is insync
3361 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3362 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3363 BUG_ON(sh->batch_head);
3364 if ((s->uptodate == disks - 1) &&
3365 (s->failed && (disk_idx == s->failed_num[0] ||
3366 disk_idx == s->failed_num[1]))) {
3367 /* have disk failed, and we're requested to fetch it;
3370 pr_debug("Computing stripe %llu block %d\n",
3371 (unsigned long long)sh->sector, disk_idx);
3372 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3373 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3374 set_bit(R5_Wantcompute, &dev->flags);
3375 sh->ops.target = disk_idx;
3376 sh->ops.target2 = -1; /* no 2nd target */
3378 /* Careful: from this point on 'uptodate' is in the eye
3379 * of raid_run_ops which services 'compute' operations
3380 * before writes. R5_Wantcompute flags a block that will
3381 * be R5_UPTODATE by the time it is needed for a
3382 * subsequent operation.
3386 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3387 /* Computing 2-failure is *very* expensive; only
3388 * do it if failed >= 2
3391 for (other = disks; other--; ) {
3392 if (other == disk_idx)
3394 if (!test_bit(R5_UPTODATE,
3395 &sh->dev[other].flags))
3399 pr_debug("Computing stripe %llu blocks %d,%d\n",
3400 (unsigned long long)sh->sector,
3402 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3403 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3404 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3405 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3406 sh->ops.target = disk_idx;
3407 sh->ops.target2 = other;
3411 } else if (test_bit(R5_Insync, &dev->flags)) {
3412 set_bit(R5_LOCKED, &dev->flags);
3413 set_bit(R5_Wantread, &dev->flags);
3415 pr_debug("Reading block %d (sync=%d)\n",
3416 disk_idx, s->syncing);
3424 * handle_stripe_fill - read or compute data to satisfy pending requests.
3426 static void handle_stripe_fill(struct stripe_head *sh,
3427 struct stripe_head_state *s,
3432 /* look for blocks to read/compute, skip this if a compute
3433 * is already in flight, or if the stripe contents are in the
3434 * midst of changing due to a write
3436 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3437 !sh->reconstruct_state)
3438 for (i = disks; i--; )
3439 if (fetch_block(sh, s, i, disks))
3441 set_bit(STRIPE_HANDLE, &sh->state);
3444 static void break_stripe_batch_list(struct stripe_head *head_sh,
3445 unsigned long handle_flags);
3446 /* handle_stripe_clean_event
3447 * any written block on an uptodate or failed drive can be returned.
3448 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3449 * never LOCKED, so we don't need to test 'failed' directly.
3451 static void handle_stripe_clean_event(struct r5conf *conf,
3452 struct stripe_head *sh, int disks, struct bio_list *return_bi)
3456 int discard_pending = 0;
3457 struct stripe_head *head_sh = sh;
3458 bool do_endio = false;
3460 for (i = disks; i--; )
3461 if (sh->dev[i].written) {
3463 if (!test_bit(R5_LOCKED, &dev->flags) &&
3464 (test_bit(R5_UPTODATE, &dev->flags) ||
3465 test_bit(R5_Discard, &dev->flags) ||
3466 test_bit(R5_SkipCopy, &dev->flags))) {
3467 /* We can return any write requests */
3468 struct bio *wbi, *wbi2;
3469 pr_debug("Return write for disc %d\n", i);
3470 if (test_and_clear_bit(R5_Discard, &dev->flags))
3471 clear_bit(R5_UPTODATE, &dev->flags);
3472 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3473 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3478 dev->page = dev->orig_page;
3480 dev->written = NULL;
3481 while (wbi && wbi->bi_iter.bi_sector <
3482 dev->sector + STRIPE_SECTORS) {
3483 wbi2 = r5_next_bio(wbi, dev->sector);
3484 if (!raid5_dec_bi_active_stripes(wbi)) {
3485 md_write_end(conf->mddev);
3486 bio_list_add(return_bi, wbi);
3490 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3492 !test_bit(STRIPE_DEGRADED, &sh->state),
3494 if (head_sh->batch_head) {
3495 sh = list_first_entry(&sh->batch_list,
3498 if (sh != head_sh) {
3505 } else if (test_bit(R5_Discard, &dev->flags))
3506 discard_pending = 1;
3507 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3508 WARN_ON(dev->page != dev->orig_page);
3511 r5l_stripe_write_finished(sh);
3513 if (!discard_pending &&
3514 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3516 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3517 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3518 if (sh->qd_idx >= 0) {
3519 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3520 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3522 /* now that discard is done we can proceed with any sync */
3523 clear_bit(STRIPE_DISCARD, &sh->state);
3525 * SCSI discard will change some bio fields and the stripe has
3526 * no updated data, so remove it from hash list and the stripe
3527 * will be reinitialized
3530 hash = sh->hash_lock_index;
3531 spin_lock_irq(conf->hash_locks + hash);
3533 spin_unlock_irq(conf->hash_locks + hash);
3534 if (head_sh->batch_head) {
3535 sh = list_first_entry(&sh->batch_list,
3536 struct stripe_head, batch_list);
3542 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3543 set_bit(STRIPE_HANDLE, &sh->state);
3547 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3548 if (atomic_dec_and_test(&conf->pending_full_writes))
3549 md_wakeup_thread(conf->mddev->thread);
3551 if (head_sh->batch_head && do_endio)
3552 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3555 static void handle_stripe_dirtying(struct r5conf *conf,
3556 struct stripe_head *sh,
3557 struct stripe_head_state *s,
3560 int rmw = 0, rcw = 0, i;
3561 sector_t recovery_cp = conf->mddev->recovery_cp;
3563 /* Check whether resync is now happening or should start.
3564 * If yes, then the array is dirty (after unclean shutdown or
3565 * initial creation), so parity in some stripes might be inconsistent.
3566 * In this case, we need to always do reconstruct-write, to ensure
3567 * that in case of drive failure or read-error correction, we
3568 * generate correct data from the parity.
3570 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3571 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3573 /* Calculate the real rcw later - for now make it
3574 * look like rcw is cheaper
3577 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3578 conf->rmw_level, (unsigned long long)recovery_cp,
3579 (unsigned long long)sh->sector);
3580 } else for (i = disks; i--; ) {
3581 /* would I have to read this buffer for read_modify_write */
3582 struct r5dev *dev = &sh->dev[i];
3583 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3584 !test_bit(R5_LOCKED, &dev->flags) &&
3585 !(test_bit(R5_UPTODATE, &dev->flags) ||
3586 test_bit(R5_Wantcompute, &dev->flags))) {
3587 if (test_bit(R5_Insync, &dev->flags))
3590 rmw += 2*disks; /* cannot read it */
3592 /* Would I have to read this buffer for reconstruct_write */
3593 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3594 i != sh->pd_idx && i != sh->qd_idx &&
3595 !test_bit(R5_LOCKED, &dev->flags) &&
3596 !(test_bit(R5_UPTODATE, &dev->flags) ||
3597 test_bit(R5_Wantcompute, &dev->flags))) {
3598 if (test_bit(R5_Insync, &dev->flags))
3604 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3605 (unsigned long long)sh->sector, rmw, rcw);
3606 set_bit(STRIPE_HANDLE, &sh->state);
3607 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3608 /* prefer read-modify-write, but need to get some data */
3609 if (conf->mddev->queue)
3610 blk_add_trace_msg(conf->mddev->queue,
3611 "raid5 rmw %llu %d",
3612 (unsigned long long)sh->sector, rmw);
3613 for (i = disks; i--; ) {
3614 struct r5dev *dev = &sh->dev[i];
3615 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3616 !test_bit(R5_LOCKED, &dev->flags) &&
3617 !(test_bit(R5_UPTODATE, &dev->flags) ||
3618 test_bit(R5_Wantcompute, &dev->flags)) &&
3619 test_bit(R5_Insync, &dev->flags)) {
3620 if (test_bit(STRIPE_PREREAD_ACTIVE,
3622 pr_debug("Read_old block %d for r-m-w\n",
3624 set_bit(R5_LOCKED, &dev->flags);
3625 set_bit(R5_Wantread, &dev->flags);
3628 set_bit(STRIPE_DELAYED, &sh->state);
3629 set_bit(STRIPE_HANDLE, &sh->state);
3634 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3635 /* want reconstruct write, but need to get some data */
3638 for (i = disks; i--; ) {
3639 struct r5dev *dev = &sh->dev[i];
3640 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3641 i != sh->pd_idx && i != sh->qd_idx &&
3642 !test_bit(R5_LOCKED, &dev->flags) &&
3643 !(test_bit(R5_UPTODATE, &dev->flags) ||
3644 test_bit(R5_Wantcompute, &dev->flags))) {
3646 if (test_bit(R5_Insync, &dev->flags) &&
3647 test_bit(STRIPE_PREREAD_ACTIVE,
3649 pr_debug("Read_old block "
3650 "%d for Reconstruct\n", i);
3651 set_bit(R5_LOCKED, &dev->flags);
3652 set_bit(R5_Wantread, &dev->flags);
3656 set_bit(STRIPE_DELAYED, &sh->state);
3657 set_bit(STRIPE_HANDLE, &sh->state);
3661 if (rcw && conf->mddev->queue)
3662 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3663 (unsigned long long)sh->sector,
3664 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3667 if (rcw > disks && rmw > disks &&
3668 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3669 set_bit(STRIPE_DELAYED, &sh->state);
3671 /* now if nothing is locked, and if we have enough data,
3672 * we can start a write request
3674 /* since handle_stripe can be called at any time we need to handle the
3675 * case where a compute block operation has been submitted and then a
3676 * subsequent call wants to start a write request. raid_run_ops only
3677 * handles the case where compute block and reconstruct are requested
3678 * simultaneously. If this is not the case then new writes need to be
3679 * held off until the compute completes.
3681 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3682 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3683 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3684 schedule_reconstruction(sh, s, rcw == 0, 0);
3687 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3688 struct stripe_head_state *s, int disks)
3690 struct r5dev *dev = NULL;
3692 BUG_ON(sh->batch_head);
3693 set_bit(STRIPE_HANDLE, &sh->state);
3695 switch (sh->check_state) {
3696 case check_state_idle:
3697 /* start a new check operation if there are no failures */
3698 if (s->failed == 0) {
3699 BUG_ON(s->uptodate != disks);
3700 sh->check_state = check_state_run;
3701 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3702 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3706 dev = &sh->dev[s->failed_num[0]];
3708 case check_state_compute_result:
3709 sh->check_state = check_state_idle;
3711 dev = &sh->dev[sh->pd_idx];
3713 /* check that a write has not made the stripe insync */
3714 if (test_bit(STRIPE_INSYNC, &sh->state))
3717 /* either failed parity check, or recovery is happening */
3718 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3719 BUG_ON(s->uptodate != disks);
3721 set_bit(R5_LOCKED, &dev->flags);
3723 set_bit(R5_Wantwrite, &dev->flags);
3725 clear_bit(STRIPE_DEGRADED, &sh->state);
3726 set_bit(STRIPE_INSYNC, &sh->state);
3728 case check_state_run:
3729 break; /* we will be called again upon completion */
3730 case check_state_check_result:
3731 sh->check_state = check_state_idle;
3733 /* if a failure occurred during the check operation, leave
3734 * STRIPE_INSYNC not set and let the stripe be handled again
3739 /* handle a successful check operation, if parity is correct
3740 * we are done. Otherwise update the mismatch count and repair
3741 * parity if !MD_RECOVERY_CHECK
3743 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3744 /* parity is correct (on disc,
3745 * not in buffer any more)
3747 set_bit(STRIPE_INSYNC, &sh->state);
3749 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3750 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3751 /* don't try to repair!! */
3752 set_bit(STRIPE_INSYNC, &sh->state);
3754 sh->check_state = check_state_compute_run;
3755 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3756 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3757 set_bit(R5_Wantcompute,
3758 &sh->dev[sh->pd_idx].flags);
3759 sh->ops.target = sh->pd_idx;
3760 sh->ops.target2 = -1;
3765 case check_state_compute_run:
3768 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3769 __func__, sh->check_state,
3770 (unsigned long long) sh->sector);
3775 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3776 struct stripe_head_state *s,
3779 int pd_idx = sh->pd_idx;
3780 int qd_idx = sh->qd_idx;
3783 BUG_ON(sh->batch_head);
3784 set_bit(STRIPE_HANDLE, &sh->state);
3786 BUG_ON(s->failed > 2);
3788 /* Want to check and possibly repair P and Q.
3789 * However there could be one 'failed' device, in which
3790 * case we can only check one of them, possibly using the
3791 * other to generate missing data
3794 switch (sh->check_state) {
3795 case check_state_idle:
3796 /* start a new check operation if there are < 2 failures */
3797 if (s->failed == s->q_failed) {
3798 /* The only possible failed device holds Q, so it
3799 * makes sense to check P (If anything else were failed,
3800 * we would have used P to recreate it).
3802 sh->check_state = check_state_run;
3804 if (!s->q_failed && s->failed < 2) {
3805 /* Q is not failed, and we didn't use it to generate
3806 * anything, so it makes sense to check it
3808 if (sh->check_state == check_state_run)
3809 sh->check_state = check_state_run_pq;
3811 sh->check_state = check_state_run_q;
3814 /* discard potentially stale zero_sum_result */
3815 sh->ops.zero_sum_result = 0;
3817 if (sh->check_state == check_state_run) {
3818 /* async_xor_zero_sum destroys the contents of P */
3819 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3822 if (sh->check_state >= check_state_run &&
3823 sh->check_state <= check_state_run_pq) {
3824 /* async_syndrome_zero_sum preserves P and Q, so
3825 * no need to mark them !uptodate here
3827 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3831 /* we have 2-disk failure */
3832 BUG_ON(s->failed != 2);
3834 case check_state_compute_result:
3835 sh->check_state = check_state_idle;
3837 /* check that a write has not made the stripe insync */
3838 if (test_bit(STRIPE_INSYNC, &sh->state))
3841 /* now write out any block on a failed drive,
3842 * or P or Q if they were recomputed
3844 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3845 if (s->failed == 2) {
3846 dev = &sh->dev[s->failed_num[1]];
3848 set_bit(R5_LOCKED, &dev->flags);
3849 set_bit(R5_Wantwrite, &dev->flags);
3851 if (s->failed >= 1) {
3852 dev = &sh->dev[s->failed_num[0]];
3854 set_bit(R5_LOCKED, &dev->flags);
3855 set_bit(R5_Wantwrite, &dev->flags);
3857 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3858 dev = &sh->dev[pd_idx];
3860 set_bit(R5_LOCKED, &dev->flags);
3861 set_bit(R5_Wantwrite, &dev->flags);
3863 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3864 dev = &sh->dev[qd_idx];
3866 set_bit(R5_LOCKED, &dev->flags);
3867 set_bit(R5_Wantwrite, &dev->flags);
3869 clear_bit(STRIPE_DEGRADED, &sh->state);
3871 set_bit(STRIPE_INSYNC, &sh->state);
3873 case check_state_run:
3874 case check_state_run_q:
3875 case check_state_run_pq:
3876 break; /* we will be called again upon completion */
3877 case check_state_check_result:
3878 sh->check_state = check_state_idle;
3880 /* handle a successful check operation, if parity is correct
3881 * we are done. Otherwise update the mismatch count and repair
3882 * parity if !MD_RECOVERY_CHECK
3884 if (sh->ops.zero_sum_result == 0) {
3885 /* both parities are correct */
3887 set_bit(STRIPE_INSYNC, &sh->state);
3889 /* in contrast to the raid5 case we can validate
3890 * parity, but still have a failure to write
3893 sh->check_state = check_state_compute_result;
3894 /* Returning at this point means that we may go
3895 * off and bring p and/or q uptodate again so
3896 * we make sure to check zero_sum_result again
3897 * to verify if p or q need writeback
3901 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3902 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3903 /* don't try to repair!! */
3904 set_bit(STRIPE_INSYNC, &sh->state);
3906 int *target = &sh->ops.target;
3908 sh->ops.target = -1;
3909 sh->ops.target2 = -1;
3910 sh->check_state = check_state_compute_run;
3911 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3912 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3913 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3914 set_bit(R5_Wantcompute,
3915 &sh->dev[pd_idx].flags);
3917 target = &sh->ops.target2;
3920 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3921 set_bit(R5_Wantcompute,
3922 &sh->dev[qd_idx].flags);
3929 case check_state_compute_run:
3932 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3933 __func__, sh->check_state,
3934 (unsigned long long) sh->sector);
3939 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3943 /* We have read all the blocks in this stripe and now we need to
3944 * copy some of them into a target stripe for expand.
3946 struct dma_async_tx_descriptor *tx = NULL;
3947 BUG_ON(sh->batch_head);
3948 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3949 for (i = 0; i < sh->disks; i++)
3950 if (i != sh->pd_idx && i != sh->qd_idx) {
3952 struct stripe_head *sh2;
3953 struct async_submit_ctl submit;
3955 sector_t bn = raid5_compute_blocknr(sh, i, 1);
3956 sector_t s = raid5_compute_sector(conf, bn, 0,
3958 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
3960 /* so far only the early blocks of this stripe
3961 * have been requested. When later blocks
3962 * get requested, we will try again
3965 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3966 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3967 /* must have already done this block */
3968 raid5_release_stripe(sh2);
3972 /* place all the copies on one channel */
3973 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3974 tx = async_memcpy(sh2->dev[dd_idx].page,
3975 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3978 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3979 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3980 for (j = 0; j < conf->raid_disks; j++)
3981 if (j != sh2->pd_idx &&
3983 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3985 if (j == conf->raid_disks) {
3986 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3987 set_bit(STRIPE_HANDLE, &sh2->state);
3989 raid5_release_stripe(sh2);
3992 /* done submitting copies, wait for them to complete */
3993 async_tx_quiesce(&tx);
3997 * handle_stripe - do things to a stripe.
3999 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4000 * state of various bits to see what needs to be done.
4002 * return some read requests which now have data
4003 * return some write requests which are safely on storage
4004 * schedule a read on some buffers
4005 * schedule a write of some buffers
4006 * return confirmation of parity correctness
4010 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4012 struct r5conf *conf = sh->raid_conf;
4013 int disks = sh->disks;
4016 int do_recovery = 0;
4018 memset(s, 0, sizeof(*s));
4020 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4021 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4022 s->failed_num[0] = -1;
4023 s->failed_num[1] = -1;
4024 s->log_failed = r5l_log_disk_error(conf);
4026 /* Now to look around and see what can be done */
4028 for (i=disks; i--; ) {
4029 struct md_rdev *rdev;
4036 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4038 dev->toread, dev->towrite, dev->written);
4039 /* maybe we can reply to a read
4041 * new wantfill requests are only permitted while
4042 * ops_complete_biofill is guaranteed to be inactive
4044 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4045 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4046 set_bit(R5_Wantfill, &dev->flags);
4048 /* now count some things */
4049 if (test_bit(R5_LOCKED, &dev->flags))
4051 if (test_bit(R5_UPTODATE, &dev->flags))
4053 if (test_bit(R5_Wantcompute, &dev->flags)) {
4055 BUG_ON(s->compute > 2);
4058 if (test_bit(R5_Wantfill, &dev->flags))
4060 else if (dev->toread)
4064 if (!test_bit(R5_OVERWRITE, &dev->flags))
4069 /* Prefer to use the replacement for reads, but only
4070 * if it is recovered enough and has no bad blocks.
4072 rdev = rcu_dereference(conf->disks[i].replacement);
4073 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4074 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4075 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4076 &first_bad, &bad_sectors))
4077 set_bit(R5_ReadRepl, &dev->flags);
4079 if (rdev && !test_bit(Faulty, &rdev->flags))
4080 set_bit(R5_NeedReplace, &dev->flags);
4082 clear_bit(R5_NeedReplace, &dev->flags);
4083 rdev = rcu_dereference(conf->disks[i].rdev);
4084 clear_bit(R5_ReadRepl, &dev->flags);
4086 if (rdev && test_bit(Faulty, &rdev->flags))
4089 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4090 &first_bad, &bad_sectors);
4091 if (s->blocked_rdev == NULL
4092 && (test_bit(Blocked, &rdev->flags)
4095 set_bit(BlockedBadBlocks,
4097 s->blocked_rdev = rdev;
4098 atomic_inc(&rdev->nr_pending);
4101 clear_bit(R5_Insync, &dev->flags);
4105 /* also not in-sync */
4106 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4107 test_bit(R5_UPTODATE, &dev->flags)) {
4108 /* treat as in-sync, but with a read error
4109 * which we can now try to correct
4111 set_bit(R5_Insync, &dev->flags);
4112 set_bit(R5_ReadError, &dev->flags);
4114 } else if (test_bit(In_sync, &rdev->flags))
4115 set_bit(R5_Insync, &dev->flags);
4116 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4117 /* in sync if before recovery_offset */
4118 set_bit(R5_Insync, &dev->flags);
4119 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4120 test_bit(R5_Expanded, &dev->flags))
4121 /* If we've reshaped into here, we assume it is Insync.
4122 * We will shortly update recovery_offset to make
4125 set_bit(R5_Insync, &dev->flags);
4127 if (test_bit(R5_WriteError, &dev->flags)) {
4128 /* This flag does not apply to '.replacement'
4129 * only to .rdev, so make sure to check that*/
4130 struct md_rdev *rdev2 = rcu_dereference(
4131 conf->disks[i].rdev);
4133 clear_bit(R5_Insync, &dev->flags);
4134 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4135 s->handle_bad_blocks = 1;
4136 atomic_inc(&rdev2->nr_pending);
4138 clear_bit(R5_WriteError, &dev->flags);
4140 if (test_bit(R5_MadeGood, &dev->flags)) {
4141 /* This flag does not apply to '.replacement'
4142 * only to .rdev, so make sure to check that*/
4143 struct md_rdev *rdev2 = rcu_dereference(
4144 conf->disks[i].rdev);
4145 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4146 s->handle_bad_blocks = 1;
4147 atomic_inc(&rdev2->nr_pending);
4149 clear_bit(R5_MadeGood, &dev->flags);
4151 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4152 struct md_rdev *rdev2 = rcu_dereference(
4153 conf->disks[i].replacement);
4154 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4155 s->handle_bad_blocks = 1;
4156 atomic_inc(&rdev2->nr_pending);
4158 clear_bit(R5_MadeGoodRepl, &dev->flags);
4160 if (!test_bit(R5_Insync, &dev->flags)) {
4161 /* The ReadError flag will just be confusing now */
4162 clear_bit(R5_ReadError, &dev->flags);
4163 clear_bit(R5_ReWrite, &dev->flags);
4165 if (test_bit(R5_ReadError, &dev->flags))
4166 clear_bit(R5_Insync, &dev->flags);
4167 if (!test_bit(R5_Insync, &dev->flags)) {
4169 s->failed_num[s->failed] = i;
4171 if (rdev && !test_bit(Faulty, &rdev->flags))
4175 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4176 /* If there is a failed device being replaced,
4177 * we must be recovering.
4178 * else if we are after recovery_cp, we must be syncing
4179 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4180 * else we can only be replacing
4181 * sync and recovery both need to read all devices, and so
4182 * use the same flag.
4185 sh->sector >= conf->mddev->recovery_cp ||
4186 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4194 static int clear_batch_ready(struct stripe_head *sh)
4196 /* Return '1' if this is a member of batch, or
4197 * '0' if it is a lone stripe or a head which can now be
4200 struct stripe_head *tmp;
4201 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4202 return (sh->batch_head && sh->batch_head != sh);
4203 spin_lock(&sh->stripe_lock);
4204 if (!sh->batch_head) {
4205 spin_unlock(&sh->stripe_lock);
4210 * this stripe could be added to a batch list before we check
4211 * BATCH_READY, skips it
4213 if (sh->batch_head != sh) {
4214 spin_unlock(&sh->stripe_lock);
4217 spin_lock(&sh->batch_lock);
4218 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4219 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4220 spin_unlock(&sh->batch_lock);
4221 spin_unlock(&sh->stripe_lock);
4224 * BATCH_READY is cleared, no new stripes can be added.
4225 * batch_list can be accessed without lock
4230 static void break_stripe_batch_list(struct stripe_head *head_sh,
4231 unsigned long handle_flags)
4233 struct stripe_head *sh, *next;
4237 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4239 list_del_init(&sh->batch_list);
4241 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4242 (1 << STRIPE_SYNCING) |
4243 (1 << STRIPE_REPLACED) |
4244 (1 << STRIPE_DELAYED) |
4245 (1 << STRIPE_BIT_DELAY) |
4246 (1 << STRIPE_FULL_WRITE) |
4247 (1 << STRIPE_BIOFILL_RUN) |
4248 (1 << STRIPE_COMPUTE_RUN) |
4249 (1 << STRIPE_OPS_REQ_PENDING) |
4250 (1 << STRIPE_DISCARD) |
4251 (1 << STRIPE_BATCH_READY) |
4252 (1 << STRIPE_BATCH_ERR) |
4253 (1 << STRIPE_BITMAP_PENDING)));
4254 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4255 (1 << STRIPE_REPLACED)));
4257 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4258 (1 << STRIPE_PREREAD_ACTIVE) |
4259 (1 << STRIPE_DEGRADED)),
4260 head_sh->state & (1 << STRIPE_INSYNC));
4262 sh->check_state = head_sh->check_state;
4263 sh->reconstruct_state = head_sh->reconstruct_state;
4264 for (i = 0; i < sh->disks; i++) {
4265 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4267 sh->dev[i].flags = head_sh->dev[i].flags &
4268 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4270 spin_lock_irq(&sh->stripe_lock);
4271 sh->batch_head = NULL;
4272 spin_unlock_irq(&sh->stripe_lock);
4273 if (handle_flags == 0 ||
4274 sh->state & handle_flags)
4275 set_bit(STRIPE_HANDLE, &sh->state);
4276 raid5_release_stripe(sh);
4278 spin_lock_irq(&head_sh->stripe_lock);
4279 head_sh->batch_head = NULL;
4280 spin_unlock_irq(&head_sh->stripe_lock);
4281 for (i = 0; i < head_sh->disks; i++)
4282 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4284 if (head_sh->state & handle_flags)
4285 set_bit(STRIPE_HANDLE, &head_sh->state);
4288 wake_up(&head_sh->raid_conf->wait_for_overlap);
4291 static void handle_stripe(struct stripe_head *sh)
4293 struct stripe_head_state s;
4294 struct r5conf *conf = sh->raid_conf;
4297 int disks = sh->disks;
4298 struct r5dev *pdev, *qdev;
4300 clear_bit(STRIPE_HANDLE, &sh->state);
4301 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4302 /* already being handled, ensure it gets handled
4303 * again when current action finishes */
4304 set_bit(STRIPE_HANDLE, &sh->state);
4308 if (clear_batch_ready(sh) ) {
4309 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4313 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4314 break_stripe_batch_list(sh, 0);
4316 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4317 spin_lock(&sh->stripe_lock);
4318 /* Cannot process 'sync' concurrently with 'discard' */
4319 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4320 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4321 set_bit(STRIPE_SYNCING, &sh->state);
4322 clear_bit(STRIPE_INSYNC, &sh->state);
4323 clear_bit(STRIPE_REPLACED, &sh->state);
4325 spin_unlock(&sh->stripe_lock);
4327 clear_bit(STRIPE_DELAYED, &sh->state);
4329 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4330 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4331 (unsigned long long)sh->sector, sh->state,
4332 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4333 sh->check_state, sh->reconstruct_state);
4335 analyse_stripe(sh, &s);
4337 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4340 if (s.handle_bad_blocks) {
4341 set_bit(STRIPE_HANDLE, &sh->state);
4345 if (unlikely(s.blocked_rdev)) {
4346 if (s.syncing || s.expanding || s.expanded ||
4347 s.replacing || s.to_write || s.written) {
4348 set_bit(STRIPE_HANDLE, &sh->state);
4351 /* There is nothing for the blocked_rdev to block */
4352 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4353 s.blocked_rdev = NULL;
4356 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4357 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4358 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4361 pr_debug("locked=%d uptodate=%d to_read=%d"
4362 " to_write=%d failed=%d failed_num=%d,%d\n",
4363 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4364 s.failed_num[0], s.failed_num[1]);
4365 /* check if the array has lost more than max_degraded devices and,
4366 * if so, some requests might need to be failed.
4368 if (s.failed > conf->max_degraded || s.log_failed) {
4369 sh->check_state = 0;
4370 sh->reconstruct_state = 0;
4371 break_stripe_batch_list(sh, 0);
4372 if (s.to_read+s.to_write+s.written)
4373 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4374 if (s.syncing + s.replacing)
4375 handle_failed_sync(conf, sh, &s);
4378 /* Now we check to see if any write operations have recently
4382 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4384 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4385 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4386 sh->reconstruct_state = reconstruct_state_idle;
4388 /* All the 'written' buffers and the parity block are ready to
4389 * be written back to disk
4391 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4392 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4393 BUG_ON(sh->qd_idx >= 0 &&
4394 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4395 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4396 for (i = disks; i--; ) {
4397 struct r5dev *dev = &sh->dev[i];
4398 if (test_bit(R5_LOCKED, &dev->flags) &&
4399 (i == sh->pd_idx || i == sh->qd_idx ||
4401 pr_debug("Writing block %d\n", i);
4402 set_bit(R5_Wantwrite, &dev->flags);
4407 if (!test_bit(R5_Insync, &dev->flags) ||
4408 ((i == sh->pd_idx || i == sh->qd_idx) &&
4410 set_bit(STRIPE_INSYNC, &sh->state);
4413 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4414 s.dec_preread_active = 1;
4418 * might be able to return some write requests if the parity blocks
4419 * are safe, or on a failed drive
4421 pdev = &sh->dev[sh->pd_idx];
4422 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4423 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4424 qdev = &sh->dev[sh->qd_idx];
4425 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4426 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4430 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4431 && !test_bit(R5_LOCKED, &pdev->flags)
4432 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4433 test_bit(R5_Discard, &pdev->flags))))) &&
4434 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4435 && !test_bit(R5_LOCKED, &qdev->flags)
4436 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4437 test_bit(R5_Discard, &qdev->flags))))))
4438 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4440 /* Now we might consider reading some blocks, either to check/generate
4441 * parity, or to satisfy requests
4442 * or to load a block that is being partially written.
4444 if (s.to_read || s.non_overwrite
4445 || (conf->level == 6 && s.to_write && s.failed)
4446 || (s.syncing && (s.uptodate + s.compute < disks))
4449 handle_stripe_fill(sh, &s, disks);
4451 /* Now to consider new write requests and what else, if anything
4452 * should be read. We do not handle new writes when:
4453 * 1/ A 'write' operation (copy+xor) is already in flight.
4454 * 2/ A 'check' operation is in flight, as it may clobber the parity
4457 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4458 handle_stripe_dirtying(conf, sh, &s, disks);
4460 /* maybe we need to check and possibly fix the parity for this stripe
4461 * Any reads will already have been scheduled, so we just see if enough
4462 * data is available. The parity check is held off while parity
4463 * dependent operations are in flight.
4465 if (sh->check_state ||
4466 (s.syncing && s.locked == 0 &&
4467 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4468 !test_bit(STRIPE_INSYNC, &sh->state))) {
4469 if (conf->level == 6)
4470 handle_parity_checks6(conf, sh, &s, disks);
4472 handle_parity_checks5(conf, sh, &s, disks);
4475 if ((s.replacing || s.syncing) && s.locked == 0
4476 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4477 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4478 /* Write out to replacement devices where possible */
4479 for (i = 0; i < conf->raid_disks; i++)
4480 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4481 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4482 set_bit(R5_WantReplace, &sh->dev[i].flags);
4483 set_bit(R5_LOCKED, &sh->dev[i].flags);
4487 set_bit(STRIPE_INSYNC, &sh->state);
4488 set_bit(STRIPE_REPLACED, &sh->state);
4490 if ((s.syncing || s.replacing) && s.locked == 0 &&
4491 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4492 test_bit(STRIPE_INSYNC, &sh->state)) {
4493 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4494 clear_bit(STRIPE_SYNCING, &sh->state);
4495 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4496 wake_up(&conf->wait_for_overlap);
4499 /* If the failed drives are just a ReadError, then we might need
4500 * to progress the repair/check process
4502 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4503 for (i = 0; i < s.failed; i++) {
4504 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4505 if (test_bit(R5_ReadError, &dev->flags)
4506 && !test_bit(R5_LOCKED, &dev->flags)
4507 && test_bit(R5_UPTODATE, &dev->flags)
4509 if (!test_bit(R5_ReWrite, &dev->flags)) {
4510 set_bit(R5_Wantwrite, &dev->flags);
4511 set_bit(R5_ReWrite, &dev->flags);
4512 set_bit(R5_LOCKED, &dev->flags);
4515 /* let's read it back */
4516 set_bit(R5_Wantread, &dev->flags);
4517 set_bit(R5_LOCKED, &dev->flags);
4523 /* Finish reconstruct operations initiated by the expansion process */
4524 if (sh->reconstruct_state == reconstruct_state_result) {
4525 struct stripe_head *sh_src
4526 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4527 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4528 /* sh cannot be written until sh_src has been read.
4529 * so arrange for sh to be delayed a little
4531 set_bit(STRIPE_DELAYED, &sh->state);
4532 set_bit(STRIPE_HANDLE, &sh->state);
4533 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4535 atomic_inc(&conf->preread_active_stripes);
4536 raid5_release_stripe(sh_src);
4540 raid5_release_stripe(sh_src);
4542 sh->reconstruct_state = reconstruct_state_idle;
4543 clear_bit(STRIPE_EXPANDING, &sh->state);
4544 for (i = conf->raid_disks; i--; ) {
4545 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4546 set_bit(R5_LOCKED, &sh->dev[i].flags);
4551 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4552 !sh->reconstruct_state) {
4553 /* Need to write out all blocks after computing parity */
4554 sh->disks = conf->raid_disks;
4555 stripe_set_idx(sh->sector, conf, 0, sh);
4556 schedule_reconstruction(sh, &s, 1, 1);
4557 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4558 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4559 atomic_dec(&conf->reshape_stripes);
4560 wake_up(&conf->wait_for_overlap);
4561 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4564 if (s.expanding && s.locked == 0 &&
4565 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4566 handle_stripe_expansion(conf, sh);
4569 /* wait for this device to become unblocked */
4570 if (unlikely(s.blocked_rdev)) {
4571 if (conf->mddev->external)
4572 md_wait_for_blocked_rdev(s.blocked_rdev,
4575 /* Internal metadata will immediately
4576 * be written by raid5d, so we don't
4577 * need to wait here.
4579 rdev_dec_pending(s.blocked_rdev,
4583 if (s.handle_bad_blocks)
4584 for (i = disks; i--; ) {
4585 struct md_rdev *rdev;
4586 struct r5dev *dev = &sh->dev[i];
4587 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4588 /* We own a safe reference to the rdev */
4589 rdev = conf->disks[i].rdev;
4590 if (!rdev_set_badblocks(rdev, sh->sector,
4592 md_error(conf->mddev, rdev);
4593 rdev_dec_pending(rdev, conf->mddev);
4595 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4596 rdev = conf->disks[i].rdev;
4597 rdev_clear_badblocks(rdev, sh->sector,
4599 rdev_dec_pending(rdev, conf->mddev);
4601 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4602 rdev = conf->disks[i].replacement;
4604 /* rdev have been moved down */
4605 rdev = conf->disks[i].rdev;
4606 rdev_clear_badblocks(rdev, sh->sector,
4608 rdev_dec_pending(rdev, conf->mddev);
4613 raid_run_ops(sh, s.ops_request);
4617 if (s.dec_preread_active) {
4618 /* We delay this until after ops_run_io so that if make_request
4619 * is waiting on a flush, it won't continue until the writes
4620 * have actually been submitted.
4622 atomic_dec(&conf->preread_active_stripes);
4623 if (atomic_read(&conf->preread_active_stripes) <
4625 md_wakeup_thread(conf->mddev->thread);
4628 if (!bio_list_empty(&s.return_bi)) {
4629 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags)) {
4630 spin_lock_irq(&conf->device_lock);
4631 bio_list_merge(&conf->return_bi, &s.return_bi);
4632 spin_unlock_irq(&conf->device_lock);
4633 md_wakeup_thread(conf->mddev->thread);
4635 return_io(&s.return_bi);
4638 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4641 static void raid5_activate_delayed(struct r5conf *conf)
4643 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4644 while (!list_empty(&conf->delayed_list)) {
4645 struct list_head *l = conf->delayed_list.next;
4646 struct stripe_head *sh;
4647 sh = list_entry(l, struct stripe_head, lru);
4649 clear_bit(STRIPE_DELAYED, &sh->state);
4650 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4651 atomic_inc(&conf->preread_active_stripes);
4652 list_add_tail(&sh->lru, &conf->hold_list);
4653 raid5_wakeup_stripe_thread(sh);
4658 static void activate_bit_delay(struct r5conf *conf,
4659 struct list_head *temp_inactive_list)
4661 /* device_lock is held */
4662 struct list_head head;
4663 list_add(&head, &conf->bitmap_list);
4664 list_del_init(&conf->bitmap_list);
4665 while (!list_empty(&head)) {
4666 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4668 list_del_init(&sh->lru);
4669 atomic_inc(&sh->count);
4670 hash = sh->hash_lock_index;
4671 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4675 static int raid5_congested(struct mddev *mddev, int bits)
4677 struct r5conf *conf = mddev->private;
4679 /* No difference between reads and writes. Just check
4680 * how busy the stripe_cache is
4683 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4687 if (atomic_read(&conf->empty_inactive_list_nr))
4693 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4695 struct r5conf *conf = mddev->private;
4696 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4697 unsigned int chunk_sectors;
4698 unsigned int bio_sectors = bio_sectors(bio);
4700 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4701 return chunk_sectors >=
4702 ((sector & (chunk_sectors - 1)) + bio_sectors);
4706 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4707 * later sampled by raid5d.
4709 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4711 unsigned long flags;
4713 spin_lock_irqsave(&conf->device_lock, flags);
4715 bi->bi_next = conf->retry_read_aligned_list;
4716 conf->retry_read_aligned_list = bi;
4718 spin_unlock_irqrestore(&conf->device_lock, flags);
4719 md_wakeup_thread(conf->mddev->thread);
4722 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4726 bi = conf->retry_read_aligned;
4728 conf->retry_read_aligned = NULL;
4731 bi = conf->retry_read_aligned_list;
4733 conf->retry_read_aligned_list = bi->bi_next;
4736 * this sets the active strip count to 1 and the processed
4737 * strip count to zero (upper 8 bits)
4739 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4746 * The "raid5_align_endio" should check if the read succeeded and if it
4747 * did, call bio_endio on the original bio (having bio_put the new bio
4749 * If the read failed..
4751 static void raid5_align_endio(struct bio *bi)
4753 struct bio* raid_bi = bi->bi_private;
4754 struct mddev *mddev;
4755 struct r5conf *conf;
4756 struct md_rdev *rdev;
4757 int error = bi->bi_error;
4761 rdev = (void*)raid_bi->bi_next;
4762 raid_bi->bi_next = NULL;
4763 mddev = rdev->mddev;
4764 conf = mddev->private;
4766 rdev_dec_pending(rdev, conf->mddev);
4769 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4772 if (atomic_dec_and_test(&conf->active_aligned_reads))
4773 wake_up(&conf->wait_for_quiescent);
4777 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4779 add_bio_to_retry(raid_bi, conf);
4782 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4784 struct r5conf *conf = mddev->private;
4786 struct bio* align_bi;
4787 struct md_rdev *rdev;
4788 sector_t end_sector;
4790 if (!in_chunk_boundary(mddev, raid_bio)) {
4791 pr_debug("%s: non aligned\n", __func__);
4795 * use bio_clone_mddev to make a copy of the bio
4797 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4801 * set bi_end_io to a new function, and set bi_private to the
4804 align_bi->bi_end_io = raid5_align_endio;
4805 align_bi->bi_private = raid_bio;
4809 align_bi->bi_iter.bi_sector =
4810 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4813 end_sector = bio_end_sector(align_bi);
4815 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4816 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4817 rdev->recovery_offset < end_sector) {
4818 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4820 (test_bit(Faulty, &rdev->flags) ||
4821 !(test_bit(In_sync, &rdev->flags) ||
4822 rdev->recovery_offset >= end_sector)))
4829 atomic_inc(&rdev->nr_pending);
4831 raid_bio->bi_next = (void*)rdev;
4832 align_bi->bi_bdev = rdev->bdev;
4833 bio_clear_flag(align_bi, BIO_SEG_VALID);
4835 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4836 bio_sectors(align_bi),
4837 &first_bad, &bad_sectors)) {
4839 rdev_dec_pending(rdev, mddev);
4843 /* No reshape active, so we can trust rdev->data_offset */
4844 align_bi->bi_iter.bi_sector += rdev->data_offset;
4846 spin_lock_irq(&conf->device_lock);
4847 wait_event_lock_irq(conf->wait_for_quiescent,
4850 atomic_inc(&conf->active_aligned_reads);
4851 spin_unlock_irq(&conf->device_lock);
4854 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4855 align_bi, disk_devt(mddev->gendisk),
4856 raid_bio->bi_iter.bi_sector);
4857 generic_make_request(align_bi);
4866 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4871 sector_t sector = raid_bio->bi_iter.bi_sector;
4872 unsigned chunk_sects = mddev->chunk_sectors;
4873 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4875 if (sectors < bio_sectors(raid_bio)) {
4876 split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4877 bio_chain(split, raid_bio);
4881 if (!raid5_read_one_chunk(mddev, split)) {
4882 if (split != raid_bio)
4883 generic_make_request(raid_bio);
4886 } while (split != raid_bio);
4891 /* __get_priority_stripe - get the next stripe to process
4893 * Full stripe writes are allowed to pass preread active stripes up until
4894 * the bypass_threshold is exceeded. In general the bypass_count
4895 * increments when the handle_list is handled before the hold_list; however, it
4896 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4897 * stripe with in flight i/o. The bypass_count will be reset when the
4898 * head of the hold_list has changed, i.e. the head was promoted to the
4901 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4903 struct stripe_head *sh = NULL, *tmp;
4904 struct list_head *handle_list = NULL;
4905 struct r5worker_group *wg = NULL;
4907 if (conf->worker_cnt_per_group == 0) {
4908 handle_list = &conf->handle_list;
4909 } else if (group != ANY_GROUP) {
4910 handle_list = &conf->worker_groups[group].handle_list;
4911 wg = &conf->worker_groups[group];
4914 for (i = 0; i < conf->group_cnt; i++) {
4915 handle_list = &conf->worker_groups[i].handle_list;
4916 wg = &conf->worker_groups[i];
4917 if (!list_empty(handle_list))
4922 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4924 list_empty(handle_list) ? "empty" : "busy",
4925 list_empty(&conf->hold_list) ? "empty" : "busy",
4926 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4928 if (!list_empty(handle_list)) {
4929 sh = list_entry(handle_list->next, typeof(*sh), lru);
4931 if (list_empty(&conf->hold_list))
4932 conf->bypass_count = 0;
4933 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4934 if (conf->hold_list.next == conf->last_hold)
4935 conf->bypass_count++;
4937 conf->last_hold = conf->hold_list.next;
4938 conf->bypass_count -= conf->bypass_threshold;
4939 if (conf->bypass_count < 0)
4940 conf->bypass_count = 0;
4943 } else if (!list_empty(&conf->hold_list) &&
4944 ((conf->bypass_threshold &&
4945 conf->bypass_count > conf->bypass_threshold) ||
4946 atomic_read(&conf->pending_full_writes) == 0)) {
4948 list_for_each_entry(tmp, &conf->hold_list, lru) {
4949 if (conf->worker_cnt_per_group == 0 ||
4950 group == ANY_GROUP ||
4951 !cpu_online(tmp->cpu) ||
4952 cpu_to_group(tmp->cpu) == group) {
4959 conf->bypass_count -= conf->bypass_threshold;
4960 if (conf->bypass_count < 0)
4961 conf->bypass_count = 0;
4973 list_del_init(&sh->lru);
4974 BUG_ON(atomic_inc_return(&sh->count) != 1);
4978 struct raid5_plug_cb {
4979 struct blk_plug_cb cb;
4980 struct list_head list;
4981 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4984 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4986 struct raid5_plug_cb *cb = container_of(
4987 blk_cb, struct raid5_plug_cb, cb);
4988 struct stripe_head *sh;
4989 struct mddev *mddev = cb->cb.data;
4990 struct r5conf *conf = mddev->private;
4994 if (cb->list.next && !list_empty(&cb->list)) {
4995 spin_lock_irq(&conf->device_lock);
4996 while (!list_empty(&cb->list)) {
4997 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4998 list_del_init(&sh->lru);
5000 * avoid race release_stripe_plug() sees
5001 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5002 * is still in our list
5004 smp_mb__before_atomic();
5005 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5007 * STRIPE_ON_RELEASE_LIST could be set here. In that
5008 * case, the count is always > 1 here
5010 hash = sh->hash_lock_index;
5011 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5014 spin_unlock_irq(&conf->device_lock);
5016 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5017 NR_STRIPE_HASH_LOCKS);
5019 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5023 static void release_stripe_plug(struct mddev *mddev,
5024 struct stripe_head *sh)
5026 struct blk_plug_cb *blk_cb = blk_check_plugged(
5027 raid5_unplug, mddev,
5028 sizeof(struct raid5_plug_cb));
5029 struct raid5_plug_cb *cb;
5032 raid5_release_stripe(sh);
5036 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5038 if (cb->list.next == NULL) {
5040 INIT_LIST_HEAD(&cb->list);
5041 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5042 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5045 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5046 list_add_tail(&sh->lru, &cb->list);
5048 raid5_release_stripe(sh);
5051 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5053 struct r5conf *conf = mddev->private;
5054 sector_t logical_sector, last_sector;
5055 struct stripe_head *sh;
5059 if (mddev->reshape_position != MaxSector)
5060 /* Skip discard while reshape is happening */
5063 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5064 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5067 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5069 stripe_sectors = conf->chunk_sectors *
5070 (conf->raid_disks - conf->max_degraded);
5071 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5073 sector_div(last_sector, stripe_sectors);
5075 logical_sector *= conf->chunk_sectors;
5076 last_sector *= conf->chunk_sectors;
5078 for (; logical_sector < last_sector;
5079 logical_sector += STRIPE_SECTORS) {
5083 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5084 prepare_to_wait(&conf->wait_for_overlap, &w,
5085 TASK_UNINTERRUPTIBLE);
5086 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5087 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5088 raid5_release_stripe(sh);
5092 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5093 spin_lock_irq(&sh->stripe_lock);
5094 for (d = 0; d < conf->raid_disks; d++) {
5095 if (d == sh->pd_idx || d == sh->qd_idx)
5097 if (sh->dev[d].towrite || sh->dev[d].toread) {
5098 set_bit(R5_Overlap, &sh->dev[d].flags);
5099 spin_unlock_irq(&sh->stripe_lock);
5100 raid5_release_stripe(sh);
5105 set_bit(STRIPE_DISCARD, &sh->state);
5106 finish_wait(&conf->wait_for_overlap, &w);
5107 sh->overwrite_disks = 0;
5108 for (d = 0; d < conf->raid_disks; d++) {
5109 if (d == sh->pd_idx || d == sh->qd_idx)
5111 sh->dev[d].towrite = bi;
5112 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5113 raid5_inc_bi_active_stripes(bi);
5114 sh->overwrite_disks++;
5116 spin_unlock_irq(&sh->stripe_lock);
5117 if (conf->mddev->bitmap) {
5119 d < conf->raid_disks - conf->max_degraded;
5121 bitmap_startwrite(mddev->bitmap,
5125 sh->bm_seq = conf->seq_flush + 1;
5126 set_bit(STRIPE_BIT_DELAY, &sh->state);
5129 set_bit(STRIPE_HANDLE, &sh->state);
5130 clear_bit(STRIPE_DELAYED, &sh->state);
5131 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5132 atomic_inc(&conf->preread_active_stripes);
5133 release_stripe_plug(mddev, sh);
5136 remaining = raid5_dec_bi_active_stripes(bi);
5137 if (remaining == 0) {
5138 md_write_end(mddev);
5143 static void make_request(struct mddev *mddev, struct bio * bi)
5145 struct r5conf *conf = mddev->private;
5147 sector_t new_sector;
5148 sector_t logical_sector, last_sector;
5149 struct stripe_head *sh;
5150 const int rw = bio_data_dir(bi);
5155 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5156 int ret = r5l_handle_flush_request(conf->log, bi);
5160 if (ret == -ENODEV) {
5161 md_flush_request(mddev, bi);
5164 /* ret == -EAGAIN, fallback */
5167 md_write_start(mddev, bi);
5170 * If array is degraded, better not do chunk aligned read because
5171 * later we might have to read it again in order to reconstruct
5172 * data on failed drives.
5174 if (rw == READ && mddev->degraded == 0 &&
5175 mddev->reshape_position == MaxSector) {
5176 bi = chunk_aligned_read(mddev, bi);
5181 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5182 make_discard_request(mddev, bi);
5186 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5187 last_sector = bio_end_sector(bi);
5189 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5191 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5192 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5198 seq = read_seqcount_begin(&conf->gen_lock);
5201 prepare_to_wait(&conf->wait_for_overlap, &w,
5202 TASK_UNINTERRUPTIBLE);
5203 if (unlikely(conf->reshape_progress != MaxSector)) {
5204 /* spinlock is needed as reshape_progress may be
5205 * 64bit on a 32bit platform, and so it might be
5206 * possible to see a half-updated value
5207 * Of course reshape_progress could change after
5208 * the lock is dropped, so once we get a reference
5209 * to the stripe that we think it is, we will have
5212 spin_lock_irq(&conf->device_lock);
5213 if (mddev->reshape_backwards
5214 ? logical_sector < conf->reshape_progress
5215 : logical_sector >= conf->reshape_progress) {
5218 if (mddev->reshape_backwards
5219 ? logical_sector < conf->reshape_safe
5220 : logical_sector >= conf->reshape_safe) {
5221 spin_unlock_irq(&conf->device_lock);
5227 spin_unlock_irq(&conf->device_lock);
5230 new_sector = raid5_compute_sector(conf, logical_sector,
5233 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5234 (unsigned long long)new_sector,
5235 (unsigned long long)logical_sector);
5237 sh = raid5_get_active_stripe(conf, new_sector, previous,
5238 (bi->bi_rw&RWA_MASK), 0);
5240 if (unlikely(previous)) {
5241 /* expansion might have moved on while waiting for a
5242 * stripe, so we must do the range check again.
5243 * Expansion could still move past after this
5244 * test, but as we are holding a reference to
5245 * 'sh', we know that if that happens,
5246 * STRIPE_EXPANDING will get set and the expansion
5247 * won't proceed until we finish with the stripe.
5250 spin_lock_irq(&conf->device_lock);
5251 if (mddev->reshape_backwards
5252 ? logical_sector >= conf->reshape_progress
5253 : logical_sector < conf->reshape_progress)
5254 /* mismatch, need to try again */
5256 spin_unlock_irq(&conf->device_lock);
5258 raid5_release_stripe(sh);
5264 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5265 /* Might have got the wrong stripe_head
5268 raid5_release_stripe(sh);
5273 logical_sector >= mddev->suspend_lo &&
5274 logical_sector < mddev->suspend_hi) {
5275 raid5_release_stripe(sh);
5276 /* As the suspend_* range is controlled by
5277 * userspace, we want an interruptible
5280 flush_signals(current);
5281 prepare_to_wait(&conf->wait_for_overlap,
5282 &w, TASK_INTERRUPTIBLE);
5283 if (logical_sector >= mddev->suspend_lo &&
5284 logical_sector < mddev->suspend_hi) {
5291 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5292 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5293 /* Stripe is busy expanding or
5294 * add failed due to overlap. Flush everything
5297 md_wakeup_thread(mddev->thread);
5298 raid5_release_stripe(sh);
5303 set_bit(STRIPE_HANDLE, &sh->state);
5304 clear_bit(STRIPE_DELAYED, &sh->state);
5305 if ((!sh->batch_head || sh == sh->batch_head) &&
5306 (bi->bi_rw & REQ_SYNC) &&
5307 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5308 atomic_inc(&conf->preread_active_stripes);
5309 release_stripe_plug(mddev, sh);
5311 /* cannot get stripe for read-ahead, just give-up */
5312 bi->bi_error = -EIO;
5316 finish_wait(&conf->wait_for_overlap, &w);
5318 remaining = raid5_dec_bi_active_stripes(bi);
5319 if (remaining == 0) {
5322 md_write_end(mddev);
5324 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5330 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5332 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5334 /* reshaping is quite different to recovery/resync so it is
5335 * handled quite separately ... here.
5337 * On each call to sync_request, we gather one chunk worth of
5338 * destination stripes and flag them as expanding.
5339 * Then we find all the source stripes and request reads.
5340 * As the reads complete, handle_stripe will copy the data
5341 * into the destination stripe and release that stripe.
5343 struct r5conf *conf = mddev->private;
5344 struct stripe_head *sh;
5345 sector_t first_sector, last_sector;
5346 int raid_disks = conf->previous_raid_disks;
5347 int data_disks = raid_disks - conf->max_degraded;
5348 int new_data_disks = conf->raid_disks - conf->max_degraded;
5351 sector_t writepos, readpos, safepos;
5352 sector_t stripe_addr;
5353 int reshape_sectors;
5354 struct list_head stripes;
5357 if (sector_nr == 0) {
5358 /* If restarting in the middle, skip the initial sectors */
5359 if (mddev->reshape_backwards &&
5360 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5361 sector_nr = raid5_size(mddev, 0, 0)
5362 - conf->reshape_progress;
5363 } else if (mddev->reshape_backwards &&
5364 conf->reshape_progress == MaxSector) {
5365 /* shouldn't happen, but just in case, finish up.*/
5366 sector_nr = MaxSector;
5367 } else if (!mddev->reshape_backwards &&
5368 conf->reshape_progress > 0)
5369 sector_nr = conf->reshape_progress;
5370 sector_div(sector_nr, new_data_disks);
5372 mddev->curr_resync_completed = sector_nr;
5373 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5380 /* We need to process a full chunk at a time.
5381 * If old and new chunk sizes differ, we need to process the
5385 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5387 /* We update the metadata at least every 10 seconds, or when
5388 * the data about to be copied would over-write the source of
5389 * the data at the front of the range. i.e. one new_stripe
5390 * along from reshape_progress new_maps to after where
5391 * reshape_safe old_maps to
5393 writepos = conf->reshape_progress;
5394 sector_div(writepos, new_data_disks);
5395 readpos = conf->reshape_progress;
5396 sector_div(readpos, data_disks);
5397 safepos = conf->reshape_safe;
5398 sector_div(safepos, data_disks);
5399 if (mddev->reshape_backwards) {
5400 BUG_ON(writepos < reshape_sectors);
5401 writepos -= reshape_sectors;
5402 readpos += reshape_sectors;
5403 safepos += reshape_sectors;
5405 writepos += reshape_sectors;
5406 /* readpos and safepos are worst-case calculations.
5407 * A negative number is overly pessimistic, and causes
5408 * obvious problems for unsigned storage. So clip to 0.
5410 readpos -= min_t(sector_t, reshape_sectors, readpos);
5411 safepos -= min_t(sector_t, reshape_sectors, safepos);
5414 /* Having calculated the 'writepos' possibly use it
5415 * to set 'stripe_addr' which is where we will write to.
5417 if (mddev->reshape_backwards) {
5418 BUG_ON(conf->reshape_progress == 0);
5419 stripe_addr = writepos;
5420 BUG_ON((mddev->dev_sectors &
5421 ~((sector_t)reshape_sectors - 1))
5422 - reshape_sectors - stripe_addr
5425 BUG_ON(writepos != sector_nr + reshape_sectors);
5426 stripe_addr = sector_nr;
5429 /* 'writepos' is the most advanced device address we might write.
5430 * 'readpos' is the least advanced device address we might read.
5431 * 'safepos' is the least address recorded in the metadata as having
5433 * If there is a min_offset_diff, these are adjusted either by
5434 * increasing the safepos/readpos if diff is negative, or
5435 * increasing writepos if diff is positive.
5436 * If 'readpos' is then behind 'writepos', there is no way that we can
5437 * ensure safety in the face of a crash - that must be done by userspace
5438 * making a backup of the data. So in that case there is no particular
5439 * rush to update metadata.
5440 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5441 * update the metadata to advance 'safepos' to match 'readpos' so that
5442 * we can be safe in the event of a crash.
5443 * So we insist on updating metadata if safepos is behind writepos and
5444 * readpos is beyond writepos.
5445 * In any case, update the metadata every 10 seconds.
5446 * Maybe that number should be configurable, but I'm not sure it is
5447 * worth it.... maybe it could be a multiple of safemode_delay???
5449 if (conf->min_offset_diff < 0) {
5450 safepos += -conf->min_offset_diff;
5451 readpos += -conf->min_offset_diff;
5453 writepos += conf->min_offset_diff;
5455 if ((mddev->reshape_backwards
5456 ? (safepos > writepos && readpos < writepos)
5457 : (safepos < writepos && readpos > writepos)) ||
5458 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5459 /* Cannot proceed until we've updated the superblock... */
5460 wait_event(conf->wait_for_overlap,
5461 atomic_read(&conf->reshape_stripes)==0
5462 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5463 if (atomic_read(&conf->reshape_stripes) != 0)
5465 mddev->reshape_position = conf->reshape_progress;
5466 mddev->curr_resync_completed = sector_nr;
5467 conf->reshape_checkpoint = jiffies;
5468 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5469 md_wakeup_thread(mddev->thread);
5470 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5471 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5472 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5474 spin_lock_irq(&conf->device_lock);
5475 conf->reshape_safe = mddev->reshape_position;
5476 spin_unlock_irq(&conf->device_lock);
5477 wake_up(&conf->wait_for_overlap);
5478 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5481 INIT_LIST_HEAD(&stripes);
5482 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5484 int skipped_disk = 0;
5485 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5486 set_bit(STRIPE_EXPANDING, &sh->state);
5487 atomic_inc(&conf->reshape_stripes);
5488 /* If any of this stripe is beyond the end of the old
5489 * array, then we need to zero those blocks
5491 for (j=sh->disks; j--;) {
5493 if (j == sh->pd_idx)
5495 if (conf->level == 6 &&
5498 s = raid5_compute_blocknr(sh, j, 0);
5499 if (s < raid5_size(mddev, 0, 0)) {
5503 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5504 set_bit(R5_Expanded, &sh->dev[j].flags);
5505 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5507 if (!skipped_disk) {
5508 set_bit(STRIPE_EXPAND_READY, &sh->state);
5509 set_bit(STRIPE_HANDLE, &sh->state);
5511 list_add(&sh->lru, &stripes);
5513 spin_lock_irq(&conf->device_lock);
5514 if (mddev->reshape_backwards)
5515 conf->reshape_progress -= reshape_sectors * new_data_disks;
5517 conf->reshape_progress += reshape_sectors * new_data_disks;
5518 spin_unlock_irq(&conf->device_lock);
5519 /* Ok, those stripe are ready. We can start scheduling
5520 * reads on the source stripes.
5521 * The source stripes are determined by mapping the first and last
5522 * block on the destination stripes.
5525 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5528 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5529 * new_data_disks - 1),
5531 if (last_sector >= mddev->dev_sectors)
5532 last_sector = mddev->dev_sectors - 1;
5533 while (first_sector <= last_sector) {
5534 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5535 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5536 set_bit(STRIPE_HANDLE, &sh->state);
5537 raid5_release_stripe(sh);
5538 first_sector += STRIPE_SECTORS;
5540 /* Now that the sources are clearly marked, we can release
5541 * the destination stripes
5543 while (!list_empty(&stripes)) {
5544 sh = list_entry(stripes.next, struct stripe_head, lru);
5545 list_del_init(&sh->lru);
5546 raid5_release_stripe(sh);
5548 /* If this takes us to the resync_max point where we have to pause,
5549 * then we need to write out the superblock.
5551 sector_nr += reshape_sectors;
5552 retn = reshape_sectors;
5554 if (mddev->curr_resync_completed > mddev->resync_max ||
5555 (sector_nr - mddev->curr_resync_completed) * 2
5556 >= mddev->resync_max - mddev->curr_resync_completed) {
5557 /* Cannot proceed until we've updated the superblock... */
5558 wait_event(conf->wait_for_overlap,
5559 atomic_read(&conf->reshape_stripes) == 0
5560 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5561 if (atomic_read(&conf->reshape_stripes) != 0)
5563 mddev->reshape_position = conf->reshape_progress;
5564 mddev->curr_resync_completed = sector_nr;
5565 conf->reshape_checkpoint = jiffies;
5566 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5567 md_wakeup_thread(mddev->thread);
5568 wait_event(mddev->sb_wait,
5569 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5570 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5571 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5573 spin_lock_irq(&conf->device_lock);
5574 conf->reshape_safe = mddev->reshape_position;
5575 spin_unlock_irq(&conf->device_lock);
5576 wake_up(&conf->wait_for_overlap);
5577 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5583 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5585 struct r5conf *conf = mddev->private;
5586 struct stripe_head *sh;
5587 sector_t max_sector = mddev->dev_sectors;
5588 sector_t sync_blocks;
5589 int still_degraded = 0;
5592 if (sector_nr >= max_sector) {
5593 /* just being told to finish up .. nothing much to do */
5595 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5600 if (mddev->curr_resync < max_sector) /* aborted */
5601 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5603 else /* completed sync */
5605 bitmap_close_sync(mddev->bitmap);
5610 /* Allow raid5_quiesce to complete */
5611 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5613 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5614 return reshape_request(mddev, sector_nr, skipped);
5616 /* No need to check resync_max as we never do more than one
5617 * stripe, and as resync_max will always be on a chunk boundary,
5618 * if the check in md_do_sync didn't fire, there is no chance
5619 * of overstepping resync_max here
5622 /* if there is too many failed drives and we are trying
5623 * to resync, then assert that we are finished, because there is
5624 * nothing we can do.
5626 if (mddev->degraded >= conf->max_degraded &&
5627 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5628 sector_t rv = mddev->dev_sectors - sector_nr;
5632 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5634 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5635 sync_blocks >= STRIPE_SECTORS) {
5636 /* we can skip this block, and probably more */
5637 sync_blocks /= STRIPE_SECTORS;
5639 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5642 bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5644 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5646 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5647 /* make sure we don't swamp the stripe cache if someone else
5648 * is trying to get access
5650 schedule_timeout_uninterruptible(1);
5652 /* Need to check if array will still be degraded after recovery/resync
5653 * Note in case of > 1 drive failures it's possible we're rebuilding
5654 * one drive while leaving another faulty drive in array.
5657 for (i = 0; i < conf->raid_disks; i++) {
5658 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5660 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5665 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5667 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5668 set_bit(STRIPE_HANDLE, &sh->state);
5670 raid5_release_stripe(sh);
5672 return STRIPE_SECTORS;
5675 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5677 /* We may not be able to submit a whole bio at once as there
5678 * may not be enough stripe_heads available.
5679 * We cannot pre-allocate enough stripe_heads as we may need
5680 * more than exist in the cache (if we allow ever large chunks).
5681 * So we do one stripe head at a time and record in
5682 * ->bi_hw_segments how many have been done.
5684 * We *know* that this entire raid_bio is in one chunk, so
5685 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5687 struct stripe_head *sh;
5689 sector_t sector, logical_sector, last_sector;
5694 logical_sector = raid_bio->bi_iter.bi_sector &
5695 ~((sector_t)STRIPE_SECTORS-1);
5696 sector = raid5_compute_sector(conf, logical_sector,
5698 last_sector = bio_end_sector(raid_bio);
5700 for (; logical_sector < last_sector;
5701 logical_sector += STRIPE_SECTORS,
5702 sector += STRIPE_SECTORS,
5705 if (scnt < raid5_bi_processed_stripes(raid_bio))
5706 /* already done this stripe */
5709 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5712 /* failed to get a stripe - must wait */
5713 raid5_set_bi_processed_stripes(raid_bio, scnt);
5714 conf->retry_read_aligned = raid_bio;
5718 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5719 raid5_release_stripe(sh);
5720 raid5_set_bi_processed_stripes(raid_bio, scnt);
5721 conf->retry_read_aligned = raid_bio;
5725 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5727 raid5_release_stripe(sh);
5730 remaining = raid5_dec_bi_active_stripes(raid_bio);
5731 if (remaining == 0) {
5732 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5734 bio_endio(raid_bio);
5736 if (atomic_dec_and_test(&conf->active_aligned_reads))
5737 wake_up(&conf->wait_for_quiescent);
5741 static int handle_active_stripes(struct r5conf *conf, int group,
5742 struct r5worker *worker,
5743 struct list_head *temp_inactive_list)
5745 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5746 int i, batch_size = 0, hash;
5747 bool release_inactive = false;
5749 while (batch_size < MAX_STRIPE_BATCH &&
5750 (sh = __get_priority_stripe(conf, group)) != NULL)
5751 batch[batch_size++] = sh;
5753 if (batch_size == 0) {
5754 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5755 if (!list_empty(temp_inactive_list + i))
5757 if (i == NR_STRIPE_HASH_LOCKS) {
5758 spin_unlock_irq(&conf->device_lock);
5759 r5l_flush_stripe_to_raid(conf->log);
5760 spin_lock_irq(&conf->device_lock);
5763 release_inactive = true;
5765 spin_unlock_irq(&conf->device_lock);
5767 release_inactive_stripe_list(conf, temp_inactive_list,
5768 NR_STRIPE_HASH_LOCKS);
5770 r5l_flush_stripe_to_raid(conf->log);
5771 if (release_inactive) {
5772 spin_lock_irq(&conf->device_lock);
5776 for (i = 0; i < batch_size; i++)
5777 handle_stripe(batch[i]);
5778 r5l_write_stripe_run(conf->log);
5782 spin_lock_irq(&conf->device_lock);
5783 for (i = 0; i < batch_size; i++) {
5784 hash = batch[i]->hash_lock_index;
5785 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5790 static void raid5_do_work(struct work_struct *work)
5792 struct r5worker *worker = container_of(work, struct r5worker, work);
5793 struct r5worker_group *group = worker->group;
5794 struct r5conf *conf = group->conf;
5795 int group_id = group - conf->worker_groups;
5797 struct blk_plug plug;
5799 pr_debug("+++ raid5worker active\n");
5801 blk_start_plug(&plug);
5803 spin_lock_irq(&conf->device_lock);
5805 int batch_size, released;
5807 released = release_stripe_list(conf, worker->temp_inactive_list);
5809 batch_size = handle_active_stripes(conf, group_id, worker,
5810 worker->temp_inactive_list);
5811 worker->working = false;
5812 if (!batch_size && !released)
5814 handled += batch_size;
5816 pr_debug("%d stripes handled\n", handled);
5818 spin_unlock_irq(&conf->device_lock);
5819 blk_finish_plug(&plug);
5821 pr_debug("--- raid5worker inactive\n");
5825 * This is our raid5 kernel thread.
5827 * We scan the hash table for stripes which can be handled now.
5828 * During the scan, completed stripes are saved for us by the interrupt
5829 * handler, so that they will not have to wait for our next wakeup.
5831 static void raid5d(struct md_thread *thread)
5833 struct mddev *mddev = thread->mddev;
5834 struct r5conf *conf = mddev->private;
5836 struct blk_plug plug;
5838 pr_debug("+++ raid5d active\n");
5840 md_check_recovery(mddev);
5842 if (!bio_list_empty(&conf->return_bi) &&
5843 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5844 struct bio_list tmp = BIO_EMPTY_LIST;
5845 spin_lock_irq(&conf->device_lock);
5846 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5847 bio_list_merge(&tmp, &conf->return_bi);
5848 bio_list_init(&conf->return_bi);
5850 spin_unlock_irq(&conf->device_lock);
5854 blk_start_plug(&plug);
5856 spin_lock_irq(&conf->device_lock);
5859 int batch_size, released;
5861 released = release_stripe_list(conf, conf->temp_inactive_list);
5863 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5866 !list_empty(&conf->bitmap_list)) {
5867 /* Now is a good time to flush some bitmap updates */
5869 spin_unlock_irq(&conf->device_lock);
5870 bitmap_unplug(mddev->bitmap);
5871 spin_lock_irq(&conf->device_lock);
5872 conf->seq_write = conf->seq_flush;
5873 activate_bit_delay(conf, conf->temp_inactive_list);
5875 raid5_activate_delayed(conf);
5877 while ((bio = remove_bio_from_retry(conf))) {
5879 spin_unlock_irq(&conf->device_lock);
5880 ok = retry_aligned_read(conf, bio);
5881 spin_lock_irq(&conf->device_lock);
5887 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5888 conf->temp_inactive_list);
5889 if (!batch_size && !released)
5891 handled += batch_size;
5893 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5894 spin_unlock_irq(&conf->device_lock);
5895 md_check_recovery(mddev);
5896 spin_lock_irq(&conf->device_lock);
5899 pr_debug("%d stripes handled\n", handled);
5901 spin_unlock_irq(&conf->device_lock);
5902 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5903 mutex_trylock(&conf->cache_size_mutex)) {
5904 grow_one_stripe(conf, __GFP_NOWARN);
5905 /* Set flag even if allocation failed. This helps
5906 * slow down allocation requests when mem is short
5908 set_bit(R5_DID_ALLOC, &conf->cache_state);
5909 mutex_unlock(&conf->cache_size_mutex);
5912 r5l_flush_stripe_to_raid(conf->log);
5914 async_tx_issue_pending_all();
5915 blk_finish_plug(&plug);
5917 pr_debug("--- raid5d inactive\n");
5921 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5923 struct r5conf *conf;
5925 spin_lock(&mddev->lock);
5926 conf = mddev->private;
5928 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5929 spin_unlock(&mddev->lock);
5934 raid5_set_cache_size(struct mddev *mddev, int size)
5936 struct r5conf *conf = mddev->private;
5939 if (size <= 16 || size > 32768)
5942 conf->min_nr_stripes = size;
5943 mutex_lock(&conf->cache_size_mutex);
5944 while (size < conf->max_nr_stripes &&
5945 drop_one_stripe(conf))
5947 mutex_unlock(&conf->cache_size_mutex);
5950 err = md_allow_write(mddev);
5954 mutex_lock(&conf->cache_size_mutex);
5955 while (size > conf->max_nr_stripes)
5956 if (!grow_one_stripe(conf, GFP_KERNEL))
5958 mutex_unlock(&conf->cache_size_mutex);
5962 EXPORT_SYMBOL(raid5_set_cache_size);
5965 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5967 struct r5conf *conf;
5971 if (len >= PAGE_SIZE)
5973 if (kstrtoul(page, 10, &new))
5975 err = mddev_lock(mddev);
5978 conf = mddev->private;
5982 err = raid5_set_cache_size(mddev, new);
5983 mddev_unlock(mddev);
5988 static struct md_sysfs_entry
5989 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5990 raid5_show_stripe_cache_size,
5991 raid5_store_stripe_cache_size);
5994 raid5_show_rmw_level(struct mddev *mddev, char *page)
5996 struct r5conf *conf = mddev->private;
5998 return sprintf(page, "%d\n", conf->rmw_level);
6004 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6006 struct r5conf *conf = mddev->private;
6012 if (len >= PAGE_SIZE)
6015 if (kstrtoul(page, 10, &new))
6018 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6021 if (new != PARITY_DISABLE_RMW &&
6022 new != PARITY_ENABLE_RMW &&
6023 new != PARITY_PREFER_RMW)
6026 conf->rmw_level = new;
6030 static struct md_sysfs_entry
6031 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6032 raid5_show_rmw_level,
6033 raid5_store_rmw_level);
6037 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6039 struct r5conf *conf;
6041 spin_lock(&mddev->lock);
6042 conf = mddev->private;
6044 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6045 spin_unlock(&mddev->lock);
6050 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6052 struct r5conf *conf;
6056 if (len >= PAGE_SIZE)
6058 if (kstrtoul(page, 10, &new))
6061 err = mddev_lock(mddev);
6064 conf = mddev->private;
6067 else if (new > conf->min_nr_stripes)
6070 conf->bypass_threshold = new;
6071 mddev_unlock(mddev);
6075 static struct md_sysfs_entry
6076 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6078 raid5_show_preread_threshold,
6079 raid5_store_preread_threshold);
6082 raid5_show_skip_copy(struct mddev *mddev, char *page)
6084 struct r5conf *conf;
6086 spin_lock(&mddev->lock);
6087 conf = mddev->private;
6089 ret = sprintf(page, "%d\n", conf->skip_copy);
6090 spin_unlock(&mddev->lock);
6095 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6097 struct r5conf *conf;
6101 if (len >= PAGE_SIZE)
6103 if (kstrtoul(page, 10, &new))
6107 err = mddev_lock(mddev);
6110 conf = mddev->private;
6113 else if (new != conf->skip_copy) {
6114 mddev_suspend(mddev);
6115 conf->skip_copy = new;
6117 mddev->queue->backing_dev_info.capabilities |=
6118 BDI_CAP_STABLE_WRITES;
6120 mddev->queue->backing_dev_info.capabilities &=
6121 ~BDI_CAP_STABLE_WRITES;
6122 mddev_resume(mddev);
6124 mddev_unlock(mddev);
6128 static struct md_sysfs_entry
6129 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6130 raid5_show_skip_copy,
6131 raid5_store_skip_copy);
6134 stripe_cache_active_show(struct mddev *mddev, char *page)
6136 struct r5conf *conf = mddev->private;
6138 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6143 static struct md_sysfs_entry
6144 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6147 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6149 struct r5conf *conf;
6151 spin_lock(&mddev->lock);
6152 conf = mddev->private;
6154 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6155 spin_unlock(&mddev->lock);
6159 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6161 int *worker_cnt_per_group,
6162 struct r5worker_group **worker_groups);
6164 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6166 struct r5conf *conf;
6169 struct r5worker_group *new_groups, *old_groups;
6170 int group_cnt, worker_cnt_per_group;
6172 if (len >= PAGE_SIZE)
6174 if (kstrtoul(page, 10, &new))
6177 err = mddev_lock(mddev);
6180 conf = mddev->private;
6183 else if (new != conf->worker_cnt_per_group) {
6184 mddev_suspend(mddev);
6186 old_groups = conf->worker_groups;
6188 flush_workqueue(raid5_wq);
6190 err = alloc_thread_groups(conf, new,
6191 &group_cnt, &worker_cnt_per_group,
6194 spin_lock_irq(&conf->device_lock);
6195 conf->group_cnt = group_cnt;
6196 conf->worker_cnt_per_group = worker_cnt_per_group;
6197 conf->worker_groups = new_groups;
6198 spin_unlock_irq(&conf->device_lock);
6201 kfree(old_groups[0].workers);
6204 mddev_resume(mddev);
6206 mddev_unlock(mddev);
6211 static struct md_sysfs_entry
6212 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6213 raid5_show_group_thread_cnt,
6214 raid5_store_group_thread_cnt);
6216 static struct attribute *raid5_attrs[] = {
6217 &raid5_stripecache_size.attr,
6218 &raid5_stripecache_active.attr,
6219 &raid5_preread_bypass_threshold.attr,
6220 &raid5_group_thread_cnt.attr,
6221 &raid5_skip_copy.attr,
6222 &raid5_rmw_level.attr,
6225 static struct attribute_group raid5_attrs_group = {
6227 .attrs = raid5_attrs,
6230 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6232 int *worker_cnt_per_group,
6233 struct r5worker_group **worker_groups)
6237 struct r5worker *workers;
6239 *worker_cnt_per_group = cnt;
6242 *worker_groups = NULL;
6245 *group_cnt = num_possible_nodes();
6246 size = sizeof(struct r5worker) * cnt;
6247 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6248 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6249 *group_cnt, GFP_NOIO);
6250 if (!*worker_groups || !workers) {
6252 kfree(*worker_groups);
6256 for (i = 0; i < *group_cnt; i++) {
6257 struct r5worker_group *group;
6259 group = &(*worker_groups)[i];
6260 INIT_LIST_HEAD(&group->handle_list);
6262 group->workers = workers + i * cnt;
6264 for (j = 0; j < cnt; j++) {
6265 struct r5worker *worker = group->workers + j;
6266 worker->group = group;
6267 INIT_WORK(&worker->work, raid5_do_work);
6269 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6270 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6277 static void free_thread_groups(struct r5conf *conf)
6279 if (conf->worker_groups)
6280 kfree(conf->worker_groups[0].workers);
6281 kfree(conf->worker_groups);
6282 conf->worker_groups = NULL;
6286 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6288 struct r5conf *conf = mddev->private;
6291 sectors = mddev->dev_sectors;
6293 /* size is defined by the smallest of previous and new size */
6294 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6296 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6297 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6298 return sectors * (raid_disks - conf->max_degraded);
6301 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6303 safe_put_page(percpu->spare_page);
6304 if (percpu->scribble)
6305 flex_array_free(percpu->scribble);
6306 percpu->spare_page = NULL;
6307 percpu->scribble = NULL;
6310 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6312 if (conf->level == 6 && !percpu->spare_page)
6313 percpu->spare_page = alloc_page(GFP_KERNEL);
6314 if (!percpu->scribble)
6315 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6316 conf->previous_raid_disks),
6317 max(conf->chunk_sectors,
6318 conf->prev_chunk_sectors)
6322 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6323 free_scratch_buffer(conf, percpu);
6330 static void raid5_free_percpu(struct r5conf *conf)
6337 #ifdef CONFIG_HOTPLUG_CPU
6338 unregister_cpu_notifier(&conf->cpu_notify);
6342 for_each_possible_cpu(cpu)
6343 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6346 free_percpu(conf->percpu);
6349 static void free_conf(struct r5conf *conf)
6352 r5l_exit_log(conf->log);
6353 if (conf->shrinker.seeks)
6354 unregister_shrinker(&conf->shrinker);
6356 free_thread_groups(conf);
6357 shrink_stripes(conf);
6358 raid5_free_percpu(conf);
6360 kfree(conf->stripe_hashtbl);
6364 #ifdef CONFIG_HOTPLUG_CPU
6365 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6368 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6369 long cpu = (long)hcpu;
6370 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6373 case CPU_UP_PREPARE:
6374 case CPU_UP_PREPARE_FROZEN:
6375 if (alloc_scratch_buffer(conf, percpu)) {
6376 pr_err("%s: failed memory allocation for cpu%ld\n",
6378 return notifier_from_errno(-ENOMEM);
6382 case CPU_DEAD_FROZEN:
6383 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6392 static int raid5_alloc_percpu(struct r5conf *conf)
6397 conf->percpu = alloc_percpu(struct raid5_percpu);
6401 #ifdef CONFIG_HOTPLUG_CPU
6402 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6403 conf->cpu_notify.priority = 0;
6404 err = register_cpu_notifier(&conf->cpu_notify);
6410 for_each_present_cpu(cpu) {
6411 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6413 pr_err("%s: failed memory allocation for cpu%ld\n",
6421 conf->scribble_disks = max(conf->raid_disks,
6422 conf->previous_raid_disks);
6423 conf->scribble_sectors = max(conf->chunk_sectors,
6424 conf->prev_chunk_sectors);
6429 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6430 struct shrink_control *sc)
6432 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6433 unsigned long ret = SHRINK_STOP;
6435 if (mutex_trylock(&conf->cache_size_mutex)) {
6437 while (ret < sc->nr_to_scan &&
6438 conf->max_nr_stripes > conf->min_nr_stripes) {
6439 if (drop_one_stripe(conf) == 0) {
6445 mutex_unlock(&conf->cache_size_mutex);
6450 static unsigned long raid5_cache_count(struct shrinker *shrink,
6451 struct shrink_control *sc)
6453 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6455 if (conf->max_nr_stripes < conf->min_nr_stripes)
6456 /* unlikely, but not impossible */
6458 return conf->max_nr_stripes - conf->min_nr_stripes;
6461 static struct r5conf *setup_conf(struct mddev *mddev)
6463 struct r5conf *conf;
6464 int raid_disk, memory, max_disks;
6465 struct md_rdev *rdev;
6466 struct disk_info *disk;
6469 int group_cnt, worker_cnt_per_group;
6470 struct r5worker_group *new_group;
6472 if (mddev->new_level != 5
6473 && mddev->new_level != 4
6474 && mddev->new_level != 6) {
6475 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6476 mdname(mddev), mddev->new_level);
6477 return ERR_PTR(-EIO);
6479 if ((mddev->new_level == 5
6480 && !algorithm_valid_raid5(mddev->new_layout)) ||
6481 (mddev->new_level == 6
6482 && !algorithm_valid_raid6(mddev->new_layout))) {
6483 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6484 mdname(mddev), mddev->new_layout);
6485 return ERR_PTR(-EIO);
6487 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6488 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6489 mdname(mddev), mddev->raid_disks);
6490 return ERR_PTR(-EINVAL);
6493 if (!mddev->new_chunk_sectors ||
6494 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6495 !is_power_of_2(mddev->new_chunk_sectors)) {
6496 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6497 mdname(mddev), mddev->new_chunk_sectors << 9);
6498 return ERR_PTR(-EINVAL);
6501 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6504 /* Don't enable multi-threading by default*/
6505 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6507 conf->group_cnt = group_cnt;
6508 conf->worker_cnt_per_group = worker_cnt_per_group;
6509 conf->worker_groups = new_group;
6512 spin_lock_init(&conf->device_lock);
6513 seqcount_init(&conf->gen_lock);
6514 mutex_init(&conf->cache_size_mutex);
6515 init_waitqueue_head(&conf->wait_for_quiescent);
6516 init_waitqueue_head(&conf->wait_for_stripe);
6517 init_waitqueue_head(&conf->wait_for_overlap);
6518 INIT_LIST_HEAD(&conf->handle_list);
6519 INIT_LIST_HEAD(&conf->hold_list);
6520 INIT_LIST_HEAD(&conf->delayed_list);
6521 INIT_LIST_HEAD(&conf->bitmap_list);
6522 bio_list_init(&conf->return_bi);
6523 init_llist_head(&conf->released_stripes);
6524 atomic_set(&conf->active_stripes, 0);
6525 atomic_set(&conf->preread_active_stripes, 0);
6526 atomic_set(&conf->active_aligned_reads, 0);
6527 conf->bypass_threshold = BYPASS_THRESHOLD;
6528 conf->recovery_disabled = mddev->recovery_disabled - 1;
6530 conf->raid_disks = mddev->raid_disks;
6531 if (mddev->reshape_position == MaxSector)
6532 conf->previous_raid_disks = mddev->raid_disks;
6534 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6535 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6537 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6542 conf->mddev = mddev;
6544 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6547 /* We init hash_locks[0] separately to that it can be used
6548 * as the reference lock in the spin_lock_nest_lock() call
6549 * in lock_all_device_hash_locks_irq in order to convince
6550 * lockdep that we know what we are doing.
6552 spin_lock_init(conf->hash_locks);
6553 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6554 spin_lock_init(conf->hash_locks + i);
6556 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6557 INIT_LIST_HEAD(conf->inactive_list + i);
6559 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6560 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6562 conf->level = mddev->new_level;
6563 conf->chunk_sectors = mddev->new_chunk_sectors;
6564 if (raid5_alloc_percpu(conf) != 0)
6567 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6569 rdev_for_each(rdev, mddev) {
6570 raid_disk = rdev->raid_disk;
6571 if (raid_disk >= max_disks
6572 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6574 disk = conf->disks + raid_disk;
6576 if (test_bit(Replacement, &rdev->flags)) {
6577 if (disk->replacement)
6579 disk->replacement = rdev;
6586 if (test_bit(In_sync, &rdev->flags)) {
6587 char b[BDEVNAME_SIZE];
6588 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6590 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6591 } else if (rdev->saved_raid_disk != raid_disk)
6592 /* Cannot rely on bitmap to complete recovery */
6596 conf->level = mddev->new_level;
6597 if (conf->level == 6) {
6598 conf->max_degraded = 2;
6599 if (raid6_call.xor_syndrome)
6600 conf->rmw_level = PARITY_ENABLE_RMW;
6602 conf->rmw_level = PARITY_DISABLE_RMW;
6604 conf->max_degraded = 1;
6605 conf->rmw_level = PARITY_ENABLE_RMW;
6607 conf->algorithm = mddev->new_layout;
6608 conf->reshape_progress = mddev->reshape_position;
6609 if (conf->reshape_progress != MaxSector) {
6610 conf->prev_chunk_sectors = mddev->chunk_sectors;
6611 conf->prev_algo = mddev->layout;
6613 conf->prev_chunk_sectors = conf->chunk_sectors;
6614 conf->prev_algo = conf->algorithm;
6617 conf->min_nr_stripes = NR_STRIPES;
6618 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6619 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6620 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6621 if (grow_stripes(conf, conf->min_nr_stripes)) {
6623 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6624 mdname(mddev), memory);
6627 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6628 mdname(mddev), memory);
6630 * Losing a stripe head costs more than the time to refill it,
6631 * it reduces the queue depth and so can hurt throughput.
6632 * So set it rather large, scaled by number of devices.
6634 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6635 conf->shrinker.scan_objects = raid5_cache_scan;
6636 conf->shrinker.count_objects = raid5_cache_count;
6637 conf->shrinker.batch = 128;
6638 conf->shrinker.flags = 0;
6639 register_shrinker(&conf->shrinker);
6641 sprintf(pers_name, "raid%d", mddev->new_level);
6642 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6643 if (!conf->thread) {
6645 "md/raid:%s: couldn't allocate thread.\n",
6655 return ERR_PTR(-EIO);
6657 return ERR_PTR(-ENOMEM);
6660 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6663 case ALGORITHM_PARITY_0:
6664 if (raid_disk < max_degraded)
6667 case ALGORITHM_PARITY_N:
6668 if (raid_disk >= raid_disks - max_degraded)
6671 case ALGORITHM_PARITY_0_6:
6672 if (raid_disk == 0 ||
6673 raid_disk == raid_disks - 1)
6676 case ALGORITHM_LEFT_ASYMMETRIC_6:
6677 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6678 case ALGORITHM_LEFT_SYMMETRIC_6:
6679 case ALGORITHM_RIGHT_SYMMETRIC_6:
6680 if (raid_disk == raid_disks - 1)
6686 static int run(struct mddev *mddev)
6688 struct r5conf *conf;
6689 int working_disks = 0;
6690 int dirty_parity_disks = 0;
6691 struct md_rdev *rdev;
6692 struct md_rdev *journal_dev = NULL;
6693 sector_t reshape_offset = 0;
6695 long long min_offset_diff = 0;
6698 if (mddev->recovery_cp != MaxSector)
6699 printk(KERN_NOTICE "md/raid:%s: not clean"
6700 " -- starting background reconstruction\n",
6703 rdev_for_each(rdev, mddev) {
6706 if (test_bit(Journal, &rdev->flags)) {
6710 if (rdev->raid_disk < 0)
6712 diff = (rdev->new_data_offset - rdev->data_offset);
6714 min_offset_diff = diff;
6716 } else if (mddev->reshape_backwards &&
6717 diff < min_offset_diff)
6718 min_offset_diff = diff;
6719 else if (!mddev->reshape_backwards &&
6720 diff > min_offset_diff)
6721 min_offset_diff = diff;
6724 if (mddev->reshape_position != MaxSector) {
6725 /* Check that we can continue the reshape.
6726 * Difficulties arise if the stripe we would write to
6727 * next is at or after the stripe we would read from next.
6728 * For a reshape that changes the number of devices, this
6729 * is only possible for a very short time, and mdadm makes
6730 * sure that time appears to have past before assembling
6731 * the array. So we fail if that time hasn't passed.
6732 * For a reshape that keeps the number of devices the same
6733 * mdadm must be monitoring the reshape can keeping the
6734 * critical areas read-only and backed up. It will start
6735 * the array in read-only mode, so we check for that.
6737 sector_t here_new, here_old;
6739 int max_degraded = (mddev->level == 6 ? 2 : 1);
6744 printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6749 if (mddev->new_level != mddev->level) {
6750 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6751 "required - aborting.\n",
6755 old_disks = mddev->raid_disks - mddev->delta_disks;
6756 /* reshape_position must be on a new-stripe boundary, and one
6757 * further up in new geometry must map after here in old
6759 * If the chunk sizes are different, then as we perform reshape
6760 * in units of the largest of the two, reshape_position needs
6761 * be a multiple of the largest chunk size times new data disks.
6763 here_new = mddev->reshape_position;
6764 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6765 new_data_disks = mddev->raid_disks - max_degraded;
6766 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6767 printk(KERN_ERR "md/raid:%s: reshape_position not "
6768 "on a stripe boundary\n", mdname(mddev));
6771 reshape_offset = here_new * chunk_sectors;
6772 /* here_new is the stripe we will write to */
6773 here_old = mddev->reshape_position;
6774 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6775 /* here_old is the first stripe that we might need to read
6777 if (mddev->delta_disks == 0) {
6778 /* We cannot be sure it is safe to start an in-place
6779 * reshape. It is only safe if user-space is monitoring
6780 * and taking constant backups.
6781 * mdadm always starts a situation like this in
6782 * readonly mode so it can take control before
6783 * allowing any writes. So just check for that.
6785 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6786 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6787 /* not really in-place - so OK */;
6788 else if (mddev->ro == 0) {
6789 printk(KERN_ERR "md/raid:%s: in-place reshape "
6790 "must be started in read-only mode "
6795 } else if (mddev->reshape_backwards
6796 ? (here_new * chunk_sectors + min_offset_diff <=
6797 here_old * chunk_sectors)
6798 : (here_new * chunk_sectors >=
6799 here_old * chunk_sectors + (-min_offset_diff))) {
6800 /* Reading from the same stripe as writing to - bad */
6801 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6802 "auto-recovery - aborting.\n",
6806 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6808 /* OK, we should be able to continue; */
6810 BUG_ON(mddev->level != mddev->new_level);
6811 BUG_ON(mddev->layout != mddev->new_layout);
6812 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6813 BUG_ON(mddev->delta_disks != 0);
6816 if (mddev->private == NULL)
6817 conf = setup_conf(mddev);
6819 conf = mddev->private;
6822 return PTR_ERR(conf);
6824 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) && !journal_dev) {
6825 printk(KERN_ERR "md/raid:%s: journal disk is missing, force array readonly\n",
6828 set_disk_ro(mddev->gendisk, 1);
6831 conf->min_offset_diff = min_offset_diff;
6832 mddev->thread = conf->thread;
6833 conf->thread = NULL;
6834 mddev->private = conf;
6836 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6838 rdev = conf->disks[i].rdev;
6839 if (!rdev && conf->disks[i].replacement) {
6840 /* The replacement is all we have yet */
6841 rdev = conf->disks[i].replacement;
6842 conf->disks[i].replacement = NULL;
6843 clear_bit(Replacement, &rdev->flags);
6844 conf->disks[i].rdev = rdev;
6848 if (conf->disks[i].replacement &&
6849 conf->reshape_progress != MaxSector) {
6850 /* replacements and reshape simply do not mix. */
6851 printk(KERN_ERR "md: cannot handle concurrent "
6852 "replacement and reshape.\n");
6855 if (test_bit(In_sync, &rdev->flags)) {
6859 /* This disc is not fully in-sync. However if it
6860 * just stored parity (beyond the recovery_offset),
6861 * when we don't need to be concerned about the
6862 * array being dirty.
6863 * When reshape goes 'backwards', we never have
6864 * partially completed devices, so we only need
6865 * to worry about reshape going forwards.
6867 /* Hack because v0.91 doesn't store recovery_offset properly. */
6868 if (mddev->major_version == 0 &&
6869 mddev->minor_version > 90)
6870 rdev->recovery_offset = reshape_offset;
6872 if (rdev->recovery_offset < reshape_offset) {
6873 /* We need to check old and new layout */
6874 if (!only_parity(rdev->raid_disk,
6877 conf->max_degraded))
6880 if (!only_parity(rdev->raid_disk,
6882 conf->previous_raid_disks,
6883 conf->max_degraded))
6885 dirty_parity_disks++;
6889 * 0 for a fully functional array, 1 or 2 for a degraded array.
6891 mddev->degraded = calc_degraded(conf);
6893 if (has_failed(conf)) {
6894 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6895 " (%d/%d failed)\n",
6896 mdname(mddev), mddev->degraded, conf->raid_disks);
6900 /* device size must be a multiple of chunk size */
6901 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6902 mddev->resync_max_sectors = mddev->dev_sectors;
6904 if (mddev->degraded > dirty_parity_disks &&
6905 mddev->recovery_cp != MaxSector) {
6906 if (mddev->ok_start_degraded)
6908 "md/raid:%s: starting dirty degraded array"
6909 " - data corruption possible.\n",
6913 "md/raid:%s: cannot start dirty degraded array.\n",
6919 if (mddev->degraded == 0)
6920 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6921 " devices, algorithm %d\n", mdname(mddev), conf->level,
6922 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6925 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6926 " out of %d devices, algorithm %d\n",
6927 mdname(mddev), conf->level,
6928 mddev->raid_disks - mddev->degraded,
6929 mddev->raid_disks, mddev->new_layout);
6931 print_raid5_conf(conf);
6933 if (conf->reshape_progress != MaxSector) {
6934 conf->reshape_safe = conf->reshape_progress;
6935 atomic_set(&conf->reshape_stripes, 0);
6936 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6937 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6938 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6939 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6940 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6944 /* Ok, everything is just fine now */
6945 if (mddev->to_remove == &raid5_attrs_group)
6946 mddev->to_remove = NULL;
6947 else if (mddev->kobj.sd &&
6948 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6950 "raid5: failed to create sysfs attributes for %s\n",
6952 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6956 bool discard_supported = true;
6957 /* read-ahead size must cover two whole stripes, which
6958 * is 2 * (datadisks) * chunksize where 'n' is the
6959 * number of raid devices
6961 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6962 int stripe = data_disks *
6963 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6964 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6965 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6967 chunk_size = mddev->chunk_sectors << 9;
6968 blk_queue_io_min(mddev->queue, chunk_size);
6969 blk_queue_io_opt(mddev->queue, chunk_size *
6970 (conf->raid_disks - conf->max_degraded));
6971 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6973 * We can only discard a whole stripe. It doesn't make sense to
6974 * discard data disk but write parity disk
6976 stripe = stripe * PAGE_SIZE;
6977 /* Round up to power of 2, as discard handling
6978 * currently assumes that */
6979 while ((stripe-1) & stripe)
6980 stripe = (stripe | (stripe-1)) + 1;
6981 mddev->queue->limits.discard_alignment = stripe;
6982 mddev->queue->limits.discard_granularity = stripe;
6985 * We use 16-bit counter of active stripes in bi_phys_segments
6986 * (minus one for over-loaded initialization)
6988 blk_queue_max_hw_sectors(mddev->queue, 0xfffe * STRIPE_SECTORS);
6989 blk_queue_max_discard_sectors(mddev->queue,
6990 0xfffe * STRIPE_SECTORS);
6993 * unaligned part of discard request will be ignored, so can't
6994 * guarantee discard_zeroes_data
6996 mddev->queue->limits.discard_zeroes_data = 0;
6998 blk_queue_max_write_same_sectors(mddev->queue, 0);
7000 rdev_for_each(rdev, mddev) {
7001 disk_stack_limits(mddev->gendisk, rdev->bdev,
7002 rdev->data_offset << 9);
7003 disk_stack_limits(mddev->gendisk, rdev->bdev,
7004 rdev->new_data_offset << 9);
7006 * discard_zeroes_data is required, otherwise data
7007 * could be lost. Consider a scenario: discard a stripe
7008 * (the stripe could be inconsistent if
7009 * discard_zeroes_data is 0); write one disk of the
7010 * stripe (the stripe could be inconsistent again
7011 * depending on which disks are used to calculate
7012 * parity); the disk is broken; The stripe data of this
7015 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7016 !bdev_get_queue(rdev->bdev)->
7017 limits.discard_zeroes_data)
7018 discard_supported = false;
7019 /* Unfortunately, discard_zeroes_data is not currently
7020 * a guarantee - just a hint. So we only allow DISCARD
7021 * if the sysadmin has confirmed that only safe devices
7022 * are in use by setting a module parameter.
7024 if (!devices_handle_discard_safely) {
7025 if (discard_supported) {
7026 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7027 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7029 discard_supported = false;
7033 if (discard_supported &&
7034 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7035 mddev->queue->limits.discard_granularity >= stripe)
7036 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7039 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7044 char b[BDEVNAME_SIZE];
7046 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7047 mdname(mddev), bdevname(journal_dev->bdev, b));
7048 r5l_init_log(conf, journal_dev);
7053 md_unregister_thread(&mddev->thread);
7054 print_raid5_conf(conf);
7056 mddev->private = NULL;
7057 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7061 static void raid5_free(struct mddev *mddev, void *priv)
7063 struct r5conf *conf = priv;
7066 mddev->to_remove = &raid5_attrs_group;
7069 static void status(struct seq_file *seq, struct mddev *mddev)
7071 struct r5conf *conf = mddev->private;
7074 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7075 conf->chunk_sectors / 2, mddev->layout);
7076 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7077 for (i = 0; i < conf->raid_disks; i++)
7078 seq_printf (seq, "%s",
7079 conf->disks[i].rdev &&
7080 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
7081 seq_printf (seq, "]");
7084 static void print_raid5_conf (struct r5conf *conf)
7087 struct disk_info *tmp;
7089 printk(KERN_DEBUG "RAID conf printout:\n");
7091 printk("(conf==NULL)\n");
7094 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7096 conf->raid_disks - conf->mddev->degraded);
7098 for (i = 0; i < conf->raid_disks; i++) {
7099 char b[BDEVNAME_SIZE];
7100 tmp = conf->disks + i;
7102 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7103 i, !test_bit(Faulty, &tmp->rdev->flags),
7104 bdevname(tmp->rdev->bdev, b));
7108 static int raid5_spare_active(struct mddev *mddev)
7111 struct r5conf *conf = mddev->private;
7112 struct disk_info *tmp;
7114 unsigned long flags;
7116 for (i = 0; i < conf->raid_disks; i++) {
7117 tmp = conf->disks + i;
7118 if (tmp->replacement
7119 && tmp->replacement->recovery_offset == MaxSector
7120 && !test_bit(Faulty, &tmp->replacement->flags)
7121 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7122 /* Replacement has just become active. */
7124 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7127 /* Replaced device not technically faulty,
7128 * but we need to be sure it gets removed
7129 * and never re-added.
7131 set_bit(Faulty, &tmp->rdev->flags);
7132 sysfs_notify_dirent_safe(
7133 tmp->rdev->sysfs_state);
7135 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7136 } else if (tmp->rdev
7137 && tmp->rdev->recovery_offset == MaxSector
7138 && !test_bit(Faulty, &tmp->rdev->flags)
7139 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7141 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7144 spin_lock_irqsave(&conf->device_lock, flags);
7145 mddev->degraded = calc_degraded(conf);
7146 spin_unlock_irqrestore(&conf->device_lock, flags);
7147 print_raid5_conf(conf);
7151 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7153 struct r5conf *conf = mddev->private;
7155 int number = rdev->raid_disk;
7156 struct md_rdev **rdevp;
7157 struct disk_info *p = conf->disks + number;
7159 print_raid5_conf(conf);
7160 if (test_bit(Journal, &rdev->flags)) {
7162 * journal disk is not removable, but we need give a chance to
7163 * update superblock of other disks. Otherwise journal disk
7164 * will be considered as 'fresh'
7166 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7169 if (rdev == p->rdev)
7171 else if (rdev == p->replacement)
7172 rdevp = &p->replacement;
7176 if (number >= conf->raid_disks &&
7177 conf->reshape_progress == MaxSector)
7178 clear_bit(In_sync, &rdev->flags);
7180 if (test_bit(In_sync, &rdev->flags) ||
7181 atomic_read(&rdev->nr_pending)) {
7185 /* Only remove non-faulty devices if recovery
7188 if (!test_bit(Faulty, &rdev->flags) &&
7189 mddev->recovery_disabled != conf->recovery_disabled &&
7190 !has_failed(conf) &&
7191 (!p->replacement || p->replacement == rdev) &&
7192 number < conf->raid_disks) {
7198 if (atomic_read(&rdev->nr_pending)) {
7199 /* lost the race, try later */
7202 } else if (p->replacement) {
7203 /* We must have just cleared 'rdev' */
7204 p->rdev = p->replacement;
7205 clear_bit(Replacement, &p->replacement->flags);
7206 smp_mb(); /* Make sure other CPUs may see both as identical
7207 * but will never see neither - if they are careful
7209 p->replacement = NULL;
7210 clear_bit(WantReplacement, &rdev->flags);
7212 /* We might have just removed the Replacement as faulty-
7213 * clear the bit just in case
7215 clear_bit(WantReplacement, &rdev->flags);
7218 print_raid5_conf(conf);
7222 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7224 struct r5conf *conf = mddev->private;
7227 struct disk_info *p;
7229 int last = conf->raid_disks - 1;
7231 if (test_bit(Journal, &rdev->flags))
7233 if (mddev->recovery_disabled == conf->recovery_disabled)
7236 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7237 /* no point adding a device */
7240 if (rdev->raid_disk >= 0)
7241 first = last = rdev->raid_disk;
7244 * find the disk ... but prefer rdev->saved_raid_disk
7247 if (rdev->saved_raid_disk >= 0 &&
7248 rdev->saved_raid_disk >= first &&
7249 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7250 first = rdev->saved_raid_disk;
7252 for (disk = first; disk <= last; disk++) {
7253 p = conf->disks + disk;
7254 if (p->rdev == NULL) {
7255 clear_bit(In_sync, &rdev->flags);
7256 rdev->raid_disk = disk;
7258 if (rdev->saved_raid_disk != disk)
7260 rcu_assign_pointer(p->rdev, rdev);
7264 for (disk = first; disk <= last; disk++) {
7265 p = conf->disks + disk;
7266 if (test_bit(WantReplacement, &p->rdev->flags) &&
7267 p->replacement == NULL) {
7268 clear_bit(In_sync, &rdev->flags);
7269 set_bit(Replacement, &rdev->flags);
7270 rdev->raid_disk = disk;
7273 rcu_assign_pointer(p->replacement, rdev);
7278 print_raid5_conf(conf);
7282 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7284 /* no resync is happening, and there is enough space
7285 * on all devices, so we can resize.
7286 * We need to make sure resync covers any new space.
7287 * If the array is shrinking we should possibly wait until
7288 * any io in the removed space completes, but it hardly seems
7292 struct r5conf *conf = mddev->private;
7296 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7297 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7298 if (mddev->external_size &&
7299 mddev->array_sectors > newsize)
7301 if (mddev->bitmap) {
7302 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7306 md_set_array_sectors(mddev, newsize);
7307 set_capacity(mddev->gendisk, mddev->array_sectors);
7308 revalidate_disk(mddev->gendisk);
7309 if (sectors > mddev->dev_sectors &&
7310 mddev->recovery_cp > mddev->dev_sectors) {
7311 mddev->recovery_cp = mddev->dev_sectors;
7312 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7314 mddev->dev_sectors = sectors;
7315 mddev->resync_max_sectors = sectors;
7319 static int check_stripe_cache(struct mddev *mddev)
7321 /* Can only proceed if there are plenty of stripe_heads.
7322 * We need a minimum of one full stripe,, and for sensible progress
7323 * it is best to have about 4 times that.
7324 * If we require 4 times, then the default 256 4K stripe_heads will
7325 * allow for chunk sizes up to 256K, which is probably OK.
7326 * If the chunk size is greater, user-space should request more
7327 * stripe_heads first.
7329 struct r5conf *conf = mddev->private;
7330 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7331 > conf->min_nr_stripes ||
7332 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7333 > conf->min_nr_stripes) {
7334 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7336 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7343 static int check_reshape(struct mddev *mddev)
7345 struct r5conf *conf = mddev->private;
7349 if (mddev->delta_disks == 0 &&
7350 mddev->new_layout == mddev->layout &&
7351 mddev->new_chunk_sectors == mddev->chunk_sectors)
7352 return 0; /* nothing to do */
7353 if (has_failed(conf))
7355 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7356 /* We might be able to shrink, but the devices must
7357 * be made bigger first.
7358 * For raid6, 4 is the minimum size.
7359 * Otherwise 2 is the minimum
7362 if (mddev->level == 6)
7364 if (mddev->raid_disks + mddev->delta_disks < min)
7368 if (!check_stripe_cache(mddev))
7371 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7372 mddev->delta_disks > 0)
7373 if (resize_chunks(conf,
7374 conf->previous_raid_disks
7375 + max(0, mddev->delta_disks),
7376 max(mddev->new_chunk_sectors,
7377 mddev->chunk_sectors)
7380 return resize_stripes(conf, (conf->previous_raid_disks
7381 + mddev->delta_disks));
7384 static int raid5_start_reshape(struct mddev *mddev)
7386 struct r5conf *conf = mddev->private;
7387 struct md_rdev *rdev;
7389 unsigned long flags;
7391 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7394 if (!check_stripe_cache(mddev))
7397 if (has_failed(conf))
7400 rdev_for_each(rdev, mddev) {
7401 if (!test_bit(In_sync, &rdev->flags)
7402 && !test_bit(Faulty, &rdev->flags))
7406 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7407 /* Not enough devices even to make a degraded array
7412 /* Refuse to reduce size of the array. Any reductions in
7413 * array size must be through explicit setting of array_size
7416 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7417 < mddev->array_sectors) {
7418 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7419 "before number of disks\n", mdname(mddev));
7423 atomic_set(&conf->reshape_stripes, 0);
7424 spin_lock_irq(&conf->device_lock);
7425 write_seqcount_begin(&conf->gen_lock);
7426 conf->previous_raid_disks = conf->raid_disks;
7427 conf->raid_disks += mddev->delta_disks;
7428 conf->prev_chunk_sectors = conf->chunk_sectors;
7429 conf->chunk_sectors = mddev->new_chunk_sectors;
7430 conf->prev_algo = conf->algorithm;
7431 conf->algorithm = mddev->new_layout;
7433 /* Code that selects data_offset needs to see the generation update
7434 * if reshape_progress has been set - so a memory barrier needed.
7437 if (mddev->reshape_backwards)
7438 conf->reshape_progress = raid5_size(mddev, 0, 0);
7440 conf->reshape_progress = 0;
7441 conf->reshape_safe = conf->reshape_progress;
7442 write_seqcount_end(&conf->gen_lock);
7443 spin_unlock_irq(&conf->device_lock);
7445 /* Now make sure any requests that proceeded on the assumption
7446 * the reshape wasn't running - like Discard or Read - have
7449 mddev_suspend(mddev);
7450 mddev_resume(mddev);
7452 /* Add some new drives, as many as will fit.
7453 * We know there are enough to make the newly sized array work.
7454 * Don't add devices if we are reducing the number of
7455 * devices in the array. This is because it is not possible
7456 * to correctly record the "partially reconstructed" state of
7457 * such devices during the reshape and confusion could result.
7459 if (mddev->delta_disks >= 0) {
7460 rdev_for_each(rdev, mddev)
7461 if (rdev->raid_disk < 0 &&
7462 !test_bit(Faulty, &rdev->flags)) {
7463 if (raid5_add_disk(mddev, rdev) == 0) {
7465 >= conf->previous_raid_disks)
7466 set_bit(In_sync, &rdev->flags);
7468 rdev->recovery_offset = 0;
7470 if (sysfs_link_rdev(mddev, rdev))
7471 /* Failure here is OK */;
7473 } else if (rdev->raid_disk >= conf->previous_raid_disks
7474 && !test_bit(Faulty, &rdev->flags)) {
7475 /* This is a spare that was manually added */
7476 set_bit(In_sync, &rdev->flags);
7479 /* When a reshape changes the number of devices,
7480 * ->degraded is measured against the larger of the
7481 * pre and post number of devices.
7483 spin_lock_irqsave(&conf->device_lock, flags);
7484 mddev->degraded = calc_degraded(conf);
7485 spin_unlock_irqrestore(&conf->device_lock, flags);
7487 mddev->raid_disks = conf->raid_disks;
7488 mddev->reshape_position = conf->reshape_progress;
7489 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7491 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7492 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7493 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7494 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7495 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7496 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7498 if (!mddev->sync_thread) {
7499 mddev->recovery = 0;
7500 spin_lock_irq(&conf->device_lock);
7501 write_seqcount_begin(&conf->gen_lock);
7502 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7503 mddev->new_chunk_sectors =
7504 conf->chunk_sectors = conf->prev_chunk_sectors;
7505 mddev->new_layout = conf->algorithm = conf->prev_algo;
7506 rdev_for_each(rdev, mddev)
7507 rdev->new_data_offset = rdev->data_offset;
7509 conf->generation --;
7510 conf->reshape_progress = MaxSector;
7511 mddev->reshape_position = MaxSector;
7512 write_seqcount_end(&conf->gen_lock);
7513 spin_unlock_irq(&conf->device_lock);
7516 conf->reshape_checkpoint = jiffies;
7517 md_wakeup_thread(mddev->sync_thread);
7518 md_new_event(mddev);
7522 /* This is called from the reshape thread and should make any
7523 * changes needed in 'conf'
7525 static void end_reshape(struct r5conf *conf)
7528 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7529 struct md_rdev *rdev;
7531 spin_lock_irq(&conf->device_lock);
7532 conf->previous_raid_disks = conf->raid_disks;
7533 rdev_for_each(rdev, conf->mddev)
7534 rdev->data_offset = rdev->new_data_offset;
7536 conf->reshape_progress = MaxSector;
7537 conf->mddev->reshape_position = MaxSector;
7538 spin_unlock_irq(&conf->device_lock);
7539 wake_up(&conf->wait_for_overlap);
7541 /* read-ahead size must cover two whole stripes, which is
7542 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7544 if (conf->mddev->queue) {
7545 int data_disks = conf->raid_disks - conf->max_degraded;
7546 int stripe = data_disks * ((conf->chunk_sectors << 9)
7548 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7549 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7554 /* This is called from the raid5d thread with mddev_lock held.
7555 * It makes config changes to the device.
7557 static void raid5_finish_reshape(struct mddev *mddev)
7559 struct r5conf *conf = mddev->private;
7561 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7563 if (mddev->delta_disks > 0) {
7564 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7565 set_capacity(mddev->gendisk, mddev->array_sectors);
7566 revalidate_disk(mddev->gendisk);
7569 spin_lock_irq(&conf->device_lock);
7570 mddev->degraded = calc_degraded(conf);
7571 spin_unlock_irq(&conf->device_lock);
7572 for (d = conf->raid_disks ;
7573 d < conf->raid_disks - mddev->delta_disks;
7575 struct md_rdev *rdev = conf->disks[d].rdev;
7577 clear_bit(In_sync, &rdev->flags);
7578 rdev = conf->disks[d].replacement;
7580 clear_bit(In_sync, &rdev->flags);
7583 mddev->layout = conf->algorithm;
7584 mddev->chunk_sectors = conf->chunk_sectors;
7585 mddev->reshape_position = MaxSector;
7586 mddev->delta_disks = 0;
7587 mddev->reshape_backwards = 0;
7591 static void raid5_quiesce(struct mddev *mddev, int state)
7593 struct r5conf *conf = mddev->private;
7596 case 2: /* resume for a suspend */
7597 wake_up(&conf->wait_for_overlap);
7600 case 1: /* stop all writes */
7601 lock_all_device_hash_locks_irq(conf);
7602 /* '2' tells resync/reshape to pause so that all
7603 * active stripes can drain
7606 wait_event_cmd(conf->wait_for_quiescent,
7607 atomic_read(&conf->active_stripes) == 0 &&
7608 atomic_read(&conf->active_aligned_reads) == 0,
7609 unlock_all_device_hash_locks_irq(conf),
7610 lock_all_device_hash_locks_irq(conf));
7612 unlock_all_device_hash_locks_irq(conf);
7613 /* allow reshape to continue */
7614 wake_up(&conf->wait_for_overlap);
7617 case 0: /* re-enable writes */
7618 lock_all_device_hash_locks_irq(conf);
7620 wake_up(&conf->wait_for_quiescent);
7621 wake_up(&conf->wait_for_overlap);
7622 unlock_all_device_hash_locks_irq(conf);
7625 r5l_quiesce(conf->log, state);
7628 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7630 struct r0conf *raid0_conf = mddev->private;
7633 /* for raid0 takeover only one zone is supported */
7634 if (raid0_conf->nr_strip_zones > 1) {
7635 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7637 return ERR_PTR(-EINVAL);
7640 sectors = raid0_conf->strip_zone[0].zone_end;
7641 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7642 mddev->dev_sectors = sectors;
7643 mddev->new_level = level;
7644 mddev->new_layout = ALGORITHM_PARITY_N;
7645 mddev->new_chunk_sectors = mddev->chunk_sectors;
7646 mddev->raid_disks += 1;
7647 mddev->delta_disks = 1;
7648 /* make sure it will be not marked as dirty */
7649 mddev->recovery_cp = MaxSector;
7651 return setup_conf(mddev);
7654 static void *raid5_takeover_raid1(struct mddev *mddev)
7658 if (mddev->raid_disks != 2 ||
7659 mddev->degraded > 1)
7660 return ERR_PTR(-EINVAL);
7662 /* Should check if there are write-behind devices? */
7664 chunksect = 64*2; /* 64K by default */
7666 /* The array must be an exact multiple of chunksize */
7667 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7670 if ((chunksect<<9) < STRIPE_SIZE)
7671 /* array size does not allow a suitable chunk size */
7672 return ERR_PTR(-EINVAL);
7674 mddev->new_level = 5;
7675 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7676 mddev->new_chunk_sectors = chunksect;
7678 return setup_conf(mddev);
7681 static void *raid5_takeover_raid6(struct mddev *mddev)
7685 switch (mddev->layout) {
7686 case ALGORITHM_LEFT_ASYMMETRIC_6:
7687 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7689 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7690 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7692 case ALGORITHM_LEFT_SYMMETRIC_6:
7693 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7695 case ALGORITHM_RIGHT_SYMMETRIC_6:
7696 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7698 case ALGORITHM_PARITY_0_6:
7699 new_layout = ALGORITHM_PARITY_0;
7701 case ALGORITHM_PARITY_N:
7702 new_layout = ALGORITHM_PARITY_N;
7705 return ERR_PTR(-EINVAL);
7707 mddev->new_level = 5;
7708 mddev->new_layout = new_layout;
7709 mddev->delta_disks = -1;
7710 mddev->raid_disks -= 1;
7711 return setup_conf(mddev);
7714 static int raid5_check_reshape(struct mddev *mddev)
7716 /* For a 2-drive array, the layout and chunk size can be changed
7717 * immediately as not restriping is needed.
7718 * For larger arrays we record the new value - after validation
7719 * to be used by a reshape pass.
7721 struct r5conf *conf = mddev->private;
7722 int new_chunk = mddev->new_chunk_sectors;
7724 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7726 if (new_chunk > 0) {
7727 if (!is_power_of_2(new_chunk))
7729 if (new_chunk < (PAGE_SIZE>>9))
7731 if (mddev->array_sectors & (new_chunk-1))
7732 /* not factor of array size */
7736 /* They look valid */
7738 if (mddev->raid_disks == 2) {
7739 /* can make the change immediately */
7740 if (mddev->new_layout >= 0) {
7741 conf->algorithm = mddev->new_layout;
7742 mddev->layout = mddev->new_layout;
7744 if (new_chunk > 0) {
7745 conf->chunk_sectors = new_chunk ;
7746 mddev->chunk_sectors = new_chunk;
7748 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7749 md_wakeup_thread(mddev->thread);
7751 return check_reshape(mddev);
7754 static int raid6_check_reshape(struct mddev *mddev)
7756 int new_chunk = mddev->new_chunk_sectors;
7758 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7760 if (new_chunk > 0) {
7761 if (!is_power_of_2(new_chunk))
7763 if (new_chunk < (PAGE_SIZE >> 9))
7765 if (mddev->array_sectors & (new_chunk-1))
7766 /* not factor of array size */
7770 /* They look valid */
7771 return check_reshape(mddev);
7774 static void *raid5_takeover(struct mddev *mddev)
7776 /* raid5 can take over:
7777 * raid0 - if there is only one strip zone - make it a raid4 layout
7778 * raid1 - if there are two drives. We need to know the chunk size
7779 * raid4 - trivial - just use a raid4 layout.
7780 * raid6 - Providing it is a *_6 layout
7782 if (mddev->level == 0)
7783 return raid45_takeover_raid0(mddev, 5);
7784 if (mddev->level == 1)
7785 return raid5_takeover_raid1(mddev);
7786 if (mddev->level == 4) {
7787 mddev->new_layout = ALGORITHM_PARITY_N;
7788 mddev->new_level = 5;
7789 return setup_conf(mddev);
7791 if (mddev->level == 6)
7792 return raid5_takeover_raid6(mddev);
7794 return ERR_PTR(-EINVAL);
7797 static void *raid4_takeover(struct mddev *mddev)
7799 /* raid4 can take over:
7800 * raid0 - if there is only one strip zone
7801 * raid5 - if layout is right
7803 if (mddev->level == 0)
7804 return raid45_takeover_raid0(mddev, 4);
7805 if (mddev->level == 5 &&
7806 mddev->layout == ALGORITHM_PARITY_N) {
7807 mddev->new_layout = 0;
7808 mddev->new_level = 4;
7809 return setup_conf(mddev);
7811 return ERR_PTR(-EINVAL);
7814 static struct md_personality raid5_personality;
7816 static void *raid6_takeover(struct mddev *mddev)
7818 /* Currently can only take over a raid5. We map the
7819 * personality to an equivalent raid6 personality
7820 * with the Q block at the end.
7824 if (mddev->pers != &raid5_personality)
7825 return ERR_PTR(-EINVAL);
7826 if (mddev->degraded > 1)
7827 return ERR_PTR(-EINVAL);
7828 if (mddev->raid_disks > 253)
7829 return ERR_PTR(-EINVAL);
7830 if (mddev->raid_disks < 3)
7831 return ERR_PTR(-EINVAL);
7833 switch (mddev->layout) {
7834 case ALGORITHM_LEFT_ASYMMETRIC:
7835 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7837 case ALGORITHM_RIGHT_ASYMMETRIC:
7838 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7840 case ALGORITHM_LEFT_SYMMETRIC:
7841 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7843 case ALGORITHM_RIGHT_SYMMETRIC:
7844 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7846 case ALGORITHM_PARITY_0:
7847 new_layout = ALGORITHM_PARITY_0_6;
7849 case ALGORITHM_PARITY_N:
7850 new_layout = ALGORITHM_PARITY_N;
7853 return ERR_PTR(-EINVAL);
7855 mddev->new_level = 6;
7856 mddev->new_layout = new_layout;
7857 mddev->delta_disks = 1;
7858 mddev->raid_disks += 1;
7859 return setup_conf(mddev);
7862 static struct md_personality raid6_personality =
7866 .owner = THIS_MODULE,
7867 .make_request = make_request,
7871 .error_handler = error,
7872 .hot_add_disk = raid5_add_disk,
7873 .hot_remove_disk= raid5_remove_disk,
7874 .spare_active = raid5_spare_active,
7875 .sync_request = sync_request,
7876 .resize = raid5_resize,
7878 .check_reshape = raid6_check_reshape,
7879 .start_reshape = raid5_start_reshape,
7880 .finish_reshape = raid5_finish_reshape,
7881 .quiesce = raid5_quiesce,
7882 .takeover = raid6_takeover,
7883 .congested = raid5_congested,
7885 static struct md_personality raid5_personality =
7889 .owner = THIS_MODULE,
7890 .make_request = make_request,
7894 .error_handler = error,
7895 .hot_add_disk = raid5_add_disk,
7896 .hot_remove_disk= raid5_remove_disk,
7897 .spare_active = raid5_spare_active,
7898 .sync_request = sync_request,
7899 .resize = raid5_resize,
7901 .check_reshape = raid5_check_reshape,
7902 .start_reshape = raid5_start_reshape,
7903 .finish_reshape = raid5_finish_reshape,
7904 .quiesce = raid5_quiesce,
7905 .takeover = raid5_takeover,
7906 .congested = raid5_congested,
7909 static struct md_personality raid4_personality =
7913 .owner = THIS_MODULE,
7914 .make_request = make_request,
7918 .error_handler = error,
7919 .hot_add_disk = raid5_add_disk,
7920 .hot_remove_disk= raid5_remove_disk,
7921 .spare_active = raid5_spare_active,
7922 .sync_request = sync_request,
7923 .resize = raid5_resize,
7925 .check_reshape = raid5_check_reshape,
7926 .start_reshape = raid5_start_reshape,
7927 .finish_reshape = raid5_finish_reshape,
7928 .quiesce = raid5_quiesce,
7929 .takeover = raid4_takeover,
7930 .congested = raid5_congested,
7933 static int __init raid5_init(void)
7935 raid5_wq = alloc_workqueue("raid5wq",
7936 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7939 register_md_personality(&raid6_personality);
7940 register_md_personality(&raid5_personality);
7941 register_md_personality(&raid4_personality);
7945 static void raid5_exit(void)
7947 unregister_md_personality(&raid6_personality);
7948 unregister_md_personality(&raid5_personality);
7949 unregister_md_personality(&raid4_personality);
7950 destroy_workqueue(raid5_wq);
7953 module_init(raid5_init);
7954 module_exit(raid5_exit);
7955 MODULE_LICENSE("GPL");
7956 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7957 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7958 MODULE_ALIAS("md-raid5");
7959 MODULE_ALIAS("md-raid4");
7960 MODULE_ALIAS("md-level-5");
7961 MODULE_ALIAS("md-level-4");
7962 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7963 MODULE_ALIAS("md-raid6");
7964 MODULE_ALIAS("md-level-6");
7966 /* This used to be two separate modules, they were: */
7967 MODULE_ALIAS("raid5");
7968 MODULE_ALIAS("raid6");