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;
1933 percpu = per_cpu_ptr(conf->percpu, cpu);
1934 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1935 ops_run_biofill(sh);
1939 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1941 tx = ops_run_compute5(sh, percpu);
1943 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1944 tx = ops_run_compute6_1(sh, percpu);
1946 tx = ops_run_compute6_2(sh, percpu);
1948 /* terminate the chain if reconstruct is not set to be run */
1949 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1953 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1955 tx = ops_run_prexor5(sh, percpu, tx);
1957 tx = ops_run_prexor6(sh, percpu, tx);
1960 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1961 tx = ops_run_biodrain(sh, tx);
1965 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1967 ops_run_reconstruct5(sh, percpu, tx);
1969 ops_run_reconstruct6(sh, percpu, tx);
1972 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1973 if (sh->check_state == check_state_run)
1974 ops_run_check_p(sh, percpu);
1975 else if (sh->check_state == check_state_run_q)
1976 ops_run_check_pq(sh, percpu, 0);
1977 else if (sh->check_state == check_state_run_pq)
1978 ops_run_check_pq(sh, percpu, 1);
1983 if (overlap_clear && !sh->batch_head)
1984 for (i = disks; i--; ) {
1985 struct r5dev *dev = &sh->dev[i];
1986 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1987 wake_up(&sh->raid_conf->wait_for_overlap);
1992 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1994 struct stripe_head *sh;
1996 sh = kmem_cache_zalloc(sc, gfp);
1998 spin_lock_init(&sh->stripe_lock);
1999 spin_lock_init(&sh->batch_lock);
2000 INIT_LIST_HEAD(&sh->batch_list);
2001 INIT_LIST_HEAD(&sh->lru);
2002 atomic_set(&sh->count, 1);
2006 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2008 struct stripe_head *sh;
2010 sh = alloc_stripe(conf->slab_cache, gfp);
2014 sh->raid_conf = conf;
2016 if (grow_buffers(sh, gfp)) {
2018 kmem_cache_free(conf->slab_cache, sh);
2021 sh->hash_lock_index =
2022 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2023 /* we just created an active stripe so... */
2024 atomic_inc(&conf->active_stripes);
2026 raid5_release_stripe(sh);
2027 conf->max_nr_stripes++;
2031 static int grow_stripes(struct r5conf *conf, int num)
2033 struct kmem_cache *sc;
2034 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2036 if (conf->mddev->gendisk)
2037 sprintf(conf->cache_name[0],
2038 "raid%d-%s", conf->level, mdname(conf->mddev));
2040 sprintf(conf->cache_name[0],
2041 "raid%d-%p", conf->level, conf->mddev);
2042 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2044 conf->active_name = 0;
2045 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2046 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2050 conf->slab_cache = sc;
2051 conf->pool_size = devs;
2053 if (!grow_one_stripe(conf, GFP_KERNEL))
2060 * scribble_len - return the required size of the scribble region
2061 * @num - total number of disks in the array
2063 * The size must be enough to contain:
2064 * 1/ a struct page pointer for each device in the array +2
2065 * 2/ room to convert each entry in (1) to its corresponding dma
2066 * (dma_map_page()) or page (page_address()) address.
2068 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2069 * calculate over all devices (not just the data blocks), using zeros in place
2070 * of the P and Q blocks.
2072 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2074 struct flex_array *ret;
2077 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2078 ret = flex_array_alloc(len, cnt, flags);
2081 /* always prealloc all elements, so no locking is required */
2082 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2083 flex_array_free(ret);
2089 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2095 * Never shrink. And mddev_suspend() could deadlock if this is called
2096 * from raid5d. In that case, scribble_disks and scribble_sectors
2097 * should equal to new_disks and new_sectors
2099 if (conf->scribble_disks >= new_disks &&
2100 conf->scribble_sectors >= new_sectors)
2102 mddev_suspend(conf->mddev);
2104 for_each_present_cpu(cpu) {
2105 struct raid5_percpu *percpu;
2106 struct flex_array *scribble;
2108 percpu = per_cpu_ptr(conf->percpu, cpu);
2109 scribble = scribble_alloc(new_disks,
2110 new_sectors / STRIPE_SECTORS,
2114 flex_array_free(percpu->scribble);
2115 percpu->scribble = scribble;
2122 mddev_resume(conf->mddev);
2124 conf->scribble_disks = new_disks;
2125 conf->scribble_sectors = new_sectors;
2130 static int resize_stripes(struct r5conf *conf, int newsize)
2132 /* Make all the stripes able to hold 'newsize' devices.
2133 * New slots in each stripe get 'page' set to a new page.
2135 * This happens in stages:
2136 * 1/ create a new kmem_cache and allocate the required number of
2138 * 2/ gather all the old stripe_heads and transfer the pages across
2139 * to the new stripe_heads. This will have the side effect of
2140 * freezing the array as once all stripe_heads have been collected,
2141 * no IO will be possible. Old stripe heads are freed once their
2142 * pages have been transferred over, and the old kmem_cache is
2143 * freed when all stripes are done.
2144 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2145 * we simple return a failre status - no need to clean anything up.
2146 * 4/ allocate new pages for the new slots in the new stripe_heads.
2147 * If this fails, we don't bother trying the shrink the
2148 * stripe_heads down again, we just leave them as they are.
2149 * As each stripe_head is processed the new one is released into
2152 * Once step2 is started, we cannot afford to wait for a write,
2153 * so we use GFP_NOIO allocations.
2155 struct stripe_head *osh, *nsh;
2156 LIST_HEAD(newstripes);
2157 struct disk_info *ndisks;
2159 struct kmem_cache *sc;
2163 if (newsize <= conf->pool_size)
2164 return 0; /* never bother to shrink */
2166 err = md_allow_write(conf->mddev);
2171 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2172 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2177 /* Need to ensure auto-resizing doesn't interfere */
2178 mutex_lock(&conf->cache_size_mutex);
2180 for (i = conf->max_nr_stripes; i; i--) {
2181 nsh = alloc_stripe(sc, GFP_KERNEL);
2185 nsh->raid_conf = conf;
2186 list_add(&nsh->lru, &newstripes);
2189 /* didn't get enough, give up */
2190 while (!list_empty(&newstripes)) {
2191 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2192 list_del(&nsh->lru);
2193 kmem_cache_free(sc, nsh);
2195 kmem_cache_destroy(sc);
2196 mutex_unlock(&conf->cache_size_mutex);
2199 /* Step 2 - Must use GFP_NOIO now.
2200 * OK, we have enough stripes, start collecting inactive
2201 * stripes and copying them over
2205 list_for_each_entry(nsh, &newstripes, lru) {
2206 lock_device_hash_lock(conf, hash);
2207 wait_event_exclusive_cmd(conf->wait_for_stripe[hash],
2208 !list_empty(conf->inactive_list + hash),
2209 unlock_device_hash_lock(conf, hash),
2210 lock_device_hash_lock(conf, hash));
2211 osh = get_free_stripe(conf, hash);
2212 unlock_device_hash_lock(conf, hash);
2214 for(i=0; i<conf->pool_size; i++) {
2215 nsh->dev[i].page = osh->dev[i].page;
2216 nsh->dev[i].orig_page = osh->dev[i].page;
2218 nsh->hash_lock_index = hash;
2219 kmem_cache_free(conf->slab_cache, osh);
2221 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2222 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2227 kmem_cache_destroy(conf->slab_cache);
2230 * At this point, we are holding all the stripes so the array
2231 * is completely stalled, so now is a good time to resize
2232 * conf->disks and the scribble region
2234 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2236 for (i=0; i<conf->raid_disks; i++)
2237 ndisks[i] = conf->disks[i];
2239 conf->disks = ndisks;
2243 mutex_unlock(&conf->cache_size_mutex);
2244 /* Step 4, return new stripes to service */
2245 while(!list_empty(&newstripes)) {
2246 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2247 list_del_init(&nsh->lru);
2249 for (i=conf->raid_disks; i < newsize; i++)
2250 if (nsh->dev[i].page == NULL) {
2251 struct page *p = alloc_page(GFP_NOIO);
2252 nsh->dev[i].page = p;
2253 nsh->dev[i].orig_page = p;
2257 raid5_release_stripe(nsh);
2259 /* critical section pass, GFP_NOIO no longer needed */
2261 conf->slab_cache = sc;
2262 conf->active_name = 1-conf->active_name;
2264 conf->pool_size = newsize;
2268 static int drop_one_stripe(struct r5conf *conf)
2270 struct stripe_head *sh;
2271 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2273 spin_lock_irq(conf->hash_locks + hash);
2274 sh = get_free_stripe(conf, hash);
2275 spin_unlock_irq(conf->hash_locks + hash);
2278 BUG_ON(atomic_read(&sh->count));
2280 kmem_cache_free(conf->slab_cache, sh);
2281 atomic_dec(&conf->active_stripes);
2282 conf->max_nr_stripes--;
2286 static void shrink_stripes(struct r5conf *conf)
2288 while (conf->max_nr_stripes &&
2289 drop_one_stripe(conf))
2292 kmem_cache_destroy(conf->slab_cache);
2293 conf->slab_cache = NULL;
2296 static void raid5_end_read_request(struct bio * bi)
2298 struct stripe_head *sh = bi->bi_private;
2299 struct r5conf *conf = sh->raid_conf;
2300 int disks = sh->disks, i;
2301 char b[BDEVNAME_SIZE];
2302 struct md_rdev *rdev = NULL;
2305 for (i=0 ; i<disks; i++)
2306 if (bi == &sh->dev[i].req)
2309 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2310 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2316 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2317 /* If replacement finished while this request was outstanding,
2318 * 'replacement' might be NULL already.
2319 * In that case it moved down to 'rdev'.
2320 * rdev is not removed until all requests are finished.
2322 rdev = conf->disks[i].replacement;
2324 rdev = conf->disks[i].rdev;
2326 if (use_new_offset(conf, sh))
2327 s = sh->sector + rdev->new_data_offset;
2329 s = sh->sector + rdev->data_offset;
2330 if (!bi->bi_error) {
2331 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2332 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2333 /* Note that this cannot happen on a
2334 * replacement device. We just fail those on
2339 "md/raid:%s: read error corrected"
2340 " (%lu sectors at %llu on %s)\n",
2341 mdname(conf->mddev), STRIPE_SECTORS,
2342 (unsigned long long)s,
2343 bdevname(rdev->bdev, b));
2344 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2345 clear_bit(R5_ReadError, &sh->dev[i].flags);
2346 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2347 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2348 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2350 if (atomic_read(&rdev->read_errors))
2351 atomic_set(&rdev->read_errors, 0);
2353 const char *bdn = bdevname(rdev->bdev, b);
2357 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2358 atomic_inc(&rdev->read_errors);
2359 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2362 "md/raid:%s: read error on replacement device "
2363 "(sector %llu on %s).\n",
2364 mdname(conf->mddev),
2365 (unsigned long long)s,
2367 else if (conf->mddev->degraded >= conf->max_degraded) {
2371 "md/raid:%s: read error not correctable "
2372 "(sector %llu on %s).\n",
2373 mdname(conf->mddev),
2374 (unsigned long long)s,
2376 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2381 "md/raid:%s: read error NOT corrected!! "
2382 "(sector %llu on %s).\n",
2383 mdname(conf->mddev),
2384 (unsigned long long)s,
2386 } else if (atomic_read(&rdev->read_errors)
2387 > conf->max_nr_stripes)
2389 "md/raid:%s: Too many read errors, failing device %s.\n",
2390 mdname(conf->mddev), bdn);
2393 if (set_bad && test_bit(In_sync, &rdev->flags)
2394 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2397 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2398 set_bit(R5_ReadError, &sh->dev[i].flags);
2399 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2401 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2403 clear_bit(R5_ReadError, &sh->dev[i].flags);
2404 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2406 && test_bit(In_sync, &rdev->flags)
2407 && rdev_set_badblocks(
2408 rdev, sh->sector, STRIPE_SECTORS, 0)))
2409 md_error(conf->mddev, rdev);
2412 rdev_dec_pending(rdev, conf->mddev);
2413 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2414 set_bit(STRIPE_HANDLE, &sh->state);
2415 raid5_release_stripe(sh);
2418 static void raid5_end_write_request(struct bio *bi)
2420 struct stripe_head *sh = bi->bi_private;
2421 struct r5conf *conf = sh->raid_conf;
2422 int disks = sh->disks, i;
2423 struct md_rdev *uninitialized_var(rdev);
2426 int replacement = 0;
2428 for (i = 0 ; i < disks; i++) {
2429 if (bi == &sh->dev[i].req) {
2430 rdev = conf->disks[i].rdev;
2433 if (bi == &sh->dev[i].rreq) {
2434 rdev = conf->disks[i].replacement;
2438 /* rdev was removed and 'replacement'
2439 * replaced it. rdev is not removed
2440 * until all requests are finished.
2442 rdev = conf->disks[i].rdev;
2446 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2447 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2456 md_error(conf->mddev, rdev);
2457 else if (is_badblock(rdev, sh->sector,
2459 &first_bad, &bad_sectors))
2460 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2463 set_bit(STRIPE_DEGRADED, &sh->state);
2464 set_bit(WriteErrorSeen, &rdev->flags);
2465 set_bit(R5_WriteError, &sh->dev[i].flags);
2466 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2467 set_bit(MD_RECOVERY_NEEDED,
2468 &rdev->mddev->recovery);
2469 } else if (is_badblock(rdev, sh->sector,
2471 &first_bad, &bad_sectors)) {
2472 set_bit(R5_MadeGood, &sh->dev[i].flags);
2473 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2474 /* That was a successful write so make
2475 * sure it looks like we already did
2478 set_bit(R5_ReWrite, &sh->dev[i].flags);
2481 rdev_dec_pending(rdev, conf->mddev);
2483 if (sh->batch_head && bi->bi_error && !replacement)
2484 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2486 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2487 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2488 set_bit(STRIPE_HANDLE, &sh->state);
2489 raid5_release_stripe(sh);
2491 if (sh->batch_head && sh != sh->batch_head)
2492 raid5_release_stripe(sh->batch_head);
2495 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2497 struct r5dev *dev = &sh->dev[i];
2499 bio_init(&dev->req);
2500 dev->req.bi_io_vec = &dev->vec;
2501 dev->req.bi_max_vecs = 1;
2502 dev->req.bi_private = sh;
2504 bio_init(&dev->rreq);
2505 dev->rreq.bi_io_vec = &dev->rvec;
2506 dev->rreq.bi_max_vecs = 1;
2507 dev->rreq.bi_private = sh;
2510 dev->sector = raid5_compute_blocknr(sh, i, previous);
2513 static void error(struct mddev *mddev, struct md_rdev *rdev)
2515 char b[BDEVNAME_SIZE];
2516 struct r5conf *conf = mddev->private;
2517 unsigned long flags;
2518 pr_debug("raid456: error called\n");
2520 spin_lock_irqsave(&conf->device_lock, flags);
2521 clear_bit(In_sync, &rdev->flags);
2522 mddev->degraded = calc_degraded(conf);
2523 spin_unlock_irqrestore(&conf->device_lock, flags);
2524 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2526 set_bit(Blocked, &rdev->flags);
2527 set_bit(Faulty, &rdev->flags);
2528 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2529 set_bit(MD_CHANGE_PENDING, &mddev->flags);
2531 "md/raid:%s: Disk failure on %s, disabling device.\n"
2532 "md/raid:%s: Operation continuing on %d devices.\n",
2534 bdevname(rdev->bdev, b),
2536 conf->raid_disks - mddev->degraded);
2540 * Input: a 'big' sector number,
2541 * Output: index of the data and parity disk, and the sector # in them.
2543 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2544 int previous, int *dd_idx,
2545 struct stripe_head *sh)
2547 sector_t stripe, stripe2;
2548 sector_t chunk_number;
2549 unsigned int chunk_offset;
2552 sector_t new_sector;
2553 int algorithm = previous ? conf->prev_algo
2555 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2556 : conf->chunk_sectors;
2557 int raid_disks = previous ? conf->previous_raid_disks
2559 int data_disks = raid_disks - conf->max_degraded;
2561 /* First compute the information on this sector */
2564 * Compute the chunk number and the sector offset inside the chunk
2566 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2567 chunk_number = r_sector;
2570 * Compute the stripe number
2572 stripe = chunk_number;
2573 *dd_idx = sector_div(stripe, data_disks);
2576 * Select the parity disk based on the user selected algorithm.
2578 pd_idx = qd_idx = -1;
2579 switch(conf->level) {
2581 pd_idx = data_disks;
2584 switch (algorithm) {
2585 case ALGORITHM_LEFT_ASYMMETRIC:
2586 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2587 if (*dd_idx >= pd_idx)
2590 case ALGORITHM_RIGHT_ASYMMETRIC:
2591 pd_idx = sector_div(stripe2, raid_disks);
2592 if (*dd_idx >= pd_idx)
2595 case ALGORITHM_LEFT_SYMMETRIC:
2596 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2597 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2599 case ALGORITHM_RIGHT_SYMMETRIC:
2600 pd_idx = sector_div(stripe2, raid_disks);
2601 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2603 case ALGORITHM_PARITY_0:
2607 case ALGORITHM_PARITY_N:
2608 pd_idx = data_disks;
2616 switch (algorithm) {
2617 case ALGORITHM_LEFT_ASYMMETRIC:
2618 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2619 qd_idx = pd_idx + 1;
2620 if (pd_idx == raid_disks-1) {
2621 (*dd_idx)++; /* Q D D D P */
2623 } else if (*dd_idx >= pd_idx)
2624 (*dd_idx) += 2; /* D D P Q D */
2626 case ALGORITHM_RIGHT_ASYMMETRIC:
2627 pd_idx = sector_div(stripe2, raid_disks);
2628 qd_idx = pd_idx + 1;
2629 if (pd_idx == raid_disks-1) {
2630 (*dd_idx)++; /* Q D D D P */
2632 } else if (*dd_idx >= pd_idx)
2633 (*dd_idx) += 2; /* D D P Q D */
2635 case ALGORITHM_LEFT_SYMMETRIC:
2636 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2637 qd_idx = (pd_idx + 1) % raid_disks;
2638 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2640 case ALGORITHM_RIGHT_SYMMETRIC:
2641 pd_idx = sector_div(stripe2, raid_disks);
2642 qd_idx = (pd_idx + 1) % raid_disks;
2643 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2646 case ALGORITHM_PARITY_0:
2651 case ALGORITHM_PARITY_N:
2652 pd_idx = data_disks;
2653 qd_idx = data_disks + 1;
2656 case ALGORITHM_ROTATING_ZERO_RESTART:
2657 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2658 * of blocks for computing Q is different.
2660 pd_idx = sector_div(stripe2, raid_disks);
2661 qd_idx = pd_idx + 1;
2662 if (pd_idx == raid_disks-1) {
2663 (*dd_idx)++; /* Q D D D P */
2665 } else if (*dd_idx >= pd_idx)
2666 (*dd_idx) += 2; /* D D P Q D */
2670 case ALGORITHM_ROTATING_N_RESTART:
2671 /* Same a left_asymmetric, by first stripe is
2672 * D D D P Q rather than
2676 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2677 qd_idx = pd_idx + 1;
2678 if (pd_idx == raid_disks-1) {
2679 (*dd_idx)++; /* Q D D D P */
2681 } else if (*dd_idx >= pd_idx)
2682 (*dd_idx) += 2; /* D D P Q D */
2686 case ALGORITHM_ROTATING_N_CONTINUE:
2687 /* Same as left_symmetric but Q is before P */
2688 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2689 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2690 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2694 case ALGORITHM_LEFT_ASYMMETRIC_6:
2695 /* RAID5 left_asymmetric, with Q on last device */
2696 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2697 if (*dd_idx >= pd_idx)
2699 qd_idx = raid_disks - 1;
2702 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2703 pd_idx = sector_div(stripe2, raid_disks-1);
2704 if (*dd_idx >= pd_idx)
2706 qd_idx = raid_disks - 1;
2709 case ALGORITHM_LEFT_SYMMETRIC_6:
2710 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2711 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2712 qd_idx = raid_disks - 1;
2715 case ALGORITHM_RIGHT_SYMMETRIC_6:
2716 pd_idx = sector_div(stripe2, raid_disks-1);
2717 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2718 qd_idx = raid_disks - 1;
2721 case ALGORITHM_PARITY_0_6:
2724 qd_idx = raid_disks - 1;
2734 sh->pd_idx = pd_idx;
2735 sh->qd_idx = qd_idx;
2736 sh->ddf_layout = ddf_layout;
2739 * Finally, compute the new sector number
2741 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2745 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2747 struct r5conf *conf = sh->raid_conf;
2748 int raid_disks = sh->disks;
2749 int data_disks = raid_disks - conf->max_degraded;
2750 sector_t new_sector = sh->sector, check;
2751 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2752 : conf->chunk_sectors;
2753 int algorithm = previous ? conf->prev_algo
2757 sector_t chunk_number;
2758 int dummy1, dd_idx = i;
2760 struct stripe_head sh2;
2762 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2763 stripe = new_sector;
2765 if (i == sh->pd_idx)
2767 switch(conf->level) {
2770 switch (algorithm) {
2771 case ALGORITHM_LEFT_ASYMMETRIC:
2772 case ALGORITHM_RIGHT_ASYMMETRIC:
2776 case ALGORITHM_LEFT_SYMMETRIC:
2777 case ALGORITHM_RIGHT_SYMMETRIC:
2780 i -= (sh->pd_idx + 1);
2782 case ALGORITHM_PARITY_0:
2785 case ALGORITHM_PARITY_N:
2792 if (i == sh->qd_idx)
2793 return 0; /* It is the Q disk */
2794 switch (algorithm) {
2795 case ALGORITHM_LEFT_ASYMMETRIC:
2796 case ALGORITHM_RIGHT_ASYMMETRIC:
2797 case ALGORITHM_ROTATING_ZERO_RESTART:
2798 case ALGORITHM_ROTATING_N_RESTART:
2799 if (sh->pd_idx == raid_disks-1)
2800 i--; /* Q D D D P */
2801 else if (i > sh->pd_idx)
2802 i -= 2; /* D D P Q D */
2804 case ALGORITHM_LEFT_SYMMETRIC:
2805 case ALGORITHM_RIGHT_SYMMETRIC:
2806 if (sh->pd_idx == raid_disks-1)
2807 i--; /* Q D D D P */
2812 i -= (sh->pd_idx + 2);
2815 case ALGORITHM_PARITY_0:
2818 case ALGORITHM_PARITY_N:
2820 case ALGORITHM_ROTATING_N_CONTINUE:
2821 /* Like left_symmetric, but P is before Q */
2822 if (sh->pd_idx == 0)
2823 i--; /* P D D D Q */
2828 i -= (sh->pd_idx + 1);
2831 case ALGORITHM_LEFT_ASYMMETRIC_6:
2832 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2836 case ALGORITHM_LEFT_SYMMETRIC_6:
2837 case ALGORITHM_RIGHT_SYMMETRIC_6:
2839 i += data_disks + 1;
2840 i -= (sh->pd_idx + 1);
2842 case ALGORITHM_PARITY_0_6:
2851 chunk_number = stripe * data_disks + i;
2852 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2854 check = raid5_compute_sector(conf, r_sector,
2855 previous, &dummy1, &sh2);
2856 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2857 || sh2.qd_idx != sh->qd_idx) {
2858 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2859 mdname(conf->mddev));
2866 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2867 int rcw, int expand)
2869 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2870 struct r5conf *conf = sh->raid_conf;
2871 int level = conf->level;
2875 for (i = disks; i--; ) {
2876 struct r5dev *dev = &sh->dev[i];
2879 set_bit(R5_LOCKED, &dev->flags);
2880 set_bit(R5_Wantdrain, &dev->flags);
2882 clear_bit(R5_UPTODATE, &dev->flags);
2886 /* if we are not expanding this is a proper write request, and
2887 * there will be bios with new data to be drained into the
2892 /* False alarm, nothing to do */
2894 sh->reconstruct_state = reconstruct_state_drain_run;
2895 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2897 sh->reconstruct_state = reconstruct_state_run;
2899 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2901 if (s->locked + conf->max_degraded == disks)
2902 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2903 atomic_inc(&conf->pending_full_writes);
2905 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2906 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2907 BUG_ON(level == 6 &&
2908 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2909 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2911 for (i = disks; i--; ) {
2912 struct r5dev *dev = &sh->dev[i];
2913 if (i == pd_idx || i == qd_idx)
2917 (test_bit(R5_UPTODATE, &dev->flags) ||
2918 test_bit(R5_Wantcompute, &dev->flags))) {
2919 set_bit(R5_Wantdrain, &dev->flags);
2920 set_bit(R5_LOCKED, &dev->flags);
2921 clear_bit(R5_UPTODATE, &dev->flags);
2926 /* False alarm - nothing to do */
2928 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2929 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2930 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2931 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2934 /* keep the parity disk(s) locked while asynchronous operations
2937 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2938 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2942 int qd_idx = sh->qd_idx;
2943 struct r5dev *dev = &sh->dev[qd_idx];
2945 set_bit(R5_LOCKED, &dev->flags);
2946 clear_bit(R5_UPTODATE, &dev->flags);
2950 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2951 __func__, (unsigned long long)sh->sector,
2952 s->locked, s->ops_request);
2956 * Each stripe/dev can have one or more bion attached.
2957 * toread/towrite point to the first in a chain.
2958 * The bi_next chain must be in order.
2960 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2961 int forwrite, int previous)
2964 struct r5conf *conf = sh->raid_conf;
2967 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2968 (unsigned long long)bi->bi_iter.bi_sector,
2969 (unsigned long long)sh->sector);
2972 * If several bio share a stripe. The bio bi_phys_segments acts as a
2973 * reference count to avoid race. The reference count should already be
2974 * increased before this function is called (for example, in
2975 * make_request()), so other bio sharing this stripe will not free the
2976 * stripe. If a stripe is owned by one stripe, the stripe lock will
2979 spin_lock_irq(&sh->stripe_lock);
2980 /* Don't allow new IO added to stripes in batch list */
2984 bip = &sh->dev[dd_idx].towrite;
2988 bip = &sh->dev[dd_idx].toread;
2989 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2990 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2992 bip = & (*bip)->bi_next;
2994 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2997 if (!forwrite || previous)
2998 clear_bit(STRIPE_BATCH_READY, &sh->state);
3000 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3004 raid5_inc_bi_active_stripes(bi);
3007 /* check if page is covered */
3008 sector_t sector = sh->dev[dd_idx].sector;
3009 for (bi=sh->dev[dd_idx].towrite;
3010 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3011 bi && bi->bi_iter.bi_sector <= sector;
3012 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3013 if (bio_end_sector(bi) >= sector)
3014 sector = bio_end_sector(bi);
3016 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3017 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3018 sh->overwrite_disks++;
3021 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3022 (unsigned long long)(*bip)->bi_iter.bi_sector,
3023 (unsigned long long)sh->sector, dd_idx);
3025 if (conf->mddev->bitmap && firstwrite) {
3026 /* Cannot hold spinlock over bitmap_startwrite,
3027 * but must ensure this isn't added to a batch until
3028 * we have added to the bitmap and set bm_seq.
3029 * So set STRIPE_BITMAP_PENDING to prevent
3031 * If multiple add_stripe_bio() calls race here they
3032 * much all set STRIPE_BITMAP_PENDING. So only the first one
3033 * to complete "bitmap_startwrite" gets to set
3034 * STRIPE_BIT_DELAY. This is important as once a stripe
3035 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3038 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3039 spin_unlock_irq(&sh->stripe_lock);
3040 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3042 spin_lock_irq(&sh->stripe_lock);
3043 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3044 if (!sh->batch_head) {
3045 sh->bm_seq = conf->seq_flush+1;
3046 set_bit(STRIPE_BIT_DELAY, &sh->state);
3049 spin_unlock_irq(&sh->stripe_lock);
3051 if (stripe_can_batch(sh))
3052 stripe_add_to_batch_list(conf, sh);
3056 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3057 spin_unlock_irq(&sh->stripe_lock);
3061 static void end_reshape(struct r5conf *conf);
3063 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3064 struct stripe_head *sh)
3066 int sectors_per_chunk =
3067 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3069 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3070 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3072 raid5_compute_sector(conf,
3073 stripe * (disks - conf->max_degraded)
3074 *sectors_per_chunk + chunk_offset,
3080 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3081 struct stripe_head_state *s, int disks,
3082 struct bio_list *return_bi)
3085 BUG_ON(sh->batch_head);
3086 for (i = disks; i--; ) {
3090 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3091 struct md_rdev *rdev;
3093 rdev = rcu_dereference(conf->disks[i].rdev);
3094 if (rdev && test_bit(In_sync, &rdev->flags))
3095 atomic_inc(&rdev->nr_pending);
3100 if (!rdev_set_badblocks(
3104 md_error(conf->mddev, rdev);
3105 rdev_dec_pending(rdev, conf->mddev);
3108 spin_lock_irq(&sh->stripe_lock);
3109 /* fail all writes first */
3110 bi = sh->dev[i].towrite;
3111 sh->dev[i].towrite = NULL;
3112 sh->overwrite_disks = 0;
3113 spin_unlock_irq(&sh->stripe_lock);
3117 r5l_stripe_write_finished(sh);
3119 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3120 wake_up(&conf->wait_for_overlap);
3122 while (bi && bi->bi_iter.bi_sector <
3123 sh->dev[i].sector + STRIPE_SECTORS) {
3124 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3126 bi->bi_error = -EIO;
3127 if (!raid5_dec_bi_active_stripes(bi)) {
3128 md_write_end(conf->mddev);
3129 bio_list_add(return_bi, bi);
3134 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3135 STRIPE_SECTORS, 0, 0);
3137 /* and fail all 'written' */
3138 bi = sh->dev[i].written;
3139 sh->dev[i].written = NULL;
3140 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3141 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3142 sh->dev[i].page = sh->dev[i].orig_page;
3145 if (bi) bitmap_end = 1;
3146 while (bi && bi->bi_iter.bi_sector <
3147 sh->dev[i].sector + STRIPE_SECTORS) {
3148 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3150 bi->bi_error = -EIO;
3151 if (!raid5_dec_bi_active_stripes(bi)) {
3152 md_write_end(conf->mddev);
3153 bio_list_add(return_bi, bi);
3158 /* fail any reads if this device is non-operational and
3159 * the data has not reached the cache yet.
3161 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3162 s->failed > conf->max_degraded &&
3163 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3164 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3165 spin_lock_irq(&sh->stripe_lock);
3166 bi = sh->dev[i].toread;
3167 sh->dev[i].toread = NULL;
3168 spin_unlock_irq(&sh->stripe_lock);
3169 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3170 wake_up(&conf->wait_for_overlap);
3173 while (bi && bi->bi_iter.bi_sector <
3174 sh->dev[i].sector + STRIPE_SECTORS) {
3175 struct bio *nextbi =
3176 r5_next_bio(bi, sh->dev[i].sector);
3178 bi->bi_error = -EIO;
3179 if (!raid5_dec_bi_active_stripes(bi))
3180 bio_list_add(return_bi, bi);
3185 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3186 STRIPE_SECTORS, 0, 0);
3187 /* If we were in the middle of a write the parity block might
3188 * still be locked - so just clear all R5_LOCKED flags
3190 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3195 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3196 if (atomic_dec_and_test(&conf->pending_full_writes))
3197 md_wakeup_thread(conf->mddev->thread);
3201 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3202 struct stripe_head_state *s)
3207 BUG_ON(sh->batch_head);
3208 clear_bit(STRIPE_SYNCING, &sh->state);
3209 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3210 wake_up(&conf->wait_for_overlap);
3213 /* There is nothing more to do for sync/check/repair.
3214 * Don't even need to abort as that is handled elsewhere
3215 * if needed, and not always wanted e.g. if there is a known
3217 * For recover/replace we need to record a bad block on all
3218 * non-sync devices, or abort the recovery
3220 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3221 /* During recovery devices cannot be removed, so
3222 * locking and refcounting of rdevs is not needed
3224 for (i = 0; i < conf->raid_disks; i++) {
3225 struct md_rdev *rdev = conf->disks[i].rdev;
3227 && !test_bit(Faulty, &rdev->flags)
3228 && !test_bit(In_sync, &rdev->flags)
3229 && !rdev_set_badblocks(rdev, sh->sector,
3232 rdev = conf->disks[i].replacement;
3234 && !test_bit(Faulty, &rdev->flags)
3235 && !test_bit(In_sync, &rdev->flags)
3236 && !rdev_set_badblocks(rdev, sh->sector,
3241 conf->recovery_disabled =
3242 conf->mddev->recovery_disabled;
3244 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3247 static int want_replace(struct stripe_head *sh, int disk_idx)
3249 struct md_rdev *rdev;
3251 /* Doing recovery so rcu locking not required */
3252 rdev = sh->raid_conf->disks[disk_idx].replacement;
3254 && !test_bit(Faulty, &rdev->flags)
3255 && !test_bit(In_sync, &rdev->flags)
3256 && (rdev->recovery_offset <= sh->sector
3257 || rdev->mddev->recovery_cp <= sh->sector))
3263 /* fetch_block - checks the given member device to see if its data needs
3264 * to be read or computed to satisfy a request.
3266 * Returns 1 when no more member devices need to be checked, otherwise returns
3267 * 0 to tell the loop in handle_stripe_fill to continue
3270 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3271 int disk_idx, int disks)
3273 struct r5dev *dev = &sh->dev[disk_idx];
3274 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3275 &sh->dev[s->failed_num[1]] };
3279 if (test_bit(R5_LOCKED, &dev->flags) ||
3280 test_bit(R5_UPTODATE, &dev->flags))
3281 /* No point reading this as we already have it or have
3282 * decided to get it.
3287 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3288 /* We need this block to directly satisfy a request */
3291 if (s->syncing || s->expanding ||
3292 (s->replacing && want_replace(sh, disk_idx)))
3293 /* When syncing, or expanding we read everything.
3294 * When replacing, we need the replaced block.
3298 if ((s->failed >= 1 && fdev[0]->toread) ||
3299 (s->failed >= 2 && fdev[1]->toread))
3300 /* If we want to read from a failed device, then
3301 * we need to actually read every other device.
3305 /* Sometimes neither read-modify-write nor reconstruct-write
3306 * cycles can work. In those cases we read every block we
3307 * can. Then the parity-update is certain to have enough to
3309 * This can only be a problem when we need to write something,
3310 * and some device has failed. If either of those tests
3311 * fail we need look no further.
3313 if (!s->failed || !s->to_write)
3316 if (test_bit(R5_Insync, &dev->flags) &&
3317 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3318 /* Pre-reads at not permitted until after short delay
3319 * to gather multiple requests. However if this
3320 * device is no Insync, the block could only be be computed
3321 * and there is no need to delay that.
3325 for (i = 0; i < s->failed && i < 2; i++) {
3326 if (fdev[i]->towrite &&
3327 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3328 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3329 /* If we have a partial write to a failed
3330 * device, then we will need to reconstruct
3331 * the content of that device, so all other
3332 * devices must be read.
3337 /* If we are forced to do a reconstruct-write, either because
3338 * the current RAID6 implementation only supports that, or
3339 * or because parity cannot be trusted and we are currently
3340 * recovering it, there is extra need to be careful.
3341 * If one of the devices that we would need to read, because
3342 * it is not being overwritten (and maybe not written at all)
3343 * is missing/faulty, then we need to read everything we can.
3345 if (sh->raid_conf->level != 6 &&
3346 sh->sector < sh->raid_conf->mddev->recovery_cp)
3347 /* reconstruct-write isn't being forced */
3349 for (i = 0; i < s->failed && i < 2; i++) {
3350 if (s->failed_num[i] != sh->pd_idx &&
3351 s->failed_num[i] != sh->qd_idx &&
3352 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3353 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3360 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3361 int disk_idx, int disks)
3363 struct r5dev *dev = &sh->dev[disk_idx];
3365 /* is the data in this block needed, and can we get it? */
3366 if (need_this_block(sh, s, disk_idx, disks)) {
3367 /* we would like to get this block, possibly by computing it,
3368 * otherwise read it if the backing disk is insync
3370 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3371 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3372 BUG_ON(sh->batch_head);
3373 if ((s->uptodate == disks - 1) &&
3374 (s->failed && (disk_idx == s->failed_num[0] ||
3375 disk_idx == s->failed_num[1]))) {
3376 /* have disk failed, and we're requested to fetch it;
3379 pr_debug("Computing stripe %llu block %d\n",
3380 (unsigned long long)sh->sector, disk_idx);
3381 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3382 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3383 set_bit(R5_Wantcompute, &dev->flags);
3384 sh->ops.target = disk_idx;
3385 sh->ops.target2 = -1; /* no 2nd target */
3387 /* Careful: from this point on 'uptodate' is in the eye
3388 * of raid_run_ops which services 'compute' operations
3389 * before writes. R5_Wantcompute flags a block that will
3390 * be R5_UPTODATE by the time it is needed for a
3391 * subsequent operation.
3395 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3396 /* Computing 2-failure is *very* expensive; only
3397 * do it if failed >= 2
3400 for (other = disks; other--; ) {
3401 if (other == disk_idx)
3403 if (!test_bit(R5_UPTODATE,
3404 &sh->dev[other].flags))
3408 pr_debug("Computing stripe %llu blocks %d,%d\n",
3409 (unsigned long long)sh->sector,
3411 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3412 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3413 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3414 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3415 sh->ops.target = disk_idx;
3416 sh->ops.target2 = other;
3420 } else if (test_bit(R5_Insync, &dev->flags)) {
3421 set_bit(R5_LOCKED, &dev->flags);
3422 set_bit(R5_Wantread, &dev->flags);
3424 pr_debug("Reading block %d (sync=%d)\n",
3425 disk_idx, s->syncing);
3433 * handle_stripe_fill - read or compute data to satisfy pending requests.
3435 static void handle_stripe_fill(struct stripe_head *sh,
3436 struct stripe_head_state *s,
3441 /* look for blocks to read/compute, skip this if a compute
3442 * is already in flight, or if the stripe contents are in the
3443 * midst of changing due to a write
3445 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3446 !sh->reconstruct_state)
3447 for (i = disks; i--; )
3448 if (fetch_block(sh, s, i, disks))
3450 set_bit(STRIPE_HANDLE, &sh->state);
3453 static void break_stripe_batch_list(struct stripe_head *head_sh,
3454 unsigned long handle_flags);
3455 /* handle_stripe_clean_event
3456 * any written block on an uptodate or failed drive can be returned.
3457 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3458 * never LOCKED, so we don't need to test 'failed' directly.
3460 static void handle_stripe_clean_event(struct r5conf *conf,
3461 struct stripe_head *sh, int disks, struct bio_list *return_bi)
3465 int discard_pending = 0;
3466 struct stripe_head *head_sh = sh;
3467 bool do_endio = false;
3469 for (i = disks; i--; )
3470 if (sh->dev[i].written) {
3472 if (!test_bit(R5_LOCKED, &dev->flags) &&
3473 (test_bit(R5_UPTODATE, &dev->flags) ||
3474 test_bit(R5_Discard, &dev->flags) ||
3475 test_bit(R5_SkipCopy, &dev->flags))) {
3476 /* We can return any write requests */
3477 struct bio *wbi, *wbi2;
3478 pr_debug("Return write for disc %d\n", i);
3479 if (test_and_clear_bit(R5_Discard, &dev->flags))
3480 clear_bit(R5_UPTODATE, &dev->flags);
3481 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3482 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3487 dev->page = dev->orig_page;
3489 dev->written = NULL;
3490 while (wbi && wbi->bi_iter.bi_sector <
3491 dev->sector + STRIPE_SECTORS) {
3492 wbi2 = r5_next_bio(wbi, dev->sector);
3493 if (!raid5_dec_bi_active_stripes(wbi)) {
3494 md_write_end(conf->mddev);
3495 bio_list_add(return_bi, wbi);
3499 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3501 !test_bit(STRIPE_DEGRADED, &sh->state),
3503 if (head_sh->batch_head) {
3504 sh = list_first_entry(&sh->batch_list,
3507 if (sh != head_sh) {
3514 } else if (test_bit(R5_Discard, &dev->flags))
3515 discard_pending = 1;
3516 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3517 WARN_ON(dev->page != dev->orig_page);
3520 r5l_stripe_write_finished(sh);
3522 if (!discard_pending &&
3523 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3525 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3526 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3527 if (sh->qd_idx >= 0) {
3528 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3529 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3531 /* now that discard is done we can proceed with any sync */
3532 clear_bit(STRIPE_DISCARD, &sh->state);
3534 * SCSI discard will change some bio fields and the stripe has
3535 * no updated data, so remove it from hash list and the stripe
3536 * will be reinitialized
3539 hash = sh->hash_lock_index;
3540 spin_lock_irq(conf->hash_locks + hash);
3542 spin_unlock_irq(conf->hash_locks + hash);
3543 if (head_sh->batch_head) {
3544 sh = list_first_entry(&sh->batch_list,
3545 struct stripe_head, batch_list);
3551 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3552 set_bit(STRIPE_HANDLE, &sh->state);
3556 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3557 if (atomic_dec_and_test(&conf->pending_full_writes))
3558 md_wakeup_thread(conf->mddev->thread);
3560 if (head_sh->batch_head && do_endio)
3561 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3564 static void handle_stripe_dirtying(struct r5conf *conf,
3565 struct stripe_head *sh,
3566 struct stripe_head_state *s,
3569 int rmw = 0, rcw = 0, i;
3570 sector_t recovery_cp = conf->mddev->recovery_cp;
3572 /* Check whether resync is now happening or should start.
3573 * If yes, then the array is dirty (after unclean shutdown or
3574 * initial creation), so parity in some stripes might be inconsistent.
3575 * In this case, we need to always do reconstruct-write, to ensure
3576 * that in case of drive failure or read-error correction, we
3577 * generate correct data from the parity.
3579 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3580 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3582 /* Calculate the real rcw later - for now make it
3583 * look like rcw is cheaper
3586 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3587 conf->rmw_level, (unsigned long long)recovery_cp,
3588 (unsigned long long)sh->sector);
3589 } else for (i = disks; i--; ) {
3590 /* would I have to read this buffer for read_modify_write */
3591 struct r5dev *dev = &sh->dev[i];
3592 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3593 !test_bit(R5_LOCKED, &dev->flags) &&
3594 !(test_bit(R5_UPTODATE, &dev->flags) ||
3595 test_bit(R5_Wantcompute, &dev->flags))) {
3596 if (test_bit(R5_Insync, &dev->flags))
3599 rmw += 2*disks; /* cannot read it */
3601 /* Would I have to read this buffer for reconstruct_write */
3602 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3603 i != sh->pd_idx && i != sh->qd_idx &&
3604 !test_bit(R5_LOCKED, &dev->flags) &&
3605 !(test_bit(R5_UPTODATE, &dev->flags) ||
3606 test_bit(R5_Wantcompute, &dev->flags))) {
3607 if (test_bit(R5_Insync, &dev->flags))
3613 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3614 (unsigned long long)sh->sector, rmw, rcw);
3615 set_bit(STRIPE_HANDLE, &sh->state);
3616 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3617 /* prefer read-modify-write, but need to get some data */
3618 if (conf->mddev->queue)
3619 blk_add_trace_msg(conf->mddev->queue,
3620 "raid5 rmw %llu %d",
3621 (unsigned long long)sh->sector, rmw);
3622 for (i = disks; i--; ) {
3623 struct r5dev *dev = &sh->dev[i];
3624 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3625 !test_bit(R5_LOCKED, &dev->flags) &&
3626 !(test_bit(R5_UPTODATE, &dev->flags) ||
3627 test_bit(R5_Wantcompute, &dev->flags)) &&
3628 test_bit(R5_Insync, &dev->flags)) {
3629 if (test_bit(STRIPE_PREREAD_ACTIVE,
3631 pr_debug("Read_old block %d for r-m-w\n",
3633 set_bit(R5_LOCKED, &dev->flags);
3634 set_bit(R5_Wantread, &dev->flags);
3637 set_bit(STRIPE_DELAYED, &sh->state);
3638 set_bit(STRIPE_HANDLE, &sh->state);
3643 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3644 /* want reconstruct write, but need to get some data */
3647 for (i = disks; i--; ) {
3648 struct r5dev *dev = &sh->dev[i];
3649 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3650 i != sh->pd_idx && i != sh->qd_idx &&
3651 !test_bit(R5_LOCKED, &dev->flags) &&
3652 !(test_bit(R5_UPTODATE, &dev->flags) ||
3653 test_bit(R5_Wantcompute, &dev->flags))) {
3655 if (test_bit(R5_Insync, &dev->flags) &&
3656 test_bit(STRIPE_PREREAD_ACTIVE,
3658 pr_debug("Read_old block "
3659 "%d for Reconstruct\n", i);
3660 set_bit(R5_LOCKED, &dev->flags);
3661 set_bit(R5_Wantread, &dev->flags);
3665 set_bit(STRIPE_DELAYED, &sh->state);
3666 set_bit(STRIPE_HANDLE, &sh->state);
3670 if (rcw && conf->mddev->queue)
3671 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3672 (unsigned long long)sh->sector,
3673 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3676 if (rcw > disks && rmw > disks &&
3677 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3678 set_bit(STRIPE_DELAYED, &sh->state);
3680 /* now if nothing is locked, and if we have enough data,
3681 * we can start a write request
3683 /* since handle_stripe can be called at any time we need to handle the
3684 * case where a compute block operation has been submitted and then a
3685 * subsequent call wants to start a write request. raid_run_ops only
3686 * handles the case where compute block and reconstruct are requested
3687 * simultaneously. If this is not the case then new writes need to be
3688 * held off until the compute completes.
3690 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3691 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3692 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3693 schedule_reconstruction(sh, s, rcw == 0, 0);
3696 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3697 struct stripe_head_state *s, int disks)
3699 struct r5dev *dev = NULL;
3701 BUG_ON(sh->batch_head);
3702 set_bit(STRIPE_HANDLE, &sh->state);
3704 switch (sh->check_state) {
3705 case check_state_idle:
3706 /* start a new check operation if there are no failures */
3707 if (s->failed == 0) {
3708 BUG_ON(s->uptodate != disks);
3709 sh->check_state = check_state_run;
3710 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3711 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3715 dev = &sh->dev[s->failed_num[0]];
3717 case check_state_compute_result:
3718 sh->check_state = check_state_idle;
3720 dev = &sh->dev[sh->pd_idx];
3722 /* check that a write has not made the stripe insync */
3723 if (test_bit(STRIPE_INSYNC, &sh->state))
3726 /* either failed parity check, or recovery is happening */
3727 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3728 BUG_ON(s->uptodate != disks);
3730 set_bit(R5_LOCKED, &dev->flags);
3732 set_bit(R5_Wantwrite, &dev->flags);
3734 clear_bit(STRIPE_DEGRADED, &sh->state);
3735 set_bit(STRIPE_INSYNC, &sh->state);
3737 case check_state_run:
3738 break; /* we will be called again upon completion */
3739 case check_state_check_result:
3740 sh->check_state = check_state_idle;
3742 /* if a failure occurred during the check operation, leave
3743 * STRIPE_INSYNC not set and let the stripe be handled again
3748 /* handle a successful check operation, if parity is correct
3749 * we are done. Otherwise update the mismatch count and repair
3750 * parity if !MD_RECOVERY_CHECK
3752 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3753 /* parity is correct (on disc,
3754 * not in buffer any more)
3756 set_bit(STRIPE_INSYNC, &sh->state);
3758 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3759 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3760 /* don't try to repair!! */
3761 set_bit(STRIPE_INSYNC, &sh->state);
3763 sh->check_state = check_state_compute_run;
3764 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3765 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3766 set_bit(R5_Wantcompute,
3767 &sh->dev[sh->pd_idx].flags);
3768 sh->ops.target = sh->pd_idx;
3769 sh->ops.target2 = -1;
3774 case check_state_compute_run:
3777 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3778 __func__, sh->check_state,
3779 (unsigned long long) sh->sector);
3784 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3785 struct stripe_head_state *s,
3788 int pd_idx = sh->pd_idx;
3789 int qd_idx = sh->qd_idx;
3792 BUG_ON(sh->batch_head);
3793 set_bit(STRIPE_HANDLE, &sh->state);
3795 BUG_ON(s->failed > 2);
3797 /* Want to check and possibly repair P and Q.
3798 * However there could be one 'failed' device, in which
3799 * case we can only check one of them, possibly using the
3800 * other to generate missing data
3803 switch (sh->check_state) {
3804 case check_state_idle:
3805 /* start a new check operation if there are < 2 failures */
3806 if (s->failed == s->q_failed) {
3807 /* The only possible failed device holds Q, so it
3808 * makes sense to check P (If anything else were failed,
3809 * we would have used P to recreate it).
3811 sh->check_state = check_state_run;
3813 if (!s->q_failed && s->failed < 2) {
3814 /* Q is not failed, and we didn't use it to generate
3815 * anything, so it makes sense to check it
3817 if (sh->check_state == check_state_run)
3818 sh->check_state = check_state_run_pq;
3820 sh->check_state = check_state_run_q;
3823 /* discard potentially stale zero_sum_result */
3824 sh->ops.zero_sum_result = 0;
3826 if (sh->check_state == check_state_run) {
3827 /* async_xor_zero_sum destroys the contents of P */
3828 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3831 if (sh->check_state >= check_state_run &&
3832 sh->check_state <= check_state_run_pq) {
3833 /* async_syndrome_zero_sum preserves P and Q, so
3834 * no need to mark them !uptodate here
3836 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3840 /* we have 2-disk failure */
3841 BUG_ON(s->failed != 2);
3843 case check_state_compute_result:
3844 sh->check_state = check_state_idle;
3846 /* check that a write has not made the stripe insync */
3847 if (test_bit(STRIPE_INSYNC, &sh->state))
3850 /* now write out any block on a failed drive,
3851 * or P or Q if they were recomputed
3853 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3854 if (s->failed == 2) {
3855 dev = &sh->dev[s->failed_num[1]];
3857 set_bit(R5_LOCKED, &dev->flags);
3858 set_bit(R5_Wantwrite, &dev->flags);
3860 if (s->failed >= 1) {
3861 dev = &sh->dev[s->failed_num[0]];
3863 set_bit(R5_LOCKED, &dev->flags);
3864 set_bit(R5_Wantwrite, &dev->flags);
3866 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3867 dev = &sh->dev[pd_idx];
3869 set_bit(R5_LOCKED, &dev->flags);
3870 set_bit(R5_Wantwrite, &dev->flags);
3872 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3873 dev = &sh->dev[qd_idx];
3875 set_bit(R5_LOCKED, &dev->flags);
3876 set_bit(R5_Wantwrite, &dev->flags);
3878 clear_bit(STRIPE_DEGRADED, &sh->state);
3880 set_bit(STRIPE_INSYNC, &sh->state);
3882 case check_state_run:
3883 case check_state_run_q:
3884 case check_state_run_pq:
3885 break; /* we will be called again upon completion */
3886 case check_state_check_result:
3887 sh->check_state = check_state_idle;
3889 /* handle a successful check operation, if parity is correct
3890 * we are done. Otherwise update the mismatch count and repair
3891 * parity if !MD_RECOVERY_CHECK
3893 if (sh->ops.zero_sum_result == 0) {
3894 /* both parities are correct */
3896 set_bit(STRIPE_INSYNC, &sh->state);
3898 /* in contrast to the raid5 case we can validate
3899 * parity, but still have a failure to write
3902 sh->check_state = check_state_compute_result;
3903 /* Returning at this point means that we may go
3904 * off and bring p and/or q uptodate again so
3905 * we make sure to check zero_sum_result again
3906 * to verify if p or q need writeback
3910 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3911 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3912 /* don't try to repair!! */
3913 set_bit(STRIPE_INSYNC, &sh->state);
3915 int *target = &sh->ops.target;
3917 sh->ops.target = -1;
3918 sh->ops.target2 = -1;
3919 sh->check_state = check_state_compute_run;
3920 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3921 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3922 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3923 set_bit(R5_Wantcompute,
3924 &sh->dev[pd_idx].flags);
3926 target = &sh->ops.target2;
3929 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3930 set_bit(R5_Wantcompute,
3931 &sh->dev[qd_idx].flags);
3938 case check_state_compute_run:
3941 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3942 __func__, sh->check_state,
3943 (unsigned long long) sh->sector);
3948 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3952 /* We have read all the blocks in this stripe and now we need to
3953 * copy some of them into a target stripe for expand.
3955 struct dma_async_tx_descriptor *tx = NULL;
3956 BUG_ON(sh->batch_head);
3957 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3958 for (i = 0; i < sh->disks; i++)
3959 if (i != sh->pd_idx && i != sh->qd_idx) {
3961 struct stripe_head *sh2;
3962 struct async_submit_ctl submit;
3964 sector_t bn = raid5_compute_blocknr(sh, i, 1);
3965 sector_t s = raid5_compute_sector(conf, bn, 0,
3967 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
3969 /* so far only the early blocks of this stripe
3970 * have been requested. When later blocks
3971 * get requested, we will try again
3974 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3975 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3976 /* must have already done this block */
3977 raid5_release_stripe(sh2);
3981 /* place all the copies on one channel */
3982 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3983 tx = async_memcpy(sh2->dev[dd_idx].page,
3984 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3987 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3988 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3989 for (j = 0; j < conf->raid_disks; j++)
3990 if (j != sh2->pd_idx &&
3992 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3994 if (j == conf->raid_disks) {
3995 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3996 set_bit(STRIPE_HANDLE, &sh2->state);
3998 raid5_release_stripe(sh2);
4001 /* done submitting copies, wait for them to complete */
4002 async_tx_quiesce(&tx);
4006 * handle_stripe - do things to a stripe.
4008 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4009 * state of various bits to see what needs to be done.
4011 * return some read requests which now have data
4012 * return some write requests which are safely on storage
4013 * schedule a read on some buffers
4014 * schedule a write of some buffers
4015 * return confirmation of parity correctness
4019 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4021 struct r5conf *conf = sh->raid_conf;
4022 int disks = sh->disks;
4025 int do_recovery = 0;
4027 memset(s, 0, sizeof(*s));
4029 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4030 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4031 s->failed_num[0] = -1;
4032 s->failed_num[1] = -1;
4033 s->log_failed = r5l_log_disk_error(conf);
4035 /* Now to look around and see what can be done */
4037 for (i=disks; i--; ) {
4038 struct md_rdev *rdev;
4045 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4047 dev->toread, dev->towrite, dev->written);
4048 /* maybe we can reply to a read
4050 * new wantfill requests are only permitted while
4051 * ops_complete_biofill is guaranteed to be inactive
4053 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4054 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4055 set_bit(R5_Wantfill, &dev->flags);
4057 /* now count some things */
4058 if (test_bit(R5_LOCKED, &dev->flags))
4060 if (test_bit(R5_UPTODATE, &dev->flags))
4062 if (test_bit(R5_Wantcompute, &dev->flags)) {
4064 BUG_ON(s->compute > 2);
4067 if (test_bit(R5_Wantfill, &dev->flags))
4069 else if (dev->toread)
4073 if (!test_bit(R5_OVERWRITE, &dev->flags))
4078 /* Prefer to use the replacement for reads, but only
4079 * if it is recovered enough and has no bad blocks.
4081 rdev = rcu_dereference(conf->disks[i].replacement);
4082 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4083 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4084 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4085 &first_bad, &bad_sectors))
4086 set_bit(R5_ReadRepl, &dev->flags);
4088 if (rdev && !test_bit(Faulty, &rdev->flags))
4089 set_bit(R5_NeedReplace, &dev->flags);
4091 clear_bit(R5_NeedReplace, &dev->flags);
4092 rdev = rcu_dereference(conf->disks[i].rdev);
4093 clear_bit(R5_ReadRepl, &dev->flags);
4095 if (rdev && test_bit(Faulty, &rdev->flags))
4098 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4099 &first_bad, &bad_sectors);
4100 if (s->blocked_rdev == NULL
4101 && (test_bit(Blocked, &rdev->flags)
4104 set_bit(BlockedBadBlocks,
4106 s->blocked_rdev = rdev;
4107 atomic_inc(&rdev->nr_pending);
4110 clear_bit(R5_Insync, &dev->flags);
4114 /* also not in-sync */
4115 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4116 test_bit(R5_UPTODATE, &dev->flags)) {
4117 /* treat as in-sync, but with a read error
4118 * which we can now try to correct
4120 set_bit(R5_Insync, &dev->flags);
4121 set_bit(R5_ReadError, &dev->flags);
4123 } else if (test_bit(In_sync, &rdev->flags))
4124 set_bit(R5_Insync, &dev->flags);
4125 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4126 /* in sync if before recovery_offset */
4127 set_bit(R5_Insync, &dev->flags);
4128 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4129 test_bit(R5_Expanded, &dev->flags))
4130 /* If we've reshaped into here, we assume it is Insync.
4131 * We will shortly update recovery_offset to make
4134 set_bit(R5_Insync, &dev->flags);
4136 if (test_bit(R5_WriteError, &dev->flags)) {
4137 /* This flag does not apply to '.replacement'
4138 * only to .rdev, so make sure to check that*/
4139 struct md_rdev *rdev2 = rcu_dereference(
4140 conf->disks[i].rdev);
4142 clear_bit(R5_Insync, &dev->flags);
4143 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4144 s->handle_bad_blocks = 1;
4145 atomic_inc(&rdev2->nr_pending);
4147 clear_bit(R5_WriteError, &dev->flags);
4149 if (test_bit(R5_MadeGood, &dev->flags)) {
4150 /* This flag does not apply to '.replacement'
4151 * only to .rdev, so make sure to check that*/
4152 struct md_rdev *rdev2 = rcu_dereference(
4153 conf->disks[i].rdev);
4154 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4155 s->handle_bad_blocks = 1;
4156 atomic_inc(&rdev2->nr_pending);
4158 clear_bit(R5_MadeGood, &dev->flags);
4160 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4161 struct md_rdev *rdev2 = rcu_dereference(
4162 conf->disks[i].replacement);
4163 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4164 s->handle_bad_blocks = 1;
4165 atomic_inc(&rdev2->nr_pending);
4167 clear_bit(R5_MadeGoodRepl, &dev->flags);
4169 if (!test_bit(R5_Insync, &dev->flags)) {
4170 /* The ReadError flag will just be confusing now */
4171 clear_bit(R5_ReadError, &dev->flags);
4172 clear_bit(R5_ReWrite, &dev->flags);
4174 if (test_bit(R5_ReadError, &dev->flags))
4175 clear_bit(R5_Insync, &dev->flags);
4176 if (!test_bit(R5_Insync, &dev->flags)) {
4178 s->failed_num[s->failed] = i;
4180 if (rdev && !test_bit(Faulty, &rdev->flags))
4184 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4185 /* If there is a failed device being replaced,
4186 * we must be recovering.
4187 * else if we are after recovery_cp, we must be syncing
4188 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4189 * else we can only be replacing
4190 * sync and recovery both need to read all devices, and so
4191 * use the same flag.
4194 sh->sector >= conf->mddev->recovery_cp ||
4195 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4203 static int clear_batch_ready(struct stripe_head *sh)
4205 /* Return '1' if this is a member of batch, or
4206 * '0' if it is a lone stripe or a head which can now be
4209 struct stripe_head *tmp;
4210 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4211 return (sh->batch_head && sh->batch_head != sh);
4212 spin_lock(&sh->stripe_lock);
4213 if (!sh->batch_head) {
4214 spin_unlock(&sh->stripe_lock);
4219 * this stripe could be added to a batch list before we check
4220 * BATCH_READY, skips it
4222 if (sh->batch_head != sh) {
4223 spin_unlock(&sh->stripe_lock);
4226 spin_lock(&sh->batch_lock);
4227 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4228 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4229 spin_unlock(&sh->batch_lock);
4230 spin_unlock(&sh->stripe_lock);
4233 * BATCH_READY is cleared, no new stripes can be added.
4234 * batch_list can be accessed without lock
4239 static void break_stripe_batch_list(struct stripe_head *head_sh,
4240 unsigned long handle_flags)
4242 struct stripe_head *sh, *next;
4246 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4248 list_del_init(&sh->batch_list);
4250 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4251 (1 << STRIPE_SYNCING) |
4252 (1 << STRIPE_REPLACED) |
4253 (1 << STRIPE_PREREAD_ACTIVE) |
4254 (1 << STRIPE_DELAYED) |
4255 (1 << STRIPE_BIT_DELAY) |
4256 (1 << STRIPE_FULL_WRITE) |
4257 (1 << STRIPE_BIOFILL_RUN) |
4258 (1 << STRIPE_COMPUTE_RUN) |
4259 (1 << STRIPE_OPS_REQ_PENDING) |
4260 (1 << STRIPE_DISCARD) |
4261 (1 << STRIPE_BATCH_READY) |
4262 (1 << STRIPE_BATCH_ERR) |
4263 (1 << STRIPE_BITMAP_PENDING)));
4264 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4265 (1 << STRIPE_REPLACED)));
4267 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4268 (1 << STRIPE_DEGRADED)),
4269 head_sh->state & (1 << STRIPE_INSYNC));
4271 sh->check_state = head_sh->check_state;
4272 sh->reconstruct_state = head_sh->reconstruct_state;
4273 for (i = 0; i < sh->disks; i++) {
4274 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4276 sh->dev[i].flags = head_sh->dev[i].flags &
4277 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4279 spin_lock_irq(&sh->stripe_lock);
4280 sh->batch_head = NULL;
4281 spin_unlock_irq(&sh->stripe_lock);
4282 if (handle_flags == 0 ||
4283 sh->state & handle_flags)
4284 set_bit(STRIPE_HANDLE, &sh->state);
4285 raid5_release_stripe(sh);
4287 spin_lock_irq(&head_sh->stripe_lock);
4288 head_sh->batch_head = NULL;
4289 spin_unlock_irq(&head_sh->stripe_lock);
4290 for (i = 0; i < head_sh->disks; i++)
4291 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4293 if (head_sh->state & handle_flags)
4294 set_bit(STRIPE_HANDLE, &head_sh->state);
4297 wake_up(&head_sh->raid_conf->wait_for_overlap);
4300 static void handle_stripe(struct stripe_head *sh)
4302 struct stripe_head_state s;
4303 struct r5conf *conf = sh->raid_conf;
4306 int disks = sh->disks;
4307 struct r5dev *pdev, *qdev;
4309 clear_bit(STRIPE_HANDLE, &sh->state);
4310 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4311 /* already being handled, ensure it gets handled
4312 * again when current action finishes */
4313 set_bit(STRIPE_HANDLE, &sh->state);
4317 if (clear_batch_ready(sh) ) {
4318 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4322 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4323 break_stripe_batch_list(sh, 0);
4325 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4326 spin_lock(&sh->stripe_lock);
4327 /* Cannot process 'sync' concurrently with 'discard' */
4328 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4329 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4330 set_bit(STRIPE_SYNCING, &sh->state);
4331 clear_bit(STRIPE_INSYNC, &sh->state);
4332 clear_bit(STRIPE_REPLACED, &sh->state);
4334 spin_unlock(&sh->stripe_lock);
4336 clear_bit(STRIPE_DELAYED, &sh->state);
4338 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4339 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4340 (unsigned long long)sh->sector, sh->state,
4341 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4342 sh->check_state, sh->reconstruct_state);
4344 analyse_stripe(sh, &s);
4346 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4349 if (s.handle_bad_blocks) {
4350 set_bit(STRIPE_HANDLE, &sh->state);
4354 if (unlikely(s.blocked_rdev)) {
4355 if (s.syncing || s.expanding || s.expanded ||
4356 s.replacing || s.to_write || s.written) {
4357 set_bit(STRIPE_HANDLE, &sh->state);
4360 /* There is nothing for the blocked_rdev to block */
4361 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4362 s.blocked_rdev = NULL;
4365 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4366 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4367 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4370 pr_debug("locked=%d uptodate=%d to_read=%d"
4371 " to_write=%d failed=%d failed_num=%d,%d\n",
4372 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4373 s.failed_num[0], s.failed_num[1]);
4374 /* check if the array has lost more than max_degraded devices and,
4375 * if so, some requests might need to be failed.
4377 if (s.failed > conf->max_degraded || s.log_failed) {
4378 sh->check_state = 0;
4379 sh->reconstruct_state = 0;
4380 break_stripe_batch_list(sh, 0);
4381 if (s.to_read+s.to_write+s.written)
4382 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4383 if (s.syncing + s.replacing)
4384 handle_failed_sync(conf, sh, &s);
4387 /* Now we check to see if any write operations have recently
4391 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4393 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4394 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4395 sh->reconstruct_state = reconstruct_state_idle;
4397 /* All the 'written' buffers and the parity block are ready to
4398 * be written back to disk
4400 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4401 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4402 BUG_ON(sh->qd_idx >= 0 &&
4403 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4404 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4405 for (i = disks; i--; ) {
4406 struct r5dev *dev = &sh->dev[i];
4407 if (test_bit(R5_LOCKED, &dev->flags) &&
4408 (i == sh->pd_idx || i == sh->qd_idx ||
4410 pr_debug("Writing block %d\n", i);
4411 set_bit(R5_Wantwrite, &dev->flags);
4416 if (!test_bit(R5_Insync, &dev->flags) ||
4417 ((i == sh->pd_idx || i == sh->qd_idx) &&
4419 set_bit(STRIPE_INSYNC, &sh->state);
4422 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4423 s.dec_preread_active = 1;
4427 * might be able to return some write requests if the parity blocks
4428 * are safe, or on a failed drive
4430 pdev = &sh->dev[sh->pd_idx];
4431 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4432 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4433 qdev = &sh->dev[sh->qd_idx];
4434 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4435 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4439 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4440 && !test_bit(R5_LOCKED, &pdev->flags)
4441 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4442 test_bit(R5_Discard, &pdev->flags))))) &&
4443 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4444 && !test_bit(R5_LOCKED, &qdev->flags)
4445 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4446 test_bit(R5_Discard, &qdev->flags))))))
4447 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4449 /* Now we might consider reading some blocks, either to check/generate
4450 * parity, or to satisfy requests
4451 * or to load a block that is being partially written.
4453 if (s.to_read || s.non_overwrite
4454 || (conf->level == 6 && s.to_write && s.failed)
4455 || (s.syncing && (s.uptodate + s.compute < disks))
4458 handle_stripe_fill(sh, &s, disks);
4460 /* Now to consider new write requests and what else, if anything
4461 * should be read. We do not handle new writes when:
4462 * 1/ A 'write' operation (copy+xor) is already in flight.
4463 * 2/ A 'check' operation is in flight, as it may clobber the parity
4466 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4467 handle_stripe_dirtying(conf, sh, &s, disks);
4469 /* maybe we need to check and possibly fix the parity for this stripe
4470 * Any reads will already have been scheduled, so we just see if enough
4471 * data is available. The parity check is held off while parity
4472 * dependent operations are in flight.
4474 if (sh->check_state ||
4475 (s.syncing && s.locked == 0 &&
4476 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4477 !test_bit(STRIPE_INSYNC, &sh->state))) {
4478 if (conf->level == 6)
4479 handle_parity_checks6(conf, sh, &s, disks);
4481 handle_parity_checks5(conf, sh, &s, disks);
4484 if ((s.replacing || s.syncing) && s.locked == 0
4485 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4486 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4487 /* Write out to replacement devices where possible */
4488 for (i = 0; i < conf->raid_disks; i++)
4489 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4490 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4491 set_bit(R5_WantReplace, &sh->dev[i].flags);
4492 set_bit(R5_LOCKED, &sh->dev[i].flags);
4496 set_bit(STRIPE_INSYNC, &sh->state);
4497 set_bit(STRIPE_REPLACED, &sh->state);
4499 if ((s.syncing || s.replacing) && s.locked == 0 &&
4500 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4501 test_bit(STRIPE_INSYNC, &sh->state)) {
4502 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4503 clear_bit(STRIPE_SYNCING, &sh->state);
4504 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4505 wake_up(&conf->wait_for_overlap);
4508 /* If the failed drives are just a ReadError, then we might need
4509 * to progress the repair/check process
4511 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4512 for (i = 0; i < s.failed; i++) {
4513 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4514 if (test_bit(R5_ReadError, &dev->flags)
4515 && !test_bit(R5_LOCKED, &dev->flags)
4516 && test_bit(R5_UPTODATE, &dev->flags)
4518 if (!test_bit(R5_ReWrite, &dev->flags)) {
4519 set_bit(R5_Wantwrite, &dev->flags);
4520 set_bit(R5_ReWrite, &dev->flags);
4521 set_bit(R5_LOCKED, &dev->flags);
4524 /* let's read it back */
4525 set_bit(R5_Wantread, &dev->flags);
4526 set_bit(R5_LOCKED, &dev->flags);
4532 /* Finish reconstruct operations initiated by the expansion process */
4533 if (sh->reconstruct_state == reconstruct_state_result) {
4534 struct stripe_head *sh_src
4535 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4536 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4537 /* sh cannot be written until sh_src has been read.
4538 * so arrange for sh to be delayed a little
4540 set_bit(STRIPE_DELAYED, &sh->state);
4541 set_bit(STRIPE_HANDLE, &sh->state);
4542 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4544 atomic_inc(&conf->preread_active_stripes);
4545 raid5_release_stripe(sh_src);
4549 raid5_release_stripe(sh_src);
4551 sh->reconstruct_state = reconstruct_state_idle;
4552 clear_bit(STRIPE_EXPANDING, &sh->state);
4553 for (i = conf->raid_disks; i--; ) {
4554 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4555 set_bit(R5_LOCKED, &sh->dev[i].flags);
4560 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4561 !sh->reconstruct_state) {
4562 /* Need to write out all blocks after computing parity */
4563 sh->disks = conf->raid_disks;
4564 stripe_set_idx(sh->sector, conf, 0, sh);
4565 schedule_reconstruction(sh, &s, 1, 1);
4566 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4567 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4568 atomic_dec(&conf->reshape_stripes);
4569 wake_up(&conf->wait_for_overlap);
4570 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4573 if (s.expanding && s.locked == 0 &&
4574 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4575 handle_stripe_expansion(conf, sh);
4578 /* wait for this device to become unblocked */
4579 if (unlikely(s.blocked_rdev)) {
4580 if (conf->mddev->external)
4581 md_wait_for_blocked_rdev(s.blocked_rdev,
4584 /* Internal metadata will immediately
4585 * be written by raid5d, so we don't
4586 * need to wait here.
4588 rdev_dec_pending(s.blocked_rdev,
4592 if (s.handle_bad_blocks)
4593 for (i = disks; i--; ) {
4594 struct md_rdev *rdev;
4595 struct r5dev *dev = &sh->dev[i];
4596 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4597 /* We own a safe reference to the rdev */
4598 rdev = conf->disks[i].rdev;
4599 if (!rdev_set_badblocks(rdev, sh->sector,
4601 md_error(conf->mddev, rdev);
4602 rdev_dec_pending(rdev, conf->mddev);
4604 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4605 rdev = conf->disks[i].rdev;
4606 rdev_clear_badblocks(rdev, sh->sector,
4608 rdev_dec_pending(rdev, conf->mddev);
4610 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4611 rdev = conf->disks[i].replacement;
4613 /* rdev have been moved down */
4614 rdev = conf->disks[i].rdev;
4615 rdev_clear_badblocks(rdev, sh->sector,
4617 rdev_dec_pending(rdev, conf->mddev);
4622 raid_run_ops(sh, s.ops_request);
4626 if (s.dec_preread_active) {
4627 /* We delay this until after ops_run_io so that if make_request
4628 * is waiting on a flush, it won't continue until the writes
4629 * have actually been submitted.
4631 atomic_dec(&conf->preread_active_stripes);
4632 if (atomic_read(&conf->preread_active_stripes) <
4634 md_wakeup_thread(conf->mddev->thread);
4637 if (!bio_list_empty(&s.return_bi)) {
4638 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags)) {
4639 spin_lock_irq(&conf->device_lock);
4640 bio_list_merge(&conf->return_bi, &s.return_bi);
4641 spin_unlock_irq(&conf->device_lock);
4642 md_wakeup_thread(conf->mddev->thread);
4644 return_io(&s.return_bi);
4647 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4650 static void raid5_activate_delayed(struct r5conf *conf)
4652 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4653 while (!list_empty(&conf->delayed_list)) {
4654 struct list_head *l = conf->delayed_list.next;
4655 struct stripe_head *sh;
4656 sh = list_entry(l, struct stripe_head, lru);
4658 clear_bit(STRIPE_DELAYED, &sh->state);
4659 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4660 atomic_inc(&conf->preread_active_stripes);
4661 list_add_tail(&sh->lru, &conf->hold_list);
4662 raid5_wakeup_stripe_thread(sh);
4667 static void activate_bit_delay(struct r5conf *conf,
4668 struct list_head *temp_inactive_list)
4670 /* device_lock is held */
4671 struct list_head head;
4672 list_add(&head, &conf->bitmap_list);
4673 list_del_init(&conf->bitmap_list);
4674 while (!list_empty(&head)) {
4675 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4677 list_del_init(&sh->lru);
4678 atomic_inc(&sh->count);
4679 hash = sh->hash_lock_index;
4680 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4684 static int raid5_congested(struct mddev *mddev, int bits)
4686 struct r5conf *conf = mddev->private;
4688 /* No difference between reads and writes. Just check
4689 * how busy the stripe_cache is
4692 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4696 if (atomic_read(&conf->empty_inactive_list_nr))
4702 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4704 struct r5conf *conf = mddev->private;
4705 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4706 unsigned int chunk_sectors;
4707 unsigned int bio_sectors = bio_sectors(bio);
4709 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4710 return chunk_sectors >=
4711 ((sector & (chunk_sectors - 1)) + bio_sectors);
4715 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4716 * later sampled by raid5d.
4718 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4720 unsigned long flags;
4722 spin_lock_irqsave(&conf->device_lock, flags);
4724 bi->bi_next = conf->retry_read_aligned_list;
4725 conf->retry_read_aligned_list = bi;
4727 spin_unlock_irqrestore(&conf->device_lock, flags);
4728 md_wakeup_thread(conf->mddev->thread);
4731 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4735 bi = conf->retry_read_aligned;
4737 conf->retry_read_aligned = NULL;
4740 bi = conf->retry_read_aligned_list;
4742 conf->retry_read_aligned_list = bi->bi_next;
4745 * this sets the active strip count to 1 and the processed
4746 * strip count to zero (upper 8 bits)
4748 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4755 * The "raid5_align_endio" should check if the read succeeded and if it
4756 * did, call bio_endio on the original bio (having bio_put the new bio
4758 * If the read failed..
4760 static void raid5_align_endio(struct bio *bi)
4762 struct bio* raid_bi = bi->bi_private;
4763 struct mddev *mddev;
4764 struct r5conf *conf;
4765 struct md_rdev *rdev;
4766 int error = bi->bi_error;
4770 rdev = (void*)raid_bi->bi_next;
4771 raid_bi->bi_next = NULL;
4772 mddev = rdev->mddev;
4773 conf = mddev->private;
4775 rdev_dec_pending(rdev, conf->mddev);
4778 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4781 if (atomic_dec_and_test(&conf->active_aligned_reads))
4782 wake_up(&conf->wait_for_quiescent);
4786 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4788 add_bio_to_retry(raid_bi, conf);
4791 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4793 struct r5conf *conf = mddev->private;
4795 struct bio* align_bi;
4796 struct md_rdev *rdev;
4797 sector_t end_sector;
4799 if (!in_chunk_boundary(mddev, raid_bio)) {
4800 pr_debug("%s: non aligned\n", __func__);
4804 * use bio_clone_mddev to make a copy of the bio
4806 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4810 * set bi_end_io to a new function, and set bi_private to the
4813 align_bi->bi_end_io = raid5_align_endio;
4814 align_bi->bi_private = raid_bio;
4818 align_bi->bi_iter.bi_sector =
4819 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4822 end_sector = bio_end_sector(align_bi);
4824 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4825 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4826 rdev->recovery_offset < end_sector) {
4827 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4829 (test_bit(Faulty, &rdev->flags) ||
4830 !(test_bit(In_sync, &rdev->flags) ||
4831 rdev->recovery_offset >= end_sector)))
4838 atomic_inc(&rdev->nr_pending);
4840 raid_bio->bi_next = (void*)rdev;
4841 align_bi->bi_bdev = rdev->bdev;
4842 bio_clear_flag(align_bi, BIO_SEG_VALID);
4844 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4845 bio_sectors(align_bi),
4846 &first_bad, &bad_sectors)) {
4848 rdev_dec_pending(rdev, mddev);
4852 /* No reshape active, so we can trust rdev->data_offset */
4853 align_bi->bi_iter.bi_sector += rdev->data_offset;
4855 spin_lock_irq(&conf->device_lock);
4856 wait_event_lock_irq(conf->wait_for_quiescent,
4859 atomic_inc(&conf->active_aligned_reads);
4860 spin_unlock_irq(&conf->device_lock);
4863 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4864 align_bi, disk_devt(mddev->gendisk),
4865 raid_bio->bi_iter.bi_sector);
4866 generic_make_request(align_bi);
4875 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4880 sector_t sector = raid_bio->bi_iter.bi_sector;
4881 unsigned chunk_sects = mddev->chunk_sectors;
4882 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4884 if (sectors < bio_sectors(raid_bio)) {
4885 split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4886 bio_chain(split, raid_bio);
4890 if (!raid5_read_one_chunk(mddev, split)) {
4891 if (split != raid_bio)
4892 generic_make_request(raid_bio);
4895 } while (split != raid_bio);
4900 /* __get_priority_stripe - get the next stripe to process
4902 * Full stripe writes are allowed to pass preread active stripes up until
4903 * the bypass_threshold is exceeded. In general the bypass_count
4904 * increments when the handle_list is handled before the hold_list; however, it
4905 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4906 * stripe with in flight i/o. The bypass_count will be reset when the
4907 * head of the hold_list has changed, i.e. the head was promoted to the
4910 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4912 struct stripe_head *sh = NULL, *tmp;
4913 struct list_head *handle_list = NULL;
4914 struct r5worker_group *wg = NULL;
4916 if (conf->worker_cnt_per_group == 0) {
4917 handle_list = &conf->handle_list;
4918 } else if (group != ANY_GROUP) {
4919 handle_list = &conf->worker_groups[group].handle_list;
4920 wg = &conf->worker_groups[group];
4923 for (i = 0; i < conf->group_cnt; i++) {
4924 handle_list = &conf->worker_groups[i].handle_list;
4925 wg = &conf->worker_groups[i];
4926 if (!list_empty(handle_list))
4931 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4933 list_empty(handle_list) ? "empty" : "busy",
4934 list_empty(&conf->hold_list) ? "empty" : "busy",
4935 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4937 if (!list_empty(handle_list)) {
4938 sh = list_entry(handle_list->next, typeof(*sh), lru);
4940 if (list_empty(&conf->hold_list))
4941 conf->bypass_count = 0;
4942 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4943 if (conf->hold_list.next == conf->last_hold)
4944 conf->bypass_count++;
4946 conf->last_hold = conf->hold_list.next;
4947 conf->bypass_count -= conf->bypass_threshold;
4948 if (conf->bypass_count < 0)
4949 conf->bypass_count = 0;
4952 } else if (!list_empty(&conf->hold_list) &&
4953 ((conf->bypass_threshold &&
4954 conf->bypass_count > conf->bypass_threshold) ||
4955 atomic_read(&conf->pending_full_writes) == 0)) {
4957 list_for_each_entry(tmp, &conf->hold_list, lru) {
4958 if (conf->worker_cnt_per_group == 0 ||
4959 group == ANY_GROUP ||
4960 !cpu_online(tmp->cpu) ||
4961 cpu_to_group(tmp->cpu) == group) {
4968 conf->bypass_count -= conf->bypass_threshold;
4969 if (conf->bypass_count < 0)
4970 conf->bypass_count = 0;
4982 list_del_init(&sh->lru);
4983 BUG_ON(atomic_inc_return(&sh->count) != 1);
4987 struct raid5_plug_cb {
4988 struct blk_plug_cb cb;
4989 struct list_head list;
4990 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4993 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4995 struct raid5_plug_cb *cb = container_of(
4996 blk_cb, struct raid5_plug_cb, cb);
4997 struct stripe_head *sh;
4998 struct mddev *mddev = cb->cb.data;
4999 struct r5conf *conf = mddev->private;
5003 if (cb->list.next && !list_empty(&cb->list)) {
5004 spin_lock_irq(&conf->device_lock);
5005 while (!list_empty(&cb->list)) {
5006 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5007 list_del_init(&sh->lru);
5009 * avoid race release_stripe_plug() sees
5010 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5011 * is still in our list
5013 smp_mb__before_atomic();
5014 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5016 * STRIPE_ON_RELEASE_LIST could be set here. In that
5017 * case, the count is always > 1 here
5019 hash = sh->hash_lock_index;
5020 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5023 spin_unlock_irq(&conf->device_lock);
5025 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5026 NR_STRIPE_HASH_LOCKS);
5028 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5032 static void release_stripe_plug(struct mddev *mddev,
5033 struct stripe_head *sh)
5035 struct blk_plug_cb *blk_cb = blk_check_plugged(
5036 raid5_unplug, mddev,
5037 sizeof(struct raid5_plug_cb));
5038 struct raid5_plug_cb *cb;
5041 raid5_release_stripe(sh);
5045 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5047 if (cb->list.next == NULL) {
5049 INIT_LIST_HEAD(&cb->list);
5050 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5051 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5054 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5055 list_add_tail(&sh->lru, &cb->list);
5057 raid5_release_stripe(sh);
5060 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5062 struct r5conf *conf = mddev->private;
5063 sector_t logical_sector, last_sector;
5064 struct stripe_head *sh;
5068 if (mddev->reshape_position != MaxSector)
5069 /* Skip discard while reshape is happening */
5072 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5073 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5076 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5078 stripe_sectors = conf->chunk_sectors *
5079 (conf->raid_disks - conf->max_degraded);
5080 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5082 sector_div(last_sector, stripe_sectors);
5084 logical_sector *= conf->chunk_sectors;
5085 last_sector *= conf->chunk_sectors;
5087 for (; logical_sector < last_sector;
5088 logical_sector += STRIPE_SECTORS) {
5092 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5093 prepare_to_wait(&conf->wait_for_overlap, &w,
5094 TASK_UNINTERRUPTIBLE);
5095 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5096 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5097 raid5_release_stripe(sh);
5101 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5102 spin_lock_irq(&sh->stripe_lock);
5103 for (d = 0; d < conf->raid_disks; d++) {
5104 if (d == sh->pd_idx || d == sh->qd_idx)
5106 if (sh->dev[d].towrite || sh->dev[d].toread) {
5107 set_bit(R5_Overlap, &sh->dev[d].flags);
5108 spin_unlock_irq(&sh->stripe_lock);
5109 raid5_release_stripe(sh);
5114 set_bit(STRIPE_DISCARD, &sh->state);
5115 finish_wait(&conf->wait_for_overlap, &w);
5116 sh->overwrite_disks = 0;
5117 for (d = 0; d < conf->raid_disks; d++) {
5118 if (d == sh->pd_idx || d == sh->qd_idx)
5120 sh->dev[d].towrite = bi;
5121 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5122 raid5_inc_bi_active_stripes(bi);
5123 sh->overwrite_disks++;
5125 spin_unlock_irq(&sh->stripe_lock);
5126 if (conf->mddev->bitmap) {
5128 d < conf->raid_disks - conf->max_degraded;
5130 bitmap_startwrite(mddev->bitmap,
5134 sh->bm_seq = conf->seq_flush + 1;
5135 set_bit(STRIPE_BIT_DELAY, &sh->state);
5138 set_bit(STRIPE_HANDLE, &sh->state);
5139 clear_bit(STRIPE_DELAYED, &sh->state);
5140 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5141 atomic_inc(&conf->preread_active_stripes);
5142 release_stripe_plug(mddev, sh);
5145 remaining = raid5_dec_bi_active_stripes(bi);
5146 if (remaining == 0) {
5147 md_write_end(mddev);
5152 static void make_request(struct mddev *mddev, struct bio * bi)
5154 struct r5conf *conf = mddev->private;
5156 sector_t new_sector;
5157 sector_t logical_sector, last_sector;
5158 struct stripe_head *sh;
5159 const int rw = bio_data_dir(bi);
5164 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5165 int ret = r5l_handle_flush_request(conf->log, bi);
5169 if (ret == -ENODEV) {
5170 md_flush_request(mddev, bi);
5173 /* ret == -EAGAIN, fallback */
5176 md_write_start(mddev, bi);
5179 * If array is degraded, better not do chunk aligned read because
5180 * later we might have to read it again in order to reconstruct
5181 * data on failed drives.
5183 if (rw == READ && mddev->degraded == 0 &&
5184 mddev->reshape_position == MaxSector) {
5185 bi = chunk_aligned_read(mddev, bi);
5190 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5191 make_discard_request(mddev, bi);
5195 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5196 last_sector = bio_end_sector(bi);
5198 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5200 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5201 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5207 seq = read_seqcount_begin(&conf->gen_lock);
5210 prepare_to_wait(&conf->wait_for_overlap, &w,
5211 TASK_UNINTERRUPTIBLE);
5212 if (unlikely(conf->reshape_progress != MaxSector)) {
5213 /* spinlock is needed as reshape_progress may be
5214 * 64bit on a 32bit platform, and so it might be
5215 * possible to see a half-updated value
5216 * Of course reshape_progress could change after
5217 * the lock is dropped, so once we get a reference
5218 * to the stripe that we think it is, we will have
5221 spin_lock_irq(&conf->device_lock);
5222 if (mddev->reshape_backwards
5223 ? logical_sector < conf->reshape_progress
5224 : logical_sector >= conf->reshape_progress) {
5227 if (mddev->reshape_backwards
5228 ? logical_sector < conf->reshape_safe
5229 : logical_sector >= conf->reshape_safe) {
5230 spin_unlock_irq(&conf->device_lock);
5236 spin_unlock_irq(&conf->device_lock);
5239 new_sector = raid5_compute_sector(conf, logical_sector,
5242 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5243 (unsigned long long)new_sector,
5244 (unsigned long long)logical_sector);
5246 sh = raid5_get_active_stripe(conf, new_sector, previous,
5247 (bi->bi_rw&RWA_MASK), 0);
5249 if (unlikely(previous)) {
5250 /* expansion might have moved on while waiting for a
5251 * stripe, so we must do the range check again.
5252 * Expansion could still move past after this
5253 * test, but as we are holding a reference to
5254 * 'sh', we know that if that happens,
5255 * STRIPE_EXPANDING will get set and the expansion
5256 * won't proceed until we finish with the stripe.
5259 spin_lock_irq(&conf->device_lock);
5260 if (mddev->reshape_backwards
5261 ? logical_sector >= conf->reshape_progress
5262 : logical_sector < conf->reshape_progress)
5263 /* mismatch, need to try again */
5265 spin_unlock_irq(&conf->device_lock);
5267 raid5_release_stripe(sh);
5273 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5274 /* Might have got the wrong stripe_head
5277 raid5_release_stripe(sh);
5282 logical_sector >= mddev->suspend_lo &&
5283 logical_sector < mddev->suspend_hi) {
5284 raid5_release_stripe(sh);
5285 /* As the suspend_* range is controlled by
5286 * userspace, we want an interruptible
5289 flush_signals(current);
5290 prepare_to_wait(&conf->wait_for_overlap,
5291 &w, TASK_INTERRUPTIBLE);
5292 if (logical_sector >= mddev->suspend_lo &&
5293 logical_sector < mddev->suspend_hi) {
5300 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5301 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5302 /* Stripe is busy expanding or
5303 * add failed due to overlap. Flush everything
5306 md_wakeup_thread(mddev->thread);
5307 raid5_release_stripe(sh);
5312 set_bit(STRIPE_HANDLE, &sh->state);
5313 clear_bit(STRIPE_DELAYED, &sh->state);
5314 if ((!sh->batch_head || sh == sh->batch_head) &&
5315 (bi->bi_rw & REQ_SYNC) &&
5316 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5317 atomic_inc(&conf->preread_active_stripes);
5318 release_stripe_plug(mddev, sh);
5320 /* cannot get stripe for read-ahead, just give-up */
5321 bi->bi_error = -EIO;
5325 finish_wait(&conf->wait_for_overlap, &w);
5327 remaining = raid5_dec_bi_active_stripes(bi);
5328 if (remaining == 0) {
5331 md_write_end(mddev);
5333 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5339 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5341 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5343 /* reshaping is quite different to recovery/resync so it is
5344 * handled quite separately ... here.
5346 * On each call to sync_request, we gather one chunk worth of
5347 * destination stripes and flag them as expanding.
5348 * Then we find all the source stripes and request reads.
5349 * As the reads complete, handle_stripe will copy the data
5350 * into the destination stripe and release that stripe.
5352 struct r5conf *conf = mddev->private;
5353 struct stripe_head *sh;
5354 sector_t first_sector, last_sector;
5355 int raid_disks = conf->previous_raid_disks;
5356 int data_disks = raid_disks - conf->max_degraded;
5357 int new_data_disks = conf->raid_disks - conf->max_degraded;
5360 sector_t writepos, readpos, safepos;
5361 sector_t stripe_addr;
5362 int reshape_sectors;
5363 struct list_head stripes;
5366 if (sector_nr == 0) {
5367 /* If restarting in the middle, skip the initial sectors */
5368 if (mddev->reshape_backwards &&
5369 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5370 sector_nr = raid5_size(mddev, 0, 0)
5371 - conf->reshape_progress;
5372 } else if (mddev->reshape_backwards &&
5373 conf->reshape_progress == MaxSector) {
5374 /* shouldn't happen, but just in case, finish up.*/
5375 sector_nr = MaxSector;
5376 } else if (!mddev->reshape_backwards &&
5377 conf->reshape_progress > 0)
5378 sector_nr = conf->reshape_progress;
5379 sector_div(sector_nr, new_data_disks);
5381 mddev->curr_resync_completed = sector_nr;
5382 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5389 /* We need to process a full chunk at a time.
5390 * If old and new chunk sizes differ, we need to process the
5394 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5396 /* We update the metadata at least every 10 seconds, or when
5397 * the data about to be copied would over-write the source of
5398 * the data at the front of the range. i.e. one new_stripe
5399 * along from reshape_progress new_maps to after where
5400 * reshape_safe old_maps to
5402 writepos = conf->reshape_progress;
5403 sector_div(writepos, new_data_disks);
5404 readpos = conf->reshape_progress;
5405 sector_div(readpos, data_disks);
5406 safepos = conf->reshape_safe;
5407 sector_div(safepos, data_disks);
5408 if (mddev->reshape_backwards) {
5409 BUG_ON(writepos < reshape_sectors);
5410 writepos -= reshape_sectors;
5411 readpos += reshape_sectors;
5412 safepos += reshape_sectors;
5414 writepos += reshape_sectors;
5415 /* readpos and safepos are worst-case calculations.
5416 * A negative number is overly pessimistic, and causes
5417 * obvious problems for unsigned storage. So clip to 0.
5419 readpos -= min_t(sector_t, reshape_sectors, readpos);
5420 safepos -= min_t(sector_t, reshape_sectors, safepos);
5423 /* Having calculated the 'writepos' possibly use it
5424 * to set 'stripe_addr' which is where we will write to.
5426 if (mddev->reshape_backwards) {
5427 BUG_ON(conf->reshape_progress == 0);
5428 stripe_addr = writepos;
5429 BUG_ON((mddev->dev_sectors &
5430 ~((sector_t)reshape_sectors - 1))
5431 - reshape_sectors - stripe_addr
5434 BUG_ON(writepos != sector_nr + reshape_sectors);
5435 stripe_addr = sector_nr;
5438 /* 'writepos' is the most advanced device address we might write.
5439 * 'readpos' is the least advanced device address we might read.
5440 * 'safepos' is the least address recorded in the metadata as having
5442 * If there is a min_offset_diff, these are adjusted either by
5443 * increasing the safepos/readpos if diff is negative, or
5444 * increasing writepos if diff is positive.
5445 * If 'readpos' is then behind 'writepos', there is no way that we can
5446 * ensure safety in the face of a crash - that must be done by userspace
5447 * making a backup of the data. So in that case there is no particular
5448 * rush to update metadata.
5449 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5450 * update the metadata to advance 'safepos' to match 'readpos' so that
5451 * we can be safe in the event of a crash.
5452 * So we insist on updating metadata if safepos is behind writepos and
5453 * readpos is beyond writepos.
5454 * In any case, update the metadata every 10 seconds.
5455 * Maybe that number should be configurable, but I'm not sure it is
5456 * worth it.... maybe it could be a multiple of safemode_delay???
5458 if (conf->min_offset_diff < 0) {
5459 safepos += -conf->min_offset_diff;
5460 readpos += -conf->min_offset_diff;
5462 writepos += conf->min_offset_diff;
5464 if ((mddev->reshape_backwards
5465 ? (safepos > writepos && readpos < writepos)
5466 : (safepos < writepos && readpos > writepos)) ||
5467 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5468 /* Cannot proceed until we've updated the superblock... */
5469 wait_event(conf->wait_for_overlap,
5470 atomic_read(&conf->reshape_stripes)==0
5471 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5472 if (atomic_read(&conf->reshape_stripes) != 0)
5474 mddev->reshape_position = conf->reshape_progress;
5475 mddev->curr_resync_completed = sector_nr;
5476 conf->reshape_checkpoint = jiffies;
5477 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5478 md_wakeup_thread(mddev->thread);
5479 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5480 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5481 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5483 spin_lock_irq(&conf->device_lock);
5484 conf->reshape_safe = mddev->reshape_position;
5485 spin_unlock_irq(&conf->device_lock);
5486 wake_up(&conf->wait_for_overlap);
5487 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5490 INIT_LIST_HEAD(&stripes);
5491 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5493 int skipped_disk = 0;
5494 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5495 set_bit(STRIPE_EXPANDING, &sh->state);
5496 atomic_inc(&conf->reshape_stripes);
5497 /* If any of this stripe is beyond the end of the old
5498 * array, then we need to zero those blocks
5500 for (j=sh->disks; j--;) {
5502 if (j == sh->pd_idx)
5504 if (conf->level == 6 &&
5507 s = raid5_compute_blocknr(sh, j, 0);
5508 if (s < raid5_size(mddev, 0, 0)) {
5512 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5513 set_bit(R5_Expanded, &sh->dev[j].flags);
5514 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5516 if (!skipped_disk) {
5517 set_bit(STRIPE_EXPAND_READY, &sh->state);
5518 set_bit(STRIPE_HANDLE, &sh->state);
5520 list_add(&sh->lru, &stripes);
5522 spin_lock_irq(&conf->device_lock);
5523 if (mddev->reshape_backwards)
5524 conf->reshape_progress -= reshape_sectors * new_data_disks;
5526 conf->reshape_progress += reshape_sectors * new_data_disks;
5527 spin_unlock_irq(&conf->device_lock);
5528 /* Ok, those stripe are ready. We can start scheduling
5529 * reads on the source stripes.
5530 * The source stripes are determined by mapping the first and last
5531 * block on the destination stripes.
5534 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5537 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5538 * new_data_disks - 1),
5540 if (last_sector >= mddev->dev_sectors)
5541 last_sector = mddev->dev_sectors - 1;
5542 while (first_sector <= last_sector) {
5543 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5544 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5545 set_bit(STRIPE_HANDLE, &sh->state);
5546 raid5_release_stripe(sh);
5547 first_sector += STRIPE_SECTORS;
5549 /* Now that the sources are clearly marked, we can release
5550 * the destination stripes
5552 while (!list_empty(&stripes)) {
5553 sh = list_entry(stripes.next, struct stripe_head, lru);
5554 list_del_init(&sh->lru);
5555 raid5_release_stripe(sh);
5557 /* If this takes us to the resync_max point where we have to pause,
5558 * then we need to write out the superblock.
5560 sector_nr += reshape_sectors;
5561 retn = reshape_sectors;
5563 if (mddev->curr_resync_completed > mddev->resync_max ||
5564 (sector_nr - mddev->curr_resync_completed) * 2
5565 >= mddev->resync_max - mddev->curr_resync_completed) {
5566 /* Cannot proceed until we've updated the superblock... */
5567 wait_event(conf->wait_for_overlap,
5568 atomic_read(&conf->reshape_stripes) == 0
5569 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5570 if (atomic_read(&conf->reshape_stripes) != 0)
5572 mddev->reshape_position = conf->reshape_progress;
5573 mddev->curr_resync_completed = sector_nr;
5574 conf->reshape_checkpoint = jiffies;
5575 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5576 md_wakeup_thread(mddev->thread);
5577 wait_event(mddev->sb_wait,
5578 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5579 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5580 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5582 spin_lock_irq(&conf->device_lock);
5583 conf->reshape_safe = mddev->reshape_position;
5584 spin_unlock_irq(&conf->device_lock);
5585 wake_up(&conf->wait_for_overlap);
5586 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5592 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5594 struct r5conf *conf = mddev->private;
5595 struct stripe_head *sh;
5596 sector_t max_sector = mddev->dev_sectors;
5597 sector_t sync_blocks;
5598 int still_degraded = 0;
5601 if (sector_nr >= max_sector) {
5602 /* just being told to finish up .. nothing much to do */
5604 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5609 if (mddev->curr_resync < max_sector) /* aborted */
5610 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5612 else /* completed sync */
5614 bitmap_close_sync(mddev->bitmap);
5619 /* Allow raid5_quiesce to complete */
5620 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5622 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5623 return reshape_request(mddev, sector_nr, skipped);
5625 /* No need to check resync_max as we never do more than one
5626 * stripe, and as resync_max will always be on a chunk boundary,
5627 * if the check in md_do_sync didn't fire, there is no chance
5628 * of overstepping resync_max here
5631 /* if there is too many failed drives and we are trying
5632 * to resync, then assert that we are finished, because there is
5633 * nothing we can do.
5635 if (mddev->degraded >= conf->max_degraded &&
5636 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5637 sector_t rv = mddev->dev_sectors - sector_nr;
5641 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5643 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5644 sync_blocks >= STRIPE_SECTORS) {
5645 /* we can skip this block, and probably more */
5646 sync_blocks /= STRIPE_SECTORS;
5648 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5651 bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5653 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5655 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5656 /* make sure we don't swamp the stripe cache if someone else
5657 * is trying to get access
5659 schedule_timeout_uninterruptible(1);
5661 /* Need to check if array will still be degraded after recovery/resync
5662 * Note in case of > 1 drive failures it's possible we're rebuilding
5663 * one drive while leaving another faulty drive in array.
5666 for (i = 0; i < conf->raid_disks; i++) {
5667 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5669 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5674 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5676 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5677 set_bit(STRIPE_HANDLE, &sh->state);
5679 raid5_release_stripe(sh);
5681 return STRIPE_SECTORS;
5684 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5686 /* We may not be able to submit a whole bio at once as there
5687 * may not be enough stripe_heads available.
5688 * We cannot pre-allocate enough stripe_heads as we may need
5689 * more than exist in the cache (if we allow ever large chunks).
5690 * So we do one stripe head at a time and record in
5691 * ->bi_hw_segments how many have been done.
5693 * We *know* that this entire raid_bio is in one chunk, so
5694 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5696 struct stripe_head *sh;
5698 sector_t sector, logical_sector, last_sector;
5703 logical_sector = raid_bio->bi_iter.bi_sector &
5704 ~((sector_t)STRIPE_SECTORS-1);
5705 sector = raid5_compute_sector(conf, logical_sector,
5707 last_sector = bio_end_sector(raid_bio);
5709 for (; logical_sector < last_sector;
5710 logical_sector += STRIPE_SECTORS,
5711 sector += STRIPE_SECTORS,
5714 if (scnt < raid5_bi_processed_stripes(raid_bio))
5715 /* already done this stripe */
5718 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5721 /* failed to get a stripe - must wait */
5722 raid5_set_bi_processed_stripes(raid_bio, scnt);
5723 conf->retry_read_aligned = raid_bio;
5727 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5728 raid5_release_stripe(sh);
5729 raid5_set_bi_processed_stripes(raid_bio, scnt);
5730 conf->retry_read_aligned = raid_bio;
5734 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5736 raid5_release_stripe(sh);
5739 remaining = raid5_dec_bi_active_stripes(raid_bio);
5740 if (remaining == 0) {
5741 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5743 bio_endio(raid_bio);
5745 if (atomic_dec_and_test(&conf->active_aligned_reads))
5746 wake_up(&conf->wait_for_quiescent);
5750 static int handle_active_stripes(struct r5conf *conf, int group,
5751 struct r5worker *worker,
5752 struct list_head *temp_inactive_list)
5754 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5755 int i, batch_size = 0, hash;
5756 bool release_inactive = false;
5758 while (batch_size < MAX_STRIPE_BATCH &&
5759 (sh = __get_priority_stripe(conf, group)) != NULL)
5760 batch[batch_size++] = sh;
5762 if (batch_size == 0) {
5763 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5764 if (!list_empty(temp_inactive_list + i))
5766 if (i == NR_STRIPE_HASH_LOCKS) {
5767 spin_unlock_irq(&conf->device_lock);
5768 r5l_flush_stripe_to_raid(conf->log);
5769 spin_lock_irq(&conf->device_lock);
5772 release_inactive = true;
5774 spin_unlock_irq(&conf->device_lock);
5776 release_inactive_stripe_list(conf, temp_inactive_list,
5777 NR_STRIPE_HASH_LOCKS);
5779 r5l_flush_stripe_to_raid(conf->log);
5780 if (release_inactive) {
5781 spin_lock_irq(&conf->device_lock);
5785 for (i = 0; i < batch_size; i++)
5786 handle_stripe(batch[i]);
5787 r5l_write_stripe_run(conf->log);
5791 spin_lock_irq(&conf->device_lock);
5792 for (i = 0; i < batch_size; i++) {
5793 hash = batch[i]->hash_lock_index;
5794 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5799 static void raid5_do_work(struct work_struct *work)
5801 struct r5worker *worker = container_of(work, struct r5worker, work);
5802 struct r5worker_group *group = worker->group;
5803 struct r5conf *conf = group->conf;
5804 int group_id = group - conf->worker_groups;
5806 struct blk_plug plug;
5808 pr_debug("+++ raid5worker active\n");
5810 blk_start_plug(&plug);
5812 spin_lock_irq(&conf->device_lock);
5814 int batch_size, released;
5816 released = release_stripe_list(conf, worker->temp_inactive_list);
5818 batch_size = handle_active_stripes(conf, group_id, worker,
5819 worker->temp_inactive_list);
5820 worker->working = false;
5821 if (!batch_size && !released)
5823 handled += batch_size;
5825 pr_debug("%d stripes handled\n", handled);
5827 spin_unlock_irq(&conf->device_lock);
5828 blk_finish_plug(&plug);
5830 pr_debug("--- raid5worker inactive\n");
5834 * This is our raid5 kernel thread.
5836 * We scan the hash table for stripes which can be handled now.
5837 * During the scan, completed stripes are saved for us by the interrupt
5838 * handler, so that they will not have to wait for our next wakeup.
5840 static void raid5d(struct md_thread *thread)
5842 struct mddev *mddev = thread->mddev;
5843 struct r5conf *conf = mddev->private;
5845 struct blk_plug plug;
5847 pr_debug("+++ raid5d active\n");
5849 md_check_recovery(mddev);
5851 if (!bio_list_empty(&conf->return_bi) &&
5852 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5853 struct bio_list tmp = BIO_EMPTY_LIST;
5854 spin_lock_irq(&conf->device_lock);
5855 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5856 bio_list_merge(&tmp, &conf->return_bi);
5857 bio_list_init(&conf->return_bi);
5859 spin_unlock_irq(&conf->device_lock);
5863 blk_start_plug(&plug);
5865 spin_lock_irq(&conf->device_lock);
5868 int batch_size, released;
5870 released = release_stripe_list(conf, conf->temp_inactive_list);
5872 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5875 !list_empty(&conf->bitmap_list)) {
5876 /* Now is a good time to flush some bitmap updates */
5878 spin_unlock_irq(&conf->device_lock);
5879 bitmap_unplug(mddev->bitmap);
5880 spin_lock_irq(&conf->device_lock);
5881 conf->seq_write = conf->seq_flush;
5882 activate_bit_delay(conf, conf->temp_inactive_list);
5884 raid5_activate_delayed(conf);
5886 while ((bio = remove_bio_from_retry(conf))) {
5888 spin_unlock_irq(&conf->device_lock);
5889 ok = retry_aligned_read(conf, bio);
5890 spin_lock_irq(&conf->device_lock);
5896 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5897 conf->temp_inactive_list);
5898 if (!batch_size && !released)
5900 handled += batch_size;
5902 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5903 spin_unlock_irq(&conf->device_lock);
5904 md_check_recovery(mddev);
5905 spin_lock_irq(&conf->device_lock);
5908 pr_debug("%d stripes handled\n", handled);
5910 spin_unlock_irq(&conf->device_lock);
5911 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5912 mutex_trylock(&conf->cache_size_mutex)) {
5913 grow_one_stripe(conf, __GFP_NOWARN);
5914 /* Set flag even if allocation failed. This helps
5915 * slow down allocation requests when mem is short
5917 set_bit(R5_DID_ALLOC, &conf->cache_state);
5918 mutex_unlock(&conf->cache_size_mutex);
5921 r5l_flush_stripe_to_raid(conf->log);
5923 async_tx_issue_pending_all();
5924 blk_finish_plug(&plug);
5926 pr_debug("--- raid5d inactive\n");
5930 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5932 struct r5conf *conf;
5934 spin_lock(&mddev->lock);
5935 conf = mddev->private;
5937 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5938 spin_unlock(&mddev->lock);
5943 raid5_set_cache_size(struct mddev *mddev, int size)
5945 struct r5conf *conf = mddev->private;
5948 if (size <= 16 || size > 32768)
5951 conf->min_nr_stripes = size;
5952 mutex_lock(&conf->cache_size_mutex);
5953 while (size < conf->max_nr_stripes &&
5954 drop_one_stripe(conf))
5956 mutex_unlock(&conf->cache_size_mutex);
5959 err = md_allow_write(mddev);
5963 mutex_lock(&conf->cache_size_mutex);
5964 while (size > conf->max_nr_stripes)
5965 if (!grow_one_stripe(conf, GFP_KERNEL))
5967 mutex_unlock(&conf->cache_size_mutex);
5971 EXPORT_SYMBOL(raid5_set_cache_size);
5974 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5976 struct r5conf *conf;
5980 if (len >= PAGE_SIZE)
5982 if (kstrtoul(page, 10, &new))
5984 err = mddev_lock(mddev);
5987 conf = mddev->private;
5991 err = raid5_set_cache_size(mddev, new);
5992 mddev_unlock(mddev);
5997 static struct md_sysfs_entry
5998 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5999 raid5_show_stripe_cache_size,
6000 raid5_store_stripe_cache_size);
6003 raid5_show_rmw_level(struct mddev *mddev, char *page)
6005 struct r5conf *conf = mddev->private;
6007 return sprintf(page, "%d\n", conf->rmw_level);
6013 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6015 struct r5conf *conf = mddev->private;
6021 if (len >= PAGE_SIZE)
6024 if (kstrtoul(page, 10, &new))
6027 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6030 if (new != PARITY_DISABLE_RMW &&
6031 new != PARITY_ENABLE_RMW &&
6032 new != PARITY_PREFER_RMW)
6035 conf->rmw_level = new;
6039 static struct md_sysfs_entry
6040 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6041 raid5_show_rmw_level,
6042 raid5_store_rmw_level);
6046 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6048 struct r5conf *conf;
6050 spin_lock(&mddev->lock);
6051 conf = mddev->private;
6053 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6054 spin_unlock(&mddev->lock);
6059 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6061 struct r5conf *conf;
6065 if (len >= PAGE_SIZE)
6067 if (kstrtoul(page, 10, &new))
6070 err = mddev_lock(mddev);
6073 conf = mddev->private;
6076 else if (new > conf->min_nr_stripes)
6079 conf->bypass_threshold = new;
6080 mddev_unlock(mddev);
6084 static struct md_sysfs_entry
6085 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6087 raid5_show_preread_threshold,
6088 raid5_store_preread_threshold);
6091 raid5_show_skip_copy(struct mddev *mddev, char *page)
6093 struct r5conf *conf;
6095 spin_lock(&mddev->lock);
6096 conf = mddev->private;
6098 ret = sprintf(page, "%d\n", conf->skip_copy);
6099 spin_unlock(&mddev->lock);
6104 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6106 struct r5conf *conf;
6110 if (len >= PAGE_SIZE)
6112 if (kstrtoul(page, 10, &new))
6116 err = mddev_lock(mddev);
6119 conf = mddev->private;
6122 else if (new != conf->skip_copy) {
6123 mddev_suspend(mddev);
6124 conf->skip_copy = new;
6126 mddev->queue->backing_dev_info.capabilities |=
6127 BDI_CAP_STABLE_WRITES;
6129 mddev->queue->backing_dev_info.capabilities &=
6130 ~BDI_CAP_STABLE_WRITES;
6131 mddev_resume(mddev);
6133 mddev_unlock(mddev);
6137 static struct md_sysfs_entry
6138 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6139 raid5_show_skip_copy,
6140 raid5_store_skip_copy);
6143 stripe_cache_active_show(struct mddev *mddev, char *page)
6145 struct r5conf *conf = mddev->private;
6147 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6152 static struct md_sysfs_entry
6153 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6156 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6158 struct r5conf *conf;
6160 spin_lock(&mddev->lock);
6161 conf = mddev->private;
6163 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6164 spin_unlock(&mddev->lock);
6168 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6170 int *worker_cnt_per_group,
6171 struct r5worker_group **worker_groups);
6173 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6175 struct r5conf *conf;
6178 struct r5worker_group *new_groups, *old_groups;
6179 int group_cnt, worker_cnt_per_group;
6181 if (len >= PAGE_SIZE)
6183 if (kstrtoul(page, 10, &new))
6186 err = mddev_lock(mddev);
6189 conf = mddev->private;
6192 else if (new != conf->worker_cnt_per_group) {
6193 mddev_suspend(mddev);
6195 old_groups = conf->worker_groups;
6197 flush_workqueue(raid5_wq);
6199 err = alloc_thread_groups(conf, new,
6200 &group_cnt, &worker_cnt_per_group,
6203 spin_lock_irq(&conf->device_lock);
6204 conf->group_cnt = group_cnt;
6205 conf->worker_cnt_per_group = worker_cnt_per_group;
6206 conf->worker_groups = new_groups;
6207 spin_unlock_irq(&conf->device_lock);
6210 kfree(old_groups[0].workers);
6213 mddev_resume(mddev);
6215 mddev_unlock(mddev);
6220 static struct md_sysfs_entry
6221 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6222 raid5_show_group_thread_cnt,
6223 raid5_store_group_thread_cnt);
6225 static struct attribute *raid5_attrs[] = {
6226 &raid5_stripecache_size.attr,
6227 &raid5_stripecache_active.attr,
6228 &raid5_preread_bypass_threshold.attr,
6229 &raid5_group_thread_cnt.attr,
6230 &raid5_skip_copy.attr,
6231 &raid5_rmw_level.attr,
6234 static struct attribute_group raid5_attrs_group = {
6236 .attrs = raid5_attrs,
6239 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6241 int *worker_cnt_per_group,
6242 struct r5worker_group **worker_groups)
6246 struct r5worker *workers;
6248 *worker_cnt_per_group = cnt;
6251 *worker_groups = NULL;
6254 *group_cnt = num_possible_nodes();
6255 size = sizeof(struct r5worker) * cnt;
6256 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6257 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6258 *group_cnt, GFP_NOIO);
6259 if (!*worker_groups || !workers) {
6261 kfree(*worker_groups);
6265 for (i = 0; i < *group_cnt; i++) {
6266 struct r5worker_group *group;
6268 group = &(*worker_groups)[i];
6269 INIT_LIST_HEAD(&group->handle_list);
6271 group->workers = workers + i * cnt;
6273 for (j = 0; j < cnt; j++) {
6274 struct r5worker *worker = group->workers + j;
6275 worker->group = group;
6276 INIT_WORK(&worker->work, raid5_do_work);
6278 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6279 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6286 static void free_thread_groups(struct r5conf *conf)
6288 if (conf->worker_groups)
6289 kfree(conf->worker_groups[0].workers);
6290 kfree(conf->worker_groups);
6291 conf->worker_groups = NULL;
6295 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6297 struct r5conf *conf = mddev->private;
6300 sectors = mddev->dev_sectors;
6302 /* size is defined by the smallest of previous and new size */
6303 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6305 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6306 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6307 return sectors * (raid_disks - conf->max_degraded);
6310 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6312 safe_put_page(percpu->spare_page);
6313 if (percpu->scribble)
6314 flex_array_free(percpu->scribble);
6315 percpu->spare_page = NULL;
6316 percpu->scribble = NULL;
6319 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6321 if (conf->level == 6 && !percpu->spare_page)
6322 percpu->spare_page = alloc_page(GFP_KERNEL);
6323 if (!percpu->scribble)
6324 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6325 conf->previous_raid_disks),
6326 max(conf->chunk_sectors,
6327 conf->prev_chunk_sectors)
6331 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6332 free_scratch_buffer(conf, percpu);
6339 static void raid5_free_percpu(struct r5conf *conf)
6346 #ifdef CONFIG_HOTPLUG_CPU
6347 unregister_cpu_notifier(&conf->cpu_notify);
6351 for_each_possible_cpu(cpu)
6352 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6355 free_percpu(conf->percpu);
6358 static void free_conf(struct r5conf *conf)
6361 r5l_exit_log(conf->log);
6362 if (conf->shrinker.seeks)
6363 unregister_shrinker(&conf->shrinker);
6365 free_thread_groups(conf);
6366 shrink_stripes(conf);
6367 raid5_free_percpu(conf);
6369 kfree(conf->stripe_hashtbl);
6373 #ifdef CONFIG_HOTPLUG_CPU
6374 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6377 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6378 long cpu = (long)hcpu;
6379 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6382 case CPU_UP_PREPARE:
6383 case CPU_UP_PREPARE_FROZEN:
6384 if (alloc_scratch_buffer(conf, percpu)) {
6385 pr_err("%s: failed memory allocation for cpu%ld\n",
6387 return notifier_from_errno(-ENOMEM);
6391 case CPU_DEAD_FROZEN:
6392 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6401 static int raid5_alloc_percpu(struct r5conf *conf)
6406 conf->percpu = alloc_percpu(struct raid5_percpu);
6410 #ifdef CONFIG_HOTPLUG_CPU
6411 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6412 conf->cpu_notify.priority = 0;
6413 err = register_cpu_notifier(&conf->cpu_notify);
6419 for_each_present_cpu(cpu) {
6420 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6422 pr_err("%s: failed memory allocation for cpu%ld\n",
6430 conf->scribble_disks = max(conf->raid_disks,
6431 conf->previous_raid_disks);
6432 conf->scribble_sectors = max(conf->chunk_sectors,
6433 conf->prev_chunk_sectors);
6438 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6439 struct shrink_control *sc)
6441 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6442 unsigned long ret = SHRINK_STOP;
6444 if (mutex_trylock(&conf->cache_size_mutex)) {
6446 while (ret < sc->nr_to_scan &&
6447 conf->max_nr_stripes > conf->min_nr_stripes) {
6448 if (drop_one_stripe(conf) == 0) {
6454 mutex_unlock(&conf->cache_size_mutex);
6459 static unsigned long raid5_cache_count(struct shrinker *shrink,
6460 struct shrink_control *sc)
6462 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6464 if (conf->max_nr_stripes < conf->min_nr_stripes)
6465 /* unlikely, but not impossible */
6467 return conf->max_nr_stripes - conf->min_nr_stripes;
6470 static struct r5conf *setup_conf(struct mddev *mddev)
6472 struct r5conf *conf;
6473 int raid_disk, memory, max_disks;
6474 struct md_rdev *rdev;
6475 struct disk_info *disk;
6478 int group_cnt, worker_cnt_per_group;
6479 struct r5worker_group *new_group;
6481 if (mddev->new_level != 5
6482 && mddev->new_level != 4
6483 && mddev->new_level != 6) {
6484 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6485 mdname(mddev), mddev->new_level);
6486 return ERR_PTR(-EIO);
6488 if ((mddev->new_level == 5
6489 && !algorithm_valid_raid5(mddev->new_layout)) ||
6490 (mddev->new_level == 6
6491 && !algorithm_valid_raid6(mddev->new_layout))) {
6492 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6493 mdname(mddev), mddev->new_layout);
6494 return ERR_PTR(-EIO);
6496 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6497 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6498 mdname(mddev), mddev->raid_disks);
6499 return ERR_PTR(-EINVAL);
6502 if (!mddev->new_chunk_sectors ||
6503 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6504 !is_power_of_2(mddev->new_chunk_sectors)) {
6505 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6506 mdname(mddev), mddev->new_chunk_sectors << 9);
6507 return ERR_PTR(-EINVAL);
6510 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6513 /* Don't enable multi-threading by default*/
6514 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6516 conf->group_cnt = group_cnt;
6517 conf->worker_cnt_per_group = worker_cnt_per_group;
6518 conf->worker_groups = new_group;
6521 spin_lock_init(&conf->device_lock);
6522 seqcount_init(&conf->gen_lock);
6523 mutex_init(&conf->cache_size_mutex);
6524 init_waitqueue_head(&conf->wait_for_quiescent);
6525 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
6526 init_waitqueue_head(&conf->wait_for_stripe[i]);
6528 init_waitqueue_head(&conf->wait_for_overlap);
6529 INIT_LIST_HEAD(&conf->handle_list);
6530 INIT_LIST_HEAD(&conf->hold_list);
6531 INIT_LIST_HEAD(&conf->delayed_list);
6532 INIT_LIST_HEAD(&conf->bitmap_list);
6533 bio_list_init(&conf->return_bi);
6534 init_llist_head(&conf->released_stripes);
6535 atomic_set(&conf->active_stripes, 0);
6536 atomic_set(&conf->preread_active_stripes, 0);
6537 atomic_set(&conf->active_aligned_reads, 0);
6538 conf->bypass_threshold = BYPASS_THRESHOLD;
6539 conf->recovery_disabled = mddev->recovery_disabled - 1;
6541 conf->raid_disks = mddev->raid_disks;
6542 if (mddev->reshape_position == MaxSector)
6543 conf->previous_raid_disks = mddev->raid_disks;
6545 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6546 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6548 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6553 conf->mddev = mddev;
6555 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6558 /* We init hash_locks[0] separately to that it can be used
6559 * as the reference lock in the spin_lock_nest_lock() call
6560 * in lock_all_device_hash_locks_irq in order to convince
6561 * lockdep that we know what we are doing.
6563 spin_lock_init(conf->hash_locks);
6564 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6565 spin_lock_init(conf->hash_locks + i);
6567 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6568 INIT_LIST_HEAD(conf->inactive_list + i);
6570 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6571 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6573 conf->level = mddev->new_level;
6574 conf->chunk_sectors = mddev->new_chunk_sectors;
6575 if (raid5_alloc_percpu(conf) != 0)
6578 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6580 rdev_for_each(rdev, mddev) {
6581 raid_disk = rdev->raid_disk;
6582 if (raid_disk >= max_disks
6583 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6585 disk = conf->disks + raid_disk;
6587 if (test_bit(Replacement, &rdev->flags)) {
6588 if (disk->replacement)
6590 disk->replacement = rdev;
6597 if (test_bit(In_sync, &rdev->flags)) {
6598 char b[BDEVNAME_SIZE];
6599 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6601 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6602 } else if (rdev->saved_raid_disk != raid_disk)
6603 /* Cannot rely on bitmap to complete recovery */
6607 conf->level = mddev->new_level;
6608 if (conf->level == 6) {
6609 conf->max_degraded = 2;
6610 if (raid6_call.xor_syndrome)
6611 conf->rmw_level = PARITY_ENABLE_RMW;
6613 conf->rmw_level = PARITY_DISABLE_RMW;
6615 conf->max_degraded = 1;
6616 conf->rmw_level = PARITY_ENABLE_RMW;
6618 conf->algorithm = mddev->new_layout;
6619 conf->reshape_progress = mddev->reshape_position;
6620 if (conf->reshape_progress != MaxSector) {
6621 conf->prev_chunk_sectors = mddev->chunk_sectors;
6622 conf->prev_algo = mddev->layout;
6624 conf->prev_chunk_sectors = conf->chunk_sectors;
6625 conf->prev_algo = conf->algorithm;
6628 conf->min_nr_stripes = NR_STRIPES;
6629 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6630 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6631 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6632 if (grow_stripes(conf, conf->min_nr_stripes)) {
6634 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6635 mdname(mddev), memory);
6638 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6639 mdname(mddev), memory);
6641 * Losing a stripe head costs more than the time to refill it,
6642 * it reduces the queue depth and so can hurt throughput.
6643 * So set it rather large, scaled by number of devices.
6645 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6646 conf->shrinker.scan_objects = raid5_cache_scan;
6647 conf->shrinker.count_objects = raid5_cache_count;
6648 conf->shrinker.batch = 128;
6649 conf->shrinker.flags = 0;
6650 register_shrinker(&conf->shrinker);
6652 sprintf(pers_name, "raid%d", mddev->new_level);
6653 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6654 if (!conf->thread) {
6656 "md/raid:%s: couldn't allocate thread.\n",
6666 return ERR_PTR(-EIO);
6668 return ERR_PTR(-ENOMEM);
6671 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6674 case ALGORITHM_PARITY_0:
6675 if (raid_disk < max_degraded)
6678 case ALGORITHM_PARITY_N:
6679 if (raid_disk >= raid_disks - max_degraded)
6682 case ALGORITHM_PARITY_0_6:
6683 if (raid_disk == 0 ||
6684 raid_disk == raid_disks - 1)
6687 case ALGORITHM_LEFT_ASYMMETRIC_6:
6688 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6689 case ALGORITHM_LEFT_SYMMETRIC_6:
6690 case ALGORITHM_RIGHT_SYMMETRIC_6:
6691 if (raid_disk == raid_disks - 1)
6697 static int run(struct mddev *mddev)
6699 struct r5conf *conf;
6700 int working_disks = 0;
6701 int dirty_parity_disks = 0;
6702 struct md_rdev *rdev;
6703 struct md_rdev *journal_dev = NULL;
6704 sector_t reshape_offset = 0;
6706 long long min_offset_diff = 0;
6709 if (mddev->recovery_cp != MaxSector)
6710 printk(KERN_NOTICE "md/raid:%s: not clean"
6711 " -- starting background reconstruction\n",
6714 rdev_for_each(rdev, mddev) {
6717 if (test_bit(Journal, &rdev->flags)) {
6721 if (rdev->raid_disk < 0)
6723 diff = (rdev->new_data_offset - rdev->data_offset);
6725 min_offset_diff = diff;
6727 } else if (mddev->reshape_backwards &&
6728 diff < min_offset_diff)
6729 min_offset_diff = diff;
6730 else if (!mddev->reshape_backwards &&
6731 diff > min_offset_diff)
6732 min_offset_diff = diff;
6735 if (mddev->reshape_position != MaxSector) {
6736 /* Check that we can continue the reshape.
6737 * Difficulties arise if the stripe we would write to
6738 * next is at or after the stripe we would read from next.
6739 * For a reshape that changes the number of devices, this
6740 * is only possible for a very short time, and mdadm makes
6741 * sure that time appears to have past before assembling
6742 * the array. So we fail if that time hasn't passed.
6743 * For a reshape that keeps the number of devices the same
6744 * mdadm must be monitoring the reshape can keeping the
6745 * critical areas read-only and backed up. It will start
6746 * the array in read-only mode, so we check for that.
6748 sector_t here_new, here_old;
6750 int max_degraded = (mddev->level == 6 ? 2 : 1);
6755 printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6760 if (mddev->new_level != mddev->level) {
6761 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6762 "required - aborting.\n",
6766 old_disks = mddev->raid_disks - mddev->delta_disks;
6767 /* reshape_position must be on a new-stripe boundary, and one
6768 * further up in new geometry must map after here in old
6770 * If the chunk sizes are different, then as we perform reshape
6771 * in units of the largest of the two, reshape_position needs
6772 * be a multiple of the largest chunk size times new data disks.
6774 here_new = mddev->reshape_position;
6775 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6776 new_data_disks = mddev->raid_disks - max_degraded;
6777 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6778 printk(KERN_ERR "md/raid:%s: reshape_position not "
6779 "on a stripe boundary\n", mdname(mddev));
6782 reshape_offset = here_new * chunk_sectors;
6783 /* here_new is the stripe we will write to */
6784 here_old = mddev->reshape_position;
6785 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6786 /* here_old is the first stripe that we might need to read
6788 if (mddev->delta_disks == 0) {
6789 /* We cannot be sure it is safe to start an in-place
6790 * reshape. It is only safe if user-space is monitoring
6791 * and taking constant backups.
6792 * mdadm always starts a situation like this in
6793 * readonly mode so it can take control before
6794 * allowing any writes. So just check for that.
6796 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6797 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6798 /* not really in-place - so OK */;
6799 else if (mddev->ro == 0) {
6800 printk(KERN_ERR "md/raid:%s: in-place reshape "
6801 "must be started in read-only mode "
6806 } else if (mddev->reshape_backwards
6807 ? (here_new * chunk_sectors + min_offset_diff <=
6808 here_old * chunk_sectors)
6809 : (here_new * chunk_sectors >=
6810 here_old * chunk_sectors + (-min_offset_diff))) {
6811 /* Reading from the same stripe as writing to - bad */
6812 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6813 "auto-recovery - aborting.\n",
6817 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6819 /* OK, we should be able to continue; */
6821 BUG_ON(mddev->level != mddev->new_level);
6822 BUG_ON(mddev->layout != mddev->new_layout);
6823 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6824 BUG_ON(mddev->delta_disks != 0);
6827 if (mddev->private == NULL)
6828 conf = setup_conf(mddev);
6830 conf = mddev->private;
6833 return PTR_ERR(conf);
6835 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) && !journal_dev) {
6836 printk(KERN_ERR "md/raid:%s: journal disk is missing, force array readonly\n",
6839 set_disk_ro(mddev->gendisk, 1);
6842 conf->min_offset_diff = min_offset_diff;
6843 mddev->thread = conf->thread;
6844 conf->thread = NULL;
6845 mddev->private = conf;
6847 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6849 rdev = conf->disks[i].rdev;
6850 if (!rdev && conf->disks[i].replacement) {
6851 /* The replacement is all we have yet */
6852 rdev = conf->disks[i].replacement;
6853 conf->disks[i].replacement = NULL;
6854 clear_bit(Replacement, &rdev->flags);
6855 conf->disks[i].rdev = rdev;
6859 if (conf->disks[i].replacement &&
6860 conf->reshape_progress != MaxSector) {
6861 /* replacements and reshape simply do not mix. */
6862 printk(KERN_ERR "md: cannot handle concurrent "
6863 "replacement and reshape.\n");
6866 if (test_bit(In_sync, &rdev->flags)) {
6870 /* This disc is not fully in-sync. However if it
6871 * just stored parity (beyond the recovery_offset),
6872 * when we don't need to be concerned about the
6873 * array being dirty.
6874 * When reshape goes 'backwards', we never have
6875 * partially completed devices, so we only need
6876 * to worry about reshape going forwards.
6878 /* Hack because v0.91 doesn't store recovery_offset properly. */
6879 if (mddev->major_version == 0 &&
6880 mddev->minor_version > 90)
6881 rdev->recovery_offset = reshape_offset;
6883 if (rdev->recovery_offset < reshape_offset) {
6884 /* We need to check old and new layout */
6885 if (!only_parity(rdev->raid_disk,
6888 conf->max_degraded))
6891 if (!only_parity(rdev->raid_disk,
6893 conf->previous_raid_disks,
6894 conf->max_degraded))
6896 dirty_parity_disks++;
6900 * 0 for a fully functional array, 1 or 2 for a degraded array.
6902 mddev->degraded = calc_degraded(conf);
6904 if (has_failed(conf)) {
6905 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6906 " (%d/%d failed)\n",
6907 mdname(mddev), mddev->degraded, conf->raid_disks);
6911 /* device size must be a multiple of chunk size */
6912 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6913 mddev->resync_max_sectors = mddev->dev_sectors;
6915 if (mddev->degraded > dirty_parity_disks &&
6916 mddev->recovery_cp != MaxSector) {
6917 if (mddev->ok_start_degraded)
6919 "md/raid:%s: starting dirty degraded array"
6920 " - data corruption possible.\n",
6924 "md/raid:%s: cannot start dirty degraded array.\n",
6930 if (mddev->degraded == 0)
6931 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6932 " devices, algorithm %d\n", mdname(mddev), conf->level,
6933 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6936 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6937 " out of %d devices, algorithm %d\n",
6938 mdname(mddev), conf->level,
6939 mddev->raid_disks - mddev->degraded,
6940 mddev->raid_disks, mddev->new_layout);
6942 print_raid5_conf(conf);
6944 if (conf->reshape_progress != MaxSector) {
6945 conf->reshape_safe = conf->reshape_progress;
6946 atomic_set(&conf->reshape_stripes, 0);
6947 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6948 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6949 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6950 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6951 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6955 /* Ok, everything is just fine now */
6956 if (mddev->to_remove == &raid5_attrs_group)
6957 mddev->to_remove = NULL;
6958 else if (mddev->kobj.sd &&
6959 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6961 "raid5: failed to create sysfs attributes for %s\n",
6963 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6967 bool discard_supported = true;
6968 /* read-ahead size must cover two whole stripes, which
6969 * is 2 * (datadisks) * chunksize where 'n' is the
6970 * number of raid devices
6972 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6973 int stripe = data_disks *
6974 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6975 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6976 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6978 chunk_size = mddev->chunk_sectors << 9;
6979 blk_queue_io_min(mddev->queue, chunk_size);
6980 blk_queue_io_opt(mddev->queue, chunk_size *
6981 (conf->raid_disks - conf->max_degraded));
6982 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6984 * We can only discard a whole stripe. It doesn't make sense to
6985 * discard data disk but write parity disk
6987 stripe = stripe * PAGE_SIZE;
6988 /* Round up to power of 2, as discard handling
6989 * currently assumes that */
6990 while ((stripe-1) & stripe)
6991 stripe = (stripe | (stripe-1)) + 1;
6992 mddev->queue->limits.discard_alignment = stripe;
6993 mddev->queue->limits.discard_granularity = stripe;
6995 * unaligned part of discard request will be ignored, so can't
6996 * guarantee discard_zeroes_data
6998 mddev->queue->limits.discard_zeroes_data = 0;
7000 blk_queue_max_write_same_sectors(mddev->queue, 0);
7002 rdev_for_each(rdev, mddev) {
7003 disk_stack_limits(mddev->gendisk, rdev->bdev,
7004 rdev->data_offset << 9);
7005 disk_stack_limits(mddev->gendisk, rdev->bdev,
7006 rdev->new_data_offset << 9);
7008 * discard_zeroes_data is required, otherwise data
7009 * could be lost. Consider a scenario: discard a stripe
7010 * (the stripe could be inconsistent if
7011 * discard_zeroes_data is 0); write one disk of the
7012 * stripe (the stripe could be inconsistent again
7013 * depending on which disks are used to calculate
7014 * parity); the disk is broken; The stripe data of this
7017 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7018 !bdev_get_queue(rdev->bdev)->
7019 limits.discard_zeroes_data)
7020 discard_supported = false;
7021 /* Unfortunately, discard_zeroes_data is not currently
7022 * a guarantee - just a hint. So we only allow DISCARD
7023 * if the sysadmin has confirmed that only safe devices
7024 * are in use by setting a module parameter.
7026 if (!devices_handle_discard_safely) {
7027 if (discard_supported) {
7028 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7029 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7031 discard_supported = false;
7035 if (discard_supported &&
7036 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7037 mddev->queue->limits.discard_granularity >= stripe)
7038 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7041 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7046 char b[BDEVNAME_SIZE];
7048 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7049 mdname(mddev), bdevname(journal_dev->bdev, b));
7050 r5l_init_log(conf, journal_dev);
7055 md_unregister_thread(&mddev->thread);
7056 print_raid5_conf(conf);
7058 mddev->private = NULL;
7059 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7063 static void raid5_free(struct mddev *mddev, void *priv)
7065 struct r5conf *conf = priv;
7068 mddev->to_remove = &raid5_attrs_group;
7071 static void status(struct seq_file *seq, struct mddev *mddev)
7073 struct r5conf *conf = mddev->private;
7076 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7077 conf->chunk_sectors / 2, mddev->layout);
7078 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7079 for (i = 0; i < conf->raid_disks; i++)
7080 seq_printf (seq, "%s",
7081 conf->disks[i].rdev &&
7082 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
7083 seq_printf (seq, "]");
7086 static void print_raid5_conf (struct r5conf *conf)
7089 struct disk_info *tmp;
7091 printk(KERN_DEBUG "RAID conf printout:\n");
7093 printk("(conf==NULL)\n");
7096 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7098 conf->raid_disks - conf->mddev->degraded);
7100 for (i = 0; i < conf->raid_disks; i++) {
7101 char b[BDEVNAME_SIZE];
7102 tmp = conf->disks + i;
7104 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7105 i, !test_bit(Faulty, &tmp->rdev->flags),
7106 bdevname(tmp->rdev->bdev, b));
7110 static int raid5_spare_active(struct mddev *mddev)
7113 struct r5conf *conf = mddev->private;
7114 struct disk_info *tmp;
7116 unsigned long flags;
7118 for (i = 0; i < conf->raid_disks; i++) {
7119 tmp = conf->disks + i;
7120 if (tmp->replacement
7121 && tmp->replacement->recovery_offset == MaxSector
7122 && !test_bit(Faulty, &tmp->replacement->flags)
7123 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7124 /* Replacement has just become active. */
7126 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7129 /* Replaced device not technically faulty,
7130 * but we need to be sure it gets removed
7131 * and never re-added.
7133 set_bit(Faulty, &tmp->rdev->flags);
7134 sysfs_notify_dirent_safe(
7135 tmp->rdev->sysfs_state);
7137 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7138 } else if (tmp->rdev
7139 && tmp->rdev->recovery_offset == MaxSector
7140 && !test_bit(Faulty, &tmp->rdev->flags)
7141 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7143 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7146 spin_lock_irqsave(&conf->device_lock, flags);
7147 mddev->degraded = calc_degraded(conf);
7148 spin_unlock_irqrestore(&conf->device_lock, flags);
7149 print_raid5_conf(conf);
7153 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7155 struct r5conf *conf = mddev->private;
7157 int number = rdev->raid_disk;
7158 struct md_rdev **rdevp;
7159 struct disk_info *p = conf->disks + number;
7161 print_raid5_conf(conf);
7162 if (test_bit(Journal, &rdev->flags)) {
7164 * journal disk is not removable, but we need give a chance to
7165 * update superblock of other disks. Otherwise journal disk
7166 * will be considered as 'fresh'
7168 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7171 if (rdev == p->rdev)
7173 else if (rdev == p->replacement)
7174 rdevp = &p->replacement;
7178 if (number >= conf->raid_disks &&
7179 conf->reshape_progress == MaxSector)
7180 clear_bit(In_sync, &rdev->flags);
7182 if (test_bit(In_sync, &rdev->flags) ||
7183 atomic_read(&rdev->nr_pending)) {
7187 /* Only remove non-faulty devices if recovery
7190 if (!test_bit(Faulty, &rdev->flags) &&
7191 mddev->recovery_disabled != conf->recovery_disabled &&
7192 !has_failed(conf) &&
7193 (!p->replacement || p->replacement == rdev) &&
7194 number < conf->raid_disks) {
7200 if (atomic_read(&rdev->nr_pending)) {
7201 /* lost the race, try later */
7204 } else if (p->replacement) {
7205 /* We must have just cleared 'rdev' */
7206 p->rdev = p->replacement;
7207 clear_bit(Replacement, &p->replacement->flags);
7208 smp_mb(); /* Make sure other CPUs may see both as identical
7209 * but will never see neither - if they are careful
7211 p->replacement = NULL;
7212 clear_bit(WantReplacement, &rdev->flags);
7214 /* We might have just removed the Replacement as faulty-
7215 * clear the bit just in case
7217 clear_bit(WantReplacement, &rdev->flags);
7220 print_raid5_conf(conf);
7224 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7226 struct r5conf *conf = mddev->private;
7229 struct disk_info *p;
7231 int last = conf->raid_disks - 1;
7233 if (test_bit(Journal, &rdev->flags))
7235 if (mddev->recovery_disabled == conf->recovery_disabled)
7238 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7239 /* no point adding a device */
7242 if (rdev->raid_disk >= 0)
7243 first = last = rdev->raid_disk;
7246 * find the disk ... but prefer rdev->saved_raid_disk
7249 if (rdev->saved_raid_disk >= 0 &&
7250 rdev->saved_raid_disk >= first &&
7251 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7252 first = rdev->saved_raid_disk;
7254 for (disk = first; disk <= last; disk++) {
7255 p = conf->disks + disk;
7256 if (p->rdev == NULL) {
7257 clear_bit(In_sync, &rdev->flags);
7258 rdev->raid_disk = disk;
7260 if (rdev->saved_raid_disk != disk)
7262 rcu_assign_pointer(p->rdev, rdev);
7266 for (disk = first; disk <= last; disk++) {
7267 p = conf->disks + disk;
7268 if (test_bit(WantReplacement, &p->rdev->flags) &&
7269 p->replacement == NULL) {
7270 clear_bit(In_sync, &rdev->flags);
7271 set_bit(Replacement, &rdev->flags);
7272 rdev->raid_disk = disk;
7275 rcu_assign_pointer(p->replacement, rdev);
7280 print_raid5_conf(conf);
7284 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7286 /* no resync is happening, and there is enough space
7287 * on all devices, so we can resize.
7288 * We need to make sure resync covers any new space.
7289 * If the array is shrinking we should possibly wait until
7290 * any io in the removed space completes, but it hardly seems
7294 struct r5conf *conf = mddev->private;
7298 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7299 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7300 if (mddev->external_size &&
7301 mddev->array_sectors > newsize)
7303 if (mddev->bitmap) {
7304 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7308 md_set_array_sectors(mddev, newsize);
7309 set_capacity(mddev->gendisk, mddev->array_sectors);
7310 revalidate_disk(mddev->gendisk);
7311 if (sectors > mddev->dev_sectors &&
7312 mddev->recovery_cp > mddev->dev_sectors) {
7313 mddev->recovery_cp = mddev->dev_sectors;
7314 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7316 mddev->dev_sectors = sectors;
7317 mddev->resync_max_sectors = sectors;
7321 static int check_stripe_cache(struct mddev *mddev)
7323 /* Can only proceed if there are plenty of stripe_heads.
7324 * We need a minimum of one full stripe,, and for sensible progress
7325 * it is best to have about 4 times that.
7326 * If we require 4 times, then the default 256 4K stripe_heads will
7327 * allow for chunk sizes up to 256K, which is probably OK.
7328 * If the chunk size is greater, user-space should request more
7329 * stripe_heads first.
7331 struct r5conf *conf = mddev->private;
7332 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7333 > conf->min_nr_stripes ||
7334 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7335 > conf->min_nr_stripes) {
7336 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7338 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7345 static int check_reshape(struct mddev *mddev)
7347 struct r5conf *conf = mddev->private;
7351 if (mddev->delta_disks == 0 &&
7352 mddev->new_layout == mddev->layout &&
7353 mddev->new_chunk_sectors == mddev->chunk_sectors)
7354 return 0; /* nothing to do */
7355 if (has_failed(conf))
7357 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7358 /* We might be able to shrink, but the devices must
7359 * be made bigger first.
7360 * For raid6, 4 is the minimum size.
7361 * Otherwise 2 is the minimum
7364 if (mddev->level == 6)
7366 if (mddev->raid_disks + mddev->delta_disks < min)
7370 if (!check_stripe_cache(mddev))
7373 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7374 mddev->delta_disks > 0)
7375 if (resize_chunks(conf,
7376 conf->previous_raid_disks
7377 + max(0, mddev->delta_disks),
7378 max(mddev->new_chunk_sectors,
7379 mddev->chunk_sectors)
7382 return resize_stripes(conf, (conf->previous_raid_disks
7383 + mddev->delta_disks));
7386 static int raid5_start_reshape(struct mddev *mddev)
7388 struct r5conf *conf = mddev->private;
7389 struct md_rdev *rdev;
7391 unsigned long flags;
7393 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7396 if (!check_stripe_cache(mddev))
7399 if (has_failed(conf))
7402 rdev_for_each(rdev, mddev) {
7403 if (!test_bit(In_sync, &rdev->flags)
7404 && !test_bit(Faulty, &rdev->flags))
7408 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7409 /* Not enough devices even to make a degraded array
7414 /* Refuse to reduce size of the array. Any reductions in
7415 * array size must be through explicit setting of array_size
7418 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7419 < mddev->array_sectors) {
7420 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7421 "before number of disks\n", mdname(mddev));
7425 atomic_set(&conf->reshape_stripes, 0);
7426 spin_lock_irq(&conf->device_lock);
7427 write_seqcount_begin(&conf->gen_lock);
7428 conf->previous_raid_disks = conf->raid_disks;
7429 conf->raid_disks += mddev->delta_disks;
7430 conf->prev_chunk_sectors = conf->chunk_sectors;
7431 conf->chunk_sectors = mddev->new_chunk_sectors;
7432 conf->prev_algo = conf->algorithm;
7433 conf->algorithm = mddev->new_layout;
7435 /* Code that selects data_offset needs to see the generation update
7436 * if reshape_progress has been set - so a memory barrier needed.
7439 if (mddev->reshape_backwards)
7440 conf->reshape_progress = raid5_size(mddev, 0, 0);
7442 conf->reshape_progress = 0;
7443 conf->reshape_safe = conf->reshape_progress;
7444 write_seqcount_end(&conf->gen_lock);
7445 spin_unlock_irq(&conf->device_lock);
7447 /* Now make sure any requests that proceeded on the assumption
7448 * the reshape wasn't running - like Discard or Read - have
7451 mddev_suspend(mddev);
7452 mddev_resume(mddev);
7454 /* Add some new drives, as many as will fit.
7455 * We know there are enough to make the newly sized array work.
7456 * Don't add devices if we are reducing the number of
7457 * devices in the array. This is because it is not possible
7458 * to correctly record the "partially reconstructed" state of
7459 * such devices during the reshape and confusion could result.
7461 if (mddev->delta_disks >= 0) {
7462 rdev_for_each(rdev, mddev)
7463 if (rdev->raid_disk < 0 &&
7464 !test_bit(Faulty, &rdev->flags)) {
7465 if (raid5_add_disk(mddev, rdev) == 0) {
7467 >= conf->previous_raid_disks)
7468 set_bit(In_sync, &rdev->flags);
7470 rdev->recovery_offset = 0;
7472 if (sysfs_link_rdev(mddev, rdev))
7473 /* Failure here is OK */;
7475 } else if (rdev->raid_disk >= conf->previous_raid_disks
7476 && !test_bit(Faulty, &rdev->flags)) {
7477 /* This is a spare that was manually added */
7478 set_bit(In_sync, &rdev->flags);
7481 /* When a reshape changes the number of devices,
7482 * ->degraded is measured against the larger of the
7483 * pre and post number of devices.
7485 spin_lock_irqsave(&conf->device_lock, flags);
7486 mddev->degraded = calc_degraded(conf);
7487 spin_unlock_irqrestore(&conf->device_lock, flags);
7489 mddev->raid_disks = conf->raid_disks;
7490 mddev->reshape_position = conf->reshape_progress;
7491 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7493 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7494 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7495 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7496 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7497 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7498 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7500 if (!mddev->sync_thread) {
7501 mddev->recovery = 0;
7502 spin_lock_irq(&conf->device_lock);
7503 write_seqcount_begin(&conf->gen_lock);
7504 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7505 mddev->new_chunk_sectors =
7506 conf->chunk_sectors = conf->prev_chunk_sectors;
7507 mddev->new_layout = conf->algorithm = conf->prev_algo;
7508 rdev_for_each(rdev, mddev)
7509 rdev->new_data_offset = rdev->data_offset;
7511 conf->generation --;
7512 conf->reshape_progress = MaxSector;
7513 mddev->reshape_position = MaxSector;
7514 write_seqcount_end(&conf->gen_lock);
7515 spin_unlock_irq(&conf->device_lock);
7518 conf->reshape_checkpoint = jiffies;
7519 md_wakeup_thread(mddev->sync_thread);
7520 md_new_event(mddev);
7524 /* This is called from the reshape thread and should make any
7525 * changes needed in 'conf'
7527 static void end_reshape(struct r5conf *conf)
7530 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7531 struct md_rdev *rdev;
7533 spin_lock_irq(&conf->device_lock);
7534 conf->previous_raid_disks = conf->raid_disks;
7535 rdev_for_each(rdev, conf->mddev)
7536 rdev->data_offset = rdev->new_data_offset;
7538 conf->reshape_progress = MaxSector;
7539 conf->mddev->reshape_position = MaxSector;
7540 spin_unlock_irq(&conf->device_lock);
7541 wake_up(&conf->wait_for_overlap);
7543 /* read-ahead size must cover two whole stripes, which is
7544 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7546 if (conf->mddev->queue) {
7547 int data_disks = conf->raid_disks - conf->max_degraded;
7548 int stripe = data_disks * ((conf->chunk_sectors << 9)
7550 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7551 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7556 /* This is called from the raid5d thread with mddev_lock held.
7557 * It makes config changes to the device.
7559 static void raid5_finish_reshape(struct mddev *mddev)
7561 struct r5conf *conf = mddev->private;
7563 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7565 if (mddev->delta_disks > 0) {
7566 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7567 set_capacity(mddev->gendisk, mddev->array_sectors);
7568 revalidate_disk(mddev->gendisk);
7571 spin_lock_irq(&conf->device_lock);
7572 mddev->degraded = calc_degraded(conf);
7573 spin_unlock_irq(&conf->device_lock);
7574 for (d = conf->raid_disks ;
7575 d < conf->raid_disks - mddev->delta_disks;
7577 struct md_rdev *rdev = conf->disks[d].rdev;
7579 clear_bit(In_sync, &rdev->flags);
7580 rdev = conf->disks[d].replacement;
7582 clear_bit(In_sync, &rdev->flags);
7585 mddev->layout = conf->algorithm;
7586 mddev->chunk_sectors = conf->chunk_sectors;
7587 mddev->reshape_position = MaxSector;
7588 mddev->delta_disks = 0;
7589 mddev->reshape_backwards = 0;
7593 static void raid5_quiesce(struct mddev *mddev, int state)
7595 struct r5conf *conf = mddev->private;
7598 case 2: /* resume for a suspend */
7599 wake_up(&conf->wait_for_overlap);
7602 case 1: /* stop all writes */
7603 lock_all_device_hash_locks_irq(conf);
7604 /* '2' tells resync/reshape to pause so that all
7605 * active stripes can drain
7608 wait_event_cmd(conf->wait_for_quiescent,
7609 atomic_read(&conf->active_stripes) == 0 &&
7610 atomic_read(&conf->active_aligned_reads) == 0,
7611 unlock_all_device_hash_locks_irq(conf),
7612 lock_all_device_hash_locks_irq(conf));
7614 unlock_all_device_hash_locks_irq(conf);
7615 /* allow reshape to continue */
7616 wake_up(&conf->wait_for_overlap);
7619 case 0: /* re-enable writes */
7620 lock_all_device_hash_locks_irq(conf);
7622 wake_up(&conf->wait_for_quiescent);
7623 wake_up(&conf->wait_for_overlap);
7624 unlock_all_device_hash_locks_irq(conf);
7627 r5l_quiesce(conf->log, state);
7630 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7632 struct r0conf *raid0_conf = mddev->private;
7635 /* for raid0 takeover only one zone is supported */
7636 if (raid0_conf->nr_strip_zones > 1) {
7637 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7639 return ERR_PTR(-EINVAL);
7642 sectors = raid0_conf->strip_zone[0].zone_end;
7643 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7644 mddev->dev_sectors = sectors;
7645 mddev->new_level = level;
7646 mddev->new_layout = ALGORITHM_PARITY_N;
7647 mddev->new_chunk_sectors = mddev->chunk_sectors;
7648 mddev->raid_disks += 1;
7649 mddev->delta_disks = 1;
7650 /* make sure it will be not marked as dirty */
7651 mddev->recovery_cp = MaxSector;
7653 return setup_conf(mddev);
7656 static void *raid5_takeover_raid1(struct mddev *mddev)
7660 if (mddev->raid_disks != 2 ||
7661 mddev->degraded > 1)
7662 return ERR_PTR(-EINVAL);
7664 /* Should check if there are write-behind devices? */
7666 chunksect = 64*2; /* 64K by default */
7668 /* The array must be an exact multiple of chunksize */
7669 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7672 if ((chunksect<<9) < STRIPE_SIZE)
7673 /* array size does not allow a suitable chunk size */
7674 return ERR_PTR(-EINVAL);
7676 mddev->new_level = 5;
7677 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7678 mddev->new_chunk_sectors = chunksect;
7680 return setup_conf(mddev);
7683 static void *raid5_takeover_raid6(struct mddev *mddev)
7687 switch (mddev->layout) {
7688 case ALGORITHM_LEFT_ASYMMETRIC_6:
7689 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7691 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7692 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7694 case ALGORITHM_LEFT_SYMMETRIC_6:
7695 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7697 case ALGORITHM_RIGHT_SYMMETRIC_6:
7698 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7700 case ALGORITHM_PARITY_0_6:
7701 new_layout = ALGORITHM_PARITY_0;
7703 case ALGORITHM_PARITY_N:
7704 new_layout = ALGORITHM_PARITY_N;
7707 return ERR_PTR(-EINVAL);
7709 mddev->new_level = 5;
7710 mddev->new_layout = new_layout;
7711 mddev->delta_disks = -1;
7712 mddev->raid_disks -= 1;
7713 return setup_conf(mddev);
7716 static int raid5_check_reshape(struct mddev *mddev)
7718 /* For a 2-drive array, the layout and chunk size can be changed
7719 * immediately as not restriping is needed.
7720 * For larger arrays we record the new value - after validation
7721 * to be used by a reshape pass.
7723 struct r5conf *conf = mddev->private;
7724 int new_chunk = mddev->new_chunk_sectors;
7726 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7728 if (new_chunk > 0) {
7729 if (!is_power_of_2(new_chunk))
7731 if (new_chunk < (PAGE_SIZE>>9))
7733 if (mddev->array_sectors & (new_chunk-1))
7734 /* not factor of array size */
7738 /* They look valid */
7740 if (mddev->raid_disks == 2) {
7741 /* can make the change immediately */
7742 if (mddev->new_layout >= 0) {
7743 conf->algorithm = mddev->new_layout;
7744 mddev->layout = mddev->new_layout;
7746 if (new_chunk > 0) {
7747 conf->chunk_sectors = new_chunk ;
7748 mddev->chunk_sectors = new_chunk;
7750 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7751 md_wakeup_thread(mddev->thread);
7753 return check_reshape(mddev);
7756 static int raid6_check_reshape(struct mddev *mddev)
7758 int new_chunk = mddev->new_chunk_sectors;
7760 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7762 if (new_chunk > 0) {
7763 if (!is_power_of_2(new_chunk))
7765 if (new_chunk < (PAGE_SIZE >> 9))
7767 if (mddev->array_sectors & (new_chunk-1))
7768 /* not factor of array size */
7772 /* They look valid */
7773 return check_reshape(mddev);
7776 static void *raid5_takeover(struct mddev *mddev)
7778 /* raid5 can take over:
7779 * raid0 - if there is only one strip zone - make it a raid4 layout
7780 * raid1 - if there are two drives. We need to know the chunk size
7781 * raid4 - trivial - just use a raid4 layout.
7782 * raid6 - Providing it is a *_6 layout
7784 if (mddev->level == 0)
7785 return raid45_takeover_raid0(mddev, 5);
7786 if (mddev->level == 1)
7787 return raid5_takeover_raid1(mddev);
7788 if (mddev->level == 4) {
7789 mddev->new_layout = ALGORITHM_PARITY_N;
7790 mddev->new_level = 5;
7791 return setup_conf(mddev);
7793 if (mddev->level == 6)
7794 return raid5_takeover_raid6(mddev);
7796 return ERR_PTR(-EINVAL);
7799 static void *raid4_takeover(struct mddev *mddev)
7801 /* raid4 can take over:
7802 * raid0 - if there is only one strip zone
7803 * raid5 - if layout is right
7805 if (mddev->level == 0)
7806 return raid45_takeover_raid0(mddev, 4);
7807 if (mddev->level == 5 &&
7808 mddev->layout == ALGORITHM_PARITY_N) {
7809 mddev->new_layout = 0;
7810 mddev->new_level = 4;
7811 return setup_conf(mddev);
7813 return ERR_PTR(-EINVAL);
7816 static struct md_personality raid5_personality;
7818 static void *raid6_takeover(struct mddev *mddev)
7820 /* Currently can only take over a raid5. We map the
7821 * personality to an equivalent raid6 personality
7822 * with the Q block at the end.
7826 if (mddev->pers != &raid5_personality)
7827 return ERR_PTR(-EINVAL);
7828 if (mddev->degraded > 1)
7829 return ERR_PTR(-EINVAL);
7830 if (mddev->raid_disks > 253)
7831 return ERR_PTR(-EINVAL);
7832 if (mddev->raid_disks < 3)
7833 return ERR_PTR(-EINVAL);
7835 switch (mddev->layout) {
7836 case ALGORITHM_LEFT_ASYMMETRIC:
7837 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7839 case ALGORITHM_RIGHT_ASYMMETRIC:
7840 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7842 case ALGORITHM_LEFT_SYMMETRIC:
7843 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7845 case ALGORITHM_RIGHT_SYMMETRIC:
7846 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7848 case ALGORITHM_PARITY_0:
7849 new_layout = ALGORITHM_PARITY_0_6;
7851 case ALGORITHM_PARITY_N:
7852 new_layout = ALGORITHM_PARITY_N;
7855 return ERR_PTR(-EINVAL);
7857 mddev->new_level = 6;
7858 mddev->new_layout = new_layout;
7859 mddev->delta_disks = 1;
7860 mddev->raid_disks += 1;
7861 return setup_conf(mddev);
7864 static struct md_personality raid6_personality =
7868 .owner = THIS_MODULE,
7869 .make_request = make_request,
7873 .error_handler = error,
7874 .hot_add_disk = raid5_add_disk,
7875 .hot_remove_disk= raid5_remove_disk,
7876 .spare_active = raid5_spare_active,
7877 .sync_request = sync_request,
7878 .resize = raid5_resize,
7880 .check_reshape = raid6_check_reshape,
7881 .start_reshape = raid5_start_reshape,
7882 .finish_reshape = raid5_finish_reshape,
7883 .quiesce = raid5_quiesce,
7884 .takeover = raid6_takeover,
7885 .congested = raid5_congested,
7887 static struct md_personality raid5_personality =
7891 .owner = THIS_MODULE,
7892 .make_request = make_request,
7896 .error_handler = error,
7897 .hot_add_disk = raid5_add_disk,
7898 .hot_remove_disk= raid5_remove_disk,
7899 .spare_active = raid5_spare_active,
7900 .sync_request = sync_request,
7901 .resize = raid5_resize,
7903 .check_reshape = raid5_check_reshape,
7904 .start_reshape = raid5_start_reshape,
7905 .finish_reshape = raid5_finish_reshape,
7906 .quiesce = raid5_quiesce,
7907 .takeover = raid5_takeover,
7908 .congested = raid5_congested,
7911 static struct md_personality raid4_personality =
7915 .owner = THIS_MODULE,
7916 .make_request = make_request,
7920 .error_handler = error,
7921 .hot_add_disk = raid5_add_disk,
7922 .hot_remove_disk= raid5_remove_disk,
7923 .spare_active = raid5_spare_active,
7924 .sync_request = sync_request,
7925 .resize = raid5_resize,
7927 .check_reshape = raid5_check_reshape,
7928 .start_reshape = raid5_start_reshape,
7929 .finish_reshape = raid5_finish_reshape,
7930 .quiesce = raid5_quiesce,
7931 .takeover = raid4_takeover,
7932 .congested = raid5_congested,
7935 static int __init raid5_init(void)
7937 raid5_wq = alloc_workqueue("raid5wq",
7938 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7941 register_md_personality(&raid6_personality);
7942 register_md_personality(&raid5_personality);
7943 register_md_personality(&raid4_personality);
7947 static void raid5_exit(void)
7949 unregister_md_personality(&raid6_personality);
7950 unregister_md_personality(&raid5_personality);
7951 unregister_md_personality(&raid4_personality);
7952 destroy_workqueue(raid5_wq);
7955 module_init(raid5_init);
7956 module_exit(raid5_exit);
7957 MODULE_LICENSE("GPL");
7958 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7959 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7960 MODULE_ALIAS("md-raid5");
7961 MODULE_ALIAS("md-raid4");
7962 MODULE_ALIAS("md-level-5");
7963 MODULE_ALIAS("md-level-4");
7964 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7965 MODULE_ALIAS("md-raid6");
7966 MODULE_ALIAS("md-level-6");
7968 /* This used to be two separate modules, they were: */
7969 MODULE_ALIAS("raid5");
7970 MODULE_ALIAS("raid6");