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)
2094 mddev_suspend(conf->mddev);
2096 for_each_present_cpu(cpu) {
2097 struct raid5_percpu *percpu;
2098 struct flex_array *scribble;
2100 percpu = per_cpu_ptr(conf->percpu, cpu);
2101 scribble = scribble_alloc(new_disks,
2102 new_sectors / STRIPE_SECTORS,
2106 flex_array_free(percpu->scribble);
2107 percpu->scribble = scribble;
2114 mddev_resume(conf->mddev);
2118 static int resize_stripes(struct r5conf *conf, int newsize)
2120 /* Make all the stripes able to hold 'newsize' devices.
2121 * New slots in each stripe get 'page' set to a new page.
2123 * This happens in stages:
2124 * 1/ create a new kmem_cache and allocate the required number of
2126 * 2/ gather all the old stripe_heads and transfer the pages across
2127 * to the new stripe_heads. This will have the side effect of
2128 * freezing the array as once all stripe_heads have been collected,
2129 * no IO will be possible. Old stripe heads are freed once their
2130 * pages have been transferred over, and the old kmem_cache is
2131 * freed when all stripes are done.
2132 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2133 * we simple return a failre status - no need to clean anything up.
2134 * 4/ allocate new pages for the new slots in the new stripe_heads.
2135 * If this fails, we don't bother trying the shrink the
2136 * stripe_heads down again, we just leave them as they are.
2137 * As each stripe_head is processed the new one is released into
2140 * Once step2 is started, we cannot afford to wait for a write,
2141 * so we use GFP_NOIO allocations.
2143 struct stripe_head *osh, *nsh;
2144 LIST_HEAD(newstripes);
2145 struct disk_info *ndisks;
2147 struct kmem_cache *sc;
2151 if (newsize <= conf->pool_size)
2152 return 0; /* never bother to shrink */
2154 err = md_allow_write(conf->mddev);
2159 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2160 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2165 /* Need to ensure auto-resizing doesn't interfere */
2166 mutex_lock(&conf->cache_size_mutex);
2168 for (i = conf->max_nr_stripes; i; i--) {
2169 nsh = alloc_stripe(sc, GFP_KERNEL);
2173 nsh->raid_conf = conf;
2174 list_add(&nsh->lru, &newstripes);
2177 /* didn't get enough, give up */
2178 while (!list_empty(&newstripes)) {
2179 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2180 list_del(&nsh->lru);
2181 kmem_cache_free(sc, nsh);
2183 kmem_cache_destroy(sc);
2184 mutex_unlock(&conf->cache_size_mutex);
2187 /* Step 2 - Must use GFP_NOIO now.
2188 * OK, we have enough stripes, start collecting inactive
2189 * stripes and copying them over
2193 list_for_each_entry(nsh, &newstripes, lru) {
2194 lock_device_hash_lock(conf, hash);
2195 wait_event_exclusive_cmd(conf->wait_for_stripe[hash],
2196 !list_empty(conf->inactive_list + hash),
2197 unlock_device_hash_lock(conf, hash),
2198 lock_device_hash_lock(conf, hash));
2199 osh = get_free_stripe(conf, hash);
2200 unlock_device_hash_lock(conf, hash);
2202 for(i=0; i<conf->pool_size; i++) {
2203 nsh->dev[i].page = osh->dev[i].page;
2204 nsh->dev[i].orig_page = osh->dev[i].page;
2206 nsh->hash_lock_index = hash;
2207 kmem_cache_free(conf->slab_cache, osh);
2209 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2210 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2215 kmem_cache_destroy(conf->slab_cache);
2218 * At this point, we are holding all the stripes so the array
2219 * is completely stalled, so now is a good time to resize
2220 * conf->disks and the scribble region
2222 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2224 for (i=0; i<conf->raid_disks; i++)
2225 ndisks[i] = conf->disks[i];
2227 conf->disks = ndisks;
2231 mutex_unlock(&conf->cache_size_mutex);
2232 /* Step 4, return new stripes to service */
2233 while(!list_empty(&newstripes)) {
2234 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2235 list_del_init(&nsh->lru);
2237 for (i=conf->raid_disks; i < newsize; i++)
2238 if (nsh->dev[i].page == NULL) {
2239 struct page *p = alloc_page(GFP_NOIO);
2240 nsh->dev[i].page = p;
2241 nsh->dev[i].orig_page = p;
2245 raid5_release_stripe(nsh);
2247 /* critical section pass, GFP_NOIO no longer needed */
2249 conf->slab_cache = sc;
2250 conf->active_name = 1-conf->active_name;
2252 conf->pool_size = newsize;
2256 static int drop_one_stripe(struct r5conf *conf)
2258 struct stripe_head *sh;
2259 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2261 spin_lock_irq(conf->hash_locks + hash);
2262 sh = get_free_stripe(conf, hash);
2263 spin_unlock_irq(conf->hash_locks + hash);
2266 BUG_ON(atomic_read(&sh->count));
2268 kmem_cache_free(conf->slab_cache, sh);
2269 atomic_dec(&conf->active_stripes);
2270 conf->max_nr_stripes--;
2274 static void shrink_stripes(struct r5conf *conf)
2276 while (conf->max_nr_stripes &&
2277 drop_one_stripe(conf))
2280 kmem_cache_destroy(conf->slab_cache);
2281 conf->slab_cache = NULL;
2284 static void raid5_end_read_request(struct bio * bi)
2286 struct stripe_head *sh = bi->bi_private;
2287 struct r5conf *conf = sh->raid_conf;
2288 int disks = sh->disks, i;
2289 char b[BDEVNAME_SIZE];
2290 struct md_rdev *rdev = NULL;
2293 for (i=0 ; i<disks; i++)
2294 if (bi == &sh->dev[i].req)
2297 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2298 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2304 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2305 /* If replacement finished while this request was outstanding,
2306 * 'replacement' might be NULL already.
2307 * In that case it moved down to 'rdev'.
2308 * rdev is not removed until all requests are finished.
2310 rdev = conf->disks[i].replacement;
2312 rdev = conf->disks[i].rdev;
2314 if (use_new_offset(conf, sh))
2315 s = sh->sector + rdev->new_data_offset;
2317 s = sh->sector + rdev->data_offset;
2318 if (!bi->bi_error) {
2319 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2320 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2321 /* Note that this cannot happen on a
2322 * replacement device. We just fail those on
2327 "md/raid:%s: read error corrected"
2328 " (%lu sectors at %llu on %s)\n",
2329 mdname(conf->mddev), STRIPE_SECTORS,
2330 (unsigned long long)s,
2331 bdevname(rdev->bdev, b));
2332 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2333 clear_bit(R5_ReadError, &sh->dev[i].flags);
2334 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2335 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2336 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2338 if (atomic_read(&rdev->read_errors))
2339 atomic_set(&rdev->read_errors, 0);
2341 const char *bdn = bdevname(rdev->bdev, b);
2345 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2346 atomic_inc(&rdev->read_errors);
2347 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2350 "md/raid:%s: read error on replacement device "
2351 "(sector %llu on %s).\n",
2352 mdname(conf->mddev),
2353 (unsigned long long)s,
2355 else if (conf->mddev->degraded >= conf->max_degraded) {
2359 "md/raid:%s: read error not correctable "
2360 "(sector %llu on %s).\n",
2361 mdname(conf->mddev),
2362 (unsigned long long)s,
2364 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2369 "md/raid:%s: read error NOT corrected!! "
2370 "(sector %llu on %s).\n",
2371 mdname(conf->mddev),
2372 (unsigned long long)s,
2374 } else if (atomic_read(&rdev->read_errors)
2375 > conf->max_nr_stripes)
2377 "md/raid:%s: Too many read errors, failing device %s.\n",
2378 mdname(conf->mddev), bdn);
2381 if (set_bad && test_bit(In_sync, &rdev->flags)
2382 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2385 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2386 set_bit(R5_ReadError, &sh->dev[i].flags);
2387 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2389 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2391 clear_bit(R5_ReadError, &sh->dev[i].flags);
2392 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2394 && test_bit(In_sync, &rdev->flags)
2395 && rdev_set_badblocks(
2396 rdev, sh->sector, STRIPE_SECTORS, 0)))
2397 md_error(conf->mddev, rdev);
2400 rdev_dec_pending(rdev, conf->mddev);
2401 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2402 set_bit(STRIPE_HANDLE, &sh->state);
2403 raid5_release_stripe(sh);
2406 static void raid5_end_write_request(struct bio *bi)
2408 struct stripe_head *sh = bi->bi_private;
2409 struct r5conf *conf = sh->raid_conf;
2410 int disks = sh->disks, i;
2411 struct md_rdev *uninitialized_var(rdev);
2414 int replacement = 0;
2416 for (i = 0 ; i < disks; i++) {
2417 if (bi == &sh->dev[i].req) {
2418 rdev = conf->disks[i].rdev;
2421 if (bi == &sh->dev[i].rreq) {
2422 rdev = conf->disks[i].replacement;
2426 /* rdev was removed and 'replacement'
2427 * replaced it. rdev is not removed
2428 * until all requests are finished.
2430 rdev = conf->disks[i].rdev;
2434 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2435 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2444 md_error(conf->mddev, rdev);
2445 else if (is_badblock(rdev, sh->sector,
2447 &first_bad, &bad_sectors))
2448 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2451 set_bit(STRIPE_DEGRADED, &sh->state);
2452 set_bit(WriteErrorSeen, &rdev->flags);
2453 set_bit(R5_WriteError, &sh->dev[i].flags);
2454 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2455 set_bit(MD_RECOVERY_NEEDED,
2456 &rdev->mddev->recovery);
2457 } else if (is_badblock(rdev, sh->sector,
2459 &first_bad, &bad_sectors)) {
2460 set_bit(R5_MadeGood, &sh->dev[i].flags);
2461 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2462 /* That was a successful write so make
2463 * sure it looks like we already did
2466 set_bit(R5_ReWrite, &sh->dev[i].flags);
2469 rdev_dec_pending(rdev, conf->mddev);
2471 if (sh->batch_head && bi->bi_error && !replacement)
2472 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2474 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2475 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2476 set_bit(STRIPE_HANDLE, &sh->state);
2477 raid5_release_stripe(sh);
2479 if (sh->batch_head && sh != sh->batch_head)
2480 raid5_release_stripe(sh->batch_head);
2483 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2485 struct r5dev *dev = &sh->dev[i];
2487 bio_init(&dev->req);
2488 dev->req.bi_io_vec = &dev->vec;
2489 dev->req.bi_max_vecs = 1;
2490 dev->req.bi_private = sh;
2492 bio_init(&dev->rreq);
2493 dev->rreq.bi_io_vec = &dev->rvec;
2494 dev->rreq.bi_max_vecs = 1;
2495 dev->rreq.bi_private = sh;
2498 dev->sector = raid5_compute_blocknr(sh, i, previous);
2501 static void error(struct mddev *mddev, struct md_rdev *rdev)
2503 char b[BDEVNAME_SIZE];
2504 struct r5conf *conf = mddev->private;
2505 unsigned long flags;
2506 pr_debug("raid456: error called\n");
2508 spin_lock_irqsave(&conf->device_lock, flags);
2509 clear_bit(In_sync, &rdev->flags);
2510 mddev->degraded = calc_degraded(conf);
2511 spin_unlock_irqrestore(&conf->device_lock, flags);
2512 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2514 set_bit(Blocked, &rdev->flags);
2515 set_bit(Faulty, &rdev->flags);
2516 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2517 set_bit(MD_CHANGE_PENDING, &mddev->flags);
2519 "md/raid:%s: Disk failure on %s, disabling device.\n"
2520 "md/raid:%s: Operation continuing on %d devices.\n",
2522 bdevname(rdev->bdev, b),
2524 conf->raid_disks - mddev->degraded);
2528 * Input: a 'big' sector number,
2529 * Output: index of the data and parity disk, and the sector # in them.
2531 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2532 int previous, int *dd_idx,
2533 struct stripe_head *sh)
2535 sector_t stripe, stripe2;
2536 sector_t chunk_number;
2537 unsigned int chunk_offset;
2540 sector_t new_sector;
2541 int algorithm = previous ? conf->prev_algo
2543 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2544 : conf->chunk_sectors;
2545 int raid_disks = previous ? conf->previous_raid_disks
2547 int data_disks = raid_disks - conf->max_degraded;
2549 /* First compute the information on this sector */
2552 * Compute the chunk number and the sector offset inside the chunk
2554 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2555 chunk_number = r_sector;
2558 * Compute the stripe number
2560 stripe = chunk_number;
2561 *dd_idx = sector_div(stripe, data_disks);
2564 * Select the parity disk based on the user selected algorithm.
2566 pd_idx = qd_idx = -1;
2567 switch(conf->level) {
2569 pd_idx = data_disks;
2572 switch (algorithm) {
2573 case ALGORITHM_LEFT_ASYMMETRIC:
2574 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2575 if (*dd_idx >= pd_idx)
2578 case ALGORITHM_RIGHT_ASYMMETRIC:
2579 pd_idx = sector_div(stripe2, raid_disks);
2580 if (*dd_idx >= pd_idx)
2583 case ALGORITHM_LEFT_SYMMETRIC:
2584 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2585 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2587 case ALGORITHM_RIGHT_SYMMETRIC:
2588 pd_idx = sector_div(stripe2, raid_disks);
2589 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2591 case ALGORITHM_PARITY_0:
2595 case ALGORITHM_PARITY_N:
2596 pd_idx = data_disks;
2604 switch (algorithm) {
2605 case ALGORITHM_LEFT_ASYMMETRIC:
2606 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2607 qd_idx = pd_idx + 1;
2608 if (pd_idx == raid_disks-1) {
2609 (*dd_idx)++; /* Q D D D P */
2611 } else if (*dd_idx >= pd_idx)
2612 (*dd_idx) += 2; /* D D P Q D */
2614 case ALGORITHM_RIGHT_ASYMMETRIC:
2615 pd_idx = sector_div(stripe2, raid_disks);
2616 qd_idx = pd_idx + 1;
2617 if (pd_idx == raid_disks-1) {
2618 (*dd_idx)++; /* Q D D D P */
2620 } else if (*dd_idx >= pd_idx)
2621 (*dd_idx) += 2; /* D D P Q D */
2623 case ALGORITHM_LEFT_SYMMETRIC:
2624 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2625 qd_idx = (pd_idx + 1) % raid_disks;
2626 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2628 case ALGORITHM_RIGHT_SYMMETRIC:
2629 pd_idx = sector_div(stripe2, raid_disks);
2630 qd_idx = (pd_idx + 1) % raid_disks;
2631 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2634 case ALGORITHM_PARITY_0:
2639 case ALGORITHM_PARITY_N:
2640 pd_idx = data_disks;
2641 qd_idx = data_disks + 1;
2644 case ALGORITHM_ROTATING_ZERO_RESTART:
2645 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2646 * of blocks for computing Q is different.
2648 pd_idx = sector_div(stripe2, raid_disks);
2649 qd_idx = pd_idx + 1;
2650 if (pd_idx == raid_disks-1) {
2651 (*dd_idx)++; /* Q D D D P */
2653 } else if (*dd_idx >= pd_idx)
2654 (*dd_idx) += 2; /* D D P Q D */
2658 case ALGORITHM_ROTATING_N_RESTART:
2659 /* Same a left_asymmetric, by first stripe is
2660 * D D D P Q rather than
2664 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2665 qd_idx = pd_idx + 1;
2666 if (pd_idx == raid_disks-1) {
2667 (*dd_idx)++; /* Q D D D P */
2669 } else if (*dd_idx >= pd_idx)
2670 (*dd_idx) += 2; /* D D P Q D */
2674 case ALGORITHM_ROTATING_N_CONTINUE:
2675 /* Same as left_symmetric but Q is before P */
2676 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2677 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2678 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2682 case ALGORITHM_LEFT_ASYMMETRIC_6:
2683 /* RAID5 left_asymmetric, with Q on last device */
2684 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2685 if (*dd_idx >= pd_idx)
2687 qd_idx = raid_disks - 1;
2690 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2691 pd_idx = sector_div(stripe2, raid_disks-1);
2692 if (*dd_idx >= pd_idx)
2694 qd_idx = raid_disks - 1;
2697 case ALGORITHM_LEFT_SYMMETRIC_6:
2698 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2699 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2700 qd_idx = raid_disks - 1;
2703 case ALGORITHM_RIGHT_SYMMETRIC_6:
2704 pd_idx = sector_div(stripe2, raid_disks-1);
2705 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2706 qd_idx = raid_disks - 1;
2709 case ALGORITHM_PARITY_0_6:
2712 qd_idx = raid_disks - 1;
2722 sh->pd_idx = pd_idx;
2723 sh->qd_idx = qd_idx;
2724 sh->ddf_layout = ddf_layout;
2727 * Finally, compute the new sector number
2729 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2733 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2735 struct r5conf *conf = sh->raid_conf;
2736 int raid_disks = sh->disks;
2737 int data_disks = raid_disks - conf->max_degraded;
2738 sector_t new_sector = sh->sector, check;
2739 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2740 : conf->chunk_sectors;
2741 int algorithm = previous ? conf->prev_algo
2745 sector_t chunk_number;
2746 int dummy1, dd_idx = i;
2748 struct stripe_head sh2;
2750 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2751 stripe = new_sector;
2753 if (i == sh->pd_idx)
2755 switch(conf->level) {
2758 switch (algorithm) {
2759 case ALGORITHM_LEFT_ASYMMETRIC:
2760 case ALGORITHM_RIGHT_ASYMMETRIC:
2764 case ALGORITHM_LEFT_SYMMETRIC:
2765 case ALGORITHM_RIGHT_SYMMETRIC:
2768 i -= (sh->pd_idx + 1);
2770 case ALGORITHM_PARITY_0:
2773 case ALGORITHM_PARITY_N:
2780 if (i == sh->qd_idx)
2781 return 0; /* It is the Q disk */
2782 switch (algorithm) {
2783 case ALGORITHM_LEFT_ASYMMETRIC:
2784 case ALGORITHM_RIGHT_ASYMMETRIC:
2785 case ALGORITHM_ROTATING_ZERO_RESTART:
2786 case ALGORITHM_ROTATING_N_RESTART:
2787 if (sh->pd_idx == raid_disks-1)
2788 i--; /* Q D D D P */
2789 else if (i > sh->pd_idx)
2790 i -= 2; /* D D P Q D */
2792 case ALGORITHM_LEFT_SYMMETRIC:
2793 case ALGORITHM_RIGHT_SYMMETRIC:
2794 if (sh->pd_idx == raid_disks-1)
2795 i--; /* Q D D D P */
2800 i -= (sh->pd_idx + 2);
2803 case ALGORITHM_PARITY_0:
2806 case ALGORITHM_PARITY_N:
2808 case ALGORITHM_ROTATING_N_CONTINUE:
2809 /* Like left_symmetric, but P is before Q */
2810 if (sh->pd_idx == 0)
2811 i--; /* P D D D Q */
2816 i -= (sh->pd_idx + 1);
2819 case ALGORITHM_LEFT_ASYMMETRIC_6:
2820 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2824 case ALGORITHM_LEFT_SYMMETRIC_6:
2825 case ALGORITHM_RIGHT_SYMMETRIC_6:
2827 i += data_disks + 1;
2828 i -= (sh->pd_idx + 1);
2830 case ALGORITHM_PARITY_0_6:
2839 chunk_number = stripe * data_disks + i;
2840 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2842 check = raid5_compute_sector(conf, r_sector,
2843 previous, &dummy1, &sh2);
2844 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2845 || sh2.qd_idx != sh->qd_idx) {
2846 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2847 mdname(conf->mddev));
2854 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2855 int rcw, int expand)
2857 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2858 struct r5conf *conf = sh->raid_conf;
2859 int level = conf->level;
2863 for (i = disks; i--; ) {
2864 struct r5dev *dev = &sh->dev[i];
2867 set_bit(R5_LOCKED, &dev->flags);
2868 set_bit(R5_Wantdrain, &dev->flags);
2870 clear_bit(R5_UPTODATE, &dev->flags);
2874 /* if we are not expanding this is a proper write request, and
2875 * there will be bios with new data to be drained into the
2880 /* False alarm, nothing to do */
2882 sh->reconstruct_state = reconstruct_state_drain_run;
2883 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2885 sh->reconstruct_state = reconstruct_state_run;
2887 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2889 if (s->locked + conf->max_degraded == disks)
2890 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2891 atomic_inc(&conf->pending_full_writes);
2893 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2894 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2895 BUG_ON(level == 6 &&
2896 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2897 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2899 for (i = disks; i--; ) {
2900 struct r5dev *dev = &sh->dev[i];
2901 if (i == pd_idx || i == qd_idx)
2905 (test_bit(R5_UPTODATE, &dev->flags) ||
2906 test_bit(R5_Wantcompute, &dev->flags))) {
2907 set_bit(R5_Wantdrain, &dev->flags);
2908 set_bit(R5_LOCKED, &dev->flags);
2909 clear_bit(R5_UPTODATE, &dev->flags);
2914 /* False alarm - nothing to do */
2916 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2917 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2918 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2919 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2922 /* keep the parity disk(s) locked while asynchronous operations
2925 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2926 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2930 int qd_idx = sh->qd_idx;
2931 struct r5dev *dev = &sh->dev[qd_idx];
2933 set_bit(R5_LOCKED, &dev->flags);
2934 clear_bit(R5_UPTODATE, &dev->flags);
2938 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2939 __func__, (unsigned long long)sh->sector,
2940 s->locked, s->ops_request);
2944 * Each stripe/dev can have one or more bion attached.
2945 * toread/towrite point to the first in a chain.
2946 * The bi_next chain must be in order.
2948 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2949 int forwrite, int previous)
2952 struct r5conf *conf = sh->raid_conf;
2955 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2956 (unsigned long long)bi->bi_iter.bi_sector,
2957 (unsigned long long)sh->sector);
2960 * If several bio share a stripe. The bio bi_phys_segments acts as a
2961 * reference count to avoid race. The reference count should already be
2962 * increased before this function is called (for example, in
2963 * make_request()), so other bio sharing this stripe will not free the
2964 * stripe. If a stripe is owned by one stripe, the stripe lock will
2967 spin_lock_irq(&sh->stripe_lock);
2968 /* Don't allow new IO added to stripes in batch list */
2972 bip = &sh->dev[dd_idx].towrite;
2976 bip = &sh->dev[dd_idx].toread;
2977 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2978 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2980 bip = & (*bip)->bi_next;
2982 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2985 if (!forwrite || previous)
2986 clear_bit(STRIPE_BATCH_READY, &sh->state);
2988 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2992 raid5_inc_bi_active_stripes(bi);
2995 /* check if page is covered */
2996 sector_t sector = sh->dev[dd_idx].sector;
2997 for (bi=sh->dev[dd_idx].towrite;
2998 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2999 bi && bi->bi_iter.bi_sector <= sector;
3000 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3001 if (bio_end_sector(bi) >= sector)
3002 sector = bio_end_sector(bi);
3004 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3005 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3006 sh->overwrite_disks++;
3009 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3010 (unsigned long long)(*bip)->bi_iter.bi_sector,
3011 (unsigned long long)sh->sector, dd_idx);
3013 if (conf->mddev->bitmap && firstwrite) {
3014 /* Cannot hold spinlock over bitmap_startwrite,
3015 * but must ensure this isn't added to a batch until
3016 * we have added to the bitmap and set bm_seq.
3017 * So set STRIPE_BITMAP_PENDING to prevent
3019 * If multiple add_stripe_bio() calls race here they
3020 * much all set STRIPE_BITMAP_PENDING. So only the first one
3021 * to complete "bitmap_startwrite" gets to set
3022 * STRIPE_BIT_DELAY. This is important as once a stripe
3023 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3026 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3027 spin_unlock_irq(&sh->stripe_lock);
3028 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3030 spin_lock_irq(&sh->stripe_lock);
3031 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3032 if (!sh->batch_head) {
3033 sh->bm_seq = conf->seq_flush+1;
3034 set_bit(STRIPE_BIT_DELAY, &sh->state);
3037 spin_unlock_irq(&sh->stripe_lock);
3039 if (stripe_can_batch(sh))
3040 stripe_add_to_batch_list(conf, sh);
3044 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3045 spin_unlock_irq(&sh->stripe_lock);
3049 static void end_reshape(struct r5conf *conf);
3051 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3052 struct stripe_head *sh)
3054 int sectors_per_chunk =
3055 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3057 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3058 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3060 raid5_compute_sector(conf,
3061 stripe * (disks - conf->max_degraded)
3062 *sectors_per_chunk + chunk_offset,
3068 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3069 struct stripe_head_state *s, int disks,
3070 struct bio_list *return_bi)
3073 BUG_ON(sh->batch_head);
3074 for (i = disks; i--; ) {
3078 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3079 struct md_rdev *rdev;
3081 rdev = rcu_dereference(conf->disks[i].rdev);
3082 if (rdev && test_bit(In_sync, &rdev->flags))
3083 atomic_inc(&rdev->nr_pending);
3088 if (!rdev_set_badblocks(
3092 md_error(conf->mddev, rdev);
3093 rdev_dec_pending(rdev, conf->mddev);
3096 spin_lock_irq(&sh->stripe_lock);
3097 /* fail all writes first */
3098 bi = sh->dev[i].towrite;
3099 sh->dev[i].towrite = NULL;
3100 sh->overwrite_disks = 0;
3101 spin_unlock_irq(&sh->stripe_lock);
3105 r5l_stripe_write_finished(sh);
3107 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3108 wake_up(&conf->wait_for_overlap);
3110 while (bi && bi->bi_iter.bi_sector <
3111 sh->dev[i].sector + STRIPE_SECTORS) {
3112 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3114 bi->bi_error = -EIO;
3115 if (!raid5_dec_bi_active_stripes(bi)) {
3116 md_write_end(conf->mddev);
3117 bio_list_add(return_bi, bi);
3122 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3123 STRIPE_SECTORS, 0, 0);
3125 /* and fail all 'written' */
3126 bi = sh->dev[i].written;
3127 sh->dev[i].written = NULL;
3128 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3129 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3130 sh->dev[i].page = sh->dev[i].orig_page;
3133 if (bi) bitmap_end = 1;
3134 while (bi && bi->bi_iter.bi_sector <
3135 sh->dev[i].sector + STRIPE_SECTORS) {
3136 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3138 bi->bi_error = -EIO;
3139 if (!raid5_dec_bi_active_stripes(bi)) {
3140 md_write_end(conf->mddev);
3141 bio_list_add(return_bi, bi);
3146 /* fail any reads if this device is non-operational and
3147 * the data has not reached the cache yet.
3149 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3150 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3151 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3152 spin_lock_irq(&sh->stripe_lock);
3153 bi = sh->dev[i].toread;
3154 sh->dev[i].toread = NULL;
3155 spin_unlock_irq(&sh->stripe_lock);
3156 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3157 wake_up(&conf->wait_for_overlap);
3160 while (bi && bi->bi_iter.bi_sector <
3161 sh->dev[i].sector + STRIPE_SECTORS) {
3162 struct bio *nextbi =
3163 r5_next_bio(bi, sh->dev[i].sector);
3165 bi->bi_error = -EIO;
3166 if (!raid5_dec_bi_active_stripes(bi))
3167 bio_list_add(return_bi, bi);
3172 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3173 STRIPE_SECTORS, 0, 0);
3174 /* If we were in the middle of a write the parity block might
3175 * still be locked - so just clear all R5_LOCKED flags
3177 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3182 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3183 if (atomic_dec_and_test(&conf->pending_full_writes))
3184 md_wakeup_thread(conf->mddev->thread);
3188 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3189 struct stripe_head_state *s)
3194 BUG_ON(sh->batch_head);
3195 clear_bit(STRIPE_SYNCING, &sh->state);
3196 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3197 wake_up(&conf->wait_for_overlap);
3200 /* There is nothing more to do for sync/check/repair.
3201 * Don't even need to abort as that is handled elsewhere
3202 * if needed, and not always wanted e.g. if there is a known
3204 * For recover/replace we need to record a bad block on all
3205 * non-sync devices, or abort the recovery
3207 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3208 /* During recovery devices cannot be removed, so
3209 * locking and refcounting of rdevs is not needed
3211 for (i = 0; i < conf->raid_disks; i++) {
3212 struct md_rdev *rdev = conf->disks[i].rdev;
3214 && !test_bit(Faulty, &rdev->flags)
3215 && !test_bit(In_sync, &rdev->flags)
3216 && !rdev_set_badblocks(rdev, sh->sector,
3219 rdev = conf->disks[i].replacement;
3221 && !test_bit(Faulty, &rdev->flags)
3222 && !test_bit(In_sync, &rdev->flags)
3223 && !rdev_set_badblocks(rdev, sh->sector,
3228 conf->recovery_disabled =
3229 conf->mddev->recovery_disabled;
3231 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3234 static int want_replace(struct stripe_head *sh, int disk_idx)
3236 struct md_rdev *rdev;
3238 /* Doing recovery so rcu locking not required */
3239 rdev = sh->raid_conf->disks[disk_idx].replacement;
3241 && !test_bit(Faulty, &rdev->flags)
3242 && !test_bit(In_sync, &rdev->flags)
3243 && (rdev->recovery_offset <= sh->sector
3244 || rdev->mddev->recovery_cp <= sh->sector))
3250 /* fetch_block - checks the given member device to see if its data needs
3251 * to be read or computed to satisfy a request.
3253 * Returns 1 when no more member devices need to be checked, otherwise returns
3254 * 0 to tell the loop in handle_stripe_fill to continue
3257 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3258 int disk_idx, int disks)
3260 struct r5dev *dev = &sh->dev[disk_idx];
3261 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3262 &sh->dev[s->failed_num[1]] };
3266 if (test_bit(R5_LOCKED, &dev->flags) ||
3267 test_bit(R5_UPTODATE, &dev->flags))
3268 /* No point reading this as we already have it or have
3269 * decided to get it.
3274 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3275 /* We need this block to directly satisfy a request */
3278 if (s->syncing || s->expanding ||
3279 (s->replacing && want_replace(sh, disk_idx)))
3280 /* When syncing, or expanding we read everything.
3281 * When replacing, we need the replaced block.
3285 if ((s->failed >= 1 && fdev[0]->toread) ||
3286 (s->failed >= 2 && fdev[1]->toread))
3287 /* If we want to read from a failed device, then
3288 * we need to actually read every other device.
3292 /* Sometimes neither read-modify-write nor reconstruct-write
3293 * cycles can work. In those cases we read every block we
3294 * can. Then the parity-update is certain to have enough to
3296 * This can only be a problem when we need to write something,
3297 * and some device has failed. If either of those tests
3298 * fail we need look no further.
3300 if (!s->failed || !s->to_write)
3303 if (test_bit(R5_Insync, &dev->flags) &&
3304 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3305 /* Pre-reads at not permitted until after short delay
3306 * to gather multiple requests. However if this
3307 * device is no Insync, the block could only be be computed
3308 * and there is no need to delay that.
3312 for (i = 0; i < s->failed && i < 2; i++) {
3313 if (fdev[i]->towrite &&
3314 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3315 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3316 /* If we have a partial write to a failed
3317 * device, then we will need to reconstruct
3318 * the content of that device, so all other
3319 * devices must be read.
3324 /* If we are forced to do a reconstruct-write, either because
3325 * the current RAID6 implementation only supports that, or
3326 * or because parity cannot be trusted and we are currently
3327 * recovering it, there is extra need to be careful.
3328 * If one of the devices that we would need to read, because
3329 * it is not being overwritten (and maybe not written at all)
3330 * is missing/faulty, then we need to read everything we can.
3332 if (sh->raid_conf->level != 6 &&
3333 sh->sector < sh->raid_conf->mddev->recovery_cp)
3334 /* reconstruct-write isn't being forced */
3336 for (i = 0; i < s->failed && i < 2; i++) {
3337 if (s->failed_num[i] != sh->pd_idx &&
3338 s->failed_num[i] != sh->qd_idx &&
3339 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3340 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3347 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3348 int disk_idx, int disks)
3350 struct r5dev *dev = &sh->dev[disk_idx];
3352 /* is the data in this block needed, and can we get it? */
3353 if (need_this_block(sh, s, disk_idx, disks)) {
3354 /* we would like to get this block, possibly by computing it,
3355 * otherwise read it if the backing disk is insync
3357 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3358 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3359 BUG_ON(sh->batch_head);
3360 if ((s->uptodate == disks - 1) &&
3361 (s->failed && (disk_idx == s->failed_num[0] ||
3362 disk_idx == s->failed_num[1]))) {
3363 /* have disk failed, and we're requested to fetch it;
3366 pr_debug("Computing stripe %llu block %d\n",
3367 (unsigned long long)sh->sector, disk_idx);
3368 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3369 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3370 set_bit(R5_Wantcompute, &dev->flags);
3371 sh->ops.target = disk_idx;
3372 sh->ops.target2 = -1; /* no 2nd target */
3374 /* Careful: from this point on 'uptodate' is in the eye
3375 * of raid_run_ops which services 'compute' operations
3376 * before writes. R5_Wantcompute flags a block that will
3377 * be R5_UPTODATE by the time it is needed for a
3378 * subsequent operation.
3382 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3383 /* Computing 2-failure is *very* expensive; only
3384 * do it if failed >= 2
3387 for (other = disks; other--; ) {
3388 if (other == disk_idx)
3390 if (!test_bit(R5_UPTODATE,
3391 &sh->dev[other].flags))
3395 pr_debug("Computing stripe %llu blocks %d,%d\n",
3396 (unsigned long long)sh->sector,
3398 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3399 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3400 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3401 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3402 sh->ops.target = disk_idx;
3403 sh->ops.target2 = other;
3407 } else if (test_bit(R5_Insync, &dev->flags)) {
3408 set_bit(R5_LOCKED, &dev->flags);
3409 set_bit(R5_Wantread, &dev->flags);
3411 pr_debug("Reading block %d (sync=%d)\n",
3412 disk_idx, s->syncing);
3420 * handle_stripe_fill - read or compute data to satisfy pending requests.
3422 static void handle_stripe_fill(struct stripe_head *sh,
3423 struct stripe_head_state *s,
3428 /* look for blocks to read/compute, skip this if a compute
3429 * is already in flight, or if the stripe contents are in the
3430 * midst of changing due to a write
3432 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3433 !sh->reconstruct_state)
3434 for (i = disks; i--; )
3435 if (fetch_block(sh, s, i, disks))
3437 set_bit(STRIPE_HANDLE, &sh->state);
3440 static void break_stripe_batch_list(struct stripe_head *head_sh,
3441 unsigned long handle_flags);
3442 /* handle_stripe_clean_event
3443 * any written block on an uptodate or failed drive can be returned.
3444 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3445 * never LOCKED, so we don't need to test 'failed' directly.
3447 static void handle_stripe_clean_event(struct r5conf *conf,
3448 struct stripe_head *sh, int disks, struct bio_list *return_bi)
3452 int discard_pending = 0;
3453 struct stripe_head *head_sh = sh;
3454 bool do_endio = false;
3456 for (i = disks; i--; )
3457 if (sh->dev[i].written) {
3459 if (!test_bit(R5_LOCKED, &dev->flags) &&
3460 (test_bit(R5_UPTODATE, &dev->flags) ||
3461 test_bit(R5_Discard, &dev->flags) ||
3462 test_bit(R5_SkipCopy, &dev->flags))) {
3463 /* We can return any write requests */
3464 struct bio *wbi, *wbi2;
3465 pr_debug("Return write for disc %d\n", i);
3466 if (test_and_clear_bit(R5_Discard, &dev->flags))
3467 clear_bit(R5_UPTODATE, &dev->flags);
3468 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3469 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3474 dev->page = dev->orig_page;
3476 dev->written = NULL;
3477 while (wbi && wbi->bi_iter.bi_sector <
3478 dev->sector + STRIPE_SECTORS) {
3479 wbi2 = r5_next_bio(wbi, dev->sector);
3480 if (!raid5_dec_bi_active_stripes(wbi)) {
3481 md_write_end(conf->mddev);
3482 bio_list_add(return_bi, wbi);
3486 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3488 !test_bit(STRIPE_DEGRADED, &sh->state),
3490 if (head_sh->batch_head) {
3491 sh = list_first_entry(&sh->batch_list,
3494 if (sh != head_sh) {
3501 } else if (test_bit(R5_Discard, &dev->flags))
3502 discard_pending = 1;
3503 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3504 WARN_ON(dev->page != dev->orig_page);
3507 r5l_stripe_write_finished(sh);
3509 if (!discard_pending &&
3510 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3511 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3512 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3513 if (sh->qd_idx >= 0) {
3514 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3515 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3517 /* now that discard is done we can proceed with any sync */
3518 clear_bit(STRIPE_DISCARD, &sh->state);
3520 * SCSI discard will change some bio fields and the stripe has
3521 * no updated data, so remove it from hash list and the stripe
3522 * will be reinitialized
3524 spin_lock_irq(&conf->device_lock);
3527 if (head_sh->batch_head) {
3528 sh = list_first_entry(&sh->batch_list,
3529 struct stripe_head, batch_list);
3533 spin_unlock_irq(&conf->device_lock);
3536 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3537 set_bit(STRIPE_HANDLE, &sh->state);
3541 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3542 if (atomic_dec_and_test(&conf->pending_full_writes))
3543 md_wakeup_thread(conf->mddev->thread);
3545 if (head_sh->batch_head && do_endio)
3546 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3549 static void handle_stripe_dirtying(struct r5conf *conf,
3550 struct stripe_head *sh,
3551 struct stripe_head_state *s,
3554 int rmw = 0, rcw = 0, i;
3555 sector_t recovery_cp = conf->mddev->recovery_cp;
3557 /* Check whether resync is now happening or should start.
3558 * If yes, then the array is dirty (after unclean shutdown or
3559 * initial creation), so parity in some stripes might be inconsistent.
3560 * In this case, we need to always do reconstruct-write, to ensure
3561 * that in case of drive failure or read-error correction, we
3562 * generate correct data from the parity.
3564 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3565 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3567 /* Calculate the real rcw later - for now make it
3568 * look like rcw is cheaper
3571 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3572 conf->rmw_level, (unsigned long long)recovery_cp,
3573 (unsigned long long)sh->sector);
3574 } else for (i = disks; i--; ) {
3575 /* would I have to read this buffer for read_modify_write */
3576 struct r5dev *dev = &sh->dev[i];
3577 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3578 !test_bit(R5_LOCKED, &dev->flags) &&
3579 !(test_bit(R5_UPTODATE, &dev->flags) ||
3580 test_bit(R5_Wantcompute, &dev->flags))) {
3581 if (test_bit(R5_Insync, &dev->flags))
3584 rmw += 2*disks; /* cannot read it */
3586 /* Would I have to read this buffer for reconstruct_write */
3587 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3588 i != sh->pd_idx && i != sh->qd_idx &&
3589 !test_bit(R5_LOCKED, &dev->flags) &&
3590 !(test_bit(R5_UPTODATE, &dev->flags) ||
3591 test_bit(R5_Wantcompute, &dev->flags))) {
3592 if (test_bit(R5_Insync, &dev->flags))
3598 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3599 (unsigned long long)sh->sector, rmw, rcw);
3600 set_bit(STRIPE_HANDLE, &sh->state);
3601 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3602 /* prefer read-modify-write, but need to get some data */
3603 if (conf->mddev->queue)
3604 blk_add_trace_msg(conf->mddev->queue,
3605 "raid5 rmw %llu %d",
3606 (unsigned long long)sh->sector, rmw);
3607 for (i = disks; i--; ) {
3608 struct r5dev *dev = &sh->dev[i];
3609 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3610 !test_bit(R5_LOCKED, &dev->flags) &&
3611 !(test_bit(R5_UPTODATE, &dev->flags) ||
3612 test_bit(R5_Wantcompute, &dev->flags)) &&
3613 test_bit(R5_Insync, &dev->flags)) {
3614 if (test_bit(STRIPE_PREREAD_ACTIVE,
3616 pr_debug("Read_old block %d for r-m-w\n",
3618 set_bit(R5_LOCKED, &dev->flags);
3619 set_bit(R5_Wantread, &dev->flags);
3622 set_bit(STRIPE_DELAYED, &sh->state);
3623 set_bit(STRIPE_HANDLE, &sh->state);
3628 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3629 /* want reconstruct write, but need to get some data */
3632 for (i = disks; i--; ) {
3633 struct r5dev *dev = &sh->dev[i];
3634 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3635 i != sh->pd_idx && i != sh->qd_idx &&
3636 !test_bit(R5_LOCKED, &dev->flags) &&
3637 !(test_bit(R5_UPTODATE, &dev->flags) ||
3638 test_bit(R5_Wantcompute, &dev->flags))) {
3640 if (test_bit(R5_Insync, &dev->flags) &&
3641 test_bit(STRIPE_PREREAD_ACTIVE,
3643 pr_debug("Read_old block "
3644 "%d for Reconstruct\n", i);
3645 set_bit(R5_LOCKED, &dev->flags);
3646 set_bit(R5_Wantread, &dev->flags);
3650 set_bit(STRIPE_DELAYED, &sh->state);
3651 set_bit(STRIPE_HANDLE, &sh->state);
3655 if (rcw && conf->mddev->queue)
3656 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3657 (unsigned long long)sh->sector,
3658 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3661 if (rcw > disks && rmw > disks &&
3662 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3663 set_bit(STRIPE_DELAYED, &sh->state);
3665 /* now if nothing is locked, and if we have enough data,
3666 * we can start a write request
3668 /* since handle_stripe can be called at any time we need to handle the
3669 * case where a compute block operation has been submitted and then a
3670 * subsequent call wants to start a write request. raid_run_ops only
3671 * handles the case where compute block and reconstruct are requested
3672 * simultaneously. If this is not the case then new writes need to be
3673 * held off until the compute completes.
3675 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3676 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3677 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3678 schedule_reconstruction(sh, s, rcw == 0, 0);
3681 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3682 struct stripe_head_state *s, int disks)
3684 struct r5dev *dev = NULL;
3686 BUG_ON(sh->batch_head);
3687 set_bit(STRIPE_HANDLE, &sh->state);
3689 switch (sh->check_state) {
3690 case check_state_idle:
3691 /* start a new check operation if there are no failures */
3692 if (s->failed == 0) {
3693 BUG_ON(s->uptodate != disks);
3694 sh->check_state = check_state_run;
3695 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3696 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3700 dev = &sh->dev[s->failed_num[0]];
3702 case check_state_compute_result:
3703 sh->check_state = check_state_idle;
3705 dev = &sh->dev[sh->pd_idx];
3707 /* check that a write has not made the stripe insync */
3708 if (test_bit(STRIPE_INSYNC, &sh->state))
3711 /* either failed parity check, or recovery is happening */
3712 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3713 BUG_ON(s->uptodate != disks);
3715 set_bit(R5_LOCKED, &dev->flags);
3717 set_bit(R5_Wantwrite, &dev->flags);
3719 clear_bit(STRIPE_DEGRADED, &sh->state);
3720 set_bit(STRIPE_INSYNC, &sh->state);
3722 case check_state_run:
3723 break; /* we will be called again upon completion */
3724 case check_state_check_result:
3725 sh->check_state = check_state_idle;
3727 /* if a failure occurred during the check operation, leave
3728 * STRIPE_INSYNC not set and let the stripe be handled again
3733 /* handle a successful check operation, if parity is correct
3734 * we are done. Otherwise update the mismatch count and repair
3735 * parity if !MD_RECOVERY_CHECK
3737 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3738 /* parity is correct (on disc,
3739 * not in buffer any more)
3741 set_bit(STRIPE_INSYNC, &sh->state);
3743 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3744 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3745 /* don't try to repair!! */
3746 set_bit(STRIPE_INSYNC, &sh->state);
3748 sh->check_state = check_state_compute_run;
3749 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3750 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3751 set_bit(R5_Wantcompute,
3752 &sh->dev[sh->pd_idx].flags);
3753 sh->ops.target = sh->pd_idx;
3754 sh->ops.target2 = -1;
3759 case check_state_compute_run:
3762 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3763 __func__, sh->check_state,
3764 (unsigned long long) sh->sector);
3769 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3770 struct stripe_head_state *s,
3773 int pd_idx = sh->pd_idx;
3774 int qd_idx = sh->qd_idx;
3777 BUG_ON(sh->batch_head);
3778 set_bit(STRIPE_HANDLE, &sh->state);
3780 BUG_ON(s->failed > 2);
3782 /* Want to check and possibly repair P and Q.
3783 * However there could be one 'failed' device, in which
3784 * case we can only check one of them, possibly using the
3785 * other to generate missing data
3788 switch (sh->check_state) {
3789 case check_state_idle:
3790 /* start a new check operation if there are < 2 failures */
3791 if (s->failed == s->q_failed) {
3792 /* The only possible failed device holds Q, so it
3793 * makes sense to check P (If anything else were failed,
3794 * we would have used P to recreate it).
3796 sh->check_state = check_state_run;
3798 if (!s->q_failed && s->failed < 2) {
3799 /* Q is not failed, and we didn't use it to generate
3800 * anything, so it makes sense to check it
3802 if (sh->check_state == check_state_run)
3803 sh->check_state = check_state_run_pq;
3805 sh->check_state = check_state_run_q;
3808 /* discard potentially stale zero_sum_result */
3809 sh->ops.zero_sum_result = 0;
3811 if (sh->check_state == check_state_run) {
3812 /* async_xor_zero_sum destroys the contents of P */
3813 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3816 if (sh->check_state >= check_state_run &&
3817 sh->check_state <= check_state_run_pq) {
3818 /* async_syndrome_zero_sum preserves P and Q, so
3819 * no need to mark them !uptodate here
3821 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3825 /* we have 2-disk failure */
3826 BUG_ON(s->failed != 2);
3828 case check_state_compute_result:
3829 sh->check_state = check_state_idle;
3831 /* check that a write has not made the stripe insync */
3832 if (test_bit(STRIPE_INSYNC, &sh->state))
3835 /* now write out any block on a failed drive,
3836 * or P or Q if they were recomputed
3838 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3839 if (s->failed == 2) {
3840 dev = &sh->dev[s->failed_num[1]];
3842 set_bit(R5_LOCKED, &dev->flags);
3843 set_bit(R5_Wantwrite, &dev->flags);
3845 if (s->failed >= 1) {
3846 dev = &sh->dev[s->failed_num[0]];
3848 set_bit(R5_LOCKED, &dev->flags);
3849 set_bit(R5_Wantwrite, &dev->flags);
3851 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3852 dev = &sh->dev[pd_idx];
3854 set_bit(R5_LOCKED, &dev->flags);
3855 set_bit(R5_Wantwrite, &dev->flags);
3857 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3858 dev = &sh->dev[qd_idx];
3860 set_bit(R5_LOCKED, &dev->flags);
3861 set_bit(R5_Wantwrite, &dev->flags);
3863 clear_bit(STRIPE_DEGRADED, &sh->state);
3865 set_bit(STRIPE_INSYNC, &sh->state);
3867 case check_state_run:
3868 case check_state_run_q:
3869 case check_state_run_pq:
3870 break; /* we will be called again upon completion */
3871 case check_state_check_result:
3872 sh->check_state = check_state_idle;
3874 /* handle a successful check operation, if parity is correct
3875 * we are done. Otherwise update the mismatch count and repair
3876 * parity if !MD_RECOVERY_CHECK
3878 if (sh->ops.zero_sum_result == 0) {
3879 /* both parities are correct */
3881 set_bit(STRIPE_INSYNC, &sh->state);
3883 /* in contrast to the raid5 case we can validate
3884 * parity, but still have a failure to write
3887 sh->check_state = check_state_compute_result;
3888 /* Returning at this point means that we may go
3889 * off and bring p and/or q uptodate again so
3890 * we make sure to check zero_sum_result again
3891 * to verify if p or q need writeback
3895 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3896 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3897 /* don't try to repair!! */
3898 set_bit(STRIPE_INSYNC, &sh->state);
3900 int *target = &sh->ops.target;
3902 sh->ops.target = -1;
3903 sh->ops.target2 = -1;
3904 sh->check_state = check_state_compute_run;
3905 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3906 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3907 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3908 set_bit(R5_Wantcompute,
3909 &sh->dev[pd_idx].flags);
3911 target = &sh->ops.target2;
3914 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3915 set_bit(R5_Wantcompute,
3916 &sh->dev[qd_idx].flags);
3923 case check_state_compute_run:
3926 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3927 __func__, sh->check_state,
3928 (unsigned long long) sh->sector);
3933 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3937 /* We have read all the blocks in this stripe and now we need to
3938 * copy some of them into a target stripe for expand.
3940 struct dma_async_tx_descriptor *tx = NULL;
3941 BUG_ON(sh->batch_head);
3942 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3943 for (i = 0; i < sh->disks; i++)
3944 if (i != sh->pd_idx && i != sh->qd_idx) {
3946 struct stripe_head *sh2;
3947 struct async_submit_ctl submit;
3949 sector_t bn = raid5_compute_blocknr(sh, i, 1);
3950 sector_t s = raid5_compute_sector(conf, bn, 0,
3952 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
3954 /* so far only the early blocks of this stripe
3955 * have been requested. When later blocks
3956 * get requested, we will try again
3959 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3960 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3961 /* must have already done this block */
3962 raid5_release_stripe(sh2);
3966 /* place all the copies on one channel */
3967 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3968 tx = async_memcpy(sh2->dev[dd_idx].page,
3969 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3972 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3973 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3974 for (j = 0; j < conf->raid_disks; j++)
3975 if (j != sh2->pd_idx &&
3977 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3979 if (j == conf->raid_disks) {
3980 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3981 set_bit(STRIPE_HANDLE, &sh2->state);
3983 raid5_release_stripe(sh2);
3986 /* done submitting copies, wait for them to complete */
3987 async_tx_quiesce(&tx);
3991 * handle_stripe - do things to a stripe.
3993 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3994 * state of various bits to see what needs to be done.
3996 * return some read requests which now have data
3997 * return some write requests which are safely on storage
3998 * schedule a read on some buffers
3999 * schedule a write of some buffers
4000 * return confirmation of parity correctness
4004 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4006 struct r5conf *conf = sh->raid_conf;
4007 int disks = sh->disks;
4010 int do_recovery = 0;
4012 memset(s, 0, sizeof(*s));
4014 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4015 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4016 s->failed_num[0] = -1;
4017 s->failed_num[1] = -1;
4019 /* Now to look around and see what can be done */
4021 for (i=disks; i--; ) {
4022 struct md_rdev *rdev;
4029 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4031 dev->toread, dev->towrite, dev->written);
4032 /* maybe we can reply to a read
4034 * new wantfill requests are only permitted while
4035 * ops_complete_biofill is guaranteed to be inactive
4037 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4038 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4039 set_bit(R5_Wantfill, &dev->flags);
4041 /* now count some things */
4042 if (test_bit(R5_LOCKED, &dev->flags))
4044 if (test_bit(R5_UPTODATE, &dev->flags))
4046 if (test_bit(R5_Wantcompute, &dev->flags)) {
4048 BUG_ON(s->compute > 2);
4051 if (test_bit(R5_Wantfill, &dev->flags))
4053 else if (dev->toread)
4057 if (!test_bit(R5_OVERWRITE, &dev->flags))
4062 /* Prefer to use the replacement for reads, but only
4063 * if it is recovered enough and has no bad blocks.
4065 rdev = rcu_dereference(conf->disks[i].replacement);
4066 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4067 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4068 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4069 &first_bad, &bad_sectors))
4070 set_bit(R5_ReadRepl, &dev->flags);
4072 if (rdev && !test_bit(Faulty, &rdev->flags))
4073 set_bit(R5_NeedReplace, &dev->flags);
4075 clear_bit(R5_NeedReplace, &dev->flags);
4076 rdev = rcu_dereference(conf->disks[i].rdev);
4077 clear_bit(R5_ReadRepl, &dev->flags);
4079 if (rdev && test_bit(Faulty, &rdev->flags))
4082 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4083 &first_bad, &bad_sectors);
4084 if (s->blocked_rdev == NULL
4085 && (test_bit(Blocked, &rdev->flags)
4088 set_bit(BlockedBadBlocks,
4090 s->blocked_rdev = rdev;
4091 atomic_inc(&rdev->nr_pending);
4094 clear_bit(R5_Insync, &dev->flags);
4098 /* also not in-sync */
4099 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4100 test_bit(R5_UPTODATE, &dev->flags)) {
4101 /* treat as in-sync, but with a read error
4102 * which we can now try to correct
4104 set_bit(R5_Insync, &dev->flags);
4105 set_bit(R5_ReadError, &dev->flags);
4107 } else if (test_bit(In_sync, &rdev->flags))
4108 set_bit(R5_Insync, &dev->flags);
4109 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4110 /* in sync if before recovery_offset */
4111 set_bit(R5_Insync, &dev->flags);
4112 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4113 test_bit(R5_Expanded, &dev->flags))
4114 /* If we've reshaped into here, we assume it is Insync.
4115 * We will shortly update recovery_offset to make
4118 set_bit(R5_Insync, &dev->flags);
4120 if (test_bit(R5_WriteError, &dev->flags)) {
4121 /* This flag does not apply to '.replacement'
4122 * only to .rdev, so make sure to check that*/
4123 struct md_rdev *rdev2 = rcu_dereference(
4124 conf->disks[i].rdev);
4126 clear_bit(R5_Insync, &dev->flags);
4127 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4128 s->handle_bad_blocks = 1;
4129 atomic_inc(&rdev2->nr_pending);
4131 clear_bit(R5_WriteError, &dev->flags);
4133 if (test_bit(R5_MadeGood, &dev->flags)) {
4134 /* This flag does not apply to '.replacement'
4135 * only to .rdev, so make sure to check that*/
4136 struct md_rdev *rdev2 = rcu_dereference(
4137 conf->disks[i].rdev);
4138 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4139 s->handle_bad_blocks = 1;
4140 atomic_inc(&rdev2->nr_pending);
4142 clear_bit(R5_MadeGood, &dev->flags);
4144 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4145 struct md_rdev *rdev2 = rcu_dereference(
4146 conf->disks[i].replacement);
4147 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4148 s->handle_bad_blocks = 1;
4149 atomic_inc(&rdev2->nr_pending);
4151 clear_bit(R5_MadeGoodRepl, &dev->flags);
4153 if (!test_bit(R5_Insync, &dev->flags)) {
4154 /* The ReadError flag will just be confusing now */
4155 clear_bit(R5_ReadError, &dev->flags);
4156 clear_bit(R5_ReWrite, &dev->flags);
4158 if (test_bit(R5_ReadError, &dev->flags))
4159 clear_bit(R5_Insync, &dev->flags);
4160 if (!test_bit(R5_Insync, &dev->flags)) {
4162 s->failed_num[s->failed] = i;
4164 if (rdev && !test_bit(Faulty, &rdev->flags))
4168 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4169 /* If there is a failed device being replaced,
4170 * we must be recovering.
4171 * else if we are after recovery_cp, we must be syncing
4172 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4173 * else we can only be replacing
4174 * sync and recovery both need to read all devices, and so
4175 * use the same flag.
4178 sh->sector >= conf->mddev->recovery_cp ||
4179 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4187 static int clear_batch_ready(struct stripe_head *sh)
4189 /* Return '1' if this is a member of batch, or
4190 * '0' if it is a lone stripe or a head which can now be
4193 struct stripe_head *tmp;
4194 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4195 return (sh->batch_head && sh->batch_head != sh);
4196 spin_lock(&sh->stripe_lock);
4197 if (!sh->batch_head) {
4198 spin_unlock(&sh->stripe_lock);
4203 * this stripe could be added to a batch list before we check
4204 * BATCH_READY, skips it
4206 if (sh->batch_head != sh) {
4207 spin_unlock(&sh->stripe_lock);
4210 spin_lock(&sh->batch_lock);
4211 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4212 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4213 spin_unlock(&sh->batch_lock);
4214 spin_unlock(&sh->stripe_lock);
4217 * BATCH_READY is cleared, no new stripes can be added.
4218 * batch_list can be accessed without lock
4223 static void break_stripe_batch_list(struct stripe_head *head_sh,
4224 unsigned long handle_flags)
4226 struct stripe_head *sh, *next;
4230 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4232 list_del_init(&sh->batch_list);
4234 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4235 (1 << STRIPE_SYNCING) |
4236 (1 << STRIPE_REPLACED) |
4237 (1 << STRIPE_PREREAD_ACTIVE) |
4238 (1 << STRIPE_DELAYED) |
4239 (1 << STRIPE_BIT_DELAY) |
4240 (1 << STRIPE_FULL_WRITE) |
4241 (1 << STRIPE_BIOFILL_RUN) |
4242 (1 << STRIPE_COMPUTE_RUN) |
4243 (1 << STRIPE_OPS_REQ_PENDING) |
4244 (1 << STRIPE_DISCARD) |
4245 (1 << STRIPE_BATCH_READY) |
4246 (1 << STRIPE_BATCH_ERR) |
4247 (1 << STRIPE_BITMAP_PENDING)));
4248 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4249 (1 << STRIPE_REPLACED)));
4251 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4252 (1 << STRIPE_DEGRADED)),
4253 head_sh->state & (1 << STRIPE_INSYNC));
4255 sh->check_state = head_sh->check_state;
4256 sh->reconstruct_state = head_sh->reconstruct_state;
4257 for (i = 0; i < sh->disks; i++) {
4258 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4260 sh->dev[i].flags = head_sh->dev[i].flags &
4261 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4263 spin_lock_irq(&sh->stripe_lock);
4264 sh->batch_head = NULL;
4265 spin_unlock_irq(&sh->stripe_lock);
4266 if (handle_flags == 0 ||
4267 sh->state & handle_flags)
4268 set_bit(STRIPE_HANDLE, &sh->state);
4269 raid5_release_stripe(sh);
4271 spin_lock_irq(&head_sh->stripe_lock);
4272 head_sh->batch_head = NULL;
4273 spin_unlock_irq(&head_sh->stripe_lock);
4274 for (i = 0; i < head_sh->disks; i++)
4275 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4277 if (head_sh->state & handle_flags)
4278 set_bit(STRIPE_HANDLE, &head_sh->state);
4281 wake_up(&head_sh->raid_conf->wait_for_overlap);
4284 static void handle_stripe(struct stripe_head *sh)
4286 struct stripe_head_state s;
4287 struct r5conf *conf = sh->raid_conf;
4290 int disks = sh->disks;
4291 struct r5dev *pdev, *qdev;
4293 clear_bit(STRIPE_HANDLE, &sh->state);
4294 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4295 /* already being handled, ensure it gets handled
4296 * again when current action finishes */
4297 set_bit(STRIPE_HANDLE, &sh->state);
4301 if (clear_batch_ready(sh) ) {
4302 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4306 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4307 break_stripe_batch_list(sh, 0);
4309 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4310 spin_lock(&sh->stripe_lock);
4311 /* Cannot process 'sync' concurrently with 'discard' */
4312 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4313 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4314 set_bit(STRIPE_SYNCING, &sh->state);
4315 clear_bit(STRIPE_INSYNC, &sh->state);
4316 clear_bit(STRIPE_REPLACED, &sh->state);
4318 spin_unlock(&sh->stripe_lock);
4320 clear_bit(STRIPE_DELAYED, &sh->state);
4322 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4323 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4324 (unsigned long long)sh->sector, sh->state,
4325 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4326 sh->check_state, sh->reconstruct_state);
4328 analyse_stripe(sh, &s);
4330 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4333 if (s.handle_bad_blocks) {
4334 set_bit(STRIPE_HANDLE, &sh->state);
4338 if (unlikely(s.blocked_rdev)) {
4339 if (s.syncing || s.expanding || s.expanded ||
4340 s.replacing || s.to_write || s.written) {
4341 set_bit(STRIPE_HANDLE, &sh->state);
4344 /* There is nothing for the blocked_rdev to block */
4345 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4346 s.blocked_rdev = NULL;
4349 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4350 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4351 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4354 pr_debug("locked=%d uptodate=%d to_read=%d"
4355 " to_write=%d failed=%d failed_num=%d,%d\n",
4356 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4357 s.failed_num[0], s.failed_num[1]);
4358 /* check if the array has lost more than max_degraded devices and,
4359 * if so, some requests might need to be failed.
4361 if (s.failed > conf->max_degraded) {
4362 sh->check_state = 0;
4363 sh->reconstruct_state = 0;
4364 break_stripe_batch_list(sh, 0);
4365 if (s.to_read+s.to_write+s.written)
4366 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4367 if (s.syncing + s.replacing)
4368 handle_failed_sync(conf, sh, &s);
4371 /* Now we check to see if any write operations have recently
4375 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4377 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4378 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4379 sh->reconstruct_state = reconstruct_state_idle;
4381 /* All the 'written' buffers and the parity block are ready to
4382 * be written back to disk
4384 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4385 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4386 BUG_ON(sh->qd_idx >= 0 &&
4387 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4388 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4389 for (i = disks; i--; ) {
4390 struct r5dev *dev = &sh->dev[i];
4391 if (test_bit(R5_LOCKED, &dev->flags) &&
4392 (i == sh->pd_idx || i == sh->qd_idx ||
4394 pr_debug("Writing block %d\n", i);
4395 set_bit(R5_Wantwrite, &dev->flags);
4400 if (!test_bit(R5_Insync, &dev->flags) ||
4401 ((i == sh->pd_idx || i == sh->qd_idx) &&
4403 set_bit(STRIPE_INSYNC, &sh->state);
4406 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4407 s.dec_preread_active = 1;
4411 * might be able to return some write requests if the parity blocks
4412 * are safe, or on a failed drive
4414 pdev = &sh->dev[sh->pd_idx];
4415 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4416 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4417 qdev = &sh->dev[sh->qd_idx];
4418 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4419 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4423 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4424 && !test_bit(R5_LOCKED, &pdev->flags)
4425 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4426 test_bit(R5_Discard, &pdev->flags))))) &&
4427 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4428 && !test_bit(R5_LOCKED, &qdev->flags)
4429 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4430 test_bit(R5_Discard, &qdev->flags))))))
4431 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4433 /* Now we might consider reading some blocks, either to check/generate
4434 * parity, or to satisfy requests
4435 * or to load a block that is being partially written.
4437 if (s.to_read || s.non_overwrite
4438 || (conf->level == 6 && s.to_write && s.failed)
4439 || (s.syncing && (s.uptodate + s.compute < disks))
4442 handle_stripe_fill(sh, &s, disks);
4444 /* Now to consider new write requests and what else, if anything
4445 * should be read. We do not handle new writes when:
4446 * 1/ A 'write' operation (copy+xor) is already in flight.
4447 * 2/ A 'check' operation is in flight, as it may clobber the parity
4450 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4451 handle_stripe_dirtying(conf, sh, &s, disks);
4453 /* maybe we need to check and possibly fix the parity for this stripe
4454 * Any reads will already have been scheduled, so we just see if enough
4455 * data is available. The parity check is held off while parity
4456 * dependent operations are in flight.
4458 if (sh->check_state ||
4459 (s.syncing && s.locked == 0 &&
4460 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4461 !test_bit(STRIPE_INSYNC, &sh->state))) {
4462 if (conf->level == 6)
4463 handle_parity_checks6(conf, sh, &s, disks);
4465 handle_parity_checks5(conf, sh, &s, disks);
4468 if ((s.replacing || s.syncing) && s.locked == 0
4469 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4470 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4471 /* Write out to replacement devices where possible */
4472 for (i = 0; i < conf->raid_disks; i++)
4473 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4474 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4475 set_bit(R5_WantReplace, &sh->dev[i].flags);
4476 set_bit(R5_LOCKED, &sh->dev[i].flags);
4480 set_bit(STRIPE_INSYNC, &sh->state);
4481 set_bit(STRIPE_REPLACED, &sh->state);
4483 if ((s.syncing || s.replacing) && s.locked == 0 &&
4484 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4485 test_bit(STRIPE_INSYNC, &sh->state)) {
4486 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4487 clear_bit(STRIPE_SYNCING, &sh->state);
4488 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4489 wake_up(&conf->wait_for_overlap);
4492 /* If the failed drives are just a ReadError, then we might need
4493 * to progress the repair/check process
4495 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4496 for (i = 0; i < s.failed; i++) {
4497 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4498 if (test_bit(R5_ReadError, &dev->flags)
4499 && !test_bit(R5_LOCKED, &dev->flags)
4500 && test_bit(R5_UPTODATE, &dev->flags)
4502 if (!test_bit(R5_ReWrite, &dev->flags)) {
4503 set_bit(R5_Wantwrite, &dev->flags);
4504 set_bit(R5_ReWrite, &dev->flags);
4505 set_bit(R5_LOCKED, &dev->flags);
4508 /* let's read it back */
4509 set_bit(R5_Wantread, &dev->flags);
4510 set_bit(R5_LOCKED, &dev->flags);
4516 /* Finish reconstruct operations initiated by the expansion process */
4517 if (sh->reconstruct_state == reconstruct_state_result) {
4518 struct stripe_head *sh_src
4519 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4520 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4521 /* sh cannot be written until sh_src has been read.
4522 * so arrange for sh to be delayed a little
4524 set_bit(STRIPE_DELAYED, &sh->state);
4525 set_bit(STRIPE_HANDLE, &sh->state);
4526 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4528 atomic_inc(&conf->preread_active_stripes);
4529 raid5_release_stripe(sh_src);
4533 raid5_release_stripe(sh_src);
4535 sh->reconstruct_state = reconstruct_state_idle;
4536 clear_bit(STRIPE_EXPANDING, &sh->state);
4537 for (i = conf->raid_disks; i--; ) {
4538 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4539 set_bit(R5_LOCKED, &sh->dev[i].flags);
4544 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4545 !sh->reconstruct_state) {
4546 /* Need to write out all blocks after computing parity */
4547 sh->disks = conf->raid_disks;
4548 stripe_set_idx(sh->sector, conf, 0, sh);
4549 schedule_reconstruction(sh, &s, 1, 1);
4550 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4551 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4552 atomic_dec(&conf->reshape_stripes);
4553 wake_up(&conf->wait_for_overlap);
4554 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4557 if (s.expanding && s.locked == 0 &&
4558 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4559 handle_stripe_expansion(conf, sh);
4562 /* wait for this device to become unblocked */
4563 if (unlikely(s.blocked_rdev)) {
4564 if (conf->mddev->external)
4565 md_wait_for_blocked_rdev(s.blocked_rdev,
4568 /* Internal metadata will immediately
4569 * be written by raid5d, so we don't
4570 * need to wait here.
4572 rdev_dec_pending(s.blocked_rdev,
4576 if (s.handle_bad_blocks)
4577 for (i = disks; i--; ) {
4578 struct md_rdev *rdev;
4579 struct r5dev *dev = &sh->dev[i];
4580 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4581 /* We own a safe reference to the rdev */
4582 rdev = conf->disks[i].rdev;
4583 if (!rdev_set_badblocks(rdev, sh->sector,
4585 md_error(conf->mddev, rdev);
4586 rdev_dec_pending(rdev, conf->mddev);
4588 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4589 rdev = conf->disks[i].rdev;
4590 rdev_clear_badblocks(rdev, sh->sector,
4592 rdev_dec_pending(rdev, conf->mddev);
4594 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4595 rdev = conf->disks[i].replacement;
4597 /* rdev have been moved down */
4598 rdev = conf->disks[i].rdev;
4599 rdev_clear_badblocks(rdev, sh->sector,
4601 rdev_dec_pending(rdev, conf->mddev);
4606 raid_run_ops(sh, s.ops_request);
4610 if (s.dec_preread_active) {
4611 /* We delay this until after ops_run_io so that if make_request
4612 * is waiting on a flush, it won't continue until the writes
4613 * have actually been submitted.
4615 atomic_dec(&conf->preread_active_stripes);
4616 if (atomic_read(&conf->preread_active_stripes) <
4618 md_wakeup_thread(conf->mddev->thread);
4621 if (!bio_list_empty(&s.return_bi)) {
4622 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags)) {
4623 spin_lock_irq(&conf->device_lock);
4624 bio_list_merge(&conf->return_bi, &s.return_bi);
4625 spin_unlock_irq(&conf->device_lock);
4626 md_wakeup_thread(conf->mddev->thread);
4628 return_io(&s.return_bi);
4631 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4634 static void raid5_activate_delayed(struct r5conf *conf)
4636 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4637 while (!list_empty(&conf->delayed_list)) {
4638 struct list_head *l = conf->delayed_list.next;
4639 struct stripe_head *sh;
4640 sh = list_entry(l, struct stripe_head, lru);
4642 clear_bit(STRIPE_DELAYED, &sh->state);
4643 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4644 atomic_inc(&conf->preread_active_stripes);
4645 list_add_tail(&sh->lru, &conf->hold_list);
4646 raid5_wakeup_stripe_thread(sh);
4651 static void activate_bit_delay(struct r5conf *conf,
4652 struct list_head *temp_inactive_list)
4654 /* device_lock is held */
4655 struct list_head head;
4656 list_add(&head, &conf->bitmap_list);
4657 list_del_init(&conf->bitmap_list);
4658 while (!list_empty(&head)) {
4659 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4661 list_del_init(&sh->lru);
4662 atomic_inc(&sh->count);
4663 hash = sh->hash_lock_index;
4664 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4668 static int raid5_congested(struct mddev *mddev, int bits)
4670 struct r5conf *conf = mddev->private;
4672 /* No difference between reads and writes. Just check
4673 * how busy the stripe_cache is
4676 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4680 if (atomic_read(&conf->empty_inactive_list_nr))
4686 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4688 struct r5conf *conf = mddev->private;
4689 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4690 unsigned int chunk_sectors;
4691 unsigned int bio_sectors = bio_sectors(bio);
4693 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4694 return chunk_sectors >=
4695 ((sector & (chunk_sectors - 1)) + bio_sectors);
4699 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4700 * later sampled by raid5d.
4702 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4704 unsigned long flags;
4706 spin_lock_irqsave(&conf->device_lock, flags);
4708 bi->bi_next = conf->retry_read_aligned_list;
4709 conf->retry_read_aligned_list = bi;
4711 spin_unlock_irqrestore(&conf->device_lock, flags);
4712 md_wakeup_thread(conf->mddev->thread);
4715 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4719 bi = conf->retry_read_aligned;
4721 conf->retry_read_aligned = NULL;
4724 bi = conf->retry_read_aligned_list;
4726 conf->retry_read_aligned_list = bi->bi_next;
4729 * this sets the active strip count to 1 and the processed
4730 * strip count to zero (upper 8 bits)
4732 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4739 * The "raid5_align_endio" should check if the read succeeded and if it
4740 * did, call bio_endio on the original bio (having bio_put the new bio
4742 * If the read failed..
4744 static void raid5_align_endio(struct bio *bi)
4746 struct bio* raid_bi = bi->bi_private;
4747 struct mddev *mddev;
4748 struct r5conf *conf;
4749 struct md_rdev *rdev;
4750 int error = bi->bi_error;
4754 rdev = (void*)raid_bi->bi_next;
4755 raid_bi->bi_next = NULL;
4756 mddev = rdev->mddev;
4757 conf = mddev->private;
4759 rdev_dec_pending(rdev, conf->mddev);
4762 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4765 if (atomic_dec_and_test(&conf->active_aligned_reads))
4766 wake_up(&conf->wait_for_quiescent);
4770 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4772 add_bio_to_retry(raid_bi, conf);
4775 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4777 struct r5conf *conf = mddev->private;
4779 struct bio* align_bi;
4780 struct md_rdev *rdev;
4781 sector_t end_sector;
4783 if (!in_chunk_boundary(mddev, raid_bio)) {
4784 pr_debug("%s: non aligned\n", __func__);
4788 * use bio_clone_mddev to make a copy of the bio
4790 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4794 * set bi_end_io to a new function, and set bi_private to the
4797 align_bi->bi_end_io = raid5_align_endio;
4798 align_bi->bi_private = raid_bio;
4802 align_bi->bi_iter.bi_sector =
4803 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4806 end_sector = bio_end_sector(align_bi);
4808 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4809 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4810 rdev->recovery_offset < end_sector) {
4811 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4813 (test_bit(Faulty, &rdev->flags) ||
4814 !(test_bit(In_sync, &rdev->flags) ||
4815 rdev->recovery_offset >= end_sector)))
4822 atomic_inc(&rdev->nr_pending);
4824 raid_bio->bi_next = (void*)rdev;
4825 align_bi->bi_bdev = rdev->bdev;
4826 bio_clear_flag(align_bi, BIO_SEG_VALID);
4828 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4829 bio_sectors(align_bi),
4830 &first_bad, &bad_sectors)) {
4832 rdev_dec_pending(rdev, mddev);
4836 /* No reshape active, so we can trust rdev->data_offset */
4837 align_bi->bi_iter.bi_sector += rdev->data_offset;
4839 spin_lock_irq(&conf->device_lock);
4840 wait_event_lock_irq(conf->wait_for_quiescent,
4843 atomic_inc(&conf->active_aligned_reads);
4844 spin_unlock_irq(&conf->device_lock);
4847 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4848 align_bi, disk_devt(mddev->gendisk),
4849 raid_bio->bi_iter.bi_sector);
4850 generic_make_request(align_bi);
4859 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4864 sector_t sector = raid_bio->bi_iter.bi_sector;
4865 unsigned chunk_sects = mddev->chunk_sectors;
4866 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4868 if (sectors < bio_sectors(raid_bio)) {
4869 split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4870 bio_chain(split, raid_bio);
4874 if (!raid5_read_one_chunk(mddev, split)) {
4875 if (split != raid_bio)
4876 generic_make_request(raid_bio);
4879 } while (split != raid_bio);
4884 /* __get_priority_stripe - get the next stripe to process
4886 * Full stripe writes are allowed to pass preread active stripes up until
4887 * the bypass_threshold is exceeded. In general the bypass_count
4888 * increments when the handle_list is handled before the hold_list; however, it
4889 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4890 * stripe with in flight i/o. The bypass_count will be reset when the
4891 * head of the hold_list has changed, i.e. the head was promoted to the
4894 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4896 struct stripe_head *sh = NULL, *tmp;
4897 struct list_head *handle_list = NULL;
4898 struct r5worker_group *wg = NULL;
4900 if (conf->worker_cnt_per_group == 0) {
4901 handle_list = &conf->handle_list;
4902 } else if (group != ANY_GROUP) {
4903 handle_list = &conf->worker_groups[group].handle_list;
4904 wg = &conf->worker_groups[group];
4907 for (i = 0; i < conf->group_cnt; i++) {
4908 handle_list = &conf->worker_groups[i].handle_list;
4909 wg = &conf->worker_groups[i];
4910 if (!list_empty(handle_list))
4915 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4917 list_empty(handle_list) ? "empty" : "busy",
4918 list_empty(&conf->hold_list) ? "empty" : "busy",
4919 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4921 if (!list_empty(handle_list)) {
4922 sh = list_entry(handle_list->next, typeof(*sh), lru);
4924 if (list_empty(&conf->hold_list))
4925 conf->bypass_count = 0;
4926 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4927 if (conf->hold_list.next == conf->last_hold)
4928 conf->bypass_count++;
4930 conf->last_hold = conf->hold_list.next;
4931 conf->bypass_count -= conf->bypass_threshold;
4932 if (conf->bypass_count < 0)
4933 conf->bypass_count = 0;
4936 } else if (!list_empty(&conf->hold_list) &&
4937 ((conf->bypass_threshold &&
4938 conf->bypass_count > conf->bypass_threshold) ||
4939 atomic_read(&conf->pending_full_writes) == 0)) {
4941 list_for_each_entry(tmp, &conf->hold_list, lru) {
4942 if (conf->worker_cnt_per_group == 0 ||
4943 group == ANY_GROUP ||
4944 !cpu_online(tmp->cpu) ||
4945 cpu_to_group(tmp->cpu) == group) {
4952 conf->bypass_count -= conf->bypass_threshold;
4953 if (conf->bypass_count < 0)
4954 conf->bypass_count = 0;
4966 list_del_init(&sh->lru);
4967 BUG_ON(atomic_inc_return(&sh->count) != 1);
4971 struct raid5_plug_cb {
4972 struct blk_plug_cb cb;
4973 struct list_head list;
4974 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4977 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4979 struct raid5_plug_cb *cb = container_of(
4980 blk_cb, struct raid5_plug_cb, cb);
4981 struct stripe_head *sh;
4982 struct mddev *mddev = cb->cb.data;
4983 struct r5conf *conf = mddev->private;
4987 if (cb->list.next && !list_empty(&cb->list)) {
4988 spin_lock_irq(&conf->device_lock);
4989 while (!list_empty(&cb->list)) {
4990 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4991 list_del_init(&sh->lru);
4993 * avoid race release_stripe_plug() sees
4994 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4995 * is still in our list
4997 smp_mb__before_atomic();
4998 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5000 * STRIPE_ON_RELEASE_LIST could be set here. In that
5001 * case, the count is always > 1 here
5003 hash = sh->hash_lock_index;
5004 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5007 spin_unlock_irq(&conf->device_lock);
5009 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5010 NR_STRIPE_HASH_LOCKS);
5012 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5016 static void release_stripe_plug(struct mddev *mddev,
5017 struct stripe_head *sh)
5019 struct blk_plug_cb *blk_cb = blk_check_plugged(
5020 raid5_unplug, mddev,
5021 sizeof(struct raid5_plug_cb));
5022 struct raid5_plug_cb *cb;
5025 raid5_release_stripe(sh);
5029 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5031 if (cb->list.next == NULL) {
5033 INIT_LIST_HEAD(&cb->list);
5034 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5035 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5038 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5039 list_add_tail(&sh->lru, &cb->list);
5041 raid5_release_stripe(sh);
5044 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5046 struct r5conf *conf = mddev->private;
5047 sector_t logical_sector, last_sector;
5048 struct stripe_head *sh;
5052 if (mddev->reshape_position != MaxSector)
5053 /* Skip discard while reshape is happening */
5056 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5057 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5060 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5062 stripe_sectors = conf->chunk_sectors *
5063 (conf->raid_disks - conf->max_degraded);
5064 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5066 sector_div(last_sector, stripe_sectors);
5068 logical_sector *= conf->chunk_sectors;
5069 last_sector *= conf->chunk_sectors;
5071 for (; logical_sector < last_sector;
5072 logical_sector += STRIPE_SECTORS) {
5076 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5077 prepare_to_wait(&conf->wait_for_overlap, &w,
5078 TASK_UNINTERRUPTIBLE);
5079 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5080 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5081 raid5_release_stripe(sh);
5085 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5086 spin_lock_irq(&sh->stripe_lock);
5087 for (d = 0; d < conf->raid_disks; d++) {
5088 if (d == sh->pd_idx || d == sh->qd_idx)
5090 if (sh->dev[d].towrite || sh->dev[d].toread) {
5091 set_bit(R5_Overlap, &sh->dev[d].flags);
5092 spin_unlock_irq(&sh->stripe_lock);
5093 raid5_release_stripe(sh);
5098 set_bit(STRIPE_DISCARD, &sh->state);
5099 finish_wait(&conf->wait_for_overlap, &w);
5100 sh->overwrite_disks = 0;
5101 for (d = 0; d < conf->raid_disks; d++) {
5102 if (d == sh->pd_idx || d == sh->qd_idx)
5104 sh->dev[d].towrite = bi;
5105 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5106 raid5_inc_bi_active_stripes(bi);
5107 sh->overwrite_disks++;
5109 spin_unlock_irq(&sh->stripe_lock);
5110 if (conf->mddev->bitmap) {
5112 d < conf->raid_disks - conf->max_degraded;
5114 bitmap_startwrite(mddev->bitmap,
5118 sh->bm_seq = conf->seq_flush + 1;
5119 set_bit(STRIPE_BIT_DELAY, &sh->state);
5122 set_bit(STRIPE_HANDLE, &sh->state);
5123 clear_bit(STRIPE_DELAYED, &sh->state);
5124 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5125 atomic_inc(&conf->preread_active_stripes);
5126 release_stripe_plug(mddev, sh);
5129 remaining = raid5_dec_bi_active_stripes(bi);
5130 if (remaining == 0) {
5131 md_write_end(mddev);
5136 static void make_request(struct mddev *mddev, struct bio * bi)
5138 struct r5conf *conf = mddev->private;
5140 sector_t new_sector;
5141 sector_t logical_sector, last_sector;
5142 struct stripe_head *sh;
5143 const int rw = bio_data_dir(bi);
5148 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5149 md_flush_request(mddev, bi);
5153 md_write_start(mddev, bi);
5156 * If array is degraded, better not do chunk aligned read because
5157 * later we might have to read it again in order to reconstruct
5158 * data on failed drives.
5160 if (rw == READ && mddev->degraded == 0 &&
5161 mddev->reshape_position == MaxSector) {
5162 bi = chunk_aligned_read(mddev, bi);
5167 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5168 make_discard_request(mddev, bi);
5172 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5173 last_sector = bio_end_sector(bi);
5175 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5177 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5178 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5184 seq = read_seqcount_begin(&conf->gen_lock);
5187 prepare_to_wait(&conf->wait_for_overlap, &w,
5188 TASK_UNINTERRUPTIBLE);
5189 if (unlikely(conf->reshape_progress != MaxSector)) {
5190 /* spinlock is needed as reshape_progress may be
5191 * 64bit on a 32bit platform, and so it might be
5192 * possible to see a half-updated value
5193 * Of course reshape_progress could change after
5194 * the lock is dropped, so once we get a reference
5195 * to the stripe that we think it is, we will have
5198 spin_lock_irq(&conf->device_lock);
5199 if (mddev->reshape_backwards
5200 ? logical_sector < conf->reshape_progress
5201 : logical_sector >= conf->reshape_progress) {
5204 if (mddev->reshape_backwards
5205 ? logical_sector < conf->reshape_safe
5206 : logical_sector >= conf->reshape_safe) {
5207 spin_unlock_irq(&conf->device_lock);
5213 spin_unlock_irq(&conf->device_lock);
5216 new_sector = raid5_compute_sector(conf, logical_sector,
5219 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5220 (unsigned long long)new_sector,
5221 (unsigned long long)logical_sector);
5223 sh = raid5_get_active_stripe(conf, new_sector, previous,
5224 (bi->bi_rw&RWA_MASK), 0);
5226 if (unlikely(previous)) {
5227 /* expansion might have moved on while waiting for a
5228 * stripe, so we must do the range check again.
5229 * Expansion could still move past after this
5230 * test, but as we are holding a reference to
5231 * 'sh', we know that if that happens,
5232 * STRIPE_EXPANDING will get set and the expansion
5233 * won't proceed until we finish with the stripe.
5236 spin_lock_irq(&conf->device_lock);
5237 if (mddev->reshape_backwards
5238 ? logical_sector >= conf->reshape_progress
5239 : logical_sector < conf->reshape_progress)
5240 /* mismatch, need to try again */
5242 spin_unlock_irq(&conf->device_lock);
5244 raid5_release_stripe(sh);
5250 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5251 /* Might have got the wrong stripe_head
5254 raid5_release_stripe(sh);
5259 logical_sector >= mddev->suspend_lo &&
5260 logical_sector < mddev->suspend_hi) {
5261 raid5_release_stripe(sh);
5262 /* As the suspend_* range is controlled by
5263 * userspace, we want an interruptible
5266 flush_signals(current);
5267 prepare_to_wait(&conf->wait_for_overlap,
5268 &w, TASK_INTERRUPTIBLE);
5269 if (logical_sector >= mddev->suspend_lo &&
5270 logical_sector < mddev->suspend_hi) {
5277 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5278 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5279 /* Stripe is busy expanding or
5280 * add failed due to overlap. Flush everything
5283 md_wakeup_thread(mddev->thread);
5284 raid5_release_stripe(sh);
5289 set_bit(STRIPE_HANDLE, &sh->state);
5290 clear_bit(STRIPE_DELAYED, &sh->state);
5291 if ((!sh->batch_head || sh == sh->batch_head) &&
5292 (bi->bi_rw & REQ_SYNC) &&
5293 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5294 atomic_inc(&conf->preread_active_stripes);
5295 release_stripe_plug(mddev, sh);
5297 /* cannot get stripe for read-ahead, just give-up */
5298 bi->bi_error = -EIO;
5302 finish_wait(&conf->wait_for_overlap, &w);
5304 remaining = raid5_dec_bi_active_stripes(bi);
5305 if (remaining == 0) {
5308 md_write_end(mddev);
5310 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5316 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5318 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5320 /* reshaping is quite different to recovery/resync so it is
5321 * handled quite separately ... here.
5323 * On each call to sync_request, we gather one chunk worth of
5324 * destination stripes and flag them as expanding.
5325 * Then we find all the source stripes and request reads.
5326 * As the reads complete, handle_stripe will copy the data
5327 * into the destination stripe and release that stripe.
5329 struct r5conf *conf = mddev->private;
5330 struct stripe_head *sh;
5331 sector_t first_sector, last_sector;
5332 int raid_disks = conf->previous_raid_disks;
5333 int data_disks = raid_disks - conf->max_degraded;
5334 int new_data_disks = conf->raid_disks - conf->max_degraded;
5337 sector_t writepos, readpos, safepos;
5338 sector_t stripe_addr;
5339 int reshape_sectors;
5340 struct list_head stripes;
5343 if (sector_nr == 0) {
5344 /* If restarting in the middle, skip the initial sectors */
5345 if (mddev->reshape_backwards &&
5346 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5347 sector_nr = raid5_size(mddev, 0, 0)
5348 - conf->reshape_progress;
5349 } else if (mddev->reshape_backwards &&
5350 conf->reshape_progress == MaxSector) {
5351 /* shouldn't happen, but just in case, finish up.*/
5352 sector_nr = MaxSector;
5353 } else if (!mddev->reshape_backwards &&
5354 conf->reshape_progress > 0)
5355 sector_nr = conf->reshape_progress;
5356 sector_div(sector_nr, new_data_disks);
5358 mddev->curr_resync_completed = sector_nr;
5359 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5366 /* We need to process a full chunk at a time.
5367 * If old and new chunk sizes differ, we need to process the
5371 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5373 /* We update the metadata at least every 10 seconds, or when
5374 * the data about to be copied would over-write the source of
5375 * the data at the front of the range. i.e. one new_stripe
5376 * along from reshape_progress new_maps to after where
5377 * reshape_safe old_maps to
5379 writepos = conf->reshape_progress;
5380 sector_div(writepos, new_data_disks);
5381 readpos = conf->reshape_progress;
5382 sector_div(readpos, data_disks);
5383 safepos = conf->reshape_safe;
5384 sector_div(safepos, data_disks);
5385 if (mddev->reshape_backwards) {
5386 BUG_ON(writepos < reshape_sectors);
5387 writepos -= reshape_sectors;
5388 readpos += reshape_sectors;
5389 safepos += reshape_sectors;
5391 writepos += reshape_sectors;
5392 /* readpos and safepos are worst-case calculations.
5393 * A negative number is overly pessimistic, and causes
5394 * obvious problems for unsigned storage. So clip to 0.
5396 readpos -= min_t(sector_t, reshape_sectors, readpos);
5397 safepos -= min_t(sector_t, reshape_sectors, safepos);
5400 /* Having calculated the 'writepos' possibly use it
5401 * to set 'stripe_addr' which is where we will write to.
5403 if (mddev->reshape_backwards) {
5404 BUG_ON(conf->reshape_progress == 0);
5405 stripe_addr = writepos;
5406 BUG_ON((mddev->dev_sectors &
5407 ~((sector_t)reshape_sectors - 1))
5408 - reshape_sectors - stripe_addr
5411 BUG_ON(writepos != sector_nr + reshape_sectors);
5412 stripe_addr = sector_nr;
5415 /* 'writepos' is the most advanced device address we might write.
5416 * 'readpos' is the least advanced device address we might read.
5417 * 'safepos' is the least address recorded in the metadata as having
5419 * If there is a min_offset_diff, these are adjusted either by
5420 * increasing the safepos/readpos if diff is negative, or
5421 * increasing writepos if diff is positive.
5422 * If 'readpos' is then behind 'writepos', there is no way that we can
5423 * ensure safety in the face of a crash - that must be done by userspace
5424 * making a backup of the data. So in that case there is no particular
5425 * rush to update metadata.
5426 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5427 * update the metadata to advance 'safepos' to match 'readpos' so that
5428 * we can be safe in the event of a crash.
5429 * So we insist on updating metadata if safepos is behind writepos and
5430 * readpos is beyond writepos.
5431 * In any case, update the metadata every 10 seconds.
5432 * Maybe that number should be configurable, but I'm not sure it is
5433 * worth it.... maybe it could be a multiple of safemode_delay???
5435 if (conf->min_offset_diff < 0) {
5436 safepos += -conf->min_offset_diff;
5437 readpos += -conf->min_offset_diff;
5439 writepos += conf->min_offset_diff;
5441 if ((mddev->reshape_backwards
5442 ? (safepos > writepos && readpos < writepos)
5443 : (safepos < writepos && readpos > writepos)) ||
5444 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5445 /* Cannot proceed until we've updated the superblock... */
5446 wait_event(conf->wait_for_overlap,
5447 atomic_read(&conf->reshape_stripes)==0
5448 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5449 if (atomic_read(&conf->reshape_stripes) != 0)
5451 mddev->reshape_position = conf->reshape_progress;
5452 mddev->curr_resync_completed = sector_nr;
5453 conf->reshape_checkpoint = jiffies;
5454 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5455 md_wakeup_thread(mddev->thread);
5456 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5457 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5458 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5460 spin_lock_irq(&conf->device_lock);
5461 conf->reshape_safe = mddev->reshape_position;
5462 spin_unlock_irq(&conf->device_lock);
5463 wake_up(&conf->wait_for_overlap);
5464 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5467 INIT_LIST_HEAD(&stripes);
5468 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5470 int skipped_disk = 0;
5471 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5472 set_bit(STRIPE_EXPANDING, &sh->state);
5473 atomic_inc(&conf->reshape_stripes);
5474 /* If any of this stripe is beyond the end of the old
5475 * array, then we need to zero those blocks
5477 for (j=sh->disks; j--;) {
5479 if (j == sh->pd_idx)
5481 if (conf->level == 6 &&
5484 s = raid5_compute_blocknr(sh, j, 0);
5485 if (s < raid5_size(mddev, 0, 0)) {
5489 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5490 set_bit(R5_Expanded, &sh->dev[j].flags);
5491 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5493 if (!skipped_disk) {
5494 set_bit(STRIPE_EXPAND_READY, &sh->state);
5495 set_bit(STRIPE_HANDLE, &sh->state);
5497 list_add(&sh->lru, &stripes);
5499 spin_lock_irq(&conf->device_lock);
5500 if (mddev->reshape_backwards)
5501 conf->reshape_progress -= reshape_sectors * new_data_disks;
5503 conf->reshape_progress += reshape_sectors * new_data_disks;
5504 spin_unlock_irq(&conf->device_lock);
5505 /* Ok, those stripe are ready. We can start scheduling
5506 * reads on the source stripes.
5507 * The source stripes are determined by mapping the first and last
5508 * block on the destination stripes.
5511 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5514 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5515 * new_data_disks - 1),
5517 if (last_sector >= mddev->dev_sectors)
5518 last_sector = mddev->dev_sectors - 1;
5519 while (first_sector <= last_sector) {
5520 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5521 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5522 set_bit(STRIPE_HANDLE, &sh->state);
5523 raid5_release_stripe(sh);
5524 first_sector += STRIPE_SECTORS;
5526 /* Now that the sources are clearly marked, we can release
5527 * the destination stripes
5529 while (!list_empty(&stripes)) {
5530 sh = list_entry(stripes.next, struct stripe_head, lru);
5531 list_del_init(&sh->lru);
5532 raid5_release_stripe(sh);
5534 /* If this takes us to the resync_max point where we have to pause,
5535 * then we need to write out the superblock.
5537 sector_nr += reshape_sectors;
5538 retn = reshape_sectors;
5540 if (mddev->curr_resync_completed > mddev->resync_max ||
5541 (sector_nr - mddev->curr_resync_completed) * 2
5542 >= mddev->resync_max - mddev->curr_resync_completed) {
5543 /* Cannot proceed until we've updated the superblock... */
5544 wait_event(conf->wait_for_overlap,
5545 atomic_read(&conf->reshape_stripes) == 0
5546 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5547 if (atomic_read(&conf->reshape_stripes) != 0)
5549 mddev->reshape_position = conf->reshape_progress;
5550 mddev->curr_resync_completed = sector_nr;
5551 conf->reshape_checkpoint = jiffies;
5552 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5553 md_wakeup_thread(mddev->thread);
5554 wait_event(mddev->sb_wait,
5555 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5556 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5557 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5559 spin_lock_irq(&conf->device_lock);
5560 conf->reshape_safe = mddev->reshape_position;
5561 spin_unlock_irq(&conf->device_lock);
5562 wake_up(&conf->wait_for_overlap);
5563 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5569 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5571 struct r5conf *conf = mddev->private;
5572 struct stripe_head *sh;
5573 sector_t max_sector = mddev->dev_sectors;
5574 sector_t sync_blocks;
5575 int still_degraded = 0;
5578 if (sector_nr >= max_sector) {
5579 /* just being told to finish up .. nothing much to do */
5581 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5586 if (mddev->curr_resync < max_sector) /* aborted */
5587 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5589 else /* completed sync */
5591 bitmap_close_sync(mddev->bitmap);
5596 /* Allow raid5_quiesce to complete */
5597 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5599 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5600 return reshape_request(mddev, sector_nr, skipped);
5602 /* No need to check resync_max as we never do more than one
5603 * stripe, and as resync_max will always be on a chunk boundary,
5604 * if the check in md_do_sync didn't fire, there is no chance
5605 * of overstepping resync_max here
5608 /* if there is too many failed drives and we are trying
5609 * to resync, then assert that we are finished, because there is
5610 * nothing we can do.
5612 if (mddev->degraded >= conf->max_degraded &&
5613 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5614 sector_t rv = mddev->dev_sectors - sector_nr;
5618 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5620 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5621 sync_blocks >= STRIPE_SECTORS) {
5622 /* we can skip this block, and probably more */
5623 sync_blocks /= STRIPE_SECTORS;
5625 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5628 bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5630 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5632 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5633 /* make sure we don't swamp the stripe cache if someone else
5634 * is trying to get access
5636 schedule_timeout_uninterruptible(1);
5638 /* Need to check if array will still be degraded after recovery/resync
5639 * Note in case of > 1 drive failures it's possible we're rebuilding
5640 * one drive while leaving another faulty drive in array.
5643 for (i = 0; i < conf->raid_disks; i++) {
5644 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5646 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5651 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5653 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5654 set_bit(STRIPE_HANDLE, &sh->state);
5656 raid5_release_stripe(sh);
5658 return STRIPE_SECTORS;
5661 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5663 /* We may not be able to submit a whole bio at once as there
5664 * may not be enough stripe_heads available.
5665 * We cannot pre-allocate enough stripe_heads as we may need
5666 * more than exist in the cache (if we allow ever large chunks).
5667 * So we do one stripe head at a time and record in
5668 * ->bi_hw_segments how many have been done.
5670 * We *know* that this entire raid_bio is in one chunk, so
5671 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5673 struct stripe_head *sh;
5675 sector_t sector, logical_sector, last_sector;
5680 logical_sector = raid_bio->bi_iter.bi_sector &
5681 ~((sector_t)STRIPE_SECTORS-1);
5682 sector = raid5_compute_sector(conf, logical_sector,
5684 last_sector = bio_end_sector(raid_bio);
5686 for (; logical_sector < last_sector;
5687 logical_sector += STRIPE_SECTORS,
5688 sector += STRIPE_SECTORS,
5691 if (scnt < raid5_bi_processed_stripes(raid_bio))
5692 /* already done this stripe */
5695 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5698 /* failed to get a stripe - must wait */
5699 raid5_set_bi_processed_stripes(raid_bio, scnt);
5700 conf->retry_read_aligned = raid_bio;
5704 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5705 raid5_release_stripe(sh);
5706 raid5_set_bi_processed_stripes(raid_bio, scnt);
5707 conf->retry_read_aligned = raid_bio;
5711 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5713 raid5_release_stripe(sh);
5716 remaining = raid5_dec_bi_active_stripes(raid_bio);
5717 if (remaining == 0) {
5718 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5720 bio_endio(raid_bio);
5722 if (atomic_dec_and_test(&conf->active_aligned_reads))
5723 wake_up(&conf->wait_for_quiescent);
5727 static int handle_active_stripes(struct r5conf *conf, int group,
5728 struct r5worker *worker,
5729 struct list_head *temp_inactive_list)
5731 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5732 int i, batch_size = 0, hash;
5733 bool release_inactive = false;
5735 while (batch_size < MAX_STRIPE_BATCH &&
5736 (sh = __get_priority_stripe(conf, group)) != NULL)
5737 batch[batch_size++] = sh;
5739 if (batch_size == 0) {
5740 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5741 if (!list_empty(temp_inactive_list + i))
5743 if (i == NR_STRIPE_HASH_LOCKS)
5745 release_inactive = true;
5747 spin_unlock_irq(&conf->device_lock);
5749 release_inactive_stripe_list(conf, temp_inactive_list,
5750 NR_STRIPE_HASH_LOCKS);
5752 if (release_inactive) {
5753 spin_lock_irq(&conf->device_lock);
5757 for (i = 0; i < batch_size; i++)
5758 handle_stripe(batch[i]);
5759 r5l_write_stripe_run(conf->log);
5763 spin_lock_irq(&conf->device_lock);
5764 for (i = 0; i < batch_size; i++) {
5765 hash = batch[i]->hash_lock_index;
5766 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5771 static void raid5_do_work(struct work_struct *work)
5773 struct r5worker *worker = container_of(work, struct r5worker, work);
5774 struct r5worker_group *group = worker->group;
5775 struct r5conf *conf = group->conf;
5776 int group_id = group - conf->worker_groups;
5778 struct blk_plug plug;
5780 pr_debug("+++ raid5worker active\n");
5782 blk_start_plug(&plug);
5784 spin_lock_irq(&conf->device_lock);
5786 int batch_size, released;
5788 released = release_stripe_list(conf, worker->temp_inactive_list);
5790 batch_size = handle_active_stripes(conf, group_id, worker,
5791 worker->temp_inactive_list);
5792 worker->working = false;
5793 if (!batch_size && !released)
5795 handled += batch_size;
5797 pr_debug("%d stripes handled\n", handled);
5799 spin_unlock_irq(&conf->device_lock);
5800 blk_finish_plug(&plug);
5802 pr_debug("--- raid5worker inactive\n");
5806 * This is our raid5 kernel thread.
5808 * We scan the hash table for stripes which can be handled now.
5809 * During the scan, completed stripes are saved for us by the interrupt
5810 * handler, so that they will not have to wait for our next wakeup.
5812 static void raid5d(struct md_thread *thread)
5814 struct mddev *mddev = thread->mddev;
5815 struct r5conf *conf = mddev->private;
5817 struct blk_plug plug;
5819 pr_debug("+++ raid5d active\n");
5821 md_check_recovery(mddev);
5823 if (!bio_list_empty(&conf->return_bi) &&
5824 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5825 struct bio_list tmp = BIO_EMPTY_LIST;
5826 spin_lock_irq(&conf->device_lock);
5827 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5828 bio_list_merge(&tmp, &conf->return_bi);
5829 bio_list_init(&conf->return_bi);
5831 spin_unlock_irq(&conf->device_lock);
5835 blk_start_plug(&plug);
5837 spin_lock_irq(&conf->device_lock);
5840 int batch_size, released;
5842 released = release_stripe_list(conf, conf->temp_inactive_list);
5844 clear_bit(R5_DID_ALLOC, &conf->cache_state);
5847 !list_empty(&conf->bitmap_list)) {
5848 /* Now is a good time to flush some bitmap updates */
5850 spin_unlock_irq(&conf->device_lock);
5851 bitmap_unplug(mddev->bitmap);
5852 spin_lock_irq(&conf->device_lock);
5853 conf->seq_write = conf->seq_flush;
5854 activate_bit_delay(conf, conf->temp_inactive_list);
5856 raid5_activate_delayed(conf);
5858 while ((bio = remove_bio_from_retry(conf))) {
5860 spin_unlock_irq(&conf->device_lock);
5861 ok = retry_aligned_read(conf, bio);
5862 spin_lock_irq(&conf->device_lock);
5868 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5869 conf->temp_inactive_list);
5870 if (!batch_size && !released)
5872 handled += batch_size;
5874 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5875 spin_unlock_irq(&conf->device_lock);
5876 md_check_recovery(mddev);
5877 spin_lock_irq(&conf->device_lock);
5880 pr_debug("%d stripes handled\n", handled);
5882 spin_unlock_irq(&conf->device_lock);
5883 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5884 mutex_trylock(&conf->cache_size_mutex)) {
5885 grow_one_stripe(conf, __GFP_NOWARN);
5886 /* Set flag even if allocation failed. This helps
5887 * slow down allocation requests when mem is short
5889 set_bit(R5_DID_ALLOC, &conf->cache_state);
5890 mutex_unlock(&conf->cache_size_mutex);
5893 r5l_flush_stripe_to_raid(conf->log);
5895 async_tx_issue_pending_all();
5896 blk_finish_plug(&plug);
5898 pr_debug("--- raid5d inactive\n");
5902 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5904 struct r5conf *conf;
5906 spin_lock(&mddev->lock);
5907 conf = mddev->private;
5909 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5910 spin_unlock(&mddev->lock);
5915 raid5_set_cache_size(struct mddev *mddev, int size)
5917 struct r5conf *conf = mddev->private;
5920 if (size <= 16 || size > 32768)
5923 conf->min_nr_stripes = size;
5924 mutex_lock(&conf->cache_size_mutex);
5925 while (size < conf->max_nr_stripes &&
5926 drop_one_stripe(conf))
5928 mutex_unlock(&conf->cache_size_mutex);
5931 err = md_allow_write(mddev);
5935 mutex_lock(&conf->cache_size_mutex);
5936 while (size > conf->max_nr_stripes)
5937 if (!grow_one_stripe(conf, GFP_KERNEL))
5939 mutex_unlock(&conf->cache_size_mutex);
5943 EXPORT_SYMBOL(raid5_set_cache_size);
5946 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5948 struct r5conf *conf;
5952 if (len >= PAGE_SIZE)
5954 if (kstrtoul(page, 10, &new))
5956 err = mddev_lock(mddev);
5959 conf = mddev->private;
5963 err = raid5_set_cache_size(mddev, new);
5964 mddev_unlock(mddev);
5969 static struct md_sysfs_entry
5970 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5971 raid5_show_stripe_cache_size,
5972 raid5_store_stripe_cache_size);
5975 raid5_show_rmw_level(struct mddev *mddev, char *page)
5977 struct r5conf *conf = mddev->private;
5979 return sprintf(page, "%d\n", conf->rmw_level);
5985 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
5987 struct r5conf *conf = mddev->private;
5993 if (len >= PAGE_SIZE)
5996 if (kstrtoul(page, 10, &new))
5999 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6002 if (new != PARITY_DISABLE_RMW &&
6003 new != PARITY_ENABLE_RMW &&
6004 new != PARITY_PREFER_RMW)
6007 conf->rmw_level = new;
6011 static struct md_sysfs_entry
6012 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6013 raid5_show_rmw_level,
6014 raid5_store_rmw_level);
6018 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6020 struct r5conf *conf;
6022 spin_lock(&mddev->lock);
6023 conf = mddev->private;
6025 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6026 spin_unlock(&mddev->lock);
6031 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6033 struct r5conf *conf;
6037 if (len >= PAGE_SIZE)
6039 if (kstrtoul(page, 10, &new))
6042 err = mddev_lock(mddev);
6045 conf = mddev->private;
6048 else if (new > conf->min_nr_stripes)
6051 conf->bypass_threshold = new;
6052 mddev_unlock(mddev);
6056 static struct md_sysfs_entry
6057 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6059 raid5_show_preread_threshold,
6060 raid5_store_preread_threshold);
6063 raid5_show_skip_copy(struct mddev *mddev, char *page)
6065 struct r5conf *conf;
6067 spin_lock(&mddev->lock);
6068 conf = mddev->private;
6070 ret = sprintf(page, "%d\n", conf->skip_copy);
6071 spin_unlock(&mddev->lock);
6076 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6078 struct r5conf *conf;
6082 if (len >= PAGE_SIZE)
6084 if (kstrtoul(page, 10, &new))
6088 err = mddev_lock(mddev);
6091 conf = mddev->private;
6094 else if (new != conf->skip_copy) {
6095 mddev_suspend(mddev);
6096 conf->skip_copy = new;
6098 mddev->queue->backing_dev_info.capabilities |=
6099 BDI_CAP_STABLE_WRITES;
6101 mddev->queue->backing_dev_info.capabilities &=
6102 ~BDI_CAP_STABLE_WRITES;
6103 mddev_resume(mddev);
6105 mddev_unlock(mddev);
6109 static struct md_sysfs_entry
6110 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6111 raid5_show_skip_copy,
6112 raid5_store_skip_copy);
6115 stripe_cache_active_show(struct mddev *mddev, char *page)
6117 struct r5conf *conf = mddev->private;
6119 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6124 static struct md_sysfs_entry
6125 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6128 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6130 struct r5conf *conf;
6132 spin_lock(&mddev->lock);
6133 conf = mddev->private;
6135 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6136 spin_unlock(&mddev->lock);
6140 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6142 int *worker_cnt_per_group,
6143 struct r5worker_group **worker_groups);
6145 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6147 struct r5conf *conf;
6150 struct r5worker_group *new_groups, *old_groups;
6151 int group_cnt, worker_cnt_per_group;
6153 if (len >= PAGE_SIZE)
6155 if (kstrtoul(page, 10, &new))
6158 err = mddev_lock(mddev);
6161 conf = mddev->private;
6164 else if (new != conf->worker_cnt_per_group) {
6165 mddev_suspend(mddev);
6167 old_groups = conf->worker_groups;
6169 flush_workqueue(raid5_wq);
6171 err = alloc_thread_groups(conf, new,
6172 &group_cnt, &worker_cnt_per_group,
6175 spin_lock_irq(&conf->device_lock);
6176 conf->group_cnt = group_cnt;
6177 conf->worker_cnt_per_group = worker_cnt_per_group;
6178 conf->worker_groups = new_groups;
6179 spin_unlock_irq(&conf->device_lock);
6182 kfree(old_groups[0].workers);
6185 mddev_resume(mddev);
6187 mddev_unlock(mddev);
6192 static struct md_sysfs_entry
6193 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6194 raid5_show_group_thread_cnt,
6195 raid5_store_group_thread_cnt);
6197 static struct attribute *raid5_attrs[] = {
6198 &raid5_stripecache_size.attr,
6199 &raid5_stripecache_active.attr,
6200 &raid5_preread_bypass_threshold.attr,
6201 &raid5_group_thread_cnt.attr,
6202 &raid5_skip_copy.attr,
6203 &raid5_rmw_level.attr,
6206 static struct attribute_group raid5_attrs_group = {
6208 .attrs = raid5_attrs,
6211 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6213 int *worker_cnt_per_group,
6214 struct r5worker_group **worker_groups)
6218 struct r5worker *workers;
6220 *worker_cnt_per_group = cnt;
6223 *worker_groups = NULL;
6226 *group_cnt = num_possible_nodes();
6227 size = sizeof(struct r5worker) * cnt;
6228 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6229 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6230 *group_cnt, GFP_NOIO);
6231 if (!*worker_groups || !workers) {
6233 kfree(*worker_groups);
6237 for (i = 0; i < *group_cnt; i++) {
6238 struct r5worker_group *group;
6240 group = &(*worker_groups)[i];
6241 INIT_LIST_HEAD(&group->handle_list);
6243 group->workers = workers + i * cnt;
6245 for (j = 0; j < cnt; j++) {
6246 struct r5worker *worker = group->workers + j;
6247 worker->group = group;
6248 INIT_WORK(&worker->work, raid5_do_work);
6250 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6251 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6258 static void free_thread_groups(struct r5conf *conf)
6260 if (conf->worker_groups)
6261 kfree(conf->worker_groups[0].workers);
6262 kfree(conf->worker_groups);
6263 conf->worker_groups = NULL;
6267 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6269 struct r5conf *conf = mddev->private;
6272 sectors = mddev->dev_sectors;
6274 /* size is defined by the smallest of previous and new size */
6275 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6277 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6278 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6279 return sectors * (raid_disks - conf->max_degraded);
6282 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6284 safe_put_page(percpu->spare_page);
6285 if (percpu->scribble)
6286 flex_array_free(percpu->scribble);
6287 percpu->spare_page = NULL;
6288 percpu->scribble = NULL;
6291 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6293 if (conf->level == 6 && !percpu->spare_page)
6294 percpu->spare_page = alloc_page(GFP_KERNEL);
6295 if (!percpu->scribble)
6296 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6297 conf->previous_raid_disks),
6298 max(conf->chunk_sectors,
6299 conf->prev_chunk_sectors)
6303 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6304 free_scratch_buffer(conf, percpu);
6311 static void raid5_free_percpu(struct r5conf *conf)
6318 #ifdef CONFIG_HOTPLUG_CPU
6319 unregister_cpu_notifier(&conf->cpu_notify);
6323 for_each_possible_cpu(cpu)
6324 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6327 free_percpu(conf->percpu);
6330 static void free_conf(struct r5conf *conf)
6333 r5l_exit_log(conf->log);
6334 if (conf->shrinker.seeks)
6335 unregister_shrinker(&conf->shrinker);
6337 free_thread_groups(conf);
6338 shrink_stripes(conf);
6339 raid5_free_percpu(conf);
6341 kfree(conf->stripe_hashtbl);
6345 #ifdef CONFIG_HOTPLUG_CPU
6346 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6349 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6350 long cpu = (long)hcpu;
6351 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6354 case CPU_UP_PREPARE:
6355 case CPU_UP_PREPARE_FROZEN:
6356 if (alloc_scratch_buffer(conf, percpu)) {
6357 pr_err("%s: failed memory allocation for cpu%ld\n",
6359 return notifier_from_errno(-ENOMEM);
6363 case CPU_DEAD_FROZEN:
6364 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6373 static int raid5_alloc_percpu(struct r5conf *conf)
6378 conf->percpu = alloc_percpu(struct raid5_percpu);
6382 #ifdef CONFIG_HOTPLUG_CPU
6383 conf->cpu_notify.notifier_call = raid456_cpu_notify;
6384 conf->cpu_notify.priority = 0;
6385 err = register_cpu_notifier(&conf->cpu_notify);
6391 for_each_present_cpu(cpu) {
6392 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6394 pr_err("%s: failed memory allocation for cpu%ld\n",
6404 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6405 struct shrink_control *sc)
6407 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6408 unsigned long ret = SHRINK_STOP;
6410 if (mutex_trylock(&conf->cache_size_mutex)) {
6412 while (ret < sc->nr_to_scan &&
6413 conf->max_nr_stripes > conf->min_nr_stripes) {
6414 if (drop_one_stripe(conf) == 0) {
6420 mutex_unlock(&conf->cache_size_mutex);
6425 static unsigned long raid5_cache_count(struct shrinker *shrink,
6426 struct shrink_control *sc)
6428 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6430 if (conf->max_nr_stripes < conf->min_nr_stripes)
6431 /* unlikely, but not impossible */
6433 return conf->max_nr_stripes - conf->min_nr_stripes;
6436 static struct r5conf *setup_conf(struct mddev *mddev)
6438 struct r5conf *conf;
6439 int raid_disk, memory, max_disks;
6440 struct md_rdev *rdev;
6441 struct disk_info *disk;
6444 int group_cnt, worker_cnt_per_group;
6445 struct r5worker_group *new_group;
6447 if (mddev->new_level != 5
6448 && mddev->new_level != 4
6449 && mddev->new_level != 6) {
6450 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6451 mdname(mddev), mddev->new_level);
6452 return ERR_PTR(-EIO);
6454 if ((mddev->new_level == 5
6455 && !algorithm_valid_raid5(mddev->new_layout)) ||
6456 (mddev->new_level == 6
6457 && !algorithm_valid_raid6(mddev->new_layout))) {
6458 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6459 mdname(mddev), mddev->new_layout);
6460 return ERR_PTR(-EIO);
6462 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6463 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6464 mdname(mddev), mddev->raid_disks);
6465 return ERR_PTR(-EINVAL);
6468 if (!mddev->new_chunk_sectors ||
6469 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6470 !is_power_of_2(mddev->new_chunk_sectors)) {
6471 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6472 mdname(mddev), mddev->new_chunk_sectors << 9);
6473 return ERR_PTR(-EINVAL);
6476 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6479 /* Don't enable multi-threading by default*/
6480 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6482 conf->group_cnt = group_cnt;
6483 conf->worker_cnt_per_group = worker_cnt_per_group;
6484 conf->worker_groups = new_group;
6487 spin_lock_init(&conf->device_lock);
6488 seqcount_init(&conf->gen_lock);
6489 mutex_init(&conf->cache_size_mutex);
6490 init_waitqueue_head(&conf->wait_for_quiescent);
6491 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
6492 init_waitqueue_head(&conf->wait_for_stripe[i]);
6494 init_waitqueue_head(&conf->wait_for_overlap);
6495 INIT_LIST_HEAD(&conf->handle_list);
6496 INIT_LIST_HEAD(&conf->hold_list);
6497 INIT_LIST_HEAD(&conf->delayed_list);
6498 INIT_LIST_HEAD(&conf->bitmap_list);
6499 bio_list_init(&conf->return_bi);
6500 init_llist_head(&conf->released_stripes);
6501 atomic_set(&conf->active_stripes, 0);
6502 atomic_set(&conf->preread_active_stripes, 0);
6503 atomic_set(&conf->active_aligned_reads, 0);
6504 conf->bypass_threshold = BYPASS_THRESHOLD;
6505 conf->recovery_disabled = mddev->recovery_disabled - 1;
6507 conf->raid_disks = mddev->raid_disks;
6508 if (mddev->reshape_position == MaxSector)
6509 conf->previous_raid_disks = mddev->raid_disks;
6511 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6512 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6514 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6519 conf->mddev = mddev;
6521 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6524 /* We init hash_locks[0] separately to that it can be used
6525 * as the reference lock in the spin_lock_nest_lock() call
6526 * in lock_all_device_hash_locks_irq in order to convince
6527 * lockdep that we know what we are doing.
6529 spin_lock_init(conf->hash_locks);
6530 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6531 spin_lock_init(conf->hash_locks + i);
6533 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6534 INIT_LIST_HEAD(conf->inactive_list + i);
6536 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6537 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6539 conf->level = mddev->new_level;
6540 conf->chunk_sectors = mddev->new_chunk_sectors;
6541 if (raid5_alloc_percpu(conf) != 0)
6544 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6546 rdev_for_each(rdev, mddev) {
6547 raid_disk = rdev->raid_disk;
6548 if (raid_disk >= max_disks
6551 disk = conf->disks + raid_disk;
6553 if (test_bit(Replacement, &rdev->flags)) {
6554 if (disk->replacement)
6556 disk->replacement = rdev;
6563 if (test_bit(In_sync, &rdev->flags)) {
6564 char b[BDEVNAME_SIZE];
6565 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6567 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6568 } else if (rdev->saved_raid_disk != raid_disk)
6569 /* Cannot rely on bitmap to complete recovery */
6573 conf->level = mddev->new_level;
6574 if (conf->level == 6) {
6575 conf->max_degraded = 2;
6576 if (raid6_call.xor_syndrome)
6577 conf->rmw_level = PARITY_ENABLE_RMW;
6579 conf->rmw_level = PARITY_DISABLE_RMW;
6581 conf->max_degraded = 1;
6582 conf->rmw_level = PARITY_ENABLE_RMW;
6584 conf->algorithm = mddev->new_layout;
6585 conf->reshape_progress = mddev->reshape_position;
6586 if (conf->reshape_progress != MaxSector) {
6587 conf->prev_chunk_sectors = mddev->chunk_sectors;
6588 conf->prev_algo = mddev->layout;
6590 conf->prev_chunk_sectors = conf->chunk_sectors;
6591 conf->prev_algo = conf->algorithm;
6594 conf->min_nr_stripes = NR_STRIPES;
6595 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6596 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6597 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6598 if (grow_stripes(conf, conf->min_nr_stripes)) {
6600 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6601 mdname(mddev), memory);
6604 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6605 mdname(mddev), memory);
6607 * Losing a stripe head costs more than the time to refill it,
6608 * it reduces the queue depth and so can hurt throughput.
6609 * So set it rather large, scaled by number of devices.
6611 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6612 conf->shrinker.scan_objects = raid5_cache_scan;
6613 conf->shrinker.count_objects = raid5_cache_count;
6614 conf->shrinker.batch = 128;
6615 conf->shrinker.flags = 0;
6616 register_shrinker(&conf->shrinker);
6618 sprintf(pers_name, "raid%d", mddev->new_level);
6619 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6620 if (!conf->thread) {
6622 "md/raid:%s: couldn't allocate thread.\n",
6632 return ERR_PTR(-EIO);
6634 return ERR_PTR(-ENOMEM);
6637 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6640 case ALGORITHM_PARITY_0:
6641 if (raid_disk < max_degraded)
6644 case ALGORITHM_PARITY_N:
6645 if (raid_disk >= raid_disks - max_degraded)
6648 case ALGORITHM_PARITY_0_6:
6649 if (raid_disk == 0 ||
6650 raid_disk == raid_disks - 1)
6653 case ALGORITHM_LEFT_ASYMMETRIC_6:
6654 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6655 case ALGORITHM_LEFT_SYMMETRIC_6:
6656 case ALGORITHM_RIGHT_SYMMETRIC_6:
6657 if (raid_disk == raid_disks - 1)
6663 static int run(struct mddev *mddev)
6665 struct r5conf *conf;
6666 int working_disks = 0;
6667 int dirty_parity_disks = 0;
6668 struct md_rdev *rdev;
6669 struct md_rdev *journal_dev = NULL;
6670 sector_t reshape_offset = 0;
6672 long long min_offset_diff = 0;
6675 if (mddev->recovery_cp != MaxSector)
6676 printk(KERN_NOTICE "md/raid:%s: not clean"
6677 " -- starting background reconstruction\n",
6680 rdev_for_each(rdev, mddev) {
6683 if (test_bit(Journal, &rdev->flags))
6685 if (rdev->raid_disk < 0)
6687 diff = (rdev->new_data_offset - rdev->data_offset);
6689 min_offset_diff = diff;
6691 } else if (mddev->reshape_backwards &&
6692 diff < min_offset_diff)
6693 min_offset_diff = diff;
6694 else if (!mddev->reshape_backwards &&
6695 diff > min_offset_diff)
6696 min_offset_diff = diff;
6699 if (mddev->reshape_position != MaxSector) {
6700 /* Check that we can continue the reshape.
6701 * Difficulties arise if the stripe we would write to
6702 * next is at or after the stripe we would read from next.
6703 * For a reshape that changes the number of devices, this
6704 * is only possible for a very short time, and mdadm makes
6705 * sure that time appears to have past before assembling
6706 * the array. So we fail if that time hasn't passed.
6707 * For a reshape that keeps the number of devices the same
6708 * mdadm must be monitoring the reshape can keeping the
6709 * critical areas read-only and backed up. It will start
6710 * the array in read-only mode, so we check for that.
6712 sector_t here_new, here_old;
6714 int max_degraded = (mddev->level == 6 ? 2 : 1);
6719 printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6724 if (mddev->new_level != mddev->level) {
6725 printk(KERN_ERR "md/raid:%s: unsupported reshape "
6726 "required - aborting.\n",
6730 old_disks = mddev->raid_disks - mddev->delta_disks;
6731 /* reshape_position must be on a new-stripe boundary, and one
6732 * further up in new geometry must map after here in old
6734 * If the chunk sizes are different, then as we perform reshape
6735 * in units of the largest of the two, reshape_position needs
6736 * be a multiple of the largest chunk size times new data disks.
6738 here_new = mddev->reshape_position;
6739 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6740 new_data_disks = mddev->raid_disks - max_degraded;
6741 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6742 printk(KERN_ERR "md/raid:%s: reshape_position not "
6743 "on a stripe boundary\n", mdname(mddev));
6746 reshape_offset = here_new * chunk_sectors;
6747 /* here_new is the stripe we will write to */
6748 here_old = mddev->reshape_position;
6749 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6750 /* here_old is the first stripe that we might need to read
6752 if (mddev->delta_disks == 0) {
6753 /* We cannot be sure it is safe to start an in-place
6754 * reshape. It is only safe if user-space is monitoring
6755 * and taking constant backups.
6756 * mdadm always starts a situation like this in
6757 * readonly mode so it can take control before
6758 * allowing any writes. So just check for that.
6760 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6761 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6762 /* not really in-place - so OK */;
6763 else if (mddev->ro == 0) {
6764 printk(KERN_ERR "md/raid:%s: in-place reshape "
6765 "must be started in read-only mode "
6770 } else if (mddev->reshape_backwards
6771 ? (here_new * chunk_sectors + min_offset_diff <=
6772 here_old * chunk_sectors)
6773 : (here_new * chunk_sectors >=
6774 here_old * chunk_sectors + (-min_offset_diff))) {
6775 /* Reading from the same stripe as writing to - bad */
6776 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6777 "auto-recovery - aborting.\n",
6781 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6783 /* OK, we should be able to continue; */
6785 BUG_ON(mddev->level != mddev->new_level);
6786 BUG_ON(mddev->layout != mddev->new_layout);
6787 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6788 BUG_ON(mddev->delta_disks != 0);
6791 if (mddev->private == NULL)
6792 conf = setup_conf(mddev);
6794 conf = mddev->private;
6797 return PTR_ERR(conf);
6799 conf->min_offset_diff = min_offset_diff;
6800 mddev->thread = conf->thread;
6801 conf->thread = NULL;
6802 mddev->private = conf;
6804 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6806 rdev = conf->disks[i].rdev;
6807 if (!rdev && conf->disks[i].replacement) {
6808 /* The replacement is all we have yet */
6809 rdev = conf->disks[i].replacement;
6810 conf->disks[i].replacement = NULL;
6811 clear_bit(Replacement, &rdev->flags);
6812 conf->disks[i].rdev = rdev;
6816 if (conf->disks[i].replacement &&
6817 conf->reshape_progress != MaxSector) {
6818 /* replacements and reshape simply do not mix. */
6819 printk(KERN_ERR "md: cannot handle concurrent "
6820 "replacement and reshape.\n");
6823 if (test_bit(In_sync, &rdev->flags)) {
6827 /* This disc is not fully in-sync. However if it
6828 * just stored parity (beyond the recovery_offset),
6829 * when we don't need to be concerned about the
6830 * array being dirty.
6831 * When reshape goes 'backwards', we never have
6832 * partially completed devices, so we only need
6833 * to worry about reshape going forwards.
6835 /* Hack because v0.91 doesn't store recovery_offset properly. */
6836 if (mddev->major_version == 0 &&
6837 mddev->minor_version > 90)
6838 rdev->recovery_offset = reshape_offset;
6840 if (rdev->recovery_offset < reshape_offset) {
6841 /* We need to check old and new layout */
6842 if (!only_parity(rdev->raid_disk,
6845 conf->max_degraded))
6848 if (!only_parity(rdev->raid_disk,
6850 conf->previous_raid_disks,
6851 conf->max_degraded))
6853 dirty_parity_disks++;
6857 * 0 for a fully functional array, 1 or 2 for a degraded array.
6859 mddev->degraded = calc_degraded(conf);
6861 if (has_failed(conf)) {
6862 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6863 " (%d/%d failed)\n",
6864 mdname(mddev), mddev->degraded, conf->raid_disks);
6868 /* device size must be a multiple of chunk size */
6869 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6870 mddev->resync_max_sectors = mddev->dev_sectors;
6872 if (mddev->degraded > dirty_parity_disks &&
6873 mddev->recovery_cp != MaxSector) {
6874 if (mddev->ok_start_degraded)
6876 "md/raid:%s: starting dirty degraded array"
6877 " - data corruption possible.\n",
6881 "md/raid:%s: cannot start dirty degraded array.\n",
6887 if (mddev->degraded == 0)
6888 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6889 " devices, algorithm %d\n", mdname(mddev), conf->level,
6890 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6893 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6894 " out of %d devices, algorithm %d\n",
6895 mdname(mddev), conf->level,
6896 mddev->raid_disks - mddev->degraded,
6897 mddev->raid_disks, mddev->new_layout);
6899 print_raid5_conf(conf);
6901 if (conf->reshape_progress != MaxSector) {
6902 conf->reshape_safe = conf->reshape_progress;
6903 atomic_set(&conf->reshape_stripes, 0);
6904 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6905 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6906 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6907 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6908 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6912 /* Ok, everything is just fine now */
6913 if (mddev->to_remove == &raid5_attrs_group)
6914 mddev->to_remove = NULL;
6915 else if (mddev->kobj.sd &&
6916 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6918 "raid5: failed to create sysfs attributes for %s\n",
6920 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6924 bool discard_supported = true;
6925 /* read-ahead size must cover two whole stripes, which
6926 * is 2 * (datadisks) * chunksize where 'n' is the
6927 * number of raid devices
6929 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6930 int stripe = data_disks *
6931 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6932 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6933 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6935 chunk_size = mddev->chunk_sectors << 9;
6936 blk_queue_io_min(mddev->queue, chunk_size);
6937 blk_queue_io_opt(mddev->queue, chunk_size *
6938 (conf->raid_disks - conf->max_degraded));
6939 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6941 * We can only discard a whole stripe. It doesn't make sense to
6942 * discard data disk but write parity disk
6944 stripe = stripe * PAGE_SIZE;
6945 /* Round up to power of 2, as discard handling
6946 * currently assumes that */
6947 while ((stripe-1) & stripe)
6948 stripe = (stripe | (stripe-1)) + 1;
6949 mddev->queue->limits.discard_alignment = stripe;
6950 mddev->queue->limits.discard_granularity = stripe;
6952 * unaligned part of discard request will be ignored, so can't
6953 * guarantee discard_zeroes_data
6955 mddev->queue->limits.discard_zeroes_data = 0;
6957 blk_queue_max_write_same_sectors(mddev->queue, 0);
6959 rdev_for_each(rdev, mddev) {
6960 disk_stack_limits(mddev->gendisk, rdev->bdev,
6961 rdev->data_offset << 9);
6962 disk_stack_limits(mddev->gendisk, rdev->bdev,
6963 rdev->new_data_offset << 9);
6965 * discard_zeroes_data is required, otherwise data
6966 * could be lost. Consider a scenario: discard a stripe
6967 * (the stripe could be inconsistent if
6968 * discard_zeroes_data is 0); write one disk of the
6969 * stripe (the stripe could be inconsistent again
6970 * depending on which disks are used to calculate
6971 * parity); the disk is broken; The stripe data of this
6974 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6975 !bdev_get_queue(rdev->bdev)->
6976 limits.discard_zeroes_data)
6977 discard_supported = false;
6978 /* Unfortunately, discard_zeroes_data is not currently
6979 * a guarantee - just a hint. So we only allow DISCARD
6980 * if the sysadmin has confirmed that only safe devices
6981 * are in use by setting a module parameter.
6983 if (!devices_handle_discard_safely) {
6984 if (discard_supported) {
6985 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6986 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6988 discard_supported = false;
6992 if (discard_supported &&
6993 mddev->queue->limits.max_discard_sectors >= stripe &&
6994 mddev->queue->limits.discard_granularity >= stripe)
6995 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6998 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7003 char b[BDEVNAME_SIZE];
7005 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7006 mdname(mddev), bdevname(journal_dev->bdev, b));
7007 r5l_init_log(conf, journal_dev);
7012 md_unregister_thread(&mddev->thread);
7013 print_raid5_conf(conf);
7015 mddev->private = NULL;
7016 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7020 static void raid5_free(struct mddev *mddev, void *priv)
7022 struct r5conf *conf = priv;
7025 mddev->to_remove = &raid5_attrs_group;
7028 static void status(struct seq_file *seq, struct mddev *mddev)
7030 struct r5conf *conf = mddev->private;
7033 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7034 conf->chunk_sectors / 2, mddev->layout);
7035 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7036 for (i = 0; i < conf->raid_disks; i++)
7037 seq_printf (seq, "%s",
7038 conf->disks[i].rdev &&
7039 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
7040 seq_printf (seq, "]");
7043 static void print_raid5_conf (struct r5conf *conf)
7046 struct disk_info *tmp;
7048 printk(KERN_DEBUG "RAID conf printout:\n");
7050 printk("(conf==NULL)\n");
7053 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7055 conf->raid_disks - conf->mddev->degraded);
7057 for (i = 0; i < conf->raid_disks; i++) {
7058 char b[BDEVNAME_SIZE];
7059 tmp = conf->disks + i;
7061 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7062 i, !test_bit(Faulty, &tmp->rdev->flags),
7063 bdevname(tmp->rdev->bdev, b));
7067 static int raid5_spare_active(struct mddev *mddev)
7070 struct r5conf *conf = mddev->private;
7071 struct disk_info *tmp;
7073 unsigned long flags;
7075 for (i = 0; i < conf->raid_disks; i++) {
7076 tmp = conf->disks + i;
7077 if (tmp->replacement
7078 && tmp->replacement->recovery_offset == MaxSector
7079 && !test_bit(Faulty, &tmp->replacement->flags)
7080 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7081 /* Replacement has just become active. */
7083 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7086 /* Replaced device not technically faulty,
7087 * but we need to be sure it gets removed
7088 * and never re-added.
7090 set_bit(Faulty, &tmp->rdev->flags);
7091 sysfs_notify_dirent_safe(
7092 tmp->rdev->sysfs_state);
7094 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7095 } else if (tmp->rdev
7096 && tmp->rdev->recovery_offset == MaxSector
7097 && !test_bit(Faulty, &tmp->rdev->flags)
7098 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7100 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7103 spin_lock_irqsave(&conf->device_lock, flags);
7104 mddev->degraded = calc_degraded(conf);
7105 spin_unlock_irqrestore(&conf->device_lock, flags);
7106 print_raid5_conf(conf);
7110 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7112 struct r5conf *conf = mddev->private;
7114 int number = rdev->raid_disk;
7115 struct md_rdev **rdevp;
7116 struct disk_info *p = conf->disks + number;
7118 print_raid5_conf(conf);
7119 if (rdev == p->rdev)
7121 else if (rdev == p->replacement)
7122 rdevp = &p->replacement;
7126 if (number >= conf->raid_disks &&
7127 conf->reshape_progress == MaxSector)
7128 clear_bit(In_sync, &rdev->flags);
7130 if (test_bit(In_sync, &rdev->flags) ||
7131 atomic_read(&rdev->nr_pending)) {
7135 /* Only remove non-faulty devices if recovery
7138 if (!test_bit(Faulty, &rdev->flags) &&
7139 mddev->recovery_disabled != conf->recovery_disabled &&
7140 !has_failed(conf) &&
7141 (!p->replacement || p->replacement == rdev) &&
7142 number < conf->raid_disks) {
7148 if (atomic_read(&rdev->nr_pending)) {
7149 /* lost the race, try later */
7152 } else if (p->replacement) {
7153 /* We must have just cleared 'rdev' */
7154 p->rdev = p->replacement;
7155 clear_bit(Replacement, &p->replacement->flags);
7156 smp_mb(); /* Make sure other CPUs may see both as identical
7157 * but will never see neither - if they are careful
7159 p->replacement = NULL;
7160 clear_bit(WantReplacement, &rdev->flags);
7162 /* We might have just removed the Replacement as faulty-
7163 * clear the bit just in case
7165 clear_bit(WantReplacement, &rdev->flags);
7168 print_raid5_conf(conf);
7172 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7174 struct r5conf *conf = mddev->private;
7177 struct disk_info *p;
7179 int last = conf->raid_disks - 1;
7181 if (mddev->recovery_disabled == conf->recovery_disabled)
7184 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7185 /* no point adding a device */
7188 if (rdev->raid_disk >= 0)
7189 first = last = rdev->raid_disk;
7192 * find the disk ... but prefer rdev->saved_raid_disk
7195 if (rdev->saved_raid_disk >= 0 &&
7196 rdev->saved_raid_disk >= first &&
7197 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7198 first = rdev->saved_raid_disk;
7200 for (disk = first; disk <= last; disk++) {
7201 p = conf->disks + disk;
7202 if (p->rdev == NULL) {
7203 clear_bit(In_sync, &rdev->flags);
7204 rdev->raid_disk = disk;
7206 if (rdev->saved_raid_disk != disk)
7208 rcu_assign_pointer(p->rdev, rdev);
7212 for (disk = first; disk <= last; disk++) {
7213 p = conf->disks + disk;
7214 if (test_bit(WantReplacement, &p->rdev->flags) &&
7215 p->replacement == NULL) {
7216 clear_bit(In_sync, &rdev->flags);
7217 set_bit(Replacement, &rdev->flags);
7218 rdev->raid_disk = disk;
7221 rcu_assign_pointer(p->replacement, rdev);
7226 print_raid5_conf(conf);
7230 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7232 /* no resync is happening, and there is enough space
7233 * on all devices, so we can resize.
7234 * We need to make sure resync covers any new space.
7235 * If the array is shrinking we should possibly wait until
7236 * any io in the removed space completes, but it hardly seems
7240 struct r5conf *conf = mddev->private;
7244 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7245 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7246 if (mddev->external_size &&
7247 mddev->array_sectors > newsize)
7249 if (mddev->bitmap) {
7250 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7254 md_set_array_sectors(mddev, newsize);
7255 set_capacity(mddev->gendisk, mddev->array_sectors);
7256 revalidate_disk(mddev->gendisk);
7257 if (sectors > mddev->dev_sectors &&
7258 mddev->recovery_cp > mddev->dev_sectors) {
7259 mddev->recovery_cp = mddev->dev_sectors;
7260 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7262 mddev->dev_sectors = sectors;
7263 mddev->resync_max_sectors = sectors;
7267 static int check_stripe_cache(struct mddev *mddev)
7269 /* Can only proceed if there are plenty of stripe_heads.
7270 * We need a minimum of one full stripe,, and for sensible progress
7271 * it is best to have about 4 times that.
7272 * If we require 4 times, then the default 256 4K stripe_heads will
7273 * allow for chunk sizes up to 256K, which is probably OK.
7274 * If the chunk size is greater, user-space should request more
7275 * stripe_heads first.
7277 struct r5conf *conf = mddev->private;
7278 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7279 > conf->min_nr_stripes ||
7280 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7281 > conf->min_nr_stripes) {
7282 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7284 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7291 static int check_reshape(struct mddev *mddev)
7293 struct r5conf *conf = mddev->private;
7297 if (mddev->delta_disks == 0 &&
7298 mddev->new_layout == mddev->layout &&
7299 mddev->new_chunk_sectors == mddev->chunk_sectors)
7300 return 0; /* nothing to do */
7301 if (has_failed(conf))
7303 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7304 /* We might be able to shrink, but the devices must
7305 * be made bigger first.
7306 * For raid6, 4 is the minimum size.
7307 * Otherwise 2 is the minimum
7310 if (mddev->level == 6)
7312 if (mddev->raid_disks + mddev->delta_disks < min)
7316 if (!check_stripe_cache(mddev))
7319 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7320 mddev->delta_disks > 0)
7321 if (resize_chunks(conf,
7322 conf->previous_raid_disks
7323 + max(0, mddev->delta_disks),
7324 max(mddev->new_chunk_sectors,
7325 mddev->chunk_sectors)
7328 return resize_stripes(conf, (conf->previous_raid_disks
7329 + mddev->delta_disks));
7332 static int raid5_start_reshape(struct mddev *mddev)
7334 struct r5conf *conf = mddev->private;
7335 struct md_rdev *rdev;
7337 unsigned long flags;
7339 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7342 if (!check_stripe_cache(mddev))
7345 if (has_failed(conf))
7348 rdev_for_each(rdev, mddev) {
7349 if (!test_bit(In_sync, &rdev->flags)
7350 && !test_bit(Faulty, &rdev->flags))
7354 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7355 /* Not enough devices even to make a degraded array
7360 /* Refuse to reduce size of the array. Any reductions in
7361 * array size must be through explicit setting of array_size
7364 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7365 < mddev->array_sectors) {
7366 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7367 "before number of disks\n", mdname(mddev));
7371 atomic_set(&conf->reshape_stripes, 0);
7372 spin_lock_irq(&conf->device_lock);
7373 write_seqcount_begin(&conf->gen_lock);
7374 conf->previous_raid_disks = conf->raid_disks;
7375 conf->raid_disks += mddev->delta_disks;
7376 conf->prev_chunk_sectors = conf->chunk_sectors;
7377 conf->chunk_sectors = mddev->new_chunk_sectors;
7378 conf->prev_algo = conf->algorithm;
7379 conf->algorithm = mddev->new_layout;
7381 /* Code that selects data_offset needs to see the generation update
7382 * if reshape_progress has been set - so a memory barrier needed.
7385 if (mddev->reshape_backwards)
7386 conf->reshape_progress = raid5_size(mddev, 0, 0);
7388 conf->reshape_progress = 0;
7389 conf->reshape_safe = conf->reshape_progress;
7390 write_seqcount_end(&conf->gen_lock);
7391 spin_unlock_irq(&conf->device_lock);
7393 /* Now make sure any requests that proceeded on the assumption
7394 * the reshape wasn't running - like Discard or Read - have
7397 mddev_suspend(mddev);
7398 mddev_resume(mddev);
7400 /* Add some new drives, as many as will fit.
7401 * We know there are enough to make the newly sized array work.
7402 * Don't add devices if we are reducing the number of
7403 * devices in the array. This is because it is not possible
7404 * to correctly record the "partially reconstructed" state of
7405 * such devices during the reshape and confusion could result.
7407 if (mddev->delta_disks >= 0) {
7408 rdev_for_each(rdev, mddev)
7409 if (rdev->raid_disk < 0 &&
7410 !test_bit(Faulty, &rdev->flags)) {
7411 if (raid5_add_disk(mddev, rdev) == 0) {
7413 >= conf->previous_raid_disks)
7414 set_bit(In_sync, &rdev->flags);
7416 rdev->recovery_offset = 0;
7418 if (sysfs_link_rdev(mddev, rdev))
7419 /* Failure here is OK */;
7421 } else if (rdev->raid_disk >= conf->previous_raid_disks
7422 && !test_bit(Faulty, &rdev->flags)) {
7423 /* This is a spare that was manually added */
7424 set_bit(In_sync, &rdev->flags);
7427 /* When a reshape changes the number of devices,
7428 * ->degraded is measured against the larger of the
7429 * pre and post number of devices.
7431 spin_lock_irqsave(&conf->device_lock, flags);
7432 mddev->degraded = calc_degraded(conf);
7433 spin_unlock_irqrestore(&conf->device_lock, flags);
7435 mddev->raid_disks = conf->raid_disks;
7436 mddev->reshape_position = conf->reshape_progress;
7437 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7439 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7440 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7441 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7442 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7443 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7444 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7446 if (!mddev->sync_thread) {
7447 mddev->recovery = 0;
7448 spin_lock_irq(&conf->device_lock);
7449 write_seqcount_begin(&conf->gen_lock);
7450 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7451 mddev->new_chunk_sectors =
7452 conf->chunk_sectors = conf->prev_chunk_sectors;
7453 mddev->new_layout = conf->algorithm = conf->prev_algo;
7454 rdev_for_each(rdev, mddev)
7455 rdev->new_data_offset = rdev->data_offset;
7457 conf->generation --;
7458 conf->reshape_progress = MaxSector;
7459 mddev->reshape_position = MaxSector;
7460 write_seqcount_end(&conf->gen_lock);
7461 spin_unlock_irq(&conf->device_lock);
7464 conf->reshape_checkpoint = jiffies;
7465 md_wakeup_thread(mddev->sync_thread);
7466 md_new_event(mddev);
7470 /* This is called from the reshape thread and should make any
7471 * changes needed in 'conf'
7473 static void end_reshape(struct r5conf *conf)
7476 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7477 struct md_rdev *rdev;
7479 spin_lock_irq(&conf->device_lock);
7480 conf->previous_raid_disks = conf->raid_disks;
7481 rdev_for_each(rdev, conf->mddev)
7482 rdev->data_offset = rdev->new_data_offset;
7484 conf->reshape_progress = MaxSector;
7485 conf->mddev->reshape_position = MaxSector;
7486 spin_unlock_irq(&conf->device_lock);
7487 wake_up(&conf->wait_for_overlap);
7489 /* read-ahead size must cover two whole stripes, which is
7490 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7492 if (conf->mddev->queue) {
7493 int data_disks = conf->raid_disks - conf->max_degraded;
7494 int stripe = data_disks * ((conf->chunk_sectors << 9)
7496 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7497 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7502 /* This is called from the raid5d thread with mddev_lock held.
7503 * It makes config changes to the device.
7505 static void raid5_finish_reshape(struct mddev *mddev)
7507 struct r5conf *conf = mddev->private;
7509 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7511 if (mddev->delta_disks > 0) {
7512 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7513 set_capacity(mddev->gendisk, mddev->array_sectors);
7514 revalidate_disk(mddev->gendisk);
7517 spin_lock_irq(&conf->device_lock);
7518 mddev->degraded = calc_degraded(conf);
7519 spin_unlock_irq(&conf->device_lock);
7520 for (d = conf->raid_disks ;
7521 d < conf->raid_disks - mddev->delta_disks;
7523 struct md_rdev *rdev = conf->disks[d].rdev;
7525 clear_bit(In_sync, &rdev->flags);
7526 rdev = conf->disks[d].replacement;
7528 clear_bit(In_sync, &rdev->flags);
7531 mddev->layout = conf->algorithm;
7532 mddev->chunk_sectors = conf->chunk_sectors;
7533 mddev->reshape_position = MaxSector;
7534 mddev->delta_disks = 0;
7535 mddev->reshape_backwards = 0;
7539 static void raid5_quiesce(struct mddev *mddev, int state)
7541 struct r5conf *conf = mddev->private;
7544 case 2: /* resume for a suspend */
7545 wake_up(&conf->wait_for_overlap);
7548 case 1: /* stop all writes */
7549 lock_all_device_hash_locks_irq(conf);
7550 /* '2' tells resync/reshape to pause so that all
7551 * active stripes can drain
7554 wait_event_cmd(conf->wait_for_quiescent,
7555 atomic_read(&conf->active_stripes) == 0 &&
7556 atomic_read(&conf->active_aligned_reads) == 0,
7557 unlock_all_device_hash_locks_irq(conf),
7558 lock_all_device_hash_locks_irq(conf));
7560 unlock_all_device_hash_locks_irq(conf);
7561 /* allow reshape to continue */
7562 wake_up(&conf->wait_for_overlap);
7565 case 0: /* re-enable writes */
7566 lock_all_device_hash_locks_irq(conf);
7568 wake_up(&conf->wait_for_quiescent);
7569 wake_up(&conf->wait_for_overlap);
7570 unlock_all_device_hash_locks_irq(conf);
7575 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7577 struct r0conf *raid0_conf = mddev->private;
7580 /* for raid0 takeover only one zone is supported */
7581 if (raid0_conf->nr_strip_zones > 1) {
7582 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7584 return ERR_PTR(-EINVAL);
7587 sectors = raid0_conf->strip_zone[0].zone_end;
7588 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7589 mddev->dev_sectors = sectors;
7590 mddev->new_level = level;
7591 mddev->new_layout = ALGORITHM_PARITY_N;
7592 mddev->new_chunk_sectors = mddev->chunk_sectors;
7593 mddev->raid_disks += 1;
7594 mddev->delta_disks = 1;
7595 /* make sure it will be not marked as dirty */
7596 mddev->recovery_cp = MaxSector;
7598 return setup_conf(mddev);
7601 static void *raid5_takeover_raid1(struct mddev *mddev)
7605 if (mddev->raid_disks != 2 ||
7606 mddev->degraded > 1)
7607 return ERR_PTR(-EINVAL);
7609 /* Should check if there are write-behind devices? */
7611 chunksect = 64*2; /* 64K by default */
7613 /* The array must be an exact multiple of chunksize */
7614 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7617 if ((chunksect<<9) < STRIPE_SIZE)
7618 /* array size does not allow a suitable chunk size */
7619 return ERR_PTR(-EINVAL);
7621 mddev->new_level = 5;
7622 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7623 mddev->new_chunk_sectors = chunksect;
7625 return setup_conf(mddev);
7628 static void *raid5_takeover_raid6(struct mddev *mddev)
7632 switch (mddev->layout) {
7633 case ALGORITHM_LEFT_ASYMMETRIC_6:
7634 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7636 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7637 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7639 case ALGORITHM_LEFT_SYMMETRIC_6:
7640 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7642 case ALGORITHM_RIGHT_SYMMETRIC_6:
7643 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7645 case ALGORITHM_PARITY_0_6:
7646 new_layout = ALGORITHM_PARITY_0;
7648 case ALGORITHM_PARITY_N:
7649 new_layout = ALGORITHM_PARITY_N;
7652 return ERR_PTR(-EINVAL);
7654 mddev->new_level = 5;
7655 mddev->new_layout = new_layout;
7656 mddev->delta_disks = -1;
7657 mddev->raid_disks -= 1;
7658 return setup_conf(mddev);
7661 static int raid5_check_reshape(struct mddev *mddev)
7663 /* For a 2-drive array, the layout and chunk size can be changed
7664 * immediately as not restriping is needed.
7665 * For larger arrays we record the new value - after validation
7666 * to be used by a reshape pass.
7668 struct r5conf *conf = mddev->private;
7669 int new_chunk = mddev->new_chunk_sectors;
7671 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7673 if (new_chunk > 0) {
7674 if (!is_power_of_2(new_chunk))
7676 if (new_chunk < (PAGE_SIZE>>9))
7678 if (mddev->array_sectors & (new_chunk-1))
7679 /* not factor of array size */
7683 /* They look valid */
7685 if (mddev->raid_disks == 2) {
7686 /* can make the change immediately */
7687 if (mddev->new_layout >= 0) {
7688 conf->algorithm = mddev->new_layout;
7689 mddev->layout = mddev->new_layout;
7691 if (new_chunk > 0) {
7692 conf->chunk_sectors = new_chunk ;
7693 mddev->chunk_sectors = new_chunk;
7695 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7696 md_wakeup_thread(mddev->thread);
7698 return check_reshape(mddev);
7701 static int raid6_check_reshape(struct mddev *mddev)
7703 int new_chunk = mddev->new_chunk_sectors;
7705 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7707 if (new_chunk > 0) {
7708 if (!is_power_of_2(new_chunk))
7710 if (new_chunk < (PAGE_SIZE >> 9))
7712 if (mddev->array_sectors & (new_chunk-1))
7713 /* not factor of array size */
7717 /* They look valid */
7718 return check_reshape(mddev);
7721 static void *raid5_takeover(struct mddev *mddev)
7723 /* raid5 can take over:
7724 * raid0 - if there is only one strip zone - make it a raid4 layout
7725 * raid1 - if there are two drives. We need to know the chunk size
7726 * raid4 - trivial - just use a raid4 layout.
7727 * raid6 - Providing it is a *_6 layout
7729 if (mddev->level == 0)
7730 return raid45_takeover_raid0(mddev, 5);
7731 if (mddev->level == 1)
7732 return raid5_takeover_raid1(mddev);
7733 if (mddev->level == 4) {
7734 mddev->new_layout = ALGORITHM_PARITY_N;
7735 mddev->new_level = 5;
7736 return setup_conf(mddev);
7738 if (mddev->level == 6)
7739 return raid5_takeover_raid6(mddev);
7741 return ERR_PTR(-EINVAL);
7744 static void *raid4_takeover(struct mddev *mddev)
7746 /* raid4 can take over:
7747 * raid0 - if there is only one strip zone
7748 * raid5 - if layout is right
7750 if (mddev->level == 0)
7751 return raid45_takeover_raid0(mddev, 4);
7752 if (mddev->level == 5 &&
7753 mddev->layout == ALGORITHM_PARITY_N) {
7754 mddev->new_layout = 0;
7755 mddev->new_level = 4;
7756 return setup_conf(mddev);
7758 return ERR_PTR(-EINVAL);
7761 static struct md_personality raid5_personality;
7763 static void *raid6_takeover(struct mddev *mddev)
7765 /* Currently can only take over a raid5. We map the
7766 * personality to an equivalent raid6 personality
7767 * with the Q block at the end.
7771 if (mddev->pers != &raid5_personality)
7772 return ERR_PTR(-EINVAL);
7773 if (mddev->degraded > 1)
7774 return ERR_PTR(-EINVAL);
7775 if (mddev->raid_disks > 253)
7776 return ERR_PTR(-EINVAL);
7777 if (mddev->raid_disks < 3)
7778 return ERR_PTR(-EINVAL);
7780 switch (mddev->layout) {
7781 case ALGORITHM_LEFT_ASYMMETRIC:
7782 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7784 case ALGORITHM_RIGHT_ASYMMETRIC:
7785 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7787 case ALGORITHM_LEFT_SYMMETRIC:
7788 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7790 case ALGORITHM_RIGHT_SYMMETRIC:
7791 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7793 case ALGORITHM_PARITY_0:
7794 new_layout = ALGORITHM_PARITY_0_6;
7796 case ALGORITHM_PARITY_N:
7797 new_layout = ALGORITHM_PARITY_N;
7800 return ERR_PTR(-EINVAL);
7802 mddev->new_level = 6;
7803 mddev->new_layout = new_layout;
7804 mddev->delta_disks = 1;
7805 mddev->raid_disks += 1;
7806 return setup_conf(mddev);
7809 static struct md_personality raid6_personality =
7813 .owner = THIS_MODULE,
7814 .make_request = make_request,
7818 .error_handler = error,
7819 .hot_add_disk = raid5_add_disk,
7820 .hot_remove_disk= raid5_remove_disk,
7821 .spare_active = raid5_spare_active,
7822 .sync_request = sync_request,
7823 .resize = raid5_resize,
7825 .check_reshape = raid6_check_reshape,
7826 .start_reshape = raid5_start_reshape,
7827 .finish_reshape = raid5_finish_reshape,
7828 .quiesce = raid5_quiesce,
7829 .takeover = raid6_takeover,
7830 .congested = raid5_congested,
7832 static struct md_personality raid5_personality =
7836 .owner = THIS_MODULE,
7837 .make_request = make_request,
7841 .error_handler = error,
7842 .hot_add_disk = raid5_add_disk,
7843 .hot_remove_disk= raid5_remove_disk,
7844 .spare_active = raid5_spare_active,
7845 .sync_request = sync_request,
7846 .resize = raid5_resize,
7848 .check_reshape = raid5_check_reshape,
7849 .start_reshape = raid5_start_reshape,
7850 .finish_reshape = raid5_finish_reshape,
7851 .quiesce = raid5_quiesce,
7852 .takeover = raid5_takeover,
7853 .congested = raid5_congested,
7856 static struct md_personality raid4_personality =
7860 .owner = THIS_MODULE,
7861 .make_request = make_request,
7865 .error_handler = error,
7866 .hot_add_disk = raid5_add_disk,
7867 .hot_remove_disk= raid5_remove_disk,
7868 .spare_active = raid5_spare_active,
7869 .sync_request = sync_request,
7870 .resize = raid5_resize,
7872 .check_reshape = raid5_check_reshape,
7873 .start_reshape = raid5_start_reshape,
7874 .finish_reshape = raid5_finish_reshape,
7875 .quiesce = raid5_quiesce,
7876 .takeover = raid4_takeover,
7877 .congested = raid5_congested,
7880 static int __init raid5_init(void)
7882 raid5_wq = alloc_workqueue("raid5wq",
7883 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7886 register_md_personality(&raid6_personality);
7887 register_md_personality(&raid5_personality);
7888 register_md_personality(&raid4_personality);
7892 static void raid5_exit(void)
7894 unregister_md_personality(&raid6_personality);
7895 unregister_md_personality(&raid5_personality);
7896 unregister_md_personality(&raid4_personality);
7897 destroy_workqueue(raid5_wq);
7900 module_init(raid5_init);
7901 module_exit(raid5_exit);
7902 MODULE_LICENSE("GPL");
7903 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7904 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7905 MODULE_ALIAS("md-raid5");
7906 MODULE_ALIAS("md-raid4");
7907 MODULE_ALIAS("md-level-5");
7908 MODULE_ALIAS("md-level-4");
7909 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7910 MODULE_ALIAS("md-raid6");
7911 MODULE_ALIAS("md-level-6");
7913 /* This used to be two separate modules, they were: */
7914 MODULE_ALIAS("raid5");
7915 MODULE_ALIAS("raid6");