2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
58 #include <trace/events/block.h>
65 #define cpu_to_group(cpu) cpu_to_node(cpu)
66 #define ANY_GROUP NUMA_NO_NODE
68 static bool devices_handle_discard_safely = false;
69 module_param(devices_handle_discard_safely, bool, 0644);
70 MODULE_PARM_DESC(devices_handle_discard_safely,
71 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
72 static struct workqueue_struct *raid5_wq;
77 #define NR_STRIPES 256
78 #define STRIPE_SIZE PAGE_SIZE
79 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
80 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
81 #define IO_THRESHOLD 1
82 #define BYPASS_THRESHOLD 1
83 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
84 #define HASH_MASK (NR_HASH - 1)
85 #define MAX_STRIPE_BATCH 8
87 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
89 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
90 return &conf->stripe_hashtbl[hash];
93 static inline int stripe_hash_locks_hash(sector_t sect)
95 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
98 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
100 spin_lock_irq(conf->hash_locks + hash);
101 spin_lock(&conf->device_lock);
104 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
106 spin_unlock(&conf->device_lock);
107 spin_unlock_irq(conf->hash_locks + hash);
110 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
114 spin_lock(conf->hash_locks);
115 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
116 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
117 spin_lock(&conf->device_lock);
120 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
123 spin_unlock(&conf->device_lock);
124 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
125 spin_unlock(conf->hash_locks + i - 1);
129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
130 * order without overlap. There may be several bio's per stripe+device, and
131 * a bio could span several devices.
132 * When walking this list for a particular stripe+device, we must never proceed
133 * beyond a bio that extends past this device, as the next bio might no longer
135 * This function is used to determine the 'next' bio in the list, given the sector
136 * of the current stripe+device
138 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
140 int sectors = bio_sectors(bio);
141 if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
148 * We maintain a biased count of active stripes in the bottom 16 bits of
149 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
151 static inline int raid5_bi_processed_stripes(struct bio *bio)
153 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
154 return (atomic_read(segments) >> 16) & 0xffff;
157 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
159 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
160 return atomic_sub_return(1, segments) & 0xffff;
163 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
165 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
166 atomic_inc(segments);
169 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
172 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
176 old = atomic_read(segments);
177 new = (old & 0xffff) | (cnt << 16);
178 } while (atomic_cmpxchg(segments, old, new) != old);
181 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
183 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
184 atomic_set(segments, cnt);
187 /* Find first data disk in a raid6 stripe */
188 static inline int raid6_d0(struct stripe_head *sh)
191 /* ddf always start from first device */
193 /* md starts just after Q block */
194 if (sh->qd_idx == sh->disks - 1)
197 return sh->qd_idx + 1;
199 static inline int raid6_next_disk(int disk, int raid_disks)
202 return (disk < raid_disks) ? disk : 0;
205 /* When walking through the disks in a raid5, starting at raid6_d0,
206 * We need to map each disk to a 'slot', where the data disks are slot
207 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
208 * is raid_disks-1. This help does that mapping.
210 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
211 int *count, int syndrome_disks)
217 if (idx == sh->pd_idx)
218 return syndrome_disks;
219 if (idx == sh->qd_idx)
220 return syndrome_disks + 1;
226 static void return_io(struct bio_list *return_bi)
229 while ((bi = bio_list_pop(return_bi)) != NULL) {
230 bi->bi_iter.bi_size = 0;
231 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
237 static void print_raid5_conf (struct r5conf *conf);
239 static int stripe_operations_active(struct stripe_head *sh)
241 return sh->check_state || sh->reconstruct_state ||
242 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
243 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
246 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
248 struct r5conf *conf = sh->raid_conf;
249 struct r5worker_group *group;
251 int i, cpu = sh->cpu;
253 if (!cpu_online(cpu)) {
254 cpu = cpumask_any(cpu_online_mask);
258 if (list_empty(&sh->lru)) {
259 struct r5worker_group *group;
260 group = conf->worker_groups + cpu_to_group(cpu);
261 list_add_tail(&sh->lru, &group->handle_list);
262 group->stripes_cnt++;
266 if (conf->worker_cnt_per_group == 0) {
267 md_wakeup_thread(conf->mddev->thread);
271 group = conf->worker_groups + cpu_to_group(sh->cpu);
273 group->workers[0].working = true;
274 /* at least one worker should run to avoid race */
275 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
277 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
278 /* wakeup more workers */
279 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
280 if (group->workers[i].working == false) {
281 group->workers[i].working = true;
282 queue_work_on(sh->cpu, raid5_wq,
283 &group->workers[i].work);
289 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
290 struct list_head *temp_inactive_list)
292 BUG_ON(!list_empty(&sh->lru));
293 BUG_ON(atomic_read(&conf->active_stripes)==0);
294 if (test_bit(STRIPE_HANDLE, &sh->state)) {
295 if (test_bit(STRIPE_DELAYED, &sh->state) &&
296 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
297 list_add_tail(&sh->lru, &conf->delayed_list);
298 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
299 sh->bm_seq - conf->seq_write > 0)
300 list_add_tail(&sh->lru, &conf->bitmap_list);
302 clear_bit(STRIPE_DELAYED, &sh->state);
303 clear_bit(STRIPE_BIT_DELAY, &sh->state);
304 if (conf->worker_cnt_per_group == 0) {
305 list_add_tail(&sh->lru, &conf->handle_list);
307 raid5_wakeup_stripe_thread(sh);
311 md_wakeup_thread(conf->mddev->thread);
313 BUG_ON(stripe_operations_active(sh));
314 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
315 if (atomic_dec_return(&conf->preread_active_stripes)
317 md_wakeup_thread(conf->mddev->thread);
318 atomic_dec(&conf->active_stripes);
319 if (!test_bit(STRIPE_EXPANDING, &sh->state))
320 list_add_tail(&sh->lru, temp_inactive_list);
324 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
325 struct list_head *temp_inactive_list)
327 if (atomic_dec_and_test(&sh->count))
328 do_release_stripe(conf, sh, temp_inactive_list);
332 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
334 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
335 * given time. Adding stripes only takes device lock, while deleting stripes
336 * only takes hash lock.
338 static void release_inactive_stripe_list(struct r5conf *conf,
339 struct list_head *temp_inactive_list,
343 unsigned long do_wakeup = 0;
347 if (hash == NR_STRIPE_HASH_LOCKS) {
348 size = NR_STRIPE_HASH_LOCKS;
349 hash = NR_STRIPE_HASH_LOCKS - 1;
353 struct list_head *list = &temp_inactive_list[size - 1];
356 * We don't hold any lock here yet, raid5_get_active_stripe() might
357 * remove stripes from the list
359 if (!list_empty_careful(list)) {
360 spin_lock_irqsave(conf->hash_locks + hash, flags);
361 if (list_empty(conf->inactive_list + hash) &&
363 atomic_dec(&conf->empty_inactive_list_nr);
364 list_splice_tail_init(list, conf->inactive_list + hash);
365 do_wakeup |= 1 << hash;
366 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
372 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
373 if (do_wakeup & (1 << i))
374 wake_up(&conf->wait_for_stripe[i]);
378 if (atomic_read(&conf->active_stripes) == 0)
379 wake_up(&conf->wait_for_quiescent);
380 if (conf->retry_read_aligned)
381 md_wakeup_thread(conf->mddev->thread);
385 /* should hold conf->device_lock already */
386 static int release_stripe_list(struct r5conf *conf,
387 struct list_head *temp_inactive_list)
389 struct stripe_head *sh;
391 struct llist_node *head;
393 head = llist_del_all(&conf->released_stripes);
394 head = llist_reverse_order(head);
398 sh = llist_entry(head, struct stripe_head, release_list);
399 head = llist_next(head);
400 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
402 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
404 * Don't worry the bit is set here, because if the bit is set
405 * again, the count is always > 1. This is true for
406 * STRIPE_ON_UNPLUG_LIST bit too.
408 hash = sh->hash_lock_index;
409 __release_stripe(conf, sh, &temp_inactive_list[hash]);
416 void raid5_release_stripe(struct stripe_head *sh)
418 struct r5conf *conf = sh->raid_conf;
420 struct list_head list;
424 /* Avoid release_list until the last reference.
426 if (atomic_add_unless(&sh->count, -1, 1))
429 if (unlikely(!conf->mddev->thread) ||
430 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
432 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
434 md_wakeup_thread(conf->mddev->thread);
437 local_irq_save(flags);
438 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
439 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
440 INIT_LIST_HEAD(&list);
441 hash = sh->hash_lock_index;
442 do_release_stripe(conf, sh, &list);
443 spin_unlock(&conf->device_lock);
444 release_inactive_stripe_list(conf, &list, hash);
446 local_irq_restore(flags);
449 static inline void remove_hash(struct stripe_head *sh)
451 pr_debug("remove_hash(), stripe %llu\n",
452 (unsigned long long)sh->sector);
454 hlist_del_init(&sh->hash);
457 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
459 struct hlist_head *hp = stripe_hash(conf, sh->sector);
461 pr_debug("insert_hash(), stripe %llu\n",
462 (unsigned long long)sh->sector);
464 hlist_add_head(&sh->hash, hp);
467 /* find an idle stripe, make sure it is unhashed, and return it. */
468 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
470 struct stripe_head *sh = NULL;
471 struct list_head *first;
473 if (list_empty(conf->inactive_list + hash))
475 first = (conf->inactive_list + hash)->next;
476 sh = list_entry(first, struct stripe_head, lru);
477 list_del_init(first);
479 atomic_inc(&conf->active_stripes);
480 BUG_ON(hash != sh->hash_lock_index);
481 if (list_empty(conf->inactive_list + hash))
482 atomic_inc(&conf->empty_inactive_list_nr);
487 static void shrink_buffers(struct stripe_head *sh)
491 int num = sh->raid_conf->pool_size;
493 for (i = 0; i < num ; i++) {
494 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
498 sh->dev[i].page = NULL;
503 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
506 int num = sh->raid_conf->pool_size;
508 for (i = 0; i < num; i++) {
511 if (!(page = alloc_page(gfp))) {
514 sh->dev[i].page = page;
515 sh->dev[i].orig_page = page;
520 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
521 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
522 struct stripe_head *sh);
524 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
526 struct r5conf *conf = sh->raid_conf;
529 BUG_ON(atomic_read(&sh->count) != 0);
530 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
531 BUG_ON(stripe_operations_active(sh));
532 BUG_ON(sh->batch_head);
534 pr_debug("init_stripe called, stripe %llu\n",
535 (unsigned long long)sector);
537 seq = read_seqcount_begin(&conf->gen_lock);
538 sh->generation = conf->generation - previous;
539 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
541 stripe_set_idx(sector, conf, previous, sh);
544 for (i = sh->disks; i--; ) {
545 struct r5dev *dev = &sh->dev[i];
547 if (dev->toread || dev->read || dev->towrite || dev->written ||
548 test_bit(R5_LOCKED, &dev->flags)) {
549 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
550 (unsigned long long)sh->sector, i, dev->toread,
551 dev->read, dev->towrite, dev->written,
552 test_bit(R5_LOCKED, &dev->flags));
556 raid5_build_block(sh, i, previous);
558 if (read_seqcount_retry(&conf->gen_lock, seq))
560 sh->overwrite_disks = 0;
561 insert_hash(conf, sh);
562 sh->cpu = smp_processor_id();
563 set_bit(STRIPE_BATCH_READY, &sh->state);
566 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
569 struct stripe_head *sh;
571 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
572 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
573 if (sh->sector == sector && sh->generation == generation)
575 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
580 * Need to check if array has failed when deciding whether to:
582 * - remove non-faulty devices
585 * This determination is simple when no reshape is happening.
586 * However if there is a reshape, we need to carefully check
587 * both the before and after sections.
588 * This is because some failed devices may only affect one
589 * of the two sections, and some non-in_sync devices may
590 * be insync in the section most affected by failed devices.
592 static int calc_degraded(struct r5conf *conf)
594 int degraded, degraded2;
599 for (i = 0; i < conf->previous_raid_disks; i++) {
600 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
601 if (rdev && test_bit(Faulty, &rdev->flags))
602 rdev = rcu_dereference(conf->disks[i].replacement);
603 if (!rdev || test_bit(Faulty, &rdev->flags))
605 else if (test_bit(In_sync, &rdev->flags))
608 /* not in-sync or faulty.
609 * If the reshape increases the number of devices,
610 * this is being recovered by the reshape, so
611 * this 'previous' section is not in_sync.
612 * If the number of devices is being reduced however,
613 * the device can only be part of the array if
614 * we are reverting a reshape, so this section will
617 if (conf->raid_disks >= conf->previous_raid_disks)
621 if (conf->raid_disks == conf->previous_raid_disks)
625 for (i = 0; i < conf->raid_disks; i++) {
626 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
627 if (rdev && test_bit(Faulty, &rdev->flags))
628 rdev = rcu_dereference(conf->disks[i].replacement);
629 if (!rdev || test_bit(Faulty, &rdev->flags))
631 else if (test_bit(In_sync, &rdev->flags))
634 /* not in-sync or faulty.
635 * If reshape increases the number of devices, this
636 * section has already been recovered, else it
637 * almost certainly hasn't.
639 if (conf->raid_disks <= conf->previous_raid_disks)
643 if (degraded2 > degraded)
648 static int has_failed(struct r5conf *conf)
652 if (conf->mddev->reshape_position == MaxSector)
653 return conf->mddev->degraded > conf->max_degraded;
655 degraded = calc_degraded(conf);
656 if (degraded > conf->max_degraded)
662 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
663 int previous, int noblock, int noquiesce)
665 struct stripe_head *sh;
666 int hash = stripe_hash_locks_hash(sector);
668 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
670 spin_lock_irq(conf->hash_locks + hash);
673 wait_event_lock_irq(conf->wait_for_quiescent,
674 conf->quiesce == 0 || noquiesce,
675 *(conf->hash_locks + hash));
676 sh = __find_stripe(conf, sector, conf->generation - previous);
678 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
679 sh = get_free_stripe(conf, hash);
680 if (!sh && !test_bit(R5_DID_ALLOC,
682 set_bit(R5_ALLOC_MORE,
685 if (noblock && sh == NULL)
688 set_bit(R5_INACTIVE_BLOCKED,
690 wait_event_exclusive_cmd(
691 conf->wait_for_stripe[hash],
692 !list_empty(conf->inactive_list + hash) &&
693 (atomic_read(&conf->active_stripes)
694 < (conf->max_nr_stripes * 3 / 4)
695 || !test_bit(R5_INACTIVE_BLOCKED,
696 &conf->cache_state)),
697 spin_unlock_irq(conf->hash_locks + hash),
698 spin_lock_irq(conf->hash_locks + hash));
699 clear_bit(R5_INACTIVE_BLOCKED,
702 init_stripe(sh, sector, previous);
703 atomic_inc(&sh->count);
705 } else if (!atomic_inc_not_zero(&sh->count)) {
706 spin_lock(&conf->device_lock);
707 if (!atomic_read(&sh->count)) {
708 if (!test_bit(STRIPE_HANDLE, &sh->state))
709 atomic_inc(&conf->active_stripes);
710 BUG_ON(list_empty(&sh->lru) &&
711 !test_bit(STRIPE_EXPANDING, &sh->state));
712 list_del_init(&sh->lru);
714 sh->group->stripes_cnt--;
718 atomic_inc(&sh->count);
719 spin_unlock(&conf->device_lock);
721 } while (sh == NULL);
723 if (!list_empty(conf->inactive_list + hash))
724 wake_up(&conf->wait_for_stripe[hash]);
726 spin_unlock_irq(conf->hash_locks + hash);
730 static bool is_full_stripe_write(struct stripe_head *sh)
732 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
733 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
736 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
740 spin_lock(&sh2->stripe_lock);
741 spin_lock_nested(&sh1->stripe_lock, 1);
743 spin_lock(&sh1->stripe_lock);
744 spin_lock_nested(&sh2->stripe_lock, 1);
748 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
750 spin_unlock(&sh1->stripe_lock);
751 spin_unlock(&sh2->stripe_lock);
755 /* Only freshly new full stripe normal write stripe can be added to a batch list */
756 static bool stripe_can_batch(struct stripe_head *sh)
758 struct r5conf *conf = sh->raid_conf;
762 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
763 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
764 is_full_stripe_write(sh);
767 /* we only do back search */
768 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
770 struct stripe_head *head;
771 sector_t head_sector, tmp_sec;
775 if (!stripe_can_batch(sh))
777 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
778 tmp_sec = sh->sector;
779 if (!sector_div(tmp_sec, conf->chunk_sectors))
781 head_sector = sh->sector - STRIPE_SECTORS;
783 hash = stripe_hash_locks_hash(head_sector);
784 spin_lock_irq(conf->hash_locks + hash);
785 head = __find_stripe(conf, head_sector, conf->generation);
786 if (head && !atomic_inc_not_zero(&head->count)) {
787 spin_lock(&conf->device_lock);
788 if (!atomic_read(&head->count)) {
789 if (!test_bit(STRIPE_HANDLE, &head->state))
790 atomic_inc(&conf->active_stripes);
791 BUG_ON(list_empty(&head->lru) &&
792 !test_bit(STRIPE_EXPANDING, &head->state));
793 list_del_init(&head->lru);
795 head->group->stripes_cnt--;
799 atomic_inc(&head->count);
800 spin_unlock(&conf->device_lock);
802 spin_unlock_irq(conf->hash_locks + hash);
806 if (!stripe_can_batch(head))
809 lock_two_stripes(head, sh);
810 /* clear_batch_ready clear the flag */
811 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
818 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
820 if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
823 if (head->batch_head) {
824 spin_lock(&head->batch_head->batch_lock);
825 /* This batch list is already running */
826 if (!stripe_can_batch(head)) {
827 spin_unlock(&head->batch_head->batch_lock);
832 * at this point, head's BATCH_READY could be cleared, but we
833 * can still add the stripe to batch list
835 list_add(&sh->batch_list, &head->batch_list);
836 spin_unlock(&head->batch_head->batch_lock);
838 sh->batch_head = head->batch_head;
840 head->batch_head = head;
841 sh->batch_head = head->batch_head;
842 spin_lock(&head->batch_lock);
843 list_add_tail(&sh->batch_list, &head->batch_list);
844 spin_unlock(&head->batch_lock);
847 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
848 if (atomic_dec_return(&conf->preread_active_stripes)
850 md_wakeup_thread(conf->mddev->thread);
852 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
853 int seq = sh->bm_seq;
854 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
855 sh->batch_head->bm_seq > seq)
856 seq = sh->batch_head->bm_seq;
857 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
858 sh->batch_head->bm_seq = seq;
861 atomic_inc(&sh->count);
863 unlock_two_stripes(head, sh);
865 raid5_release_stripe(head);
868 /* Determine if 'data_offset' or 'new_data_offset' should be used
869 * in this stripe_head.
871 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
873 sector_t progress = conf->reshape_progress;
874 /* Need a memory barrier to make sure we see the value
875 * of conf->generation, or ->data_offset that was set before
876 * reshape_progress was updated.
879 if (progress == MaxSector)
881 if (sh->generation == conf->generation - 1)
883 /* We are in a reshape, and this is a new-generation stripe,
884 * so use new_data_offset.
890 raid5_end_read_request(struct bio *bi);
892 raid5_end_write_request(struct bio *bi);
894 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
896 struct r5conf *conf = sh->raid_conf;
897 int i, disks = sh->disks;
898 struct stripe_head *head_sh = sh;
902 if (r5l_write_stripe(conf->log, sh) == 0)
904 for (i = disks; i--; ) {
906 int replace_only = 0;
907 struct bio *bi, *rbi;
908 struct md_rdev *rdev, *rrdev = NULL;
911 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
912 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
916 if (test_bit(R5_Discard, &sh->dev[i].flags))
918 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
920 else if (test_and_clear_bit(R5_WantReplace,
921 &sh->dev[i].flags)) {
926 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
930 bi = &sh->dev[i].req;
931 rbi = &sh->dev[i].rreq; /* For writing to replacement */
934 rrdev = rcu_dereference(conf->disks[i].replacement);
935 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
936 rdev = rcu_dereference(conf->disks[i].rdev);
945 /* We raced and saw duplicates */
948 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
953 if (rdev && test_bit(Faulty, &rdev->flags))
956 atomic_inc(&rdev->nr_pending);
957 if (rrdev && test_bit(Faulty, &rrdev->flags))
960 atomic_inc(&rrdev->nr_pending);
963 /* We have already checked bad blocks for reads. Now
964 * need to check for writes. We never accept write errors
965 * on the replacement, so we don't to check rrdev.
967 while ((rw & WRITE) && rdev &&
968 test_bit(WriteErrorSeen, &rdev->flags)) {
971 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
972 &first_bad, &bad_sectors);
977 set_bit(BlockedBadBlocks, &rdev->flags);
978 if (!conf->mddev->external &&
979 conf->mddev->flags) {
980 /* It is very unlikely, but we might
981 * still need to write out the
982 * bad block log - better give it
984 md_check_recovery(conf->mddev);
987 * Because md_wait_for_blocked_rdev
988 * will dec nr_pending, we must
989 * increment it first.
991 atomic_inc(&rdev->nr_pending);
992 md_wait_for_blocked_rdev(rdev, conf->mddev);
994 /* Acknowledged bad block - skip the write */
995 rdev_dec_pending(rdev, conf->mddev);
1001 if (s->syncing || s->expanding || s->expanded
1003 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1005 set_bit(STRIPE_IO_STARTED, &sh->state);
1008 bi->bi_bdev = rdev->bdev;
1010 bi->bi_end_io = (rw & WRITE)
1011 ? raid5_end_write_request
1012 : raid5_end_read_request;
1013 bi->bi_private = sh;
1015 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1016 __func__, (unsigned long long)sh->sector,
1018 atomic_inc(&sh->count);
1020 atomic_inc(&head_sh->count);
1021 if (use_new_offset(conf, sh))
1022 bi->bi_iter.bi_sector = (sh->sector
1023 + rdev->new_data_offset);
1025 bi->bi_iter.bi_sector = (sh->sector
1026 + rdev->data_offset);
1027 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1028 bi->bi_rw |= REQ_NOMERGE;
1030 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1031 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1032 sh->dev[i].vec.bv_page = sh->dev[i].page;
1034 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1035 bi->bi_io_vec[0].bv_offset = 0;
1036 bi->bi_iter.bi_size = STRIPE_SIZE;
1038 * If this is discard request, set bi_vcnt 0. We don't
1039 * want to confuse SCSI because SCSI will replace payload
1041 if (rw & REQ_DISCARD)
1044 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1046 if (conf->mddev->gendisk)
1047 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1048 bi, disk_devt(conf->mddev->gendisk),
1050 generic_make_request(bi);
1053 if (s->syncing || s->expanding || s->expanded
1055 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1057 set_bit(STRIPE_IO_STARTED, &sh->state);
1060 rbi->bi_bdev = rrdev->bdev;
1062 BUG_ON(!(rw & WRITE));
1063 rbi->bi_end_io = raid5_end_write_request;
1064 rbi->bi_private = sh;
1066 pr_debug("%s: for %llu schedule op %ld on "
1067 "replacement disc %d\n",
1068 __func__, (unsigned long long)sh->sector,
1070 atomic_inc(&sh->count);
1072 atomic_inc(&head_sh->count);
1073 if (use_new_offset(conf, sh))
1074 rbi->bi_iter.bi_sector = (sh->sector
1075 + rrdev->new_data_offset);
1077 rbi->bi_iter.bi_sector = (sh->sector
1078 + rrdev->data_offset);
1079 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1080 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1081 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1083 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1084 rbi->bi_io_vec[0].bv_offset = 0;
1085 rbi->bi_iter.bi_size = STRIPE_SIZE;
1087 * If this is discard request, set bi_vcnt 0. We don't
1088 * want to confuse SCSI because SCSI will replace payload
1090 if (rw & REQ_DISCARD)
1092 if (conf->mddev->gendisk)
1093 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1094 rbi, disk_devt(conf->mddev->gendisk),
1096 generic_make_request(rbi);
1098 if (!rdev && !rrdev) {
1100 set_bit(STRIPE_DEGRADED, &sh->state);
1101 pr_debug("skip op %ld on disc %d for sector %llu\n",
1102 bi->bi_rw, i, (unsigned long long)sh->sector);
1103 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1104 set_bit(STRIPE_HANDLE, &sh->state);
1107 if (!head_sh->batch_head)
1109 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1116 static struct dma_async_tx_descriptor *
1117 async_copy_data(int frombio, struct bio *bio, struct page **page,
1118 sector_t sector, struct dma_async_tx_descriptor *tx,
1119 struct stripe_head *sh)
1122 struct bvec_iter iter;
1123 struct page *bio_page;
1125 struct async_submit_ctl submit;
1126 enum async_tx_flags flags = 0;
1128 if (bio->bi_iter.bi_sector >= sector)
1129 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1131 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1134 flags |= ASYNC_TX_FENCE;
1135 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1137 bio_for_each_segment(bvl, bio, iter) {
1138 int len = bvl.bv_len;
1142 if (page_offset < 0) {
1143 b_offset = -page_offset;
1144 page_offset += b_offset;
1148 if (len > 0 && page_offset + len > STRIPE_SIZE)
1149 clen = STRIPE_SIZE - page_offset;
1154 b_offset += bvl.bv_offset;
1155 bio_page = bvl.bv_page;
1157 if (sh->raid_conf->skip_copy &&
1158 b_offset == 0 && page_offset == 0 &&
1159 clen == STRIPE_SIZE)
1162 tx = async_memcpy(*page, bio_page, page_offset,
1163 b_offset, clen, &submit);
1165 tx = async_memcpy(bio_page, *page, b_offset,
1166 page_offset, clen, &submit);
1168 /* chain the operations */
1169 submit.depend_tx = tx;
1171 if (clen < len) /* hit end of page */
1179 static void ops_complete_biofill(void *stripe_head_ref)
1181 struct stripe_head *sh = stripe_head_ref;
1182 struct bio_list return_bi = BIO_EMPTY_LIST;
1185 pr_debug("%s: stripe %llu\n", __func__,
1186 (unsigned long long)sh->sector);
1188 /* clear completed biofills */
1189 for (i = sh->disks; i--; ) {
1190 struct r5dev *dev = &sh->dev[i];
1192 /* acknowledge completion of a biofill operation */
1193 /* and check if we need to reply to a read request,
1194 * new R5_Wantfill requests are held off until
1195 * !STRIPE_BIOFILL_RUN
1197 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1198 struct bio *rbi, *rbi2;
1203 while (rbi && rbi->bi_iter.bi_sector <
1204 dev->sector + STRIPE_SECTORS) {
1205 rbi2 = r5_next_bio(rbi, dev->sector);
1206 if (!raid5_dec_bi_active_stripes(rbi))
1207 bio_list_add(&return_bi, rbi);
1212 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1214 return_io(&return_bi);
1216 set_bit(STRIPE_HANDLE, &sh->state);
1217 raid5_release_stripe(sh);
1220 static void ops_run_biofill(struct stripe_head *sh)
1222 struct dma_async_tx_descriptor *tx = NULL;
1223 struct async_submit_ctl submit;
1226 BUG_ON(sh->batch_head);
1227 pr_debug("%s: stripe %llu\n", __func__,
1228 (unsigned long long)sh->sector);
1230 for (i = sh->disks; i--; ) {
1231 struct r5dev *dev = &sh->dev[i];
1232 if (test_bit(R5_Wantfill, &dev->flags)) {
1234 spin_lock_irq(&sh->stripe_lock);
1235 dev->read = rbi = dev->toread;
1237 spin_unlock_irq(&sh->stripe_lock);
1238 while (rbi && rbi->bi_iter.bi_sector <
1239 dev->sector + STRIPE_SECTORS) {
1240 tx = async_copy_data(0, rbi, &dev->page,
1241 dev->sector, tx, sh);
1242 rbi = r5_next_bio(rbi, dev->sector);
1247 atomic_inc(&sh->count);
1248 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1249 async_trigger_callback(&submit);
1252 static void mark_target_uptodate(struct stripe_head *sh, int target)
1259 tgt = &sh->dev[target];
1260 set_bit(R5_UPTODATE, &tgt->flags);
1261 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1262 clear_bit(R5_Wantcompute, &tgt->flags);
1265 static void ops_complete_compute(void *stripe_head_ref)
1267 struct stripe_head *sh = stripe_head_ref;
1269 pr_debug("%s: stripe %llu\n", __func__,
1270 (unsigned long long)sh->sector);
1272 /* mark the computed target(s) as uptodate */
1273 mark_target_uptodate(sh, sh->ops.target);
1274 mark_target_uptodate(sh, sh->ops.target2);
1276 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1277 if (sh->check_state == check_state_compute_run)
1278 sh->check_state = check_state_compute_result;
1279 set_bit(STRIPE_HANDLE, &sh->state);
1280 raid5_release_stripe(sh);
1283 /* return a pointer to the address conversion region of the scribble buffer */
1284 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1285 struct raid5_percpu *percpu, int i)
1289 addr = flex_array_get(percpu->scribble, i);
1290 return addr + sizeof(struct page *) * (sh->disks + 2);
1293 /* return a pointer to the address conversion region of the scribble buffer */
1294 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1298 addr = flex_array_get(percpu->scribble, i);
1302 static struct dma_async_tx_descriptor *
1303 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1305 int disks = sh->disks;
1306 struct page **xor_srcs = to_addr_page(percpu, 0);
1307 int target = sh->ops.target;
1308 struct r5dev *tgt = &sh->dev[target];
1309 struct page *xor_dest = tgt->page;
1311 struct dma_async_tx_descriptor *tx;
1312 struct async_submit_ctl submit;
1315 BUG_ON(sh->batch_head);
1317 pr_debug("%s: stripe %llu block: %d\n",
1318 __func__, (unsigned long long)sh->sector, target);
1319 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1321 for (i = disks; i--; )
1323 xor_srcs[count++] = sh->dev[i].page;
1325 atomic_inc(&sh->count);
1327 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1328 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1329 if (unlikely(count == 1))
1330 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1332 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1337 /* set_syndrome_sources - populate source buffers for gen_syndrome
1338 * @srcs - (struct page *) array of size sh->disks
1339 * @sh - stripe_head to parse
1341 * Populates srcs in proper layout order for the stripe and returns the
1342 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1343 * destination buffer is recorded in srcs[count] and the Q destination
1344 * is recorded in srcs[count+1]].
1346 static int set_syndrome_sources(struct page **srcs,
1347 struct stripe_head *sh,
1350 int disks = sh->disks;
1351 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1352 int d0_idx = raid6_d0(sh);
1356 for (i = 0; i < disks; i++)
1362 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1363 struct r5dev *dev = &sh->dev[i];
1365 if (i == sh->qd_idx || i == sh->pd_idx ||
1366 (srctype == SYNDROME_SRC_ALL) ||
1367 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1368 test_bit(R5_Wantdrain, &dev->flags)) ||
1369 (srctype == SYNDROME_SRC_WRITTEN &&
1371 srcs[slot] = sh->dev[i].page;
1372 i = raid6_next_disk(i, disks);
1373 } while (i != d0_idx);
1375 return syndrome_disks;
1378 static struct dma_async_tx_descriptor *
1379 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1381 int disks = sh->disks;
1382 struct page **blocks = to_addr_page(percpu, 0);
1384 int qd_idx = sh->qd_idx;
1385 struct dma_async_tx_descriptor *tx;
1386 struct async_submit_ctl submit;
1392 BUG_ON(sh->batch_head);
1393 if (sh->ops.target < 0)
1394 target = sh->ops.target2;
1395 else if (sh->ops.target2 < 0)
1396 target = sh->ops.target;
1398 /* we should only have one valid target */
1401 pr_debug("%s: stripe %llu block: %d\n",
1402 __func__, (unsigned long long)sh->sector, target);
1404 tgt = &sh->dev[target];
1405 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1408 atomic_inc(&sh->count);
1410 if (target == qd_idx) {
1411 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1412 blocks[count] = NULL; /* regenerating p is not necessary */
1413 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1414 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1415 ops_complete_compute, sh,
1416 to_addr_conv(sh, percpu, 0));
1417 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1419 /* Compute any data- or p-drive using XOR */
1421 for (i = disks; i-- ; ) {
1422 if (i == target || i == qd_idx)
1424 blocks[count++] = sh->dev[i].page;
1427 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1428 NULL, ops_complete_compute, sh,
1429 to_addr_conv(sh, percpu, 0));
1430 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1436 static struct dma_async_tx_descriptor *
1437 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1439 int i, count, disks = sh->disks;
1440 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1441 int d0_idx = raid6_d0(sh);
1442 int faila = -1, failb = -1;
1443 int target = sh->ops.target;
1444 int target2 = sh->ops.target2;
1445 struct r5dev *tgt = &sh->dev[target];
1446 struct r5dev *tgt2 = &sh->dev[target2];
1447 struct dma_async_tx_descriptor *tx;
1448 struct page **blocks = to_addr_page(percpu, 0);
1449 struct async_submit_ctl submit;
1451 BUG_ON(sh->batch_head);
1452 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1453 __func__, (unsigned long long)sh->sector, target, target2);
1454 BUG_ON(target < 0 || target2 < 0);
1455 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1456 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1458 /* we need to open-code set_syndrome_sources to handle the
1459 * slot number conversion for 'faila' and 'failb'
1461 for (i = 0; i < disks ; i++)
1466 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1468 blocks[slot] = sh->dev[i].page;
1474 i = raid6_next_disk(i, disks);
1475 } while (i != d0_idx);
1477 BUG_ON(faila == failb);
1480 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1481 __func__, (unsigned long long)sh->sector, faila, failb);
1483 atomic_inc(&sh->count);
1485 if (failb == syndrome_disks+1) {
1486 /* Q disk is one of the missing disks */
1487 if (faila == syndrome_disks) {
1488 /* Missing P+Q, just recompute */
1489 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1490 ops_complete_compute, sh,
1491 to_addr_conv(sh, percpu, 0));
1492 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1493 STRIPE_SIZE, &submit);
1497 int qd_idx = sh->qd_idx;
1499 /* Missing D+Q: recompute D from P, then recompute Q */
1500 if (target == qd_idx)
1501 data_target = target2;
1503 data_target = target;
1506 for (i = disks; i-- ; ) {
1507 if (i == data_target || i == qd_idx)
1509 blocks[count++] = sh->dev[i].page;
1511 dest = sh->dev[data_target].page;
1512 init_async_submit(&submit,
1513 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1515 to_addr_conv(sh, percpu, 0));
1516 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1519 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1520 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1521 ops_complete_compute, sh,
1522 to_addr_conv(sh, percpu, 0));
1523 return async_gen_syndrome(blocks, 0, count+2,
1524 STRIPE_SIZE, &submit);
1527 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1528 ops_complete_compute, sh,
1529 to_addr_conv(sh, percpu, 0));
1530 if (failb == syndrome_disks) {
1531 /* We're missing D+P. */
1532 return async_raid6_datap_recov(syndrome_disks+2,
1536 /* We're missing D+D. */
1537 return async_raid6_2data_recov(syndrome_disks+2,
1538 STRIPE_SIZE, faila, failb,
1544 static void ops_complete_prexor(void *stripe_head_ref)
1546 struct stripe_head *sh = stripe_head_ref;
1548 pr_debug("%s: stripe %llu\n", __func__,
1549 (unsigned long long)sh->sector);
1552 static struct dma_async_tx_descriptor *
1553 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1554 struct dma_async_tx_descriptor *tx)
1556 int disks = sh->disks;
1557 struct page **xor_srcs = to_addr_page(percpu, 0);
1558 int count = 0, pd_idx = sh->pd_idx, i;
1559 struct async_submit_ctl submit;
1561 /* existing parity data subtracted */
1562 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1564 BUG_ON(sh->batch_head);
1565 pr_debug("%s: stripe %llu\n", __func__,
1566 (unsigned long long)sh->sector);
1568 for (i = disks; i--; ) {
1569 struct r5dev *dev = &sh->dev[i];
1570 /* Only process blocks that are known to be uptodate */
1571 if (test_bit(R5_Wantdrain, &dev->flags))
1572 xor_srcs[count++] = dev->page;
1575 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1576 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1577 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1582 static struct dma_async_tx_descriptor *
1583 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1584 struct dma_async_tx_descriptor *tx)
1586 struct page **blocks = to_addr_page(percpu, 0);
1588 struct async_submit_ctl submit;
1590 pr_debug("%s: stripe %llu\n", __func__,
1591 (unsigned long long)sh->sector);
1593 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1595 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1596 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1597 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1602 static struct dma_async_tx_descriptor *
1603 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1605 int disks = sh->disks;
1607 struct stripe_head *head_sh = sh;
1609 pr_debug("%s: stripe %llu\n", __func__,
1610 (unsigned long long)sh->sector);
1612 for (i = disks; i--; ) {
1617 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1622 spin_lock_irq(&sh->stripe_lock);
1623 chosen = dev->towrite;
1624 dev->towrite = NULL;
1625 sh->overwrite_disks = 0;
1626 BUG_ON(dev->written);
1627 wbi = dev->written = chosen;
1628 spin_unlock_irq(&sh->stripe_lock);
1629 WARN_ON(dev->page != dev->orig_page);
1631 while (wbi && wbi->bi_iter.bi_sector <
1632 dev->sector + STRIPE_SECTORS) {
1633 if (wbi->bi_rw & REQ_FUA)
1634 set_bit(R5_WantFUA, &dev->flags);
1635 if (wbi->bi_rw & REQ_SYNC)
1636 set_bit(R5_SyncIO, &dev->flags);
1637 if (wbi->bi_rw & REQ_DISCARD)
1638 set_bit(R5_Discard, &dev->flags);
1640 tx = async_copy_data(1, wbi, &dev->page,
1641 dev->sector, tx, sh);
1642 if (dev->page != dev->orig_page) {
1643 set_bit(R5_SkipCopy, &dev->flags);
1644 clear_bit(R5_UPTODATE, &dev->flags);
1645 clear_bit(R5_OVERWRITE, &dev->flags);
1648 wbi = r5_next_bio(wbi, dev->sector);
1651 if (head_sh->batch_head) {
1652 sh = list_first_entry(&sh->batch_list,
1665 static void ops_complete_reconstruct(void *stripe_head_ref)
1667 struct stripe_head *sh = stripe_head_ref;
1668 int disks = sh->disks;
1669 int pd_idx = sh->pd_idx;
1670 int qd_idx = sh->qd_idx;
1672 bool fua = false, sync = false, discard = false;
1674 pr_debug("%s: stripe %llu\n", __func__,
1675 (unsigned long long)sh->sector);
1677 for (i = disks; i--; ) {
1678 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1679 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1680 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1683 for (i = disks; i--; ) {
1684 struct r5dev *dev = &sh->dev[i];
1686 if (dev->written || i == pd_idx || i == qd_idx) {
1687 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1688 set_bit(R5_UPTODATE, &dev->flags);
1690 set_bit(R5_WantFUA, &dev->flags);
1692 set_bit(R5_SyncIO, &dev->flags);
1696 if (sh->reconstruct_state == reconstruct_state_drain_run)
1697 sh->reconstruct_state = reconstruct_state_drain_result;
1698 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1699 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1701 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1702 sh->reconstruct_state = reconstruct_state_result;
1705 set_bit(STRIPE_HANDLE, &sh->state);
1706 raid5_release_stripe(sh);
1710 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1711 struct dma_async_tx_descriptor *tx)
1713 int disks = sh->disks;
1714 struct page **xor_srcs;
1715 struct async_submit_ctl submit;
1716 int count, pd_idx = sh->pd_idx, i;
1717 struct page *xor_dest;
1719 unsigned long flags;
1721 struct stripe_head *head_sh = sh;
1724 pr_debug("%s: stripe %llu\n", __func__,
1725 (unsigned long long)sh->sector);
1727 for (i = 0; i < sh->disks; i++) {
1730 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1733 if (i >= sh->disks) {
1734 atomic_inc(&sh->count);
1735 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1736 ops_complete_reconstruct(sh);
1741 xor_srcs = to_addr_page(percpu, j);
1742 /* check if prexor is active which means only process blocks
1743 * that are part of a read-modify-write (written)
1745 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1747 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1748 for (i = disks; i--; ) {
1749 struct r5dev *dev = &sh->dev[i];
1750 if (head_sh->dev[i].written)
1751 xor_srcs[count++] = dev->page;
1754 xor_dest = sh->dev[pd_idx].page;
1755 for (i = disks; i--; ) {
1756 struct r5dev *dev = &sh->dev[i];
1758 xor_srcs[count++] = dev->page;
1762 /* 1/ if we prexor'd then the dest is reused as a source
1763 * 2/ if we did not prexor then we are redoing the parity
1764 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1765 * for the synchronous xor case
1767 last_stripe = !head_sh->batch_head ||
1768 list_first_entry(&sh->batch_list,
1769 struct stripe_head, batch_list) == head_sh;
1771 flags = ASYNC_TX_ACK |
1772 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1774 atomic_inc(&head_sh->count);
1775 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1776 to_addr_conv(sh, percpu, j));
1778 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1779 init_async_submit(&submit, flags, tx, NULL, NULL,
1780 to_addr_conv(sh, percpu, j));
1783 if (unlikely(count == 1))
1784 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1786 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1789 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1796 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1797 struct dma_async_tx_descriptor *tx)
1799 struct async_submit_ctl submit;
1800 struct page **blocks;
1801 int count, i, j = 0;
1802 struct stripe_head *head_sh = sh;
1805 unsigned long txflags;
1807 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1809 for (i = 0; i < sh->disks; i++) {
1810 if (sh->pd_idx == i || sh->qd_idx == i)
1812 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1815 if (i >= sh->disks) {
1816 atomic_inc(&sh->count);
1817 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1818 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1819 ops_complete_reconstruct(sh);
1824 blocks = to_addr_page(percpu, j);
1826 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1827 synflags = SYNDROME_SRC_WRITTEN;
1828 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1830 synflags = SYNDROME_SRC_ALL;
1831 txflags = ASYNC_TX_ACK;
1834 count = set_syndrome_sources(blocks, sh, synflags);
1835 last_stripe = !head_sh->batch_head ||
1836 list_first_entry(&sh->batch_list,
1837 struct stripe_head, batch_list) == head_sh;
1840 atomic_inc(&head_sh->count);
1841 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1842 head_sh, to_addr_conv(sh, percpu, j));
1844 init_async_submit(&submit, 0, tx, NULL, NULL,
1845 to_addr_conv(sh, percpu, j));
1846 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1849 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1855 static void ops_complete_check(void *stripe_head_ref)
1857 struct stripe_head *sh = stripe_head_ref;
1859 pr_debug("%s: stripe %llu\n", __func__,
1860 (unsigned long long)sh->sector);
1862 sh->check_state = check_state_check_result;
1863 set_bit(STRIPE_HANDLE, &sh->state);
1864 raid5_release_stripe(sh);
1867 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1869 int disks = sh->disks;
1870 int pd_idx = sh->pd_idx;
1871 int qd_idx = sh->qd_idx;
1872 struct page *xor_dest;
1873 struct page **xor_srcs = to_addr_page(percpu, 0);
1874 struct dma_async_tx_descriptor *tx;
1875 struct async_submit_ctl submit;
1879 pr_debug("%s: stripe %llu\n", __func__,
1880 (unsigned long long)sh->sector);
1882 BUG_ON(sh->batch_head);
1884 xor_dest = sh->dev[pd_idx].page;
1885 xor_srcs[count++] = xor_dest;
1886 for (i = disks; i--; ) {
1887 if (i == pd_idx || i == qd_idx)
1889 xor_srcs[count++] = sh->dev[i].page;
1892 init_async_submit(&submit, 0, NULL, NULL, NULL,
1893 to_addr_conv(sh, percpu, 0));
1894 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1895 &sh->ops.zero_sum_result, &submit);
1897 atomic_inc(&sh->count);
1898 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1899 tx = async_trigger_callback(&submit);
1902 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1904 struct page **srcs = to_addr_page(percpu, 0);
1905 struct async_submit_ctl submit;
1908 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1909 (unsigned long long)sh->sector, checkp);
1911 BUG_ON(sh->batch_head);
1912 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1916 atomic_inc(&sh->count);
1917 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1918 sh, to_addr_conv(sh, percpu, 0));
1919 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1920 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1923 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1925 int overlap_clear = 0, i, disks = sh->disks;
1926 struct dma_async_tx_descriptor *tx = NULL;
1927 struct r5conf *conf = sh->raid_conf;
1928 int level = conf->level;
1929 struct raid5_percpu *percpu;
1932 cpu = get_cpu_light();
1933 percpu = per_cpu_ptr(conf->percpu, cpu);
1934 spin_lock(&percpu->lock);
1935 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1936 ops_run_biofill(sh);
1940 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1942 tx = ops_run_compute5(sh, percpu);
1944 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1945 tx = ops_run_compute6_1(sh, percpu);
1947 tx = ops_run_compute6_2(sh, percpu);
1949 /* terminate the chain if reconstruct is not set to be run */
1950 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1954 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1956 tx = ops_run_prexor5(sh, percpu, tx);
1958 tx = ops_run_prexor6(sh, percpu, tx);
1961 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1962 tx = ops_run_biodrain(sh, tx);
1966 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1968 ops_run_reconstruct5(sh, percpu, tx);
1970 ops_run_reconstruct6(sh, percpu, tx);
1973 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1974 if (sh->check_state == check_state_run)
1975 ops_run_check_p(sh, percpu);
1976 else if (sh->check_state == check_state_run_q)
1977 ops_run_check_pq(sh, percpu, 0);
1978 else if (sh->check_state == check_state_run_pq)
1979 ops_run_check_pq(sh, percpu, 1);
1984 if (overlap_clear && !sh->batch_head)
1985 for (i = disks; i--; ) {
1986 struct r5dev *dev = &sh->dev[i];
1987 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1988 wake_up(&sh->raid_conf->wait_for_overlap);
1990 spin_unlock(&percpu->lock);
1994 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1996 struct stripe_head *sh;
1998 sh = kmem_cache_zalloc(sc, gfp);
2000 spin_lock_init(&sh->stripe_lock);
2001 spin_lock_init(&sh->batch_lock);
2002 INIT_LIST_HEAD(&sh->batch_list);
2003 INIT_LIST_HEAD(&sh->lru);
2004 atomic_set(&sh->count, 1);
2008 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2010 struct stripe_head *sh;
2012 sh = alloc_stripe(conf->slab_cache, gfp);
2016 sh->raid_conf = conf;
2018 if (grow_buffers(sh, gfp)) {
2020 kmem_cache_free(conf->slab_cache, sh);
2023 sh->hash_lock_index =
2024 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2025 /* we just created an active stripe so... */
2026 atomic_inc(&conf->active_stripes);
2028 raid5_release_stripe(sh);
2029 conf->max_nr_stripes++;
2033 static int grow_stripes(struct r5conf *conf, int num)
2035 struct kmem_cache *sc;
2036 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2038 if (conf->mddev->gendisk)
2039 sprintf(conf->cache_name[0],
2040 "raid%d-%s", conf->level, mdname(conf->mddev));
2042 sprintf(conf->cache_name[0],
2043 "raid%d-%p", conf->level, conf->mddev);
2044 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2046 conf->active_name = 0;
2047 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2048 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2052 conf->slab_cache = sc;
2053 conf->pool_size = devs;
2055 if (!grow_one_stripe(conf, GFP_KERNEL))
2062 * scribble_len - return the required size of the scribble region
2063 * @num - total number of disks in the array
2065 * The size must be enough to contain:
2066 * 1/ a struct page pointer for each device in the array +2
2067 * 2/ room to convert each entry in (1) to its corresponding dma
2068 * (dma_map_page()) or page (page_address()) address.
2070 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2071 * calculate over all devices (not just the data blocks), using zeros in place
2072 * of the P and Q blocks.
2074 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2076 struct flex_array *ret;
2079 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2080 ret = flex_array_alloc(len, cnt, flags);
2083 /* always prealloc all elements, so no locking is required */
2084 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2085 flex_array_free(ret);
2091 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2096 mddev_suspend(conf->mddev);
2098 for_each_present_cpu(cpu) {
2099 struct raid5_percpu *percpu;
2100 struct flex_array *scribble;
2102 percpu = per_cpu_ptr(conf->percpu, cpu);
2103 scribble = scribble_alloc(new_disks,
2104 new_sectors / STRIPE_SECTORS,
2108 flex_array_free(percpu->scribble);
2109 percpu->scribble = scribble;
2116 mddev_resume(conf->mddev);
2120 static int resize_stripes(struct r5conf *conf, int newsize)
2122 /* Make all the stripes able to hold 'newsize' devices.
2123 * New slots in each stripe get 'page' set to a new page.
2125 * This happens in stages:
2126 * 1/ create a new kmem_cache and allocate the required number of
2128 * 2/ gather all the old stripe_heads and transfer the pages across
2129 * to the new stripe_heads. This will have the side effect of
2130 * freezing the array as once all stripe_heads have been collected,
2131 * no IO will be possible. Old stripe heads are freed once their
2132 * pages have been transferred over, and the old kmem_cache is
2133 * freed when all stripes are done.
2134 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2135 * we simple return a failre status - no need to clean anything up.
2136 * 4/ allocate new pages for the new slots in the new stripe_heads.
2137 * If this fails, we don't bother trying the shrink the
2138 * stripe_heads down again, we just leave them as they are.
2139 * As each stripe_head is processed the new one is released into
2142 * Once step2 is started, we cannot afford to wait for a write,
2143 * so we use GFP_NOIO allocations.
2145 struct stripe_head *osh, *nsh;
2146 LIST_HEAD(newstripes);
2147 struct disk_info *ndisks;
2149 struct kmem_cache *sc;
2153 if (newsize <= conf->pool_size)
2154 return 0; /* never bother to shrink */
2156 err = md_allow_write(conf->mddev);
2161 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2162 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2167 /* Need to ensure auto-resizing doesn't interfere */
2168 mutex_lock(&conf->cache_size_mutex);
2170 for (i = conf->max_nr_stripes; i; i--) {
2171 nsh = alloc_stripe(sc, GFP_KERNEL);
2175 nsh->raid_conf = conf;
2176 list_add(&nsh->lru, &newstripes);
2179 /* didn't get enough, give up */
2180 while (!list_empty(&newstripes)) {
2181 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2182 list_del(&nsh->lru);
2183 kmem_cache_free(sc, nsh);
2185 kmem_cache_destroy(sc);
2186 mutex_unlock(&conf->cache_size_mutex);
2189 /* Step 2 - Must use GFP_NOIO now.
2190 * OK, we have enough stripes, start collecting inactive
2191 * stripes and copying them over
2195 list_for_each_entry(nsh, &newstripes, lru) {
2196 lock_device_hash_lock(conf, hash);
2197 wait_event_exclusive_cmd(conf->wait_for_stripe[hash],
2198 !list_empty(conf->inactive_list + hash),
2199 unlock_device_hash_lock(conf, hash),
2200 lock_device_hash_lock(conf, hash));
2201 osh = get_free_stripe(conf, hash);
2202 unlock_device_hash_lock(conf, hash);
2204 for(i=0; i<conf->pool_size; i++) {
2205 nsh->dev[i].page = osh->dev[i].page;
2206 nsh->dev[i].orig_page = osh->dev[i].page;
2208 nsh->hash_lock_index = hash;
2209 kmem_cache_free(conf->slab_cache, osh);
2211 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2212 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2217 kmem_cache_destroy(conf->slab_cache);
2220 * At this point, we are holding all the stripes so the array
2221 * is completely stalled, so now is a good time to resize
2222 * conf->disks and the scribble region
2224 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2226 for (i=0; i<conf->raid_disks; i++)
2227 ndisks[i] = conf->disks[i];
2229 conf->disks = ndisks;
2233 mutex_unlock(&conf->cache_size_mutex);
2234 /* Step 4, return new stripes to service */
2235 while(!list_empty(&newstripes)) {
2236 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2237 list_del_init(&nsh->lru);
2239 for (i=conf->raid_disks; i < newsize; i++)
2240 if (nsh->dev[i].page == NULL) {
2241 struct page *p = alloc_page(GFP_NOIO);
2242 nsh->dev[i].page = p;
2243 nsh->dev[i].orig_page = p;
2247 raid5_release_stripe(nsh);
2249 /* critical section pass, GFP_NOIO no longer needed */
2251 conf->slab_cache = sc;
2252 conf->active_name = 1-conf->active_name;
2254 conf->pool_size = newsize;
2258 static int drop_one_stripe(struct r5conf *conf)
2260 struct stripe_head *sh;
2261 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2263 spin_lock_irq(conf->hash_locks + hash);
2264 sh = get_free_stripe(conf, hash);
2265 spin_unlock_irq(conf->hash_locks + hash);
2268 BUG_ON(atomic_read(&sh->count));
2270 kmem_cache_free(conf->slab_cache, sh);
2271 atomic_dec(&conf->active_stripes);
2272 conf->max_nr_stripes--;
2276 static void shrink_stripes(struct r5conf *conf)
2278 while (conf->max_nr_stripes &&
2279 drop_one_stripe(conf))
2282 kmem_cache_destroy(conf->slab_cache);
2283 conf->slab_cache = NULL;
2286 static void raid5_end_read_request(struct bio * bi)
2288 struct stripe_head *sh = bi->bi_private;
2289 struct r5conf *conf = sh->raid_conf;
2290 int disks = sh->disks, i;
2291 char b[BDEVNAME_SIZE];
2292 struct md_rdev *rdev = NULL;
2295 for (i=0 ; i<disks; i++)
2296 if (bi == &sh->dev[i].req)
2299 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2300 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2306 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2307 /* If replacement finished while this request was outstanding,
2308 * 'replacement' might be NULL already.
2309 * In that case it moved down to 'rdev'.
2310 * rdev is not removed until all requests are finished.
2312 rdev = conf->disks[i].replacement;
2314 rdev = conf->disks[i].rdev;
2316 if (use_new_offset(conf, sh))
2317 s = sh->sector + rdev->new_data_offset;
2319 s = sh->sector + rdev->data_offset;
2320 if (!bi->bi_error) {
2321 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2322 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2323 /* Note that this cannot happen on a
2324 * replacement device. We just fail those on
2329 "md/raid:%s: read error corrected"
2330 " (%lu sectors at %llu on %s)\n",
2331 mdname(conf->mddev), STRIPE_SECTORS,
2332 (unsigned long long)s,
2333 bdevname(rdev->bdev, b));
2334 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2335 clear_bit(R5_ReadError, &sh->dev[i].flags);
2336 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2337 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2338 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2340 if (atomic_read(&rdev->read_errors))
2341 atomic_set(&rdev->read_errors, 0);
2343 const char *bdn = bdevname(rdev->bdev, b);
2347 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2348 atomic_inc(&rdev->read_errors);
2349 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2352 "md/raid:%s: read error on replacement device "
2353 "(sector %llu on %s).\n",
2354 mdname(conf->mddev),
2355 (unsigned long long)s,
2357 else if (conf->mddev->degraded >= conf->max_degraded) {
2361 "md/raid:%s: read error not correctable "
2362 "(sector %llu on %s).\n",
2363 mdname(conf->mddev),
2364 (unsigned long long)s,
2366 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2371 "md/raid:%s: read error NOT corrected!! "
2372 "(sector %llu on %s).\n",
2373 mdname(conf->mddev),
2374 (unsigned long long)s,
2376 } else if (atomic_read(&rdev->read_errors)
2377 > conf->max_nr_stripes)
2379 "md/raid:%s: Too many read errors, failing device %s.\n",
2380 mdname(conf->mddev), bdn);
2383 if (set_bad && test_bit(In_sync, &rdev->flags)
2384 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2387 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2388 set_bit(R5_ReadError, &sh->dev[i].flags);
2389 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2391 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2393 clear_bit(R5_ReadError, &sh->dev[i].flags);
2394 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2396 && test_bit(In_sync, &rdev->flags)
2397 && rdev_set_badblocks(
2398 rdev, sh->sector, STRIPE_SECTORS, 0)))
2399 md_error(conf->mddev, rdev);
2402 rdev_dec_pending(rdev, conf->mddev);
2403 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2404 set_bit(STRIPE_HANDLE, &sh->state);
2405 raid5_release_stripe(sh);
2408 static void raid5_end_write_request(struct bio *bi)
2410 struct stripe_head *sh = bi->bi_private;
2411 struct r5conf *conf = sh->raid_conf;
2412 int disks = sh->disks, i;
2413 struct md_rdev *uninitialized_var(rdev);
2416 int replacement = 0;
2418 for (i = 0 ; i < disks; i++) {
2419 if (bi == &sh->dev[i].req) {
2420 rdev = conf->disks[i].rdev;
2423 if (bi == &sh->dev[i].rreq) {
2424 rdev = conf->disks[i].replacement;
2428 /* rdev was removed and 'replacement'
2429 * replaced it. rdev is not removed
2430 * until all requests are finished.
2432 rdev = conf->disks[i].rdev;
2436 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2437 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2446 md_error(conf->mddev, rdev);
2447 else if (is_badblock(rdev, sh->sector,
2449 &first_bad, &bad_sectors))
2450 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2453 set_bit(STRIPE_DEGRADED, &sh->state);
2454 set_bit(WriteErrorSeen, &rdev->flags);
2455 set_bit(R5_WriteError, &sh->dev[i].flags);
2456 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2457 set_bit(MD_RECOVERY_NEEDED,
2458 &rdev->mddev->recovery);
2459 } else if (is_badblock(rdev, sh->sector,
2461 &first_bad, &bad_sectors)) {
2462 set_bit(R5_MadeGood, &sh->dev[i].flags);
2463 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2464 /* That was a successful write so make
2465 * sure it looks like we already did
2468 set_bit(R5_ReWrite, &sh->dev[i].flags);
2471 rdev_dec_pending(rdev, conf->mddev);
2473 if (sh->batch_head && bi->bi_error && !replacement)
2474 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2476 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2477 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2478 set_bit(STRIPE_HANDLE, &sh->state);
2479 raid5_release_stripe(sh);
2481 if (sh->batch_head && sh != sh->batch_head)
2482 raid5_release_stripe(sh->batch_head);
2485 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2487 struct r5dev *dev = &sh->dev[i];
2489 bio_init(&dev->req);
2490 dev->req.bi_io_vec = &dev->vec;
2491 dev->req.bi_max_vecs = 1;
2492 dev->req.bi_private = sh;
2494 bio_init(&dev->rreq);
2495 dev->rreq.bi_io_vec = &dev->rvec;
2496 dev->rreq.bi_max_vecs = 1;
2497 dev->rreq.bi_private = sh;
2500 dev->sector = raid5_compute_blocknr(sh, i, previous);
2503 static void error(struct mddev *mddev, struct md_rdev *rdev)
2505 char b[BDEVNAME_SIZE];
2506 struct r5conf *conf = mddev->private;
2507 unsigned long flags;
2508 pr_debug("raid456: error called\n");
2510 spin_lock_irqsave(&conf->device_lock, flags);
2511 clear_bit(In_sync, &rdev->flags);
2512 mddev->degraded = calc_degraded(conf);
2513 spin_unlock_irqrestore(&conf->device_lock, flags);
2514 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2516 set_bit(Blocked, &rdev->flags);
2517 set_bit(Faulty, &rdev->flags);
2518 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2519 set_bit(MD_CHANGE_PENDING, &mddev->flags);
2521 "md/raid:%s: Disk failure on %s, disabling device.\n"
2522 "md/raid:%s: Operation continuing on %d devices.\n",
2524 bdevname(rdev->bdev, b),
2526 conf->raid_disks - mddev->degraded);
2530 * Input: a 'big' sector number,
2531 * Output: index of the data and parity disk, and the sector # in them.
2533 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2534 int previous, int *dd_idx,
2535 struct stripe_head *sh)
2537 sector_t stripe, stripe2;
2538 sector_t chunk_number;
2539 unsigned int chunk_offset;
2542 sector_t new_sector;
2543 int algorithm = previous ? conf->prev_algo
2545 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2546 : conf->chunk_sectors;
2547 int raid_disks = previous ? conf->previous_raid_disks
2549 int data_disks = raid_disks - conf->max_degraded;
2551 /* First compute the information on this sector */
2554 * Compute the chunk number and the sector offset inside the chunk
2556 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2557 chunk_number = r_sector;
2560 * Compute the stripe number
2562 stripe = chunk_number;
2563 *dd_idx = sector_div(stripe, data_disks);
2566 * Select the parity disk based on the user selected algorithm.
2568 pd_idx = qd_idx = -1;
2569 switch(conf->level) {
2571 pd_idx = data_disks;
2574 switch (algorithm) {
2575 case ALGORITHM_LEFT_ASYMMETRIC:
2576 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2577 if (*dd_idx >= pd_idx)
2580 case ALGORITHM_RIGHT_ASYMMETRIC:
2581 pd_idx = sector_div(stripe2, raid_disks);
2582 if (*dd_idx >= pd_idx)
2585 case ALGORITHM_LEFT_SYMMETRIC:
2586 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2587 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2589 case ALGORITHM_RIGHT_SYMMETRIC:
2590 pd_idx = sector_div(stripe2, raid_disks);
2591 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2593 case ALGORITHM_PARITY_0:
2597 case ALGORITHM_PARITY_N:
2598 pd_idx = data_disks;
2606 switch (algorithm) {
2607 case ALGORITHM_LEFT_ASYMMETRIC:
2608 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2609 qd_idx = pd_idx + 1;
2610 if (pd_idx == raid_disks-1) {
2611 (*dd_idx)++; /* Q D D D P */
2613 } else if (*dd_idx >= pd_idx)
2614 (*dd_idx) += 2; /* D D P Q D */
2616 case ALGORITHM_RIGHT_ASYMMETRIC:
2617 pd_idx = sector_div(stripe2, raid_disks);
2618 qd_idx = pd_idx + 1;
2619 if (pd_idx == raid_disks-1) {
2620 (*dd_idx)++; /* Q D D D P */
2622 } else if (*dd_idx >= pd_idx)
2623 (*dd_idx) += 2; /* D D P Q D */
2625 case ALGORITHM_LEFT_SYMMETRIC:
2626 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2627 qd_idx = (pd_idx + 1) % raid_disks;
2628 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2630 case ALGORITHM_RIGHT_SYMMETRIC:
2631 pd_idx = sector_div(stripe2, raid_disks);
2632 qd_idx = (pd_idx + 1) % raid_disks;
2633 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2636 case ALGORITHM_PARITY_0:
2641 case ALGORITHM_PARITY_N:
2642 pd_idx = data_disks;
2643 qd_idx = data_disks + 1;
2646 case ALGORITHM_ROTATING_ZERO_RESTART:
2647 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2648 * of blocks for computing Q is different.
2650 pd_idx = sector_div(stripe2, raid_disks);
2651 qd_idx = pd_idx + 1;
2652 if (pd_idx == raid_disks-1) {
2653 (*dd_idx)++; /* Q D D D P */
2655 } else if (*dd_idx >= pd_idx)
2656 (*dd_idx) += 2; /* D D P Q D */
2660 case ALGORITHM_ROTATING_N_RESTART:
2661 /* Same a left_asymmetric, by first stripe is
2662 * D D D P Q rather than
2666 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2667 qd_idx = pd_idx + 1;
2668 if (pd_idx == raid_disks-1) {
2669 (*dd_idx)++; /* Q D D D P */
2671 } else if (*dd_idx >= pd_idx)
2672 (*dd_idx) += 2; /* D D P Q D */
2676 case ALGORITHM_ROTATING_N_CONTINUE:
2677 /* Same as left_symmetric but Q is before P */
2678 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2679 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2680 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2684 case ALGORITHM_LEFT_ASYMMETRIC_6:
2685 /* RAID5 left_asymmetric, with Q on last device */
2686 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2687 if (*dd_idx >= pd_idx)
2689 qd_idx = raid_disks - 1;
2692 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2693 pd_idx = sector_div(stripe2, raid_disks-1);
2694 if (*dd_idx >= pd_idx)
2696 qd_idx = raid_disks - 1;
2699 case ALGORITHM_LEFT_SYMMETRIC_6:
2700 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2701 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2702 qd_idx = raid_disks - 1;
2705 case ALGORITHM_RIGHT_SYMMETRIC_6:
2706 pd_idx = sector_div(stripe2, raid_disks-1);
2707 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2708 qd_idx = raid_disks - 1;
2711 case ALGORITHM_PARITY_0_6:
2714 qd_idx = raid_disks - 1;
2724 sh->pd_idx = pd_idx;
2725 sh->qd_idx = qd_idx;
2726 sh->ddf_layout = ddf_layout;
2729 * Finally, compute the new sector number
2731 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2735 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2737 struct r5conf *conf = sh->raid_conf;
2738 int raid_disks = sh->disks;
2739 int data_disks = raid_disks - conf->max_degraded;
2740 sector_t new_sector = sh->sector, check;
2741 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2742 : conf->chunk_sectors;
2743 int algorithm = previous ? conf->prev_algo
2747 sector_t chunk_number;
2748 int dummy1, dd_idx = i;
2750 struct stripe_head sh2;
2752 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2753 stripe = new_sector;
2755 if (i == sh->pd_idx)
2757 switch(conf->level) {
2760 switch (algorithm) {
2761 case ALGORITHM_LEFT_ASYMMETRIC:
2762 case ALGORITHM_RIGHT_ASYMMETRIC:
2766 case ALGORITHM_LEFT_SYMMETRIC:
2767 case ALGORITHM_RIGHT_SYMMETRIC:
2770 i -= (sh->pd_idx + 1);
2772 case ALGORITHM_PARITY_0:
2775 case ALGORITHM_PARITY_N:
2782 if (i == sh->qd_idx)
2783 return 0; /* It is the Q disk */
2784 switch (algorithm) {
2785 case ALGORITHM_LEFT_ASYMMETRIC:
2786 case ALGORITHM_RIGHT_ASYMMETRIC:
2787 case ALGORITHM_ROTATING_ZERO_RESTART:
2788 case ALGORITHM_ROTATING_N_RESTART:
2789 if (sh->pd_idx == raid_disks-1)
2790 i--; /* Q D D D P */
2791 else if (i > sh->pd_idx)
2792 i -= 2; /* D D P Q D */
2794 case ALGORITHM_LEFT_SYMMETRIC:
2795 case ALGORITHM_RIGHT_SYMMETRIC:
2796 if (sh->pd_idx == raid_disks-1)
2797 i--; /* Q D D D P */
2802 i -= (sh->pd_idx + 2);
2805 case ALGORITHM_PARITY_0:
2808 case ALGORITHM_PARITY_N:
2810 case ALGORITHM_ROTATING_N_CONTINUE:
2811 /* Like left_symmetric, but P is before Q */
2812 if (sh->pd_idx == 0)
2813 i--; /* P D D D Q */
2818 i -= (sh->pd_idx + 1);
2821 case ALGORITHM_LEFT_ASYMMETRIC_6:
2822 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2826 case ALGORITHM_LEFT_SYMMETRIC_6:
2827 case ALGORITHM_RIGHT_SYMMETRIC_6:
2829 i += data_disks + 1;
2830 i -= (sh->pd_idx + 1);
2832 case ALGORITHM_PARITY_0_6:
2841 chunk_number = stripe * data_disks + i;
2842 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2844 check = raid5_compute_sector(conf, r_sector,
2845 previous, &dummy1, &sh2);
2846 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2847 || sh2.qd_idx != sh->qd_idx) {
2848 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2849 mdname(conf->mddev));
2856 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2857 int rcw, int expand)
2859 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2860 struct r5conf *conf = sh->raid_conf;
2861 int level = conf->level;
2865 for (i = disks; i--; ) {
2866 struct r5dev *dev = &sh->dev[i];
2869 set_bit(R5_LOCKED, &dev->flags);
2870 set_bit(R5_Wantdrain, &dev->flags);
2872 clear_bit(R5_UPTODATE, &dev->flags);
2876 /* if we are not expanding this is a proper write request, and
2877 * there will be bios with new data to be drained into the
2882 /* False alarm, nothing to do */
2884 sh->reconstruct_state = reconstruct_state_drain_run;
2885 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2887 sh->reconstruct_state = reconstruct_state_run;
2889 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2891 if (s->locked + conf->max_degraded == disks)
2892 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2893 atomic_inc(&conf->pending_full_writes);
2895 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2896 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2897 BUG_ON(level == 6 &&
2898 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2899 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2901 for (i = disks; i--; ) {
2902 struct r5dev *dev = &sh->dev[i];
2903 if (i == pd_idx || i == qd_idx)
2907 (test_bit(R5_UPTODATE, &dev->flags) ||
2908 test_bit(R5_Wantcompute, &dev->flags))) {
2909 set_bit(R5_Wantdrain, &dev->flags);
2910 set_bit(R5_LOCKED, &dev->flags);
2911 clear_bit(R5_UPTODATE, &dev->flags);
2916 /* False alarm - nothing to do */
2918 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2919 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2920 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2921 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2924 /* keep the parity disk(s) locked while asynchronous operations
2927 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2928 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2932 int qd_idx = sh->qd_idx;
2933 struct r5dev *dev = &sh->dev[qd_idx];
2935 set_bit(R5_LOCKED, &dev->flags);
2936 clear_bit(R5_UPTODATE, &dev->flags);
2940 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2941 __func__, (unsigned long long)sh->sector,
2942 s->locked, s->ops_request);
2946 * Each stripe/dev can have one or more bion attached.
2947 * toread/towrite point to the first in a chain.
2948 * The bi_next chain must be in order.
2950 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2951 int forwrite, int previous)
2954 struct r5conf *conf = sh->raid_conf;
2957 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2958 (unsigned long long)bi->bi_iter.bi_sector,
2959 (unsigned long long)sh->sector);
2962 * If several bio share a stripe. The bio bi_phys_segments acts as a
2963 * reference count to avoid race. The reference count should already be
2964 * increased before this function is called (for example, in
2965 * make_request()), so other bio sharing this stripe will not free the
2966 * stripe. If a stripe is owned by one stripe, the stripe lock will
2969 spin_lock_irq(&sh->stripe_lock);
2970 /* Don't allow new IO added to stripes in batch list */
2974 bip = &sh->dev[dd_idx].towrite;
2978 bip = &sh->dev[dd_idx].toread;
2979 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2980 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2982 bip = & (*bip)->bi_next;
2984 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2987 if (!forwrite || previous)
2988 clear_bit(STRIPE_BATCH_READY, &sh->state);
2990 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2994 raid5_inc_bi_active_stripes(bi);
2997 /* check if page is covered */
2998 sector_t sector = sh->dev[dd_idx].sector;
2999 for (bi=sh->dev[dd_idx].towrite;
3000 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3001 bi && bi->bi_iter.bi_sector <= sector;
3002 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3003 if (bio_end_sector(bi) >= sector)
3004 sector = bio_end_sector(bi);
3006 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3007 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3008 sh->overwrite_disks++;
3011 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3012 (unsigned long long)(*bip)->bi_iter.bi_sector,
3013 (unsigned long long)sh->sector, dd_idx);
3015 if (conf->mddev->bitmap && firstwrite) {
3016 /* Cannot hold spinlock over bitmap_startwrite,
3017 * but must ensure this isn't added to a batch until
3018 * we have added to the bitmap and set bm_seq.
3019 * So set STRIPE_BITMAP_PENDING to prevent
3021 * If multiple add_stripe_bio() calls race here they
3022 * much all set STRIPE_BITMAP_PENDING. So only the first one
3023 * to complete "bitmap_startwrite" gets to set
3024 * STRIPE_BIT_DELAY. This is important as once a stripe
3025 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3028 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3029 spin_unlock_irq(&sh->stripe_lock);
3030 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3032 spin_lock_irq(&sh->stripe_lock);
3033 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3034 if (!sh->batch_head) {
3035 sh->bm_seq = conf->seq_flush+1;
3036 set_bit(STRIPE_BIT_DELAY, &sh->state);
3039 spin_unlock_irq(&sh->stripe_lock);
3041 if (stripe_can_batch(sh))
3042 stripe_add_to_batch_list(conf, sh);
3046 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3047 spin_unlock_irq(&sh->stripe_lock);
3051 static void end_reshape(struct r5conf *conf);
3053 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3054 struct stripe_head *sh)
3056 int sectors_per_chunk =
3057 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3059 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3060 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3062 raid5_compute_sector(conf,
3063 stripe * (disks - conf->max_degraded)
3064 *sectors_per_chunk + chunk_offset,
3070 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3071 struct stripe_head_state *s, int disks,
3072 struct bio_list *return_bi)
3075 BUG_ON(sh->batch_head);
3076 for (i = disks; i--; ) {
3080 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3081 struct md_rdev *rdev;
3083 rdev = rcu_dereference(conf->disks[i].rdev);
3084 if (rdev && test_bit(In_sync, &rdev->flags))
3085 atomic_inc(&rdev->nr_pending);
3090 if (!rdev_set_badblocks(
3094 md_error(conf->mddev, rdev);
3095 rdev_dec_pending(rdev, conf->mddev);
3098 spin_lock_irq(&sh->stripe_lock);
3099 /* fail all writes first */
3100 bi = sh->dev[i].towrite;
3101 sh->dev[i].towrite = NULL;
3102 sh->overwrite_disks = 0;
3103 spin_unlock_irq(&sh->stripe_lock);
3107 r5l_stripe_write_finished(sh);
3109 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3110 wake_up(&conf->wait_for_overlap);
3112 while (bi && bi->bi_iter.bi_sector <
3113 sh->dev[i].sector + STRIPE_SECTORS) {
3114 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3116 bi->bi_error = -EIO;
3117 if (!raid5_dec_bi_active_stripes(bi)) {
3118 md_write_end(conf->mddev);
3119 bio_list_add(return_bi, bi);
3124 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3125 STRIPE_SECTORS, 0, 0);
3127 /* and fail all 'written' */
3128 bi = sh->dev[i].written;
3129 sh->dev[i].written = NULL;
3130 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3131 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3132 sh->dev[i].page = sh->dev[i].orig_page;
3135 if (bi) bitmap_end = 1;
3136 while (bi && bi->bi_iter.bi_sector <
3137 sh->dev[i].sector + STRIPE_SECTORS) {
3138 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3140 bi->bi_error = -EIO;
3141 if (!raid5_dec_bi_active_stripes(bi)) {
3142 md_write_end(conf->mddev);
3143 bio_list_add(return_bi, bi);
3148 /* fail any reads if this device is non-operational and
3149 * the data has not reached the cache yet.
3151 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3152 s->failed > conf->max_degraded &&
3153 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3154 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3155 spin_lock_irq(&sh->stripe_lock);
3156 bi = sh->dev[i].toread;
3157 sh->dev[i].toread = NULL;
3158 spin_unlock_irq(&sh->stripe_lock);
3159 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3160 wake_up(&conf->wait_for_overlap);
3163 while (bi && bi->bi_iter.bi_sector <
3164 sh->dev[i].sector + STRIPE_SECTORS) {
3165 struct bio *nextbi =
3166 r5_next_bio(bi, sh->dev[i].sector);
3168 bi->bi_error = -EIO;
3169 if (!raid5_dec_bi_active_stripes(bi))
3170 bio_list_add(return_bi, bi);
3175 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3176 STRIPE_SECTORS, 0, 0);
3177 /* If we were in the middle of a write the parity block might
3178 * still be locked - so just clear all R5_LOCKED flags
3180 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3185 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3186 if (atomic_dec_and_test(&conf->pending_full_writes))
3187 md_wakeup_thread(conf->mddev->thread);
3191 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3192 struct stripe_head_state *s)
3197 BUG_ON(sh->batch_head);
3198 clear_bit(STRIPE_SYNCING, &sh->state);
3199 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3200 wake_up(&conf->wait_for_overlap);
3203 /* There is nothing more to do for sync/check/repair.
3204 * Don't even need to abort as that is handled elsewhere
3205 * if needed, and not always wanted e.g. if there is a known
3207 * For recover/replace we need to record a bad block on all
3208 * non-sync devices, or abort the recovery
3210 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3211 /* During recovery devices cannot be removed, so
3212 * locking and refcounting of rdevs is not needed
3214 for (i = 0; i < conf->raid_disks; i++) {
3215 struct md_rdev *rdev = conf->disks[i].rdev;
3217 && !test_bit(Faulty, &rdev->flags)
3218 && !test_bit(In_sync, &rdev->flags)
3219 && !rdev_set_badblocks(rdev, sh->sector,
3222 rdev = conf->disks[i].replacement;
3224 && !test_bit(Faulty, &rdev->flags)
3225 && !test_bit(In_sync, &rdev->flags)
3226 && !rdev_set_badblocks(rdev, sh->sector,
3231 conf->recovery_disabled =
3232 conf->mddev->recovery_disabled;
3234 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3237 static int want_replace(struct stripe_head *sh, int disk_idx)
3239 struct md_rdev *rdev;
3241 /* Doing recovery so rcu locking not required */
3242 rdev = sh->raid_conf->disks[disk_idx].replacement;
3244 && !test_bit(Faulty, &rdev->flags)
3245 && !test_bit(In_sync, &rdev->flags)
3246 && (rdev->recovery_offset <= sh->sector
3247 || rdev->mddev->recovery_cp <= sh->sector))
3253 /* fetch_block - checks the given member device to see if its data needs
3254 * to be read or computed to satisfy a request.
3256 * Returns 1 when no more member devices need to be checked, otherwise returns
3257 * 0 to tell the loop in handle_stripe_fill to continue
3260 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3261 int disk_idx, int disks)
3263 struct r5dev *dev = &sh->dev[disk_idx];
3264 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3265 &sh->dev[s->failed_num[1]] };
3269 if (test_bit(R5_LOCKED, &dev->flags) ||
3270 test_bit(R5_UPTODATE, &dev->flags))
3271 /* No point reading this as we already have it or have
3272 * decided to get it.
3277 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3278 /* We need this block to directly satisfy a request */
3281 if (s->syncing || s->expanding ||
3282 (s->replacing && want_replace(sh, disk_idx)))
3283 /* When syncing, or expanding we read everything.
3284 * When replacing, we need the replaced block.
3288 if ((s->failed >= 1 && fdev[0]->toread) ||
3289 (s->failed >= 2 && fdev[1]->toread))
3290 /* If we want to read from a failed device, then
3291 * we need to actually read every other device.
3295 /* Sometimes neither read-modify-write nor reconstruct-write
3296 * cycles can work. In those cases we read every block we
3297 * can. Then the parity-update is certain to have enough to
3299 * This can only be a problem when we need to write something,
3300 * and some device has failed. If either of those tests
3301 * fail we need look no further.
3303 if (!s->failed || !s->to_write)
3306 if (test_bit(R5_Insync, &dev->flags) &&
3307 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3308 /* Pre-reads at not permitted until after short delay
3309 * to gather multiple requests. However if this
3310 * device is no Insync, the block could only be be computed
3311 * and there is no need to delay that.
3315 for (i = 0; i < s->failed && i < 2; i++) {
3316 if (fdev[i]->towrite &&
3317 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3318 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3319 /* If we have a partial write to a failed
3320 * device, then we will need to reconstruct
3321 * the content of that device, so all other
3322 * devices must be read.
3327 /* If we are forced to do a reconstruct-write, either because
3328 * the current RAID6 implementation only supports that, or
3329 * or because parity cannot be trusted and we are currently
3330 * recovering it, there is extra need to be careful.
3331 * If one of the devices that we would need to read, because
3332 * it is not being overwritten (and maybe not written at all)
3333 * is missing/faulty, then we need to read everything we can.
3335 if (sh->raid_conf->level != 6 &&
3336 sh->sector < sh->raid_conf->mddev->recovery_cp)
3337 /* reconstruct-write isn't being forced */
3339 for (i = 0; i < s->failed && i < 2; i++) {
3340 if (s->failed_num[i] != sh->pd_idx &&
3341 s->failed_num[i] != sh->qd_idx &&
3342 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3343 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3350 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3351 int disk_idx, int disks)
3353 struct r5dev *dev = &sh->dev[disk_idx];
3355 /* is the data in this block needed, and can we get it? */
3356 if (need_this_block(sh, s, disk_idx, disks)) {
3357 /* we would like to get this block, possibly by computing it,
3358 * otherwise read it if the backing disk is insync
3360 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3361 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3362 BUG_ON(sh->batch_head);
3363 if ((s->uptodate == disks - 1) &&
3364 (s->failed && (disk_idx == s->failed_num[0] ||
3365 disk_idx == s->failed_num[1]))) {
3366 /* have disk failed, and we're requested to fetch it;
3369 pr_debug("Computing stripe %llu block %d\n",
3370 (unsigned long long)sh->sector, disk_idx);
3371 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3372 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3373 set_bit(R5_Wantcompute, &dev->flags);
3374 sh->ops.target = disk_idx;
3375 sh->ops.target2 = -1; /* no 2nd target */
3377 /* Careful: from this point on 'uptodate' is in the eye
3378 * of raid_run_ops which services 'compute' operations
3379 * before writes. R5_Wantcompute flags a block that will
3380 * be R5_UPTODATE by the time it is needed for a
3381 * subsequent operation.
3385 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3386 /* Computing 2-failure is *very* expensive; only
3387 * do it if failed >= 2
3390 for (other = disks; other--; ) {
3391 if (other == disk_idx)
3393 if (!test_bit(R5_UPTODATE,
3394 &sh->dev[other].flags))
3398 pr_debug("Computing stripe %llu blocks %d,%d\n",
3399 (unsigned long long)sh->sector,
3401 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3402 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3403 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3404 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3405 sh->ops.target = disk_idx;
3406 sh->ops.target2 = other;
3410 } else if (test_bit(R5_Insync, &dev->flags)) {
3411 set_bit(R5_LOCKED, &dev->flags);
3412 set_bit(R5_Wantread, &dev->flags);
3414 pr_debug("Reading block %d (sync=%d)\n",
3415 disk_idx, s->syncing);
3423 * handle_stripe_fill - read or compute data to satisfy pending requests.
3425 static void handle_stripe_fill(struct stripe_head *sh,
3426 struct stripe_head_state *s,
3431 /* look for blocks to read/compute, skip this if a compute
3432 * is already in flight, or if the stripe contents are in the
3433 * midst of changing due to a write
3435 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3436 !sh->reconstruct_state)
3437 for (i = disks; i--; )
3438 if (fetch_block(sh, s, i, disks))
3440 set_bit(STRIPE_HANDLE, &sh->state);
3443 static void break_stripe_batch_list(struct stripe_head *head_sh,
3444 unsigned long handle_flags);
3445 /* handle_stripe_clean_event
3446 * any written block on an uptodate or failed drive can be returned.
3447 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3448 * never LOCKED, so we don't need to test 'failed' directly.
3450 static void handle_stripe_clean_event(struct r5conf *conf,
3451 struct stripe_head *sh, int disks, struct bio_list *return_bi)
3455 int discard_pending = 0;
3456 struct stripe_head *head_sh = sh;
3457 bool do_endio = false;
3459 for (i = disks; i--; )
3460 if (sh->dev[i].written) {
3462 if (!test_bit(R5_LOCKED, &dev->flags) &&
3463 (test_bit(R5_UPTODATE, &dev->flags) ||
3464 test_bit(R5_Discard, &dev->flags) ||
3465 test_bit(R5_SkipCopy, &dev->flags))) {
3466 /* We can return any write requests */
3467 struct bio *wbi, *wbi2;
3468 pr_debug("Return write for disc %d\n", i);
3469 if (test_and_clear_bit(R5_Discard, &dev->flags))
3470 clear_bit(R5_UPTODATE, &dev->flags);
3471 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3472 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3477 dev->page = dev->orig_page;
3479 dev->written = NULL;
3480 while (wbi && wbi->bi_iter.bi_sector <
3481 dev->sector + STRIPE_SECTORS) {
3482 wbi2 = r5_next_bio(wbi, dev->sector);
3483 if (!raid5_dec_bi_active_stripes(wbi)) {
3484 md_write_end(conf->mddev);
3485 bio_list_add(return_bi, wbi);
3489 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3491 !test_bit(STRIPE_DEGRADED, &sh->state),
3493 if (head_sh->batch_head) {
3494 sh = list_first_entry(&sh->batch_list,
3497 if (sh != head_sh) {
3504 } else if (test_bit(R5_Discard, &dev->flags))
3505 discard_pending = 1;
3506 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3507 WARN_ON(dev->page != dev->orig_page);
3510 r5l_stripe_write_finished(sh);
3512 if (!discard_pending &&
3513 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3515 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3516 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3517 if (sh->qd_idx >= 0) {
3518 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3519 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3521 /* now that discard is done we can proceed with any sync */
3522 clear_bit(STRIPE_DISCARD, &sh->state);
3524 * SCSI discard will change some bio fields and the stripe has
3525 * no updated data, so remove it from hash list and the stripe
3526 * will be reinitialized
3529 hash = sh->hash_lock_index;
3530 spin_lock_irq(conf->hash_locks + hash);
3532 spin_unlock_irq(conf->hash_locks + hash);
3533 if (head_sh->batch_head) {
3534 sh = list_first_entry(&sh->batch_list,
3535 struct stripe_head, batch_list);
3541 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3542 set_bit(STRIPE_HANDLE, &sh->state);
3546 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3547 if (atomic_dec_and_test(&conf->pending_full_writes))
3548 md_wakeup_thread(conf->mddev->thread);
3550 if (head_sh->batch_head && do_endio)
3551 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3554 static void handle_stripe_dirtying(struct r5conf *conf,
3555 struct stripe_head *sh,
3556 struct stripe_head_state *s,
3559 int rmw = 0, rcw = 0, i;
3560 sector_t recovery_cp = conf->mddev->recovery_cp;
3562 /* Check whether resync is now happening or should start.
3563 * If yes, then the array is dirty (after unclean shutdown or
3564 * initial creation), so parity in some stripes might be inconsistent.
3565 * In this case, we need to always do reconstruct-write, to ensure
3566 * that in case of drive failure or read-error correction, we
3567 * generate correct data from the parity.
3569 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3570 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3572 /* Calculate the real rcw later - for now make it
3573 * look like rcw is cheaper
3576 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3577 conf->rmw_level, (unsigned long long)recovery_cp,
3578 (unsigned long long)sh->sector);
3579 } else for (i = disks; i--; ) {
3580 /* would I have to read this buffer for read_modify_write */
3581 struct r5dev *dev = &sh->dev[i];
3582 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3583 !test_bit(R5_LOCKED, &dev->flags) &&
3584 !(test_bit(R5_UPTODATE, &dev->flags) ||
3585 test_bit(R5_Wantcompute, &dev->flags))) {
3586 if (test_bit(R5_Insync, &dev->flags))
3589 rmw += 2*disks; /* cannot read it */
3591 /* Would I have to read this buffer for reconstruct_write */
3592 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3593 i != sh->pd_idx && i != sh->qd_idx &&
3594 !test_bit(R5_LOCKED, &dev->flags) &&
3595 !(test_bit(R5_UPTODATE, &dev->flags) ||
3596 test_bit(R5_Wantcompute, &dev->flags))) {
3597 if (test_bit(R5_Insync, &dev->flags))
3603 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3604 (unsigned long long)sh->sector, rmw, rcw);
3605 set_bit(STRIPE_HANDLE, &sh->state);
3606 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3607 /* prefer read-modify-write, but need to get some data */
3608 if (conf->mddev->queue)
3609 blk_add_trace_msg(conf->mddev->queue,
3610 "raid5 rmw %llu %d",
3611 (unsigned long long)sh->sector, rmw);
3612 for (i = disks; i--; ) {
3613 struct r5dev *dev = &sh->dev[i];
3614 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3615 !test_bit(R5_LOCKED, &dev->flags) &&
3616 !(test_bit(R5_UPTODATE, &dev->flags) ||
3617 test_bit(R5_Wantcompute, &dev->flags)) &&
3618 test_bit(R5_Insync, &dev->flags)) {
3619 if (test_bit(STRIPE_PREREAD_ACTIVE,
3621 pr_debug("Read_old block %d for r-m-w\n",
3623 set_bit(R5_LOCKED, &dev->flags);
3624 set_bit(R5_Wantread, &dev->flags);
3627 set_bit(STRIPE_DELAYED, &sh->state);
3628 set_bit(STRIPE_HANDLE, &sh->state);
3633 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3634 /* want reconstruct write, but need to get some data */
3637 for (i = disks; i--; ) {
3638 struct r5dev *dev = &sh->dev[i];
3639 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3640 i != sh->pd_idx && i != sh->qd_idx &&
3641 !test_bit(R5_LOCKED, &dev->flags) &&
3642 !(test_bit(R5_UPTODATE, &dev->flags) ||
3643 test_bit(R5_Wantcompute, &dev->flags))) {
3645 if (test_bit(R5_Insync, &dev->flags) &&
3646 test_bit(STRIPE_PREREAD_ACTIVE,
3648 pr_debug("Read_old block "
3649 "%d for Reconstruct\n", i);
3650 set_bit(R5_LOCKED, &dev->flags);
3651 set_bit(R5_Wantread, &dev->flags);
3655 set_bit(STRIPE_DELAYED, &sh->state);
3656 set_bit(STRIPE_HANDLE, &sh->state);
3660 if (rcw && conf->mddev->queue)
3661 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3662 (unsigned long long)sh->sector,
3663 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3666 if (rcw > disks && rmw > disks &&
3667 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3668 set_bit(STRIPE_DELAYED, &sh->state);
3670 /* now if nothing is locked, and if we have enough data,
3671 * we can start a write request
3673 /* since handle_stripe can be called at any time we need to handle the
3674 * case where a compute block operation has been submitted and then a
3675 * subsequent call wants to start a write request. raid_run_ops only
3676 * handles the case where compute block and reconstruct are requested
3677 * simultaneously. If this is not the case then new writes need to be
3678 * held off until the compute completes.
3680 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3681 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3682 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3683 schedule_reconstruction(sh, s, rcw == 0, 0);
3686 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3687 struct stripe_head_state *s, int disks)
3689 struct r5dev *dev = NULL;
3691 BUG_ON(sh->batch_head);
3692 set_bit(STRIPE_HANDLE, &sh->state);
3694 switch (sh->check_state) {
3695 case check_state_idle:
3696 /* start a new check operation if there are no failures */
3697 if (s->failed == 0) {
3698 BUG_ON(s->uptodate != disks);
3699 sh->check_state = check_state_run;
3700 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3701 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3705 dev = &sh->dev[s->failed_num[0]];
3707 case check_state_compute_result:
3708 sh->check_state = check_state_idle;
3710 dev = &sh->dev[sh->pd_idx];
3712 /* check that a write has not made the stripe insync */
3713 if (test_bit(STRIPE_INSYNC, &sh->state))
3716 /* either failed parity check, or recovery is happening */
3717 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3718 BUG_ON(s->uptodate != disks);
3720 set_bit(R5_LOCKED, &dev->flags);
3722 set_bit(R5_Wantwrite, &dev->flags);
3724 clear_bit(STRIPE_DEGRADED, &sh->state);
3725 set_bit(STRIPE_INSYNC, &sh->state);
3727 case check_state_run:
3728 break; /* we will be called again upon completion */
3729 case check_state_check_result:
3730 sh->check_state = check_state_idle;
3732 /* if a failure occurred during the check operation, leave
3733 * STRIPE_INSYNC not set and let the stripe be handled again
3738 /* handle a successful check operation, if parity is correct
3739 * we are done. Otherwise update the mismatch count and repair
3740 * parity if !MD_RECOVERY_CHECK
3742 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3743 /* parity is correct (on disc,
3744 * not in buffer any more)
3746 set_bit(STRIPE_INSYNC, &sh->state);
3748 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3749 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3750 /* don't try to repair!! */
3751 set_bit(STRIPE_INSYNC, &sh->state);
3753 sh->check_state = check_state_compute_run;
3754 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3755 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3756 set_bit(R5_Wantcompute,
3757 &sh->dev[sh->pd_idx].flags);
3758 sh->ops.target = sh->pd_idx;
3759 sh->ops.target2 = -1;
3764 case check_state_compute_run:
3767 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3768 __func__, sh->check_state,
3769 (unsigned long long) sh->sector);
3774 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3775 struct stripe_head_state *s,
3778 int pd_idx = sh->pd_idx;
3779 int qd_idx = sh->qd_idx;
3782 BUG_ON(sh->batch_head);
3783 set_bit(STRIPE_HANDLE, &sh->state);
3785 BUG_ON(s->failed > 2);
3787 /* Want to check and possibly repair P and Q.
3788 * However there could be one 'failed' device, in which
3789 * case we can only check one of them, possibly using the
3790 * other to generate missing data
3793 switch (sh->check_state) {
3794 case check_state_idle:
3795 /* start a new check operation if there are < 2 failures */
3796 if (s->failed == s->q_failed) {
3797 /* The only possible failed device holds Q, so it
3798 * makes sense to check P (If anything else were failed,
3799 * we would have used P to recreate it).
3801 sh->check_state = check_state_run;
3803 if (!s->q_failed && s->failed < 2) {
3804 /* Q is not failed, and we didn't use it to generate
3805 * anything, so it makes sense to check it
3807 if (sh->check_state == check_state_run)
3808 sh->check_state = check_state_run_pq;
3810 sh->check_state = check_state_run_q;
3813 /* discard potentially stale zero_sum_result */
3814 sh->ops.zero_sum_result = 0;
3816 if (sh->check_state == check_state_run) {
3817 /* async_xor_zero_sum destroys the contents of P */
3818 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3821 if (sh->check_state >= check_state_run &&
3822 sh->check_state <= check_state_run_pq) {
3823 /* async_syndrome_zero_sum preserves P and Q, so
3824 * no need to mark them !uptodate here
3826 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3830 /* we have 2-disk failure */
3831 BUG_ON(s->failed != 2);
3833 case check_state_compute_result:
3834 sh->check_state = check_state_idle;
3836 /* check that a write has not made the stripe insync */
3837 if (test_bit(STRIPE_INSYNC, &sh->state))
3840 /* now write out any block on a failed drive,
3841 * or P or Q if they were recomputed
3843 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3844 if (s->failed == 2) {
3845 dev = &sh->dev[s->failed_num[1]];
3847 set_bit(R5_LOCKED, &dev->flags);
3848 set_bit(R5_Wantwrite, &dev->flags);
3850 if (s->failed >= 1) {
3851 dev = &sh->dev[s->failed_num[0]];
3853 set_bit(R5_LOCKED, &dev->flags);
3854 set_bit(R5_Wantwrite, &dev->flags);
3856 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3857 dev = &sh->dev[pd_idx];
3859 set_bit(R5_LOCKED, &dev->flags);
3860 set_bit(R5_Wantwrite, &dev->flags);
3862 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3863 dev = &sh->dev[qd_idx];
3865 set_bit(R5_LOCKED, &dev->flags);
3866 set_bit(R5_Wantwrite, &dev->flags);
3868 clear_bit(STRIPE_DEGRADED, &sh->state);
3870 set_bit(STRIPE_INSYNC, &sh->state);
3872 case check_state_run:
3873 case check_state_run_q:
3874 case check_state_run_pq:
3875 break; /* we will be called again upon completion */
3876 case check_state_check_result:
3877 sh->check_state = check_state_idle;
3879 /* handle a successful check operation, if parity is correct
3880 * we are done. Otherwise update the mismatch count and repair
3881 * parity if !MD_RECOVERY_CHECK
3883 if (sh->ops.zero_sum_result == 0) {
3884 /* both parities are correct */
3886 set_bit(STRIPE_INSYNC, &sh->state);
3888 /* in contrast to the raid5 case we can validate
3889 * parity, but still have a failure to write
3892 sh->check_state = check_state_compute_result;
3893 /* Returning at this point means that we may go
3894 * off and bring p and/or q uptodate again so
3895 * we make sure to check zero_sum_result again
3896 * to verify if p or q need writeback
3900 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3901 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3902 /* don't try to repair!! */
3903 set_bit(STRIPE_INSYNC, &sh->state);
3905 int *target = &sh->ops.target;
3907 sh->ops.target = -1;
3908 sh->ops.target2 = -1;
3909 sh->check_state = check_state_compute_run;
3910 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3911 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3912 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3913 set_bit(R5_Wantcompute,
3914 &sh->dev[pd_idx].flags);
3916 target = &sh->ops.target2;
3919 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3920 set_bit(R5_Wantcompute,
3921 &sh->dev[qd_idx].flags);
3928 case check_state_compute_run:
3931 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3932 __func__, sh->check_state,
3933 (unsigned long long) sh->sector);
3938 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3942 /* We have read all the blocks in this stripe and now we need to
3943 * copy some of them into a target stripe for expand.
3945 struct dma_async_tx_descriptor *tx = NULL;
3946 BUG_ON(sh->batch_head);
3947 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3948 for (i = 0; i < sh->disks; i++)
3949 if (i != sh->pd_idx && i != sh->qd_idx) {
3951 struct stripe_head *sh2;
3952 struct async_submit_ctl submit;
3954 sector_t bn = raid5_compute_blocknr(sh, i, 1);
3955 sector_t s = raid5_compute_sector(conf, bn, 0,
3957 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
3959 /* so far only the early blocks of this stripe
3960 * have been requested. When later blocks
3961 * get requested, we will try again
3964 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3965 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3966 /* must have already done this block */
3967 raid5_release_stripe(sh2);
3971 /* place all the copies on one channel */
3972 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3973 tx = async_memcpy(sh2->dev[dd_idx].page,
3974 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3977 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3978 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3979 for (j = 0; j < conf->raid_disks; j++)
3980 if (j != sh2->pd_idx &&
3982 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3984 if (j == conf->raid_disks) {
3985 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3986 set_bit(STRIPE_HANDLE, &sh2->state);
3988 raid5_release_stripe(sh2);
3991 /* done submitting copies, wait for them to complete */
3992 async_tx_quiesce(&tx);
3996 * handle_stripe - do things to a stripe.
3998 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3999 * state of various bits to see what needs to be done.
4001 * return some read requests which now have data
4002 * return some write requests which are safely on storage
4003 * schedule a read on some buffers
4004 * schedule a write of some buffers
4005 * return confirmation of parity correctness
4009 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4011 struct r5conf *conf = sh->raid_conf;
4012 int disks = sh->disks;
4015 int do_recovery = 0;
4017 memset(s, 0, sizeof(*s));
4019 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4020 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4021 s->failed_num[0] = -1;
4022 s->failed_num[1] = -1;
4023 s->log_failed = r5l_log_disk_error(conf);
4025 /* Now to look around and see what can be done */
4027 for (i=disks; i--; ) {
4028 struct md_rdev *rdev;
4035 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4037 dev->toread, dev->towrite, dev->written);
4038 /* maybe we can reply to a read
4040 * new wantfill requests are only permitted while
4041 * ops_complete_biofill is guaranteed to be inactive
4043 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4044 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4045 set_bit(R5_Wantfill, &dev->flags);
4047 /* now count some things */
4048 if (test_bit(R5_LOCKED, &dev->flags))
4050 if (test_bit(R5_UPTODATE, &dev->flags))
4052 if (test_bit(R5_Wantcompute, &dev->flags)) {
4054 BUG_ON(s->compute > 2);
4057 if (test_bit(R5_Wantfill, &dev->flags))
4059 else if (dev->toread)
4063 if (!test_bit(R5_OVERWRITE, &dev->flags))
4068 /* Prefer to use the replacement for reads, but only
4069 * if it is recovered enough and has no bad blocks.
4071 rdev = rcu_dereference(conf->disks[i].replacement);
4072 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4073 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4074 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4075 &first_bad, &bad_sectors))
4076 set_bit(R5_ReadRepl, &dev->flags);
4078 if (rdev && !test_bit(Faulty, &rdev->flags))
4079 set_bit(R5_NeedReplace, &dev->flags);
4081 clear_bit(R5_NeedReplace, &dev->flags);
4082 rdev = rcu_dereference(conf->disks[i].rdev);
4083 clear_bit(R5_ReadRepl, &dev->flags);
4085 if (rdev && test_bit(Faulty, &rdev->flags))
4088 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4089 &first_bad, &bad_sectors);
4090 if (s->blocked_rdev == NULL
4091 && (test_bit(Blocked, &rdev->flags)
4094 set_bit(BlockedBadBlocks,
4096 s->blocked_rdev = rdev;
4097 atomic_inc(&rdev->nr_pending);
4100 clear_bit(R5_Insync, &dev->flags);
4104 /* also not in-sync */
4105 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4106 test_bit(R5_UPTODATE, &dev->flags)) {
4107 /* treat as in-sync, but with a read error
4108 * which we can now try to correct
4110 set_bit(R5_Insync, &dev->flags);
4111 set_bit(R5_ReadError, &dev->flags);
4113 } else if (test_bit(In_sync, &rdev->flags))
4114 set_bit(R5_Insync, &dev->flags);
4115 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4116 /* in sync if before recovery_offset */
4117 set_bit(R5_Insync, &dev->flags);
4118 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4119 test_bit(R5_Expanded, &dev->flags))
4120 /* If we've reshaped into here, we assume it is Insync.
4121 * We will shortly update recovery_offset to make
4124 set_bit(R5_Insync, &dev->flags);
4126 if (test_bit(R5_WriteError, &dev->flags)) {
4127 /* This flag does not apply to '.replacement'
4128 * only to .rdev, so make sure to check that*/
4129 struct md_rdev *rdev2 = rcu_dereference(
4130 conf->disks[i].rdev);
4132 clear_bit(R5_Insync, &dev->flags);
4133 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4134 s->handle_bad_blocks = 1;
4135 atomic_inc(&rdev2->nr_pending);
4137 clear_bit(R5_WriteError, &dev->flags);
4139 if (test_bit(R5_MadeGood, &dev->flags)) {
4140 /* This flag does not apply to '.replacement'
4141 * only to .rdev, so make sure to check that*/
4142 struct md_rdev *rdev2 = rcu_dereference(
4143 conf->disks[i].rdev);
4144 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4145 s->handle_bad_blocks = 1;
4146 atomic_inc(&rdev2->nr_pending);
4148 clear_bit(R5_MadeGood, &dev->flags);
4150 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4151 struct md_rdev *rdev2 = rcu_dereference(
4152 conf->disks[i].replacement);
4153 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4154 s->handle_bad_blocks = 1;
4155 atomic_inc(&rdev2->nr_pending);
4157 clear_bit(R5_MadeGoodRepl, &dev->flags);
4159 if (!test_bit(R5_Insync, &dev->flags)) {
4160 /* The ReadError flag will just be confusing now */
4161 clear_bit(R5_ReadError, &dev->flags);
4162 clear_bit(R5_ReWrite, &dev->flags);
4164 if (test_bit(R5_ReadError, &dev->flags))
4165 clear_bit(R5_Insync, &dev->flags);
4166 if (!test_bit(R5_Insync, &dev->flags)) {
4168 s->failed_num[s->failed] = i;
4170 if (rdev && !test_bit(Faulty, &rdev->flags))
4174 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4175 /* If there is a failed device being replaced,
4176 * we must be recovering.
4177 * else if we are after recovery_cp, we must be syncing
4178 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4179 * else we can only be replacing
4180 * sync and recovery both need to read all devices, and so
4181 * use the same flag.
4184 sh->sector >= conf->mddev->recovery_cp ||
4185 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4193 static int clear_batch_ready(struct stripe_head *sh)
4195 /* Return '1' if this is a member of batch, or
4196 * '0' if it is a lone stripe or a head which can now be
4199 struct stripe_head *tmp;
4200 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4201 return (sh->batch_head && sh->batch_head != sh);
4202 spin_lock(&sh->stripe_lock);
4203 if (!sh->batch_head) {
4204 spin_unlock(&sh->stripe_lock);
4209 * this stripe could be added to a batch list before we check
4210 * BATCH_READY, skips it
4212 if (sh->batch_head != sh) {
4213 spin_unlock(&sh->stripe_lock);
4216 spin_lock(&sh->batch_lock);
4217 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4218 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4219 spin_unlock(&sh->batch_lock);
4220 spin_unlock(&sh->stripe_lock);
4223 * BATCH_READY is cleared, no new stripes can be added.
4224 * batch_list can be accessed without lock
4229 static void break_stripe_batch_list(struct stripe_head *head_sh,
4230 unsigned long handle_flags)
4232 struct stripe_head *sh, *next;
4236 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4238 list_del_init(&sh->batch_list);
4240 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4241 (1 << STRIPE_SYNCING) |
4242 (1 << STRIPE_REPLACED) |
4243 (1 << STRIPE_PREREAD_ACTIVE) |
4244 (1 << STRIPE_DELAYED) |
4245 (1 << STRIPE_BIT_DELAY) |
4246 (1 << STRIPE_FULL_WRITE) |
4247 (1 << STRIPE_BIOFILL_RUN) |
4248 (1 << STRIPE_COMPUTE_RUN) |
4249 (1 << STRIPE_OPS_REQ_PENDING) |
4250 (1 << STRIPE_DISCARD) |
4251 (1 << STRIPE_BATCH_READY) |
4252 (1 << STRIPE_BATCH_ERR) |
4253 (1 << STRIPE_BITMAP_PENDING)));
4254 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4255 (1 << STRIPE_REPLACED)));
4257 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4258 (1 << STRIPE_DEGRADED)),
4259 head_sh->state & (1 << STRIPE_INSYNC));
4261 sh->check_state = head_sh->check_state;
4262 sh->reconstruct_state = head_sh->reconstruct_state;
4263 for (i = 0; i < sh->disks; i++) {
4264 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4266 sh->dev[i].flags = head_sh->dev[i].flags &
4267 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4269 spin_lock_irq(&sh->stripe_lock);
4270 sh->batch_head = NULL;
4271 spin_unlock_irq(&sh->stripe_lock);
4272 if (handle_flags == 0 ||
4273 sh->state & handle_flags)
4274 set_bit(STRIPE_HANDLE, &sh->state);
4275 raid5_release_stripe(sh);
4277 spin_lock_irq(&head_sh->stripe_lock);
4278 head_sh->batch_head = NULL;
4279 spin_unlock_irq(&head_sh->stripe_lock);
4280 for (i = 0; i < head_sh->disks; i++)
4281 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4283 if (head_sh->state & handle_flags)
4284 set_bit(STRIPE_HANDLE, &head_sh->state);
4287 wake_up(&head_sh->raid_conf->wait_for_overlap);
4290 static void handle_stripe(struct stripe_head *sh)
4292 struct stripe_head_state s;
4293 struct r5conf *conf = sh->raid_conf;
4296 int disks = sh->disks;
4297 struct r5dev *pdev, *qdev;
4299 clear_bit(STRIPE_HANDLE, &sh->state);
4300 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4301 /* already being handled, ensure it gets handled
4302 * again when current action finishes */
4303 set_bit(STRIPE_HANDLE, &sh->state);
4307 if (clear_batch_ready(sh) ) {
4308 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4312 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4313 break_stripe_batch_list(sh, 0);
4315 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4316 spin_lock(&sh->stripe_lock);
4317 /* Cannot process 'sync' concurrently with 'discard' */
4318 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4319 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4320 set_bit(STRIPE_SYNCING, &sh->state);
4321 clear_bit(STRIPE_INSYNC, &sh->state);
4322 clear_bit(STRIPE_REPLACED, &sh->state);
4324 spin_unlock(&sh->stripe_lock);
4326 clear_bit(STRIPE_DELAYED, &sh->state);
4328 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4329 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4330 (unsigned long long)sh->sector, sh->state,
4331 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4332 sh->check_state, sh->reconstruct_state);
4334 analyse_stripe(sh, &s);
4336 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4339 if (s.handle_bad_blocks) {
4340 set_bit(STRIPE_HANDLE, &sh->state);
4344 if (unlikely(s.blocked_rdev)) {
4345 if (s.syncing || s.expanding || s.expanded ||
4346 s.replacing || s.to_write || s.written) {
4347 set_bit(STRIPE_HANDLE, &sh->state);
4350 /* There is nothing for the blocked_rdev to block */
4351 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4352 s.blocked_rdev = NULL;
4355 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4356 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4357 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4360 pr_debug("locked=%d uptodate=%d to_read=%d"
4361 " to_write=%d failed=%d failed_num=%d,%d\n",
4362 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4363 s.failed_num[0], s.failed_num[1]);
4364 /* check if the array has lost more than max_degraded devices and,
4365 * if so, some requests might need to be failed.
4367 if (s.failed > conf->max_degraded || s.log_failed) {
4368 sh->check_state = 0;
4369 sh->reconstruct_state = 0;
4370 break_stripe_batch_list(sh, 0);
4371 if (s.to_read+s.to_write+s.written)
4372 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4373 if (s.syncing + s.replacing)
4374 handle_failed_sync(conf, sh, &s);
4377 /* Now we check to see if any write operations have recently
4381 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4383 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4384 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4385 sh->reconstruct_state = reconstruct_state_idle;
4387 /* All the 'written' buffers and the parity block are ready to
4388 * be written back to disk
4390 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4391 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4392 BUG_ON(sh->qd_idx >= 0 &&
4393 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4394 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4395 for (i = disks; i--; ) {
4396 struct r5dev *dev = &sh->dev[i];
4397 if (test_bit(R5_LOCKED, &dev->flags) &&
4398 (i == sh->pd_idx || i == sh->qd_idx ||
4400 pr_debug("Writing block %d\n", i);
4401 set_bit(R5_Wantwrite, &dev->flags);
4406 if (!test_bit(R5_Insync, &dev->flags) ||
4407 ((i == sh->pd_idx || i == sh->qd_idx) &&
4409 set_bit(STRIPE_INSYNC, &sh->state);
4412 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4413 s.dec_preread_active = 1;
4417 * might be able to return some write requests if the parity blocks
4418 * are safe, or on a failed drive
4420 pdev = &sh->dev[sh->pd_idx];
4421 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4422 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4423 qdev = &sh->dev[sh->qd_idx];
4424 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4425 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4429 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4430 && !test_bit(R5_LOCKED, &pdev->flags)
4431 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4432 test_bit(R5_Discard, &pdev->flags))))) &&
4433 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4434 && !test_bit(R5_LOCKED, &qdev->flags)
4435 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4436 test_bit(R5_Discard, &qdev->flags))))))
4437 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4439 /* Now we might consider reading some blocks, either to check/generate
4440 * parity, or to satisfy requests
4441 * or to load a block that is being partially written.
4443 if (s.to_read || s.non_overwrite
4444 || (conf->level == 6 && s.to_write && s.failed)
4445 || (s.syncing && (s.uptodate + s.compute < disks))
4448 handle_stripe_fill(sh, &s, disks);
4450 /* Now to consider new write requests and what else, if anything
4451 * should be read. We do not handle new writes when:
4452 * 1/ A 'write' operation (copy+xor) is already in flight.
4453 * 2/ A 'check' operation is in flight, as it may clobber the parity
4456 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4457 handle_stripe_dirtying(conf, sh, &s, disks);
4459 /* maybe we need to check and possibly fix the parity for this stripe
4460 * Any reads will already have been scheduled, so we just see if enough
4461 * data is available. The parity check is held off while parity
4462 * dependent operations are in flight.
4464 if (sh->check_state ||
4465 (s.syncing && s.locked == 0 &&
4466 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4467 !test_bit(STRIPE_INSYNC, &sh->state))) {
4468 if (conf->level == 6)
4469 handle_parity_checks6(conf, sh, &s, disks);
4471 handle_parity_checks5(conf, sh, &s, disks);
4474 if ((s.replacing || s.syncing) && s.locked == 0
4475 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4476 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4477 /* Write out to replacement devices where possible */
4478 for (i = 0; i < conf->raid_disks; i++)
4479 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4480 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4481 set_bit(R5_WantReplace, &sh->dev[i].flags);
4482 set_bit(R5_LOCKED, &sh->dev[i].flags);
4486 set_bit(STRIPE_INSYNC, &sh->state);
4487 set_bit(STRIPE_REPLACED, &sh->state);
4489 if ((s.syncing || s.replacing) && s.locked == 0 &&
4490 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4491 test_bit(STRIPE_INSYNC, &sh->state)) {
4492 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4493 clear_bit(STRIPE_SYNCING, &sh->state);
4494 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4495 wake_up(&conf->wait_for_overlap);
4498 /* If the failed drives are just a ReadError, then we might need
4499 * to progress the repair/check process
4501 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4502 for (i = 0; i < s.failed; i++) {
4503 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4504 if (test_bit(R5_ReadError, &dev->flags)
4505 && !test_bit(R5_LOCKED, &dev->flags)
4506 && test_bit(R5_UPTODATE, &dev->flags)
4508 if (!test_bit(R5_ReWrite, &dev->flags)) {
4509 set_bit(R5_Wantwrite, &dev->flags);
4510 set_bit(R5_ReWrite, &dev->flags);
4511 set_bit(R5_LOCKED, &dev->flags);
4514 /* let's read it back */
4515 set_bit(R5_Wantread, &dev->flags);
4516 set_bit(R5_LOCKED, &dev->flags);
4522 /* Finish reconstruct operations initiated by the expansion process */
4523 if (sh->reconstruct_state == reconstruct_state_result) {
4524 struct stripe_head *sh_src
4525 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4526 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4527 /* sh cannot be written until sh_src has been read.
4528 * so arrange for sh to be delayed a little
4530 set_bit(STRIPE_DELAYED, &sh->state);
4531 set_bit(STRIPE_HANDLE, &sh->state);
4532 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4534 atomic_inc(&conf->preread_active_stripes);
4535 raid5_release_stripe(sh_src);
4539 raid5_release_stripe(sh_src);
4541 sh->reconstruct_state = reconstruct_state_idle;
4542 clear_bit(STRIPE_EXPANDING, &sh->state);
4543 for (i = conf->raid_disks; i--; ) {
4544 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4545 set_bit(R5_LOCKED, &sh->dev[i].flags);
4550 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4551 !sh->reconstruct_state) {
4552 /* Need to write out all blocks after computing parity */
4553 sh->disks = conf->raid_disks;
4554 stripe_set_idx(sh->sector, conf, 0, sh);
4555 schedule_reconstruction(sh, &s, 1, 1);
4556 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4557 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4558 atomic_dec(&conf->reshape_stripes);
4559 wake_up(&conf->wait_for_overlap);
4560 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4563 if (s.expanding && s.locked == 0 &&
4564 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4565 handle_stripe_expansion(conf, sh);
4568 /* wait for this device to become unblocked */
4569 if (unlikely(s.blocked_rdev)) {
4570 if (conf->mddev->external)
4571 md_wait_for_blocked_rdev(s.blocked_rdev,
4574 /* Internal metadata will immediately
4575 * be written by raid5d, so we don't
4576 * need to wait here.
4578 rdev_dec_pending(s.blocked_rdev,
4582 if (s.handle_bad_blocks)
4583 for (i = disks; i--; ) {
4584 struct md_rdev *rdev;
4585 struct r5dev *dev = &sh->dev[i];
4586 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4587 /* We own a safe reference to the rdev */
4588 rdev = conf->disks[i].rdev;
4589 if (!rdev_set_badblocks(rdev, sh->sector,
4591 md_error(conf->mddev, rdev);
4592 rdev_dec_pending(rdev, conf->mddev);
4594 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4595 rdev = conf->disks[i].rdev;
4596 rdev_clear_badblocks(rdev, sh->sector,
4598 rdev_dec_pending(rdev, conf->mddev);
4600 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4601 rdev = conf->disks[i].replacement;
4603 /* rdev have been moved down */
4604 rdev = conf->disks[i].rdev;
4605 rdev_clear_badblocks(rdev, sh->sector,
4607 rdev_dec_pending(rdev, conf->mddev);
4612 raid_run_ops(sh, s.ops_request);
4616 if (s.dec_preread_active) {
4617 /* We delay this until after ops_run_io so that if make_request
4618 * is waiting on a flush, it won't continue until the writes
4619 * have actually been submitted.
4621 atomic_dec(&conf->preread_active_stripes);
4622 if (atomic_read(&conf->preread_active_stripes) <
4624 md_wakeup_thread(conf->mddev->thread);
4627 if (!bio_list_empty(&s.return_bi)) {
4628 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags)) {
4629 spin_lock_irq(&conf->device_lock);
4630 bio_list_merge(&conf->return_bi, &s.return_bi);
4631 spin_unlock_irq(&conf->device_lock);
4632 md_wakeup_thread(conf->mddev->thread);
4634 return_io(&s.return_bi);
4637 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4640 static void raid5_activate_delayed(struct r5conf *conf)
4642 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4643 while (!list_empty(&conf->delayed_list)) {
4644 struct list_head *l = conf->delayed_list.next;
4645 struct stripe_head *sh;
4646 sh = list_entry(l, struct stripe_head, lru);
4648 clear_bit(STRIPE_DELAYED, &sh->state);
4649 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4650 atomic_inc(&conf->preread_active_stripes);
4651 list_add_tail(&sh->lru, &conf->hold_list);
4652 raid5_wakeup_stripe_thread(sh);
4657 static void activate_bit_delay(struct r5conf *conf,
4658 struct list_head *temp_inactive_list)
4660 /* device_lock is held */
4661 struct list_head head;
4662 list_add(&head, &conf->bitmap_list);
4663 list_del_init(&conf->bitmap_list);
4664 while (!list_empty(&head)) {
4665 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4667 list_del_init(&sh->lru);
4668 atomic_inc(&sh->count);
4669 hash = sh->hash_lock_index;
4670 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4674 static int raid5_congested(struct mddev *mddev, int bits)
4676 struct r5conf *conf = mddev->private;
4678 /* No difference between reads and writes. Just check
4679 * how busy the stripe_cache is
4682 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4686 if (atomic_read(&conf->empty_inactive_list_nr))
4692 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4694 struct r5conf *conf = mddev->private;
4695 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4696 unsigned int chunk_sectors;
4697 unsigned int bio_sectors = bio_sectors(bio);
4699 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4700 return chunk_sectors >=
4701 ((sector & (chunk_sectors - 1)) + bio_sectors);
4705 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4706 * later sampled by raid5d.
4708 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4710 unsigned long flags;
4712 spin_lock_irqsave(&conf->device_lock, flags);
4714 bi->bi_next = conf->retry_read_aligned_list;
4715 conf->retry_read_aligned_list = bi;
4717 spin_unlock_irqrestore(&conf->device_lock, flags);
4718 md_wakeup_thread(conf->mddev->thread);
4721 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4725 bi = conf->retry_read_aligned;
4727 conf->retry_read_aligned = NULL;
4730 bi = conf->retry_read_aligned_list;
4732 conf->retry_read_aligned_list = bi->bi_next;
4735 * this sets the active strip count to 1 and the processed
4736 * strip count to zero (upper 8 bits)
4738 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4745 * The "raid5_align_endio" should check if the read succeeded and if it
4746 * did, call bio_endio on the original bio (having bio_put the new bio
4748 * If the read failed..
4750 static void raid5_align_endio(struct bio *bi)
4752 struct bio* raid_bi = bi->bi_private;
4753 struct mddev *mddev;
4754 struct r5conf *conf;
4755 struct md_rdev *rdev;
4756 int error = bi->bi_error;
4760 rdev = (void*)raid_bi->bi_next;
4761 raid_bi->bi_next = NULL;
4762 mddev = rdev->mddev;
4763 conf = mddev->private;
4765 rdev_dec_pending(rdev, conf->mddev);
4768 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4771 if (atomic_dec_and_test(&conf->active_aligned_reads))
4772 wake_up(&conf->wait_for_quiescent);
4776 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4778 add_bio_to_retry(raid_bi, conf);
4781 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4783 struct r5conf *conf = mddev->private;
4785 struct bio* align_bi;
4786 struct md_rdev *rdev;
4787 sector_t end_sector;
4789 if (!in_chunk_boundary(mddev, raid_bio)) {
4790 pr_debug("%s: non aligned\n", __func__);
4794 * use bio_clone_mddev to make a copy of the bio
4796 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4800 * set bi_end_io to a new function, and set bi_private to the
4803 align_bi->bi_end_io = raid5_align_endio;
4804 align_bi->bi_private = raid_bio;
4808 align_bi->bi_iter.bi_sector =
4809 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4812 end_sector = bio_end_sector(align_bi);
4814 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4815 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4816 rdev->recovery_offset < end_sector) {
4817 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4819 (test_bit(Faulty, &rdev->flags) ||
4820 !(test_bit(In_sync, &rdev->flags) ||
4821 rdev->recovery_offset >= end_sector)))
4828 atomic_inc(&rdev->nr_pending);
4830 raid_bio->bi_next = (void*)rdev;
4831 align_bi->bi_bdev = rdev->bdev;
4832 bio_clear_flag(align_bi, BIO_SEG_VALID);
4834 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4835 bio_sectors(align_bi),
4836 &first_bad, &bad_sectors)) {
4838 rdev_dec_pending(rdev, mddev);
4842 /* No reshape active, so we can trust rdev->data_offset */
4843 align_bi->bi_iter.bi_sector += rdev->data_offset;
4845 spin_lock_irq(&conf->device_lock);
4846 wait_event_lock_irq(conf->wait_for_quiescent,
4849 atomic_inc(&conf->active_aligned_reads);
4850 spin_unlock_irq(&conf->device_lock);
4853 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4854 align_bi, disk_devt(mddev->gendisk),
4855 raid_bio->bi_iter.bi_sector);
4856 generic_make_request(align_bi);
4865 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4870 sector_t sector = raid_bio->bi_iter.bi_sector;
4871 unsigned chunk_sects = mddev->chunk_sectors;
4872 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4874 if (sectors < bio_sectors(raid_bio)) {
4875 split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4876 bio_chain(split, raid_bio);
4880 if (!raid5_read_one_chunk(mddev, split)) {
4881 if (split != raid_bio)
4882 generic_make_request(raid_bio);
4885 } while (split != raid_bio);
4890 /* __get_priority_stripe - get the next stripe to process
4892 * Full stripe writes are allowed to pass preread active stripes up until
4893 * the bypass_threshold is exceeded. In general the bypass_count
4894 * increments when the handle_list is handled before the hold_list; however, it
4895 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4896 * stripe with in flight i/o. The bypass_count will be reset when the
4897 * head of the hold_list has changed, i.e. the head was promoted to the
4900 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4902 struct stripe_head *sh = NULL, *tmp;
4903 struct list_head *handle_list = NULL;
4904 struct r5worker_group *wg = NULL;
4906 if (conf->worker_cnt_per_group == 0) {
4907 handle_list = &conf->handle_list;
4908 } else if (group != ANY_GROUP) {
4909 handle_list = &conf->worker_groups[group].handle_list;
4910 wg = &conf->worker_groups[group];
4913 for (i = 0; i < conf->group_cnt; i++) {
4914 handle_list = &conf->worker_groups[i].handle_list;
4915 wg = &conf->worker_groups[i];
4916 if (!list_empty(handle_list))
4921 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4923 list_empty(handle_list) ? "empty" : "busy",
4924 list_empty(&conf->hold_list) ? "empty" : "busy",
4925 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4927 if (!list_empty(handle_list)) {
4928 sh = list_entry(handle_list->next, typeof(*sh), lru);
4930 if (list_empty(&conf->hold_list))
4931 conf->bypass_count = 0;
4932 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4933 if (conf->hold_list.next == conf->last_hold)
4934 conf->bypass_count++;
4936 conf->last_hold = conf->hold_list.next;
4937 conf->bypass_count -= conf->bypass_threshold;
4938 if (conf->bypass_count < 0)
4939 conf->bypass_count = 0;
4942 } else if (!list_empty(&conf->hold_list) &&
4943 ((conf->bypass_threshold &&
4944 conf->bypass_count > conf->bypass_threshold) ||
4945 atomic_read(&conf->pending_full_writes) == 0)) {
4947 list_for_each_entry(tmp, &conf->hold_list, lru) {
4948 if (conf->worker_cnt_per_group == 0 ||
4949 group == ANY_GROUP ||
4950 !cpu_online(tmp->cpu) ||
4951 cpu_to_group(tmp->cpu) == group) {
4958 conf->bypass_count -= conf->bypass_threshold;
4959 if (conf->bypass_count < 0)
4960 conf->bypass_count = 0;
4972 list_del_init(&sh->lru);
4973 BUG_ON(atomic_inc_return(&sh->count) != 1);
4977 struct raid5_plug_cb {
4978 struct blk_plug_cb cb;
4979 struct list_head list;
4980 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4983 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4985 struct raid5_plug_cb *cb = container_of(
4986 blk_cb, struct raid5_plug_cb, cb);
4987 struct stripe_head *sh;
4988 struct mddev *mddev = cb->cb.data;
4989 struct r5conf *conf = mddev->private;
4993 if (cb->list.next && !list_empty(&cb->list)) {
4994 spin_lock_irq(&conf->device_lock);
4995 while (!list_empty(&cb->list)) {
4996 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4997 list_del_init(&sh->lru);
4999 * avoid race release_stripe_plug() sees
5000 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5001 * is still in our list
5003 smp_mb__before_atomic();
5004 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5006 * STRIPE_ON_RELEASE_LIST could be set here. In that
5007 * case, the count is always > 1 here
5009 hash = sh->hash_lock_index;
5010 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5013 spin_unlock_irq(&conf->device_lock);
5015 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5016 NR_STRIPE_HASH_LOCKS);
5018 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5022 static void release_stripe_plug(struct mddev *mddev,
5023 struct stripe_head *sh)
5025 struct blk_plug_cb *blk_cb = blk_check_plugged(
5026 raid5_unplug, mddev,
5027 sizeof(struct raid5_plug_cb));
5028 struct raid5_plug_cb *cb;
5031 raid5_release_stripe(sh);
5035 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5037 if (cb->list.next == NULL) {
5039 INIT_LIST_HEAD(&cb->list);
5040 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5041 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5044 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5045 list_add_tail(&sh->lru, &cb->list);
5047 raid5_release_stripe(sh);
5050 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5052 struct r5conf *conf = mddev->private;
5053 sector_t logical_sector, last_sector;
5054 struct stripe_head *sh;
5058 if (mddev->reshape_position != MaxSector)
5059 /* Skip discard while reshape is happening */
5062 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5063 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5066 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5068 stripe_sectors = conf->chunk_sectors *
5069 (conf->raid_disks - conf->max_degraded);
5070 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5072 sector_div(last_sector, stripe_sectors);
5074 logical_sector *= conf->chunk_sectors;
5075 last_sector *= conf->chunk_sectors;
5077 for (; logical_sector < last_sector;
5078 logical_sector += STRIPE_SECTORS) {
5082 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5083 prepare_to_wait(&conf->wait_for_overlap, &w,
5084 TASK_UNINTERRUPTIBLE);
5085 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5086 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5087 raid5_release_stripe(sh);
5091 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5092 spin_lock_irq(&sh->stripe_lock);
5093 for (d = 0; d < conf->raid_disks; d++) {
5094 if (d == sh->pd_idx || d == sh->qd_idx)
5096 if (sh->dev[d].towrite || sh->dev[d].toread) {
5097 set_bit(R5_Overlap, &sh->dev[d].flags);
5098 spin_unlock_irq(&sh->stripe_lock);
5099 raid5_release_stripe(sh);
5104 set_bit(STRIPE_DISCARD, &sh->state);
5105 finish_wait(&conf->wait_for_overlap, &w);
5106 sh->overwrite_disks = 0;
5107 for (d = 0; d < conf->raid_disks; d++) {
5108 if (d == sh->pd_idx || d == sh->qd_idx)
5110 sh->dev[d].towrite = bi;
5111 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5112 raid5_inc_bi_active_stripes(bi);
5113 sh->overwrite_disks++;
5115 spin_unlock_irq(&sh->stripe_lock);
5116 if (conf->mddev->bitmap) {
5118 d < conf->raid_disks - conf->max_degraded;
5120 bitmap_startwrite(mddev->bitmap,
5124 sh->bm_seq = conf->seq_flush + 1;
5125 set_bit(STRIPE_BIT_DELAY, &sh->state);
5128 set_bit(STRIPE_HANDLE, &sh->state);
5129 clear_bit(STRIPE_DELAYED, &sh->state);
5130 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5131 atomic_inc(&conf->preread_active_stripes);
5132 release_stripe_plug(mddev, sh);
5135 remaining = raid5_dec_bi_active_stripes(bi);
5136 if (remaining == 0) {
5137 md_write_end(mddev);
5142 static void make_request(struct mddev *mddev, struct bio * bi)
5144 struct r5conf *conf = mddev->private;
5146 sector_t new_sector;
5147 sector_t logical_sector, last_sector;
5148 struct stripe_head *sh;
5149 const int rw = bio_data_dir(bi);
5154 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5155 int ret = r5l_handle_flush_request(conf->log, bi);
5159 if (ret == -ENODEV) {
5160 md_flush_request(mddev, bi);
5163 /* ret == -EAGAIN, fallback */
5166 md_write_start(mddev, bi);
5169 * If array is degraded, better not do chunk aligned read because
5170 * later we might have to read it again in order to reconstruct
5171 * data on failed drives.
5173 if (rw == READ && mddev->degraded == 0 &&
5174 mddev->reshape_position == MaxSector) {
5175 bi = chunk_aligned_read(mddev, bi);
5180 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5181 make_discard_request(mddev, bi);
5185 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5186 last_sector = bio_end_sector(bi);
5188 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5190 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5191 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5197 seq = read_seqcount_begin(&conf->gen_lock);
5200 prepare_to_wait(&conf->wait_for_overlap, &w,
5201 TASK_UNINTERRUPTIBLE);
5202 if (unlikely(conf->reshape_progress != MaxSector)) {
5203 /* spinlock is needed as reshape_progress may be
5204 * 64bit on a 32bit platform, and so it might be
5205 * possible to see a half-updated value
5206 * Of course reshape_progress could change after
5207 * the lock is dropped, so once we get a reference
5208 * to the stripe that we think it is, we will have
5211 spin_lock_irq(&conf->device_lock);
5212 if (mddev->reshape_backwards
5213 ? logical_sector < conf->reshape_progress
5214 : logical_sector >= conf->reshape_progress) {
5217 if (mddev->reshape_backwards
5218 ? logical_sector < conf->reshape_safe
5219 : logical_sector >= conf->reshape_safe) {
5220 spin_unlock_irq(&conf->device_lock);
5226 spin_unlock_irq(&conf->device_lock);
5229 new_sector = raid5_compute_sector(conf, logical_sector,
5232 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5233 (unsigned long long)new_sector,
5234 (unsigned long long)logical_sector);
5236 sh = raid5_get_active_stripe(conf, new_sector, previous,
5237 (bi->bi_rw&RWA_MASK), 0);
5239 if (unlikely(previous)) {
5240 /* expansion might have moved on while waiting for a
5241 * stripe, so we must do the range check again.
5242 * Expansion could still move past after this
5243 * test, but as we are holding a reference to
5244 * 'sh', we know that if that happens,
5245 * STRIPE_EXPANDING will get set and the expansion
5246 * won't proceed until we finish with the stripe.
5249 spin_lock_irq(&conf->device_lock);
5250 if (mddev->reshape_backwards
5251 ? logical_sector >= conf->reshape_progress
5252 : logical_sector < conf->reshape_progress)
5253 /* mismatch, need to try again */
5255 spin_unlock_irq(&conf->device_lock);
5257 raid5_release_stripe(sh);
5263 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5264 /* Might have got the wrong stripe_head
5267 raid5_release_stripe(sh);
5272 logical_sector >= mddev->suspend_lo &&
5273 logical_sector < mddev->suspend_hi) {
5274 raid5_release_stripe(sh);
5275 /* As the suspend_* range is controlled by
5276 * userspace, we want an interruptible
5279 flush_signals(current);
5280 prepare_to_wait(&conf->wait_for_overlap,
5281 &w, TASK_INTERRUPTIBLE);
5282 if (logical_sector >= mddev->suspend_lo &&
5283 logical_sector < mddev->suspend_hi) {
5290 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5291 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5292 /* Stripe is busy expanding or
5293 * add failed due to overlap. Flush everything
5296 md_wakeup_thread(mddev->thread);
5297 raid5_release_stripe(sh);
5302 set_bit(STRIPE_HANDLE, &sh->state);
5303 clear_bit(STRIPE_DELAYED, &sh->state);
5304 if ((!sh->batch_head || sh == sh->batch_head) &&
5305 (bi->bi_rw & REQ_SYNC) &&
5306 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5307 atomic_inc(&conf->preread_active_stripes);
5308 release_stripe_plug(mddev, sh);
5310 /* cannot get stripe for read-ahead, just give-up */
5311 bi->bi_error = -EIO;
5315 finish_wait(&conf->wait_for_overlap, &w);
5317 remaining = raid5_dec_bi_active_stripes(bi);
5318 if (remaining == 0) {
5321 md_write_end(mddev);
5323 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5329 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5331 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5333 /* reshaping is quite different to recovery/resync so it is
5334 * handled quite separately ... here.
5336 * On each call to sync_request, we gather one chunk worth of
5337 * destination stripes and flag them as expanding.
5338 * Then we find all the source stripes and request reads.
5339 * As the reads complete, handle_stripe will copy the data
5340 * into the destination stripe and release that stripe.
5342 struct r5conf *conf = mddev->private;
5343 struct stripe_head *sh;
5344 sector_t first_sector, last_sector;
5345 int raid_disks = conf->previous_raid_disks;
5346 int data_disks = raid_disks - conf->max_degraded;
5347 int new_data_disks = conf->raid_disks - conf->max_degraded;
5350 sector_t writepos, readpos, safepos;
5351 sector_t stripe_addr;
5352 int reshape_sectors;
5353 struct list_head stripes;
5356 if (sector_nr == 0) {
5357 /* If restarting in the middle, skip the initial sectors */
5358 if (mddev->reshape_backwards &&
5359 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5360 sector_nr = raid5_size(mddev, 0, 0)
5361 - conf->reshape_progress;
5362 } else if (mddev->reshape_backwards &&
5363 conf->reshape_progress == MaxSector) {
5364 /* shouldn't happen, but just in case, finish up.*/
5365 sector_nr = MaxSector;
5366 } else if (!mddev->reshape_backwards &&
5367 conf->reshape_progress > 0)
5368 sector_nr = conf->reshape_progress;
5369 sector_div(sector_nr, new_data_disks);
5371 mddev->curr_resync_completed = sector_nr;
5372 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5379 /* We need to process a full chunk at a time.
5380 * If old and new chunk sizes differ, we need to process the
5384 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5386 /* We update the metadata at least every 10 seconds, or when
5387 * the data about to be copied would over-write the source of
5388 * the data at the front of the range. i.e. one new_stripe
5389 * along from reshape_progress new_maps to after where
5390 * reshape_safe old_maps to
5392 writepos = conf->reshape_progress;
5393 sector_div(writepos, new_data_disks);
5394 readpos = conf->reshape_progress;
5395 sector_div(readpos, data_disks);
5396 safepos = conf->reshape_safe;
5397 sector_div(safepos, data_disks);
5398 if (mddev->reshape_backwards) {
5399 BUG_ON(writepos < reshape_sectors);
5400 writepos -= reshape_sectors;
5401 readpos += reshape_sectors;
5402 safepos += reshape_sectors;
5404 writepos += reshape_sectors;
5405 /* readpos and safepos are worst-case calculations.
5406 * A negative number is overly pessimistic, and causes
5407 * obvious problems for unsigned storage. So clip to 0.
5409 readpos -= min_t(sector_t, reshape_sectors, readpos);
5410 safepos -= min_t(sector_t, reshape_sectors, safepos);
5413 /* Having calculated the 'writepos' possibly use it
5414 * to set 'stripe_addr' which is where we will write to.
5416 if (mddev->reshape_backwards) {
5417 BUG_ON(conf->reshape_progress == 0);
5418 stripe_addr = writepos;
5419 BUG_ON((mddev->dev_sectors &
5420 ~((sector_t)reshape_sectors - 1))
5421 - reshape_sectors - stripe_addr
5424 BUG_ON(writepos != sector_nr + reshape_sectors);
5425 stripe_addr = sector_nr;
5428 /* 'writepos' is the most advanced device address we might write.
5429 * 'readpos' is the least advanced device address we might read.
5430 * 'safepos' is the least address recorded in the metadata as having
5432 * If there is a min_offset_diff, these are adjusted either by
5433 * increasing the safepos/readpos if diff is negative, or
5434 * increasing writepos if diff is positive.
5435 * If 'readpos' is then behind 'writepos', there is no way that we can
5436 * ensure safety in the face of a crash - that must be done by userspace
5437 * making a backup of the data. So in that case there is no particular
5438 * rush to update metadata.
5439 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5440 * update the metadata to advance 'safepos' to match 'readpos' so that
5441 * we can be safe in the event of a crash.
5442 * So we insist on updating metadata if safepos is behind writepos and
5443 * readpos is beyond writepos.
5444 * In any case, update the metadata every 10 seconds.
5445 * Maybe that number should be configurable, but I'm not sure it is
5446 * worth it.... maybe it could be a multiple of safemode_delay???
5448 if (conf->min_offset_diff < 0) {
5449 safepos += -conf->min_offset_diff;
5450 readpos += -conf->min_offset_diff;
5452 writepos += conf->min_offset_diff;
5454 if ((mddev->reshape_backwards
5455 ? (safepos > writepos && readpos < writepos)
5456 : (safepos < writepos && readpos > writepos)) ||
5457 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5458 /* Cannot proceed until we've updated the superblock... */
5459 wait_event(conf->wait_for_overlap,
5460 atomic_read(&conf->reshape_stripes)==0
5461 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5462 if (atomic_read(&conf->reshape_stripes) != 0)
5464 mddev->reshape_position = conf->reshape_progress;
5465 mddev->curr_resync_completed = sector_nr;
5466 conf->reshape_checkpoint = jiffies;
5467 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5468 md_wakeup_thread(mddev->thread);
5469 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5470 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5471 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5473 spin_lock_irq(&conf->device_lock);
5474 conf->reshape_safe = mddev->reshape_position;
5475 spin_unlock_irq(&conf->device_lock);
5476 wake_up(&conf->wait_for_overlap);
5477 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5480 INIT_LIST_HEAD(&stripes);
5481 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5483 int skipped_disk = 0;
5484 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5485 set_bit(STRIPE_EXPANDING, &sh->state);
5486 atomic_inc(&conf->reshape_stripes);
5487 /* If any of this stripe is beyond the end of the old
5488 * array, then we need to zero those blocks
5490 for (j=sh->disks; j--;) {
5492 if (j == sh->pd_idx)
5494 if (conf->level == 6 &&
5497 s = raid5_compute_blocknr(sh, j, 0);
5498 if (s < raid5_size(mddev, 0, 0)) {
5502 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5503 set_bit(R5_Expanded, &sh->dev[j].flags);
5504 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5506 if (!skipped_disk) {
5507 set_bit(STRIPE_EXPAND_READY, &sh->state);
5508 set_bit(STRIPE_HANDLE, &sh->state);
5510 list_add(&sh->lru, &stripes);
5512 spin_lock_irq(&conf->device_lock);
5513 if (mddev->reshape_backwards)
5514 conf->reshape_progress -= reshape_sectors * new_data_disks;
5516 conf->reshape_progress += reshape_sectors * new_data_disks;
5517 spin_unlock_irq(&conf->device_lock);
5518 /* Ok, those stripe are ready. We can start scheduling
5519 * reads on the source stripes.
5520 * The source stripes are determined by mapping the first and last
5521 * block on the destination stripes.
5524 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5527 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5528 * new_data_disks - 1),
5530 if (last_sector >= mddev->dev_sectors)
5531 last_sector = mddev->dev_sectors - 1;
5532 while (first_sector <= last_sector) {
5533 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5534 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5535 set_bit(STRIPE_HANDLE, &sh->state);
5536 raid5_release_stripe(sh);
5537 first_sector += STRIPE_SECTORS;
5539 /* Now that the sources are clearly marked, we can release
5540 * the destination stripes
5542 while (!list_empty(&stripes)) {
5543 sh = list_entry(stripes.next, struct stripe_head, lru);
5544 list_del_init(&sh->lru);
5545 raid5_release_stripe(sh);
5547 /* If this takes us to the resync_max point where we have to pause,
5548 * then we need to write out the superblock.
5550 sector_nr += reshape_sectors;
5551 retn = reshape_sectors;
5553 if (mddev->curr_resync_completed > mddev->resync_max ||
5554 (sector_nr - mddev->curr_resync_completed) * 2
5555 >= mddev->resync_max - mddev->curr_resync_completed) {
5556 /* Cannot proceed until we've updated the superblock... */
5557 wait_event(conf->wait_for_overlap,
5558 atomic_read(&conf->reshape_stripes) == 0
5559 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5560 if (atomic_read(&conf->reshape_stripes) != 0)
5562 mddev->reshape_position = conf->reshape_progress;
5563 mddev->curr_resync_completed = sector_nr;
5564 conf->reshape_checkpoint = jiffies;
5565 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5566 md_wakeup_thread(mddev->thread);
5567 wait_event(mddev->sb_wait,
5568 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5569 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5570 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5572 spin_lock_irq(&conf->device_lock);
5573 conf->reshape_safe = mddev->reshape_position;
5574 spin_unlock_irq(&conf->device_lock);
5575 wake_up(&conf->wait_for_overlap);
5576 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5582 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5584 struct r5conf *conf = mddev->private;
5585 struct stripe_head *sh;
5586 sector_t max_sector = mddev->dev_sectors;
5587 sector_t sync_blocks;
5588 int still_degraded = 0;
5591 if (sector_nr >= max_sector) {
5592 /* just being told to finish up .. nothing much to do */
5594 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5599 if (mddev->curr_resync < max_sector) /* aborted */
5600 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5602 else /* completed sync */
5604 bitmap_close_sync(mddev->bitmap);
5609 /* Allow raid5_quiesce to complete */
5610 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5612 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5613 return reshape_request(mddev, sector_nr, skipped);
5615 /* No need to check resync_max as we never do more than one
5616 * stripe, and as resync_max will always be on a chunk boundary,
5617 * if the check in md_do_sync didn't fire, there is no chance
5618 * of overstepping resync_max here
5621 /* if there is too many failed drives and we are trying
5622 * to resync, then assert that we are finished, because there is
5623 * nothing we can do.
5625 if (mddev->degraded >= conf->max_degraded &&
5626 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5627 sector_t rv = mddev->dev_sectors - sector_nr;
5631 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5633 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5634 sync_blocks >= STRIPE_SECTORS) {
5635 /* we can skip this block, and probably more */
5636 sync_blocks /= STRIPE_SECTORS;
5638 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5641 bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5643 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5645 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5646 /* make sure we don't swamp the stripe cache if someone else
5647 * is trying to get access
5649 schedule_timeout_uninterruptible(1);
5651 /* Need to check if array will still be degraded after recovery/resync
5652 * Note in case of > 1 drive failures it's possible we're rebuilding
5653 * one drive while leaving another faulty drive in array.
5656 for (i = 0; i < conf->raid_disks; i++) {
5657 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5659 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5664 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5666 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5667 set_bit(STRIPE_HANDLE, &sh->state);
5669 raid5_release_stripe(sh);
5671 return STRIPE_SECTORS;
5674 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5676 /* We may not be able to submit a whole bio at once as there
5677 * may not be enough stripe_heads available.
5678 * We cannot pre-allocate enough stripe_heads as we may need
5679 * more than exist in the cache (if we allow ever large chunks).
5680 * So we do one stripe head at a time and record in
5681 * ->bi_hw_segments how many have been done.
5683 * We *know* that this entire raid_bio is in one chunk, so
5684 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5686 struct stripe_head *sh;
5688 sector_t sector, logical_sector, last_sector;
5693 logical_sector = raid_bio->bi_iter.bi_sector &
5694 ~((sector_t)STRIPE_SECTORS-1);
5695 sector = raid5_compute_sector(conf, logical_sector,
5697 last_sector = bio_end_sector(raid_bio);
5699 for (; logical_sector < last_sector;
5700 logical_sector += STRIPE_SECTORS,
5701 sector += STRIPE_SECTORS,
5704 if (scnt < raid5_bi_processed_stripes(raid_bio))
5705 /* already done this stripe */
5708 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5711 /* failed to get a stripe - must wait */
5712 raid5_set_bi_processed_stripes(raid_bio, scnt);
5713 conf->retry_read_aligned = raid_bio;
5717 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5718 raid5_release_stripe(sh);
5719 raid5_set_bi_processed_stripes(raid_bio, scnt);
5720 conf->retry_read_aligned = raid_bio;
5724 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5726 raid5_release_stripe(sh);
5729 remaining = raid5_dec_bi_active_stripes(raid_bio);
5730 if (remaining == 0) {
5731 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5733 bio_endio(raid_bio);
5735 if (atomic_dec_and_test(&conf->active_aligned_reads))
5736 wake_up(&conf->wait_for_quiescent);
5740 static int handle_active_stripes(struct r5conf *conf, int group,
5741 struct r5worker *worker,
5742 struct list_head *temp_inactive_list)
5744 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5745 int i, batch_size = 0, hash;
5746 bool release_inactive = false;
5748 while (batch_size < MAX_STRIPE_BATCH &&
5749 (sh = __get_priority_stripe(conf, group)) != NULL)
5750 batch[batch_size++] = sh;
5752 if (batch_size == 0) {
5753 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5754 if (!list_empty(temp_inactive_list + i))
5756 if (i == NR_STRIPE_HASH_LOCKS) {
5757 spin_unlock_irq(&conf->device_lock);
5758 r5l_flush_stripe_to_raid(conf->log);
5759 spin_lock_irq(&conf->device_lock);
5762 release_inactive = true;
5764 spin_unlock_irq(&conf->device_lock);
5766 release_inactive_stripe_list(conf, temp_inactive_list,
5767 NR_STRIPE_HASH_LOCKS);
5769 r5l_flush_stripe_to_raid(conf->log);
5770 if (release_inactive) {
5771 spin_lock_irq(&conf->device_lock);
5775 for (i = 0; i < batch_size; i++)
5776 handle_stripe(batch[i]);
5777 r5l_write_stripe_run(conf->log);
5781 spin_lock_irq(&conf->device_lock);
5782 for (i = 0; i < batch_size; i++) {
5783 hash = batch[i]->hash_lock_index;
5784 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5789 static void raid5_do_work(struct work_struct *work)
5791 struct r5worker *worker = container_of(work, struct r5worker, work);
5792 struct r5worker_group *group = worker->group;
5793 struct r5conf *conf = group->conf;
5794 int group_id = group - conf->worker_groups;
5796 struct blk_plug plug;
5798 pr_debug("+++ raid5worker active\n");
5800 blk_start_plug(&plug);
5802 spin_lock_irq(&conf->device_lock);
5804 int batch_size, released;
5806 released = release_stripe_list(conf, worker->temp_inactive_list);
5808 batch_size = handle_active_stripes(conf, group_id, worker,
5809 worker->temp_inactive_list);
5810 worker->working = false;
5811 if (!batch_size && !released)
5813 handled += batch_size;
5815 pr_debug("%d stripes handled\n", handled);
5817 spin_unlock_irq(&conf->device_lock);
5818 blk_finish_plug(&plug);
5820 pr_debug("--- raid5worker inactive\n");
5824 * This is our raid5 kernel thread.
5826 * We scan the hash table for stripes which can be handled now.
5827 * During the scan, completed stripes are saved for us by the interrupt
5828 * handler, so that they will not have to wait for our next wakeup.
5830 static void raid5d(struct md_thread *thread)
5832 struct mddev *mddev = thread->mddev;
5833 struct r5conf *conf = mddev->private;
5835 struct blk_plug plug;
5837 pr_debug("+++ raid5d active\n");
5839 md_check_recovery(mddev);
5841 if (!bio_list_empty(&conf->return_bi) &&
5842 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5843 struct bio_list tmp = BIO_EMPTY_LIST;
5844 spin_lock_irq(&conf->device_lock);
5845 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5846 bio_list_merge(&tmp, &conf->return_bi);
5847 bio_list_init(&conf->return_bi);
5849 spin_unlock_irq(&conf->device_lock);
5853 blk_start_plug(&plug);
5855 spin_lock_irq(&conf->device_lock);
5858 int batch_size, released;
5860 released = release_stripe_list(conf, conf->temp_inactive_list);
5862 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5865 !list_empty(&conf->bitmap_list)) {
5866 /* Now is a good time to flush some bitmap updates */
5868 spin_unlock_irq(&conf->device_lock);
5869 bitmap_unplug(mddev->bitmap);
5870 spin_lock_irq(&conf->device_lock);
5871 conf->seq_write = conf->seq_flush;
5872 activate_bit_delay(conf, conf->temp_inactive_list);
5874 raid5_activate_delayed(conf);
5876 while ((bio = remove_bio_from_retry(conf))) {
5878 spin_unlock_irq(&conf->device_lock);
5879 ok = retry_aligned_read(conf, bio);
5880 spin_lock_irq(&conf->device_lock);
5886 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5887 conf->temp_inactive_list);
5888 if (!batch_size && !released)
5890 handled += batch_size;
5892 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5893 spin_unlock_irq(&conf->device_lock);
5894 md_check_recovery(mddev);
5895 spin_lock_irq(&conf->device_lock);
5898 pr_debug("%d stripes handled\n", handled);
5900 spin_unlock_irq(&conf->device_lock);
5901 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5902 mutex_trylock(&conf->cache_size_mutex)) {
5903 grow_one_stripe(conf, __GFP_NOWARN);
5904 /* Set flag even if allocation failed. This helps
5905 * slow down allocation requests when mem is short
5907 set_bit(R5_DID_ALLOC, &conf->cache_state);
5908 mutex_unlock(&conf->cache_size_mutex);
5911 r5l_flush_stripe_to_raid(conf->log);
5913 async_tx_issue_pending_all();
5914 blk_finish_plug(&plug);
5916 pr_debug("--- raid5d inactive\n");
5920 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5922 struct r5conf *conf;
5924 spin_lock(&mddev->lock);
5925 conf = mddev->private;
5927 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5928 spin_unlock(&mddev->lock);
5933 raid5_set_cache_size(struct mddev *mddev, int size)
5935 struct r5conf *conf = mddev->private;
5938 if (size <= 16 || size > 32768)
5941 conf->min_nr_stripes = size;
5942 mutex_lock(&conf->cache_size_mutex);
5943 while (size < conf->max_nr_stripes &&
5944 drop_one_stripe(conf))
5946 mutex_unlock(&conf->cache_size_mutex);
5949 err = md_allow_write(mddev);
5953 mutex_lock(&conf->cache_size_mutex);
5954 while (size > conf->max_nr_stripes)
5955 if (!grow_one_stripe(conf, GFP_KERNEL))
5957 mutex_unlock(&conf->cache_size_mutex);
5961 EXPORT_SYMBOL(raid5_set_cache_size);
5964 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5966 struct r5conf *conf;
5970 if (len >= PAGE_SIZE)
5972 if (kstrtoul(page, 10, &new))
5974 err = mddev_lock(mddev);
5977 conf = mddev->private;
5981 err = raid5_set_cache_size(mddev, new);
5982 mddev_unlock(mddev);
5987 static struct md_sysfs_entry
5988 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5989 raid5_show_stripe_cache_size,
5990 raid5_store_stripe_cache_size);
5993 raid5_show_rmw_level(struct mddev *mddev, char *page)
5995 struct r5conf *conf = mddev->private;
5997 return sprintf(page, "%d\n", conf->rmw_level);
6003 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6005 struct r5conf *conf = mddev->private;
6011 if (len >= PAGE_SIZE)
6014 if (kstrtoul(page, 10, &new))
6017 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6020 if (new != PARITY_DISABLE_RMW &&
6021 new != PARITY_ENABLE_RMW &&
6022 new != PARITY_PREFER_RMW)
6025 conf->rmw_level = new;
6029 static struct md_sysfs_entry
6030 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6031 raid5_show_rmw_level,
6032 raid5_store_rmw_level);
6036 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6038 struct r5conf *conf;
6040 spin_lock(&mddev->lock);
6041 conf = mddev->private;
6043 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6044 spin_unlock(&mddev->lock);
6049 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6051 struct r5conf *conf;
6055 if (len >= PAGE_SIZE)
6057 if (kstrtoul(page, 10, &new))
6060 err = mddev_lock(mddev);
6063 conf = mddev->private;
6066 else if (new > conf->min_nr_stripes)
6069 conf->bypass_threshold = new;
6070 mddev_unlock(mddev);
6074 static struct md_sysfs_entry
6075 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6077 raid5_show_preread_threshold,
6078 raid5_store_preread_threshold);
6081 raid5_show_skip_copy(struct mddev *mddev, char *page)
6083 struct r5conf *conf;
6085 spin_lock(&mddev->lock);
6086 conf = mddev->private;
6088 ret = sprintf(page, "%d\n", conf->skip_copy);
6089 spin_unlock(&mddev->lock);
6094 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6096 struct r5conf *conf;
6100 if (len >= PAGE_SIZE)
6102 if (kstrtoul(page, 10, &new))
6106 err = mddev_lock(mddev);
6109 conf = mddev->private;
6112 else if (new != conf->skip_copy) {
6113 mddev_suspend(mddev);
6114 conf->skip_copy = new;
6116 mddev->queue->backing_dev_info.capabilities |=
6117 BDI_CAP_STABLE_WRITES;
6119 mddev->queue->backing_dev_info.capabilities &=
6120 ~BDI_CAP_STABLE_WRITES;
6121 mddev_resume(mddev);
6123 mddev_unlock(mddev);
6127 static struct md_sysfs_entry
6128 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6129 raid5_show_skip_copy,
6130 raid5_store_skip_copy);
6133 stripe_cache_active_show(struct mddev *mddev, char *page)
6135 struct r5conf *conf = mddev->private;
6137 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6142 static struct md_sysfs_entry
6143 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6146 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6148 struct r5conf *conf;
6150 spin_lock(&mddev->lock);
6151 conf = mddev->private;
6153 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6154 spin_unlock(&mddev->lock);
6158 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6160 int *worker_cnt_per_group,
6161 struct r5worker_group **worker_groups);
6163 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6165 struct r5conf *conf;
6168 struct r5worker_group *new_groups, *old_groups;
6169 int group_cnt, worker_cnt_per_group;
6171 if (len >= PAGE_SIZE)
6173 if (kstrtoul(page, 10, &new))
6176 err = mddev_lock(mddev);
6179 conf = mddev->private;
6182 else if (new != conf->worker_cnt_per_group) {
6183 mddev_suspend(mddev);
6185 old_groups = conf->worker_groups;
6187 flush_workqueue(raid5_wq);
6189 err = alloc_thread_groups(conf, new,
6190 &group_cnt, &worker_cnt_per_group,
6193 spin_lock_irq(&conf->device_lock);
6194 conf->group_cnt = group_cnt;
6195 conf->worker_cnt_per_group = worker_cnt_per_group;
6196 conf->worker_groups = new_groups;
6197 spin_unlock_irq(&conf->device_lock);
6200 kfree(old_groups[0].workers);
6203 mddev_resume(mddev);
6205 mddev_unlock(mddev);
6210 static struct md_sysfs_entry
6211 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6212 raid5_show_group_thread_cnt,
6213 raid5_store_group_thread_cnt);
6215 static struct attribute *raid5_attrs[] = {
6216 &raid5_stripecache_size.attr,
6217 &raid5_stripecache_active.attr,
6218 &raid5_preread_bypass_threshold.attr,
6219 &raid5_group_thread_cnt.attr,
6220 &raid5_skip_copy.attr,
6221 &raid5_rmw_level.attr,
6224 static struct attribute_group raid5_attrs_group = {
6226 .attrs = raid5_attrs,
6229 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6231 int *worker_cnt_per_group,
6232 struct r5worker_group **worker_groups)
6236 struct r5worker *workers;
6238 *worker_cnt_per_group = cnt;
6241 *worker_groups = NULL;
6244 *group_cnt = num_possible_nodes();
6245 size = sizeof(struct r5worker) * cnt;
6246 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6247 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6248 *group_cnt, GFP_NOIO);
6249 if (!*worker_groups || !workers) {
6251 kfree(*worker_groups);
6255 for (i = 0; i < *group_cnt; i++) {
6256 struct r5worker_group *group;
6258 group = &(*worker_groups)[i];
6259 INIT_LIST_HEAD(&group->handle_list);
6261 group->workers = workers + i * cnt;
6263 for (j = 0; j < cnt; j++) {
6264 struct r5worker *worker = group->workers + j;
6265 worker->group = group;
6266 INIT_WORK(&worker->work, raid5_do_work);
6268 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6269 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6276 static void free_thread_groups(struct r5conf *conf)
6278 if (conf->worker_groups)
6279 kfree(conf->worker_groups[0].workers);
6280 kfree(conf->worker_groups);
6281 conf->worker_groups = NULL;
6285 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6287 struct r5conf *conf = mddev->private;
6290 sectors = mddev->dev_sectors;
6292 /* size is defined by the smallest of previous and new size */
6293 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6295 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6296 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6297 return sectors * (raid_disks - conf->max_degraded);
6300 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6302 safe_put_page(percpu->spare_page);
6303 if (percpu->scribble)
6304 flex_array_free(percpu->scribble);
6305 percpu->spare_page = NULL;
6306 percpu->scribble = NULL;
6309 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6311 if (conf->level == 6 && !percpu->spare_page)
6312 percpu->spare_page = alloc_page(GFP_KERNEL);
6313 if (!percpu->scribble)
6314 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6315 conf->previous_raid_disks),
6316 max(conf->chunk_sectors,
6317 conf->prev_chunk_sectors)
6321 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6322 free_scratch_buffer(conf, percpu);
6329 static void raid5_free_percpu(struct r5conf *conf)
6336 #ifdef CONFIG_HOTPLUG_CPU
6337 unregister_cpu_notifier(&conf->cpu_notify);
6341 for_each_possible_cpu(cpu)
6342 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6345 free_percpu(conf->percpu);
6348 static void free_conf(struct r5conf *conf)
6351 r5l_exit_log(conf->log);
6352 if (conf->shrinker.seeks)
6353 unregister_shrinker(&conf->shrinker);
6355 free_thread_groups(conf);
6356 shrink_stripes(conf);
6357 raid5_free_percpu(conf);
6359 kfree(conf->stripe_hashtbl);
6363 #ifdef CONFIG_HOTPLUG_CPU
6364 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6367 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6368 long cpu = (long)hcpu;
6369 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6372 case CPU_UP_PREPARE:
6373 case CPU_UP_PREPARE_FROZEN:
6374 if (alloc_scratch_buffer(conf, percpu)) {
6375 pr_err("%s: failed memory allocation for cpu%ld\n",
6377 return notifier_from_errno(-ENOMEM);
6381 case CPU_DEAD_FROZEN:
6382 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6391 static int raid5_alloc_percpu(struct r5conf *conf)
6396 conf->percpu = alloc_percpu(struct raid5_percpu);
6400 #ifdef CONFIG_HOTPLUG_CPU
6401 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6402 conf->cpu_notify.priority = 0;
6403 err = register_cpu_notifier(&conf->cpu_notify);
6409 for_each_present_cpu(cpu) {
6410 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6412 pr_err("%s: failed memory allocation for cpu%ld\n",
6416 spin_lock_init(&per_cpu_ptr(conf->percpu, cpu)->lock);
6423 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6424 struct shrink_control *sc)
6426 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6427 unsigned long ret = SHRINK_STOP;
6429 if (mutex_trylock(&conf->cache_size_mutex)) {
6431 while (ret < sc->nr_to_scan &&
6432 conf->max_nr_stripes > conf->min_nr_stripes) {
6433 if (drop_one_stripe(conf) == 0) {
6439 mutex_unlock(&conf->cache_size_mutex);
6444 static unsigned long raid5_cache_count(struct shrinker *shrink,
6445 struct shrink_control *sc)
6447 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6449 if (conf->max_nr_stripes < conf->min_nr_stripes)
6450 /* unlikely, but not impossible */
6452 return conf->max_nr_stripes - conf->min_nr_stripes;
6455 static struct r5conf *setup_conf(struct mddev *mddev)
6457 struct r5conf *conf;
6458 int raid_disk, memory, max_disks;
6459 struct md_rdev *rdev;
6460 struct disk_info *disk;
6463 int group_cnt, worker_cnt_per_group;
6464 struct r5worker_group *new_group;
6466 if (mddev->new_level != 5
6467 && mddev->new_level != 4
6468 && mddev->new_level != 6) {
6469 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6470 mdname(mddev), mddev->new_level);
6471 return ERR_PTR(-EIO);
6473 if ((mddev->new_level == 5
6474 && !algorithm_valid_raid5(mddev->new_layout)) ||
6475 (mddev->new_level == 6
6476 && !algorithm_valid_raid6(mddev->new_layout))) {
6477 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6478 mdname(mddev), mddev->new_layout);
6479 return ERR_PTR(-EIO);
6481 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6482 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6483 mdname(mddev), mddev->raid_disks);
6484 return ERR_PTR(-EINVAL);
6487 if (!mddev->new_chunk_sectors ||
6488 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6489 !is_power_of_2(mddev->new_chunk_sectors)) {
6490 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6491 mdname(mddev), mddev->new_chunk_sectors << 9);
6492 return ERR_PTR(-EINVAL);
6495 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6498 /* Don't enable multi-threading by default*/
6499 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6501 conf->group_cnt = group_cnt;
6502 conf->worker_cnt_per_group = worker_cnt_per_group;
6503 conf->worker_groups = new_group;
6506 spin_lock_init(&conf->device_lock);
6507 seqcount_init(&conf->gen_lock);
6508 mutex_init(&conf->cache_size_mutex);
6509 init_waitqueue_head(&conf->wait_for_quiescent);
6510 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
6511 init_waitqueue_head(&conf->wait_for_stripe[i]);
6513 init_waitqueue_head(&conf->wait_for_overlap);
6514 INIT_LIST_HEAD(&conf->handle_list);
6515 INIT_LIST_HEAD(&conf->hold_list);
6516 INIT_LIST_HEAD(&conf->delayed_list);
6517 INIT_LIST_HEAD(&conf->bitmap_list);
6518 bio_list_init(&conf->return_bi);
6519 init_llist_head(&conf->released_stripes);
6520 atomic_set(&conf->active_stripes, 0);
6521 atomic_set(&conf->preread_active_stripes, 0);
6522 atomic_set(&conf->active_aligned_reads, 0);
6523 conf->bypass_threshold = BYPASS_THRESHOLD;
6524 conf->recovery_disabled = mddev->recovery_disabled - 1;
6526 conf->raid_disks = mddev->raid_disks;
6527 if (mddev->reshape_position == MaxSector)
6528 conf->previous_raid_disks = mddev->raid_disks;
6530 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6531 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6533 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6538 conf->mddev = mddev;
6540 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6543 /* We init hash_locks[0] separately to that it can be used
6544 * as the reference lock in the spin_lock_nest_lock() call
6545 * in lock_all_device_hash_locks_irq in order to convince
6546 * lockdep that we know what we are doing.
6548 spin_lock_init(conf->hash_locks);
6549 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6550 spin_lock_init(conf->hash_locks + i);
6552 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6553 INIT_LIST_HEAD(conf->inactive_list + i);
6555 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6556 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6558 conf->level = mddev->new_level;
6559 conf->chunk_sectors = mddev->new_chunk_sectors;
6560 if (raid5_alloc_percpu(conf) != 0)
6563 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6565 rdev_for_each(rdev, mddev) {
6566 raid_disk = rdev->raid_disk;
6567 if (raid_disk >= max_disks
6568 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6570 disk = conf->disks + raid_disk;
6572 if (test_bit(Replacement, &rdev->flags)) {
6573 if (disk->replacement)
6575 disk->replacement = rdev;
6582 if (test_bit(In_sync, &rdev->flags)) {
6583 char b[BDEVNAME_SIZE];
6584 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6586 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6587 } else if (rdev->saved_raid_disk != raid_disk)
6588 /* Cannot rely on bitmap to complete recovery */
6592 conf->level = mddev->new_level;
6593 if (conf->level == 6) {
6594 conf->max_degraded = 2;
6595 if (raid6_call.xor_syndrome)
6596 conf->rmw_level = PARITY_ENABLE_RMW;
6598 conf->rmw_level = PARITY_DISABLE_RMW;
6600 conf->max_degraded = 1;
6601 conf->rmw_level = PARITY_ENABLE_RMW;
6603 conf->algorithm = mddev->new_layout;
6604 conf->reshape_progress = mddev->reshape_position;
6605 if (conf->reshape_progress != MaxSector) {
6606 conf->prev_chunk_sectors = mddev->chunk_sectors;
6607 conf->prev_algo = mddev->layout;
6609 conf->prev_chunk_sectors = conf->chunk_sectors;
6610 conf->prev_algo = conf->algorithm;
6613 conf->min_nr_stripes = NR_STRIPES;
6614 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6615 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6616 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6617 if (grow_stripes(conf, conf->min_nr_stripes)) {
6619 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6620 mdname(mddev), memory);
6623 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6624 mdname(mddev), memory);
6626 * Losing a stripe head costs more than the time to refill it,
6627 * it reduces the queue depth and so can hurt throughput.
6628 * So set it rather large, scaled by number of devices.
6630 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6631 conf->shrinker.scan_objects = raid5_cache_scan;
6632 conf->shrinker.count_objects = raid5_cache_count;
6633 conf->shrinker.batch = 128;
6634 conf->shrinker.flags = 0;
6635 register_shrinker(&conf->shrinker);
6637 sprintf(pers_name, "raid%d", mddev->new_level);
6638 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6639 if (!conf->thread) {
6641 "md/raid:%s: couldn't allocate thread.\n",
6651 return ERR_PTR(-EIO);
6653 return ERR_PTR(-ENOMEM);
6656 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6659 case ALGORITHM_PARITY_0:
6660 if (raid_disk < max_degraded)
6663 case ALGORITHM_PARITY_N:
6664 if (raid_disk >= raid_disks - max_degraded)
6667 case ALGORITHM_PARITY_0_6:
6668 if (raid_disk == 0 ||
6669 raid_disk == raid_disks - 1)
6672 case ALGORITHM_LEFT_ASYMMETRIC_6:
6673 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6674 case ALGORITHM_LEFT_SYMMETRIC_6:
6675 case ALGORITHM_RIGHT_SYMMETRIC_6:
6676 if (raid_disk == raid_disks - 1)
6682 static int run(struct mddev *mddev)
6684 struct r5conf *conf;
6685 int working_disks = 0;
6686 int dirty_parity_disks = 0;
6687 struct md_rdev *rdev;
6688 struct md_rdev *journal_dev = NULL;
6689 sector_t reshape_offset = 0;
6691 long long min_offset_diff = 0;
6694 if (mddev->recovery_cp != MaxSector)
6695 printk(KERN_NOTICE "md/raid:%s: not clean"
6696 " -- starting background reconstruction\n",
6699 rdev_for_each(rdev, mddev) {
6702 if (test_bit(Journal, &rdev->flags)) {
6706 if (rdev->raid_disk < 0)
6708 diff = (rdev->new_data_offset - rdev->data_offset);
6710 min_offset_diff = diff;
6712 } else if (mddev->reshape_backwards &&
6713 diff < min_offset_diff)
6714 min_offset_diff = diff;
6715 else if (!mddev->reshape_backwards &&
6716 diff > min_offset_diff)
6717 min_offset_diff = diff;
6720 if (mddev->reshape_position != MaxSector) {
6721 /* Check that we can continue the reshape.
6722 * Difficulties arise if the stripe we would write to
6723 * next is at or after the stripe we would read from next.
6724 * For a reshape that changes the number of devices, this
6725 * is only possible for a very short time, and mdadm makes
6726 * sure that time appears to have past before assembling
6727 * the array. So we fail if that time hasn't passed.
6728 * For a reshape that keeps the number of devices the same
6729 * mdadm must be monitoring the reshape can keeping the
6730 * critical areas read-only and backed up. It will start
6731 * the array in read-only mode, so we check for that.
6733 sector_t here_new, here_old;
6735 int max_degraded = (mddev->level == 6 ? 2 : 1);
6740 printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6745 if (mddev->new_level != mddev->level) {
6746 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6747 "required - aborting.\n",
6751 old_disks = mddev->raid_disks - mddev->delta_disks;
6752 /* reshape_position must be on a new-stripe boundary, and one
6753 * further up in new geometry must map after here in old
6755 * If the chunk sizes are different, then as we perform reshape
6756 * in units of the largest of the two, reshape_position needs
6757 * be a multiple of the largest chunk size times new data disks.
6759 here_new = mddev->reshape_position;
6760 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6761 new_data_disks = mddev->raid_disks - max_degraded;
6762 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6763 printk(KERN_ERR "md/raid:%s: reshape_position not "
6764 "on a stripe boundary\n", mdname(mddev));
6767 reshape_offset = here_new * chunk_sectors;
6768 /* here_new is the stripe we will write to */
6769 here_old = mddev->reshape_position;
6770 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6771 /* here_old is the first stripe that we might need to read
6773 if (mddev->delta_disks == 0) {
6774 /* We cannot be sure it is safe to start an in-place
6775 * reshape. It is only safe if user-space is monitoring
6776 * and taking constant backups.
6777 * mdadm always starts a situation like this in
6778 * readonly mode so it can take control before
6779 * allowing any writes. So just check for that.
6781 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6782 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6783 /* not really in-place - so OK */;
6784 else if (mddev->ro == 0) {
6785 printk(KERN_ERR "md/raid:%s: in-place reshape "
6786 "must be started in read-only mode "
6791 } else if (mddev->reshape_backwards
6792 ? (here_new * chunk_sectors + min_offset_diff <=
6793 here_old * chunk_sectors)
6794 : (here_new * chunk_sectors >=
6795 here_old * chunk_sectors + (-min_offset_diff))) {
6796 /* Reading from the same stripe as writing to - bad */
6797 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6798 "auto-recovery - aborting.\n",
6802 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6804 /* OK, we should be able to continue; */
6806 BUG_ON(mddev->level != mddev->new_level);
6807 BUG_ON(mddev->layout != mddev->new_layout);
6808 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6809 BUG_ON(mddev->delta_disks != 0);
6812 if (mddev->private == NULL)
6813 conf = setup_conf(mddev);
6815 conf = mddev->private;
6818 return PTR_ERR(conf);
6820 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) && !journal_dev) {
6821 printk(KERN_ERR "md/raid:%s: journal disk is missing, force array readonly\n",
6824 set_disk_ro(mddev->gendisk, 1);
6827 conf->min_offset_diff = min_offset_diff;
6828 mddev->thread = conf->thread;
6829 conf->thread = NULL;
6830 mddev->private = conf;
6832 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6834 rdev = conf->disks[i].rdev;
6835 if (!rdev && conf->disks[i].replacement) {
6836 /* The replacement is all we have yet */
6837 rdev = conf->disks[i].replacement;
6838 conf->disks[i].replacement = NULL;
6839 clear_bit(Replacement, &rdev->flags);
6840 conf->disks[i].rdev = rdev;
6844 if (conf->disks[i].replacement &&
6845 conf->reshape_progress != MaxSector) {
6846 /* replacements and reshape simply do not mix. */
6847 printk(KERN_ERR "md: cannot handle concurrent "
6848 "replacement and reshape.\n");
6851 if (test_bit(In_sync, &rdev->flags)) {
6855 /* This disc is not fully in-sync. However if it
6856 * just stored parity (beyond the recovery_offset),
6857 * when we don't need to be concerned about the
6858 * array being dirty.
6859 * When reshape goes 'backwards', we never have
6860 * partially completed devices, so we only need
6861 * to worry about reshape going forwards.
6863 /* Hack because v0.91 doesn't store recovery_offset properly. */
6864 if (mddev->major_version == 0 &&
6865 mddev->minor_version > 90)
6866 rdev->recovery_offset = reshape_offset;
6868 if (rdev->recovery_offset < reshape_offset) {
6869 /* We need to check old and new layout */
6870 if (!only_parity(rdev->raid_disk,
6873 conf->max_degraded))
6876 if (!only_parity(rdev->raid_disk,
6878 conf->previous_raid_disks,
6879 conf->max_degraded))
6881 dirty_parity_disks++;
6885 * 0 for a fully functional array, 1 or 2 for a degraded array.
6887 mddev->degraded = calc_degraded(conf);
6889 if (has_failed(conf)) {
6890 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6891 " (%d/%d failed)\n",
6892 mdname(mddev), mddev->degraded, conf->raid_disks);
6896 /* device size must be a multiple of chunk size */
6897 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6898 mddev->resync_max_sectors = mddev->dev_sectors;
6900 if (mddev->degraded > dirty_parity_disks &&
6901 mddev->recovery_cp != MaxSector) {
6902 if (mddev->ok_start_degraded)
6904 "md/raid:%s: starting dirty degraded array"
6905 " - data corruption possible.\n",
6909 "md/raid:%s: cannot start dirty degraded array.\n",
6915 if (mddev->degraded == 0)
6916 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6917 " devices, algorithm %d\n", mdname(mddev), conf->level,
6918 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6921 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6922 " out of %d devices, algorithm %d\n",
6923 mdname(mddev), conf->level,
6924 mddev->raid_disks - mddev->degraded,
6925 mddev->raid_disks, mddev->new_layout);
6927 print_raid5_conf(conf);
6929 if (conf->reshape_progress != MaxSector) {
6930 conf->reshape_safe = conf->reshape_progress;
6931 atomic_set(&conf->reshape_stripes, 0);
6932 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6933 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6934 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6935 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6936 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6940 /* Ok, everything is just fine now */
6941 if (mddev->to_remove == &raid5_attrs_group)
6942 mddev->to_remove = NULL;
6943 else if (mddev->kobj.sd &&
6944 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6946 "raid5: failed to create sysfs attributes for %s\n",
6948 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6952 bool discard_supported = true;
6953 /* read-ahead size must cover two whole stripes, which
6954 * is 2 * (datadisks) * chunksize where 'n' is the
6955 * number of raid devices
6957 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6958 int stripe = data_disks *
6959 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6960 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6961 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6963 chunk_size = mddev->chunk_sectors << 9;
6964 blk_queue_io_min(mddev->queue, chunk_size);
6965 blk_queue_io_opt(mddev->queue, chunk_size *
6966 (conf->raid_disks - conf->max_degraded));
6967 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6969 * We can only discard a whole stripe. It doesn't make sense to
6970 * discard data disk but write parity disk
6972 stripe = stripe * PAGE_SIZE;
6973 /* Round up to power of 2, as discard handling
6974 * currently assumes that */
6975 while ((stripe-1) & stripe)
6976 stripe = (stripe | (stripe-1)) + 1;
6977 mddev->queue->limits.discard_alignment = stripe;
6978 mddev->queue->limits.discard_granularity = stripe;
6980 * unaligned part of discard request will be ignored, so can't
6981 * guarantee discard_zeroes_data
6983 mddev->queue->limits.discard_zeroes_data = 0;
6985 blk_queue_max_write_same_sectors(mddev->queue, 0);
6987 rdev_for_each(rdev, mddev) {
6988 disk_stack_limits(mddev->gendisk, rdev->bdev,
6989 rdev->data_offset << 9);
6990 disk_stack_limits(mddev->gendisk, rdev->bdev,
6991 rdev->new_data_offset << 9);
6993 * discard_zeroes_data is required, otherwise data
6994 * could be lost. Consider a scenario: discard a stripe
6995 * (the stripe could be inconsistent if
6996 * discard_zeroes_data is 0); write one disk of the
6997 * stripe (the stripe could be inconsistent again
6998 * depending on which disks are used to calculate
6999 * parity); the disk is broken; The stripe data of this
7002 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7003 !bdev_get_queue(rdev->bdev)->
7004 limits.discard_zeroes_data)
7005 discard_supported = false;
7006 /* Unfortunately, discard_zeroes_data is not currently
7007 * a guarantee - just a hint. So we only allow DISCARD
7008 * if the sysadmin has confirmed that only safe devices
7009 * are in use by setting a module parameter.
7011 if (!devices_handle_discard_safely) {
7012 if (discard_supported) {
7013 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7014 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7016 discard_supported = false;
7020 if (discard_supported &&
7021 mddev->queue->limits.max_discard_sectors >= stripe &&
7022 mddev->queue->limits.discard_granularity >= stripe)
7023 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7026 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7031 char b[BDEVNAME_SIZE];
7033 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7034 mdname(mddev), bdevname(journal_dev->bdev, b));
7035 r5l_init_log(conf, journal_dev);
7040 md_unregister_thread(&mddev->thread);
7041 print_raid5_conf(conf);
7043 mddev->private = NULL;
7044 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7048 static void raid5_free(struct mddev *mddev, void *priv)
7050 struct r5conf *conf = priv;
7053 mddev->to_remove = &raid5_attrs_group;
7056 static void status(struct seq_file *seq, struct mddev *mddev)
7058 struct r5conf *conf = mddev->private;
7061 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7062 conf->chunk_sectors / 2, mddev->layout);
7063 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7064 for (i = 0; i < conf->raid_disks; i++)
7065 seq_printf (seq, "%s",
7066 conf->disks[i].rdev &&
7067 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
7068 seq_printf (seq, "]");
7071 static void print_raid5_conf (struct r5conf *conf)
7074 struct disk_info *tmp;
7076 printk(KERN_DEBUG "RAID conf printout:\n");
7078 printk("(conf==NULL)\n");
7081 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7083 conf->raid_disks - conf->mddev->degraded);
7085 for (i = 0; i < conf->raid_disks; i++) {
7086 char b[BDEVNAME_SIZE];
7087 tmp = conf->disks + i;
7089 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7090 i, !test_bit(Faulty, &tmp->rdev->flags),
7091 bdevname(tmp->rdev->bdev, b));
7095 static int raid5_spare_active(struct mddev *mddev)
7098 struct r5conf *conf = mddev->private;
7099 struct disk_info *tmp;
7101 unsigned long flags;
7103 for (i = 0; i < conf->raid_disks; i++) {
7104 tmp = conf->disks + i;
7105 if (tmp->replacement
7106 && tmp->replacement->recovery_offset == MaxSector
7107 && !test_bit(Faulty, &tmp->replacement->flags)
7108 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7109 /* Replacement has just become active. */
7111 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7114 /* Replaced device not technically faulty,
7115 * but we need to be sure it gets removed
7116 * and never re-added.
7118 set_bit(Faulty, &tmp->rdev->flags);
7119 sysfs_notify_dirent_safe(
7120 tmp->rdev->sysfs_state);
7122 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7123 } else if (tmp->rdev
7124 && tmp->rdev->recovery_offset == MaxSector
7125 && !test_bit(Faulty, &tmp->rdev->flags)
7126 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7128 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7131 spin_lock_irqsave(&conf->device_lock, flags);
7132 mddev->degraded = calc_degraded(conf);
7133 spin_unlock_irqrestore(&conf->device_lock, flags);
7134 print_raid5_conf(conf);
7138 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7140 struct r5conf *conf = mddev->private;
7142 int number = rdev->raid_disk;
7143 struct md_rdev **rdevp;
7144 struct disk_info *p = conf->disks + number;
7146 print_raid5_conf(conf);
7147 if (test_bit(Journal, &rdev->flags)) {
7149 * journal disk is not removable, but we need give a chance to
7150 * update superblock of other disks. Otherwise journal disk
7151 * will be considered as 'fresh'
7153 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7156 if (rdev == p->rdev)
7158 else if (rdev == p->replacement)
7159 rdevp = &p->replacement;
7163 if (number >= conf->raid_disks &&
7164 conf->reshape_progress == MaxSector)
7165 clear_bit(In_sync, &rdev->flags);
7167 if (test_bit(In_sync, &rdev->flags) ||
7168 atomic_read(&rdev->nr_pending)) {
7172 /* Only remove non-faulty devices if recovery
7175 if (!test_bit(Faulty, &rdev->flags) &&
7176 mddev->recovery_disabled != conf->recovery_disabled &&
7177 !has_failed(conf) &&
7178 (!p->replacement || p->replacement == rdev) &&
7179 number < conf->raid_disks) {
7185 if (atomic_read(&rdev->nr_pending)) {
7186 /* lost the race, try later */
7189 } else if (p->replacement) {
7190 /* We must have just cleared 'rdev' */
7191 p->rdev = p->replacement;
7192 clear_bit(Replacement, &p->replacement->flags);
7193 smp_mb(); /* Make sure other CPUs may see both as identical
7194 * but will never see neither - if they are careful
7196 p->replacement = NULL;
7197 clear_bit(WantReplacement, &rdev->flags);
7199 /* We might have just removed the Replacement as faulty-
7200 * clear the bit just in case
7202 clear_bit(WantReplacement, &rdev->flags);
7205 print_raid5_conf(conf);
7209 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7211 struct r5conf *conf = mddev->private;
7214 struct disk_info *p;
7216 int last = conf->raid_disks - 1;
7218 if (test_bit(Journal, &rdev->flags))
7220 if (mddev->recovery_disabled == conf->recovery_disabled)
7223 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7224 /* no point adding a device */
7227 if (rdev->raid_disk >= 0)
7228 first = last = rdev->raid_disk;
7231 * find the disk ... but prefer rdev->saved_raid_disk
7234 if (rdev->saved_raid_disk >= 0 &&
7235 rdev->saved_raid_disk >= first &&
7236 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7237 first = rdev->saved_raid_disk;
7239 for (disk = first; disk <= last; disk++) {
7240 p = conf->disks + disk;
7241 if (p->rdev == NULL) {
7242 clear_bit(In_sync, &rdev->flags);
7243 rdev->raid_disk = disk;
7245 if (rdev->saved_raid_disk != disk)
7247 rcu_assign_pointer(p->rdev, rdev);
7251 for (disk = first; disk <= last; disk++) {
7252 p = conf->disks + disk;
7253 if (test_bit(WantReplacement, &p->rdev->flags) &&
7254 p->replacement == NULL) {
7255 clear_bit(In_sync, &rdev->flags);
7256 set_bit(Replacement, &rdev->flags);
7257 rdev->raid_disk = disk;
7260 rcu_assign_pointer(p->replacement, rdev);
7265 print_raid5_conf(conf);
7269 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7271 /* no resync is happening, and there is enough space
7272 * on all devices, so we can resize.
7273 * We need to make sure resync covers any new space.
7274 * If the array is shrinking we should possibly wait until
7275 * any io in the removed space completes, but it hardly seems
7279 struct r5conf *conf = mddev->private;
7283 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7284 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7285 if (mddev->external_size &&
7286 mddev->array_sectors > newsize)
7288 if (mddev->bitmap) {
7289 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7293 md_set_array_sectors(mddev, newsize);
7294 set_capacity(mddev->gendisk, mddev->array_sectors);
7295 revalidate_disk(mddev->gendisk);
7296 if (sectors > mddev->dev_sectors &&
7297 mddev->recovery_cp > mddev->dev_sectors) {
7298 mddev->recovery_cp = mddev->dev_sectors;
7299 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7301 mddev->dev_sectors = sectors;
7302 mddev->resync_max_sectors = sectors;
7306 static int check_stripe_cache(struct mddev *mddev)
7308 /* Can only proceed if there are plenty of stripe_heads.
7309 * We need a minimum of one full stripe,, and for sensible progress
7310 * it is best to have about 4 times that.
7311 * If we require 4 times, then the default 256 4K stripe_heads will
7312 * allow for chunk sizes up to 256K, which is probably OK.
7313 * If the chunk size is greater, user-space should request more
7314 * stripe_heads first.
7316 struct r5conf *conf = mddev->private;
7317 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7318 > conf->min_nr_stripes ||
7319 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7320 > conf->min_nr_stripes) {
7321 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7323 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7330 static int check_reshape(struct mddev *mddev)
7332 struct r5conf *conf = mddev->private;
7336 if (mddev->delta_disks == 0 &&
7337 mddev->new_layout == mddev->layout &&
7338 mddev->new_chunk_sectors == mddev->chunk_sectors)
7339 return 0; /* nothing to do */
7340 if (has_failed(conf))
7342 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7343 /* We might be able to shrink, but the devices must
7344 * be made bigger first.
7345 * For raid6, 4 is the minimum size.
7346 * Otherwise 2 is the minimum
7349 if (mddev->level == 6)
7351 if (mddev->raid_disks + mddev->delta_disks < min)
7355 if (!check_stripe_cache(mddev))
7358 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7359 mddev->delta_disks > 0)
7360 if (resize_chunks(conf,
7361 conf->previous_raid_disks
7362 + max(0, mddev->delta_disks),
7363 max(mddev->new_chunk_sectors,
7364 mddev->chunk_sectors)
7367 return resize_stripes(conf, (conf->previous_raid_disks
7368 + mddev->delta_disks));
7371 static int raid5_start_reshape(struct mddev *mddev)
7373 struct r5conf *conf = mddev->private;
7374 struct md_rdev *rdev;
7376 unsigned long flags;
7378 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7381 if (!check_stripe_cache(mddev))
7384 if (has_failed(conf))
7387 rdev_for_each(rdev, mddev) {
7388 if (!test_bit(In_sync, &rdev->flags)
7389 && !test_bit(Faulty, &rdev->flags))
7393 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7394 /* Not enough devices even to make a degraded array
7399 /* Refuse to reduce size of the array. Any reductions in
7400 * array size must be through explicit setting of array_size
7403 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7404 < mddev->array_sectors) {
7405 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7406 "before number of disks\n", mdname(mddev));
7410 atomic_set(&conf->reshape_stripes, 0);
7411 spin_lock_irq(&conf->device_lock);
7412 write_seqcount_begin(&conf->gen_lock);
7413 conf->previous_raid_disks = conf->raid_disks;
7414 conf->raid_disks += mddev->delta_disks;
7415 conf->prev_chunk_sectors = conf->chunk_sectors;
7416 conf->chunk_sectors = mddev->new_chunk_sectors;
7417 conf->prev_algo = conf->algorithm;
7418 conf->algorithm = mddev->new_layout;
7420 /* Code that selects data_offset needs to see the generation update
7421 * if reshape_progress has been set - so a memory barrier needed.
7424 if (mddev->reshape_backwards)
7425 conf->reshape_progress = raid5_size(mddev, 0, 0);
7427 conf->reshape_progress = 0;
7428 conf->reshape_safe = conf->reshape_progress;
7429 write_seqcount_end(&conf->gen_lock);
7430 spin_unlock_irq(&conf->device_lock);
7432 /* Now make sure any requests that proceeded on the assumption
7433 * the reshape wasn't running - like Discard or Read - have
7436 mddev_suspend(mddev);
7437 mddev_resume(mddev);
7439 /* Add some new drives, as many as will fit.
7440 * We know there are enough to make the newly sized array work.
7441 * Don't add devices if we are reducing the number of
7442 * devices in the array. This is because it is not possible
7443 * to correctly record the "partially reconstructed" state of
7444 * such devices during the reshape and confusion could result.
7446 if (mddev->delta_disks >= 0) {
7447 rdev_for_each(rdev, mddev)
7448 if (rdev->raid_disk < 0 &&
7449 !test_bit(Faulty, &rdev->flags)) {
7450 if (raid5_add_disk(mddev, rdev) == 0) {
7452 >= conf->previous_raid_disks)
7453 set_bit(In_sync, &rdev->flags);
7455 rdev->recovery_offset = 0;
7457 if (sysfs_link_rdev(mddev, rdev))
7458 /* Failure here is OK */;
7460 } else if (rdev->raid_disk >= conf->previous_raid_disks
7461 && !test_bit(Faulty, &rdev->flags)) {
7462 /* This is a spare that was manually added */
7463 set_bit(In_sync, &rdev->flags);
7466 /* When a reshape changes the number of devices,
7467 * ->degraded is measured against the larger of the
7468 * pre and post number of devices.
7470 spin_lock_irqsave(&conf->device_lock, flags);
7471 mddev->degraded = calc_degraded(conf);
7472 spin_unlock_irqrestore(&conf->device_lock, flags);
7474 mddev->raid_disks = conf->raid_disks;
7475 mddev->reshape_position = conf->reshape_progress;
7476 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7478 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7479 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7480 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7481 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7482 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7483 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7485 if (!mddev->sync_thread) {
7486 mddev->recovery = 0;
7487 spin_lock_irq(&conf->device_lock);
7488 write_seqcount_begin(&conf->gen_lock);
7489 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7490 mddev->new_chunk_sectors =
7491 conf->chunk_sectors = conf->prev_chunk_sectors;
7492 mddev->new_layout = conf->algorithm = conf->prev_algo;
7493 rdev_for_each(rdev, mddev)
7494 rdev->new_data_offset = rdev->data_offset;
7496 conf->generation --;
7497 conf->reshape_progress = MaxSector;
7498 mddev->reshape_position = MaxSector;
7499 write_seqcount_end(&conf->gen_lock);
7500 spin_unlock_irq(&conf->device_lock);
7503 conf->reshape_checkpoint = jiffies;
7504 md_wakeup_thread(mddev->sync_thread);
7505 md_new_event(mddev);
7509 /* This is called from the reshape thread and should make any
7510 * changes needed in 'conf'
7512 static void end_reshape(struct r5conf *conf)
7515 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7516 struct md_rdev *rdev;
7518 spin_lock_irq(&conf->device_lock);
7519 conf->previous_raid_disks = conf->raid_disks;
7520 rdev_for_each(rdev, conf->mddev)
7521 rdev->data_offset = rdev->new_data_offset;
7523 conf->reshape_progress = MaxSector;
7524 conf->mddev->reshape_position = MaxSector;
7525 spin_unlock_irq(&conf->device_lock);
7526 wake_up(&conf->wait_for_overlap);
7528 /* read-ahead size must cover two whole stripes, which is
7529 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7531 if (conf->mddev->queue) {
7532 int data_disks = conf->raid_disks - conf->max_degraded;
7533 int stripe = data_disks * ((conf->chunk_sectors << 9)
7535 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7536 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7541 /* This is called from the raid5d thread with mddev_lock held.
7542 * It makes config changes to the device.
7544 static void raid5_finish_reshape(struct mddev *mddev)
7546 struct r5conf *conf = mddev->private;
7548 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7550 if (mddev->delta_disks > 0) {
7551 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7552 set_capacity(mddev->gendisk, mddev->array_sectors);
7553 revalidate_disk(mddev->gendisk);
7556 spin_lock_irq(&conf->device_lock);
7557 mddev->degraded = calc_degraded(conf);
7558 spin_unlock_irq(&conf->device_lock);
7559 for (d = conf->raid_disks ;
7560 d < conf->raid_disks - mddev->delta_disks;
7562 struct md_rdev *rdev = conf->disks[d].rdev;
7564 clear_bit(In_sync, &rdev->flags);
7565 rdev = conf->disks[d].replacement;
7567 clear_bit(In_sync, &rdev->flags);
7570 mddev->layout = conf->algorithm;
7571 mddev->chunk_sectors = conf->chunk_sectors;
7572 mddev->reshape_position = MaxSector;
7573 mddev->delta_disks = 0;
7574 mddev->reshape_backwards = 0;
7578 static void raid5_quiesce(struct mddev *mddev, int state)
7580 struct r5conf *conf = mddev->private;
7583 case 2: /* resume for a suspend */
7584 wake_up(&conf->wait_for_overlap);
7587 case 1: /* stop all writes */
7588 lock_all_device_hash_locks_irq(conf);
7589 /* '2' tells resync/reshape to pause so that all
7590 * active stripes can drain
7593 wait_event_cmd(conf->wait_for_quiescent,
7594 atomic_read(&conf->active_stripes) == 0 &&
7595 atomic_read(&conf->active_aligned_reads) == 0,
7596 unlock_all_device_hash_locks_irq(conf),
7597 lock_all_device_hash_locks_irq(conf));
7599 unlock_all_device_hash_locks_irq(conf);
7600 /* allow reshape to continue */
7601 wake_up(&conf->wait_for_overlap);
7604 case 0: /* re-enable writes */
7605 lock_all_device_hash_locks_irq(conf);
7607 wake_up(&conf->wait_for_quiescent);
7608 wake_up(&conf->wait_for_overlap);
7609 unlock_all_device_hash_locks_irq(conf);
7612 r5l_quiesce(conf->log, state);
7615 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7617 struct r0conf *raid0_conf = mddev->private;
7620 /* for raid0 takeover only one zone is supported */
7621 if (raid0_conf->nr_strip_zones > 1) {
7622 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7624 return ERR_PTR(-EINVAL);
7627 sectors = raid0_conf->strip_zone[0].zone_end;
7628 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7629 mddev->dev_sectors = sectors;
7630 mddev->new_level = level;
7631 mddev->new_layout = ALGORITHM_PARITY_N;
7632 mddev->new_chunk_sectors = mddev->chunk_sectors;
7633 mddev->raid_disks += 1;
7634 mddev->delta_disks = 1;
7635 /* make sure it will be not marked as dirty */
7636 mddev->recovery_cp = MaxSector;
7638 return setup_conf(mddev);
7641 static void *raid5_takeover_raid1(struct mddev *mddev)
7645 if (mddev->raid_disks != 2 ||
7646 mddev->degraded > 1)
7647 return ERR_PTR(-EINVAL);
7649 /* Should check if there are write-behind devices? */
7651 chunksect = 64*2; /* 64K by default */
7653 /* The array must be an exact multiple of chunksize */
7654 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7657 if ((chunksect<<9) < STRIPE_SIZE)
7658 /* array size does not allow a suitable chunk size */
7659 return ERR_PTR(-EINVAL);
7661 mddev->new_level = 5;
7662 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7663 mddev->new_chunk_sectors = chunksect;
7665 return setup_conf(mddev);
7668 static void *raid5_takeover_raid6(struct mddev *mddev)
7672 switch (mddev->layout) {
7673 case ALGORITHM_LEFT_ASYMMETRIC_6:
7674 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7676 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7677 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7679 case ALGORITHM_LEFT_SYMMETRIC_6:
7680 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7682 case ALGORITHM_RIGHT_SYMMETRIC_6:
7683 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7685 case ALGORITHM_PARITY_0_6:
7686 new_layout = ALGORITHM_PARITY_0;
7688 case ALGORITHM_PARITY_N:
7689 new_layout = ALGORITHM_PARITY_N;
7692 return ERR_PTR(-EINVAL);
7694 mddev->new_level = 5;
7695 mddev->new_layout = new_layout;
7696 mddev->delta_disks = -1;
7697 mddev->raid_disks -= 1;
7698 return setup_conf(mddev);
7701 static int raid5_check_reshape(struct mddev *mddev)
7703 /* For a 2-drive array, the layout and chunk size can be changed
7704 * immediately as not restriping is needed.
7705 * For larger arrays we record the new value - after validation
7706 * to be used by a reshape pass.
7708 struct r5conf *conf = mddev->private;
7709 int new_chunk = mddev->new_chunk_sectors;
7711 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7713 if (new_chunk > 0) {
7714 if (!is_power_of_2(new_chunk))
7716 if (new_chunk < (PAGE_SIZE>>9))
7718 if (mddev->array_sectors & (new_chunk-1))
7719 /* not factor of array size */
7723 /* They look valid */
7725 if (mddev->raid_disks == 2) {
7726 /* can make the change immediately */
7727 if (mddev->new_layout >= 0) {
7728 conf->algorithm = mddev->new_layout;
7729 mddev->layout = mddev->new_layout;
7731 if (new_chunk > 0) {
7732 conf->chunk_sectors = new_chunk ;
7733 mddev->chunk_sectors = new_chunk;
7735 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7736 md_wakeup_thread(mddev->thread);
7738 return check_reshape(mddev);
7741 static int raid6_check_reshape(struct mddev *mddev)
7743 int new_chunk = mddev->new_chunk_sectors;
7745 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7747 if (new_chunk > 0) {
7748 if (!is_power_of_2(new_chunk))
7750 if (new_chunk < (PAGE_SIZE >> 9))
7752 if (mddev->array_sectors & (new_chunk-1))
7753 /* not factor of array size */
7757 /* They look valid */
7758 return check_reshape(mddev);
7761 static void *raid5_takeover(struct mddev *mddev)
7763 /* raid5 can take over:
7764 * raid0 - if there is only one strip zone - make it a raid4 layout
7765 * raid1 - if there are two drives. We need to know the chunk size
7766 * raid4 - trivial - just use a raid4 layout.
7767 * raid6 - Providing it is a *_6 layout
7769 if (mddev->level == 0)
7770 return raid45_takeover_raid0(mddev, 5);
7771 if (mddev->level == 1)
7772 return raid5_takeover_raid1(mddev);
7773 if (mddev->level == 4) {
7774 mddev->new_layout = ALGORITHM_PARITY_N;
7775 mddev->new_level = 5;
7776 return setup_conf(mddev);
7778 if (mddev->level == 6)
7779 return raid5_takeover_raid6(mddev);
7781 return ERR_PTR(-EINVAL);
7784 static void *raid4_takeover(struct mddev *mddev)
7786 /* raid4 can take over:
7787 * raid0 - if there is only one strip zone
7788 * raid5 - if layout is right
7790 if (mddev->level == 0)
7791 return raid45_takeover_raid0(mddev, 4);
7792 if (mddev->level == 5 &&
7793 mddev->layout == ALGORITHM_PARITY_N) {
7794 mddev->new_layout = 0;
7795 mddev->new_level = 4;
7796 return setup_conf(mddev);
7798 return ERR_PTR(-EINVAL);
7801 static struct md_personality raid5_personality;
7803 static void *raid6_takeover(struct mddev *mddev)
7805 /* Currently can only take over a raid5. We map the
7806 * personality to an equivalent raid6 personality
7807 * with the Q block at the end.
7811 if (mddev->pers != &raid5_personality)
7812 return ERR_PTR(-EINVAL);
7813 if (mddev->degraded > 1)
7814 return ERR_PTR(-EINVAL);
7815 if (mddev->raid_disks > 253)
7816 return ERR_PTR(-EINVAL);
7817 if (mddev->raid_disks < 3)
7818 return ERR_PTR(-EINVAL);
7820 switch (mddev->layout) {
7821 case ALGORITHM_LEFT_ASYMMETRIC:
7822 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7824 case ALGORITHM_RIGHT_ASYMMETRIC:
7825 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7827 case ALGORITHM_LEFT_SYMMETRIC:
7828 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7830 case ALGORITHM_RIGHT_SYMMETRIC:
7831 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7833 case ALGORITHM_PARITY_0:
7834 new_layout = ALGORITHM_PARITY_0_6;
7836 case ALGORITHM_PARITY_N:
7837 new_layout = ALGORITHM_PARITY_N;
7840 return ERR_PTR(-EINVAL);
7842 mddev->new_level = 6;
7843 mddev->new_layout = new_layout;
7844 mddev->delta_disks = 1;
7845 mddev->raid_disks += 1;
7846 return setup_conf(mddev);
7849 static struct md_personality raid6_personality =
7853 .owner = THIS_MODULE,
7854 .make_request = make_request,
7858 .error_handler = error,
7859 .hot_add_disk = raid5_add_disk,
7860 .hot_remove_disk= raid5_remove_disk,
7861 .spare_active = raid5_spare_active,
7862 .sync_request = sync_request,
7863 .resize = raid5_resize,
7865 .check_reshape = raid6_check_reshape,
7866 .start_reshape = raid5_start_reshape,
7867 .finish_reshape = raid5_finish_reshape,
7868 .quiesce = raid5_quiesce,
7869 .takeover = raid6_takeover,
7870 .congested = raid5_congested,
7872 static struct md_personality raid5_personality =
7876 .owner = THIS_MODULE,
7877 .make_request = make_request,
7881 .error_handler = error,
7882 .hot_add_disk = raid5_add_disk,
7883 .hot_remove_disk= raid5_remove_disk,
7884 .spare_active = raid5_spare_active,
7885 .sync_request = sync_request,
7886 .resize = raid5_resize,
7888 .check_reshape = raid5_check_reshape,
7889 .start_reshape = raid5_start_reshape,
7890 .finish_reshape = raid5_finish_reshape,
7891 .quiesce = raid5_quiesce,
7892 .takeover = raid5_takeover,
7893 .congested = raid5_congested,
7896 static struct md_personality raid4_personality =
7900 .owner = THIS_MODULE,
7901 .make_request = make_request,
7905 .error_handler = error,
7906 .hot_add_disk = raid5_add_disk,
7907 .hot_remove_disk= raid5_remove_disk,
7908 .spare_active = raid5_spare_active,
7909 .sync_request = sync_request,
7910 .resize = raid5_resize,
7912 .check_reshape = raid5_check_reshape,
7913 .start_reshape = raid5_start_reshape,
7914 .finish_reshape = raid5_finish_reshape,
7915 .quiesce = raid5_quiesce,
7916 .takeover = raid4_takeover,
7917 .congested = raid5_congested,
7920 static int __init raid5_init(void)
7922 raid5_wq = alloc_workqueue("raid5wq",
7923 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7926 register_md_personality(&raid6_personality);
7927 register_md_personality(&raid5_personality);
7928 register_md_personality(&raid4_personality);
7932 static void raid5_exit(void)
7934 unregister_md_personality(&raid6_personality);
7935 unregister_md_personality(&raid5_personality);
7936 unregister_md_personality(&raid4_personality);
7937 destroy_workqueue(raid5_wq);
7940 module_init(raid5_init);
7941 module_exit(raid5_exit);
7942 MODULE_LICENSE("GPL");
7943 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7944 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7945 MODULE_ALIAS("md-raid5");
7946 MODULE_ALIAS("md-raid4");
7947 MODULE_ALIAS("md-level-5");
7948 MODULE_ALIAS("md-level-4");
7949 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7950 MODULE_ALIAS("md-raid6");
7951 MODULE_ALIAS("md-level-6");
7953 /* This used to be two separate modules, they were: */
7954 MODULE_ALIAS("raid5");
7955 MODULE_ALIAS("raid6");