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 *return_bi)
228 struct bio *bi = return_bi;
231 return_bi = bi->bi_next;
233 bi->bi_iter.bi_size = 0;
234 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
241 static void print_raid5_conf (struct r5conf *conf);
243 static int stripe_operations_active(struct stripe_head *sh)
245 return sh->check_state || sh->reconstruct_state ||
246 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
247 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
250 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
252 struct r5conf *conf = sh->raid_conf;
253 struct r5worker_group *group;
255 int i, cpu = sh->cpu;
257 if (!cpu_online(cpu)) {
258 cpu = cpumask_any(cpu_online_mask);
262 if (list_empty(&sh->lru)) {
263 struct r5worker_group *group;
264 group = conf->worker_groups + cpu_to_group(cpu);
265 list_add_tail(&sh->lru, &group->handle_list);
266 group->stripes_cnt++;
270 if (conf->worker_cnt_per_group == 0) {
271 md_wakeup_thread(conf->mddev->thread);
275 group = conf->worker_groups + cpu_to_group(sh->cpu);
277 group->workers[0].working = true;
278 /* at least one worker should run to avoid race */
279 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
281 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
282 /* wakeup more workers */
283 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
284 if (group->workers[i].working == false) {
285 group->workers[i].working = true;
286 queue_work_on(sh->cpu, raid5_wq,
287 &group->workers[i].work);
293 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
294 struct list_head *temp_inactive_list)
296 BUG_ON(!list_empty(&sh->lru));
297 BUG_ON(atomic_read(&conf->active_stripes)==0);
298 if (test_bit(STRIPE_HANDLE, &sh->state)) {
299 if (test_bit(STRIPE_DELAYED, &sh->state) &&
300 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
301 list_add_tail(&sh->lru, &conf->delayed_list);
302 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
303 sh->bm_seq - conf->seq_write > 0)
304 list_add_tail(&sh->lru, &conf->bitmap_list);
306 clear_bit(STRIPE_DELAYED, &sh->state);
307 clear_bit(STRIPE_BIT_DELAY, &sh->state);
308 if (conf->worker_cnt_per_group == 0) {
309 list_add_tail(&sh->lru, &conf->handle_list);
311 raid5_wakeup_stripe_thread(sh);
315 md_wakeup_thread(conf->mddev->thread);
317 BUG_ON(stripe_operations_active(sh));
318 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
319 if (atomic_dec_return(&conf->preread_active_stripes)
321 md_wakeup_thread(conf->mddev->thread);
322 atomic_dec(&conf->active_stripes);
323 if (!test_bit(STRIPE_EXPANDING, &sh->state))
324 list_add_tail(&sh->lru, temp_inactive_list);
328 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
329 struct list_head *temp_inactive_list)
331 if (atomic_dec_and_test(&sh->count))
332 do_release_stripe(conf, sh, temp_inactive_list);
336 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
338 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
339 * given time. Adding stripes only takes device lock, while deleting stripes
340 * only takes hash lock.
342 static void release_inactive_stripe_list(struct r5conf *conf,
343 struct list_head *temp_inactive_list,
347 bool do_wakeup = false;
350 if (hash == NR_STRIPE_HASH_LOCKS) {
351 size = NR_STRIPE_HASH_LOCKS;
352 hash = NR_STRIPE_HASH_LOCKS - 1;
356 struct list_head *list = &temp_inactive_list[size - 1];
359 * We don't hold any lock here yet, get_active_stripe() might
360 * remove stripes from the list
362 if (!list_empty_careful(list)) {
363 spin_lock_irqsave(conf->hash_locks + hash, flags);
364 if (list_empty(conf->inactive_list + hash) &&
366 atomic_dec(&conf->empty_inactive_list_nr);
367 list_splice_tail_init(list, conf->inactive_list + hash);
369 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
376 wake_up(&conf->wait_for_stripe);
377 if (conf->retry_read_aligned)
378 md_wakeup_thread(conf->mddev->thread);
382 /* should hold conf->device_lock already */
383 static int release_stripe_list(struct r5conf *conf,
384 struct list_head *temp_inactive_list)
386 struct stripe_head *sh;
388 struct llist_node *head;
390 head = llist_del_all(&conf->released_stripes);
391 head = llist_reverse_order(head);
395 sh = llist_entry(head, struct stripe_head, release_list);
396 head = llist_next(head);
397 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
399 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
401 * Don't worry the bit is set here, because if the bit is set
402 * again, the count is always > 1. This is true for
403 * STRIPE_ON_UNPLUG_LIST bit too.
405 hash = sh->hash_lock_index;
406 __release_stripe(conf, sh, &temp_inactive_list[hash]);
413 static void release_stripe(struct stripe_head *sh)
415 struct r5conf *conf = sh->raid_conf;
417 struct list_head list;
421 /* Avoid release_list until the last reference.
423 if (atomic_add_unless(&sh->count, -1, 1))
426 if (unlikely(!conf->mddev->thread) ||
427 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
429 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
431 md_wakeup_thread(conf->mddev->thread);
434 local_irq_save(flags);
435 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
436 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
437 INIT_LIST_HEAD(&list);
438 hash = sh->hash_lock_index;
439 do_release_stripe(conf, sh, &list);
440 spin_unlock(&conf->device_lock);
441 release_inactive_stripe_list(conf, &list, hash);
443 local_irq_restore(flags);
446 static inline void remove_hash(struct stripe_head *sh)
448 pr_debug("remove_hash(), stripe %llu\n",
449 (unsigned long long)sh->sector);
451 hlist_del_init(&sh->hash);
454 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
456 struct hlist_head *hp = stripe_hash(conf, sh->sector);
458 pr_debug("insert_hash(), stripe %llu\n",
459 (unsigned long long)sh->sector);
461 hlist_add_head(&sh->hash, hp);
464 /* find an idle stripe, make sure it is unhashed, and return it. */
465 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
467 struct stripe_head *sh = NULL;
468 struct list_head *first;
470 if (list_empty(conf->inactive_list + hash))
472 first = (conf->inactive_list + hash)->next;
473 sh = list_entry(first, struct stripe_head, lru);
474 list_del_init(first);
476 atomic_inc(&conf->active_stripes);
477 BUG_ON(hash != sh->hash_lock_index);
478 if (list_empty(conf->inactive_list + hash))
479 atomic_inc(&conf->empty_inactive_list_nr);
484 static void shrink_buffers(struct stripe_head *sh)
488 int num = sh->raid_conf->pool_size;
490 for (i = 0; i < num ; i++) {
491 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
495 sh->dev[i].page = NULL;
500 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
503 int num = sh->raid_conf->pool_size;
505 for (i = 0; i < num; i++) {
508 if (!(page = alloc_page(gfp))) {
511 sh->dev[i].page = page;
512 sh->dev[i].orig_page = page;
517 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
518 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
519 struct stripe_head *sh);
521 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
523 struct r5conf *conf = sh->raid_conf;
526 BUG_ON(atomic_read(&sh->count) != 0);
527 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
528 BUG_ON(stripe_operations_active(sh));
529 BUG_ON(sh->batch_head);
531 pr_debug("init_stripe called, stripe %llu\n",
532 (unsigned long long)sector);
534 seq = read_seqcount_begin(&conf->gen_lock);
535 sh->generation = conf->generation - previous;
536 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
538 stripe_set_idx(sector, conf, previous, sh);
541 for (i = sh->disks; i--; ) {
542 struct r5dev *dev = &sh->dev[i];
544 if (dev->toread || dev->read || dev->towrite || dev->written ||
545 test_bit(R5_LOCKED, &dev->flags)) {
546 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
547 (unsigned long long)sh->sector, i, dev->toread,
548 dev->read, dev->towrite, dev->written,
549 test_bit(R5_LOCKED, &dev->flags));
553 raid5_build_block(sh, i, previous);
555 if (read_seqcount_retry(&conf->gen_lock, seq))
557 sh->overwrite_disks = 0;
558 insert_hash(conf, sh);
559 sh->cpu = smp_processor_id();
560 set_bit(STRIPE_BATCH_READY, &sh->state);
563 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
566 struct stripe_head *sh;
568 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
569 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
570 if (sh->sector == sector && sh->generation == generation)
572 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
577 * Need to check if array has failed when deciding whether to:
579 * - remove non-faulty devices
582 * This determination is simple when no reshape is happening.
583 * However if there is a reshape, we need to carefully check
584 * both the before and after sections.
585 * This is because some failed devices may only affect one
586 * of the two sections, and some non-in_sync devices may
587 * be insync in the section most affected by failed devices.
589 static int calc_degraded(struct r5conf *conf)
591 int degraded, degraded2;
596 for (i = 0; i < conf->previous_raid_disks; i++) {
597 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
598 if (rdev && test_bit(Faulty, &rdev->flags))
599 rdev = rcu_dereference(conf->disks[i].replacement);
600 if (!rdev || test_bit(Faulty, &rdev->flags))
602 else if (test_bit(In_sync, &rdev->flags))
605 /* not in-sync or faulty.
606 * If the reshape increases the number of devices,
607 * this is being recovered by the reshape, so
608 * this 'previous' section is not in_sync.
609 * If the number of devices is being reduced however,
610 * the device can only be part of the array if
611 * we are reverting a reshape, so this section will
614 if (conf->raid_disks >= conf->previous_raid_disks)
618 if (conf->raid_disks == conf->previous_raid_disks)
622 for (i = 0; i < conf->raid_disks; i++) {
623 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
624 if (rdev && test_bit(Faulty, &rdev->flags))
625 rdev = rcu_dereference(conf->disks[i].replacement);
626 if (!rdev || test_bit(Faulty, &rdev->flags))
628 else if (test_bit(In_sync, &rdev->flags))
631 /* not in-sync or faulty.
632 * If reshape increases the number of devices, this
633 * section has already been recovered, else it
634 * almost certainly hasn't.
636 if (conf->raid_disks <= conf->previous_raid_disks)
640 if (degraded2 > degraded)
645 static int has_failed(struct r5conf *conf)
649 if (conf->mddev->reshape_position == MaxSector)
650 return conf->mddev->degraded > conf->max_degraded;
652 degraded = calc_degraded(conf);
653 if (degraded > conf->max_degraded)
658 static struct stripe_head *
659 get_active_stripe(struct r5conf *conf, sector_t sector,
660 int previous, int noblock, int noquiesce)
662 struct stripe_head *sh;
663 int hash = stripe_hash_locks_hash(sector);
665 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
667 spin_lock_irq(conf->hash_locks + hash);
670 wait_event_lock_irq(conf->wait_for_stripe,
671 conf->quiesce == 0 || noquiesce,
672 *(conf->hash_locks + hash));
673 sh = __find_stripe(conf, sector, conf->generation - previous);
675 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
676 sh = get_free_stripe(conf, hash);
677 if (!sh && llist_empty(&conf->released_stripes) &&
678 !test_bit(R5_DID_ALLOC, &conf->cache_state))
679 set_bit(R5_ALLOC_MORE,
682 if (noblock && sh == NULL)
685 set_bit(R5_INACTIVE_BLOCKED,
688 conf->wait_for_stripe,
689 !list_empty(conf->inactive_list + hash) &&
690 (atomic_read(&conf->active_stripes)
691 < (conf->max_nr_stripes * 3 / 4)
692 || !test_bit(R5_INACTIVE_BLOCKED,
693 &conf->cache_state)),
694 *(conf->hash_locks + hash));
695 clear_bit(R5_INACTIVE_BLOCKED,
698 init_stripe(sh, sector, previous);
699 atomic_inc(&sh->count);
701 } else if (!atomic_inc_not_zero(&sh->count)) {
702 spin_lock(&conf->device_lock);
703 if (!atomic_read(&sh->count)) {
704 if (!test_bit(STRIPE_HANDLE, &sh->state))
705 atomic_inc(&conf->active_stripes);
706 BUG_ON(list_empty(&sh->lru) &&
707 !test_bit(STRIPE_EXPANDING, &sh->state));
708 list_del_init(&sh->lru);
710 sh->group->stripes_cnt--;
714 atomic_inc(&sh->count);
715 spin_unlock(&conf->device_lock);
717 } while (sh == NULL);
719 spin_unlock_irq(conf->hash_locks + hash);
723 static bool is_full_stripe_write(struct stripe_head *sh)
725 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
726 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
729 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
733 spin_lock(&sh2->stripe_lock);
734 spin_lock_nested(&sh1->stripe_lock, 1);
736 spin_lock(&sh1->stripe_lock);
737 spin_lock_nested(&sh2->stripe_lock, 1);
741 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
743 spin_unlock(&sh1->stripe_lock);
744 spin_unlock(&sh2->stripe_lock);
748 /* Only freshly new full stripe normal write stripe can be added to a batch list */
749 static bool stripe_can_batch(struct stripe_head *sh)
751 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
752 is_full_stripe_write(sh);
755 /* we only do back search */
756 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
758 struct stripe_head *head;
759 sector_t head_sector, tmp_sec;
763 if (!stripe_can_batch(sh))
765 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
766 tmp_sec = sh->sector;
767 if (!sector_div(tmp_sec, conf->chunk_sectors))
769 head_sector = sh->sector - STRIPE_SECTORS;
771 hash = stripe_hash_locks_hash(head_sector);
772 spin_lock_irq(conf->hash_locks + hash);
773 head = __find_stripe(conf, head_sector, conf->generation);
774 if (head && !atomic_inc_not_zero(&head->count)) {
775 spin_lock(&conf->device_lock);
776 if (!atomic_read(&head->count)) {
777 if (!test_bit(STRIPE_HANDLE, &head->state))
778 atomic_inc(&conf->active_stripes);
779 BUG_ON(list_empty(&head->lru) &&
780 !test_bit(STRIPE_EXPANDING, &head->state));
781 list_del_init(&head->lru);
783 head->group->stripes_cnt--;
787 atomic_inc(&head->count);
788 spin_unlock(&conf->device_lock);
790 spin_unlock_irq(conf->hash_locks + hash);
794 if (!stripe_can_batch(head))
797 lock_two_stripes(head, sh);
798 /* clear_batch_ready clear the flag */
799 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
806 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
808 if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
811 if (head->batch_head) {
812 spin_lock(&head->batch_head->batch_lock);
813 /* This batch list is already running */
814 if (!stripe_can_batch(head)) {
815 spin_unlock(&head->batch_head->batch_lock);
820 * at this point, head's BATCH_READY could be cleared, but we
821 * can still add the stripe to batch list
823 list_add(&sh->batch_list, &head->batch_list);
824 spin_unlock(&head->batch_head->batch_lock);
826 sh->batch_head = head->batch_head;
828 head->batch_head = head;
829 sh->batch_head = head->batch_head;
830 spin_lock(&head->batch_lock);
831 list_add_tail(&sh->batch_list, &head->batch_list);
832 spin_unlock(&head->batch_lock);
835 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
836 if (atomic_dec_return(&conf->preread_active_stripes)
838 md_wakeup_thread(conf->mddev->thread);
840 atomic_inc(&sh->count);
842 unlock_two_stripes(head, sh);
844 release_stripe(head);
847 /* Determine if 'data_offset' or 'new_data_offset' should be used
848 * in this stripe_head.
850 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
852 sector_t progress = conf->reshape_progress;
853 /* Need a memory barrier to make sure we see the value
854 * of conf->generation, or ->data_offset that was set before
855 * reshape_progress was updated.
858 if (progress == MaxSector)
860 if (sh->generation == conf->generation - 1)
862 /* We are in a reshape, and this is a new-generation stripe,
863 * so use new_data_offset.
869 raid5_end_read_request(struct bio *bi, int error);
871 raid5_end_write_request(struct bio *bi, int error);
873 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
875 struct r5conf *conf = sh->raid_conf;
876 int i, disks = sh->disks;
877 struct stripe_head *head_sh = sh;
881 for (i = disks; i--; ) {
883 int replace_only = 0;
884 struct bio *bi, *rbi;
885 struct md_rdev *rdev, *rrdev = NULL;
888 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
889 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
893 if (test_bit(R5_Discard, &sh->dev[i].flags))
895 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
897 else if (test_and_clear_bit(R5_WantReplace,
898 &sh->dev[i].flags)) {
903 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
907 bi = &sh->dev[i].req;
908 rbi = &sh->dev[i].rreq; /* For writing to replacement */
911 rrdev = rcu_dereference(conf->disks[i].replacement);
912 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
913 rdev = rcu_dereference(conf->disks[i].rdev);
922 /* We raced and saw duplicates */
925 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
930 if (rdev && test_bit(Faulty, &rdev->flags))
933 atomic_inc(&rdev->nr_pending);
934 if (rrdev && test_bit(Faulty, &rrdev->flags))
937 atomic_inc(&rrdev->nr_pending);
940 /* We have already checked bad blocks for reads. Now
941 * need to check for writes. We never accept write errors
942 * on the replacement, so we don't to check rrdev.
944 while ((rw & WRITE) && rdev &&
945 test_bit(WriteErrorSeen, &rdev->flags)) {
948 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
949 &first_bad, &bad_sectors);
954 set_bit(BlockedBadBlocks, &rdev->flags);
955 if (!conf->mddev->external &&
956 conf->mddev->flags) {
957 /* It is very unlikely, but we might
958 * still need to write out the
959 * bad block log - better give it
961 md_check_recovery(conf->mddev);
964 * Because md_wait_for_blocked_rdev
965 * will dec nr_pending, we must
966 * increment it first.
968 atomic_inc(&rdev->nr_pending);
969 md_wait_for_blocked_rdev(rdev, conf->mddev);
971 /* Acknowledged bad block - skip the write */
972 rdev_dec_pending(rdev, conf->mddev);
978 if (s->syncing || s->expanding || s->expanded
980 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
982 set_bit(STRIPE_IO_STARTED, &sh->state);
985 bi->bi_bdev = rdev->bdev;
987 bi->bi_end_io = (rw & WRITE)
988 ? raid5_end_write_request
989 : raid5_end_read_request;
992 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
993 __func__, (unsigned long long)sh->sector,
995 atomic_inc(&sh->count);
997 atomic_inc(&head_sh->count);
998 if (use_new_offset(conf, sh))
999 bi->bi_iter.bi_sector = (sh->sector
1000 + rdev->new_data_offset);
1002 bi->bi_iter.bi_sector = (sh->sector
1003 + rdev->data_offset);
1004 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1005 bi->bi_rw |= REQ_NOMERGE;
1007 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1008 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1009 sh->dev[i].vec.bv_page = sh->dev[i].page;
1011 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1012 bi->bi_io_vec[0].bv_offset = 0;
1013 bi->bi_iter.bi_size = STRIPE_SIZE;
1015 * If this is discard request, set bi_vcnt 0. We don't
1016 * want to confuse SCSI because SCSI will replace payload
1018 if (rw & REQ_DISCARD)
1021 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1023 if (conf->mddev->gendisk)
1024 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1025 bi, disk_devt(conf->mddev->gendisk),
1027 generic_make_request(bi);
1030 if (s->syncing || s->expanding || s->expanded
1032 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1034 set_bit(STRIPE_IO_STARTED, &sh->state);
1037 rbi->bi_bdev = rrdev->bdev;
1039 BUG_ON(!(rw & WRITE));
1040 rbi->bi_end_io = raid5_end_write_request;
1041 rbi->bi_private = sh;
1043 pr_debug("%s: for %llu schedule op %ld on "
1044 "replacement disc %d\n",
1045 __func__, (unsigned long long)sh->sector,
1047 atomic_inc(&sh->count);
1049 atomic_inc(&head_sh->count);
1050 if (use_new_offset(conf, sh))
1051 rbi->bi_iter.bi_sector = (sh->sector
1052 + rrdev->new_data_offset);
1054 rbi->bi_iter.bi_sector = (sh->sector
1055 + rrdev->data_offset);
1056 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1057 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1058 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1060 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1061 rbi->bi_io_vec[0].bv_offset = 0;
1062 rbi->bi_iter.bi_size = STRIPE_SIZE;
1064 * If this is discard request, set bi_vcnt 0. We don't
1065 * want to confuse SCSI because SCSI will replace payload
1067 if (rw & REQ_DISCARD)
1069 if (conf->mddev->gendisk)
1070 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1071 rbi, disk_devt(conf->mddev->gendisk),
1073 generic_make_request(rbi);
1075 if (!rdev && !rrdev) {
1077 set_bit(STRIPE_DEGRADED, &sh->state);
1078 pr_debug("skip op %ld on disc %d for sector %llu\n",
1079 bi->bi_rw, i, (unsigned long long)sh->sector);
1080 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1082 set_bit(STRIPE_BATCH_ERR,
1083 &sh->batch_head->state);
1084 set_bit(STRIPE_HANDLE, &sh->state);
1087 if (!head_sh->batch_head)
1089 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1096 static struct dma_async_tx_descriptor *
1097 async_copy_data(int frombio, struct bio *bio, struct page **page,
1098 sector_t sector, struct dma_async_tx_descriptor *tx,
1099 struct stripe_head *sh)
1102 struct bvec_iter iter;
1103 struct page *bio_page;
1105 struct async_submit_ctl submit;
1106 enum async_tx_flags flags = 0;
1108 if (bio->bi_iter.bi_sector >= sector)
1109 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1111 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1114 flags |= ASYNC_TX_FENCE;
1115 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1117 bio_for_each_segment(bvl, bio, iter) {
1118 int len = bvl.bv_len;
1122 if (page_offset < 0) {
1123 b_offset = -page_offset;
1124 page_offset += b_offset;
1128 if (len > 0 && page_offset + len > STRIPE_SIZE)
1129 clen = STRIPE_SIZE - page_offset;
1134 b_offset += bvl.bv_offset;
1135 bio_page = bvl.bv_page;
1137 if (sh->raid_conf->skip_copy &&
1138 b_offset == 0 && page_offset == 0 &&
1139 clen == STRIPE_SIZE)
1142 tx = async_memcpy(*page, bio_page, page_offset,
1143 b_offset, clen, &submit);
1145 tx = async_memcpy(bio_page, *page, b_offset,
1146 page_offset, clen, &submit);
1148 /* chain the operations */
1149 submit.depend_tx = tx;
1151 if (clen < len) /* hit end of page */
1159 static void ops_complete_biofill(void *stripe_head_ref)
1161 struct stripe_head *sh = stripe_head_ref;
1162 struct bio *return_bi = NULL;
1165 pr_debug("%s: stripe %llu\n", __func__,
1166 (unsigned long long)sh->sector);
1168 /* clear completed biofills */
1169 for (i = sh->disks; i--; ) {
1170 struct r5dev *dev = &sh->dev[i];
1172 /* acknowledge completion of a biofill operation */
1173 /* and check if we need to reply to a read request,
1174 * new R5_Wantfill requests are held off until
1175 * !STRIPE_BIOFILL_RUN
1177 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1178 struct bio *rbi, *rbi2;
1183 while (rbi && rbi->bi_iter.bi_sector <
1184 dev->sector + STRIPE_SECTORS) {
1185 rbi2 = r5_next_bio(rbi, dev->sector);
1186 if (!raid5_dec_bi_active_stripes(rbi)) {
1187 rbi->bi_next = return_bi;
1194 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1196 return_io(return_bi);
1198 set_bit(STRIPE_HANDLE, &sh->state);
1202 static void ops_run_biofill(struct stripe_head *sh)
1204 struct dma_async_tx_descriptor *tx = NULL;
1205 struct async_submit_ctl submit;
1208 BUG_ON(sh->batch_head);
1209 pr_debug("%s: stripe %llu\n", __func__,
1210 (unsigned long long)sh->sector);
1212 for (i = sh->disks; i--; ) {
1213 struct r5dev *dev = &sh->dev[i];
1214 if (test_bit(R5_Wantfill, &dev->flags)) {
1216 spin_lock_irq(&sh->stripe_lock);
1217 dev->read = rbi = dev->toread;
1219 spin_unlock_irq(&sh->stripe_lock);
1220 while (rbi && rbi->bi_iter.bi_sector <
1221 dev->sector + STRIPE_SECTORS) {
1222 tx = async_copy_data(0, rbi, &dev->page,
1223 dev->sector, tx, sh);
1224 rbi = r5_next_bio(rbi, dev->sector);
1229 atomic_inc(&sh->count);
1230 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1231 async_trigger_callback(&submit);
1234 static void mark_target_uptodate(struct stripe_head *sh, int target)
1241 tgt = &sh->dev[target];
1242 set_bit(R5_UPTODATE, &tgt->flags);
1243 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1244 clear_bit(R5_Wantcompute, &tgt->flags);
1247 static void ops_complete_compute(void *stripe_head_ref)
1249 struct stripe_head *sh = stripe_head_ref;
1251 pr_debug("%s: stripe %llu\n", __func__,
1252 (unsigned long long)sh->sector);
1254 /* mark the computed target(s) as uptodate */
1255 mark_target_uptodate(sh, sh->ops.target);
1256 mark_target_uptodate(sh, sh->ops.target2);
1258 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1259 if (sh->check_state == check_state_compute_run)
1260 sh->check_state = check_state_compute_result;
1261 set_bit(STRIPE_HANDLE, &sh->state);
1265 /* return a pointer to the address conversion region of the scribble buffer */
1266 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1267 struct raid5_percpu *percpu, int i)
1271 addr = flex_array_get(percpu->scribble, i);
1272 return addr + sizeof(struct page *) * (sh->disks + 2);
1275 /* return a pointer to the address conversion region of the scribble buffer */
1276 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1280 addr = flex_array_get(percpu->scribble, i);
1284 static struct dma_async_tx_descriptor *
1285 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1287 int disks = sh->disks;
1288 struct page **xor_srcs = to_addr_page(percpu, 0);
1289 int target = sh->ops.target;
1290 struct r5dev *tgt = &sh->dev[target];
1291 struct page *xor_dest = tgt->page;
1293 struct dma_async_tx_descriptor *tx;
1294 struct async_submit_ctl submit;
1297 BUG_ON(sh->batch_head);
1299 pr_debug("%s: stripe %llu block: %d\n",
1300 __func__, (unsigned long long)sh->sector, target);
1301 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1303 for (i = disks; i--; )
1305 xor_srcs[count++] = sh->dev[i].page;
1307 atomic_inc(&sh->count);
1309 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1310 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1311 if (unlikely(count == 1))
1312 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1314 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1319 /* set_syndrome_sources - populate source buffers for gen_syndrome
1320 * @srcs - (struct page *) array of size sh->disks
1321 * @sh - stripe_head to parse
1323 * Populates srcs in proper layout order for the stripe and returns the
1324 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1325 * destination buffer is recorded in srcs[count] and the Q destination
1326 * is recorded in srcs[count+1]].
1328 static int set_syndrome_sources(struct page **srcs,
1329 struct stripe_head *sh,
1332 int disks = sh->disks;
1333 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1334 int d0_idx = raid6_d0(sh);
1338 for (i = 0; i < disks; i++)
1344 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1345 struct r5dev *dev = &sh->dev[i];
1347 if (i == sh->qd_idx || i == sh->pd_idx ||
1348 (srctype == SYNDROME_SRC_ALL) ||
1349 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1350 test_bit(R5_Wantdrain, &dev->flags)) ||
1351 (srctype == SYNDROME_SRC_WRITTEN &&
1353 srcs[slot] = sh->dev[i].page;
1354 i = raid6_next_disk(i, disks);
1355 } while (i != d0_idx);
1357 return syndrome_disks;
1360 static struct dma_async_tx_descriptor *
1361 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1363 int disks = sh->disks;
1364 struct page **blocks = to_addr_page(percpu, 0);
1366 int qd_idx = sh->qd_idx;
1367 struct dma_async_tx_descriptor *tx;
1368 struct async_submit_ctl submit;
1374 BUG_ON(sh->batch_head);
1375 if (sh->ops.target < 0)
1376 target = sh->ops.target2;
1377 else if (sh->ops.target2 < 0)
1378 target = sh->ops.target;
1380 /* we should only have one valid target */
1383 pr_debug("%s: stripe %llu block: %d\n",
1384 __func__, (unsigned long long)sh->sector, target);
1386 tgt = &sh->dev[target];
1387 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1390 atomic_inc(&sh->count);
1392 if (target == qd_idx) {
1393 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1394 blocks[count] = NULL; /* regenerating p is not necessary */
1395 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1396 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1397 ops_complete_compute, sh,
1398 to_addr_conv(sh, percpu, 0));
1399 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1401 /* Compute any data- or p-drive using XOR */
1403 for (i = disks; i-- ; ) {
1404 if (i == target || i == qd_idx)
1406 blocks[count++] = sh->dev[i].page;
1409 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1410 NULL, ops_complete_compute, sh,
1411 to_addr_conv(sh, percpu, 0));
1412 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1418 static struct dma_async_tx_descriptor *
1419 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1421 int i, count, disks = sh->disks;
1422 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1423 int d0_idx = raid6_d0(sh);
1424 int faila = -1, failb = -1;
1425 int target = sh->ops.target;
1426 int target2 = sh->ops.target2;
1427 struct r5dev *tgt = &sh->dev[target];
1428 struct r5dev *tgt2 = &sh->dev[target2];
1429 struct dma_async_tx_descriptor *tx;
1430 struct page **blocks = to_addr_page(percpu, 0);
1431 struct async_submit_ctl submit;
1433 BUG_ON(sh->batch_head);
1434 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1435 __func__, (unsigned long long)sh->sector, target, target2);
1436 BUG_ON(target < 0 || target2 < 0);
1437 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1438 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1440 /* we need to open-code set_syndrome_sources to handle the
1441 * slot number conversion for 'faila' and 'failb'
1443 for (i = 0; i < disks ; i++)
1448 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1450 blocks[slot] = sh->dev[i].page;
1456 i = raid6_next_disk(i, disks);
1457 } while (i != d0_idx);
1459 BUG_ON(faila == failb);
1462 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1463 __func__, (unsigned long long)sh->sector, faila, failb);
1465 atomic_inc(&sh->count);
1467 if (failb == syndrome_disks+1) {
1468 /* Q disk is one of the missing disks */
1469 if (faila == syndrome_disks) {
1470 /* Missing P+Q, just recompute */
1471 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1472 ops_complete_compute, sh,
1473 to_addr_conv(sh, percpu, 0));
1474 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1475 STRIPE_SIZE, &submit);
1479 int qd_idx = sh->qd_idx;
1481 /* Missing D+Q: recompute D from P, then recompute Q */
1482 if (target == qd_idx)
1483 data_target = target2;
1485 data_target = target;
1488 for (i = disks; i-- ; ) {
1489 if (i == data_target || i == qd_idx)
1491 blocks[count++] = sh->dev[i].page;
1493 dest = sh->dev[data_target].page;
1494 init_async_submit(&submit,
1495 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1497 to_addr_conv(sh, percpu, 0));
1498 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1501 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1502 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1503 ops_complete_compute, sh,
1504 to_addr_conv(sh, percpu, 0));
1505 return async_gen_syndrome(blocks, 0, count+2,
1506 STRIPE_SIZE, &submit);
1509 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1510 ops_complete_compute, sh,
1511 to_addr_conv(sh, percpu, 0));
1512 if (failb == syndrome_disks) {
1513 /* We're missing D+P. */
1514 return async_raid6_datap_recov(syndrome_disks+2,
1518 /* We're missing D+D. */
1519 return async_raid6_2data_recov(syndrome_disks+2,
1520 STRIPE_SIZE, faila, failb,
1526 static void ops_complete_prexor(void *stripe_head_ref)
1528 struct stripe_head *sh = stripe_head_ref;
1530 pr_debug("%s: stripe %llu\n", __func__,
1531 (unsigned long long)sh->sector);
1534 static struct dma_async_tx_descriptor *
1535 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1536 struct dma_async_tx_descriptor *tx)
1538 int disks = sh->disks;
1539 struct page **xor_srcs = to_addr_page(percpu, 0);
1540 int count = 0, pd_idx = sh->pd_idx, i;
1541 struct async_submit_ctl submit;
1543 /* existing parity data subtracted */
1544 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1546 BUG_ON(sh->batch_head);
1547 pr_debug("%s: stripe %llu\n", __func__,
1548 (unsigned long long)sh->sector);
1550 for (i = disks; i--; ) {
1551 struct r5dev *dev = &sh->dev[i];
1552 /* Only process blocks that are known to be uptodate */
1553 if (test_bit(R5_Wantdrain, &dev->flags))
1554 xor_srcs[count++] = dev->page;
1557 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1558 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1559 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1564 static struct dma_async_tx_descriptor *
1565 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1566 struct dma_async_tx_descriptor *tx)
1568 struct page **blocks = to_addr_page(percpu, 0);
1570 struct async_submit_ctl submit;
1572 pr_debug("%s: stripe %llu\n", __func__,
1573 (unsigned long long)sh->sector);
1575 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1577 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1578 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1579 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1584 static struct dma_async_tx_descriptor *
1585 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1587 int disks = sh->disks;
1589 struct stripe_head *head_sh = sh;
1591 pr_debug("%s: stripe %llu\n", __func__,
1592 (unsigned long long)sh->sector);
1594 for (i = disks; i--; ) {
1599 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1604 spin_lock_irq(&sh->stripe_lock);
1605 chosen = dev->towrite;
1606 dev->towrite = NULL;
1607 sh->overwrite_disks = 0;
1608 BUG_ON(dev->written);
1609 wbi = dev->written = chosen;
1610 spin_unlock_irq(&sh->stripe_lock);
1611 WARN_ON(dev->page != dev->orig_page);
1613 while (wbi && wbi->bi_iter.bi_sector <
1614 dev->sector + STRIPE_SECTORS) {
1615 if (wbi->bi_rw & REQ_FUA)
1616 set_bit(R5_WantFUA, &dev->flags);
1617 if (wbi->bi_rw & REQ_SYNC)
1618 set_bit(R5_SyncIO, &dev->flags);
1619 if (wbi->bi_rw & REQ_DISCARD)
1620 set_bit(R5_Discard, &dev->flags);
1622 tx = async_copy_data(1, wbi, &dev->page,
1623 dev->sector, tx, sh);
1624 if (dev->page != dev->orig_page) {
1625 set_bit(R5_SkipCopy, &dev->flags);
1626 clear_bit(R5_UPTODATE, &dev->flags);
1627 clear_bit(R5_OVERWRITE, &dev->flags);
1630 wbi = r5_next_bio(wbi, dev->sector);
1633 if (head_sh->batch_head) {
1634 sh = list_first_entry(&sh->batch_list,
1647 static void ops_complete_reconstruct(void *stripe_head_ref)
1649 struct stripe_head *sh = stripe_head_ref;
1650 int disks = sh->disks;
1651 int pd_idx = sh->pd_idx;
1652 int qd_idx = sh->qd_idx;
1654 bool fua = false, sync = false, discard = false;
1656 pr_debug("%s: stripe %llu\n", __func__,
1657 (unsigned long long)sh->sector);
1659 for (i = disks; i--; ) {
1660 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1661 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1662 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1665 for (i = disks; i--; ) {
1666 struct r5dev *dev = &sh->dev[i];
1668 if (dev->written || i == pd_idx || i == qd_idx) {
1669 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1670 set_bit(R5_UPTODATE, &dev->flags);
1672 set_bit(R5_WantFUA, &dev->flags);
1674 set_bit(R5_SyncIO, &dev->flags);
1678 if (sh->reconstruct_state == reconstruct_state_drain_run)
1679 sh->reconstruct_state = reconstruct_state_drain_result;
1680 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1681 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1683 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1684 sh->reconstruct_state = reconstruct_state_result;
1687 set_bit(STRIPE_HANDLE, &sh->state);
1692 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1693 struct dma_async_tx_descriptor *tx)
1695 int disks = sh->disks;
1696 struct page **xor_srcs;
1697 struct async_submit_ctl submit;
1698 int count, pd_idx = sh->pd_idx, i;
1699 struct page *xor_dest;
1701 unsigned long flags;
1703 struct stripe_head *head_sh = sh;
1706 pr_debug("%s: stripe %llu\n", __func__,
1707 (unsigned long long)sh->sector);
1709 for (i = 0; i < sh->disks; i++) {
1712 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1715 if (i >= sh->disks) {
1716 atomic_inc(&sh->count);
1717 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1718 ops_complete_reconstruct(sh);
1723 xor_srcs = to_addr_page(percpu, j);
1724 /* check if prexor is active which means only process blocks
1725 * that are part of a read-modify-write (written)
1727 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1729 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1730 for (i = disks; i--; ) {
1731 struct r5dev *dev = &sh->dev[i];
1732 if (head_sh->dev[i].written)
1733 xor_srcs[count++] = dev->page;
1736 xor_dest = sh->dev[pd_idx].page;
1737 for (i = disks; i--; ) {
1738 struct r5dev *dev = &sh->dev[i];
1740 xor_srcs[count++] = dev->page;
1744 /* 1/ if we prexor'd then the dest is reused as a source
1745 * 2/ if we did not prexor then we are redoing the parity
1746 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1747 * for the synchronous xor case
1749 last_stripe = !head_sh->batch_head ||
1750 list_first_entry(&sh->batch_list,
1751 struct stripe_head, batch_list) == head_sh;
1753 flags = ASYNC_TX_ACK |
1754 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1756 atomic_inc(&head_sh->count);
1757 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1758 to_addr_conv(sh, percpu, j));
1760 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1761 init_async_submit(&submit, flags, tx, NULL, NULL,
1762 to_addr_conv(sh, percpu, j));
1765 if (unlikely(count == 1))
1766 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1768 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1771 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1778 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1779 struct dma_async_tx_descriptor *tx)
1781 struct async_submit_ctl submit;
1782 struct page **blocks;
1783 int count, i, j = 0;
1784 struct stripe_head *head_sh = sh;
1787 unsigned long txflags;
1789 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1791 for (i = 0; i < sh->disks; i++) {
1792 if (sh->pd_idx == i || sh->qd_idx == i)
1794 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1797 if (i >= sh->disks) {
1798 atomic_inc(&sh->count);
1799 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1800 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1801 ops_complete_reconstruct(sh);
1806 blocks = to_addr_page(percpu, j);
1808 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1809 synflags = SYNDROME_SRC_WRITTEN;
1810 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1812 synflags = SYNDROME_SRC_ALL;
1813 txflags = ASYNC_TX_ACK;
1816 count = set_syndrome_sources(blocks, sh, synflags);
1817 last_stripe = !head_sh->batch_head ||
1818 list_first_entry(&sh->batch_list,
1819 struct stripe_head, batch_list) == head_sh;
1822 atomic_inc(&head_sh->count);
1823 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1824 head_sh, to_addr_conv(sh, percpu, j));
1826 init_async_submit(&submit, 0, tx, NULL, NULL,
1827 to_addr_conv(sh, percpu, j));
1828 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1831 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1837 static void ops_complete_check(void *stripe_head_ref)
1839 struct stripe_head *sh = stripe_head_ref;
1841 pr_debug("%s: stripe %llu\n", __func__,
1842 (unsigned long long)sh->sector);
1844 sh->check_state = check_state_check_result;
1845 set_bit(STRIPE_HANDLE, &sh->state);
1849 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1851 int disks = sh->disks;
1852 int pd_idx = sh->pd_idx;
1853 int qd_idx = sh->qd_idx;
1854 struct page *xor_dest;
1855 struct page **xor_srcs = to_addr_page(percpu, 0);
1856 struct dma_async_tx_descriptor *tx;
1857 struct async_submit_ctl submit;
1861 pr_debug("%s: stripe %llu\n", __func__,
1862 (unsigned long long)sh->sector);
1864 BUG_ON(sh->batch_head);
1866 xor_dest = sh->dev[pd_idx].page;
1867 xor_srcs[count++] = xor_dest;
1868 for (i = disks; i--; ) {
1869 if (i == pd_idx || i == qd_idx)
1871 xor_srcs[count++] = sh->dev[i].page;
1874 init_async_submit(&submit, 0, NULL, NULL, NULL,
1875 to_addr_conv(sh, percpu, 0));
1876 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1877 &sh->ops.zero_sum_result, &submit);
1879 atomic_inc(&sh->count);
1880 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1881 tx = async_trigger_callback(&submit);
1884 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1886 struct page **srcs = to_addr_page(percpu, 0);
1887 struct async_submit_ctl submit;
1890 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1891 (unsigned long long)sh->sector, checkp);
1893 BUG_ON(sh->batch_head);
1894 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1898 atomic_inc(&sh->count);
1899 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1900 sh, to_addr_conv(sh, percpu, 0));
1901 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1902 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1905 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1907 int overlap_clear = 0, i, disks = sh->disks;
1908 struct dma_async_tx_descriptor *tx = NULL;
1909 struct r5conf *conf = sh->raid_conf;
1910 int level = conf->level;
1911 struct raid5_percpu *percpu;
1915 percpu = per_cpu_ptr(conf->percpu, cpu);
1916 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1917 ops_run_biofill(sh);
1921 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1923 tx = ops_run_compute5(sh, percpu);
1925 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1926 tx = ops_run_compute6_1(sh, percpu);
1928 tx = ops_run_compute6_2(sh, percpu);
1930 /* terminate the chain if reconstruct is not set to be run */
1931 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1935 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1937 tx = ops_run_prexor5(sh, percpu, tx);
1939 tx = ops_run_prexor6(sh, percpu, tx);
1942 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1943 tx = ops_run_biodrain(sh, tx);
1947 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1949 ops_run_reconstruct5(sh, percpu, tx);
1951 ops_run_reconstruct6(sh, percpu, tx);
1954 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1955 if (sh->check_state == check_state_run)
1956 ops_run_check_p(sh, percpu);
1957 else if (sh->check_state == check_state_run_q)
1958 ops_run_check_pq(sh, percpu, 0);
1959 else if (sh->check_state == check_state_run_pq)
1960 ops_run_check_pq(sh, percpu, 1);
1965 if (overlap_clear && !sh->batch_head)
1966 for (i = disks; i--; ) {
1967 struct r5dev *dev = &sh->dev[i];
1968 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1969 wake_up(&sh->raid_conf->wait_for_overlap);
1974 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1976 struct stripe_head *sh;
1978 sh = kmem_cache_zalloc(sc, gfp);
1980 spin_lock_init(&sh->stripe_lock);
1981 spin_lock_init(&sh->batch_lock);
1982 INIT_LIST_HEAD(&sh->batch_list);
1983 INIT_LIST_HEAD(&sh->lru);
1984 atomic_set(&sh->count, 1);
1988 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
1990 struct stripe_head *sh;
1992 sh = alloc_stripe(conf->slab_cache, gfp);
1996 sh->raid_conf = conf;
1998 if (grow_buffers(sh, gfp)) {
2000 kmem_cache_free(conf->slab_cache, sh);
2003 sh->hash_lock_index =
2004 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2005 /* we just created an active stripe so... */
2006 atomic_inc(&conf->active_stripes);
2009 conf->max_nr_stripes++;
2013 static int grow_stripes(struct r5conf *conf, int num)
2015 struct kmem_cache *sc;
2016 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2018 if (conf->mddev->gendisk)
2019 sprintf(conf->cache_name[0],
2020 "raid%d-%s", conf->level, mdname(conf->mddev));
2022 sprintf(conf->cache_name[0],
2023 "raid%d-%p", conf->level, conf->mddev);
2024 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2026 conf->active_name = 0;
2027 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2028 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2032 conf->slab_cache = sc;
2033 conf->pool_size = devs;
2035 if (!grow_one_stripe(conf, GFP_KERNEL))
2042 * scribble_len - return the required size of the scribble region
2043 * @num - total number of disks in the array
2045 * The size must be enough to contain:
2046 * 1/ a struct page pointer for each device in the array +2
2047 * 2/ room to convert each entry in (1) to its corresponding dma
2048 * (dma_map_page()) or page (page_address()) address.
2050 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2051 * calculate over all devices (not just the data blocks), using zeros in place
2052 * of the P and Q blocks.
2054 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2056 struct flex_array *ret;
2059 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2060 ret = flex_array_alloc(len, cnt, flags);
2063 /* always prealloc all elements, so no locking is required */
2064 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2065 flex_array_free(ret);
2071 static int resize_stripes(struct r5conf *conf, int newsize)
2073 /* Make all the stripes able to hold 'newsize' devices.
2074 * New slots in each stripe get 'page' set to a new page.
2076 * This happens in stages:
2077 * 1/ create a new kmem_cache and allocate the required number of
2079 * 2/ gather all the old stripe_heads and transfer the pages across
2080 * to the new stripe_heads. This will have the side effect of
2081 * freezing the array as once all stripe_heads have been collected,
2082 * no IO will be possible. Old stripe heads are freed once their
2083 * pages have been transferred over, and the old kmem_cache is
2084 * freed when all stripes are done.
2085 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2086 * we simple return a failre status - no need to clean anything up.
2087 * 4/ allocate new pages for the new slots in the new stripe_heads.
2088 * If this fails, we don't bother trying the shrink the
2089 * stripe_heads down again, we just leave them as they are.
2090 * As each stripe_head is processed the new one is released into
2093 * Once step2 is started, we cannot afford to wait for a write,
2094 * so we use GFP_NOIO allocations.
2096 struct stripe_head *osh, *nsh;
2097 LIST_HEAD(newstripes);
2098 struct disk_info *ndisks;
2101 struct kmem_cache *sc;
2105 if (newsize <= conf->pool_size)
2106 return 0; /* never bother to shrink */
2108 err = md_allow_write(conf->mddev);
2113 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2114 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2119 for (i = conf->max_nr_stripes; i; i--) {
2120 nsh = alloc_stripe(sc, GFP_KERNEL);
2124 nsh->raid_conf = conf;
2125 list_add(&nsh->lru, &newstripes);
2128 /* didn't get enough, give up */
2129 while (!list_empty(&newstripes)) {
2130 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2131 list_del(&nsh->lru);
2132 kmem_cache_free(sc, nsh);
2134 kmem_cache_destroy(sc);
2137 /* Step 2 - Must use GFP_NOIO now.
2138 * OK, we have enough stripes, start collecting inactive
2139 * stripes and copying them over
2143 list_for_each_entry(nsh, &newstripes, lru) {
2144 lock_device_hash_lock(conf, hash);
2145 wait_event_cmd(conf->wait_for_stripe,
2146 !list_empty(conf->inactive_list + hash),
2147 unlock_device_hash_lock(conf, hash),
2148 lock_device_hash_lock(conf, hash));
2149 osh = get_free_stripe(conf, hash);
2150 unlock_device_hash_lock(conf, hash);
2152 for(i=0; i<conf->pool_size; i++) {
2153 nsh->dev[i].page = osh->dev[i].page;
2154 nsh->dev[i].orig_page = osh->dev[i].page;
2156 nsh->hash_lock_index = hash;
2157 kmem_cache_free(conf->slab_cache, osh);
2159 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2160 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2165 kmem_cache_destroy(conf->slab_cache);
2168 * At this point, we are holding all the stripes so the array
2169 * is completely stalled, so now is a good time to resize
2170 * conf->disks and the scribble region
2172 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2174 for (i=0; i<conf->raid_disks; i++)
2175 ndisks[i] = conf->disks[i];
2177 conf->disks = ndisks;
2182 for_each_present_cpu(cpu) {
2183 struct raid5_percpu *percpu;
2184 struct flex_array *scribble;
2186 percpu = per_cpu_ptr(conf->percpu, cpu);
2187 scribble = scribble_alloc(newsize, conf->chunk_sectors /
2188 STRIPE_SECTORS, GFP_NOIO);
2191 flex_array_free(percpu->scribble);
2192 percpu->scribble = scribble;
2200 /* Step 4, return new stripes to service */
2201 while(!list_empty(&newstripes)) {
2202 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2203 list_del_init(&nsh->lru);
2205 for (i=conf->raid_disks; i < newsize; i++)
2206 if (nsh->dev[i].page == NULL) {
2207 struct page *p = alloc_page(GFP_NOIO);
2208 nsh->dev[i].page = p;
2209 nsh->dev[i].orig_page = p;
2213 release_stripe(nsh);
2215 /* critical section pass, GFP_NOIO no longer needed */
2217 conf->slab_cache = sc;
2218 conf->active_name = 1-conf->active_name;
2219 conf->pool_size = newsize;
2223 static int drop_one_stripe(struct r5conf *conf)
2225 struct stripe_head *sh;
2226 int hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
2228 spin_lock_irq(conf->hash_locks + hash);
2229 sh = get_free_stripe(conf, hash);
2230 spin_unlock_irq(conf->hash_locks + hash);
2233 BUG_ON(atomic_read(&sh->count));
2235 kmem_cache_free(conf->slab_cache, sh);
2236 atomic_dec(&conf->active_stripes);
2237 conf->max_nr_stripes--;
2241 static void shrink_stripes(struct r5conf *conf)
2243 while (conf->max_nr_stripes &&
2244 drop_one_stripe(conf))
2247 if (conf->slab_cache)
2248 kmem_cache_destroy(conf->slab_cache);
2249 conf->slab_cache = NULL;
2252 static void raid5_end_read_request(struct bio * bi, int error)
2254 struct stripe_head *sh = bi->bi_private;
2255 struct r5conf *conf = sh->raid_conf;
2256 int disks = sh->disks, i;
2257 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2258 char b[BDEVNAME_SIZE];
2259 struct md_rdev *rdev = NULL;
2262 for (i=0 ; i<disks; i++)
2263 if (bi == &sh->dev[i].req)
2266 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2267 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2273 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2274 /* If replacement finished while this request was outstanding,
2275 * 'replacement' might be NULL already.
2276 * In that case it moved down to 'rdev'.
2277 * rdev is not removed until all requests are finished.
2279 rdev = conf->disks[i].replacement;
2281 rdev = conf->disks[i].rdev;
2283 if (use_new_offset(conf, sh))
2284 s = sh->sector + rdev->new_data_offset;
2286 s = sh->sector + rdev->data_offset;
2288 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2289 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2290 /* Note that this cannot happen on a
2291 * replacement device. We just fail those on
2296 "md/raid:%s: read error corrected"
2297 " (%lu sectors at %llu on %s)\n",
2298 mdname(conf->mddev), STRIPE_SECTORS,
2299 (unsigned long long)s,
2300 bdevname(rdev->bdev, b));
2301 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2302 clear_bit(R5_ReadError, &sh->dev[i].flags);
2303 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2304 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2305 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2307 if (atomic_read(&rdev->read_errors))
2308 atomic_set(&rdev->read_errors, 0);
2310 const char *bdn = bdevname(rdev->bdev, b);
2314 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2315 atomic_inc(&rdev->read_errors);
2316 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2319 "md/raid:%s: read error on replacement device "
2320 "(sector %llu on %s).\n",
2321 mdname(conf->mddev),
2322 (unsigned long long)s,
2324 else if (conf->mddev->degraded >= conf->max_degraded) {
2328 "md/raid:%s: read error not correctable "
2329 "(sector %llu on %s).\n",
2330 mdname(conf->mddev),
2331 (unsigned long long)s,
2333 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2338 "md/raid:%s: read error NOT corrected!! "
2339 "(sector %llu on %s).\n",
2340 mdname(conf->mddev),
2341 (unsigned long long)s,
2343 } else if (atomic_read(&rdev->read_errors)
2344 > conf->max_nr_stripes)
2346 "md/raid:%s: Too many read errors, failing device %s.\n",
2347 mdname(conf->mddev), bdn);
2350 if (set_bad && test_bit(In_sync, &rdev->flags)
2351 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2354 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2355 set_bit(R5_ReadError, &sh->dev[i].flags);
2356 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2358 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2360 clear_bit(R5_ReadError, &sh->dev[i].flags);
2361 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2363 && test_bit(In_sync, &rdev->flags)
2364 && rdev_set_badblocks(
2365 rdev, sh->sector, STRIPE_SECTORS, 0)))
2366 md_error(conf->mddev, rdev);
2369 rdev_dec_pending(rdev, conf->mddev);
2370 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2371 set_bit(STRIPE_HANDLE, &sh->state);
2375 static void raid5_end_write_request(struct bio *bi, int error)
2377 struct stripe_head *sh = bi->bi_private;
2378 struct r5conf *conf = sh->raid_conf;
2379 int disks = sh->disks, i;
2380 struct md_rdev *uninitialized_var(rdev);
2381 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2384 int replacement = 0;
2386 for (i = 0 ; i < disks; i++) {
2387 if (bi == &sh->dev[i].req) {
2388 rdev = conf->disks[i].rdev;
2391 if (bi == &sh->dev[i].rreq) {
2392 rdev = conf->disks[i].replacement;
2396 /* rdev was removed and 'replacement'
2397 * replaced it. rdev is not removed
2398 * until all requests are finished.
2400 rdev = conf->disks[i].rdev;
2404 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2405 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2414 md_error(conf->mddev, rdev);
2415 else if (is_badblock(rdev, sh->sector,
2417 &first_bad, &bad_sectors))
2418 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2421 set_bit(STRIPE_DEGRADED, &sh->state);
2422 set_bit(WriteErrorSeen, &rdev->flags);
2423 set_bit(R5_WriteError, &sh->dev[i].flags);
2424 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2425 set_bit(MD_RECOVERY_NEEDED,
2426 &rdev->mddev->recovery);
2427 } else if (is_badblock(rdev, sh->sector,
2429 &first_bad, &bad_sectors)) {
2430 set_bit(R5_MadeGood, &sh->dev[i].flags);
2431 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2432 /* That was a successful write so make
2433 * sure it looks like we already did
2436 set_bit(R5_ReWrite, &sh->dev[i].flags);
2439 rdev_dec_pending(rdev, conf->mddev);
2441 if (sh->batch_head && !uptodate)
2442 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2444 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2445 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2446 set_bit(STRIPE_HANDLE, &sh->state);
2449 if (sh->batch_head && sh != sh->batch_head)
2450 release_stripe(sh->batch_head);
2453 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2455 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2457 struct r5dev *dev = &sh->dev[i];
2459 bio_init(&dev->req);
2460 dev->req.bi_io_vec = &dev->vec;
2461 dev->req.bi_max_vecs = 1;
2462 dev->req.bi_private = sh;
2464 bio_init(&dev->rreq);
2465 dev->rreq.bi_io_vec = &dev->rvec;
2466 dev->rreq.bi_max_vecs = 1;
2467 dev->rreq.bi_private = sh;
2470 dev->sector = compute_blocknr(sh, i, previous);
2473 static void error(struct mddev *mddev, struct md_rdev *rdev)
2475 char b[BDEVNAME_SIZE];
2476 struct r5conf *conf = mddev->private;
2477 unsigned long flags;
2478 pr_debug("raid456: error called\n");
2480 spin_lock_irqsave(&conf->device_lock, flags);
2481 clear_bit(In_sync, &rdev->flags);
2482 mddev->degraded = calc_degraded(conf);
2483 spin_unlock_irqrestore(&conf->device_lock, flags);
2484 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2486 set_bit(Blocked, &rdev->flags);
2487 set_bit(Faulty, &rdev->flags);
2488 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2490 "md/raid:%s: Disk failure on %s, disabling device.\n"
2491 "md/raid:%s: Operation continuing on %d devices.\n",
2493 bdevname(rdev->bdev, b),
2495 conf->raid_disks - mddev->degraded);
2499 * Input: a 'big' sector number,
2500 * Output: index of the data and parity disk, and the sector # in them.
2502 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2503 int previous, int *dd_idx,
2504 struct stripe_head *sh)
2506 sector_t stripe, stripe2;
2507 sector_t chunk_number;
2508 unsigned int chunk_offset;
2511 sector_t new_sector;
2512 int algorithm = previous ? conf->prev_algo
2514 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2515 : conf->chunk_sectors;
2516 int raid_disks = previous ? conf->previous_raid_disks
2518 int data_disks = raid_disks - conf->max_degraded;
2520 /* First compute the information on this sector */
2523 * Compute the chunk number and the sector offset inside the chunk
2525 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2526 chunk_number = r_sector;
2529 * Compute the stripe number
2531 stripe = chunk_number;
2532 *dd_idx = sector_div(stripe, data_disks);
2535 * Select the parity disk based on the user selected algorithm.
2537 pd_idx = qd_idx = -1;
2538 switch(conf->level) {
2540 pd_idx = data_disks;
2543 switch (algorithm) {
2544 case ALGORITHM_LEFT_ASYMMETRIC:
2545 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2546 if (*dd_idx >= pd_idx)
2549 case ALGORITHM_RIGHT_ASYMMETRIC:
2550 pd_idx = sector_div(stripe2, raid_disks);
2551 if (*dd_idx >= pd_idx)
2554 case ALGORITHM_LEFT_SYMMETRIC:
2555 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2556 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2558 case ALGORITHM_RIGHT_SYMMETRIC:
2559 pd_idx = sector_div(stripe2, raid_disks);
2560 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2562 case ALGORITHM_PARITY_0:
2566 case ALGORITHM_PARITY_N:
2567 pd_idx = data_disks;
2575 switch (algorithm) {
2576 case ALGORITHM_LEFT_ASYMMETRIC:
2577 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2578 qd_idx = pd_idx + 1;
2579 if (pd_idx == raid_disks-1) {
2580 (*dd_idx)++; /* Q D D D P */
2582 } else if (*dd_idx >= pd_idx)
2583 (*dd_idx) += 2; /* D D P Q D */
2585 case ALGORITHM_RIGHT_ASYMMETRIC:
2586 pd_idx = sector_div(stripe2, raid_disks);
2587 qd_idx = pd_idx + 1;
2588 if (pd_idx == raid_disks-1) {
2589 (*dd_idx)++; /* Q D D D P */
2591 } else if (*dd_idx >= pd_idx)
2592 (*dd_idx) += 2; /* D D P Q D */
2594 case ALGORITHM_LEFT_SYMMETRIC:
2595 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2596 qd_idx = (pd_idx + 1) % raid_disks;
2597 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2599 case ALGORITHM_RIGHT_SYMMETRIC:
2600 pd_idx = sector_div(stripe2, raid_disks);
2601 qd_idx = (pd_idx + 1) % raid_disks;
2602 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2605 case ALGORITHM_PARITY_0:
2610 case ALGORITHM_PARITY_N:
2611 pd_idx = data_disks;
2612 qd_idx = data_disks + 1;
2615 case ALGORITHM_ROTATING_ZERO_RESTART:
2616 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2617 * of blocks for computing Q is different.
2619 pd_idx = sector_div(stripe2, raid_disks);
2620 qd_idx = pd_idx + 1;
2621 if (pd_idx == raid_disks-1) {
2622 (*dd_idx)++; /* Q D D D P */
2624 } else if (*dd_idx >= pd_idx)
2625 (*dd_idx) += 2; /* D D P Q D */
2629 case ALGORITHM_ROTATING_N_RESTART:
2630 /* Same a left_asymmetric, by first stripe is
2631 * D D D P Q rather than
2635 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2636 qd_idx = pd_idx + 1;
2637 if (pd_idx == raid_disks-1) {
2638 (*dd_idx)++; /* Q D D D P */
2640 } else if (*dd_idx >= pd_idx)
2641 (*dd_idx) += 2; /* D D P Q D */
2645 case ALGORITHM_ROTATING_N_CONTINUE:
2646 /* Same as left_symmetric but Q is before P */
2647 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2648 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2649 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2653 case ALGORITHM_LEFT_ASYMMETRIC_6:
2654 /* RAID5 left_asymmetric, with Q on last device */
2655 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2656 if (*dd_idx >= pd_idx)
2658 qd_idx = raid_disks - 1;
2661 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2662 pd_idx = sector_div(stripe2, raid_disks-1);
2663 if (*dd_idx >= pd_idx)
2665 qd_idx = raid_disks - 1;
2668 case ALGORITHM_LEFT_SYMMETRIC_6:
2669 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2670 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2671 qd_idx = raid_disks - 1;
2674 case ALGORITHM_RIGHT_SYMMETRIC_6:
2675 pd_idx = sector_div(stripe2, raid_disks-1);
2676 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2677 qd_idx = raid_disks - 1;
2680 case ALGORITHM_PARITY_0_6:
2683 qd_idx = raid_disks - 1;
2693 sh->pd_idx = pd_idx;
2694 sh->qd_idx = qd_idx;
2695 sh->ddf_layout = ddf_layout;
2698 * Finally, compute the new sector number
2700 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2704 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2706 struct r5conf *conf = sh->raid_conf;
2707 int raid_disks = sh->disks;
2708 int data_disks = raid_disks - conf->max_degraded;
2709 sector_t new_sector = sh->sector, check;
2710 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2711 : conf->chunk_sectors;
2712 int algorithm = previous ? conf->prev_algo
2716 sector_t chunk_number;
2717 int dummy1, dd_idx = i;
2719 struct stripe_head sh2;
2721 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2722 stripe = new_sector;
2724 if (i == sh->pd_idx)
2726 switch(conf->level) {
2729 switch (algorithm) {
2730 case ALGORITHM_LEFT_ASYMMETRIC:
2731 case ALGORITHM_RIGHT_ASYMMETRIC:
2735 case ALGORITHM_LEFT_SYMMETRIC:
2736 case ALGORITHM_RIGHT_SYMMETRIC:
2739 i -= (sh->pd_idx + 1);
2741 case ALGORITHM_PARITY_0:
2744 case ALGORITHM_PARITY_N:
2751 if (i == sh->qd_idx)
2752 return 0; /* It is the Q disk */
2753 switch (algorithm) {
2754 case ALGORITHM_LEFT_ASYMMETRIC:
2755 case ALGORITHM_RIGHT_ASYMMETRIC:
2756 case ALGORITHM_ROTATING_ZERO_RESTART:
2757 case ALGORITHM_ROTATING_N_RESTART:
2758 if (sh->pd_idx == raid_disks-1)
2759 i--; /* Q D D D P */
2760 else if (i > sh->pd_idx)
2761 i -= 2; /* D D P Q D */
2763 case ALGORITHM_LEFT_SYMMETRIC:
2764 case ALGORITHM_RIGHT_SYMMETRIC:
2765 if (sh->pd_idx == raid_disks-1)
2766 i--; /* Q D D D P */
2771 i -= (sh->pd_idx + 2);
2774 case ALGORITHM_PARITY_0:
2777 case ALGORITHM_PARITY_N:
2779 case ALGORITHM_ROTATING_N_CONTINUE:
2780 /* Like left_symmetric, but P is before Q */
2781 if (sh->pd_idx == 0)
2782 i--; /* P D D D Q */
2787 i -= (sh->pd_idx + 1);
2790 case ALGORITHM_LEFT_ASYMMETRIC_6:
2791 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2795 case ALGORITHM_LEFT_SYMMETRIC_6:
2796 case ALGORITHM_RIGHT_SYMMETRIC_6:
2798 i += data_disks + 1;
2799 i -= (sh->pd_idx + 1);
2801 case ALGORITHM_PARITY_0_6:
2810 chunk_number = stripe * data_disks + i;
2811 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2813 check = raid5_compute_sector(conf, r_sector,
2814 previous, &dummy1, &sh2);
2815 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2816 || sh2.qd_idx != sh->qd_idx) {
2817 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2818 mdname(conf->mddev));
2825 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2826 int rcw, int expand)
2828 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2829 struct r5conf *conf = sh->raid_conf;
2830 int level = conf->level;
2834 for (i = disks; i--; ) {
2835 struct r5dev *dev = &sh->dev[i];
2838 set_bit(R5_LOCKED, &dev->flags);
2839 set_bit(R5_Wantdrain, &dev->flags);
2841 clear_bit(R5_UPTODATE, &dev->flags);
2845 /* if we are not expanding this is a proper write request, and
2846 * there will be bios with new data to be drained into the
2851 /* False alarm, nothing to do */
2853 sh->reconstruct_state = reconstruct_state_drain_run;
2854 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2856 sh->reconstruct_state = reconstruct_state_run;
2858 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2860 if (s->locked + conf->max_degraded == disks)
2861 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2862 atomic_inc(&conf->pending_full_writes);
2864 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2865 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2866 BUG_ON(level == 6 &&
2867 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2868 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2870 for (i = disks; i--; ) {
2871 struct r5dev *dev = &sh->dev[i];
2872 if (i == pd_idx || i == qd_idx)
2876 (test_bit(R5_UPTODATE, &dev->flags) ||
2877 test_bit(R5_Wantcompute, &dev->flags))) {
2878 set_bit(R5_Wantdrain, &dev->flags);
2879 set_bit(R5_LOCKED, &dev->flags);
2880 clear_bit(R5_UPTODATE, &dev->flags);
2885 /* False alarm - nothing to do */
2887 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2888 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2889 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2890 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2893 /* keep the parity disk(s) locked while asynchronous operations
2896 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2897 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2901 int qd_idx = sh->qd_idx;
2902 struct r5dev *dev = &sh->dev[qd_idx];
2904 set_bit(R5_LOCKED, &dev->flags);
2905 clear_bit(R5_UPTODATE, &dev->flags);
2909 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2910 __func__, (unsigned long long)sh->sector,
2911 s->locked, s->ops_request);
2915 * Each stripe/dev can have one or more bion attached.
2916 * toread/towrite point to the first in a chain.
2917 * The bi_next chain must be in order.
2919 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2920 int forwrite, int previous)
2923 struct r5conf *conf = sh->raid_conf;
2926 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2927 (unsigned long long)bi->bi_iter.bi_sector,
2928 (unsigned long long)sh->sector);
2931 * If several bio share a stripe. The bio bi_phys_segments acts as a
2932 * reference count to avoid race. The reference count should already be
2933 * increased before this function is called (for example, in
2934 * make_request()), so other bio sharing this stripe will not free the
2935 * stripe. If a stripe is owned by one stripe, the stripe lock will
2938 spin_lock_irq(&sh->stripe_lock);
2939 /* Don't allow new IO added to stripes in batch list */
2943 bip = &sh->dev[dd_idx].towrite;
2947 bip = &sh->dev[dd_idx].toread;
2948 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2949 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2951 bip = & (*bip)->bi_next;
2953 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2956 if (!forwrite || previous)
2957 clear_bit(STRIPE_BATCH_READY, &sh->state);
2959 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2963 raid5_inc_bi_active_stripes(bi);
2966 /* check if page is covered */
2967 sector_t sector = sh->dev[dd_idx].sector;
2968 for (bi=sh->dev[dd_idx].towrite;
2969 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2970 bi && bi->bi_iter.bi_sector <= sector;
2971 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2972 if (bio_end_sector(bi) >= sector)
2973 sector = bio_end_sector(bi);
2975 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2976 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
2977 sh->overwrite_disks++;
2980 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2981 (unsigned long long)(*bip)->bi_iter.bi_sector,
2982 (unsigned long long)sh->sector, dd_idx);
2983 spin_unlock_irq(&sh->stripe_lock);
2985 if (conf->mddev->bitmap && firstwrite) {
2986 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2988 sh->bm_seq = conf->seq_flush+1;
2989 set_bit(STRIPE_BIT_DELAY, &sh->state);
2992 if (stripe_can_batch(sh))
2993 stripe_add_to_batch_list(conf, sh);
2997 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2998 spin_unlock_irq(&sh->stripe_lock);
3002 static void end_reshape(struct r5conf *conf);
3004 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3005 struct stripe_head *sh)
3007 int sectors_per_chunk =
3008 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3010 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3011 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3013 raid5_compute_sector(conf,
3014 stripe * (disks - conf->max_degraded)
3015 *sectors_per_chunk + chunk_offset,
3021 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3022 struct stripe_head_state *s, int disks,
3023 struct bio **return_bi)
3026 BUG_ON(sh->batch_head);
3027 for (i = disks; i--; ) {
3031 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3032 struct md_rdev *rdev;
3034 rdev = rcu_dereference(conf->disks[i].rdev);
3035 if (rdev && test_bit(In_sync, &rdev->flags))
3036 atomic_inc(&rdev->nr_pending);
3041 if (!rdev_set_badblocks(
3045 md_error(conf->mddev, rdev);
3046 rdev_dec_pending(rdev, conf->mddev);
3049 spin_lock_irq(&sh->stripe_lock);
3050 /* fail all writes first */
3051 bi = sh->dev[i].towrite;
3052 sh->dev[i].towrite = NULL;
3053 sh->overwrite_disks = 0;
3054 spin_unlock_irq(&sh->stripe_lock);
3058 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3059 wake_up(&conf->wait_for_overlap);
3061 while (bi && bi->bi_iter.bi_sector <
3062 sh->dev[i].sector + STRIPE_SECTORS) {
3063 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3064 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3065 if (!raid5_dec_bi_active_stripes(bi)) {
3066 md_write_end(conf->mddev);
3067 bi->bi_next = *return_bi;
3073 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3074 STRIPE_SECTORS, 0, 0);
3076 /* and fail all 'written' */
3077 bi = sh->dev[i].written;
3078 sh->dev[i].written = NULL;
3079 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3080 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3081 sh->dev[i].page = sh->dev[i].orig_page;
3084 if (bi) bitmap_end = 1;
3085 while (bi && bi->bi_iter.bi_sector <
3086 sh->dev[i].sector + STRIPE_SECTORS) {
3087 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3088 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3089 if (!raid5_dec_bi_active_stripes(bi)) {
3090 md_write_end(conf->mddev);
3091 bi->bi_next = *return_bi;
3097 /* fail any reads if this device is non-operational and
3098 * the data has not reached the cache yet.
3100 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3101 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3102 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3103 spin_lock_irq(&sh->stripe_lock);
3104 bi = sh->dev[i].toread;
3105 sh->dev[i].toread = NULL;
3106 spin_unlock_irq(&sh->stripe_lock);
3107 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3108 wake_up(&conf->wait_for_overlap);
3109 while (bi && bi->bi_iter.bi_sector <
3110 sh->dev[i].sector + STRIPE_SECTORS) {
3111 struct bio *nextbi =
3112 r5_next_bio(bi, sh->dev[i].sector);
3113 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3114 if (!raid5_dec_bi_active_stripes(bi)) {
3115 bi->bi_next = *return_bi;
3122 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3123 STRIPE_SECTORS, 0, 0);
3124 /* If we were in the middle of a write the parity block might
3125 * still be locked - so just clear all R5_LOCKED flags
3127 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3130 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3131 if (atomic_dec_and_test(&conf->pending_full_writes))
3132 md_wakeup_thread(conf->mddev->thread);
3136 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3137 struct stripe_head_state *s)
3142 BUG_ON(sh->batch_head);
3143 clear_bit(STRIPE_SYNCING, &sh->state);
3144 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3145 wake_up(&conf->wait_for_overlap);
3148 /* There is nothing more to do for sync/check/repair.
3149 * Don't even need to abort as that is handled elsewhere
3150 * if needed, and not always wanted e.g. if there is a known
3152 * For recover/replace we need to record a bad block on all
3153 * non-sync devices, or abort the recovery
3155 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3156 /* During recovery devices cannot be removed, so
3157 * locking and refcounting of rdevs is not needed
3159 for (i = 0; i < conf->raid_disks; i++) {
3160 struct md_rdev *rdev = conf->disks[i].rdev;
3162 && !test_bit(Faulty, &rdev->flags)
3163 && !test_bit(In_sync, &rdev->flags)
3164 && !rdev_set_badblocks(rdev, sh->sector,
3167 rdev = conf->disks[i].replacement;
3169 && !test_bit(Faulty, &rdev->flags)
3170 && !test_bit(In_sync, &rdev->flags)
3171 && !rdev_set_badblocks(rdev, sh->sector,
3176 conf->recovery_disabled =
3177 conf->mddev->recovery_disabled;
3179 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3182 static int want_replace(struct stripe_head *sh, int disk_idx)
3184 struct md_rdev *rdev;
3186 /* Doing recovery so rcu locking not required */
3187 rdev = sh->raid_conf->disks[disk_idx].replacement;
3189 && !test_bit(Faulty, &rdev->flags)
3190 && !test_bit(In_sync, &rdev->flags)
3191 && (rdev->recovery_offset <= sh->sector
3192 || rdev->mddev->recovery_cp <= sh->sector))
3198 /* fetch_block - checks the given member device to see if its data needs
3199 * to be read or computed to satisfy a request.
3201 * Returns 1 when no more member devices need to be checked, otherwise returns
3202 * 0 to tell the loop in handle_stripe_fill to continue
3205 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3206 int disk_idx, int disks)
3208 struct r5dev *dev = &sh->dev[disk_idx];
3209 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3210 &sh->dev[s->failed_num[1]] };
3214 if (test_bit(R5_LOCKED, &dev->flags) ||
3215 test_bit(R5_UPTODATE, &dev->flags))
3216 /* No point reading this as we already have it or have
3217 * decided to get it.
3222 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3223 /* We need this block to directly satisfy a request */
3226 if (s->syncing || s->expanding ||
3227 (s->replacing && want_replace(sh, disk_idx)))
3228 /* When syncing, or expanding we read everything.
3229 * When replacing, we need the replaced block.
3233 if ((s->failed >= 1 && fdev[0]->toread) ||
3234 (s->failed >= 2 && fdev[1]->toread))
3235 /* If we want to read from a failed device, then
3236 * we need to actually read every other device.
3240 /* Sometimes neither read-modify-write nor reconstruct-write
3241 * cycles can work. In those cases we read every block we
3242 * can. Then the parity-update is certain to have enough to
3244 * This can only be a problem when we need to write something,
3245 * and some device has failed. If either of those tests
3246 * fail we need look no further.
3248 if (!s->failed || !s->to_write)
3251 if (test_bit(R5_Insync, &dev->flags) &&
3252 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3253 /* Pre-reads at not permitted until after short delay
3254 * to gather multiple requests. However if this
3255 * device is no Insync, the block could only be be computed
3256 * and there is no need to delay that.
3260 for (i = 0; i < s->failed; i++) {
3261 if (fdev[i]->towrite &&
3262 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3263 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3264 /* If we have a partial write to a failed
3265 * device, then we will need to reconstruct
3266 * the content of that device, so all other
3267 * devices must be read.
3272 /* If we are forced to do a reconstruct-write, either because
3273 * the current RAID6 implementation only supports that, or
3274 * or because parity cannot be trusted and we are currently
3275 * recovering it, there is extra need to be careful.
3276 * If one of the devices that we would need to read, because
3277 * it is not being overwritten (and maybe not written at all)
3278 * is missing/faulty, then we need to read everything we can.
3280 if (sh->raid_conf->level != 6 &&
3281 sh->sector < sh->raid_conf->mddev->recovery_cp)
3282 /* reconstruct-write isn't being forced */
3284 for (i = 0; i < s->failed; i++) {
3285 if (s->failed_num[i] != sh->pd_idx &&
3286 s->failed_num[i] != sh->qd_idx &&
3287 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3288 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3295 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3296 int disk_idx, int disks)
3298 struct r5dev *dev = &sh->dev[disk_idx];
3300 /* is the data in this block needed, and can we get it? */
3301 if (need_this_block(sh, s, disk_idx, disks)) {
3302 /* we would like to get this block, possibly by computing it,
3303 * otherwise read it if the backing disk is insync
3305 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3306 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3307 BUG_ON(sh->batch_head);
3308 if ((s->uptodate == disks - 1) &&
3309 (s->failed && (disk_idx == s->failed_num[0] ||
3310 disk_idx == s->failed_num[1]))) {
3311 /* have disk failed, and we're requested to fetch it;
3314 pr_debug("Computing stripe %llu block %d\n",
3315 (unsigned long long)sh->sector, disk_idx);
3316 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3317 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3318 set_bit(R5_Wantcompute, &dev->flags);
3319 sh->ops.target = disk_idx;
3320 sh->ops.target2 = -1; /* no 2nd target */
3322 /* Careful: from this point on 'uptodate' is in the eye
3323 * of raid_run_ops which services 'compute' operations
3324 * before writes. R5_Wantcompute flags a block that will
3325 * be R5_UPTODATE by the time it is needed for a
3326 * subsequent operation.
3330 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3331 /* Computing 2-failure is *very* expensive; only
3332 * do it if failed >= 2
3335 for (other = disks; other--; ) {
3336 if (other == disk_idx)
3338 if (!test_bit(R5_UPTODATE,
3339 &sh->dev[other].flags))
3343 pr_debug("Computing stripe %llu blocks %d,%d\n",
3344 (unsigned long long)sh->sector,
3346 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3347 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3348 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3349 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3350 sh->ops.target = disk_idx;
3351 sh->ops.target2 = other;
3355 } else if (test_bit(R5_Insync, &dev->flags)) {
3356 set_bit(R5_LOCKED, &dev->flags);
3357 set_bit(R5_Wantread, &dev->flags);
3359 pr_debug("Reading block %d (sync=%d)\n",
3360 disk_idx, s->syncing);
3368 * handle_stripe_fill - read or compute data to satisfy pending requests.
3370 static void handle_stripe_fill(struct stripe_head *sh,
3371 struct stripe_head_state *s,
3376 /* look for blocks to read/compute, skip this if a compute
3377 * is already in flight, or if the stripe contents are in the
3378 * midst of changing due to a write
3380 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3381 !sh->reconstruct_state)
3382 for (i = disks; i--; )
3383 if (fetch_block(sh, s, i, disks))
3385 set_bit(STRIPE_HANDLE, &sh->state);
3388 /* handle_stripe_clean_event
3389 * any written block on an uptodate or failed drive can be returned.
3390 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3391 * never LOCKED, so we don't need to test 'failed' directly.
3393 static void handle_stripe_clean_event(struct r5conf *conf,
3394 struct stripe_head *sh, int disks, struct bio **return_bi)
3398 int discard_pending = 0;
3399 struct stripe_head *head_sh = sh;
3400 bool do_endio = false;
3403 for (i = disks; i--; )
3404 if (sh->dev[i].written) {
3406 if (!test_bit(R5_LOCKED, &dev->flags) &&
3407 (test_bit(R5_UPTODATE, &dev->flags) ||
3408 test_bit(R5_Discard, &dev->flags) ||
3409 test_bit(R5_SkipCopy, &dev->flags))) {
3410 /* We can return any write requests */
3411 struct bio *wbi, *wbi2;
3412 pr_debug("Return write for disc %d\n", i);
3413 if (test_and_clear_bit(R5_Discard, &dev->flags))
3414 clear_bit(R5_UPTODATE, &dev->flags);
3415 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3416 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3421 dev->page = dev->orig_page;
3423 dev->written = NULL;
3424 while (wbi && wbi->bi_iter.bi_sector <
3425 dev->sector + STRIPE_SECTORS) {
3426 wbi2 = r5_next_bio(wbi, dev->sector);
3427 if (!raid5_dec_bi_active_stripes(wbi)) {
3428 md_write_end(conf->mddev);
3429 wbi->bi_next = *return_bi;
3434 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3436 !test_bit(STRIPE_DEGRADED, &sh->state),
3438 if (head_sh->batch_head) {
3439 sh = list_first_entry(&sh->batch_list,
3442 if (sh != head_sh) {
3449 } else if (test_bit(R5_Discard, &dev->flags))
3450 discard_pending = 1;
3451 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3452 WARN_ON(dev->page != dev->orig_page);
3454 if (!discard_pending &&
3455 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3456 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3457 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3458 if (sh->qd_idx >= 0) {
3459 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3460 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3462 /* now that discard is done we can proceed with any sync */
3463 clear_bit(STRIPE_DISCARD, &sh->state);
3465 * SCSI discard will change some bio fields and the stripe has
3466 * no updated data, so remove it from hash list and the stripe
3467 * will be reinitialized
3469 spin_lock_irq(&conf->device_lock);
3472 if (head_sh->batch_head) {
3473 sh = list_first_entry(&sh->batch_list,
3474 struct stripe_head, batch_list);
3478 spin_unlock_irq(&conf->device_lock);
3481 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3482 set_bit(STRIPE_HANDLE, &sh->state);
3486 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3487 if (atomic_dec_and_test(&conf->pending_full_writes))
3488 md_wakeup_thread(conf->mddev->thread);
3490 if (!head_sh->batch_head || !do_endio)
3492 for (i = 0; i < head_sh->disks; i++) {
3493 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
3496 while (!list_empty(&head_sh->batch_list)) {
3498 sh = list_first_entry(&head_sh->batch_list,
3499 struct stripe_head, batch_list);
3500 list_del_init(&sh->batch_list);
3502 set_mask_bits(&sh->state, ~STRIPE_EXPAND_SYNC_FLAG,
3503 head_sh->state & ~((1 << STRIPE_ACTIVE) |
3504 (1 << STRIPE_PREREAD_ACTIVE) |
3505 STRIPE_EXPAND_SYNC_FLAG));
3506 sh->check_state = head_sh->check_state;
3507 sh->reconstruct_state = head_sh->reconstruct_state;
3508 for (i = 0; i < sh->disks; i++) {
3509 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3511 sh->dev[i].flags = head_sh->dev[i].flags;
3514 spin_lock_irq(&sh->stripe_lock);
3515 sh->batch_head = NULL;
3516 spin_unlock_irq(&sh->stripe_lock);
3517 if (sh->state & STRIPE_EXPAND_SYNC_FLAG)
3518 set_bit(STRIPE_HANDLE, &sh->state);
3522 spin_lock_irq(&head_sh->stripe_lock);
3523 head_sh->batch_head = NULL;
3524 spin_unlock_irq(&head_sh->stripe_lock);
3525 wake_up_nr(&conf->wait_for_overlap, wakeup_nr);
3526 if (head_sh->state & STRIPE_EXPAND_SYNC_FLAG)
3527 set_bit(STRIPE_HANDLE, &head_sh->state);
3530 static void handle_stripe_dirtying(struct r5conf *conf,
3531 struct stripe_head *sh,
3532 struct stripe_head_state *s,
3535 int rmw = 0, rcw = 0, i;
3536 sector_t recovery_cp = conf->mddev->recovery_cp;
3538 /* Check whether resync is now happening or should start.
3539 * If yes, then the array is dirty (after unclean shutdown or
3540 * initial creation), so parity in some stripes might be inconsistent.
3541 * In this case, we need to always do reconstruct-write, to ensure
3542 * that in case of drive failure or read-error correction, we
3543 * generate correct data from the parity.
3545 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3546 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3548 /* Calculate the real rcw later - for now make it
3549 * look like rcw is cheaper
3552 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3553 conf->rmw_level, (unsigned long long)recovery_cp,
3554 (unsigned long long)sh->sector);
3555 } else for (i = disks; i--; ) {
3556 /* would I have to read this buffer for read_modify_write */
3557 struct r5dev *dev = &sh->dev[i];
3558 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3559 !test_bit(R5_LOCKED, &dev->flags) &&
3560 !(test_bit(R5_UPTODATE, &dev->flags) ||
3561 test_bit(R5_Wantcompute, &dev->flags))) {
3562 if (test_bit(R5_Insync, &dev->flags))
3565 rmw += 2*disks; /* cannot read it */
3567 /* Would I have to read this buffer for reconstruct_write */
3568 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3569 i != sh->pd_idx && i != sh->qd_idx &&
3570 !test_bit(R5_LOCKED, &dev->flags) &&
3571 !(test_bit(R5_UPTODATE, &dev->flags) ||
3572 test_bit(R5_Wantcompute, &dev->flags))) {
3573 if (test_bit(R5_Insync, &dev->flags))
3579 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3580 (unsigned long long)sh->sector, rmw, rcw);
3581 set_bit(STRIPE_HANDLE, &sh->state);
3582 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3583 /* prefer read-modify-write, but need to get some data */
3584 if (conf->mddev->queue)
3585 blk_add_trace_msg(conf->mddev->queue,
3586 "raid5 rmw %llu %d",
3587 (unsigned long long)sh->sector, rmw);
3588 for (i = disks; i--; ) {
3589 struct r5dev *dev = &sh->dev[i];
3590 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3591 !test_bit(R5_LOCKED, &dev->flags) &&
3592 !(test_bit(R5_UPTODATE, &dev->flags) ||
3593 test_bit(R5_Wantcompute, &dev->flags)) &&
3594 test_bit(R5_Insync, &dev->flags)) {
3595 if (test_bit(STRIPE_PREREAD_ACTIVE,
3597 pr_debug("Read_old block %d for r-m-w\n",
3599 set_bit(R5_LOCKED, &dev->flags);
3600 set_bit(R5_Wantread, &dev->flags);
3603 set_bit(STRIPE_DELAYED, &sh->state);
3604 set_bit(STRIPE_HANDLE, &sh->state);
3609 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3610 /* want reconstruct write, but need to get some data */
3613 for (i = disks; i--; ) {
3614 struct r5dev *dev = &sh->dev[i];
3615 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3616 i != sh->pd_idx && i != sh->qd_idx &&
3617 !test_bit(R5_LOCKED, &dev->flags) &&
3618 !(test_bit(R5_UPTODATE, &dev->flags) ||
3619 test_bit(R5_Wantcompute, &dev->flags))) {
3621 if (test_bit(R5_Insync, &dev->flags) &&
3622 test_bit(STRIPE_PREREAD_ACTIVE,
3624 pr_debug("Read_old block "
3625 "%d for Reconstruct\n", i);
3626 set_bit(R5_LOCKED, &dev->flags);
3627 set_bit(R5_Wantread, &dev->flags);
3631 set_bit(STRIPE_DELAYED, &sh->state);
3632 set_bit(STRIPE_HANDLE, &sh->state);
3636 if (rcw && conf->mddev->queue)
3637 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3638 (unsigned long long)sh->sector,
3639 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3642 if (rcw > disks && rmw > disks &&
3643 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3644 set_bit(STRIPE_DELAYED, &sh->state);
3646 /* now if nothing is locked, and if we have enough data,
3647 * we can start a write request
3649 /* since handle_stripe can be called at any time we need to handle the
3650 * case where a compute block operation has been submitted and then a
3651 * subsequent call wants to start a write request. raid_run_ops only
3652 * handles the case where compute block and reconstruct are requested
3653 * simultaneously. If this is not the case then new writes need to be
3654 * held off until the compute completes.
3656 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3657 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3658 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3659 schedule_reconstruction(sh, s, rcw == 0, 0);
3662 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3663 struct stripe_head_state *s, int disks)
3665 struct r5dev *dev = NULL;
3667 BUG_ON(sh->batch_head);
3668 set_bit(STRIPE_HANDLE, &sh->state);
3670 switch (sh->check_state) {
3671 case check_state_idle:
3672 /* start a new check operation if there are no failures */
3673 if (s->failed == 0) {
3674 BUG_ON(s->uptodate != disks);
3675 sh->check_state = check_state_run;
3676 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3677 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3681 dev = &sh->dev[s->failed_num[0]];
3683 case check_state_compute_result:
3684 sh->check_state = check_state_idle;
3686 dev = &sh->dev[sh->pd_idx];
3688 /* check that a write has not made the stripe insync */
3689 if (test_bit(STRIPE_INSYNC, &sh->state))
3692 /* either failed parity check, or recovery is happening */
3693 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3694 BUG_ON(s->uptodate != disks);
3696 set_bit(R5_LOCKED, &dev->flags);
3698 set_bit(R5_Wantwrite, &dev->flags);
3700 clear_bit(STRIPE_DEGRADED, &sh->state);
3701 set_bit(STRIPE_INSYNC, &sh->state);
3703 case check_state_run:
3704 break; /* we will be called again upon completion */
3705 case check_state_check_result:
3706 sh->check_state = check_state_idle;
3708 /* if a failure occurred during the check operation, leave
3709 * STRIPE_INSYNC not set and let the stripe be handled again
3714 /* handle a successful check operation, if parity is correct
3715 * we are done. Otherwise update the mismatch count and repair
3716 * parity if !MD_RECOVERY_CHECK
3718 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3719 /* parity is correct (on disc,
3720 * not in buffer any more)
3722 set_bit(STRIPE_INSYNC, &sh->state);
3724 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3725 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3726 /* don't try to repair!! */
3727 set_bit(STRIPE_INSYNC, &sh->state);
3729 sh->check_state = check_state_compute_run;
3730 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3731 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3732 set_bit(R5_Wantcompute,
3733 &sh->dev[sh->pd_idx].flags);
3734 sh->ops.target = sh->pd_idx;
3735 sh->ops.target2 = -1;
3740 case check_state_compute_run:
3743 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3744 __func__, sh->check_state,
3745 (unsigned long long) sh->sector);
3750 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3751 struct stripe_head_state *s,
3754 int pd_idx = sh->pd_idx;
3755 int qd_idx = sh->qd_idx;
3758 BUG_ON(sh->batch_head);
3759 set_bit(STRIPE_HANDLE, &sh->state);
3761 BUG_ON(s->failed > 2);
3763 /* Want to check and possibly repair P and Q.
3764 * However there could be one 'failed' device, in which
3765 * case we can only check one of them, possibly using the
3766 * other to generate missing data
3769 switch (sh->check_state) {
3770 case check_state_idle:
3771 /* start a new check operation if there are < 2 failures */
3772 if (s->failed == s->q_failed) {
3773 /* The only possible failed device holds Q, so it
3774 * makes sense to check P (If anything else were failed,
3775 * we would have used P to recreate it).
3777 sh->check_state = check_state_run;
3779 if (!s->q_failed && s->failed < 2) {
3780 /* Q is not failed, and we didn't use it to generate
3781 * anything, so it makes sense to check it
3783 if (sh->check_state == check_state_run)
3784 sh->check_state = check_state_run_pq;
3786 sh->check_state = check_state_run_q;
3789 /* discard potentially stale zero_sum_result */
3790 sh->ops.zero_sum_result = 0;
3792 if (sh->check_state == check_state_run) {
3793 /* async_xor_zero_sum destroys the contents of P */
3794 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3797 if (sh->check_state >= check_state_run &&
3798 sh->check_state <= check_state_run_pq) {
3799 /* async_syndrome_zero_sum preserves P and Q, so
3800 * no need to mark them !uptodate here
3802 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3806 /* we have 2-disk failure */
3807 BUG_ON(s->failed != 2);
3809 case check_state_compute_result:
3810 sh->check_state = check_state_idle;
3812 /* check that a write has not made the stripe insync */
3813 if (test_bit(STRIPE_INSYNC, &sh->state))
3816 /* now write out any block on a failed drive,
3817 * or P or Q if they were recomputed
3819 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3820 if (s->failed == 2) {
3821 dev = &sh->dev[s->failed_num[1]];
3823 set_bit(R5_LOCKED, &dev->flags);
3824 set_bit(R5_Wantwrite, &dev->flags);
3826 if (s->failed >= 1) {
3827 dev = &sh->dev[s->failed_num[0]];
3829 set_bit(R5_LOCKED, &dev->flags);
3830 set_bit(R5_Wantwrite, &dev->flags);
3832 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3833 dev = &sh->dev[pd_idx];
3835 set_bit(R5_LOCKED, &dev->flags);
3836 set_bit(R5_Wantwrite, &dev->flags);
3838 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3839 dev = &sh->dev[qd_idx];
3841 set_bit(R5_LOCKED, &dev->flags);
3842 set_bit(R5_Wantwrite, &dev->flags);
3844 clear_bit(STRIPE_DEGRADED, &sh->state);
3846 set_bit(STRIPE_INSYNC, &sh->state);
3848 case check_state_run:
3849 case check_state_run_q:
3850 case check_state_run_pq:
3851 break; /* we will be called again upon completion */
3852 case check_state_check_result:
3853 sh->check_state = check_state_idle;
3855 /* handle a successful check operation, if parity is correct
3856 * we are done. Otherwise update the mismatch count and repair
3857 * parity if !MD_RECOVERY_CHECK
3859 if (sh->ops.zero_sum_result == 0) {
3860 /* both parities are correct */
3862 set_bit(STRIPE_INSYNC, &sh->state);
3864 /* in contrast to the raid5 case we can validate
3865 * parity, but still have a failure to write
3868 sh->check_state = check_state_compute_result;
3869 /* Returning at this point means that we may go
3870 * off and bring p and/or q uptodate again so
3871 * we make sure to check zero_sum_result again
3872 * to verify if p or q need writeback
3876 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3877 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3878 /* don't try to repair!! */
3879 set_bit(STRIPE_INSYNC, &sh->state);
3881 int *target = &sh->ops.target;
3883 sh->ops.target = -1;
3884 sh->ops.target2 = -1;
3885 sh->check_state = check_state_compute_run;
3886 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3887 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3888 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3889 set_bit(R5_Wantcompute,
3890 &sh->dev[pd_idx].flags);
3892 target = &sh->ops.target2;
3895 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3896 set_bit(R5_Wantcompute,
3897 &sh->dev[qd_idx].flags);
3904 case check_state_compute_run:
3907 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3908 __func__, sh->check_state,
3909 (unsigned long long) sh->sector);
3914 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3918 /* We have read all the blocks in this stripe and now we need to
3919 * copy some of them into a target stripe for expand.
3921 struct dma_async_tx_descriptor *tx = NULL;
3922 BUG_ON(sh->batch_head);
3923 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3924 for (i = 0; i < sh->disks; i++)
3925 if (i != sh->pd_idx && i != sh->qd_idx) {
3927 struct stripe_head *sh2;
3928 struct async_submit_ctl submit;
3930 sector_t bn = compute_blocknr(sh, i, 1);
3931 sector_t s = raid5_compute_sector(conf, bn, 0,
3933 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3935 /* so far only the early blocks of this stripe
3936 * have been requested. When later blocks
3937 * get requested, we will try again
3940 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3941 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3942 /* must have already done this block */
3943 release_stripe(sh2);
3947 /* place all the copies on one channel */
3948 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3949 tx = async_memcpy(sh2->dev[dd_idx].page,
3950 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3953 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3954 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3955 for (j = 0; j < conf->raid_disks; j++)
3956 if (j != sh2->pd_idx &&
3958 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3960 if (j == conf->raid_disks) {
3961 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3962 set_bit(STRIPE_HANDLE, &sh2->state);
3964 release_stripe(sh2);
3967 /* done submitting copies, wait for them to complete */
3968 async_tx_quiesce(&tx);
3972 * handle_stripe - do things to a stripe.
3974 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3975 * state of various bits to see what needs to be done.
3977 * return some read requests which now have data
3978 * return some write requests which are safely on storage
3979 * schedule a read on some buffers
3980 * schedule a write of some buffers
3981 * return confirmation of parity correctness
3985 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3987 struct r5conf *conf = sh->raid_conf;
3988 int disks = sh->disks;
3991 int do_recovery = 0;
3993 memset(s, 0, sizeof(*s));
3995 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
3996 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
3997 s->failed_num[0] = -1;
3998 s->failed_num[1] = -1;
4000 /* Now to look around and see what can be done */
4002 for (i=disks; i--; ) {
4003 struct md_rdev *rdev;
4010 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4012 dev->toread, dev->towrite, dev->written);
4013 /* maybe we can reply to a read
4015 * new wantfill requests are only permitted while
4016 * ops_complete_biofill is guaranteed to be inactive
4018 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4019 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4020 set_bit(R5_Wantfill, &dev->flags);
4022 /* now count some things */
4023 if (test_bit(R5_LOCKED, &dev->flags))
4025 if (test_bit(R5_UPTODATE, &dev->flags))
4027 if (test_bit(R5_Wantcompute, &dev->flags)) {
4029 BUG_ON(s->compute > 2);
4032 if (test_bit(R5_Wantfill, &dev->flags))
4034 else if (dev->toread)
4038 if (!test_bit(R5_OVERWRITE, &dev->flags))
4043 /* Prefer to use the replacement for reads, but only
4044 * if it is recovered enough and has no bad blocks.
4046 rdev = rcu_dereference(conf->disks[i].replacement);
4047 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4048 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4049 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4050 &first_bad, &bad_sectors))
4051 set_bit(R5_ReadRepl, &dev->flags);
4054 set_bit(R5_NeedReplace, &dev->flags);
4055 rdev = rcu_dereference(conf->disks[i].rdev);
4056 clear_bit(R5_ReadRepl, &dev->flags);
4058 if (rdev && test_bit(Faulty, &rdev->flags))
4061 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4062 &first_bad, &bad_sectors);
4063 if (s->blocked_rdev == NULL
4064 && (test_bit(Blocked, &rdev->flags)
4067 set_bit(BlockedBadBlocks,
4069 s->blocked_rdev = rdev;
4070 atomic_inc(&rdev->nr_pending);
4073 clear_bit(R5_Insync, &dev->flags);
4077 /* also not in-sync */
4078 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4079 test_bit(R5_UPTODATE, &dev->flags)) {
4080 /* treat as in-sync, but with a read error
4081 * which we can now try to correct
4083 set_bit(R5_Insync, &dev->flags);
4084 set_bit(R5_ReadError, &dev->flags);
4086 } else if (test_bit(In_sync, &rdev->flags))
4087 set_bit(R5_Insync, &dev->flags);
4088 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4089 /* in sync if before recovery_offset */
4090 set_bit(R5_Insync, &dev->flags);
4091 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4092 test_bit(R5_Expanded, &dev->flags))
4093 /* If we've reshaped into here, we assume it is Insync.
4094 * We will shortly update recovery_offset to make
4097 set_bit(R5_Insync, &dev->flags);
4099 if (test_bit(R5_WriteError, &dev->flags)) {
4100 /* This flag does not apply to '.replacement'
4101 * only to .rdev, so make sure to check that*/
4102 struct md_rdev *rdev2 = rcu_dereference(
4103 conf->disks[i].rdev);
4105 clear_bit(R5_Insync, &dev->flags);
4106 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4107 s->handle_bad_blocks = 1;
4108 atomic_inc(&rdev2->nr_pending);
4110 clear_bit(R5_WriteError, &dev->flags);
4112 if (test_bit(R5_MadeGood, &dev->flags)) {
4113 /* This flag does not apply to '.replacement'
4114 * only to .rdev, so make sure to check that*/
4115 struct md_rdev *rdev2 = rcu_dereference(
4116 conf->disks[i].rdev);
4117 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4118 s->handle_bad_blocks = 1;
4119 atomic_inc(&rdev2->nr_pending);
4121 clear_bit(R5_MadeGood, &dev->flags);
4123 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4124 struct md_rdev *rdev2 = rcu_dereference(
4125 conf->disks[i].replacement);
4126 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4127 s->handle_bad_blocks = 1;
4128 atomic_inc(&rdev2->nr_pending);
4130 clear_bit(R5_MadeGoodRepl, &dev->flags);
4132 if (!test_bit(R5_Insync, &dev->flags)) {
4133 /* The ReadError flag will just be confusing now */
4134 clear_bit(R5_ReadError, &dev->flags);
4135 clear_bit(R5_ReWrite, &dev->flags);
4137 if (test_bit(R5_ReadError, &dev->flags))
4138 clear_bit(R5_Insync, &dev->flags);
4139 if (!test_bit(R5_Insync, &dev->flags)) {
4141 s->failed_num[s->failed] = i;
4143 if (rdev && !test_bit(Faulty, &rdev->flags))
4147 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4148 /* If there is a failed device being replaced,
4149 * we must be recovering.
4150 * else if we are after recovery_cp, we must be syncing
4151 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4152 * else we can only be replacing
4153 * sync and recovery both need to read all devices, and so
4154 * use the same flag.
4157 sh->sector >= conf->mddev->recovery_cp ||
4158 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4166 static int clear_batch_ready(struct stripe_head *sh)
4168 struct stripe_head *tmp;
4169 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4171 spin_lock(&sh->stripe_lock);
4172 if (!sh->batch_head) {
4173 spin_unlock(&sh->stripe_lock);
4178 * this stripe could be added to a batch list before we check
4179 * BATCH_READY, skips it
4181 if (sh->batch_head != sh) {
4182 spin_unlock(&sh->stripe_lock);
4185 spin_lock(&sh->batch_lock);
4186 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4187 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4188 spin_unlock(&sh->batch_lock);
4189 spin_unlock(&sh->stripe_lock);
4192 * BATCH_READY is cleared, no new stripes can be added.
4193 * batch_list can be accessed without lock
4198 static void check_break_stripe_batch_list(struct stripe_head *sh)
4200 struct stripe_head *head_sh, *next;
4203 if (!test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4208 sh = list_first_entry(&sh->batch_list,
4209 struct stripe_head, batch_list);
4210 BUG_ON(sh == head_sh);
4211 } while (!test_bit(STRIPE_DEGRADED, &sh->state));
4213 while (sh != head_sh) {
4214 next = list_first_entry(&sh->batch_list,
4215 struct stripe_head, batch_list);
4216 list_del_init(&sh->batch_list);
4218 set_mask_bits(&sh->state, ~STRIPE_EXPAND_SYNC_FLAG,
4219 head_sh->state & ~((1 << STRIPE_ACTIVE) |
4220 (1 << STRIPE_PREREAD_ACTIVE) |
4221 (1 << STRIPE_DEGRADED) |
4222 STRIPE_EXPAND_SYNC_FLAG));
4223 sh->check_state = head_sh->check_state;
4224 sh->reconstruct_state = head_sh->reconstruct_state;
4225 for (i = 0; i < sh->disks; i++)
4226 sh->dev[i].flags = head_sh->dev[i].flags &
4227 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4229 spin_lock_irq(&sh->stripe_lock);
4230 sh->batch_head = NULL;
4231 spin_unlock_irq(&sh->stripe_lock);
4233 set_bit(STRIPE_HANDLE, &sh->state);
4240 static void handle_stripe(struct stripe_head *sh)
4242 struct stripe_head_state s;
4243 struct r5conf *conf = sh->raid_conf;
4246 int disks = sh->disks;
4247 struct r5dev *pdev, *qdev;
4249 clear_bit(STRIPE_HANDLE, &sh->state);
4250 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4251 /* already being handled, ensure it gets handled
4252 * again when current action finishes */
4253 set_bit(STRIPE_HANDLE, &sh->state);
4257 if (clear_batch_ready(sh) ) {
4258 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4262 check_break_stripe_batch_list(sh);
4264 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4265 spin_lock(&sh->stripe_lock);
4266 /* Cannot process 'sync' concurrently with 'discard' */
4267 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4268 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4269 set_bit(STRIPE_SYNCING, &sh->state);
4270 clear_bit(STRIPE_INSYNC, &sh->state);
4271 clear_bit(STRIPE_REPLACED, &sh->state);
4273 spin_unlock(&sh->stripe_lock);
4275 clear_bit(STRIPE_DELAYED, &sh->state);
4277 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4278 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4279 (unsigned long long)sh->sector, sh->state,
4280 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4281 sh->check_state, sh->reconstruct_state);
4283 analyse_stripe(sh, &s);
4285 if (s.handle_bad_blocks) {
4286 set_bit(STRIPE_HANDLE, &sh->state);
4290 if (unlikely(s.blocked_rdev)) {
4291 if (s.syncing || s.expanding || s.expanded ||
4292 s.replacing || s.to_write || s.written) {
4293 set_bit(STRIPE_HANDLE, &sh->state);
4296 /* There is nothing for the blocked_rdev to block */
4297 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4298 s.blocked_rdev = NULL;
4301 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4302 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4303 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4306 pr_debug("locked=%d uptodate=%d to_read=%d"
4307 " to_write=%d failed=%d failed_num=%d,%d\n",
4308 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4309 s.failed_num[0], s.failed_num[1]);
4310 /* check if the array has lost more than max_degraded devices and,
4311 * if so, some requests might need to be failed.
4313 if (s.failed > conf->max_degraded) {
4314 sh->check_state = 0;
4315 sh->reconstruct_state = 0;
4316 if (s.to_read+s.to_write+s.written)
4317 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4318 if (s.syncing + s.replacing)
4319 handle_failed_sync(conf, sh, &s);
4322 /* Now we check to see if any write operations have recently
4326 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4328 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4329 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4330 sh->reconstruct_state = reconstruct_state_idle;
4332 /* All the 'written' buffers and the parity block are ready to
4333 * be written back to disk
4335 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4336 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4337 BUG_ON(sh->qd_idx >= 0 &&
4338 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4339 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4340 for (i = disks; i--; ) {
4341 struct r5dev *dev = &sh->dev[i];
4342 if (test_bit(R5_LOCKED, &dev->flags) &&
4343 (i == sh->pd_idx || i == sh->qd_idx ||
4345 pr_debug("Writing block %d\n", i);
4346 set_bit(R5_Wantwrite, &dev->flags);
4351 if (!test_bit(R5_Insync, &dev->flags) ||
4352 ((i == sh->pd_idx || i == sh->qd_idx) &&
4354 set_bit(STRIPE_INSYNC, &sh->state);
4357 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4358 s.dec_preread_active = 1;
4362 * might be able to return some write requests if the parity blocks
4363 * are safe, or on a failed drive
4365 pdev = &sh->dev[sh->pd_idx];
4366 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4367 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4368 qdev = &sh->dev[sh->qd_idx];
4369 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4370 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4374 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4375 && !test_bit(R5_LOCKED, &pdev->flags)
4376 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4377 test_bit(R5_Discard, &pdev->flags))))) &&
4378 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4379 && !test_bit(R5_LOCKED, &qdev->flags)
4380 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4381 test_bit(R5_Discard, &qdev->flags))))))
4382 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4384 /* Now we might consider reading some blocks, either to check/generate
4385 * parity, or to satisfy requests
4386 * or to load a block that is being partially written.
4388 if (s.to_read || s.non_overwrite
4389 || (conf->level == 6 && s.to_write && s.failed)
4390 || (s.syncing && (s.uptodate + s.compute < disks))
4393 handle_stripe_fill(sh, &s, disks);
4395 /* Now to consider new write requests and what else, if anything
4396 * should be read. We do not handle new writes when:
4397 * 1/ A 'write' operation (copy+xor) is already in flight.
4398 * 2/ A 'check' operation is in flight, as it may clobber the parity
4401 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4402 handle_stripe_dirtying(conf, sh, &s, disks);
4404 /* maybe we need to check and possibly fix the parity for this stripe
4405 * Any reads will already have been scheduled, so we just see if enough
4406 * data is available. The parity check is held off while parity
4407 * dependent operations are in flight.
4409 if (sh->check_state ||
4410 (s.syncing && s.locked == 0 &&
4411 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4412 !test_bit(STRIPE_INSYNC, &sh->state))) {
4413 if (conf->level == 6)
4414 handle_parity_checks6(conf, sh, &s, disks);
4416 handle_parity_checks5(conf, sh, &s, disks);
4419 if ((s.replacing || s.syncing) && s.locked == 0
4420 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4421 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4422 /* Write out to replacement devices where possible */
4423 for (i = 0; i < conf->raid_disks; i++)
4424 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4425 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4426 set_bit(R5_WantReplace, &sh->dev[i].flags);
4427 set_bit(R5_LOCKED, &sh->dev[i].flags);
4431 set_bit(STRIPE_INSYNC, &sh->state);
4432 set_bit(STRIPE_REPLACED, &sh->state);
4434 if ((s.syncing || s.replacing) && s.locked == 0 &&
4435 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4436 test_bit(STRIPE_INSYNC, &sh->state)) {
4437 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4438 clear_bit(STRIPE_SYNCING, &sh->state);
4439 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4440 wake_up(&conf->wait_for_overlap);
4443 /* If the failed drives are just a ReadError, then we might need
4444 * to progress the repair/check process
4446 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4447 for (i = 0; i < s.failed; i++) {
4448 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4449 if (test_bit(R5_ReadError, &dev->flags)
4450 && !test_bit(R5_LOCKED, &dev->flags)
4451 && test_bit(R5_UPTODATE, &dev->flags)
4453 if (!test_bit(R5_ReWrite, &dev->flags)) {
4454 set_bit(R5_Wantwrite, &dev->flags);
4455 set_bit(R5_ReWrite, &dev->flags);
4456 set_bit(R5_LOCKED, &dev->flags);
4459 /* let's read it back */
4460 set_bit(R5_Wantread, &dev->flags);
4461 set_bit(R5_LOCKED, &dev->flags);
4467 /* Finish reconstruct operations initiated by the expansion process */
4468 if (sh->reconstruct_state == reconstruct_state_result) {
4469 struct stripe_head *sh_src
4470 = get_active_stripe(conf, sh->sector, 1, 1, 1);
4471 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4472 /* sh cannot be written until sh_src has been read.
4473 * so arrange for sh to be delayed a little
4475 set_bit(STRIPE_DELAYED, &sh->state);
4476 set_bit(STRIPE_HANDLE, &sh->state);
4477 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4479 atomic_inc(&conf->preread_active_stripes);
4480 release_stripe(sh_src);
4484 release_stripe(sh_src);
4486 sh->reconstruct_state = reconstruct_state_idle;
4487 clear_bit(STRIPE_EXPANDING, &sh->state);
4488 for (i = conf->raid_disks; i--; ) {
4489 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4490 set_bit(R5_LOCKED, &sh->dev[i].flags);
4495 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4496 !sh->reconstruct_state) {
4497 /* Need to write out all blocks after computing parity */
4498 sh->disks = conf->raid_disks;
4499 stripe_set_idx(sh->sector, conf, 0, sh);
4500 schedule_reconstruction(sh, &s, 1, 1);
4501 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4502 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4503 atomic_dec(&conf->reshape_stripes);
4504 wake_up(&conf->wait_for_overlap);
4505 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4508 if (s.expanding && s.locked == 0 &&
4509 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4510 handle_stripe_expansion(conf, sh);
4513 /* wait for this device to become unblocked */
4514 if (unlikely(s.blocked_rdev)) {
4515 if (conf->mddev->external)
4516 md_wait_for_blocked_rdev(s.blocked_rdev,
4519 /* Internal metadata will immediately
4520 * be written by raid5d, so we don't
4521 * need to wait here.
4523 rdev_dec_pending(s.blocked_rdev,
4527 if (s.handle_bad_blocks)
4528 for (i = disks; i--; ) {
4529 struct md_rdev *rdev;
4530 struct r5dev *dev = &sh->dev[i];
4531 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4532 /* We own a safe reference to the rdev */
4533 rdev = conf->disks[i].rdev;
4534 if (!rdev_set_badblocks(rdev, sh->sector,
4536 md_error(conf->mddev, rdev);
4537 rdev_dec_pending(rdev, conf->mddev);
4539 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4540 rdev = conf->disks[i].rdev;
4541 rdev_clear_badblocks(rdev, sh->sector,
4543 rdev_dec_pending(rdev, conf->mddev);
4545 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4546 rdev = conf->disks[i].replacement;
4548 /* rdev have been moved down */
4549 rdev = conf->disks[i].rdev;
4550 rdev_clear_badblocks(rdev, sh->sector,
4552 rdev_dec_pending(rdev, conf->mddev);
4557 raid_run_ops(sh, s.ops_request);
4561 if (s.dec_preread_active) {
4562 /* We delay this until after ops_run_io so that if make_request
4563 * is waiting on a flush, it won't continue until the writes
4564 * have actually been submitted.
4566 atomic_dec(&conf->preread_active_stripes);
4567 if (atomic_read(&conf->preread_active_stripes) <
4569 md_wakeup_thread(conf->mddev->thread);
4572 return_io(s.return_bi);
4574 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4577 static void raid5_activate_delayed(struct r5conf *conf)
4579 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4580 while (!list_empty(&conf->delayed_list)) {
4581 struct list_head *l = conf->delayed_list.next;
4582 struct stripe_head *sh;
4583 sh = list_entry(l, struct stripe_head, lru);
4585 clear_bit(STRIPE_DELAYED, &sh->state);
4586 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4587 atomic_inc(&conf->preread_active_stripes);
4588 list_add_tail(&sh->lru, &conf->hold_list);
4589 raid5_wakeup_stripe_thread(sh);
4594 static void activate_bit_delay(struct r5conf *conf,
4595 struct list_head *temp_inactive_list)
4597 /* device_lock is held */
4598 struct list_head head;
4599 list_add(&head, &conf->bitmap_list);
4600 list_del_init(&conf->bitmap_list);
4601 while (!list_empty(&head)) {
4602 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4604 list_del_init(&sh->lru);
4605 atomic_inc(&sh->count);
4606 hash = sh->hash_lock_index;
4607 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4611 static int raid5_congested(struct mddev *mddev, int bits)
4613 struct r5conf *conf = mddev->private;
4615 /* No difference between reads and writes. Just check
4616 * how busy the stripe_cache is
4619 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4623 if (atomic_read(&conf->empty_inactive_list_nr))
4629 /* We want read requests to align with chunks where possible,
4630 * but write requests don't need to.
4632 static int raid5_mergeable_bvec(struct mddev *mddev,
4633 struct bvec_merge_data *bvm,
4634 struct bio_vec *biovec)
4636 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4638 unsigned int chunk_sectors = mddev->chunk_sectors;
4639 unsigned int bio_sectors = bvm->bi_size >> 9;
4642 * always allow writes to be mergeable, read as well if array
4643 * is degraded as we'll go through stripe cache anyway.
4645 if ((bvm->bi_rw & 1) == WRITE || mddev->degraded)
4646 return biovec->bv_len;
4648 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4649 chunk_sectors = mddev->new_chunk_sectors;
4650 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4651 if (max < 0) max = 0;
4652 if (max <= biovec->bv_len && bio_sectors == 0)
4653 return biovec->bv_len;
4658 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4660 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4661 unsigned int chunk_sectors = mddev->chunk_sectors;
4662 unsigned int bio_sectors = bio_sectors(bio);
4664 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4665 chunk_sectors = mddev->new_chunk_sectors;
4666 return chunk_sectors >=
4667 ((sector & (chunk_sectors - 1)) + bio_sectors);
4671 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4672 * later sampled by raid5d.
4674 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4676 unsigned long flags;
4678 spin_lock_irqsave(&conf->device_lock, flags);
4680 bi->bi_next = conf->retry_read_aligned_list;
4681 conf->retry_read_aligned_list = bi;
4683 spin_unlock_irqrestore(&conf->device_lock, flags);
4684 md_wakeup_thread(conf->mddev->thread);
4687 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4691 bi = conf->retry_read_aligned;
4693 conf->retry_read_aligned = NULL;
4696 bi = conf->retry_read_aligned_list;
4698 conf->retry_read_aligned_list = bi->bi_next;
4701 * this sets the active strip count to 1 and the processed
4702 * strip count to zero (upper 8 bits)
4704 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4711 * The "raid5_align_endio" should check if the read succeeded and if it
4712 * did, call bio_endio on the original bio (having bio_put the new bio
4714 * If the read failed..
4716 static void raid5_align_endio(struct bio *bi, int error)
4718 struct bio* raid_bi = bi->bi_private;
4719 struct mddev *mddev;
4720 struct r5conf *conf;
4721 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4722 struct md_rdev *rdev;
4726 rdev = (void*)raid_bi->bi_next;
4727 raid_bi->bi_next = NULL;
4728 mddev = rdev->mddev;
4729 conf = mddev->private;
4731 rdev_dec_pending(rdev, conf->mddev);
4733 if (!error && uptodate) {
4734 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4736 bio_endio(raid_bi, 0);
4737 if (atomic_dec_and_test(&conf->active_aligned_reads))
4738 wake_up(&conf->wait_for_stripe);
4742 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4744 add_bio_to_retry(raid_bi, conf);
4747 static int bio_fits_rdev(struct bio *bi)
4749 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4751 if (bio_sectors(bi) > queue_max_sectors(q))
4753 blk_recount_segments(q, bi);
4754 if (bi->bi_phys_segments > queue_max_segments(q))
4757 if (q->merge_bvec_fn)
4758 /* it's too hard to apply the merge_bvec_fn at this stage,
4766 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4768 struct r5conf *conf = mddev->private;
4770 struct bio* align_bi;
4771 struct md_rdev *rdev;
4772 sector_t end_sector;
4774 if (!in_chunk_boundary(mddev, raid_bio)) {
4775 pr_debug("chunk_aligned_read : non aligned\n");
4779 * use bio_clone_mddev to make a copy of the bio
4781 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4785 * set bi_end_io to a new function, and set bi_private to the
4788 align_bi->bi_end_io = raid5_align_endio;
4789 align_bi->bi_private = raid_bio;
4793 align_bi->bi_iter.bi_sector =
4794 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4797 end_sector = bio_end_sector(align_bi);
4799 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4800 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4801 rdev->recovery_offset < end_sector) {
4802 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4804 (test_bit(Faulty, &rdev->flags) ||
4805 !(test_bit(In_sync, &rdev->flags) ||
4806 rdev->recovery_offset >= end_sector)))
4813 atomic_inc(&rdev->nr_pending);
4815 raid_bio->bi_next = (void*)rdev;
4816 align_bi->bi_bdev = rdev->bdev;
4817 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
4819 if (!bio_fits_rdev(align_bi) ||
4820 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4821 bio_sectors(align_bi),
4822 &first_bad, &bad_sectors)) {
4823 /* too big in some way, or has a known bad block */
4825 rdev_dec_pending(rdev, mddev);
4829 /* No reshape active, so we can trust rdev->data_offset */
4830 align_bi->bi_iter.bi_sector += rdev->data_offset;
4832 spin_lock_irq(&conf->device_lock);
4833 wait_event_lock_irq(conf->wait_for_stripe,
4836 atomic_inc(&conf->active_aligned_reads);
4837 spin_unlock_irq(&conf->device_lock);
4840 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4841 align_bi, disk_devt(mddev->gendisk),
4842 raid_bio->bi_iter.bi_sector);
4843 generic_make_request(align_bi);
4852 /* __get_priority_stripe - get the next stripe to process
4854 * Full stripe writes are allowed to pass preread active stripes up until
4855 * the bypass_threshold is exceeded. In general the bypass_count
4856 * increments when the handle_list is handled before the hold_list; however, it
4857 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4858 * stripe with in flight i/o. The bypass_count will be reset when the
4859 * head of the hold_list has changed, i.e. the head was promoted to the
4862 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4864 struct stripe_head *sh = NULL, *tmp;
4865 struct list_head *handle_list = NULL;
4866 struct r5worker_group *wg = NULL;
4868 if (conf->worker_cnt_per_group == 0) {
4869 handle_list = &conf->handle_list;
4870 } else if (group != ANY_GROUP) {
4871 handle_list = &conf->worker_groups[group].handle_list;
4872 wg = &conf->worker_groups[group];
4875 for (i = 0; i < conf->group_cnt; i++) {
4876 handle_list = &conf->worker_groups[i].handle_list;
4877 wg = &conf->worker_groups[i];
4878 if (!list_empty(handle_list))
4883 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4885 list_empty(handle_list) ? "empty" : "busy",
4886 list_empty(&conf->hold_list) ? "empty" : "busy",
4887 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4889 if (!list_empty(handle_list)) {
4890 sh = list_entry(handle_list->next, typeof(*sh), lru);
4892 if (list_empty(&conf->hold_list))
4893 conf->bypass_count = 0;
4894 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4895 if (conf->hold_list.next == conf->last_hold)
4896 conf->bypass_count++;
4898 conf->last_hold = conf->hold_list.next;
4899 conf->bypass_count -= conf->bypass_threshold;
4900 if (conf->bypass_count < 0)
4901 conf->bypass_count = 0;
4904 } else if (!list_empty(&conf->hold_list) &&
4905 ((conf->bypass_threshold &&
4906 conf->bypass_count > conf->bypass_threshold) ||
4907 atomic_read(&conf->pending_full_writes) == 0)) {
4909 list_for_each_entry(tmp, &conf->hold_list, lru) {
4910 if (conf->worker_cnt_per_group == 0 ||
4911 group == ANY_GROUP ||
4912 !cpu_online(tmp->cpu) ||
4913 cpu_to_group(tmp->cpu) == group) {
4920 conf->bypass_count -= conf->bypass_threshold;
4921 if (conf->bypass_count < 0)
4922 conf->bypass_count = 0;
4934 list_del_init(&sh->lru);
4935 BUG_ON(atomic_inc_return(&sh->count) != 1);
4939 struct raid5_plug_cb {
4940 struct blk_plug_cb cb;
4941 struct list_head list;
4942 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4945 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4947 struct raid5_plug_cb *cb = container_of(
4948 blk_cb, struct raid5_plug_cb, cb);
4949 struct stripe_head *sh;
4950 struct mddev *mddev = cb->cb.data;
4951 struct r5conf *conf = mddev->private;
4955 if (cb->list.next && !list_empty(&cb->list)) {
4956 spin_lock_irq(&conf->device_lock);
4957 while (!list_empty(&cb->list)) {
4958 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4959 list_del_init(&sh->lru);
4961 * avoid race release_stripe_plug() sees
4962 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4963 * is still in our list
4965 smp_mb__before_atomic();
4966 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4968 * STRIPE_ON_RELEASE_LIST could be set here. In that
4969 * case, the count is always > 1 here
4971 hash = sh->hash_lock_index;
4972 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
4975 spin_unlock_irq(&conf->device_lock);
4977 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4978 NR_STRIPE_HASH_LOCKS);
4980 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4984 static void release_stripe_plug(struct mddev *mddev,
4985 struct stripe_head *sh)
4987 struct blk_plug_cb *blk_cb = blk_check_plugged(
4988 raid5_unplug, mddev,
4989 sizeof(struct raid5_plug_cb));
4990 struct raid5_plug_cb *cb;
4997 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4999 if (cb->list.next == NULL) {
5001 INIT_LIST_HEAD(&cb->list);
5002 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5003 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5006 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5007 list_add_tail(&sh->lru, &cb->list);
5012 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5014 struct r5conf *conf = mddev->private;
5015 sector_t logical_sector, last_sector;
5016 struct stripe_head *sh;
5020 if (mddev->reshape_position != MaxSector)
5021 /* Skip discard while reshape is happening */
5024 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5025 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5028 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5030 stripe_sectors = conf->chunk_sectors *
5031 (conf->raid_disks - conf->max_degraded);
5032 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5034 sector_div(last_sector, stripe_sectors);
5036 logical_sector *= conf->chunk_sectors;
5037 last_sector *= conf->chunk_sectors;
5039 for (; logical_sector < last_sector;
5040 logical_sector += STRIPE_SECTORS) {
5044 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
5045 prepare_to_wait(&conf->wait_for_overlap, &w,
5046 TASK_UNINTERRUPTIBLE);
5047 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5048 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5053 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5054 spin_lock_irq(&sh->stripe_lock);
5055 for (d = 0; d < conf->raid_disks; d++) {
5056 if (d == sh->pd_idx || d == sh->qd_idx)
5058 if (sh->dev[d].towrite || sh->dev[d].toread) {
5059 set_bit(R5_Overlap, &sh->dev[d].flags);
5060 spin_unlock_irq(&sh->stripe_lock);
5066 set_bit(STRIPE_DISCARD, &sh->state);
5067 finish_wait(&conf->wait_for_overlap, &w);
5068 sh->overwrite_disks = 0;
5069 for (d = 0; d < conf->raid_disks; d++) {
5070 if (d == sh->pd_idx || d == sh->qd_idx)
5072 sh->dev[d].towrite = bi;
5073 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5074 raid5_inc_bi_active_stripes(bi);
5075 sh->overwrite_disks++;
5077 spin_unlock_irq(&sh->stripe_lock);
5078 if (conf->mddev->bitmap) {
5080 d < conf->raid_disks - conf->max_degraded;
5082 bitmap_startwrite(mddev->bitmap,
5086 sh->bm_seq = conf->seq_flush + 1;
5087 set_bit(STRIPE_BIT_DELAY, &sh->state);
5090 set_bit(STRIPE_HANDLE, &sh->state);
5091 clear_bit(STRIPE_DELAYED, &sh->state);
5092 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5093 atomic_inc(&conf->preread_active_stripes);
5094 release_stripe_plug(mddev, sh);
5097 remaining = raid5_dec_bi_active_stripes(bi);
5098 if (remaining == 0) {
5099 md_write_end(mddev);
5104 static void make_request(struct mddev *mddev, struct bio * bi)
5106 struct r5conf *conf = mddev->private;
5108 sector_t new_sector;
5109 sector_t logical_sector, last_sector;
5110 struct stripe_head *sh;
5111 const int rw = bio_data_dir(bi);
5116 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5117 md_flush_request(mddev, bi);
5121 md_write_start(mddev, bi);
5124 * If array is degraded, better not do chunk aligned read because
5125 * later we might have to read it again in order to reconstruct
5126 * data on failed drives.
5128 if (rw == READ && mddev->degraded == 0 &&
5129 mddev->reshape_position == MaxSector &&
5130 chunk_aligned_read(mddev,bi))
5133 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5134 make_discard_request(mddev, bi);
5138 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5139 last_sector = bio_end_sector(bi);
5141 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5143 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5144 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5150 seq = read_seqcount_begin(&conf->gen_lock);
5153 prepare_to_wait(&conf->wait_for_overlap, &w,
5154 TASK_UNINTERRUPTIBLE);
5155 if (unlikely(conf->reshape_progress != MaxSector)) {
5156 /* spinlock is needed as reshape_progress may be
5157 * 64bit on a 32bit platform, and so it might be
5158 * possible to see a half-updated value
5159 * Of course reshape_progress could change after
5160 * the lock is dropped, so once we get a reference
5161 * to the stripe that we think it is, we will have
5164 spin_lock_irq(&conf->device_lock);
5165 if (mddev->reshape_backwards
5166 ? logical_sector < conf->reshape_progress
5167 : logical_sector >= conf->reshape_progress) {
5170 if (mddev->reshape_backwards
5171 ? logical_sector < conf->reshape_safe
5172 : logical_sector >= conf->reshape_safe) {
5173 spin_unlock_irq(&conf->device_lock);
5179 spin_unlock_irq(&conf->device_lock);
5182 new_sector = raid5_compute_sector(conf, logical_sector,
5185 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5186 (unsigned long long)new_sector,
5187 (unsigned long long)logical_sector);
5189 sh = get_active_stripe(conf, new_sector, previous,
5190 (bi->bi_rw&RWA_MASK), 0);
5192 if (unlikely(previous)) {
5193 /* expansion might have moved on while waiting for a
5194 * stripe, so we must do the range check again.
5195 * Expansion could still move past after this
5196 * test, but as we are holding a reference to
5197 * 'sh', we know that if that happens,
5198 * STRIPE_EXPANDING will get set and the expansion
5199 * won't proceed until we finish with the stripe.
5202 spin_lock_irq(&conf->device_lock);
5203 if (mddev->reshape_backwards
5204 ? logical_sector >= conf->reshape_progress
5205 : logical_sector < conf->reshape_progress)
5206 /* mismatch, need to try again */
5208 spin_unlock_irq(&conf->device_lock);
5216 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5217 /* Might have got the wrong stripe_head
5225 logical_sector >= mddev->suspend_lo &&
5226 logical_sector < mddev->suspend_hi) {
5228 /* As the suspend_* range is controlled by
5229 * userspace, we want an interruptible
5232 flush_signals(current);
5233 prepare_to_wait(&conf->wait_for_overlap,
5234 &w, TASK_INTERRUPTIBLE);
5235 if (logical_sector >= mddev->suspend_lo &&
5236 logical_sector < mddev->suspend_hi) {
5243 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5244 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5245 /* Stripe is busy expanding or
5246 * add failed due to overlap. Flush everything
5249 md_wakeup_thread(mddev->thread);
5255 set_bit(STRIPE_HANDLE, &sh->state);
5256 clear_bit(STRIPE_DELAYED, &sh->state);
5257 if ((!sh->batch_head || sh == sh->batch_head) &&
5258 (bi->bi_rw & REQ_SYNC) &&
5259 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5260 atomic_inc(&conf->preread_active_stripes);
5261 release_stripe_plug(mddev, sh);
5263 /* cannot get stripe for read-ahead, just give-up */
5264 clear_bit(BIO_UPTODATE, &bi->bi_flags);
5268 finish_wait(&conf->wait_for_overlap, &w);
5270 remaining = raid5_dec_bi_active_stripes(bi);
5271 if (remaining == 0) {
5274 md_write_end(mddev);
5276 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5282 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5284 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5286 /* reshaping is quite different to recovery/resync so it is
5287 * handled quite separately ... here.
5289 * On each call to sync_request, we gather one chunk worth of
5290 * destination stripes and flag them as expanding.
5291 * Then we find all the source stripes and request reads.
5292 * As the reads complete, handle_stripe will copy the data
5293 * into the destination stripe and release that stripe.
5295 struct r5conf *conf = mddev->private;
5296 struct stripe_head *sh;
5297 sector_t first_sector, last_sector;
5298 int raid_disks = conf->previous_raid_disks;
5299 int data_disks = raid_disks - conf->max_degraded;
5300 int new_data_disks = conf->raid_disks - conf->max_degraded;
5303 sector_t writepos, readpos, safepos;
5304 sector_t stripe_addr;
5305 int reshape_sectors;
5306 struct list_head stripes;
5308 if (sector_nr == 0) {
5309 /* If restarting in the middle, skip the initial sectors */
5310 if (mddev->reshape_backwards &&
5311 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5312 sector_nr = raid5_size(mddev, 0, 0)
5313 - conf->reshape_progress;
5314 } else if (!mddev->reshape_backwards &&
5315 conf->reshape_progress > 0)
5316 sector_nr = conf->reshape_progress;
5317 sector_div(sector_nr, new_data_disks);
5319 mddev->curr_resync_completed = sector_nr;
5320 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5326 /* We need to process a full chunk at a time.
5327 * If old and new chunk sizes differ, we need to process the
5330 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
5331 reshape_sectors = mddev->new_chunk_sectors;
5333 reshape_sectors = mddev->chunk_sectors;
5335 /* We update the metadata at least every 10 seconds, or when
5336 * the data about to be copied would over-write the source of
5337 * the data at the front of the range. i.e. one new_stripe
5338 * along from reshape_progress new_maps to after where
5339 * reshape_safe old_maps to
5341 writepos = conf->reshape_progress;
5342 sector_div(writepos, new_data_disks);
5343 readpos = conf->reshape_progress;
5344 sector_div(readpos, data_disks);
5345 safepos = conf->reshape_safe;
5346 sector_div(safepos, data_disks);
5347 if (mddev->reshape_backwards) {
5348 writepos -= min_t(sector_t, reshape_sectors, writepos);
5349 readpos += reshape_sectors;
5350 safepos += reshape_sectors;
5352 writepos += reshape_sectors;
5353 readpos -= min_t(sector_t, reshape_sectors, readpos);
5354 safepos -= min_t(sector_t, reshape_sectors, safepos);
5357 /* Having calculated the 'writepos' possibly use it
5358 * to set 'stripe_addr' which is where we will write to.
5360 if (mddev->reshape_backwards) {
5361 BUG_ON(conf->reshape_progress == 0);
5362 stripe_addr = writepos;
5363 BUG_ON((mddev->dev_sectors &
5364 ~((sector_t)reshape_sectors - 1))
5365 - reshape_sectors - stripe_addr
5368 BUG_ON(writepos != sector_nr + reshape_sectors);
5369 stripe_addr = sector_nr;
5372 /* 'writepos' is the most advanced device address we might write.
5373 * 'readpos' is the least advanced device address we might read.
5374 * 'safepos' is the least address recorded in the metadata as having
5376 * If there is a min_offset_diff, these are adjusted either by
5377 * increasing the safepos/readpos if diff is negative, or
5378 * increasing writepos if diff is positive.
5379 * If 'readpos' is then behind 'writepos', there is no way that we can
5380 * ensure safety in the face of a crash - that must be done by userspace
5381 * making a backup of the data. So in that case there is no particular
5382 * rush to update metadata.
5383 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5384 * update the metadata to advance 'safepos' to match 'readpos' so that
5385 * we can be safe in the event of a crash.
5386 * So we insist on updating metadata if safepos is behind writepos and
5387 * readpos is beyond writepos.
5388 * In any case, update the metadata every 10 seconds.
5389 * Maybe that number should be configurable, but I'm not sure it is
5390 * worth it.... maybe it could be a multiple of safemode_delay???
5392 if (conf->min_offset_diff < 0) {
5393 safepos += -conf->min_offset_diff;
5394 readpos += -conf->min_offset_diff;
5396 writepos += conf->min_offset_diff;
5398 if ((mddev->reshape_backwards
5399 ? (safepos > writepos && readpos < writepos)
5400 : (safepos < writepos && readpos > writepos)) ||
5401 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5402 /* Cannot proceed until we've updated the superblock... */
5403 wait_event(conf->wait_for_overlap,
5404 atomic_read(&conf->reshape_stripes)==0
5405 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5406 if (atomic_read(&conf->reshape_stripes) != 0)
5408 mddev->reshape_position = conf->reshape_progress;
5409 mddev->curr_resync_completed = sector_nr;
5410 conf->reshape_checkpoint = jiffies;
5411 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5412 md_wakeup_thread(mddev->thread);
5413 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5414 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5415 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5417 spin_lock_irq(&conf->device_lock);
5418 conf->reshape_safe = mddev->reshape_position;
5419 spin_unlock_irq(&conf->device_lock);
5420 wake_up(&conf->wait_for_overlap);
5421 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5424 INIT_LIST_HEAD(&stripes);
5425 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5427 int skipped_disk = 0;
5428 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5429 set_bit(STRIPE_EXPANDING, &sh->state);
5430 atomic_inc(&conf->reshape_stripes);
5431 /* If any of this stripe is beyond the end of the old
5432 * array, then we need to zero those blocks
5434 for (j=sh->disks; j--;) {
5436 if (j == sh->pd_idx)
5438 if (conf->level == 6 &&
5441 s = compute_blocknr(sh, j, 0);
5442 if (s < raid5_size(mddev, 0, 0)) {
5446 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5447 set_bit(R5_Expanded, &sh->dev[j].flags);
5448 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5450 if (!skipped_disk) {
5451 set_bit(STRIPE_EXPAND_READY, &sh->state);
5452 set_bit(STRIPE_HANDLE, &sh->state);
5454 list_add(&sh->lru, &stripes);
5456 spin_lock_irq(&conf->device_lock);
5457 if (mddev->reshape_backwards)
5458 conf->reshape_progress -= reshape_sectors * new_data_disks;
5460 conf->reshape_progress += reshape_sectors * new_data_disks;
5461 spin_unlock_irq(&conf->device_lock);
5462 /* Ok, those stripe are ready. We can start scheduling
5463 * reads on the source stripes.
5464 * The source stripes are determined by mapping the first and last
5465 * block on the destination stripes.
5468 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5471 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5472 * new_data_disks - 1),
5474 if (last_sector >= mddev->dev_sectors)
5475 last_sector = mddev->dev_sectors - 1;
5476 while (first_sector <= last_sector) {
5477 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
5478 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5479 set_bit(STRIPE_HANDLE, &sh->state);
5481 first_sector += STRIPE_SECTORS;
5483 /* Now that the sources are clearly marked, we can release
5484 * the destination stripes
5486 while (!list_empty(&stripes)) {
5487 sh = list_entry(stripes.next, struct stripe_head, lru);
5488 list_del_init(&sh->lru);
5491 /* If this takes us to the resync_max point where we have to pause,
5492 * then we need to write out the superblock.
5494 sector_nr += reshape_sectors;
5495 if ((sector_nr - mddev->curr_resync_completed) * 2
5496 >= mddev->resync_max - mddev->curr_resync_completed) {
5497 /* Cannot proceed until we've updated the superblock... */
5498 wait_event(conf->wait_for_overlap,
5499 atomic_read(&conf->reshape_stripes) == 0
5500 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5501 if (atomic_read(&conf->reshape_stripes) != 0)
5503 mddev->reshape_position = conf->reshape_progress;
5504 mddev->curr_resync_completed = sector_nr;
5505 conf->reshape_checkpoint = jiffies;
5506 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5507 md_wakeup_thread(mddev->thread);
5508 wait_event(mddev->sb_wait,
5509 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5510 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5511 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5513 spin_lock_irq(&conf->device_lock);
5514 conf->reshape_safe = mddev->reshape_position;
5515 spin_unlock_irq(&conf->device_lock);
5516 wake_up(&conf->wait_for_overlap);
5517 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5520 return reshape_sectors;
5523 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5525 struct r5conf *conf = mddev->private;
5526 struct stripe_head *sh;
5527 sector_t max_sector = mddev->dev_sectors;
5528 sector_t sync_blocks;
5529 int still_degraded = 0;
5532 if (sector_nr >= max_sector) {
5533 /* just being told to finish up .. nothing much to do */
5535 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5540 if (mddev->curr_resync < max_sector) /* aborted */
5541 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5543 else /* completed sync */
5545 bitmap_close_sync(mddev->bitmap);
5550 /* Allow raid5_quiesce to complete */
5551 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5553 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5554 return reshape_request(mddev, sector_nr, skipped);
5556 /* No need to check resync_max as we never do more than one
5557 * stripe, and as resync_max will always be on a chunk boundary,
5558 * if the check in md_do_sync didn't fire, there is no chance
5559 * of overstepping resync_max here
5562 /* if there is too many failed drives and we are trying
5563 * to resync, then assert that we are finished, because there is
5564 * nothing we can do.
5566 if (mddev->degraded >= conf->max_degraded &&
5567 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5568 sector_t rv = mddev->dev_sectors - sector_nr;
5572 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5574 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5575 sync_blocks >= STRIPE_SECTORS) {
5576 /* we can skip this block, and probably more */
5577 sync_blocks /= STRIPE_SECTORS;
5579 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5582 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5584 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5586 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5587 /* make sure we don't swamp the stripe cache if someone else
5588 * is trying to get access
5590 schedule_timeout_uninterruptible(1);
5592 /* Need to check if array will still be degraded after recovery/resync
5593 * Note in case of > 1 drive failures it's possible we're rebuilding
5594 * one drive while leaving another faulty drive in array.
5597 for (i = 0; i < conf->raid_disks; i++) {
5598 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5600 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5605 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5607 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5608 set_bit(STRIPE_HANDLE, &sh->state);
5612 return STRIPE_SECTORS;
5615 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5617 /* We may not be able to submit a whole bio at once as there
5618 * may not be enough stripe_heads available.
5619 * We cannot pre-allocate enough stripe_heads as we may need
5620 * more than exist in the cache (if we allow ever large chunks).
5621 * So we do one stripe head at a time and record in
5622 * ->bi_hw_segments how many have been done.
5624 * We *know* that this entire raid_bio is in one chunk, so
5625 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5627 struct stripe_head *sh;
5629 sector_t sector, logical_sector, last_sector;
5634 logical_sector = raid_bio->bi_iter.bi_sector &
5635 ~((sector_t)STRIPE_SECTORS-1);
5636 sector = raid5_compute_sector(conf, logical_sector,
5638 last_sector = bio_end_sector(raid_bio);
5640 for (; logical_sector < last_sector;
5641 logical_sector += STRIPE_SECTORS,
5642 sector += STRIPE_SECTORS,
5645 if (scnt < raid5_bi_processed_stripes(raid_bio))
5646 /* already done this stripe */
5649 sh = get_active_stripe(conf, sector, 0, 1, 1);
5652 /* failed to get a stripe - must wait */
5653 raid5_set_bi_processed_stripes(raid_bio, scnt);
5654 conf->retry_read_aligned = raid_bio;
5658 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5660 raid5_set_bi_processed_stripes(raid_bio, scnt);
5661 conf->retry_read_aligned = raid_bio;
5665 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5670 remaining = raid5_dec_bi_active_stripes(raid_bio);
5671 if (remaining == 0) {
5672 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5674 bio_endio(raid_bio, 0);
5676 if (atomic_dec_and_test(&conf->active_aligned_reads))
5677 wake_up(&conf->wait_for_stripe);
5681 static int handle_active_stripes(struct r5conf *conf, int group,
5682 struct r5worker *worker,
5683 struct list_head *temp_inactive_list)
5685 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5686 int i, batch_size = 0, hash;
5687 bool release_inactive = false;
5689 while (batch_size < MAX_STRIPE_BATCH &&
5690 (sh = __get_priority_stripe(conf, group)) != NULL)
5691 batch[batch_size++] = sh;
5693 if (batch_size == 0) {
5694 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5695 if (!list_empty(temp_inactive_list + i))
5697 if (i == NR_STRIPE_HASH_LOCKS)
5699 release_inactive = true;
5701 spin_unlock_irq(&conf->device_lock);
5703 release_inactive_stripe_list(conf, temp_inactive_list,
5704 NR_STRIPE_HASH_LOCKS);
5706 if (release_inactive) {
5707 spin_lock_irq(&conf->device_lock);
5711 for (i = 0; i < batch_size; i++)
5712 handle_stripe(batch[i]);
5716 spin_lock_irq(&conf->device_lock);
5717 for (i = 0; i < batch_size; i++) {
5718 hash = batch[i]->hash_lock_index;
5719 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5724 static void raid5_do_work(struct work_struct *work)
5726 struct r5worker *worker = container_of(work, struct r5worker, work);
5727 struct r5worker_group *group = worker->group;
5728 struct r5conf *conf = group->conf;
5729 int group_id = group - conf->worker_groups;
5731 struct blk_plug plug;
5733 pr_debug("+++ raid5worker active\n");
5735 blk_start_plug(&plug);
5737 spin_lock_irq(&conf->device_lock);
5739 int batch_size, released;
5741 released = release_stripe_list(conf, worker->temp_inactive_list);
5743 batch_size = handle_active_stripes(conf, group_id, worker,
5744 worker->temp_inactive_list);
5745 worker->working = false;
5746 if (!batch_size && !released)
5748 handled += batch_size;
5750 pr_debug("%d stripes handled\n", handled);
5752 spin_unlock_irq(&conf->device_lock);
5753 blk_finish_plug(&plug);
5755 pr_debug("--- raid5worker inactive\n");
5759 * This is our raid5 kernel thread.
5761 * We scan the hash table for stripes which can be handled now.
5762 * During the scan, completed stripes are saved for us by the interrupt
5763 * handler, so that they will not have to wait for our next wakeup.
5765 static void raid5d(struct md_thread *thread)
5767 struct mddev *mddev = thread->mddev;
5768 struct r5conf *conf = mddev->private;
5770 struct blk_plug plug;
5772 pr_debug("+++ raid5d active\n");
5774 md_check_recovery(mddev);
5776 blk_start_plug(&plug);
5778 spin_lock_irq(&conf->device_lock);
5781 int batch_size, released;
5783 released = release_stripe_list(conf, conf->temp_inactive_list);
5785 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5788 !list_empty(&conf->bitmap_list)) {
5789 /* Now is a good time to flush some bitmap updates */
5791 spin_unlock_irq(&conf->device_lock);
5792 bitmap_unplug(mddev->bitmap);
5793 spin_lock_irq(&conf->device_lock);
5794 conf->seq_write = conf->seq_flush;
5795 activate_bit_delay(conf, conf->temp_inactive_list);
5797 raid5_activate_delayed(conf);
5799 while ((bio = remove_bio_from_retry(conf))) {
5801 spin_unlock_irq(&conf->device_lock);
5802 ok = retry_aligned_read(conf, bio);
5803 spin_lock_irq(&conf->device_lock);
5809 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5810 conf->temp_inactive_list);
5811 if (!batch_size && !released)
5813 handled += batch_size;
5815 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5816 spin_unlock_irq(&conf->device_lock);
5817 md_check_recovery(mddev);
5818 spin_lock_irq(&conf->device_lock);
5821 pr_debug("%d stripes handled\n", handled);
5823 spin_unlock_irq(&conf->device_lock);
5824 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state)) {
5825 grow_one_stripe(conf, __GFP_NOWARN);
5826 /* Set flag even if allocation failed. This helps
5827 * slow down allocation requests when mem is short
5829 set_bit(R5_DID_ALLOC, &conf->cache_state);
5832 async_tx_issue_pending_all();
5833 blk_finish_plug(&plug);
5835 pr_debug("--- raid5d inactive\n");
5839 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5841 struct r5conf *conf;
5843 spin_lock(&mddev->lock);
5844 conf = mddev->private;
5846 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5847 spin_unlock(&mddev->lock);
5852 raid5_set_cache_size(struct mddev *mddev, int size)
5854 struct r5conf *conf = mddev->private;
5857 if (size <= 16 || size > 32768)
5860 conf->min_nr_stripes = size;
5861 while (size < conf->max_nr_stripes &&
5862 drop_one_stripe(conf))
5866 err = md_allow_write(mddev);
5870 while (size > conf->max_nr_stripes)
5871 if (!grow_one_stripe(conf, GFP_KERNEL))
5876 EXPORT_SYMBOL(raid5_set_cache_size);
5879 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5881 struct r5conf *conf;
5885 if (len >= PAGE_SIZE)
5887 if (kstrtoul(page, 10, &new))
5889 err = mddev_lock(mddev);
5892 conf = mddev->private;
5896 err = raid5_set_cache_size(mddev, new);
5897 mddev_unlock(mddev);
5902 static struct md_sysfs_entry
5903 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5904 raid5_show_stripe_cache_size,
5905 raid5_store_stripe_cache_size);
5908 raid5_show_rmw_level(struct mddev *mddev, char *page)
5910 struct r5conf *conf = mddev->private;
5912 return sprintf(page, "%d\n", conf->rmw_level);
5918 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
5920 struct r5conf *conf = mddev->private;
5926 if (len >= PAGE_SIZE)
5929 if (kstrtoul(page, 10, &new))
5932 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
5935 if (new != PARITY_DISABLE_RMW &&
5936 new != PARITY_ENABLE_RMW &&
5937 new != PARITY_PREFER_RMW)
5940 conf->rmw_level = new;
5944 static struct md_sysfs_entry
5945 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
5946 raid5_show_rmw_level,
5947 raid5_store_rmw_level);
5951 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5953 struct r5conf *conf;
5955 spin_lock(&mddev->lock);
5956 conf = mddev->private;
5958 ret = sprintf(page, "%d\n", conf->bypass_threshold);
5959 spin_unlock(&mddev->lock);
5964 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
5966 struct r5conf *conf;
5970 if (len >= PAGE_SIZE)
5972 if (kstrtoul(page, 10, &new))
5975 err = mddev_lock(mddev);
5978 conf = mddev->private;
5981 else if (new > conf->min_nr_stripes)
5984 conf->bypass_threshold = new;
5985 mddev_unlock(mddev);
5989 static struct md_sysfs_entry
5990 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5992 raid5_show_preread_threshold,
5993 raid5_store_preread_threshold);
5996 raid5_show_skip_copy(struct mddev *mddev, char *page)
5998 struct r5conf *conf;
6000 spin_lock(&mddev->lock);
6001 conf = mddev->private;
6003 ret = sprintf(page, "%d\n", conf->skip_copy);
6004 spin_unlock(&mddev->lock);
6009 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6011 struct r5conf *conf;
6015 if (len >= PAGE_SIZE)
6017 if (kstrtoul(page, 10, &new))
6021 err = mddev_lock(mddev);
6024 conf = mddev->private;
6027 else if (new != conf->skip_copy) {
6028 mddev_suspend(mddev);
6029 conf->skip_copy = new;
6031 mddev->queue->backing_dev_info.capabilities |=
6032 BDI_CAP_STABLE_WRITES;
6034 mddev->queue->backing_dev_info.capabilities &=
6035 ~BDI_CAP_STABLE_WRITES;
6036 mddev_resume(mddev);
6038 mddev_unlock(mddev);
6042 static struct md_sysfs_entry
6043 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6044 raid5_show_skip_copy,
6045 raid5_store_skip_copy);
6048 stripe_cache_active_show(struct mddev *mddev, char *page)
6050 struct r5conf *conf = mddev->private;
6052 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6057 static struct md_sysfs_entry
6058 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6061 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6063 struct r5conf *conf;
6065 spin_lock(&mddev->lock);
6066 conf = mddev->private;
6068 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6069 spin_unlock(&mddev->lock);
6073 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6075 int *worker_cnt_per_group,
6076 struct r5worker_group **worker_groups);
6078 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6080 struct r5conf *conf;
6083 struct r5worker_group *new_groups, *old_groups;
6084 int group_cnt, worker_cnt_per_group;
6086 if (len >= PAGE_SIZE)
6088 if (kstrtoul(page, 10, &new))
6091 err = mddev_lock(mddev);
6094 conf = mddev->private;
6097 else if (new != conf->worker_cnt_per_group) {
6098 mddev_suspend(mddev);
6100 old_groups = conf->worker_groups;
6102 flush_workqueue(raid5_wq);
6104 err = alloc_thread_groups(conf, new,
6105 &group_cnt, &worker_cnt_per_group,
6108 spin_lock_irq(&conf->device_lock);
6109 conf->group_cnt = group_cnt;
6110 conf->worker_cnt_per_group = worker_cnt_per_group;
6111 conf->worker_groups = new_groups;
6112 spin_unlock_irq(&conf->device_lock);
6115 kfree(old_groups[0].workers);
6118 mddev_resume(mddev);
6120 mddev_unlock(mddev);
6125 static struct md_sysfs_entry
6126 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6127 raid5_show_group_thread_cnt,
6128 raid5_store_group_thread_cnt);
6130 static struct attribute *raid5_attrs[] = {
6131 &raid5_stripecache_size.attr,
6132 &raid5_stripecache_active.attr,
6133 &raid5_preread_bypass_threshold.attr,
6134 &raid5_group_thread_cnt.attr,
6135 &raid5_skip_copy.attr,
6136 &raid5_rmw_level.attr,
6139 static struct attribute_group raid5_attrs_group = {
6141 .attrs = raid5_attrs,
6144 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6146 int *worker_cnt_per_group,
6147 struct r5worker_group **worker_groups)
6151 struct r5worker *workers;
6153 *worker_cnt_per_group = cnt;
6156 *worker_groups = NULL;
6159 *group_cnt = num_possible_nodes();
6160 size = sizeof(struct r5worker) * cnt;
6161 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6162 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6163 *group_cnt, GFP_NOIO);
6164 if (!*worker_groups || !workers) {
6166 kfree(*worker_groups);
6170 for (i = 0; i < *group_cnt; i++) {
6171 struct r5worker_group *group;
6173 group = &(*worker_groups)[i];
6174 INIT_LIST_HEAD(&group->handle_list);
6176 group->workers = workers + i * cnt;
6178 for (j = 0; j < cnt; j++) {
6179 struct r5worker *worker = group->workers + j;
6180 worker->group = group;
6181 INIT_WORK(&worker->work, raid5_do_work);
6183 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6184 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6191 static void free_thread_groups(struct r5conf *conf)
6193 if (conf->worker_groups)
6194 kfree(conf->worker_groups[0].workers);
6195 kfree(conf->worker_groups);
6196 conf->worker_groups = NULL;
6200 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6202 struct r5conf *conf = mddev->private;
6205 sectors = mddev->dev_sectors;
6207 /* size is defined by the smallest of previous and new size */
6208 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6210 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6211 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
6212 return sectors * (raid_disks - conf->max_degraded);
6215 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6217 safe_put_page(percpu->spare_page);
6218 if (percpu->scribble)
6219 flex_array_free(percpu->scribble);
6220 percpu->spare_page = NULL;
6221 percpu->scribble = NULL;
6224 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6226 if (conf->level == 6 && !percpu->spare_page)
6227 percpu->spare_page = alloc_page(GFP_KERNEL);
6228 if (!percpu->scribble)
6229 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6230 conf->previous_raid_disks), conf->chunk_sectors /
6231 STRIPE_SECTORS, GFP_KERNEL);
6233 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6234 free_scratch_buffer(conf, percpu);
6241 static void raid5_free_percpu(struct r5conf *conf)
6248 #ifdef CONFIG_HOTPLUG_CPU
6249 unregister_cpu_notifier(&conf->cpu_notify);
6253 for_each_possible_cpu(cpu)
6254 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6257 free_percpu(conf->percpu);
6260 static void free_conf(struct r5conf *conf)
6262 if (conf->shrinker.seeks)
6263 unregister_shrinker(&conf->shrinker);
6264 free_thread_groups(conf);
6265 shrink_stripes(conf);
6266 raid5_free_percpu(conf);
6268 kfree(conf->stripe_hashtbl);
6272 #ifdef CONFIG_HOTPLUG_CPU
6273 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6276 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6277 long cpu = (long)hcpu;
6278 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6281 case CPU_UP_PREPARE:
6282 case CPU_UP_PREPARE_FROZEN:
6283 if (alloc_scratch_buffer(conf, percpu)) {
6284 pr_err("%s: failed memory allocation for cpu%ld\n",
6286 return notifier_from_errno(-ENOMEM);
6290 case CPU_DEAD_FROZEN:
6291 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6300 static int raid5_alloc_percpu(struct r5conf *conf)
6305 conf->percpu = alloc_percpu(struct raid5_percpu);
6309 #ifdef CONFIG_HOTPLUG_CPU
6310 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6311 conf->cpu_notify.priority = 0;
6312 err = register_cpu_notifier(&conf->cpu_notify);
6318 for_each_present_cpu(cpu) {
6319 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6321 pr_err("%s: failed memory allocation for cpu%ld\n",
6331 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6332 struct shrink_control *sc)
6334 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6336 while (ret < sc->nr_to_scan) {
6337 if (drop_one_stripe(conf) == 0)
6344 static unsigned long raid5_cache_count(struct shrinker *shrink,
6345 struct shrink_control *sc)
6347 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6349 if (conf->max_nr_stripes < conf->min_nr_stripes)
6350 /* unlikely, but not impossible */
6352 return conf->max_nr_stripes - conf->min_nr_stripes;
6355 static struct r5conf *setup_conf(struct mddev *mddev)
6357 struct r5conf *conf;
6358 int raid_disk, memory, max_disks;
6359 struct md_rdev *rdev;
6360 struct disk_info *disk;
6363 int group_cnt, worker_cnt_per_group;
6364 struct r5worker_group *new_group;
6366 if (mddev->new_level != 5
6367 && mddev->new_level != 4
6368 && mddev->new_level != 6) {
6369 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6370 mdname(mddev), mddev->new_level);
6371 return ERR_PTR(-EIO);
6373 if ((mddev->new_level == 5
6374 && !algorithm_valid_raid5(mddev->new_layout)) ||
6375 (mddev->new_level == 6
6376 && !algorithm_valid_raid6(mddev->new_layout))) {
6377 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6378 mdname(mddev), mddev->new_layout);
6379 return ERR_PTR(-EIO);
6381 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6382 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6383 mdname(mddev), mddev->raid_disks);
6384 return ERR_PTR(-EINVAL);
6387 if (!mddev->new_chunk_sectors ||
6388 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6389 !is_power_of_2(mddev->new_chunk_sectors)) {
6390 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6391 mdname(mddev), mddev->new_chunk_sectors << 9);
6392 return ERR_PTR(-EINVAL);
6395 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6398 /* Don't enable multi-threading by default*/
6399 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6401 conf->group_cnt = group_cnt;
6402 conf->worker_cnt_per_group = worker_cnt_per_group;
6403 conf->worker_groups = new_group;
6406 spin_lock_init(&conf->device_lock);
6407 seqcount_init(&conf->gen_lock);
6408 init_waitqueue_head(&conf->wait_for_stripe);
6409 init_waitqueue_head(&conf->wait_for_overlap);
6410 INIT_LIST_HEAD(&conf->handle_list);
6411 INIT_LIST_HEAD(&conf->hold_list);
6412 INIT_LIST_HEAD(&conf->delayed_list);
6413 INIT_LIST_HEAD(&conf->bitmap_list);
6414 init_llist_head(&conf->released_stripes);
6415 atomic_set(&conf->active_stripes, 0);
6416 atomic_set(&conf->preread_active_stripes, 0);
6417 atomic_set(&conf->active_aligned_reads, 0);
6418 conf->bypass_threshold = BYPASS_THRESHOLD;
6419 conf->recovery_disabled = mddev->recovery_disabled - 1;
6421 conf->raid_disks = mddev->raid_disks;
6422 if (mddev->reshape_position == MaxSector)
6423 conf->previous_raid_disks = mddev->raid_disks;
6425 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6426 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6428 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6433 conf->mddev = mddev;
6435 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6438 /* We init hash_locks[0] separately to that it can be used
6439 * as the reference lock in the spin_lock_nest_lock() call
6440 * in lock_all_device_hash_locks_irq in order to convince
6441 * lockdep that we know what we are doing.
6443 spin_lock_init(conf->hash_locks);
6444 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6445 spin_lock_init(conf->hash_locks + i);
6447 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6448 INIT_LIST_HEAD(conf->inactive_list + i);
6450 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6451 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6453 conf->level = mddev->new_level;
6454 conf->chunk_sectors = mddev->new_chunk_sectors;
6455 if (raid5_alloc_percpu(conf) != 0)
6458 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6460 rdev_for_each(rdev, mddev) {
6461 raid_disk = rdev->raid_disk;
6462 if (raid_disk >= max_disks
6465 disk = conf->disks + raid_disk;
6467 if (test_bit(Replacement, &rdev->flags)) {
6468 if (disk->replacement)
6470 disk->replacement = rdev;
6477 if (test_bit(In_sync, &rdev->flags)) {
6478 char b[BDEVNAME_SIZE];
6479 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6481 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6482 } else if (rdev->saved_raid_disk != raid_disk)
6483 /* Cannot rely on bitmap to complete recovery */
6487 conf->level = mddev->new_level;
6488 if (conf->level == 6) {
6489 conf->max_degraded = 2;
6490 if (raid6_call.xor_syndrome)
6491 conf->rmw_level = PARITY_ENABLE_RMW;
6493 conf->rmw_level = PARITY_DISABLE_RMW;
6495 conf->max_degraded = 1;
6496 conf->rmw_level = PARITY_ENABLE_RMW;
6498 conf->algorithm = mddev->new_layout;
6499 conf->reshape_progress = mddev->reshape_position;
6500 if (conf->reshape_progress != MaxSector) {
6501 conf->prev_chunk_sectors = mddev->chunk_sectors;
6502 conf->prev_algo = mddev->layout;
6505 conf->min_nr_stripes = NR_STRIPES;
6506 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6507 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6508 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6509 if (grow_stripes(conf, conf->min_nr_stripes)) {
6511 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6512 mdname(mddev), memory);
6515 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6516 mdname(mddev), memory);
6518 * Losing a stripe head costs more than the time to refill it,
6519 * it reduces the queue depth and so can hurt throughput.
6520 * So set it rather large, scaled by number of devices.
6522 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6523 conf->shrinker.scan_objects = raid5_cache_scan;
6524 conf->shrinker.count_objects = raid5_cache_count;
6525 conf->shrinker.batch = 128;
6526 conf->shrinker.flags = 0;
6527 register_shrinker(&conf->shrinker);
6529 sprintf(pers_name, "raid%d", mddev->new_level);
6530 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6531 if (!conf->thread) {
6533 "md/raid:%s: couldn't allocate thread.\n",
6543 return ERR_PTR(-EIO);
6545 return ERR_PTR(-ENOMEM);
6548 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6551 case ALGORITHM_PARITY_0:
6552 if (raid_disk < max_degraded)
6555 case ALGORITHM_PARITY_N:
6556 if (raid_disk >= raid_disks - max_degraded)
6559 case ALGORITHM_PARITY_0_6:
6560 if (raid_disk == 0 ||
6561 raid_disk == raid_disks - 1)
6564 case ALGORITHM_LEFT_ASYMMETRIC_6:
6565 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6566 case ALGORITHM_LEFT_SYMMETRIC_6:
6567 case ALGORITHM_RIGHT_SYMMETRIC_6:
6568 if (raid_disk == raid_disks - 1)
6574 static int run(struct mddev *mddev)
6576 struct r5conf *conf;
6577 int working_disks = 0;
6578 int dirty_parity_disks = 0;
6579 struct md_rdev *rdev;
6580 sector_t reshape_offset = 0;
6582 long long min_offset_diff = 0;
6585 if (mddev->recovery_cp != MaxSector)
6586 printk(KERN_NOTICE "md/raid:%s: not clean"
6587 " -- starting background reconstruction\n",
6590 rdev_for_each(rdev, mddev) {
6592 if (rdev->raid_disk < 0)
6594 diff = (rdev->new_data_offset - rdev->data_offset);
6596 min_offset_diff = diff;
6598 } else if (mddev->reshape_backwards &&
6599 diff < min_offset_diff)
6600 min_offset_diff = diff;
6601 else if (!mddev->reshape_backwards &&
6602 diff > min_offset_diff)
6603 min_offset_diff = diff;
6606 if (mddev->reshape_position != MaxSector) {
6607 /* Check that we can continue the reshape.
6608 * Difficulties arise if the stripe we would write to
6609 * next is at or after the stripe we would read from next.
6610 * For a reshape that changes the number of devices, this
6611 * is only possible for a very short time, and mdadm makes
6612 * sure that time appears to have past before assembling
6613 * the array. So we fail if that time hasn't passed.
6614 * For a reshape that keeps the number of devices the same
6615 * mdadm must be monitoring the reshape can keeping the
6616 * critical areas read-only and backed up. It will start
6617 * the array in read-only mode, so we check for that.
6619 sector_t here_new, here_old;
6621 int max_degraded = (mddev->level == 6 ? 2 : 1);
6623 if (mddev->new_level != mddev->level) {
6624 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6625 "required - aborting.\n",
6629 old_disks = mddev->raid_disks - mddev->delta_disks;
6630 /* reshape_position must be on a new-stripe boundary, and one
6631 * further up in new geometry must map after here in old
6634 here_new = mddev->reshape_position;
6635 if (sector_div(here_new, mddev->new_chunk_sectors *
6636 (mddev->raid_disks - max_degraded))) {
6637 printk(KERN_ERR "md/raid:%s: reshape_position not "
6638 "on a stripe boundary\n", mdname(mddev));
6641 reshape_offset = here_new * mddev->new_chunk_sectors;
6642 /* here_new is the stripe we will write to */
6643 here_old = mddev->reshape_position;
6644 sector_div(here_old, mddev->chunk_sectors *
6645 (old_disks-max_degraded));
6646 /* here_old is the first stripe that we might need to read
6648 if (mddev->delta_disks == 0) {
6649 if ((here_new * mddev->new_chunk_sectors !=
6650 here_old * mddev->chunk_sectors)) {
6651 printk(KERN_ERR "md/raid:%s: reshape position is"
6652 " confused - aborting\n", mdname(mddev));
6655 /* We cannot be sure it is safe to start an in-place
6656 * reshape. It is only safe if user-space is monitoring
6657 * and taking constant backups.
6658 * mdadm always starts a situation like this in
6659 * readonly mode so it can take control before
6660 * allowing any writes. So just check for that.
6662 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6663 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6664 /* not really in-place - so OK */;
6665 else if (mddev->ro == 0) {
6666 printk(KERN_ERR "md/raid:%s: in-place reshape "
6667 "must be started in read-only mode "
6672 } else if (mddev->reshape_backwards
6673 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6674 here_old * mddev->chunk_sectors)
6675 : (here_new * mddev->new_chunk_sectors >=
6676 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6677 /* Reading from the same stripe as writing to - bad */
6678 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6679 "auto-recovery - aborting.\n",
6683 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6685 /* OK, we should be able to continue; */
6687 BUG_ON(mddev->level != mddev->new_level);
6688 BUG_ON(mddev->layout != mddev->new_layout);
6689 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6690 BUG_ON(mddev->delta_disks != 0);
6693 if (mddev->private == NULL)
6694 conf = setup_conf(mddev);
6696 conf = mddev->private;
6699 return PTR_ERR(conf);
6701 conf->min_offset_diff = min_offset_diff;
6702 mddev->thread = conf->thread;
6703 conf->thread = NULL;
6704 mddev->private = conf;
6706 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6708 rdev = conf->disks[i].rdev;
6709 if (!rdev && conf->disks[i].replacement) {
6710 /* The replacement is all we have yet */
6711 rdev = conf->disks[i].replacement;
6712 conf->disks[i].replacement = NULL;
6713 clear_bit(Replacement, &rdev->flags);
6714 conf->disks[i].rdev = rdev;
6718 if (conf->disks[i].replacement &&
6719 conf->reshape_progress != MaxSector) {
6720 /* replacements and reshape simply do not mix. */
6721 printk(KERN_ERR "md: cannot handle concurrent "
6722 "replacement and reshape.\n");
6725 if (test_bit(In_sync, &rdev->flags)) {
6729 /* This disc is not fully in-sync. However if it
6730 * just stored parity (beyond the recovery_offset),
6731 * when we don't need to be concerned about the
6732 * array being dirty.
6733 * When reshape goes 'backwards', we never have
6734 * partially completed devices, so we only need
6735 * to worry about reshape going forwards.
6737 /* Hack because v0.91 doesn't store recovery_offset properly. */
6738 if (mddev->major_version == 0 &&
6739 mddev->minor_version > 90)
6740 rdev->recovery_offset = reshape_offset;
6742 if (rdev->recovery_offset < reshape_offset) {
6743 /* We need to check old and new layout */
6744 if (!only_parity(rdev->raid_disk,
6747 conf->max_degraded))
6750 if (!only_parity(rdev->raid_disk,
6752 conf->previous_raid_disks,
6753 conf->max_degraded))
6755 dirty_parity_disks++;
6759 * 0 for a fully functional array, 1 or 2 for a degraded array.
6761 mddev->degraded = calc_degraded(conf);
6763 if (has_failed(conf)) {
6764 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6765 " (%d/%d failed)\n",
6766 mdname(mddev), mddev->degraded, conf->raid_disks);
6770 /* device size must be a multiple of chunk size */
6771 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6772 mddev->resync_max_sectors = mddev->dev_sectors;
6774 if (mddev->degraded > dirty_parity_disks &&
6775 mddev->recovery_cp != MaxSector) {
6776 if (mddev->ok_start_degraded)
6778 "md/raid:%s: starting dirty degraded array"
6779 " - data corruption possible.\n",
6783 "md/raid:%s: cannot start dirty degraded array.\n",
6789 if (mddev->degraded == 0)
6790 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6791 " devices, algorithm %d\n", mdname(mddev), conf->level,
6792 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6795 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6796 " out of %d devices, algorithm %d\n",
6797 mdname(mddev), conf->level,
6798 mddev->raid_disks - mddev->degraded,
6799 mddev->raid_disks, mddev->new_layout);
6801 print_raid5_conf(conf);
6803 if (conf->reshape_progress != MaxSector) {
6804 conf->reshape_safe = conf->reshape_progress;
6805 atomic_set(&conf->reshape_stripes, 0);
6806 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6807 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6808 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6809 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6810 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6814 /* Ok, everything is just fine now */
6815 if (mddev->to_remove == &raid5_attrs_group)
6816 mddev->to_remove = NULL;
6817 else if (mddev->kobj.sd &&
6818 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6820 "raid5: failed to create sysfs attributes for %s\n",
6822 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6826 bool discard_supported = true;
6827 /* read-ahead size must cover two whole stripes, which
6828 * is 2 * (datadisks) * chunksize where 'n' is the
6829 * number of raid devices
6831 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6832 int stripe = data_disks *
6833 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6834 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6835 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6837 chunk_size = mddev->chunk_sectors << 9;
6838 blk_queue_io_min(mddev->queue, chunk_size);
6839 blk_queue_io_opt(mddev->queue, chunk_size *
6840 (conf->raid_disks - conf->max_degraded));
6841 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6843 * We can only discard a whole stripe. It doesn't make sense to
6844 * discard data disk but write parity disk
6846 stripe = stripe * PAGE_SIZE;
6847 /* Round up to power of 2, as discard handling
6848 * currently assumes that */
6849 while ((stripe-1) & stripe)
6850 stripe = (stripe | (stripe-1)) + 1;
6851 mddev->queue->limits.discard_alignment = stripe;
6852 mddev->queue->limits.discard_granularity = stripe;
6854 * unaligned part of discard request will be ignored, so can't
6855 * guarantee discard_zeroes_data
6857 mddev->queue->limits.discard_zeroes_data = 0;
6859 blk_queue_max_write_same_sectors(mddev->queue, 0);
6861 rdev_for_each(rdev, mddev) {
6862 disk_stack_limits(mddev->gendisk, rdev->bdev,
6863 rdev->data_offset << 9);
6864 disk_stack_limits(mddev->gendisk, rdev->bdev,
6865 rdev->new_data_offset << 9);
6867 * discard_zeroes_data is required, otherwise data
6868 * could be lost. Consider a scenario: discard a stripe
6869 * (the stripe could be inconsistent if
6870 * discard_zeroes_data is 0); write one disk of the
6871 * stripe (the stripe could be inconsistent again
6872 * depending on which disks are used to calculate
6873 * parity); the disk is broken; The stripe data of this
6876 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6877 !bdev_get_queue(rdev->bdev)->
6878 limits.discard_zeroes_data)
6879 discard_supported = false;
6880 /* Unfortunately, discard_zeroes_data is not currently
6881 * a guarantee - just a hint. So we only allow DISCARD
6882 * if the sysadmin has confirmed that only safe devices
6883 * are in use by setting a module parameter.
6885 if (!devices_handle_discard_safely) {
6886 if (discard_supported) {
6887 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6888 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6890 discard_supported = false;
6894 if (discard_supported &&
6895 mddev->queue->limits.max_discard_sectors >= stripe &&
6896 mddev->queue->limits.discard_granularity >= stripe)
6897 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6900 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6906 md_unregister_thread(&mddev->thread);
6907 print_raid5_conf(conf);
6909 mddev->private = NULL;
6910 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6914 static void raid5_free(struct mddev *mddev, void *priv)
6916 struct r5conf *conf = priv;
6919 mddev->to_remove = &raid5_attrs_group;
6922 static void status(struct seq_file *seq, struct mddev *mddev)
6924 struct r5conf *conf = mddev->private;
6927 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6928 mddev->chunk_sectors / 2, mddev->layout);
6929 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6930 for (i = 0; i < conf->raid_disks; i++)
6931 seq_printf (seq, "%s",
6932 conf->disks[i].rdev &&
6933 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6934 seq_printf (seq, "]");
6937 static void print_raid5_conf (struct r5conf *conf)
6940 struct disk_info *tmp;
6942 printk(KERN_DEBUG "RAID conf printout:\n");
6944 printk("(conf==NULL)\n");
6947 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6949 conf->raid_disks - conf->mddev->degraded);
6951 for (i = 0; i < conf->raid_disks; i++) {
6952 char b[BDEVNAME_SIZE];
6953 tmp = conf->disks + i;
6955 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6956 i, !test_bit(Faulty, &tmp->rdev->flags),
6957 bdevname(tmp->rdev->bdev, b));
6961 static int raid5_spare_active(struct mddev *mddev)
6964 struct r5conf *conf = mddev->private;
6965 struct disk_info *tmp;
6967 unsigned long flags;
6969 for (i = 0; i < conf->raid_disks; i++) {
6970 tmp = conf->disks + i;
6971 if (tmp->replacement
6972 && tmp->replacement->recovery_offset == MaxSector
6973 && !test_bit(Faulty, &tmp->replacement->flags)
6974 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
6975 /* Replacement has just become active. */
6977 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
6980 /* Replaced device not technically faulty,
6981 * but we need to be sure it gets removed
6982 * and never re-added.
6984 set_bit(Faulty, &tmp->rdev->flags);
6985 sysfs_notify_dirent_safe(
6986 tmp->rdev->sysfs_state);
6988 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
6989 } else if (tmp->rdev
6990 && tmp->rdev->recovery_offset == MaxSector
6991 && !test_bit(Faulty, &tmp->rdev->flags)
6992 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
6994 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
6997 spin_lock_irqsave(&conf->device_lock, flags);
6998 mddev->degraded = calc_degraded(conf);
6999 spin_unlock_irqrestore(&conf->device_lock, flags);
7000 print_raid5_conf(conf);
7004 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7006 struct r5conf *conf = mddev->private;
7008 int number = rdev->raid_disk;
7009 struct md_rdev **rdevp;
7010 struct disk_info *p = conf->disks + number;
7012 print_raid5_conf(conf);
7013 if (rdev == p->rdev)
7015 else if (rdev == p->replacement)
7016 rdevp = &p->replacement;
7020 if (number >= conf->raid_disks &&
7021 conf->reshape_progress == MaxSector)
7022 clear_bit(In_sync, &rdev->flags);
7024 if (test_bit(In_sync, &rdev->flags) ||
7025 atomic_read(&rdev->nr_pending)) {
7029 /* Only remove non-faulty devices if recovery
7032 if (!test_bit(Faulty, &rdev->flags) &&
7033 mddev->recovery_disabled != conf->recovery_disabled &&
7034 !has_failed(conf) &&
7035 (!p->replacement || p->replacement == rdev) &&
7036 number < conf->raid_disks) {
7042 if (atomic_read(&rdev->nr_pending)) {
7043 /* lost the race, try later */
7046 } else if (p->replacement) {
7047 /* We must have just cleared 'rdev' */
7048 p->rdev = p->replacement;
7049 clear_bit(Replacement, &p->replacement->flags);
7050 smp_mb(); /* Make sure other CPUs may see both as identical
7051 * but will never see neither - if they are careful
7053 p->replacement = NULL;
7054 clear_bit(WantReplacement, &rdev->flags);
7056 /* We might have just removed the Replacement as faulty-
7057 * clear the bit just in case
7059 clear_bit(WantReplacement, &rdev->flags);
7062 print_raid5_conf(conf);
7066 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7068 struct r5conf *conf = mddev->private;
7071 struct disk_info *p;
7073 int last = conf->raid_disks - 1;
7075 if (mddev->recovery_disabled == conf->recovery_disabled)
7078 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7079 /* no point adding a device */
7082 if (rdev->raid_disk >= 0)
7083 first = last = rdev->raid_disk;
7086 * find the disk ... but prefer rdev->saved_raid_disk
7089 if (rdev->saved_raid_disk >= 0 &&
7090 rdev->saved_raid_disk >= first &&
7091 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7092 first = rdev->saved_raid_disk;
7094 for (disk = first; disk <= last; disk++) {
7095 p = conf->disks + disk;
7096 if (p->rdev == NULL) {
7097 clear_bit(In_sync, &rdev->flags);
7098 rdev->raid_disk = disk;
7100 if (rdev->saved_raid_disk != disk)
7102 rcu_assign_pointer(p->rdev, rdev);
7106 for (disk = first; disk <= last; disk++) {
7107 p = conf->disks + disk;
7108 if (test_bit(WantReplacement, &p->rdev->flags) &&
7109 p->replacement == NULL) {
7110 clear_bit(In_sync, &rdev->flags);
7111 set_bit(Replacement, &rdev->flags);
7112 rdev->raid_disk = disk;
7115 rcu_assign_pointer(p->replacement, rdev);
7120 print_raid5_conf(conf);
7124 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7126 /* no resync is happening, and there is enough space
7127 * on all devices, so we can resize.
7128 * We need to make sure resync covers any new space.
7129 * If the array is shrinking we should possibly wait until
7130 * any io in the removed space completes, but it hardly seems
7134 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7135 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7136 if (mddev->external_size &&
7137 mddev->array_sectors > newsize)
7139 if (mddev->bitmap) {
7140 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7144 md_set_array_sectors(mddev, newsize);
7145 set_capacity(mddev->gendisk, mddev->array_sectors);
7146 revalidate_disk(mddev->gendisk);
7147 if (sectors > mddev->dev_sectors &&
7148 mddev->recovery_cp > mddev->dev_sectors) {
7149 mddev->recovery_cp = mddev->dev_sectors;
7150 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7152 mddev->dev_sectors = sectors;
7153 mddev->resync_max_sectors = sectors;
7157 static int check_stripe_cache(struct mddev *mddev)
7159 /* Can only proceed if there are plenty of stripe_heads.
7160 * We need a minimum of one full stripe,, and for sensible progress
7161 * it is best to have about 4 times that.
7162 * If we require 4 times, then the default 256 4K stripe_heads will
7163 * allow for chunk sizes up to 256K, which is probably OK.
7164 * If the chunk size is greater, user-space should request more
7165 * stripe_heads first.
7167 struct r5conf *conf = mddev->private;
7168 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7169 > conf->min_nr_stripes ||
7170 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7171 > conf->min_nr_stripes) {
7172 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7174 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7181 static int check_reshape(struct mddev *mddev)
7183 struct r5conf *conf = mddev->private;
7185 if (mddev->delta_disks == 0 &&
7186 mddev->new_layout == mddev->layout &&
7187 mddev->new_chunk_sectors == mddev->chunk_sectors)
7188 return 0; /* nothing to do */
7189 if (has_failed(conf))
7191 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7192 /* We might be able to shrink, but the devices must
7193 * be made bigger first.
7194 * For raid6, 4 is the minimum size.
7195 * Otherwise 2 is the minimum
7198 if (mddev->level == 6)
7200 if (mddev->raid_disks + mddev->delta_disks < min)
7204 if (!check_stripe_cache(mddev))
7207 return resize_stripes(conf, (conf->previous_raid_disks
7208 + mddev->delta_disks));
7211 static int raid5_start_reshape(struct mddev *mddev)
7213 struct r5conf *conf = mddev->private;
7214 struct md_rdev *rdev;
7216 unsigned long flags;
7218 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7221 if (!check_stripe_cache(mddev))
7224 if (has_failed(conf))
7227 rdev_for_each(rdev, mddev) {
7228 if (!test_bit(In_sync, &rdev->flags)
7229 && !test_bit(Faulty, &rdev->flags))
7233 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7234 /* Not enough devices even to make a degraded array
7239 /* Refuse to reduce size of the array. Any reductions in
7240 * array size must be through explicit setting of array_size
7243 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7244 < mddev->array_sectors) {
7245 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7246 "before number of disks\n", mdname(mddev));
7250 atomic_set(&conf->reshape_stripes, 0);
7251 spin_lock_irq(&conf->device_lock);
7252 write_seqcount_begin(&conf->gen_lock);
7253 conf->previous_raid_disks = conf->raid_disks;
7254 conf->raid_disks += mddev->delta_disks;
7255 conf->prev_chunk_sectors = conf->chunk_sectors;
7256 conf->chunk_sectors = mddev->new_chunk_sectors;
7257 conf->prev_algo = conf->algorithm;
7258 conf->algorithm = mddev->new_layout;
7260 /* Code that selects data_offset needs to see the generation update
7261 * if reshape_progress has been set - so a memory barrier needed.
7264 if (mddev->reshape_backwards)
7265 conf->reshape_progress = raid5_size(mddev, 0, 0);
7267 conf->reshape_progress = 0;
7268 conf->reshape_safe = conf->reshape_progress;
7269 write_seqcount_end(&conf->gen_lock);
7270 spin_unlock_irq(&conf->device_lock);
7272 /* Now make sure any requests that proceeded on the assumption
7273 * the reshape wasn't running - like Discard or Read - have
7276 mddev_suspend(mddev);
7277 mddev_resume(mddev);
7279 /* Add some new drives, as many as will fit.
7280 * We know there are enough to make the newly sized array work.
7281 * Don't add devices if we are reducing the number of
7282 * devices in the array. This is because it is not possible
7283 * to correctly record the "partially reconstructed" state of
7284 * such devices during the reshape and confusion could result.
7286 if (mddev->delta_disks >= 0) {
7287 rdev_for_each(rdev, mddev)
7288 if (rdev->raid_disk < 0 &&
7289 !test_bit(Faulty, &rdev->flags)) {
7290 if (raid5_add_disk(mddev, rdev) == 0) {
7292 >= conf->previous_raid_disks)
7293 set_bit(In_sync, &rdev->flags);
7295 rdev->recovery_offset = 0;
7297 if (sysfs_link_rdev(mddev, rdev))
7298 /* Failure here is OK */;
7300 } else if (rdev->raid_disk >= conf->previous_raid_disks
7301 && !test_bit(Faulty, &rdev->flags)) {
7302 /* This is a spare that was manually added */
7303 set_bit(In_sync, &rdev->flags);
7306 /* When a reshape changes the number of devices,
7307 * ->degraded is measured against the larger of the
7308 * pre and post number of devices.
7310 spin_lock_irqsave(&conf->device_lock, flags);
7311 mddev->degraded = calc_degraded(conf);
7312 spin_unlock_irqrestore(&conf->device_lock, flags);
7314 mddev->raid_disks = conf->raid_disks;
7315 mddev->reshape_position = conf->reshape_progress;
7316 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7318 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7319 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7320 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7321 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7322 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7324 if (!mddev->sync_thread) {
7325 mddev->recovery = 0;
7326 spin_lock_irq(&conf->device_lock);
7327 write_seqcount_begin(&conf->gen_lock);
7328 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7329 mddev->new_chunk_sectors =
7330 conf->chunk_sectors = conf->prev_chunk_sectors;
7331 mddev->new_layout = conf->algorithm = conf->prev_algo;
7332 rdev_for_each(rdev, mddev)
7333 rdev->new_data_offset = rdev->data_offset;
7335 conf->generation --;
7336 conf->reshape_progress = MaxSector;
7337 mddev->reshape_position = MaxSector;
7338 write_seqcount_end(&conf->gen_lock);
7339 spin_unlock_irq(&conf->device_lock);
7342 conf->reshape_checkpoint = jiffies;
7343 md_wakeup_thread(mddev->sync_thread);
7344 md_new_event(mddev);
7348 /* This is called from the reshape thread and should make any
7349 * changes needed in 'conf'
7351 static void end_reshape(struct r5conf *conf)
7354 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7355 struct md_rdev *rdev;
7357 spin_lock_irq(&conf->device_lock);
7358 conf->previous_raid_disks = conf->raid_disks;
7359 rdev_for_each(rdev, conf->mddev)
7360 rdev->data_offset = rdev->new_data_offset;
7362 conf->reshape_progress = MaxSector;
7363 spin_unlock_irq(&conf->device_lock);
7364 wake_up(&conf->wait_for_overlap);
7366 /* read-ahead size must cover two whole stripes, which is
7367 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7369 if (conf->mddev->queue) {
7370 int data_disks = conf->raid_disks - conf->max_degraded;
7371 int stripe = data_disks * ((conf->chunk_sectors << 9)
7373 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7374 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7379 /* This is called from the raid5d thread with mddev_lock held.
7380 * It makes config changes to the device.
7382 static void raid5_finish_reshape(struct mddev *mddev)
7384 struct r5conf *conf = mddev->private;
7386 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7388 if (mddev->delta_disks > 0) {
7389 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7390 set_capacity(mddev->gendisk, mddev->array_sectors);
7391 revalidate_disk(mddev->gendisk);
7394 spin_lock_irq(&conf->device_lock);
7395 mddev->degraded = calc_degraded(conf);
7396 spin_unlock_irq(&conf->device_lock);
7397 for (d = conf->raid_disks ;
7398 d < conf->raid_disks - mddev->delta_disks;
7400 struct md_rdev *rdev = conf->disks[d].rdev;
7402 clear_bit(In_sync, &rdev->flags);
7403 rdev = conf->disks[d].replacement;
7405 clear_bit(In_sync, &rdev->flags);
7408 mddev->layout = conf->algorithm;
7409 mddev->chunk_sectors = conf->chunk_sectors;
7410 mddev->reshape_position = MaxSector;
7411 mddev->delta_disks = 0;
7412 mddev->reshape_backwards = 0;
7416 static void raid5_quiesce(struct mddev *mddev, int state)
7418 struct r5conf *conf = mddev->private;
7421 case 2: /* resume for a suspend */
7422 wake_up(&conf->wait_for_overlap);
7425 case 1: /* stop all writes */
7426 lock_all_device_hash_locks_irq(conf);
7427 /* '2' tells resync/reshape to pause so that all
7428 * active stripes can drain
7431 wait_event_cmd(conf->wait_for_stripe,
7432 atomic_read(&conf->active_stripes) == 0 &&
7433 atomic_read(&conf->active_aligned_reads) == 0,
7434 unlock_all_device_hash_locks_irq(conf),
7435 lock_all_device_hash_locks_irq(conf));
7437 unlock_all_device_hash_locks_irq(conf);
7438 /* allow reshape to continue */
7439 wake_up(&conf->wait_for_overlap);
7442 case 0: /* re-enable writes */
7443 lock_all_device_hash_locks_irq(conf);
7445 wake_up(&conf->wait_for_stripe);
7446 wake_up(&conf->wait_for_overlap);
7447 unlock_all_device_hash_locks_irq(conf);
7452 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7454 struct r0conf *raid0_conf = mddev->private;
7457 /* for raid0 takeover only one zone is supported */
7458 if (raid0_conf->nr_strip_zones > 1) {
7459 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7461 return ERR_PTR(-EINVAL);
7464 sectors = raid0_conf->strip_zone[0].zone_end;
7465 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7466 mddev->dev_sectors = sectors;
7467 mddev->new_level = level;
7468 mddev->new_layout = ALGORITHM_PARITY_N;
7469 mddev->new_chunk_sectors = mddev->chunk_sectors;
7470 mddev->raid_disks += 1;
7471 mddev->delta_disks = 1;
7472 /* make sure it will be not marked as dirty */
7473 mddev->recovery_cp = MaxSector;
7475 return setup_conf(mddev);
7478 static void *raid5_takeover_raid1(struct mddev *mddev)
7482 if (mddev->raid_disks != 2 ||
7483 mddev->degraded > 1)
7484 return ERR_PTR(-EINVAL);
7486 /* Should check if there are write-behind devices? */
7488 chunksect = 64*2; /* 64K by default */
7490 /* The array must be an exact multiple of chunksize */
7491 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7494 if ((chunksect<<9) < STRIPE_SIZE)
7495 /* array size does not allow a suitable chunk size */
7496 return ERR_PTR(-EINVAL);
7498 mddev->new_level = 5;
7499 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7500 mddev->new_chunk_sectors = chunksect;
7502 return setup_conf(mddev);
7505 static void *raid5_takeover_raid6(struct mddev *mddev)
7509 switch (mddev->layout) {
7510 case ALGORITHM_LEFT_ASYMMETRIC_6:
7511 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7513 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7514 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7516 case ALGORITHM_LEFT_SYMMETRIC_6:
7517 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7519 case ALGORITHM_RIGHT_SYMMETRIC_6:
7520 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7522 case ALGORITHM_PARITY_0_6:
7523 new_layout = ALGORITHM_PARITY_0;
7525 case ALGORITHM_PARITY_N:
7526 new_layout = ALGORITHM_PARITY_N;
7529 return ERR_PTR(-EINVAL);
7531 mddev->new_level = 5;
7532 mddev->new_layout = new_layout;
7533 mddev->delta_disks = -1;
7534 mddev->raid_disks -= 1;
7535 return setup_conf(mddev);
7538 static int raid5_check_reshape(struct mddev *mddev)
7540 /* For a 2-drive array, the layout and chunk size can be changed
7541 * immediately as not restriping is needed.
7542 * For larger arrays we record the new value - after validation
7543 * to be used by a reshape pass.
7545 struct r5conf *conf = mddev->private;
7546 int new_chunk = mddev->new_chunk_sectors;
7548 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7550 if (new_chunk > 0) {
7551 if (!is_power_of_2(new_chunk))
7553 if (new_chunk < (PAGE_SIZE>>9))
7555 if (mddev->array_sectors & (new_chunk-1))
7556 /* not factor of array size */
7560 /* They look valid */
7562 if (mddev->raid_disks == 2) {
7563 /* can make the change immediately */
7564 if (mddev->new_layout >= 0) {
7565 conf->algorithm = mddev->new_layout;
7566 mddev->layout = mddev->new_layout;
7568 if (new_chunk > 0) {
7569 conf->chunk_sectors = new_chunk ;
7570 mddev->chunk_sectors = new_chunk;
7572 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7573 md_wakeup_thread(mddev->thread);
7575 return check_reshape(mddev);
7578 static int raid6_check_reshape(struct mddev *mddev)
7580 int new_chunk = mddev->new_chunk_sectors;
7582 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7584 if (new_chunk > 0) {
7585 if (!is_power_of_2(new_chunk))
7587 if (new_chunk < (PAGE_SIZE >> 9))
7589 if (mddev->array_sectors & (new_chunk-1))
7590 /* not factor of array size */
7594 /* They look valid */
7595 return check_reshape(mddev);
7598 static void *raid5_takeover(struct mddev *mddev)
7600 /* raid5 can take over:
7601 * raid0 - if there is only one strip zone - make it a raid4 layout
7602 * raid1 - if there are two drives. We need to know the chunk size
7603 * raid4 - trivial - just use a raid4 layout.
7604 * raid6 - Providing it is a *_6 layout
7606 if (mddev->level == 0)
7607 return raid45_takeover_raid0(mddev, 5);
7608 if (mddev->level == 1)
7609 return raid5_takeover_raid1(mddev);
7610 if (mddev->level == 4) {
7611 mddev->new_layout = ALGORITHM_PARITY_N;
7612 mddev->new_level = 5;
7613 return setup_conf(mddev);
7615 if (mddev->level == 6)
7616 return raid5_takeover_raid6(mddev);
7618 return ERR_PTR(-EINVAL);
7621 static void *raid4_takeover(struct mddev *mddev)
7623 /* raid4 can take over:
7624 * raid0 - if there is only one strip zone
7625 * raid5 - if layout is right
7627 if (mddev->level == 0)
7628 return raid45_takeover_raid0(mddev, 4);
7629 if (mddev->level == 5 &&
7630 mddev->layout == ALGORITHM_PARITY_N) {
7631 mddev->new_layout = 0;
7632 mddev->new_level = 4;
7633 return setup_conf(mddev);
7635 return ERR_PTR(-EINVAL);
7638 static struct md_personality raid5_personality;
7640 static void *raid6_takeover(struct mddev *mddev)
7642 /* Currently can only take over a raid5. We map the
7643 * personality to an equivalent raid6 personality
7644 * with the Q block at the end.
7648 if (mddev->pers != &raid5_personality)
7649 return ERR_PTR(-EINVAL);
7650 if (mddev->degraded > 1)
7651 return ERR_PTR(-EINVAL);
7652 if (mddev->raid_disks > 253)
7653 return ERR_PTR(-EINVAL);
7654 if (mddev->raid_disks < 3)
7655 return ERR_PTR(-EINVAL);
7657 switch (mddev->layout) {
7658 case ALGORITHM_LEFT_ASYMMETRIC:
7659 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7661 case ALGORITHM_RIGHT_ASYMMETRIC:
7662 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7664 case ALGORITHM_LEFT_SYMMETRIC:
7665 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7667 case ALGORITHM_RIGHT_SYMMETRIC:
7668 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7670 case ALGORITHM_PARITY_0:
7671 new_layout = ALGORITHM_PARITY_0_6;
7673 case ALGORITHM_PARITY_N:
7674 new_layout = ALGORITHM_PARITY_N;
7677 return ERR_PTR(-EINVAL);
7679 mddev->new_level = 6;
7680 mddev->new_layout = new_layout;
7681 mddev->delta_disks = 1;
7682 mddev->raid_disks += 1;
7683 return setup_conf(mddev);
7686 static struct md_personality raid6_personality =
7690 .owner = THIS_MODULE,
7691 .make_request = make_request,
7695 .error_handler = error,
7696 .hot_add_disk = raid5_add_disk,
7697 .hot_remove_disk= raid5_remove_disk,
7698 .spare_active = raid5_spare_active,
7699 .sync_request = sync_request,
7700 .resize = raid5_resize,
7702 .check_reshape = raid6_check_reshape,
7703 .start_reshape = raid5_start_reshape,
7704 .finish_reshape = raid5_finish_reshape,
7705 .quiesce = raid5_quiesce,
7706 .takeover = raid6_takeover,
7707 .congested = raid5_congested,
7708 .mergeable_bvec = raid5_mergeable_bvec,
7710 static struct md_personality raid5_personality =
7714 .owner = THIS_MODULE,
7715 .make_request = make_request,
7719 .error_handler = error,
7720 .hot_add_disk = raid5_add_disk,
7721 .hot_remove_disk= raid5_remove_disk,
7722 .spare_active = raid5_spare_active,
7723 .sync_request = sync_request,
7724 .resize = raid5_resize,
7726 .check_reshape = raid5_check_reshape,
7727 .start_reshape = raid5_start_reshape,
7728 .finish_reshape = raid5_finish_reshape,
7729 .quiesce = raid5_quiesce,
7730 .takeover = raid5_takeover,
7731 .congested = raid5_congested,
7732 .mergeable_bvec = raid5_mergeable_bvec,
7735 static struct md_personality raid4_personality =
7739 .owner = THIS_MODULE,
7740 .make_request = make_request,
7744 .error_handler = error,
7745 .hot_add_disk = raid5_add_disk,
7746 .hot_remove_disk= raid5_remove_disk,
7747 .spare_active = raid5_spare_active,
7748 .sync_request = sync_request,
7749 .resize = raid5_resize,
7751 .check_reshape = raid5_check_reshape,
7752 .start_reshape = raid5_start_reshape,
7753 .finish_reshape = raid5_finish_reshape,
7754 .quiesce = raid5_quiesce,
7755 .takeover = raid4_takeover,
7756 .congested = raid5_congested,
7757 .mergeable_bvec = raid5_mergeable_bvec,
7760 static int __init raid5_init(void)
7762 raid5_wq = alloc_workqueue("raid5wq",
7763 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7766 register_md_personality(&raid6_personality);
7767 register_md_personality(&raid5_personality);
7768 register_md_personality(&raid4_personality);
7772 static void raid5_exit(void)
7774 unregister_md_personality(&raid6_personality);
7775 unregister_md_personality(&raid5_personality);
7776 unregister_md_personality(&raid4_personality);
7777 destroy_workqueue(raid5_wq);
7780 module_init(raid5_init);
7781 module_exit(raid5_exit);
7782 MODULE_LICENSE("GPL");
7783 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7784 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7785 MODULE_ALIAS("md-raid5");
7786 MODULE_ALIAS("md-raid4");
7787 MODULE_ALIAS("md-level-5");
7788 MODULE_ALIAS("md-level-4");
7789 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7790 MODULE_ALIAS("md-raid6");
7791 MODULE_ALIAS("md-level-6");
7793 /* This used to be two separate modules, they were: */
7794 MODULE_ALIAS("raid5");
7795 MODULE_ALIAS("raid6");