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 <linux/sched/signal.h>
60 #include <trace/events/block.h>
67 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
69 #define cpu_to_group(cpu) cpu_to_node(cpu)
70 #define ANY_GROUP NUMA_NO_NODE
72 static bool devices_handle_discard_safely = false;
73 module_param(devices_handle_discard_safely, bool, 0644);
74 MODULE_PARM_DESC(devices_handle_discard_safely,
75 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
76 static struct workqueue_struct *raid5_wq;
78 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
80 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
81 return &conf->stripe_hashtbl[hash];
84 static inline int stripe_hash_locks_hash(sector_t sect)
86 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
89 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
91 spin_lock_irq(conf->hash_locks + hash);
92 spin_lock(&conf->device_lock);
95 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
97 spin_unlock(&conf->device_lock);
98 spin_unlock_irq(conf->hash_locks + hash);
101 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
105 spin_lock(conf->hash_locks);
106 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
107 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
108 spin_lock(&conf->device_lock);
111 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
114 spin_unlock(&conf->device_lock);
115 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
116 spin_unlock(conf->hash_locks + i - 1);
120 /* Find first data disk in a raid6 stripe */
121 static inline int raid6_d0(struct stripe_head *sh)
124 /* ddf always start from first device */
126 /* md starts just after Q block */
127 if (sh->qd_idx == sh->disks - 1)
130 return sh->qd_idx + 1;
132 static inline int raid6_next_disk(int disk, int raid_disks)
135 return (disk < raid_disks) ? disk : 0;
138 /* When walking through the disks in a raid5, starting at raid6_d0,
139 * We need to map each disk to a 'slot', where the data disks are slot
140 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
141 * is raid_disks-1. This help does that mapping.
143 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
144 int *count, int syndrome_disks)
150 if (idx == sh->pd_idx)
151 return syndrome_disks;
152 if (idx == sh->qd_idx)
153 return syndrome_disks + 1;
159 static void return_io(struct bio_list *return_bi)
162 while ((bi = bio_list_pop(return_bi)) != NULL) {
163 bi->bi_iter.bi_size = 0;
164 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
170 static void print_raid5_conf (struct r5conf *conf);
172 static int stripe_operations_active(struct stripe_head *sh)
174 return sh->check_state || sh->reconstruct_state ||
175 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
176 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
179 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
181 struct r5conf *conf = sh->raid_conf;
182 struct r5worker_group *group;
184 int i, cpu = sh->cpu;
186 if (!cpu_online(cpu)) {
187 cpu = cpumask_any(cpu_online_mask);
191 if (list_empty(&sh->lru)) {
192 struct r5worker_group *group;
193 group = conf->worker_groups + cpu_to_group(cpu);
194 list_add_tail(&sh->lru, &group->handle_list);
195 group->stripes_cnt++;
199 if (conf->worker_cnt_per_group == 0) {
200 md_wakeup_thread(conf->mddev->thread);
204 group = conf->worker_groups + cpu_to_group(sh->cpu);
206 group->workers[0].working = true;
207 /* at least one worker should run to avoid race */
208 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
210 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
211 /* wakeup more workers */
212 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
213 if (group->workers[i].working == false) {
214 group->workers[i].working = true;
215 queue_work_on(sh->cpu, raid5_wq,
216 &group->workers[i].work);
222 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
223 struct list_head *temp_inactive_list)
226 int injournal = 0; /* number of date pages with R5_InJournal */
228 BUG_ON(!list_empty(&sh->lru));
229 BUG_ON(atomic_read(&conf->active_stripes)==0);
231 if (r5c_is_writeback(conf->log))
232 for (i = sh->disks; i--; )
233 if (test_bit(R5_InJournal, &sh->dev[i].flags))
236 * When quiesce in r5c write back, set STRIPE_HANDLE for stripes with
237 * data in journal, so they are not released to cached lists
239 if (conf->quiesce && r5c_is_writeback(conf->log) &&
240 !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0) {
241 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
242 r5c_make_stripe_write_out(sh);
243 set_bit(STRIPE_HANDLE, &sh->state);
246 if (test_bit(STRIPE_HANDLE, &sh->state)) {
247 if (test_bit(STRIPE_DELAYED, &sh->state) &&
248 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
249 list_add_tail(&sh->lru, &conf->delayed_list);
250 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
251 sh->bm_seq - conf->seq_write > 0)
252 list_add_tail(&sh->lru, &conf->bitmap_list);
254 clear_bit(STRIPE_DELAYED, &sh->state);
255 clear_bit(STRIPE_BIT_DELAY, &sh->state);
256 if (conf->worker_cnt_per_group == 0) {
257 list_add_tail(&sh->lru, &conf->handle_list);
259 raid5_wakeup_stripe_thread(sh);
263 md_wakeup_thread(conf->mddev->thread);
265 BUG_ON(stripe_operations_active(sh));
266 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
267 if (atomic_dec_return(&conf->preread_active_stripes)
269 md_wakeup_thread(conf->mddev->thread);
270 atomic_dec(&conf->active_stripes);
271 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
272 if (!r5c_is_writeback(conf->log))
273 list_add_tail(&sh->lru, temp_inactive_list);
275 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
277 list_add_tail(&sh->lru, temp_inactive_list);
278 else if (injournal == conf->raid_disks - conf->max_degraded) {
280 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
281 atomic_inc(&conf->r5c_cached_full_stripes);
282 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
283 atomic_dec(&conf->r5c_cached_partial_stripes);
284 list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
285 r5c_check_cached_full_stripe(conf);
288 * STRIPE_R5C_PARTIAL_STRIPE is set in
289 * r5c_try_caching_write(). No need to
292 list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
298 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
299 struct list_head *temp_inactive_list)
301 if (atomic_dec_and_test(&sh->count))
302 do_release_stripe(conf, sh, temp_inactive_list);
306 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
308 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
309 * given time. Adding stripes only takes device lock, while deleting stripes
310 * only takes hash lock.
312 static void release_inactive_stripe_list(struct r5conf *conf,
313 struct list_head *temp_inactive_list,
317 bool do_wakeup = false;
320 if (hash == NR_STRIPE_HASH_LOCKS) {
321 size = NR_STRIPE_HASH_LOCKS;
322 hash = NR_STRIPE_HASH_LOCKS - 1;
326 struct list_head *list = &temp_inactive_list[size - 1];
329 * We don't hold any lock here yet, raid5_get_active_stripe() might
330 * remove stripes from the list
332 if (!list_empty_careful(list)) {
333 spin_lock_irqsave(conf->hash_locks + hash, flags);
334 if (list_empty(conf->inactive_list + hash) &&
336 atomic_dec(&conf->empty_inactive_list_nr);
337 list_splice_tail_init(list, conf->inactive_list + hash);
339 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
346 wake_up(&conf->wait_for_stripe);
347 if (atomic_read(&conf->active_stripes) == 0)
348 wake_up(&conf->wait_for_quiescent);
349 if (conf->retry_read_aligned)
350 md_wakeup_thread(conf->mddev->thread);
354 /* should hold conf->device_lock already */
355 static int release_stripe_list(struct r5conf *conf,
356 struct list_head *temp_inactive_list)
358 struct stripe_head *sh, *t;
360 struct llist_node *head;
362 head = llist_del_all(&conf->released_stripes);
363 head = llist_reverse_order(head);
364 llist_for_each_entry_safe(sh, t, head, release_list) {
367 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
369 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
371 * Don't worry the bit is set here, because if the bit is set
372 * again, the count is always > 1. This is true for
373 * STRIPE_ON_UNPLUG_LIST bit too.
375 hash = sh->hash_lock_index;
376 __release_stripe(conf, sh, &temp_inactive_list[hash]);
383 void raid5_release_stripe(struct stripe_head *sh)
385 struct r5conf *conf = sh->raid_conf;
387 struct list_head list;
391 /* Avoid release_list until the last reference.
393 if (atomic_add_unless(&sh->count, -1, 1))
396 if (unlikely(!conf->mddev->thread) ||
397 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
399 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
401 md_wakeup_thread(conf->mddev->thread);
404 local_irq_save(flags);
405 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
406 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
407 INIT_LIST_HEAD(&list);
408 hash = sh->hash_lock_index;
409 do_release_stripe(conf, sh, &list);
410 spin_unlock(&conf->device_lock);
411 release_inactive_stripe_list(conf, &list, hash);
413 local_irq_restore(flags);
416 static inline void remove_hash(struct stripe_head *sh)
418 pr_debug("remove_hash(), stripe %llu\n",
419 (unsigned long long)sh->sector);
421 hlist_del_init(&sh->hash);
424 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
426 struct hlist_head *hp = stripe_hash(conf, sh->sector);
428 pr_debug("insert_hash(), stripe %llu\n",
429 (unsigned long long)sh->sector);
431 hlist_add_head(&sh->hash, hp);
434 /* find an idle stripe, make sure it is unhashed, and return it. */
435 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
437 struct stripe_head *sh = NULL;
438 struct list_head *first;
440 if (list_empty(conf->inactive_list + hash))
442 first = (conf->inactive_list + hash)->next;
443 sh = list_entry(first, struct stripe_head, lru);
444 list_del_init(first);
446 atomic_inc(&conf->active_stripes);
447 BUG_ON(hash != sh->hash_lock_index);
448 if (list_empty(conf->inactive_list + hash))
449 atomic_inc(&conf->empty_inactive_list_nr);
454 static void shrink_buffers(struct stripe_head *sh)
458 int num = sh->raid_conf->pool_size;
460 for (i = 0; i < num ; i++) {
461 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
465 sh->dev[i].page = NULL;
470 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
473 int num = sh->raid_conf->pool_size;
475 for (i = 0; i < num; i++) {
478 if (!(page = alloc_page(gfp))) {
481 sh->dev[i].page = page;
482 sh->dev[i].orig_page = page;
487 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
488 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
489 struct stripe_head *sh);
491 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
493 struct r5conf *conf = sh->raid_conf;
496 BUG_ON(atomic_read(&sh->count) != 0);
497 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
498 BUG_ON(stripe_operations_active(sh));
499 BUG_ON(sh->batch_head);
501 pr_debug("init_stripe called, stripe %llu\n",
502 (unsigned long long)sector);
504 seq = read_seqcount_begin(&conf->gen_lock);
505 sh->generation = conf->generation - previous;
506 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
508 stripe_set_idx(sector, conf, previous, sh);
511 for (i = sh->disks; i--; ) {
512 struct r5dev *dev = &sh->dev[i];
514 if (dev->toread || dev->read || dev->towrite || dev->written ||
515 test_bit(R5_LOCKED, &dev->flags)) {
516 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
517 (unsigned long long)sh->sector, i, dev->toread,
518 dev->read, dev->towrite, dev->written,
519 test_bit(R5_LOCKED, &dev->flags));
523 raid5_build_block(sh, i, previous);
525 if (read_seqcount_retry(&conf->gen_lock, seq))
527 sh->overwrite_disks = 0;
528 insert_hash(conf, sh);
529 sh->cpu = smp_processor_id();
530 set_bit(STRIPE_BATCH_READY, &sh->state);
533 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
536 struct stripe_head *sh;
538 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
539 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
540 if (sh->sector == sector && sh->generation == generation)
542 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
547 * Need to check if array has failed when deciding whether to:
549 * - remove non-faulty devices
552 * This determination is simple when no reshape is happening.
553 * However if there is a reshape, we need to carefully check
554 * both the before and after sections.
555 * This is because some failed devices may only affect one
556 * of the two sections, and some non-in_sync devices may
557 * be insync in the section most affected by failed devices.
559 int raid5_calc_degraded(struct r5conf *conf)
561 int degraded, degraded2;
566 for (i = 0; i < conf->previous_raid_disks; i++) {
567 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
568 if (rdev && test_bit(Faulty, &rdev->flags))
569 rdev = rcu_dereference(conf->disks[i].replacement);
570 if (!rdev || test_bit(Faulty, &rdev->flags))
572 else if (test_bit(In_sync, &rdev->flags))
575 /* not in-sync or faulty.
576 * If the reshape increases the number of devices,
577 * this is being recovered by the reshape, so
578 * this 'previous' section is not in_sync.
579 * If the number of devices is being reduced however,
580 * the device can only be part of the array if
581 * we are reverting a reshape, so this section will
584 if (conf->raid_disks >= conf->previous_raid_disks)
588 if (conf->raid_disks == conf->previous_raid_disks)
592 for (i = 0; i < conf->raid_disks; i++) {
593 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
594 if (rdev && test_bit(Faulty, &rdev->flags))
595 rdev = rcu_dereference(conf->disks[i].replacement);
596 if (!rdev || test_bit(Faulty, &rdev->flags))
598 else if (test_bit(In_sync, &rdev->flags))
601 /* not in-sync or faulty.
602 * If reshape increases the number of devices, this
603 * section has already been recovered, else it
604 * almost certainly hasn't.
606 if (conf->raid_disks <= conf->previous_raid_disks)
610 if (degraded2 > degraded)
615 static int has_failed(struct r5conf *conf)
619 if (conf->mddev->reshape_position == MaxSector)
620 return conf->mddev->degraded > conf->max_degraded;
622 degraded = raid5_calc_degraded(conf);
623 if (degraded > conf->max_degraded)
629 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
630 int previous, int noblock, int noquiesce)
632 struct stripe_head *sh;
633 int hash = stripe_hash_locks_hash(sector);
634 int inc_empty_inactive_list_flag;
636 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
638 spin_lock_irq(conf->hash_locks + hash);
641 wait_event_lock_irq(conf->wait_for_quiescent,
642 conf->quiesce == 0 || noquiesce,
643 *(conf->hash_locks + hash));
644 sh = __find_stripe(conf, sector, conf->generation - previous);
646 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
647 sh = get_free_stripe(conf, hash);
648 if (!sh && !test_bit(R5_DID_ALLOC,
650 set_bit(R5_ALLOC_MORE,
653 if (noblock && sh == NULL)
656 r5c_check_stripe_cache_usage(conf);
658 set_bit(R5_INACTIVE_BLOCKED,
660 r5l_wake_reclaim(conf->log, 0);
662 conf->wait_for_stripe,
663 !list_empty(conf->inactive_list + hash) &&
664 (atomic_read(&conf->active_stripes)
665 < (conf->max_nr_stripes * 3 / 4)
666 || !test_bit(R5_INACTIVE_BLOCKED,
667 &conf->cache_state)),
668 *(conf->hash_locks + hash));
669 clear_bit(R5_INACTIVE_BLOCKED,
672 init_stripe(sh, sector, previous);
673 atomic_inc(&sh->count);
675 } else if (!atomic_inc_not_zero(&sh->count)) {
676 spin_lock(&conf->device_lock);
677 if (!atomic_read(&sh->count)) {
678 if (!test_bit(STRIPE_HANDLE, &sh->state))
679 atomic_inc(&conf->active_stripes);
680 BUG_ON(list_empty(&sh->lru) &&
681 !test_bit(STRIPE_EXPANDING, &sh->state));
682 inc_empty_inactive_list_flag = 0;
683 if (!list_empty(conf->inactive_list + hash))
684 inc_empty_inactive_list_flag = 1;
685 list_del_init(&sh->lru);
686 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
687 atomic_inc(&conf->empty_inactive_list_nr);
689 sh->group->stripes_cnt--;
693 atomic_inc(&sh->count);
694 spin_unlock(&conf->device_lock);
696 } while (sh == NULL);
698 spin_unlock_irq(conf->hash_locks + hash);
702 static bool is_full_stripe_write(struct stripe_head *sh)
704 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
705 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
708 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
712 spin_lock(&sh2->stripe_lock);
713 spin_lock_nested(&sh1->stripe_lock, 1);
715 spin_lock(&sh1->stripe_lock);
716 spin_lock_nested(&sh2->stripe_lock, 1);
720 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
722 spin_unlock(&sh1->stripe_lock);
723 spin_unlock(&sh2->stripe_lock);
727 /* Only freshly new full stripe normal write stripe can be added to a batch list */
728 static bool stripe_can_batch(struct stripe_head *sh)
730 struct r5conf *conf = sh->raid_conf;
734 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
735 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
736 is_full_stripe_write(sh);
739 /* we only do back search */
740 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
742 struct stripe_head *head;
743 sector_t head_sector, tmp_sec;
746 int inc_empty_inactive_list_flag;
748 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
749 tmp_sec = sh->sector;
750 if (!sector_div(tmp_sec, conf->chunk_sectors))
752 head_sector = sh->sector - STRIPE_SECTORS;
754 hash = stripe_hash_locks_hash(head_sector);
755 spin_lock_irq(conf->hash_locks + hash);
756 head = __find_stripe(conf, head_sector, conf->generation);
757 if (head && !atomic_inc_not_zero(&head->count)) {
758 spin_lock(&conf->device_lock);
759 if (!atomic_read(&head->count)) {
760 if (!test_bit(STRIPE_HANDLE, &head->state))
761 atomic_inc(&conf->active_stripes);
762 BUG_ON(list_empty(&head->lru) &&
763 !test_bit(STRIPE_EXPANDING, &head->state));
764 inc_empty_inactive_list_flag = 0;
765 if (!list_empty(conf->inactive_list + hash))
766 inc_empty_inactive_list_flag = 1;
767 list_del_init(&head->lru);
768 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
769 atomic_inc(&conf->empty_inactive_list_nr);
771 head->group->stripes_cnt--;
775 atomic_inc(&head->count);
776 spin_unlock(&conf->device_lock);
778 spin_unlock_irq(conf->hash_locks + hash);
782 if (!stripe_can_batch(head))
785 lock_two_stripes(head, sh);
786 /* clear_batch_ready clear the flag */
787 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
794 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
796 if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
797 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
800 if (head->batch_head) {
801 spin_lock(&head->batch_head->batch_lock);
802 /* This batch list is already running */
803 if (!stripe_can_batch(head)) {
804 spin_unlock(&head->batch_head->batch_lock);
809 * at this point, head's BATCH_READY could be cleared, but we
810 * can still add the stripe to batch list
812 list_add(&sh->batch_list, &head->batch_list);
813 spin_unlock(&head->batch_head->batch_lock);
815 sh->batch_head = head->batch_head;
817 head->batch_head = head;
818 sh->batch_head = head->batch_head;
819 spin_lock(&head->batch_lock);
820 list_add_tail(&sh->batch_list, &head->batch_list);
821 spin_unlock(&head->batch_lock);
824 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
825 if (atomic_dec_return(&conf->preread_active_stripes)
827 md_wakeup_thread(conf->mddev->thread);
829 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
830 int seq = sh->bm_seq;
831 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
832 sh->batch_head->bm_seq > seq)
833 seq = sh->batch_head->bm_seq;
834 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
835 sh->batch_head->bm_seq = seq;
838 atomic_inc(&sh->count);
840 unlock_two_stripes(head, sh);
842 raid5_release_stripe(head);
845 /* Determine if 'data_offset' or 'new_data_offset' should be used
846 * in this stripe_head.
848 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
850 sector_t progress = conf->reshape_progress;
851 /* Need a memory barrier to make sure we see the value
852 * of conf->generation, or ->data_offset that was set before
853 * reshape_progress was updated.
856 if (progress == MaxSector)
858 if (sh->generation == conf->generation - 1)
860 /* We are in a reshape, and this is a new-generation stripe,
861 * so use new_data_offset.
866 static void flush_deferred_bios(struct r5conf *conf)
871 if (!conf->batch_bio_dispatch || !conf->group_cnt)
875 spin_lock(&conf->pending_bios_lock);
876 bio_list_merge(&tmp, &conf->pending_bios);
877 bio_list_init(&conf->pending_bios);
878 spin_unlock(&conf->pending_bios_lock);
880 while ((bio = bio_list_pop(&tmp)))
881 generic_make_request(bio);
884 static void defer_bio_issue(struct r5conf *conf, struct bio *bio)
887 * change group_cnt will drain all bios, so this is safe
889 * A read generally means a read-modify-write, which usually means a
890 * randwrite, so we don't delay it
892 if (!conf->batch_bio_dispatch || !conf->group_cnt ||
893 bio_op(bio) == REQ_OP_READ) {
894 generic_make_request(bio);
897 spin_lock(&conf->pending_bios_lock);
898 bio_list_add(&conf->pending_bios, bio);
899 spin_unlock(&conf->pending_bios_lock);
900 md_wakeup_thread(conf->mddev->thread);
904 raid5_end_read_request(struct bio *bi);
906 raid5_end_write_request(struct bio *bi);
908 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
910 struct r5conf *conf = sh->raid_conf;
911 int i, disks = sh->disks;
912 struct stripe_head *head_sh = sh;
916 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
917 /* writing out phase */
918 if (s->waiting_extra_page)
920 if (r5l_write_stripe(conf->log, sh) == 0)
922 } else { /* caching phase */
923 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state)) {
924 r5c_cache_data(conf->log, sh, s);
929 for (i = disks; i--; ) {
930 int op, op_flags = 0;
931 int replace_only = 0;
932 struct bio *bi, *rbi;
933 struct md_rdev *rdev, *rrdev = NULL;
936 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
938 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
940 if (test_bit(R5_Discard, &sh->dev[i].flags))
942 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
944 else if (test_and_clear_bit(R5_WantReplace,
945 &sh->dev[i].flags)) {
950 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
951 op_flags |= REQ_SYNC;
954 bi = &sh->dev[i].req;
955 rbi = &sh->dev[i].rreq; /* For writing to replacement */
958 rrdev = rcu_dereference(conf->disks[i].replacement);
959 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
960 rdev = rcu_dereference(conf->disks[i].rdev);
965 if (op_is_write(op)) {
969 /* We raced and saw duplicates */
972 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
977 if (rdev && test_bit(Faulty, &rdev->flags))
980 atomic_inc(&rdev->nr_pending);
981 if (rrdev && test_bit(Faulty, &rrdev->flags))
984 atomic_inc(&rrdev->nr_pending);
987 /* We have already checked bad blocks for reads. Now
988 * need to check for writes. We never accept write errors
989 * on the replacement, so we don't to check rrdev.
991 while (op_is_write(op) && rdev &&
992 test_bit(WriteErrorSeen, &rdev->flags)) {
995 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
996 &first_bad, &bad_sectors);
1001 set_bit(BlockedBadBlocks, &rdev->flags);
1002 if (!conf->mddev->external &&
1003 conf->mddev->sb_flags) {
1004 /* It is very unlikely, but we might
1005 * still need to write out the
1006 * bad block log - better give it
1008 md_check_recovery(conf->mddev);
1011 * Because md_wait_for_blocked_rdev
1012 * will dec nr_pending, we must
1013 * increment it first.
1015 atomic_inc(&rdev->nr_pending);
1016 md_wait_for_blocked_rdev(rdev, conf->mddev);
1018 /* Acknowledged bad block - skip the write */
1019 rdev_dec_pending(rdev, conf->mddev);
1025 if (s->syncing || s->expanding || s->expanded
1027 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1029 set_bit(STRIPE_IO_STARTED, &sh->state);
1031 bi->bi_bdev = rdev->bdev;
1032 bio_set_op_attrs(bi, op, op_flags);
1033 bi->bi_end_io = op_is_write(op)
1034 ? raid5_end_write_request
1035 : raid5_end_read_request;
1036 bi->bi_private = sh;
1038 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1039 __func__, (unsigned long long)sh->sector,
1041 atomic_inc(&sh->count);
1043 atomic_inc(&head_sh->count);
1044 if (use_new_offset(conf, sh))
1045 bi->bi_iter.bi_sector = (sh->sector
1046 + rdev->new_data_offset);
1048 bi->bi_iter.bi_sector = (sh->sector
1049 + rdev->data_offset);
1050 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1051 bi->bi_opf |= REQ_NOMERGE;
1053 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1054 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1056 if (!op_is_write(op) &&
1057 test_bit(R5_InJournal, &sh->dev[i].flags))
1059 * issuing read for a page in journal, this
1060 * must be preparing for prexor in rmw; read
1061 * the data into orig_page
1063 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1065 sh->dev[i].vec.bv_page = sh->dev[i].page;
1067 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1068 bi->bi_io_vec[0].bv_offset = 0;
1069 bi->bi_iter.bi_size = STRIPE_SIZE;
1071 * If this is discard request, set bi_vcnt 0. We don't
1072 * want to confuse SCSI because SCSI will replace payload
1074 if (op == REQ_OP_DISCARD)
1077 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1079 if (conf->mddev->gendisk)
1080 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1081 bi, disk_devt(conf->mddev->gendisk),
1083 defer_bio_issue(conf, bi);
1086 if (s->syncing || s->expanding || s->expanded
1088 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1090 set_bit(STRIPE_IO_STARTED, &sh->state);
1092 rbi->bi_bdev = rrdev->bdev;
1093 bio_set_op_attrs(rbi, op, op_flags);
1094 BUG_ON(!op_is_write(op));
1095 rbi->bi_end_io = raid5_end_write_request;
1096 rbi->bi_private = sh;
1098 pr_debug("%s: for %llu schedule op %d on "
1099 "replacement disc %d\n",
1100 __func__, (unsigned long long)sh->sector,
1102 atomic_inc(&sh->count);
1104 atomic_inc(&head_sh->count);
1105 if (use_new_offset(conf, sh))
1106 rbi->bi_iter.bi_sector = (sh->sector
1107 + rrdev->new_data_offset);
1109 rbi->bi_iter.bi_sector = (sh->sector
1110 + rrdev->data_offset);
1111 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1112 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1113 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1115 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1116 rbi->bi_io_vec[0].bv_offset = 0;
1117 rbi->bi_iter.bi_size = STRIPE_SIZE;
1119 * If this is discard request, set bi_vcnt 0. We don't
1120 * want to confuse SCSI because SCSI will replace payload
1122 if (op == REQ_OP_DISCARD)
1124 if (conf->mddev->gendisk)
1125 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1126 rbi, disk_devt(conf->mddev->gendisk),
1128 defer_bio_issue(conf, rbi);
1130 if (!rdev && !rrdev) {
1131 if (op_is_write(op))
1132 set_bit(STRIPE_DEGRADED, &sh->state);
1133 pr_debug("skip op %d on disc %d for sector %llu\n",
1134 bi->bi_opf, i, (unsigned long long)sh->sector);
1135 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1136 set_bit(STRIPE_HANDLE, &sh->state);
1139 if (!head_sh->batch_head)
1141 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1148 static struct dma_async_tx_descriptor *
1149 async_copy_data(int frombio, struct bio *bio, struct page **page,
1150 sector_t sector, struct dma_async_tx_descriptor *tx,
1151 struct stripe_head *sh, int no_skipcopy)
1154 struct bvec_iter iter;
1155 struct page *bio_page;
1157 struct async_submit_ctl submit;
1158 enum async_tx_flags flags = 0;
1160 if (bio->bi_iter.bi_sector >= sector)
1161 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1163 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1166 flags |= ASYNC_TX_FENCE;
1167 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1169 bio_for_each_segment(bvl, bio, iter) {
1170 int len = bvl.bv_len;
1174 if (page_offset < 0) {
1175 b_offset = -page_offset;
1176 page_offset += b_offset;
1180 if (len > 0 && page_offset + len > STRIPE_SIZE)
1181 clen = STRIPE_SIZE - page_offset;
1186 b_offset += bvl.bv_offset;
1187 bio_page = bvl.bv_page;
1189 if (sh->raid_conf->skip_copy &&
1190 b_offset == 0 && page_offset == 0 &&
1191 clen == STRIPE_SIZE &&
1195 tx = async_memcpy(*page, bio_page, page_offset,
1196 b_offset, clen, &submit);
1198 tx = async_memcpy(bio_page, *page, b_offset,
1199 page_offset, clen, &submit);
1201 /* chain the operations */
1202 submit.depend_tx = tx;
1204 if (clen < len) /* hit end of page */
1212 static void ops_complete_biofill(void *stripe_head_ref)
1214 struct stripe_head *sh = stripe_head_ref;
1215 struct bio_list return_bi = BIO_EMPTY_LIST;
1218 pr_debug("%s: stripe %llu\n", __func__,
1219 (unsigned long long)sh->sector);
1221 /* clear completed biofills */
1222 for (i = sh->disks; i--; ) {
1223 struct r5dev *dev = &sh->dev[i];
1225 /* acknowledge completion of a biofill operation */
1226 /* and check if we need to reply to a read request,
1227 * new R5_Wantfill requests are held off until
1228 * !STRIPE_BIOFILL_RUN
1230 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1231 struct bio *rbi, *rbi2;
1236 while (rbi && rbi->bi_iter.bi_sector <
1237 dev->sector + STRIPE_SECTORS) {
1238 rbi2 = r5_next_bio(rbi, dev->sector);
1239 if (!raid5_dec_bi_active_stripes(rbi))
1240 bio_list_add(&return_bi, rbi);
1245 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1247 return_io(&return_bi);
1249 set_bit(STRIPE_HANDLE, &sh->state);
1250 raid5_release_stripe(sh);
1253 static void ops_run_biofill(struct stripe_head *sh)
1255 struct dma_async_tx_descriptor *tx = NULL;
1256 struct async_submit_ctl submit;
1259 BUG_ON(sh->batch_head);
1260 pr_debug("%s: stripe %llu\n", __func__,
1261 (unsigned long long)sh->sector);
1263 for (i = sh->disks; i--; ) {
1264 struct r5dev *dev = &sh->dev[i];
1265 if (test_bit(R5_Wantfill, &dev->flags)) {
1267 spin_lock_irq(&sh->stripe_lock);
1268 dev->read = rbi = dev->toread;
1270 spin_unlock_irq(&sh->stripe_lock);
1271 while (rbi && rbi->bi_iter.bi_sector <
1272 dev->sector + STRIPE_SECTORS) {
1273 tx = async_copy_data(0, rbi, &dev->page,
1274 dev->sector, tx, sh, 0);
1275 rbi = r5_next_bio(rbi, dev->sector);
1280 atomic_inc(&sh->count);
1281 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1282 async_trigger_callback(&submit);
1285 static void mark_target_uptodate(struct stripe_head *sh, int target)
1292 tgt = &sh->dev[target];
1293 set_bit(R5_UPTODATE, &tgt->flags);
1294 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1295 clear_bit(R5_Wantcompute, &tgt->flags);
1298 static void ops_complete_compute(void *stripe_head_ref)
1300 struct stripe_head *sh = stripe_head_ref;
1302 pr_debug("%s: stripe %llu\n", __func__,
1303 (unsigned long long)sh->sector);
1305 /* mark the computed target(s) as uptodate */
1306 mark_target_uptodate(sh, sh->ops.target);
1307 mark_target_uptodate(sh, sh->ops.target2);
1309 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1310 if (sh->check_state == check_state_compute_run)
1311 sh->check_state = check_state_compute_result;
1312 set_bit(STRIPE_HANDLE, &sh->state);
1313 raid5_release_stripe(sh);
1316 /* return a pointer to the address conversion region of the scribble buffer */
1317 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1318 struct raid5_percpu *percpu, int i)
1322 addr = flex_array_get(percpu->scribble, i);
1323 return addr + sizeof(struct page *) * (sh->disks + 2);
1326 /* return a pointer to the address conversion region of the scribble buffer */
1327 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1331 addr = flex_array_get(percpu->scribble, i);
1335 static struct dma_async_tx_descriptor *
1336 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1338 int disks = sh->disks;
1339 struct page **xor_srcs = to_addr_page(percpu, 0);
1340 int target = sh->ops.target;
1341 struct r5dev *tgt = &sh->dev[target];
1342 struct page *xor_dest = tgt->page;
1344 struct dma_async_tx_descriptor *tx;
1345 struct async_submit_ctl submit;
1348 BUG_ON(sh->batch_head);
1350 pr_debug("%s: stripe %llu block: %d\n",
1351 __func__, (unsigned long long)sh->sector, target);
1352 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1354 for (i = disks; i--; )
1356 xor_srcs[count++] = sh->dev[i].page;
1358 atomic_inc(&sh->count);
1360 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1361 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1362 if (unlikely(count == 1))
1363 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1365 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1370 /* set_syndrome_sources - populate source buffers for gen_syndrome
1371 * @srcs - (struct page *) array of size sh->disks
1372 * @sh - stripe_head to parse
1374 * Populates srcs in proper layout order for the stripe and returns the
1375 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1376 * destination buffer is recorded in srcs[count] and the Q destination
1377 * is recorded in srcs[count+1]].
1379 static int set_syndrome_sources(struct page **srcs,
1380 struct stripe_head *sh,
1383 int disks = sh->disks;
1384 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1385 int d0_idx = raid6_d0(sh);
1389 for (i = 0; i < disks; i++)
1395 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1396 struct r5dev *dev = &sh->dev[i];
1398 if (i == sh->qd_idx || i == sh->pd_idx ||
1399 (srctype == SYNDROME_SRC_ALL) ||
1400 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1401 (test_bit(R5_Wantdrain, &dev->flags) ||
1402 test_bit(R5_InJournal, &dev->flags))) ||
1403 (srctype == SYNDROME_SRC_WRITTEN &&
1405 test_bit(R5_InJournal, &dev->flags)))) {
1406 if (test_bit(R5_InJournal, &dev->flags))
1407 srcs[slot] = sh->dev[i].orig_page;
1409 srcs[slot] = sh->dev[i].page;
1411 i = raid6_next_disk(i, disks);
1412 } while (i != d0_idx);
1414 return syndrome_disks;
1417 static struct dma_async_tx_descriptor *
1418 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1420 int disks = sh->disks;
1421 struct page **blocks = to_addr_page(percpu, 0);
1423 int qd_idx = sh->qd_idx;
1424 struct dma_async_tx_descriptor *tx;
1425 struct async_submit_ctl submit;
1431 BUG_ON(sh->batch_head);
1432 if (sh->ops.target < 0)
1433 target = sh->ops.target2;
1434 else if (sh->ops.target2 < 0)
1435 target = sh->ops.target;
1437 /* we should only have one valid target */
1440 pr_debug("%s: stripe %llu block: %d\n",
1441 __func__, (unsigned long long)sh->sector, target);
1443 tgt = &sh->dev[target];
1444 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1447 atomic_inc(&sh->count);
1449 if (target == qd_idx) {
1450 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1451 blocks[count] = NULL; /* regenerating p is not necessary */
1452 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1453 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1454 ops_complete_compute, sh,
1455 to_addr_conv(sh, percpu, 0));
1456 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1458 /* Compute any data- or p-drive using XOR */
1460 for (i = disks; i-- ; ) {
1461 if (i == target || i == qd_idx)
1463 blocks[count++] = sh->dev[i].page;
1466 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1467 NULL, ops_complete_compute, sh,
1468 to_addr_conv(sh, percpu, 0));
1469 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1475 static struct dma_async_tx_descriptor *
1476 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1478 int i, count, disks = sh->disks;
1479 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1480 int d0_idx = raid6_d0(sh);
1481 int faila = -1, failb = -1;
1482 int target = sh->ops.target;
1483 int target2 = sh->ops.target2;
1484 struct r5dev *tgt = &sh->dev[target];
1485 struct r5dev *tgt2 = &sh->dev[target2];
1486 struct dma_async_tx_descriptor *tx;
1487 struct page **blocks = to_addr_page(percpu, 0);
1488 struct async_submit_ctl submit;
1490 BUG_ON(sh->batch_head);
1491 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1492 __func__, (unsigned long long)sh->sector, target, target2);
1493 BUG_ON(target < 0 || target2 < 0);
1494 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1495 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1497 /* we need to open-code set_syndrome_sources to handle the
1498 * slot number conversion for 'faila' and 'failb'
1500 for (i = 0; i < disks ; i++)
1505 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1507 blocks[slot] = sh->dev[i].page;
1513 i = raid6_next_disk(i, disks);
1514 } while (i != d0_idx);
1516 BUG_ON(faila == failb);
1519 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1520 __func__, (unsigned long long)sh->sector, faila, failb);
1522 atomic_inc(&sh->count);
1524 if (failb == syndrome_disks+1) {
1525 /* Q disk is one of the missing disks */
1526 if (faila == syndrome_disks) {
1527 /* Missing P+Q, just recompute */
1528 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1529 ops_complete_compute, sh,
1530 to_addr_conv(sh, percpu, 0));
1531 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1532 STRIPE_SIZE, &submit);
1536 int qd_idx = sh->qd_idx;
1538 /* Missing D+Q: recompute D from P, then recompute Q */
1539 if (target == qd_idx)
1540 data_target = target2;
1542 data_target = target;
1545 for (i = disks; i-- ; ) {
1546 if (i == data_target || i == qd_idx)
1548 blocks[count++] = sh->dev[i].page;
1550 dest = sh->dev[data_target].page;
1551 init_async_submit(&submit,
1552 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1554 to_addr_conv(sh, percpu, 0));
1555 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1558 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1559 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1560 ops_complete_compute, sh,
1561 to_addr_conv(sh, percpu, 0));
1562 return async_gen_syndrome(blocks, 0, count+2,
1563 STRIPE_SIZE, &submit);
1566 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1567 ops_complete_compute, sh,
1568 to_addr_conv(sh, percpu, 0));
1569 if (failb == syndrome_disks) {
1570 /* We're missing D+P. */
1571 return async_raid6_datap_recov(syndrome_disks+2,
1575 /* We're missing D+D. */
1576 return async_raid6_2data_recov(syndrome_disks+2,
1577 STRIPE_SIZE, faila, failb,
1583 static void ops_complete_prexor(void *stripe_head_ref)
1585 struct stripe_head *sh = stripe_head_ref;
1587 pr_debug("%s: stripe %llu\n", __func__,
1588 (unsigned long long)sh->sector);
1590 if (r5c_is_writeback(sh->raid_conf->log))
1592 * raid5-cache write back uses orig_page during prexor.
1593 * After prexor, it is time to free orig_page
1595 r5c_release_extra_page(sh);
1598 static struct dma_async_tx_descriptor *
1599 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1600 struct dma_async_tx_descriptor *tx)
1602 int disks = sh->disks;
1603 struct page **xor_srcs = to_addr_page(percpu, 0);
1604 int count = 0, pd_idx = sh->pd_idx, i;
1605 struct async_submit_ctl submit;
1607 /* existing parity data subtracted */
1608 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1610 BUG_ON(sh->batch_head);
1611 pr_debug("%s: stripe %llu\n", __func__,
1612 (unsigned long long)sh->sector);
1614 for (i = disks; i--; ) {
1615 struct r5dev *dev = &sh->dev[i];
1616 /* Only process blocks that are known to be uptodate */
1617 if (test_bit(R5_InJournal, &dev->flags))
1618 xor_srcs[count++] = dev->orig_page;
1619 else if (test_bit(R5_Wantdrain, &dev->flags))
1620 xor_srcs[count++] = dev->page;
1623 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1624 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1625 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1630 static struct dma_async_tx_descriptor *
1631 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1632 struct dma_async_tx_descriptor *tx)
1634 struct page **blocks = to_addr_page(percpu, 0);
1636 struct async_submit_ctl submit;
1638 pr_debug("%s: stripe %llu\n", __func__,
1639 (unsigned long long)sh->sector);
1641 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1643 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1644 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1645 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1650 static struct dma_async_tx_descriptor *
1651 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1653 struct r5conf *conf = sh->raid_conf;
1654 int disks = sh->disks;
1656 struct stripe_head *head_sh = sh;
1658 pr_debug("%s: stripe %llu\n", __func__,
1659 (unsigned long long)sh->sector);
1661 for (i = disks; i--; ) {
1666 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1672 * clear R5_InJournal, so when rewriting a page in
1673 * journal, it is not skipped by r5l_log_stripe()
1675 clear_bit(R5_InJournal, &dev->flags);
1676 spin_lock_irq(&sh->stripe_lock);
1677 chosen = dev->towrite;
1678 dev->towrite = NULL;
1679 sh->overwrite_disks = 0;
1680 BUG_ON(dev->written);
1681 wbi = dev->written = chosen;
1682 spin_unlock_irq(&sh->stripe_lock);
1683 WARN_ON(dev->page != dev->orig_page);
1685 while (wbi && wbi->bi_iter.bi_sector <
1686 dev->sector + STRIPE_SECTORS) {
1687 if (wbi->bi_opf & REQ_FUA)
1688 set_bit(R5_WantFUA, &dev->flags);
1689 if (wbi->bi_opf & REQ_SYNC)
1690 set_bit(R5_SyncIO, &dev->flags);
1691 if (bio_op(wbi) == REQ_OP_DISCARD)
1692 set_bit(R5_Discard, &dev->flags);
1694 tx = async_copy_data(1, wbi, &dev->page,
1695 dev->sector, tx, sh,
1696 r5c_is_writeback(conf->log));
1697 if (dev->page != dev->orig_page &&
1698 !r5c_is_writeback(conf->log)) {
1699 set_bit(R5_SkipCopy, &dev->flags);
1700 clear_bit(R5_UPTODATE, &dev->flags);
1701 clear_bit(R5_OVERWRITE, &dev->flags);
1704 wbi = r5_next_bio(wbi, dev->sector);
1707 if (head_sh->batch_head) {
1708 sh = list_first_entry(&sh->batch_list,
1721 static void ops_complete_reconstruct(void *stripe_head_ref)
1723 struct stripe_head *sh = stripe_head_ref;
1724 int disks = sh->disks;
1725 int pd_idx = sh->pd_idx;
1726 int qd_idx = sh->qd_idx;
1728 bool fua = false, sync = false, discard = false;
1730 pr_debug("%s: stripe %llu\n", __func__,
1731 (unsigned long long)sh->sector);
1733 for (i = disks; i--; ) {
1734 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1735 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1736 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1739 for (i = disks; i--; ) {
1740 struct r5dev *dev = &sh->dev[i];
1742 if (dev->written || i == pd_idx || i == qd_idx) {
1743 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1744 set_bit(R5_UPTODATE, &dev->flags);
1746 set_bit(R5_WantFUA, &dev->flags);
1748 set_bit(R5_SyncIO, &dev->flags);
1752 if (sh->reconstruct_state == reconstruct_state_drain_run)
1753 sh->reconstruct_state = reconstruct_state_drain_result;
1754 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1755 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1757 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1758 sh->reconstruct_state = reconstruct_state_result;
1761 set_bit(STRIPE_HANDLE, &sh->state);
1762 raid5_release_stripe(sh);
1766 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1767 struct dma_async_tx_descriptor *tx)
1769 int disks = sh->disks;
1770 struct page **xor_srcs;
1771 struct async_submit_ctl submit;
1772 int count, pd_idx = sh->pd_idx, i;
1773 struct page *xor_dest;
1775 unsigned long flags;
1777 struct stripe_head *head_sh = sh;
1780 pr_debug("%s: stripe %llu\n", __func__,
1781 (unsigned long long)sh->sector);
1783 for (i = 0; i < sh->disks; i++) {
1786 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1789 if (i >= sh->disks) {
1790 atomic_inc(&sh->count);
1791 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1792 ops_complete_reconstruct(sh);
1797 xor_srcs = to_addr_page(percpu, j);
1798 /* check if prexor is active which means only process blocks
1799 * that are part of a read-modify-write (written)
1801 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1803 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1804 for (i = disks; i--; ) {
1805 struct r5dev *dev = &sh->dev[i];
1806 if (head_sh->dev[i].written ||
1807 test_bit(R5_InJournal, &head_sh->dev[i].flags))
1808 xor_srcs[count++] = dev->page;
1811 xor_dest = sh->dev[pd_idx].page;
1812 for (i = disks; i--; ) {
1813 struct r5dev *dev = &sh->dev[i];
1815 xor_srcs[count++] = dev->page;
1819 /* 1/ if we prexor'd then the dest is reused as a source
1820 * 2/ if we did not prexor then we are redoing the parity
1821 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1822 * for the synchronous xor case
1824 last_stripe = !head_sh->batch_head ||
1825 list_first_entry(&sh->batch_list,
1826 struct stripe_head, batch_list) == head_sh;
1828 flags = ASYNC_TX_ACK |
1829 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1831 atomic_inc(&head_sh->count);
1832 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1833 to_addr_conv(sh, percpu, j));
1835 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1836 init_async_submit(&submit, flags, tx, NULL, NULL,
1837 to_addr_conv(sh, percpu, j));
1840 if (unlikely(count == 1))
1841 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1843 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1846 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1853 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1854 struct dma_async_tx_descriptor *tx)
1856 struct async_submit_ctl submit;
1857 struct page **blocks;
1858 int count, i, j = 0;
1859 struct stripe_head *head_sh = sh;
1862 unsigned long txflags;
1864 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1866 for (i = 0; i < sh->disks; i++) {
1867 if (sh->pd_idx == i || sh->qd_idx == i)
1869 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1872 if (i >= sh->disks) {
1873 atomic_inc(&sh->count);
1874 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1875 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1876 ops_complete_reconstruct(sh);
1881 blocks = to_addr_page(percpu, j);
1883 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1884 synflags = SYNDROME_SRC_WRITTEN;
1885 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1887 synflags = SYNDROME_SRC_ALL;
1888 txflags = ASYNC_TX_ACK;
1891 count = set_syndrome_sources(blocks, sh, synflags);
1892 last_stripe = !head_sh->batch_head ||
1893 list_first_entry(&sh->batch_list,
1894 struct stripe_head, batch_list) == head_sh;
1897 atomic_inc(&head_sh->count);
1898 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1899 head_sh, to_addr_conv(sh, percpu, j));
1901 init_async_submit(&submit, 0, tx, NULL, NULL,
1902 to_addr_conv(sh, percpu, j));
1903 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1906 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1912 static void ops_complete_check(void *stripe_head_ref)
1914 struct stripe_head *sh = stripe_head_ref;
1916 pr_debug("%s: stripe %llu\n", __func__,
1917 (unsigned long long)sh->sector);
1919 sh->check_state = check_state_check_result;
1920 set_bit(STRIPE_HANDLE, &sh->state);
1921 raid5_release_stripe(sh);
1924 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1926 int disks = sh->disks;
1927 int pd_idx = sh->pd_idx;
1928 int qd_idx = sh->qd_idx;
1929 struct page *xor_dest;
1930 struct page **xor_srcs = to_addr_page(percpu, 0);
1931 struct dma_async_tx_descriptor *tx;
1932 struct async_submit_ctl submit;
1936 pr_debug("%s: stripe %llu\n", __func__,
1937 (unsigned long long)sh->sector);
1939 BUG_ON(sh->batch_head);
1941 xor_dest = sh->dev[pd_idx].page;
1942 xor_srcs[count++] = xor_dest;
1943 for (i = disks; i--; ) {
1944 if (i == pd_idx || i == qd_idx)
1946 xor_srcs[count++] = sh->dev[i].page;
1949 init_async_submit(&submit, 0, NULL, NULL, NULL,
1950 to_addr_conv(sh, percpu, 0));
1951 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1952 &sh->ops.zero_sum_result, &submit);
1954 atomic_inc(&sh->count);
1955 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1956 tx = async_trigger_callback(&submit);
1959 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1961 struct page **srcs = to_addr_page(percpu, 0);
1962 struct async_submit_ctl submit;
1965 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1966 (unsigned long long)sh->sector, checkp);
1968 BUG_ON(sh->batch_head);
1969 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1973 atomic_inc(&sh->count);
1974 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1975 sh, to_addr_conv(sh, percpu, 0));
1976 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1977 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1980 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1982 int overlap_clear = 0, i, disks = sh->disks;
1983 struct dma_async_tx_descriptor *tx = NULL;
1984 struct r5conf *conf = sh->raid_conf;
1985 int level = conf->level;
1986 struct raid5_percpu *percpu;
1990 percpu = per_cpu_ptr(conf->percpu, cpu);
1991 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1992 ops_run_biofill(sh);
1996 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1998 tx = ops_run_compute5(sh, percpu);
2000 if (sh->ops.target2 < 0 || sh->ops.target < 0)
2001 tx = ops_run_compute6_1(sh, percpu);
2003 tx = ops_run_compute6_2(sh, percpu);
2005 /* terminate the chain if reconstruct is not set to be run */
2006 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2010 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2012 tx = ops_run_prexor5(sh, percpu, tx);
2014 tx = ops_run_prexor6(sh, percpu, tx);
2017 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2018 tx = ops_run_biodrain(sh, tx);
2022 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2024 ops_run_reconstruct5(sh, percpu, tx);
2026 ops_run_reconstruct6(sh, percpu, tx);
2029 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2030 if (sh->check_state == check_state_run)
2031 ops_run_check_p(sh, percpu);
2032 else if (sh->check_state == check_state_run_q)
2033 ops_run_check_pq(sh, percpu, 0);
2034 else if (sh->check_state == check_state_run_pq)
2035 ops_run_check_pq(sh, percpu, 1);
2040 if (overlap_clear && !sh->batch_head)
2041 for (i = disks; i--; ) {
2042 struct r5dev *dev = &sh->dev[i];
2043 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2044 wake_up(&sh->raid_conf->wait_for_overlap);
2049 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2052 struct stripe_head *sh;
2055 sh = kmem_cache_zalloc(sc, gfp);
2057 spin_lock_init(&sh->stripe_lock);
2058 spin_lock_init(&sh->batch_lock);
2059 INIT_LIST_HEAD(&sh->batch_list);
2060 INIT_LIST_HEAD(&sh->lru);
2061 INIT_LIST_HEAD(&sh->r5c);
2062 INIT_LIST_HEAD(&sh->log_list);
2063 atomic_set(&sh->count, 1);
2064 sh->log_start = MaxSector;
2065 for (i = 0; i < disks; i++) {
2066 struct r5dev *dev = &sh->dev[i];
2068 bio_init(&dev->req, &dev->vec, 1);
2069 bio_init(&dev->rreq, &dev->rvec, 1);
2074 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2076 struct stripe_head *sh;
2078 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size);
2082 sh->raid_conf = conf;
2084 if (grow_buffers(sh, gfp)) {
2086 kmem_cache_free(conf->slab_cache, sh);
2089 sh->hash_lock_index =
2090 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2091 /* we just created an active stripe so... */
2092 atomic_inc(&conf->active_stripes);
2094 raid5_release_stripe(sh);
2095 conf->max_nr_stripes++;
2099 static int grow_stripes(struct r5conf *conf, int num)
2101 struct kmem_cache *sc;
2102 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2104 if (conf->mddev->gendisk)
2105 sprintf(conf->cache_name[0],
2106 "raid%d-%s", conf->level, mdname(conf->mddev));
2108 sprintf(conf->cache_name[0],
2109 "raid%d-%p", conf->level, conf->mddev);
2110 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2112 conf->active_name = 0;
2113 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2114 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2118 conf->slab_cache = sc;
2119 conf->pool_size = devs;
2121 if (!grow_one_stripe(conf, GFP_KERNEL))
2128 * scribble_len - return the required size of the scribble region
2129 * @num - total number of disks in the array
2131 * The size must be enough to contain:
2132 * 1/ a struct page pointer for each device in the array +2
2133 * 2/ room to convert each entry in (1) to its corresponding dma
2134 * (dma_map_page()) or page (page_address()) address.
2136 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2137 * calculate over all devices (not just the data blocks), using zeros in place
2138 * of the P and Q blocks.
2140 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2142 struct flex_array *ret;
2145 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2146 ret = flex_array_alloc(len, cnt, flags);
2149 /* always prealloc all elements, so no locking is required */
2150 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2151 flex_array_free(ret);
2157 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2163 * Never shrink. And mddev_suspend() could deadlock if this is called
2164 * from raid5d. In that case, scribble_disks and scribble_sectors
2165 * should equal to new_disks and new_sectors
2167 if (conf->scribble_disks >= new_disks &&
2168 conf->scribble_sectors >= new_sectors)
2170 mddev_suspend(conf->mddev);
2172 for_each_present_cpu(cpu) {
2173 struct raid5_percpu *percpu;
2174 struct flex_array *scribble;
2176 percpu = per_cpu_ptr(conf->percpu, cpu);
2177 scribble = scribble_alloc(new_disks,
2178 new_sectors / STRIPE_SECTORS,
2182 flex_array_free(percpu->scribble);
2183 percpu->scribble = scribble;
2190 mddev_resume(conf->mddev);
2192 conf->scribble_disks = new_disks;
2193 conf->scribble_sectors = new_sectors;
2198 static int resize_stripes(struct r5conf *conf, int newsize)
2200 /* Make all the stripes able to hold 'newsize' devices.
2201 * New slots in each stripe get 'page' set to a new page.
2203 * This happens in stages:
2204 * 1/ create a new kmem_cache and allocate the required number of
2206 * 2/ gather all the old stripe_heads and transfer the pages across
2207 * to the new stripe_heads. This will have the side effect of
2208 * freezing the array as once all stripe_heads have been collected,
2209 * no IO will be possible. Old stripe heads are freed once their
2210 * pages have been transferred over, and the old kmem_cache is
2211 * freed when all stripes are done.
2212 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2213 * we simple return a failre status - no need to clean anything up.
2214 * 4/ allocate new pages for the new slots in the new stripe_heads.
2215 * If this fails, we don't bother trying the shrink the
2216 * stripe_heads down again, we just leave them as they are.
2217 * As each stripe_head is processed the new one is released into
2220 * Once step2 is started, we cannot afford to wait for a write,
2221 * so we use GFP_NOIO allocations.
2223 struct stripe_head *osh, *nsh;
2224 LIST_HEAD(newstripes);
2225 struct disk_info *ndisks;
2227 struct kmem_cache *sc;
2231 if (newsize <= conf->pool_size)
2232 return 0; /* never bother to shrink */
2234 err = md_allow_write(conf->mddev);
2239 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2240 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2245 /* Need to ensure auto-resizing doesn't interfere */
2246 mutex_lock(&conf->cache_size_mutex);
2248 for (i = conf->max_nr_stripes; i; i--) {
2249 nsh = alloc_stripe(sc, GFP_KERNEL, newsize);
2253 nsh->raid_conf = conf;
2254 list_add(&nsh->lru, &newstripes);
2257 /* didn't get enough, give up */
2258 while (!list_empty(&newstripes)) {
2259 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2260 list_del(&nsh->lru);
2261 kmem_cache_free(sc, nsh);
2263 kmem_cache_destroy(sc);
2264 mutex_unlock(&conf->cache_size_mutex);
2267 /* Step 2 - Must use GFP_NOIO now.
2268 * OK, we have enough stripes, start collecting inactive
2269 * stripes and copying them over
2273 list_for_each_entry(nsh, &newstripes, lru) {
2274 lock_device_hash_lock(conf, hash);
2275 wait_event_cmd(conf->wait_for_stripe,
2276 !list_empty(conf->inactive_list + hash),
2277 unlock_device_hash_lock(conf, hash),
2278 lock_device_hash_lock(conf, hash));
2279 osh = get_free_stripe(conf, hash);
2280 unlock_device_hash_lock(conf, hash);
2282 for(i=0; i<conf->pool_size; i++) {
2283 nsh->dev[i].page = osh->dev[i].page;
2284 nsh->dev[i].orig_page = osh->dev[i].page;
2286 nsh->hash_lock_index = hash;
2287 kmem_cache_free(conf->slab_cache, osh);
2289 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2290 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2295 kmem_cache_destroy(conf->slab_cache);
2298 * At this point, we are holding all the stripes so the array
2299 * is completely stalled, so now is a good time to resize
2300 * conf->disks and the scribble region
2302 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2304 for (i = 0; i < conf->pool_size; i++)
2305 ndisks[i] = conf->disks[i];
2307 for (i = conf->pool_size; i < newsize; i++) {
2308 ndisks[i].extra_page = alloc_page(GFP_NOIO);
2309 if (!ndisks[i].extra_page)
2314 for (i = conf->pool_size; i < newsize; i++)
2315 if (ndisks[i].extra_page)
2316 put_page(ndisks[i].extra_page);
2320 conf->disks = ndisks;
2325 mutex_unlock(&conf->cache_size_mutex);
2326 /* Step 4, return new stripes to service */
2327 while(!list_empty(&newstripes)) {
2328 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2329 list_del_init(&nsh->lru);
2331 for (i=conf->raid_disks; i < newsize; i++)
2332 if (nsh->dev[i].page == NULL) {
2333 struct page *p = alloc_page(GFP_NOIO);
2334 nsh->dev[i].page = p;
2335 nsh->dev[i].orig_page = p;
2339 raid5_release_stripe(nsh);
2341 /* critical section pass, GFP_NOIO no longer needed */
2343 conf->slab_cache = sc;
2344 conf->active_name = 1-conf->active_name;
2346 conf->pool_size = newsize;
2350 static int drop_one_stripe(struct r5conf *conf)
2352 struct stripe_head *sh;
2353 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2355 spin_lock_irq(conf->hash_locks + hash);
2356 sh = get_free_stripe(conf, hash);
2357 spin_unlock_irq(conf->hash_locks + hash);
2360 BUG_ON(atomic_read(&sh->count));
2362 kmem_cache_free(conf->slab_cache, sh);
2363 atomic_dec(&conf->active_stripes);
2364 conf->max_nr_stripes--;
2368 static void shrink_stripes(struct r5conf *conf)
2370 while (conf->max_nr_stripes &&
2371 drop_one_stripe(conf))
2374 kmem_cache_destroy(conf->slab_cache);
2375 conf->slab_cache = NULL;
2378 static void raid5_end_read_request(struct bio * bi)
2380 struct stripe_head *sh = bi->bi_private;
2381 struct r5conf *conf = sh->raid_conf;
2382 int disks = sh->disks, i;
2383 char b[BDEVNAME_SIZE];
2384 struct md_rdev *rdev = NULL;
2387 for (i=0 ; i<disks; i++)
2388 if (bi == &sh->dev[i].req)
2391 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2392 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2399 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2400 /* If replacement finished while this request was outstanding,
2401 * 'replacement' might be NULL already.
2402 * In that case it moved down to 'rdev'.
2403 * rdev is not removed until all requests are finished.
2405 rdev = conf->disks[i].replacement;
2407 rdev = conf->disks[i].rdev;
2409 if (use_new_offset(conf, sh))
2410 s = sh->sector + rdev->new_data_offset;
2412 s = sh->sector + rdev->data_offset;
2413 if (!bi->bi_error) {
2414 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2415 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2416 /* Note that this cannot happen on a
2417 * replacement device. We just fail those on
2420 pr_info_ratelimited(
2421 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2422 mdname(conf->mddev), STRIPE_SECTORS,
2423 (unsigned long long)s,
2424 bdevname(rdev->bdev, b));
2425 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2426 clear_bit(R5_ReadError, &sh->dev[i].flags);
2427 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2428 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2429 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2431 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2433 * end read for a page in journal, this
2434 * must be preparing for prexor in rmw
2436 set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2438 if (atomic_read(&rdev->read_errors))
2439 atomic_set(&rdev->read_errors, 0);
2441 const char *bdn = bdevname(rdev->bdev, b);
2445 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2446 atomic_inc(&rdev->read_errors);
2447 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2448 pr_warn_ratelimited(
2449 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2450 mdname(conf->mddev),
2451 (unsigned long long)s,
2453 else if (conf->mddev->degraded >= conf->max_degraded) {
2455 pr_warn_ratelimited(
2456 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2457 mdname(conf->mddev),
2458 (unsigned long long)s,
2460 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2463 pr_warn_ratelimited(
2464 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2465 mdname(conf->mddev),
2466 (unsigned long long)s,
2468 } else if (atomic_read(&rdev->read_errors)
2469 > conf->max_nr_stripes)
2470 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2471 mdname(conf->mddev), bdn);
2474 if (set_bad && test_bit(In_sync, &rdev->flags)
2475 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2478 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2479 set_bit(R5_ReadError, &sh->dev[i].flags);
2480 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2482 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2484 clear_bit(R5_ReadError, &sh->dev[i].flags);
2485 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2487 && test_bit(In_sync, &rdev->flags)
2488 && rdev_set_badblocks(
2489 rdev, sh->sector, STRIPE_SECTORS, 0)))
2490 md_error(conf->mddev, rdev);
2493 rdev_dec_pending(rdev, conf->mddev);
2495 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2496 set_bit(STRIPE_HANDLE, &sh->state);
2497 raid5_release_stripe(sh);
2500 static void raid5_end_write_request(struct bio *bi)
2502 struct stripe_head *sh = bi->bi_private;
2503 struct r5conf *conf = sh->raid_conf;
2504 int disks = sh->disks, i;
2505 struct md_rdev *uninitialized_var(rdev);
2508 int replacement = 0;
2510 for (i = 0 ; i < disks; i++) {
2511 if (bi == &sh->dev[i].req) {
2512 rdev = conf->disks[i].rdev;
2515 if (bi == &sh->dev[i].rreq) {
2516 rdev = conf->disks[i].replacement;
2520 /* rdev was removed and 'replacement'
2521 * replaced it. rdev is not removed
2522 * until all requests are finished.
2524 rdev = conf->disks[i].rdev;
2528 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2529 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2539 md_error(conf->mddev, rdev);
2540 else if (is_badblock(rdev, sh->sector,
2542 &first_bad, &bad_sectors))
2543 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2546 set_bit(STRIPE_DEGRADED, &sh->state);
2547 set_bit(WriteErrorSeen, &rdev->flags);
2548 set_bit(R5_WriteError, &sh->dev[i].flags);
2549 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2550 set_bit(MD_RECOVERY_NEEDED,
2551 &rdev->mddev->recovery);
2552 } else if (is_badblock(rdev, sh->sector,
2554 &first_bad, &bad_sectors)) {
2555 set_bit(R5_MadeGood, &sh->dev[i].flags);
2556 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2557 /* That was a successful write so make
2558 * sure it looks like we already did
2561 set_bit(R5_ReWrite, &sh->dev[i].flags);
2564 rdev_dec_pending(rdev, conf->mddev);
2566 if (sh->batch_head && bi->bi_error && !replacement)
2567 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2570 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2571 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2572 set_bit(STRIPE_HANDLE, &sh->state);
2573 raid5_release_stripe(sh);
2575 if (sh->batch_head && sh != sh->batch_head)
2576 raid5_release_stripe(sh->batch_head);
2579 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2581 struct r5dev *dev = &sh->dev[i];
2584 dev->sector = raid5_compute_blocknr(sh, i, previous);
2587 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2589 char b[BDEVNAME_SIZE];
2590 struct r5conf *conf = mddev->private;
2591 unsigned long flags;
2592 pr_debug("raid456: error called\n");
2594 spin_lock_irqsave(&conf->device_lock, flags);
2595 clear_bit(In_sync, &rdev->flags);
2596 mddev->degraded = raid5_calc_degraded(conf);
2597 spin_unlock_irqrestore(&conf->device_lock, flags);
2598 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2600 set_bit(Blocked, &rdev->flags);
2601 set_bit(Faulty, &rdev->flags);
2602 set_mask_bits(&mddev->sb_flags, 0,
2603 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2604 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2605 "md/raid:%s: Operation continuing on %d devices.\n",
2607 bdevname(rdev->bdev, b),
2609 conf->raid_disks - mddev->degraded);
2610 r5c_update_on_rdev_error(mddev);
2614 * Input: a 'big' sector number,
2615 * Output: index of the data and parity disk, and the sector # in them.
2617 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2618 int previous, int *dd_idx,
2619 struct stripe_head *sh)
2621 sector_t stripe, stripe2;
2622 sector_t chunk_number;
2623 unsigned int chunk_offset;
2626 sector_t new_sector;
2627 int algorithm = previous ? conf->prev_algo
2629 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2630 : conf->chunk_sectors;
2631 int raid_disks = previous ? conf->previous_raid_disks
2633 int data_disks = raid_disks - conf->max_degraded;
2635 /* First compute the information on this sector */
2638 * Compute the chunk number and the sector offset inside the chunk
2640 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2641 chunk_number = r_sector;
2644 * Compute the stripe number
2646 stripe = chunk_number;
2647 *dd_idx = sector_div(stripe, data_disks);
2650 * Select the parity disk based on the user selected algorithm.
2652 pd_idx = qd_idx = -1;
2653 switch(conf->level) {
2655 pd_idx = data_disks;
2658 switch (algorithm) {
2659 case ALGORITHM_LEFT_ASYMMETRIC:
2660 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2661 if (*dd_idx >= pd_idx)
2664 case ALGORITHM_RIGHT_ASYMMETRIC:
2665 pd_idx = sector_div(stripe2, raid_disks);
2666 if (*dd_idx >= pd_idx)
2669 case ALGORITHM_LEFT_SYMMETRIC:
2670 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2671 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2673 case ALGORITHM_RIGHT_SYMMETRIC:
2674 pd_idx = sector_div(stripe2, raid_disks);
2675 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2677 case ALGORITHM_PARITY_0:
2681 case ALGORITHM_PARITY_N:
2682 pd_idx = data_disks;
2690 switch (algorithm) {
2691 case ALGORITHM_LEFT_ASYMMETRIC:
2692 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2693 qd_idx = pd_idx + 1;
2694 if (pd_idx == raid_disks-1) {
2695 (*dd_idx)++; /* Q D D D P */
2697 } else if (*dd_idx >= pd_idx)
2698 (*dd_idx) += 2; /* D D P Q D */
2700 case ALGORITHM_RIGHT_ASYMMETRIC:
2701 pd_idx = sector_div(stripe2, raid_disks);
2702 qd_idx = pd_idx + 1;
2703 if (pd_idx == raid_disks-1) {
2704 (*dd_idx)++; /* Q D D D P */
2706 } else if (*dd_idx >= pd_idx)
2707 (*dd_idx) += 2; /* D D P Q D */
2709 case ALGORITHM_LEFT_SYMMETRIC:
2710 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2711 qd_idx = (pd_idx + 1) % raid_disks;
2712 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2714 case ALGORITHM_RIGHT_SYMMETRIC:
2715 pd_idx = sector_div(stripe2, raid_disks);
2716 qd_idx = (pd_idx + 1) % raid_disks;
2717 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2720 case ALGORITHM_PARITY_0:
2725 case ALGORITHM_PARITY_N:
2726 pd_idx = data_disks;
2727 qd_idx = data_disks + 1;
2730 case ALGORITHM_ROTATING_ZERO_RESTART:
2731 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2732 * of blocks for computing Q is different.
2734 pd_idx = sector_div(stripe2, raid_disks);
2735 qd_idx = pd_idx + 1;
2736 if (pd_idx == raid_disks-1) {
2737 (*dd_idx)++; /* Q D D D P */
2739 } else if (*dd_idx >= pd_idx)
2740 (*dd_idx) += 2; /* D D P Q D */
2744 case ALGORITHM_ROTATING_N_RESTART:
2745 /* Same a left_asymmetric, by first stripe is
2746 * D D D P Q rather than
2750 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2751 qd_idx = pd_idx + 1;
2752 if (pd_idx == raid_disks-1) {
2753 (*dd_idx)++; /* Q D D D P */
2755 } else if (*dd_idx >= pd_idx)
2756 (*dd_idx) += 2; /* D D P Q D */
2760 case ALGORITHM_ROTATING_N_CONTINUE:
2761 /* Same as left_symmetric but Q is before P */
2762 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2763 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2764 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2768 case ALGORITHM_LEFT_ASYMMETRIC_6:
2769 /* RAID5 left_asymmetric, with Q on last device */
2770 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2771 if (*dd_idx >= pd_idx)
2773 qd_idx = raid_disks - 1;
2776 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2777 pd_idx = sector_div(stripe2, raid_disks-1);
2778 if (*dd_idx >= pd_idx)
2780 qd_idx = raid_disks - 1;
2783 case ALGORITHM_LEFT_SYMMETRIC_6:
2784 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2785 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2786 qd_idx = raid_disks - 1;
2789 case ALGORITHM_RIGHT_SYMMETRIC_6:
2790 pd_idx = sector_div(stripe2, raid_disks-1);
2791 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2792 qd_idx = raid_disks - 1;
2795 case ALGORITHM_PARITY_0_6:
2798 qd_idx = raid_disks - 1;
2808 sh->pd_idx = pd_idx;
2809 sh->qd_idx = qd_idx;
2810 sh->ddf_layout = ddf_layout;
2813 * Finally, compute the new sector number
2815 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2819 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2821 struct r5conf *conf = sh->raid_conf;
2822 int raid_disks = sh->disks;
2823 int data_disks = raid_disks - conf->max_degraded;
2824 sector_t new_sector = sh->sector, check;
2825 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2826 : conf->chunk_sectors;
2827 int algorithm = previous ? conf->prev_algo
2831 sector_t chunk_number;
2832 int dummy1, dd_idx = i;
2834 struct stripe_head sh2;
2836 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2837 stripe = new_sector;
2839 if (i == sh->pd_idx)
2841 switch(conf->level) {
2844 switch (algorithm) {
2845 case ALGORITHM_LEFT_ASYMMETRIC:
2846 case ALGORITHM_RIGHT_ASYMMETRIC:
2850 case ALGORITHM_LEFT_SYMMETRIC:
2851 case ALGORITHM_RIGHT_SYMMETRIC:
2854 i -= (sh->pd_idx + 1);
2856 case ALGORITHM_PARITY_0:
2859 case ALGORITHM_PARITY_N:
2866 if (i == sh->qd_idx)
2867 return 0; /* It is the Q disk */
2868 switch (algorithm) {
2869 case ALGORITHM_LEFT_ASYMMETRIC:
2870 case ALGORITHM_RIGHT_ASYMMETRIC:
2871 case ALGORITHM_ROTATING_ZERO_RESTART:
2872 case ALGORITHM_ROTATING_N_RESTART:
2873 if (sh->pd_idx == raid_disks-1)
2874 i--; /* Q D D D P */
2875 else if (i > sh->pd_idx)
2876 i -= 2; /* D D P Q D */
2878 case ALGORITHM_LEFT_SYMMETRIC:
2879 case ALGORITHM_RIGHT_SYMMETRIC:
2880 if (sh->pd_idx == raid_disks-1)
2881 i--; /* Q D D D P */
2886 i -= (sh->pd_idx + 2);
2889 case ALGORITHM_PARITY_0:
2892 case ALGORITHM_PARITY_N:
2894 case ALGORITHM_ROTATING_N_CONTINUE:
2895 /* Like left_symmetric, but P is before Q */
2896 if (sh->pd_idx == 0)
2897 i--; /* P D D D Q */
2902 i -= (sh->pd_idx + 1);
2905 case ALGORITHM_LEFT_ASYMMETRIC_6:
2906 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2910 case ALGORITHM_LEFT_SYMMETRIC_6:
2911 case ALGORITHM_RIGHT_SYMMETRIC_6:
2913 i += data_disks + 1;
2914 i -= (sh->pd_idx + 1);
2916 case ALGORITHM_PARITY_0_6:
2925 chunk_number = stripe * data_disks + i;
2926 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2928 check = raid5_compute_sector(conf, r_sector,
2929 previous, &dummy1, &sh2);
2930 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2931 || sh2.qd_idx != sh->qd_idx) {
2932 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
2933 mdname(conf->mddev));
2940 * There are cases where we want handle_stripe_dirtying() and
2941 * schedule_reconstruction() to delay towrite to some dev of a stripe.
2943 * This function checks whether we want to delay the towrite. Specifically,
2944 * we delay the towrite when:
2946 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
2947 * stripe has data in journal (for other devices).
2949 * In this case, when reading data for the non-overwrite dev, it is
2950 * necessary to handle complex rmw of write back cache (prexor with
2951 * orig_page, and xor with page). To keep read path simple, we would
2952 * like to flush data in journal to RAID disks first, so complex rmw
2953 * is handled in the write patch (handle_stripe_dirtying).
2955 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
2957 * It is important to be able to flush all stripes in raid5-cache.
2958 * Therefore, we need reserve some space on the journal device for
2959 * these flushes. If flush operation includes pending writes to the
2960 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
2961 * for the flush out. If we exclude these pending writes from flush
2962 * operation, we only need (conf->max_degraded + 1) pages per stripe.
2963 * Therefore, excluding pending writes in these cases enables more
2964 * efficient use of the journal device.
2966 * Note: To make sure the stripe makes progress, we only delay
2967 * towrite for stripes with data already in journal (injournal > 0).
2968 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
2969 * no_space_stripes list.
2972 static inline bool delay_towrite(struct r5conf *conf,
2974 struct stripe_head_state *s)
2977 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2978 !test_bit(R5_Insync, &dev->flags) && s->injournal)
2981 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2988 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2989 int rcw, int expand)
2991 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2992 struct r5conf *conf = sh->raid_conf;
2993 int level = conf->level;
2997 * In some cases, handle_stripe_dirtying initially decided to
2998 * run rmw and allocates extra page for prexor. However, rcw is
2999 * cheaper later on. We need to free the extra page now,
3000 * because we won't be able to do that in ops_complete_prexor().
3002 r5c_release_extra_page(sh);
3004 for (i = disks; i--; ) {
3005 struct r5dev *dev = &sh->dev[i];
3007 if (dev->towrite && !delay_towrite(conf, dev, s)) {
3008 set_bit(R5_LOCKED, &dev->flags);
3009 set_bit(R5_Wantdrain, &dev->flags);
3011 clear_bit(R5_UPTODATE, &dev->flags);
3013 } else if (test_bit(R5_InJournal, &dev->flags)) {
3014 set_bit(R5_LOCKED, &dev->flags);
3018 /* if we are not expanding this is a proper write request, and
3019 * there will be bios with new data to be drained into the
3024 /* False alarm, nothing to do */
3026 sh->reconstruct_state = reconstruct_state_drain_run;
3027 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3029 sh->reconstruct_state = reconstruct_state_run;
3031 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3033 if (s->locked + conf->max_degraded == disks)
3034 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3035 atomic_inc(&conf->pending_full_writes);
3037 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3038 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3039 BUG_ON(level == 6 &&
3040 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3041 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3043 for (i = disks; i--; ) {
3044 struct r5dev *dev = &sh->dev[i];
3045 if (i == pd_idx || i == qd_idx)
3049 (test_bit(R5_UPTODATE, &dev->flags) ||
3050 test_bit(R5_Wantcompute, &dev->flags))) {
3051 set_bit(R5_Wantdrain, &dev->flags);
3052 set_bit(R5_LOCKED, &dev->flags);
3053 clear_bit(R5_UPTODATE, &dev->flags);
3055 } else if (test_bit(R5_InJournal, &dev->flags)) {
3056 set_bit(R5_LOCKED, &dev->flags);
3061 /* False alarm - nothing to do */
3063 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3064 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3065 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3066 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3069 /* keep the parity disk(s) locked while asynchronous operations
3072 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3073 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3077 int qd_idx = sh->qd_idx;
3078 struct r5dev *dev = &sh->dev[qd_idx];
3080 set_bit(R5_LOCKED, &dev->flags);
3081 clear_bit(R5_UPTODATE, &dev->flags);
3085 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3086 __func__, (unsigned long long)sh->sector,
3087 s->locked, s->ops_request);
3091 * Each stripe/dev can have one or more bion attached.
3092 * toread/towrite point to the first in a chain.
3093 * The bi_next chain must be in order.
3095 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3096 int forwrite, int previous)
3099 struct r5conf *conf = sh->raid_conf;
3102 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3103 (unsigned long long)bi->bi_iter.bi_sector,
3104 (unsigned long long)sh->sector);
3107 * If several bio share a stripe. The bio bi_phys_segments acts as a
3108 * reference count to avoid race. The reference count should already be
3109 * increased before this function is called (for example, in
3110 * raid5_make_request()), so other bio sharing this stripe will not free the
3111 * stripe. If a stripe is owned by one stripe, the stripe lock will
3114 spin_lock_irq(&sh->stripe_lock);
3115 /* Don't allow new IO added to stripes in batch list */
3119 bip = &sh->dev[dd_idx].towrite;
3123 bip = &sh->dev[dd_idx].toread;
3124 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3125 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3127 bip = & (*bip)->bi_next;
3129 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3132 if (!forwrite || previous)
3133 clear_bit(STRIPE_BATCH_READY, &sh->state);
3135 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3139 raid5_inc_bi_active_stripes(bi);
3142 /* check if page is covered */
3143 sector_t sector = sh->dev[dd_idx].sector;
3144 for (bi=sh->dev[dd_idx].towrite;
3145 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3146 bi && bi->bi_iter.bi_sector <= sector;
3147 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3148 if (bio_end_sector(bi) >= sector)
3149 sector = bio_end_sector(bi);
3151 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3152 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3153 sh->overwrite_disks++;
3156 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3157 (unsigned long long)(*bip)->bi_iter.bi_sector,
3158 (unsigned long long)sh->sector, dd_idx);
3160 if (conf->mddev->bitmap && firstwrite) {
3161 /* Cannot hold spinlock over bitmap_startwrite,
3162 * but must ensure this isn't added to a batch until
3163 * we have added to the bitmap and set bm_seq.
3164 * So set STRIPE_BITMAP_PENDING to prevent
3166 * If multiple add_stripe_bio() calls race here they
3167 * much all set STRIPE_BITMAP_PENDING. So only the first one
3168 * to complete "bitmap_startwrite" gets to set
3169 * STRIPE_BIT_DELAY. This is important as once a stripe
3170 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3173 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3174 spin_unlock_irq(&sh->stripe_lock);
3175 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3177 spin_lock_irq(&sh->stripe_lock);
3178 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3179 if (!sh->batch_head) {
3180 sh->bm_seq = conf->seq_flush+1;
3181 set_bit(STRIPE_BIT_DELAY, &sh->state);
3184 spin_unlock_irq(&sh->stripe_lock);
3186 if (stripe_can_batch(sh))
3187 stripe_add_to_batch_list(conf, sh);
3191 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3192 spin_unlock_irq(&sh->stripe_lock);
3196 static void end_reshape(struct r5conf *conf);
3198 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3199 struct stripe_head *sh)
3201 int sectors_per_chunk =
3202 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3204 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3205 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3207 raid5_compute_sector(conf,
3208 stripe * (disks - conf->max_degraded)
3209 *sectors_per_chunk + chunk_offset,
3215 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3216 struct stripe_head_state *s, int disks,
3217 struct bio_list *return_bi)
3220 BUG_ON(sh->batch_head);
3221 for (i = disks; i--; ) {
3225 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3226 struct md_rdev *rdev;
3228 rdev = rcu_dereference(conf->disks[i].rdev);
3229 if (rdev && test_bit(In_sync, &rdev->flags) &&
3230 !test_bit(Faulty, &rdev->flags))
3231 atomic_inc(&rdev->nr_pending);
3236 if (!rdev_set_badblocks(
3240 md_error(conf->mddev, rdev);
3241 rdev_dec_pending(rdev, conf->mddev);
3244 spin_lock_irq(&sh->stripe_lock);
3245 /* fail all writes first */
3246 bi = sh->dev[i].towrite;
3247 sh->dev[i].towrite = NULL;
3248 sh->overwrite_disks = 0;
3249 spin_unlock_irq(&sh->stripe_lock);
3253 r5l_stripe_write_finished(sh);
3255 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3256 wake_up(&conf->wait_for_overlap);
3258 while (bi && bi->bi_iter.bi_sector <
3259 sh->dev[i].sector + STRIPE_SECTORS) {
3260 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3262 bi->bi_error = -EIO;
3263 if (!raid5_dec_bi_active_stripes(bi)) {
3264 md_write_end(conf->mddev);
3265 bio_list_add(return_bi, bi);
3270 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3271 STRIPE_SECTORS, 0, 0);
3273 /* and fail all 'written' */
3274 bi = sh->dev[i].written;
3275 sh->dev[i].written = NULL;
3276 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3277 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3278 sh->dev[i].page = sh->dev[i].orig_page;
3281 if (bi) bitmap_end = 1;
3282 while (bi && bi->bi_iter.bi_sector <
3283 sh->dev[i].sector + STRIPE_SECTORS) {
3284 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3286 bi->bi_error = -EIO;
3287 if (!raid5_dec_bi_active_stripes(bi)) {
3288 md_write_end(conf->mddev);
3289 bio_list_add(return_bi, bi);
3294 /* fail any reads if this device is non-operational and
3295 * the data has not reached the cache yet.
3297 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3298 s->failed > conf->max_degraded &&
3299 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3300 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3301 spin_lock_irq(&sh->stripe_lock);
3302 bi = sh->dev[i].toread;
3303 sh->dev[i].toread = NULL;
3304 spin_unlock_irq(&sh->stripe_lock);
3305 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3306 wake_up(&conf->wait_for_overlap);
3309 while (bi && bi->bi_iter.bi_sector <
3310 sh->dev[i].sector + STRIPE_SECTORS) {
3311 struct bio *nextbi =
3312 r5_next_bio(bi, sh->dev[i].sector);
3314 bi->bi_error = -EIO;
3315 if (!raid5_dec_bi_active_stripes(bi))
3316 bio_list_add(return_bi, bi);
3321 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3322 STRIPE_SECTORS, 0, 0);
3323 /* If we were in the middle of a write the parity block might
3324 * still be locked - so just clear all R5_LOCKED flags
3326 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3331 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3332 if (atomic_dec_and_test(&conf->pending_full_writes))
3333 md_wakeup_thread(conf->mddev->thread);
3337 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3338 struct stripe_head_state *s)
3343 BUG_ON(sh->batch_head);
3344 clear_bit(STRIPE_SYNCING, &sh->state);
3345 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3346 wake_up(&conf->wait_for_overlap);
3349 /* There is nothing more to do for sync/check/repair.
3350 * Don't even need to abort as that is handled elsewhere
3351 * if needed, and not always wanted e.g. if there is a known
3353 * For recover/replace we need to record a bad block on all
3354 * non-sync devices, or abort the recovery
3356 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3357 /* During recovery devices cannot be removed, so
3358 * locking and refcounting of rdevs is not needed
3361 for (i = 0; i < conf->raid_disks; i++) {
3362 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3364 && !test_bit(Faulty, &rdev->flags)
3365 && !test_bit(In_sync, &rdev->flags)
3366 && !rdev_set_badblocks(rdev, sh->sector,
3369 rdev = rcu_dereference(conf->disks[i].replacement);
3371 && !test_bit(Faulty, &rdev->flags)
3372 && !test_bit(In_sync, &rdev->flags)
3373 && !rdev_set_badblocks(rdev, sh->sector,
3379 conf->recovery_disabled =
3380 conf->mddev->recovery_disabled;
3382 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3385 static int want_replace(struct stripe_head *sh, int disk_idx)
3387 struct md_rdev *rdev;
3391 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3393 && !test_bit(Faulty, &rdev->flags)
3394 && !test_bit(In_sync, &rdev->flags)
3395 && (rdev->recovery_offset <= sh->sector
3396 || rdev->mddev->recovery_cp <= sh->sector))
3402 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3403 int disk_idx, int disks)
3405 struct r5dev *dev = &sh->dev[disk_idx];
3406 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3407 &sh->dev[s->failed_num[1]] };
3411 if (test_bit(R5_LOCKED, &dev->flags) ||
3412 test_bit(R5_UPTODATE, &dev->flags))
3413 /* No point reading this as we already have it or have
3414 * decided to get it.
3419 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3420 /* We need this block to directly satisfy a request */
3423 if (s->syncing || s->expanding ||
3424 (s->replacing && want_replace(sh, disk_idx)))
3425 /* When syncing, or expanding we read everything.
3426 * When replacing, we need the replaced block.
3430 if ((s->failed >= 1 && fdev[0]->toread) ||
3431 (s->failed >= 2 && fdev[1]->toread))
3432 /* If we want to read from a failed device, then
3433 * we need to actually read every other device.
3437 /* Sometimes neither read-modify-write nor reconstruct-write
3438 * cycles can work. In those cases we read every block we
3439 * can. Then the parity-update is certain to have enough to
3441 * This can only be a problem when we need to write something,
3442 * and some device has failed. If either of those tests
3443 * fail we need look no further.
3445 if (!s->failed || !s->to_write)
3448 if (test_bit(R5_Insync, &dev->flags) &&
3449 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3450 /* Pre-reads at not permitted until after short delay
3451 * to gather multiple requests. However if this
3452 * device is no Insync, the block could only be be computed
3453 * and there is no need to delay that.
3457 for (i = 0; i < s->failed && i < 2; i++) {
3458 if (fdev[i]->towrite &&
3459 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3460 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3461 /* If we have a partial write to a failed
3462 * device, then we will need to reconstruct
3463 * the content of that device, so all other
3464 * devices must be read.
3469 /* If we are forced to do a reconstruct-write, either because
3470 * the current RAID6 implementation only supports that, or
3471 * or because parity cannot be trusted and we are currently
3472 * recovering it, there is extra need to be careful.
3473 * If one of the devices that we would need to read, because
3474 * it is not being overwritten (and maybe not written at all)
3475 * is missing/faulty, then we need to read everything we can.
3477 if (sh->raid_conf->level != 6 &&
3478 sh->sector < sh->raid_conf->mddev->recovery_cp)
3479 /* reconstruct-write isn't being forced */
3481 for (i = 0; i < s->failed && i < 2; i++) {
3482 if (s->failed_num[i] != sh->pd_idx &&
3483 s->failed_num[i] != sh->qd_idx &&
3484 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3485 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3492 /* fetch_block - checks the given member device to see if its data needs
3493 * to be read or computed to satisfy a request.
3495 * Returns 1 when no more member devices need to be checked, otherwise returns
3496 * 0 to tell the loop in handle_stripe_fill to continue
3498 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3499 int disk_idx, int disks)
3501 struct r5dev *dev = &sh->dev[disk_idx];
3503 /* is the data in this block needed, and can we get it? */
3504 if (need_this_block(sh, s, disk_idx, disks)) {
3505 /* we would like to get this block, possibly by computing it,
3506 * otherwise read it if the backing disk is insync
3508 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3509 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3510 BUG_ON(sh->batch_head);
3511 if ((s->uptodate == disks - 1) &&
3512 (s->failed && (disk_idx == s->failed_num[0] ||
3513 disk_idx == s->failed_num[1]))) {
3514 /* have disk failed, and we're requested to fetch it;
3517 pr_debug("Computing stripe %llu block %d\n",
3518 (unsigned long long)sh->sector, disk_idx);
3519 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3520 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3521 set_bit(R5_Wantcompute, &dev->flags);
3522 sh->ops.target = disk_idx;
3523 sh->ops.target2 = -1; /* no 2nd target */
3525 /* Careful: from this point on 'uptodate' is in the eye
3526 * of raid_run_ops which services 'compute' operations
3527 * before writes. R5_Wantcompute flags a block that will
3528 * be R5_UPTODATE by the time it is needed for a
3529 * subsequent operation.
3533 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3534 /* Computing 2-failure is *very* expensive; only
3535 * do it if failed >= 2
3538 for (other = disks; other--; ) {
3539 if (other == disk_idx)
3541 if (!test_bit(R5_UPTODATE,
3542 &sh->dev[other].flags))
3546 pr_debug("Computing stripe %llu blocks %d,%d\n",
3547 (unsigned long long)sh->sector,
3549 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3550 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3551 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3552 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3553 sh->ops.target = disk_idx;
3554 sh->ops.target2 = other;
3558 } else if (test_bit(R5_Insync, &dev->flags)) {
3559 set_bit(R5_LOCKED, &dev->flags);
3560 set_bit(R5_Wantread, &dev->flags);
3562 pr_debug("Reading block %d (sync=%d)\n",
3563 disk_idx, s->syncing);
3571 * handle_stripe_fill - read or compute data to satisfy pending requests.
3573 static void handle_stripe_fill(struct stripe_head *sh,
3574 struct stripe_head_state *s,
3579 /* look for blocks to read/compute, skip this if a compute
3580 * is already in flight, or if the stripe contents are in the
3581 * midst of changing due to a write
3583 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3584 !sh->reconstruct_state) {
3587 * For degraded stripe with data in journal, do not handle
3588 * read requests yet, instead, flush the stripe to raid
3589 * disks first, this avoids handling complex rmw of write
3590 * back cache (prexor with orig_page, and then xor with
3591 * page) in the read path
3593 if (s->injournal && s->failed) {
3594 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3595 r5c_make_stripe_write_out(sh);
3599 for (i = disks; i--; )
3600 if (fetch_block(sh, s, i, disks))
3604 set_bit(STRIPE_HANDLE, &sh->state);
3607 static void break_stripe_batch_list(struct stripe_head *head_sh,
3608 unsigned long handle_flags);
3609 /* handle_stripe_clean_event
3610 * any written block on an uptodate or failed drive can be returned.
3611 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3612 * never LOCKED, so we don't need to test 'failed' directly.
3614 static void handle_stripe_clean_event(struct r5conf *conf,
3615 struct stripe_head *sh, int disks, struct bio_list *return_bi)
3619 int discard_pending = 0;
3620 struct stripe_head *head_sh = sh;
3621 bool do_endio = false;
3623 for (i = disks; i--; )
3624 if (sh->dev[i].written) {
3626 if (!test_bit(R5_LOCKED, &dev->flags) &&
3627 (test_bit(R5_UPTODATE, &dev->flags) ||
3628 test_bit(R5_Discard, &dev->flags) ||
3629 test_bit(R5_SkipCopy, &dev->flags))) {
3630 /* We can return any write requests */
3631 struct bio *wbi, *wbi2;
3632 pr_debug("Return write for disc %d\n", i);
3633 if (test_and_clear_bit(R5_Discard, &dev->flags))
3634 clear_bit(R5_UPTODATE, &dev->flags);
3635 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3636 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3641 dev->page = dev->orig_page;
3643 dev->written = NULL;
3644 while (wbi && wbi->bi_iter.bi_sector <
3645 dev->sector + STRIPE_SECTORS) {
3646 wbi2 = r5_next_bio(wbi, dev->sector);
3647 if (!raid5_dec_bi_active_stripes(wbi)) {
3648 md_write_end(conf->mddev);
3649 bio_list_add(return_bi, wbi);
3653 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3655 !test_bit(STRIPE_DEGRADED, &sh->state),
3657 if (head_sh->batch_head) {
3658 sh = list_first_entry(&sh->batch_list,
3661 if (sh != head_sh) {
3668 } else if (test_bit(R5_Discard, &dev->flags))
3669 discard_pending = 1;
3672 r5l_stripe_write_finished(sh);
3674 if (!discard_pending &&
3675 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3677 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3678 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3679 if (sh->qd_idx >= 0) {
3680 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3681 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3683 /* now that discard is done we can proceed with any sync */
3684 clear_bit(STRIPE_DISCARD, &sh->state);
3686 * SCSI discard will change some bio fields and the stripe has
3687 * no updated data, so remove it from hash list and the stripe
3688 * will be reinitialized
3691 hash = sh->hash_lock_index;
3692 spin_lock_irq(conf->hash_locks + hash);
3694 spin_unlock_irq(conf->hash_locks + hash);
3695 if (head_sh->batch_head) {
3696 sh = list_first_entry(&sh->batch_list,
3697 struct stripe_head, batch_list);
3703 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3704 set_bit(STRIPE_HANDLE, &sh->state);
3708 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3709 if (atomic_dec_and_test(&conf->pending_full_writes))
3710 md_wakeup_thread(conf->mddev->thread);
3712 if (head_sh->batch_head && do_endio)
3713 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3717 * For RMW in write back cache, we need extra page in prexor to store the
3718 * old data. This page is stored in dev->orig_page.
3720 * This function checks whether we have data for prexor. The exact logic
3722 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3724 static inline bool uptodate_for_rmw(struct r5dev *dev)
3726 return (test_bit(R5_UPTODATE, &dev->flags)) &&
3727 (!test_bit(R5_InJournal, &dev->flags) ||
3728 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3731 static int handle_stripe_dirtying(struct r5conf *conf,
3732 struct stripe_head *sh,
3733 struct stripe_head_state *s,
3736 int rmw = 0, rcw = 0, i;
3737 sector_t recovery_cp = conf->mddev->recovery_cp;
3739 /* Check whether resync is now happening or should start.
3740 * If yes, then the array is dirty (after unclean shutdown or
3741 * initial creation), so parity in some stripes might be inconsistent.
3742 * In this case, we need to always do reconstruct-write, to ensure
3743 * that in case of drive failure or read-error correction, we
3744 * generate correct data from the parity.
3746 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3747 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3749 /* Calculate the real rcw later - for now make it
3750 * look like rcw is cheaper
3753 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3754 conf->rmw_level, (unsigned long long)recovery_cp,
3755 (unsigned long long)sh->sector);
3756 } else for (i = disks; i--; ) {
3757 /* would I have to read this buffer for read_modify_write */
3758 struct r5dev *dev = &sh->dev[i];
3759 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3760 i == sh->pd_idx || i == sh->qd_idx ||
3761 test_bit(R5_InJournal, &dev->flags)) &&
3762 !test_bit(R5_LOCKED, &dev->flags) &&
3763 !(uptodate_for_rmw(dev) ||
3764 test_bit(R5_Wantcompute, &dev->flags))) {
3765 if (test_bit(R5_Insync, &dev->flags))
3768 rmw += 2*disks; /* cannot read it */
3770 /* Would I have to read this buffer for reconstruct_write */
3771 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3772 i != sh->pd_idx && i != sh->qd_idx &&
3773 !test_bit(R5_LOCKED, &dev->flags) &&
3774 !(test_bit(R5_UPTODATE, &dev->flags) ||
3775 test_bit(R5_Wantcompute, &dev->flags))) {
3776 if (test_bit(R5_Insync, &dev->flags))
3783 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3784 (unsigned long long)sh->sector, sh->state, rmw, rcw);
3785 set_bit(STRIPE_HANDLE, &sh->state);
3786 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3787 /* prefer read-modify-write, but need to get some data */
3788 if (conf->mddev->queue)
3789 blk_add_trace_msg(conf->mddev->queue,
3790 "raid5 rmw %llu %d",
3791 (unsigned long long)sh->sector, rmw);
3792 for (i = disks; i--; ) {
3793 struct r5dev *dev = &sh->dev[i];
3794 if (test_bit(R5_InJournal, &dev->flags) &&
3795 dev->page == dev->orig_page &&
3796 !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3797 /* alloc page for prexor */
3798 struct page *p = alloc_page(GFP_NOIO);
3806 * alloc_page() failed, try use
3807 * disk_info->extra_page
3809 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3810 &conf->cache_state)) {
3811 r5c_use_extra_page(sh);
3815 /* extra_page in use, add to delayed_list */
3816 set_bit(STRIPE_DELAYED, &sh->state);
3817 s->waiting_extra_page = 1;
3822 for (i = disks; i--; ) {
3823 struct r5dev *dev = &sh->dev[i];
3824 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3825 i == sh->pd_idx || i == sh->qd_idx ||
3826 test_bit(R5_InJournal, &dev->flags)) &&
3827 !test_bit(R5_LOCKED, &dev->flags) &&
3828 !(uptodate_for_rmw(dev) ||
3829 test_bit(R5_Wantcompute, &dev->flags)) &&
3830 test_bit(R5_Insync, &dev->flags)) {
3831 if (test_bit(STRIPE_PREREAD_ACTIVE,
3833 pr_debug("Read_old block %d for r-m-w\n",
3835 set_bit(R5_LOCKED, &dev->flags);
3836 set_bit(R5_Wantread, &dev->flags);
3839 set_bit(STRIPE_DELAYED, &sh->state);
3840 set_bit(STRIPE_HANDLE, &sh->state);
3845 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3846 /* want reconstruct write, but need to get some data */
3849 for (i = disks; i--; ) {
3850 struct r5dev *dev = &sh->dev[i];
3851 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3852 i != sh->pd_idx && i != sh->qd_idx &&
3853 !test_bit(R5_LOCKED, &dev->flags) &&
3854 !(test_bit(R5_UPTODATE, &dev->flags) ||
3855 test_bit(R5_Wantcompute, &dev->flags))) {
3857 if (test_bit(R5_Insync, &dev->flags) &&
3858 test_bit(STRIPE_PREREAD_ACTIVE,
3860 pr_debug("Read_old block "
3861 "%d for Reconstruct\n", i);
3862 set_bit(R5_LOCKED, &dev->flags);
3863 set_bit(R5_Wantread, &dev->flags);
3867 set_bit(STRIPE_DELAYED, &sh->state);
3868 set_bit(STRIPE_HANDLE, &sh->state);
3872 if (rcw && conf->mddev->queue)
3873 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3874 (unsigned long long)sh->sector,
3875 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3878 if (rcw > disks && rmw > disks &&
3879 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3880 set_bit(STRIPE_DELAYED, &sh->state);
3882 /* now if nothing is locked, and if we have enough data,
3883 * we can start a write request
3885 /* since handle_stripe can be called at any time we need to handle the
3886 * case where a compute block operation has been submitted and then a
3887 * subsequent call wants to start a write request. raid_run_ops only
3888 * handles the case where compute block and reconstruct are requested
3889 * simultaneously. If this is not the case then new writes need to be
3890 * held off until the compute completes.
3892 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3893 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3894 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3895 schedule_reconstruction(sh, s, rcw == 0, 0);
3899 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3900 struct stripe_head_state *s, int disks)
3902 struct r5dev *dev = NULL;
3904 BUG_ON(sh->batch_head);
3905 set_bit(STRIPE_HANDLE, &sh->state);
3907 switch (sh->check_state) {
3908 case check_state_idle:
3909 /* start a new check operation if there are no failures */
3910 if (s->failed == 0) {
3911 BUG_ON(s->uptodate != disks);
3912 sh->check_state = check_state_run;
3913 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3914 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3918 dev = &sh->dev[s->failed_num[0]];
3920 case check_state_compute_result:
3921 sh->check_state = check_state_idle;
3923 dev = &sh->dev[sh->pd_idx];
3925 /* check that a write has not made the stripe insync */
3926 if (test_bit(STRIPE_INSYNC, &sh->state))
3929 /* either failed parity check, or recovery is happening */
3930 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3931 BUG_ON(s->uptodate != disks);
3933 set_bit(R5_LOCKED, &dev->flags);
3935 set_bit(R5_Wantwrite, &dev->flags);
3937 clear_bit(STRIPE_DEGRADED, &sh->state);
3938 set_bit(STRIPE_INSYNC, &sh->state);
3940 case check_state_run:
3941 break; /* we will be called again upon completion */
3942 case check_state_check_result:
3943 sh->check_state = check_state_idle;
3945 /* if a failure occurred during the check operation, leave
3946 * STRIPE_INSYNC not set and let the stripe be handled again
3951 /* handle a successful check operation, if parity is correct
3952 * we are done. Otherwise update the mismatch count and repair
3953 * parity if !MD_RECOVERY_CHECK
3955 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3956 /* parity is correct (on disc,
3957 * not in buffer any more)
3959 set_bit(STRIPE_INSYNC, &sh->state);
3961 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3962 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3963 /* don't try to repair!! */
3964 set_bit(STRIPE_INSYNC, &sh->state);
3966 sh->check_state = check_state_compute_run;
3967 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3968 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3969 set_bit(R5_Wantcompute,
3970 &sh->dev[sh->pd_idx].flags);
3971 sh->ops.target = sh->pd_idx;
3972 sh->ops.target2 = -1;
3977 case check_state_compute_run:
3980 pr_err("%s: unknown check_state: %d sector: %llu\n",
3981 __func__, sh->check_state,
3982 (unsigned long long) sh->sector);
3987 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3988 struct stripe_head_state *s,
3991 int pd_idx = sh->pd_idx;
3992 int qd_idx = sh->qd_idx;
3995 BUG_ON(sh->batch_head);
3996 set_bit(STRIPE_HANDLE, &sh->state);
3998 BUG_ON(s->failed > 2);
4000 /* Want to check and possibly repair P and Q.
4001 * However there could be one 'failed' device, in which
4002 * case we can only check one of them, possibly using the
4003 * other to generate missing data
4006 switch (sh->check_state) {
4007 case check_state_idle:
4008 /* start a new check operation if there are < 2 failures */
4009 if (s->failed == s->q_failed) {
4010 /* The only possible failed device holds Q, so it
4011 * makes sense to check P (If anything else were failed,
4012 * we would have used P to recreate it).
4014 sh->check_state = check_state_run;
4016 if (!s->q_failed && s->failed < 2) {
4017 /* Q is not failed, and we didn't use it to generate
4018 * anything, so it makes sense to check it
4020 if (sh->check_state == check_state_run)
4021 sh->check_state = check_state_run_pq;
4023 sh->check_state = check_state_run_q;
4026 /* discard potentially stale zero_sum_result */
4027 sh->ops.zero_sum_result = 0;
4029 if (sh->check_state == check_state_run) {
4030 /* async_xor_zero_sum destroys the contents of P */
4031 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4034 if (sh->check_state >= check_state_run &&
4035 sh->check_state <= check_state_run_pq) {
4036 /* async_syndrome_zero_sum preserves P and Q, so
4037 * no need to mark them !uptodate here
4039 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4043 /* we have 2-disk failure */
4044 BUG_ON(s->failed != 2);
4046 case check_state_compute_result:
4047 sh->check_state = check_state_idle;
4049 /* check that a write has not made the stripe insync */
4050 if (test_bit(STRIPE_INSYNC, &sh->state))
4053 /* now write out any block on a failed drive,
4054 * or P or Q if they were recomputed
4056 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
4057 if (s->failed == 2) {
4058 dev = &sh->dev[s->failed_num[1]];
4060 set_bit(R5_LOCKED, &dev->flags);
4061 set_bit(R5_Wantwrite, &dev->flags);
4063 if (s->failed >= 1) {
4064 dev = &sh->dev[s->failed_num[0]];
4066 set_bit(R5_LOCKED, &dev->flags);
4067 set_bit(R5_Wantwrite, &dev->flags);
4069 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4070 dev = &sh->dev[pd_idx];
4072 set_bit(R5_LOCKED, &dev->flags);
4073 set_bit(R5_Wantwrite, &dev->flags);
4075 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4076 dev = &sh->dev[qd_idx];
4078 set_bit(R5_LOCKED, &dev->flags);
4079 set_bit(R5_Wantwrite, &dev->flags);
4081 clear_bit(STRIPE_DEGRADED, &sh->state);
4083 set_bit(STRIPE_INSYNC, &sh->state);
4085 case check_state_run:
4086 case check_state_run_q:
4087 case check_state_run_pq:
4088 break; /* we will be called again upon completion */
4089 case check_state_check_result:
4090 sh->check_state = check_state_idle;
4092 /* handle a successful check operation, if parity is correct
4093 * we are done. Otherwise update the mismatch count and repair
4094 * parity if !MD_RECOVERY_CHECK
4096 if (sh->ops.zero_sum_result == 0) {
4097 /* both parities are correct */
4099 set_bit(STRIPE_INSYNC, &sh->state);
4101 /* in contrast to the raid5 case we can validate
4102 * parity, but still have a failure to write
4105 sh->check_state = check_state_compute_result;
4106 /* Returning at this point means that we may go
4107 * off and bring p and/or q uptodate again so
4108 * we make sure to check zero_sum_result again
4109 * to verify if p or q need writeback
4113 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4114 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
4115 /* don't try to repair!! */
4116 set_bit(STRIPE_INSYNC, &sh->state);
4118 int *target = &sh->ops.target;
4120 sh->ops.target = -1;
4121 sh->ops.target2 = -1;
4122 sh->check_state = check_state_compute_run;
4123 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4124 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4125 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4126 set_bit(R5_Wantcompute,
4127 &sh->dev[pd_idx].flags);
4129 target = &sh->ops.target2;
4132 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4133 set_bit(R5_Wantcompute,
4134 &sh->dev[qd_idx].flags);
4141 case check_state_compute_run:
4144 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4145 __func__, sh->check_state,
4146 (unsigned long long) sh->sector);
4151 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4155 /* We have read all the blocks in this stripe and now we need to
4156 * copy some of them into a target stripe for expand.
4158 struct dma_async_tx_descriptor *tx = NULL;
4159 BUG_ON(sh->batch_head);
4160 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4161 for (i = 0; i < sh->disks; i++)
4162 if (i != sh->pd_idx && i != sh->qd_idx) {
4164 struct stripe_head *sh2;
4165 struct async_submit_ctl submit;
4167 sector_t bn = raid5_compute_blocknr(sh, i, 1);
4168 sector_t s = raid5_compute_sector(conf, bn, 0,
4170 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4172 /* so far only the early blocks of this stripe
4173 * have been requested. When later blocks
4174 * get requested, we will try again
4177 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4178 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4179 /* must have already done this block */
4180 raid5_release_stripe(sh2);
4184 /* place all the copies on one channel */
4185 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4186 tx = async_memcpy(sh2->dev[dd_idx].page,
4187 sh->dev[i].page, 0, 0, STRIPE_SIZE,
4190 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4191 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4192 for (j = 0; j < conf->raid_disks; j++)
4193 if (j != sh2->pd_idx &&
4195 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4197 if (j == conf->raid_disks) {
4198 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4199 set_bit(STRIPE_HANDLE, &sh2->state);
4201 raid5_release_stripe(sh2);
4204 /* done submitting copies, wait for them to complete */
4205 async_tx_quiesce(&tx);
4209 * handle_stripe - do things to a stripe.
4211 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4212 * state of various bits to see what needs to be done.
4214 * return some read requests which now have data
4215 * return some write requests which are safely on storage
4216 * schedule a read on some buffers
4217 * schedule a write of some buffers
4218 * return confirmation of parity correctness
4222 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4224 struct r5conf *conf = sh->raid_conf;
4225 int disks = sh->disks;
4228 int do_recovery = 0;
4230 memset(s, 0, sizeof(*s));
4232 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4233 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4234 s->failed_num[0] = -1;
4235 s->failed_num[1] = -1;
4236 s->log_failed = r5l_log_disk_error(conf);
4238 /* Now to look around and see what can be done */
4240 for (i=disks; i--; ) {
4241 struct md_rdev *rdev;
4248 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4250 dev->toread, dev->towrite, dev->written);
4251 /* maybe we can reply to a read
4253 * new wantfill requests are only permitted while
4254 * ops_complete_biofill is guaranteed to be inactive
4256 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4257 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4258 set_bit(R5_Wantfill, &dev->flags);
4260 /* now count some things */
4261 if (test_bit(R5_LOCKED, &dev->flags))
4263 if (test_bit(R5_UPTODATE, &dev->flags))
4265 if (test_bit(R5_Wantcompute, &dev->flags)) {
4267 BUG_ON(s->compute > 2);
4270 if (test_bit(R5_Wantfill, &dev->flags))
4272 else if (dev->toread)
4276 if (!test_bit(R5_OVERWRITE, &dev->flags))
4281 /* Prefer to use the replacement for reads, but only
4282 * if it is recovered enough and has no bad blocks.
4284 rdev = rcu_dereference(conf->disks[i].replacement);
4285 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4286 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4287 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4288 &first_bad, &bad_sectors))
4289 set_bit(R5_ReadRepl, &dev->flags);
4291 if (rdev && !test_bit(Faulty, &rdev->flags))
4292 set_bit(R5_NeedReplace, &dev->flags);
4294 clear_bit(R5_NeedReplace, &dev->flags);
4295 rdev = rcu_dereference(conf->disks[i].rdev);
4296 clear_bit(R5_ReadRepl, &dev->flags);
4298 if (rdev && test_bit(Faulty, &rdev->flags))
4301 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4302 &first_bad, &bad_sectors);
4303 if (s->blocked_rdev == NULL
4304 && (test_bit(Blocked, &rdev->flags)
4307 set_bit(BlockedBadBlocks,
4309 s->blocked_rdev = rdev;
4310 atomic_inc(&rdev->nr_pending);
4313 clear_bit(R5_Insync, &dev->flags);
4317 /* also not in-sync */
4318 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4319 test_bit(R5_UPTODATE, &dev->flags)) {
4320 /* treat as in-sync, but with a read error
4321 * which we can now try to correct
4323 set_bit(R5_Insync, &dev->flags);
4324 set_bit(R5_ReadError, &dev->flags);
4326 } else if (test_bit(In_sync, &rdev->flags))
4327 set_bit(R5_Insync, &dev->flags);
4328 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4329 /* in sync if before recovery_offset */
4330 set_bit(R5_Insync, &dev->flags);
4331 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4332 test_bit(R5_Expanded, &dev->flags))
4333 /* If we've reshaped into here, we assume it is Insync.
4334 * We will shortly update recovery_offset to make
4337 set_bit(R5_Insync, &dev->flags);
4339 if (test_bit(R5_WriteError, &dev->flags)) {
4340 /* This flag does not apply to '.replacement'
4341 * only to .rdev, so make sure to check that*/
4342 struct md_rdev *rdev2 = rcu_dereference(
4343 conf->disks[i].rdev);
4345 clear_bit(R5_Insync, &dev->flags);
4346 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4347 s->handle_bad_blocks = 1;
4348 atomic_inc(&rdev2->nr_pending);
4350 clear_bit(R5_WriteError, &dev->flags);
4352 if (test_bit(R5_MadeGood, &dev->flags)) {
4353 /* This flag does not apply to '.replacement'
4354 * only to .rdev, so make sure to check that*/
4355 struct md_rdev *rdev2 = rcu_dereference(
4356 conf->disks[i].rdev);
4357 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4358 s->handle_bad_blocks = 1;
4359 atomic_inc(&rdev2->nr_pending);
4361 clear_bit(R5_MadeGood, &dev->flags);
4363 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4364 struct md_rdev *rdev2 = rcu_dereference(
4365 conf->disks[i].replacement);
4366 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4367 s->handle_bad_blocks = 1;
4368 atomic_inc(&rdev2->nr_pending);
4370 clear_bit(R5_MadeGoodRepl, &dev->flags);
4372 if (!test_bit(R5_Insync, &dev->flags)) {
4373 /* The ReadError flag will just be confusing now */
4374 clear_bit(R5_ReadError, &dev->flags);
4375 clear_bit(R5_ReWrite, &dev->flags);
4377 if (test_bit(R5_ReadError, &dev->flags))
4378 clear_bit(R5_Insync, &dev->flags);
4379 if (!test_bit(R5_Insync, &dev->flags)) {
4381 s->failed_num[s->failed] = i;
4383 if (rdev && !test_bit(Faulty, &rdev->flags))
4387 if (test_bit(R5_InJournal, &dev->flags))
4389 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4392 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4393 /* If there is a failed device being replaced,
4394 * we must be recovering.
4395 * else if we are after recovery_cp, we must be syncing
4396 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4397 * else we can only be replacing
4398 * sync and recovery both need to read all devices, and so
4399 * use the same flag.
4402 sh->sector >= conf->mddev->recovery_cp ||
4403 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4411 static int clear_batch_ready(struct stripe_head *sh)
4413 /* Return '1' if this is a member of batch, or
4414 * '0' if it is a lone stripe or a head which can now be
4417 struct stripe_head *tmp;
4418 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4419 return (sh->batch_head && sh->batch_head != sh);
4420 spin_lock(&sh->stripe_lock);
4421 if (!sh->batch_head) {
4422 spin_unlock(&sh->stripe_lock);
4427 * this stripe could be added to a batch list before we check
4428 * BATCH_READY, skips it
4430 if (sh->batch_head != sh) {
4431 spin_unlock(&sh->stripe_lock);
4434 spin_lock(&sh->batch_lock);
4435 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4436 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4437 spin_unlock(&sh->batch_lock);
4438 spin_unlock(&sh->stripe_lock);
4441 * BATCH_READY is cleared, no new stripes can be added.
4442 * batch_list can be accessed without lock
4447 static void break_stripe_batch_list(struct stripe_head *head_sh,
4448 unsigned long handle_flags)
4450 struct stripe_head *sh, *next;
4454 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4456 list_del_init(&sh->batch_list);
4458 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4459 (1 << STRIPE_SYNCING) |
4460 (1 << STRIPE_REPLACED) |
4461 (1 << STRIPE_DELAYED) |
4462 (1 << STRIPE_BIT_DELAY) |
4463 (1 << STRIPE_FULL_WRITE) |
4464 (1 << STRIPE_BIOFILL_RUN) |
4465 (1 << STRIPE_COMPUTE_RUN) |
4466 (1 << STRIPE_OPS_REQ_PENDING) |
4467 (1 << STRIPE_DISCARD) |
4468 (1 << STRIPE_BATCH_READY) |
4469 (1 << STRIPE_BATCH_ERR) |
4470 (1 << STRIPE_BITMAP_PENDING)),
4471 "stripe state: %lx\n", sh->state);
4472 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4473 (1 << STRIPE_REPLACED)),
4474 "head stripe state: %lx\n", head_sh->state);
4476 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4477 (1 << STRIPE_PREREAD_ACTIVE) |
4478 (1 << STRIPE_DEGRADED)),
4479 head_sh->state & (1 << STRIPE_INSYNC));
4481 sh->check_state = head_sh->check_state;
4482 sh->reconstruct_state = head_sh->reconstruct_state;
4483 for (i = 0; i < sh->disks; i++) {
4484 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4486 sh->dev[i].flags = head_sh->dev[i].flags &
4487 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4489 spin_lock_irq(&sh->stripe_lock);
4490 sh->batch_head = NULL;
4491 spin_unlock_irq(&sh->stripe_lock);
4492 if (handle_flags == 0 ||
4493 sh->state & handle_flags)
4494 set_bit(STRIPE_HANDLE, &sh->state);
4495 raid5_release_stripe(sh);
4497 spin_lock_irq(&head_sh->stripe_lock);
4498 head_sh->batch_head = NULL;
4499 spin_unlock_irq(&head_sh->stripe_lock);
4500 for (i = 0; i < head_sh->disks; i++)
4501 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4503 if (head_sh->state & handle_flags)
4504 set_bit(STRIPE_HANDLE, &head_sh->state);
4507 wake_up(&head_sh->raid_conf->wait_for_overlap);
4510 static void handle_stripe(struct stripe_head *sh)
4512 struct stripe_head_state s;
4513 struct r5conf *conf = sh->raid_conf;
4516 int disks = sh->disks;
4517 struct r5dev *pdev, *qdev;
4519 clear_bit(STRIPE_HANDLE, &sh->state);
4520 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4521 /* already being handled, ensure it gets handled
4522 * again when current action finishes */
4523 set_bit(STRIPE_HANDLE, &sh->state);
4527 if (clear_batch_ready(sh) ) {
4528 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4532 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4533 break_stripe_batch_list(sh, 0);
4535 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4536 spin_lock(&sh->stripe_lock);
4537 /* Cannot process 'sync' concurrently with 'discard' */
4538 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4539 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4540 set_bit(STRIPE_SYNCING, &sh->state);
4541 clear_bit(STRIPE_INSYNC, &sh->state);
4542 clear_bit(STRIPE_REPLACED, &sh->state);
4544 spin_unlock(&sh->stripe_lock);
4546 clear_bit(STRIPE_DELAYED, &sh->state);
4548 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4549 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4550 (unsigned long long)sh->sector, sh->state,
4551 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4552 sh->check_state, sh->reconstruct_state);
4554 analyse_stripe(sh, &s);
4556 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4559 if (s.handle_bad_blocks) {
4560 set_bit(STRIPE_HANDLE, &sh->state);
4564 if (unlikely(s.blocked_rdev)) {
4565 if (s.syncing || s.expanding || s.expanded ||
4566 s.replacing || s.to_write || s.written) {
4567 set_bit(STRIPE_HANDLE, &sh->state);
4570 /* There is nothing for the blocked_rdev to block */
4571 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4572 s.blocked_rdev = NULL;
4575 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4576 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4577 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4580 pr_debug("locked=%d uptodate=%d to_read=%d"
4581 " to_write=%d failed=%d failed_num=%d,%d\n",
4582 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4583 s.failed_num[0], s.failed_num[1]);
4584 /* check if the array has lost more than max_degraded devices and,
4585 * if so, some requests might need to be failed.
4587 if (s.failed > conf->max_degraded || s.log_failed) {
4588 sh->check_state = 0;
4589 sh->reconstruct_state = 0;
4590 break_stripe_batch_list(sh, 0);
4591 if (s.to_read+s.to_write+s.written)
4592 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4593 if (s.syncing + s.replacing)
4594 handle_failed_sync(conf, sh, &s);
4597 /* Now we check to see if any write operations have recently
4601 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4603 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4604 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4605 sh->reconstruct_state = reconstruct_state_idle;
4607 /* All the 'written' buffers and the parity block are ready to
4608 * be written back to disk
4610 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4611 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4612 BUG_ON(sh->qd_idx >= 0 &&
4613 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4614 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4615 for (i = disks; i--; ) {
4616 struct r5dev *dev = &sh->dev[i];
4617 if (test_bit(R5_LOCKED, &dev->flags) &&
4618 (i == sh->pd_idx || i == sh->qd_idx ||
4619 dev->written || test_bit(R5_InJournal,
4621 pr_debug("Writing block %d\n", i);
4622 set_bit(R5_Wantwrite, &dev->flags);
4627 if (!test_bit(R5_Insync, &dev->flags) ||
4628 ((i == sh->pd_idx || i == sh->qd_idx) &&
4630 set_bit(STRIPE_INSYNC, &sh->state);
4633 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4634 s.dec_preread_active = 1;
4638 * might be able to return some write requests if the parity blocks
4639 * are safe, or on a failed drive
4641 pdev = &sh->dev[sh->pd_idx];
4642 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4643 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4644 qdev = &sh->dev[sh->qd_idx];
4645 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4646 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4650 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4651 && !test_bit(R5_LOCKED, &pdev->flags)
4652 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4653 test_bit(R5_Discard, &pdev->flags))))) &&
4654 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4655 && !test_bit(R5_LOCKED, &qdev->flags)
4656 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4657 test_bit(R5_Discard, &qdev->flags))))))
4658 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4661 r5c_handle_cached_data_endio(conf, sh, disks, &s.return_bi);
4662 r5l_stripe_write_finished(sh);
4664 /* Now we might consider reading some blocks, either to check/generate
4665 * parity, or to satisfy requests
4666 * or to load a block that is being partially written.
4668 if (s.to_read || s.non_overwrite
4669 || (conf->level == 6 && s.to_write && s.failed)
4670 || (s.syncing && (s.uptodate + s.compute < disks))
4673 handle_stripe_fill(sh, &s, disks);
4676 * When the stripe finishes full journal write cycle (write to journal
4677 * and raid disk), this is the clean up procedure so it is ready for
4680 r5c_finish_stripe_write_out(conf, sh, &s);
4683 * Now to consider new write requests, cache write back and what else,
4684 * if anything should be read. We do not handle new writes when:
4685 * 1/ A 'write' operation (copy+xor) is already in flight.
4686 * 2/ A 'check' operation is in flight, as it may clobber the parity
4688 * 3/ A r5c cache log write is in flight.
4691 if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4692 if (!r5c_is_writeback(conf->log)) {
4694 handle_stripe_dirtying(conf, sh, &s, disks);
4695 } else { /* write back cache */
4698 /* First, try handle writes in caching phase */
4700 ret = r5c_try_caching_write(conf, sh, &s,
4703 * If caching phase failed: ret == -EAGAIN
4705 * stripe under reclaim: !caching && injournal
4707 * fall back to handle_stripe_dirtying()
4709 if (ret == -EAGAIN ||
4710 /* stripe under reclaim: !caching && injournal */
4711 (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4713 ret = handle_stripe_dirtying(conf, sh, &s,
4721 /* maybe we need to check and possibly fix the parity for this stripe
4722 * Any reads will already have been scheduled, so we just see if enough
4723 * data is available. The parity check is held off while parity
4724 * dependent operations are in flight.
4726 if (sh->check_state ||
4727 (s.syncing && s.locked == 0 &&
4728 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4729 !test_bit(STRIPE_INSYNC, &sh->state))) {
4730 if (conf->level == 6)
4731 handle_parity_checks6(conf, sh, &s, disks);
4733 handle_parity_checks5(conf, sh, &s, disks);
4736 if ((s.replacing || s.syncing) && s.locked == 0
4737 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4738 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4739 /* Write out to replacement devices where possible */
4740 for (i = 0; i < conf->raid_disks; i++)
4741 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4742 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4743 set_bit(R5_WantReplace, &sh->dev[i].flags);
4744 set_bit(R5_LOCKED, &sh->dev[i].flags);
4748 set_bit(STRIPE_INSYNC, &sh->state);
4749 set_bit(STRIPE_REPLACED, &sh->state);
4751 if ((s.syncing || s.replacing) && s.locked == 0 &&
4752 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4753 test_bit(STRIPE_INSYNC, &sh->state)) {
4754 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4755 clear_bit(STRIPE_SYNCING, &sh->state);
4756 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4757 wake_up(&conf->wait_for_overlap);
4760 /* If the failed drives are just a ReadError, then we might need
4761 * to progress the repair/check process
4763 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4764 for (i = 0; i < s.failed; i++) {
4765 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4766 if (test_bit(R5_ReadError, &dev->flags)
4767 && !test_bit(R5_LOCKED, &dev->flags)
4768 && test_bit(R5_UPTODATE, &dev->flags)
4770 if (!test_bit(R5_ReWrite, &dev->flags)) {
4771 set_bit(R5_Wantwrite, &dev->flags);
4772 set_bit(R5_ReWrite, &dev->flags);
4773 set_bit(R5_LOCKED, &dev->flags);
4776 /* let's read it back */
4777 set_bit(R5_Wantread, &dev->flags);
4778 set_bit(R5_LOCKED, &dev->flags);
4784 /* Finish reconstruct operations initiated by the expansion process */
4785 if (sh->reconstruct_state == reconstruct_state_result) {
4786 struct stripe_head *sh_src
4787 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4788 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4789 /* sh cannot be written until sh_src has been read.
4790 * so arrange for sh to be delayed a little
4792 set_bit(STRIPE_DELAYED, &sh->state);
4793 set_bit(STRIPE_HANDLE, &sh->state);
4794 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4796 atomic_inc(&conf->preread_active_stripes);
4797 raid5_release_stripe(sh_src);
4801 raid5_release_stripe(sh_src);
4803 sh->reconstruct_state = reconstruct_state_idle;
4804 clear_bit(STRIPE_EXPANDING, &sh->state);
4805 for (i = conf->raid_disks; i--; ) {
4806 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4807 set_bit(R5_LOCKED, &sh->dev[i].flags);
4812 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4813 !sh->reconstruct_state) {
4814 /* Need to write out all blocks after computing parity */
4815 sh->disks = conf->raid_disks;
4816 stripe_set_idx(sh->sector, conf, 0, sh);
4817 schedule_reconstruction(sh, &s, 1, 1);
4818 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4819 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4820 atomic_dec(&conf->reshape_stripes);
4821 wake_up(&conf->wait_for_overlap);
4822 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4825 if (s.expanding && s.locked == 0 &&
4826 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4827 handle_stripe_expansion(conf, sh);
4830 /* wait for this device to become unblocked */
4831 if (unlikely(s.blocked_rdev)) {
4832 if (conf->mddev->external)
4833 md_wait_for_blocked_rdev(s.blocked_rdev,
4836 /* Internal metadata will immediately
4837 * be written by raid5d, so we don't
4838 * need to wait here.
4840 rdev_dec_pending(s.blocked_rdev,
4844 if (s.handle_bad_blocks)
4845 for (i = disks; i--; ) {
4846 struct md_rdev *rdev;
4847 struct r5dev *dev = &sh->dev[i];
4848 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4849 /* We own a safe reference to the rdev */
4850 rdev = conf->disks[i].rdev;
4851 if (!rdev_set_badblocks(rdev, sh->sector,
4853 md_error(conf->mddev, rdev);
4854 rdev_dec_pending(rdev, conf->mddev);
4856 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4857 rdev = conf->disks[i].rdev;
4858 rdev_clear_badblocks(rdev, sh->sector,
4860 rdev_dec_pending(rdev, conf->mddev);
4862 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4863 rdev = conf->disks[i].replacement;
4865 /* rdev have been moved down */
4866 rdev = conf->disks[i].rdev;
4867 rdev_clear_badblocks(rdev, sh->sector,
4869 rdev_dec_pending(rdev, conf->mddev);
4874 raid_run_ops(sh, s.ops_request);
4878 if (s.dec_preread_active) {
4879 /* We delay this until after ops_run_io so that if make_request
4880 * is waiting on a flush, it won't continue until the writes
4881 * have actually been submitted.
4883 atomic_dec(&conf->preread_active_stripes);
4884 if (atomic_read(&conf->preread_active_stripes) <
4886 md_wakeup_thread(conf->mddev->thread);
4889 if (!bio_list_empty(&s.return_bi)) {
4890 if (test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4891 spin_lock_irq(&conf->device_lock);
4892 bio_list_merge(&conf->return_bi, &s.return_bi);
4893 spin_unlock_irq(&conf->device_lock);
4894 md_wakeup_thread(conf->mddev->thread);
4896 return_io(&s.return_bi);
4899 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4902 static void raid5_activate_delayed(struct r5conf *conf)
4904 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4905 while (!list_empty(&conf->delayed_list)) {
4906 struct list_head *l = conf->delayed_list.next;
4907 struct stripe_head *sh;
4908 sh = list_entry(l, struct stripe_head, lru);
4910 clear_bit(STRIPE_DELAYED, &sh->state);
4911 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4912 atomic_inc(&conf->preread_active_stripes);
4913 list_add_tail(&sh->lru, &conf->hold_list);
4914 raid5_wakeup_stripe_thread(sh);
4919 static void activate_bit_delay(struct r5conf *conf,
4920 struct list_head *temp_inactive_list)
4922 /* device_lock is held */
4923 struct list_head head;
4924 list_add(&head, &conf->bitmap_list);
4925 list_del_init(&conf->bitmap_list);
4926 while (!list_empty(&head)) {
4927 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4929 list_del_init(&sh->lru);
4930 atomic_inc(&sh->count);
4931 hash = sh->hash_lock_index;
4932 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4936 static int raid5_congested(struct mddev *mddev, int bits)
4938 struct r5conf *conf = mddev->private;
4940 /* No difference between reads and writes. Just check
4941 * how busy the stripe_cache is
4944 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4947 /* Also checks whether there is pressure on r5cache log space */
4948 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
4952 if (atomic_read(&conf->empty_inactive_list_nr))
4958 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4960 struct r5conf *conf = mddev->private;
4961 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4962 unsigned int chunk_sectors;
4963 unsigned int bio_sectors = bio_sectors(bio);
4965 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4966 return chunk_sectors >=
4967 ((sector & (chunk_sectors - 1)) + bio_sectors);
4971 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4972 * later sampled by raid5d.
4974 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4976 unsigned long flags;
4978 spin_lock_irqsave(&conf->device_lock, flags);
4980 bi->bi_next = conf->retry_read_aligned_list;
4981 conf->retry_read_aligned_list = bi;
4983 spin_unlock_irqrestore(&conf->device_lock, flags);
4984 md_wakeup_thread(conf->mddev->thread);
4987 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4991 bi = conf->retry_read_aligned;
4993 conf->retry_read_aligned = NULL;
4996 bi = conf->retry_read_aligned_list;
4998 conf->retry_read_aligned_list = bi->bi_next;
5001 * this sets the active strip count to 1 and the processed
5002 * strip count to zero (upper 8 bits)
5004 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
5011 * The "raid5_align_endio" should check if the read succeeded and if it
5012 * did, call bio_endio on the original bio (having bio_put the new bio
5014 * If the read failed..
5016 static void raid5_align_endio(struct bio *bi)
5018 struct bio* raid_bi = bi->bi_private;
5019 struct mddev *mddev;
5020 struct r5conf *conf;
5021 struct md_rdev *rdev;
5022 int error = bi->bi_error;
5026 rdev = (void*)raid_bi->bi_next;
5027 raid_bi->bi_next = NULL;
5028 mddev = rdev->mddev;
5029 conf = mddev->private;
5031 rdev_dec_pending(rdev, conf->mddev);
5034 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
5037 if (atomic_dec_and_test(&conf->active_aligned_reads))
5038 wake_up(&conf->wait_for_quiescent);
5042 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5044 add_bio_to_retry(raid_bi, conf);
5047 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5049 struct r5conf *conf = mddev->private;
5051 struct bio* align_bi;
5052 struct md_rdev *rdev;
5053 sector_t end_sector;
5055 if (!in_chunk_boundary(mddev, raid_bio)) {
5056 pr_debug("%s: non aligned\n", __func__);
5060 * use bio_clone_fast to make a copy of the bio
5062 align_bi = bio_clone_fast(raid_bio, GFP_NOIO, mddev->bio_set);
5066 * set bi_end_io to a new function, and set bi_private to the
5069 align_bi->bi_end_io = raid5_align_endio;
5070 align_bi->bi_private = raid_bio;
5074 align_bi->bi_iter.bi_sector =
5075 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5078 end_sector = bio_end_sector(align_bi);
5080 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5081 if (!rdev || test_bit(Faulty, &rdev->flags) ||
5082 rdev->recovery_offset < end_sector) {
5083 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5085 (test_bit(Faulty, &rdev->flags) ||
5086 !(test_bit(In_sync, &rdev->flags) ||
5087 rdev->recovery_offset >= end_sector)))
5091 if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5101 atomic_inc(&rdev->nr_pending);
5103 raid_bio->bi_next = (void*)rdev;
5104 align_bi->bi_bdev = rdev->bdev;
5105 bio_clear_flag(align_bi, BIO_SEG_VALID);
5107 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5108 bio_sectors(align_bi),
5109 &first_bad, &bad_sectors)) {
5111 rdev_dec_pending(rdev, mddev);
5115 /* No reshape active, so we can trust rdev->data_offset */
5116 align_bi->bi_iter.bi_sector += rdev->data_offset;
5118 spin_lock_irq(&conf->device_lock);
5119 wait_event_lock_irq(conf->wait_for_quiescent,
5122 atomic_inc(&conf->active_aligned_reads);
5123 spin_unlock_irq(&conf->device_lock);
5126 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
5127 align_bi, disk_devt(mddev->gendisk),
5128 raid_bio->bi_iter.bi_sector);
5129 generic_make_request(align_bi);
5138 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5143 sector_t sector = raid_bio->bi_iter.bi_sector;
5144 unsigned chunk_sects = mddev->chunk_sectors;
5145 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5147 if (sectors < bio_sectors(raid_bio)) {
5148 split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
5149 bio_chain(split, raid_bio);
5153 if (!raid5_read_one_chunk(mddev, split)) {
5154 if (split != raid_bio)
5155 generic_make_request(raid_bio);
5158 } while (split != raid_bio);
5163 /* __get_priority_stripe - get the next stripe to process
5165 * Full stripe writes are allowed to pass preread active stripes up until
5166 * the bypass_threshold is exceeded. In general the bypass_count
5167 * increments when the handle_list is handled before the hold_list; however, it
5168 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5169 * stripe with in flight i/o. The bypass_count will be reset when the
5170 * head of the hold_list has changed, i.e. the head was promoted to the
5173 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5175 struct stripe_head *sh = NULL, *tmp;
5176 struct list_head *handle_list = NULL;
5177 struct r5worker_group *wg = NULL;
5179 if (conf->worker_cnt_per_group == 0) {
5180 handle_list = &conf->handle_list;
5181 } else if (group != ANY_GROUP) {
5182 handle_list = &conf->worker_groups[group].handle_list;
5183 wg = &conf->worker_groups[group];
5186 for (i = 0; i < conf->group_cnt; i++) {
5187 handle_list = &conf->worker_groups[i].handle_list;
5188 wg = &conf->worker_groups[i];
5189 if (!list_empty(handle_list))
5194 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5196 list_empty(handle_list) ? "empty" : "busy",
5197 list_empty(&conf->hold_list) ? "empty" : "busy",
5198 atomic_read(&conf->pending_full_writes), conf->bypass_count);
5200 if (!list_empty(handle_list)) {
5201 sh = list_entry(handle_list->next, typeof(*sh), lru);
5203 if (list_empty(&conf->hold_list))
5204 conf->bypass_count = 0;
5205 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5206 if (conf->hold_list.next == conf->last_hold)
5207 conf->bypass_count++;
5209 conf->last_hold = conf->hold_list.next;
5210 conf->bypass_count -= conf->bypass_threshold;
5211 if (conf->bypass_count < 0)
5212 conf->bypass_count = 0;
5215 } else if (!list_empty(&conf->hold_list) &&
5216 ((conf->bypass_threshold &&
5217 conf->bypass_count > conf->bypass_threshold) ||
5218 atomic_read(&conf->pending_full_writes) == 0)) {
5220 list_for_each_entry(tmp, &conf->hold_list, lru) {
5221 if (conf->worker_cnt_per_group == 0 ||
5222 group == ANY_GROUP ||
5223 !cpu_online(tmp->cpu) ||
5224 cpu_to_group(tmp->cpu) == group) {
5231 conf->bypass_count -= conf->bypass_threshold;
5232 if (conf->bypass_count < 0)
5233 conf->bypass_count = 0;
5245 list_del_init(&sh->lru);
5246 BUG_ON(atomic_inc_return(&sh->count) != 1);
5250 struct raid5_plug_cb {
5251 struct blk_plug_cb cb;
5252 struct list_head list;
5253 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5256 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5258 struct raid5_plug_cb *cb = container_of(
5259 blk_cb, struct raid5_plug_cb, cb);
5260 struct stripe_head *sh;
5261 struct mddev *mddev = cb->cb.data;
5262 struct r5conf *conf = mddev->private;
5266 if (cb->list.next && !list_empty(&cb->list)) {
5267 spin_lock_irq(&conf->device_lock);
5268 while (!list_empty(&cb->list)) {
5269 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5270 list_del_init(&sh->lru);
5272 * avoid race release_stripe_plug() sees
5273 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5274 * is still in our list
5276 smp_mb__before_atomic();
5277 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5279 * STRIPE_ON_RELEASE_LIST could be set here. In that
5280 * case, the count is always > 1 here
5282 hash = sh->hash_lock_index;
5283 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5286 spin_unlock_irq(&conf->device_lock);
5288 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5289 NR_STRIPE_HASH_LOCKS);
5291 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5295 static void release_stripe_plug(struct mddev *mddev,
5296 struct stripe_head *sh)
5298 struct blk_plug_cb *blk_cb = blk_check_plugged(
5299 raid5_unplug, mddev,
5300 sizeof(struct raid5_plug_cb));
5301 struct raid5_plug_cb *cb;
5304 raid5_release_stripe(sh);
5308 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5310 if (cb->list.next == NULL) {
5312 INIT_LIST_HEAD(&cb->list);
5313 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5314 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5317 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5318 list_add_tail(&sh->lru, &cb->list);
5320 raid5_release_stripe(sh);
5323 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5325 struct r5conf *conf = mddev->private;
5326 sector_t logical_sector, last_sector;
5327 struct stripe_head *sh;
5331 if (mddev->reshape_position != MaxSector)
5332 /* Skip discard while reshape is happening */
5335 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5336 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5339 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5341 stripe_sectors = conf->chunk_sectors *
5342 (conf->raid_disks - conf->max_degraded);
5343 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5345 sector_div(last_sector, stripe_sectors);
5347 logical_sector *= conf->chunk_sectors;
5348 last_sector *= conf->chunk_sectors;
5350 for (; logical_sector < last_sector;
5351 logical_sector += STRIPE_SECTORS) {
5355 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5356 prepare_to_wait(&conf->wait_for_overlap, &w,
5357 TASK_UNINTERRUPTIBLE);
5358 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5359 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5360 raid5_release_stripe(sh);
5364 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5365 spin_lock_irq(&sh->stripe_lock);
5366 for (d = 0; d < conf->raid_disks; d++) {
5367 if (d == sh->pd_idx || d == sh->qd_idx)
5369 if (sh->dev[d].towrite || sh->dev[d].toread) {
5370 set_bit(R5_Overlap, &sh->dev[d].flags);
5371 spin_unlock_irq(&sh->stripe_lock);
5372 raid5_release_stripe(sh);
5377 set_bit(STRIPE_DISCARD, &sh->state);
5378 finish_wait(&conf->wait_for_overlap, &w);
5379 sh->overwrite_disks = 0;
5380 for (d = 0; d < conf->raid_disks; d++) {
5381 if (d == sh->pd_idx || d == sh->qd_idx)
5383 sh->dev[d].towrite = bi;
5384 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5385 raid5_inc_bi_active_stripes(bi);
5386 sh->overwrite_disks++;
5388 spin_unlock_irq(&sh->stripe_lock);
5389 if (conf->mddev->bitmap) {
5391 d < conf->raid_disks - conf->max_degraded;
5393 bitmap_startwrite(mddev->bitmap,
5397 sh->bm_seq = conf->seq_flush + 1;
5398 set_bit(STRIPE_BIT_DELAY, &sh->state);
5401 set_bit(STRIPE_HANDLE, &sh->state);
5402 clear_bit(STRIPE_DELAYED, &sh->state);
5403 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5404 atomic_inc(&conf->preread_active_stripes);
5405 release_stripe_plug(mddev, sh);
5408 remaining = raid5_dec_bi_active_stripes(bi);
5409 if (remaining == 0) {
5410 md_write_end(mddev);
5415 static void raid5_make_request(struct mddev *mddev, struct bio * bi)
5417 struct r5conf *conf = mddev->private;
5419 sector_t new_sector;
5420 sector_t logical_sector, last_sector;
5421 struct stripe_head *sh;
5422 const int rw = bio_data_dir(bi);
5426 bool do_flush = false;
5428 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5429 int ret = r5l_handle_flush_request(conf->log, bi);
5433 if (ret == -ENODEV) {
5434 md_flush_request(mddev, bi);
5437 /* ret == -EAGAIN, fallback */
5439 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5440 * we need to flush journal device
5442 do_flush = bi->bi_opf & REQ_PREFLUSH;
5445 md_write_start(mddev, bi);
5448 * If array is degraded, better not do chunk aligned read because
5449 * later we might have to read it again in order to reconstruct
5450 * data on failed drives.
5452 if (rw == READ && mddev->degraded == 0 &&
5453 mddev->reshape_position == MaxSector) {
5454 bi = chunk_aligned_read(mddev, bi);
5459 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5460 make_discard_request(mddev, bi);
5464 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5465 last_sector = bio_end_sector(bi);
5467 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5469 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5470 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5476 seq = read_seqcount_begin(&conf->gen_lock);
5479 prepare_to_wait(&conf->wait_for_overlap, &w,
5480 TASK_UNINTERRUPTIBLE);
5481 if (unlikely(conf->reshape_progress != MaxSector)) {
5482 /* spinlock is needed as reshape_progress may be
5483 * 64bit on a 32bit platform, and so it might be
5484 * possible to see a half-updated value
5485 * Of course reshape_progress could change after
5486 * the lock is dropped, so once we get a reference
5487 * to the stripe that we think it is, we will have
5490 spin_lock_irq(&conf->device_lock);
5491 if (mddev->reshape_backwards
5492 ? logical_sector < conf->reshape_progress
5493 : logical_sector >= conf->reshape_progress) {
5496 if (mddev->reshape_backwards
5497 ? logical_sector < conf->reshape_safe
5498 : logical_sector >= conf->reshape_safe) {
5499 spin_unlock_irq(&conf->device_lock);
5505 spin_unlock_irq(&conf->device_lock);
5508 new_sector = raid5_compute_sector(conf, logical_sector,
5511 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5512 (unsigned long long)new_sector,
5513 (unsigned long long)logical_sector);
5515 sh = raid5_get_active_stripe(conf, new_sector, previous,
5516 (bi->bi_opf & REQ_RAHEAD), 0);
5518 if (unlikely(previous)) {
5519 /* expansion might have moved on while waiting for a
5520 * stripe, so we must do the range check again.
5521 * Expansion could still move past after this
5522 * test, but as we are holding a reference to
5523 * 'sh', we know that if that happens,
5524 * STRIPE_EXPANDING will get set and the expansion
5525 * won't proceed until we finish with the stripe.
5528 spin_lock_irq(&conf->device_lock);
5529 if (mddev->reshape_backwards
5530 ? logical_sector >= conf->reshape_progress
5531 : logical_sector < conf->reshape_progress)
5532 /* mismatch, need to try again */
5534 spin_unlock_irq(&conf->device_lock);
5536 raid5_release_stripe(sh);
5542 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5543 /* Might have got the wrong stripe_head
5546 raid5_release_stripe(sh);
5551 logical_sector >= mddev->suspend_lo &&
5552 logical_sector < mddev->suspend_hi) {
5553 raid5_release_stripe(sh);
5554 /* As the suspend_* range is controlled by
5555 * userspace, we want an interruptible
5558 flush_signals(current);
5559 prepare_to_wait(&conf->wait_for_overlap,
5560 &w, TASK_INTERRUPTIBLE);
5561 if (logical_sector >= mddev->suspend_lo &&
5562 logical_sector < mddev->suspend_hi) {
5569 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5570 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5571 /* Stripe is busy expanding or
5572 * add failed due to overlap. Flush everything
5575 md_wakeup_thread(mddev->thread);
5576 raid5_release_stripe(sh);
5582 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5583 /* we only need flush for one stripe */
5587 set_bit(STRIPE_HANDLE, &sh->state);
5588 clear_bit(STRIPE_DELAYED, &sh->state);
5589 if ((!sh->batch_head || sh == sh->batch_head) &&
5590 (bi->bi_opf & REQ_SYNC) &&
5591 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5592 atomic_inc(&conf->preread_active_stripes);
5593 release_stripe_plug(mddev, sh);
5595 /* cannot get stripe for read-ahead, just give-up */
5596 bi->bi_error = -EIO;
5600 finish_wait(&conf->wait_for_overlap, &w);
5602 remaining = raid5_dec_bi_active_stripes(bi);
5603 if (remaining == 0) {
5606 md_write_end(mddev);
5608 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5614 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5616 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5618 /* reshaping is quite different to recovery/resync so it is
5619 * handled quite separately ... here.
5621 * On each call to sync_request, we gather one chunk worth of
5622 * destination stripes and flag them as expanding.
5623 * Then we find all the source stripes and request reads.
5624 * As the reads complete, handle_stripe will copy the data
5625 * into the destination stripe and release that stripe.
5627 struct r5conf *conf = mddev->private;
5628 struct stripe_head *sh;
5629 sector_t first_sector, last_sector;
5630 int raid_disks = conf->previous_raid_disks;
5631 int data_disks = raid_disks - conf->max_degraded;
5632 int new_data_disks = conf->raid_disks - conf->max_degraded;
5635 sector_t writepos, readpos, safepos;
5636 sector_t stripe_addr;
5637 int reshape_sectors;
5638 struct list_head stripes;
5641 if (sector_nr == 0) {
5642 /* If restarting in the middle, skip the initial sectors */
5643 if (mddev->reshape_backwards &&
5644 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5645 sector_nr = raid5_size(mddev, 0, 0)
5646 - conf->reshape_progress;
5647 } else if (mddev->reshape_backwards &&
5648 conf->reshape_progress == MaxSector) {
5649 /* shouldn't happen, but just in case, finish up.*/
5650 sector_nr = MaxSector;
5651 } else if (!mddev->reshape_backwards &&
5652 conf->reshape_progress > 0)
5653 sector_nr = conf->reshape_progress;
5654 sector_div(sector_nr, new_data_disks);
5656 mddev->curr_resync_completed = sector_nr;
5657 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5664 /* We need to process a full chunk at a time.
5665 * If old and new chunk sizes differ, we need to process the
5669 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5671 /* We update the metadata at least every 10 seconds, or when
5672 * the data about to be copied would over-write the source of
5673 * the data at the front of the range. i.e. one new_stripe
5674 * along from reshape_progress new_maps to after where
5675 * reshape_safe old_maps to
5677 writepos = conf->reshape_progress;
5678 sector_div(writepos, new_data_disks);
5679 readpos = conf->reshape_progress;
5680 sector_div(readpos, data_disks);
5681 safepos = conf->reshape_safe;
5682 sector_div(safepos, data_disks);
5683 if (mddev->reshape_backwards) {
5684 BUG_ON(writepos < reshape_sectors);
5685 writepos -= reshape_sectors;
5686 readpos += reshape_sectors;
5687 safepos += reshape_sectors;
5689 writepos += reshape_sectors;
5690 /* readpos and safepos are worst-case calculations.
5691 * A negative number is overly pessimistic, and causes
5692 * obvious problems for unsigned storage. So clip to 0.
5694 readpos -= min_t(sector_t, reshape_sectors, readpos);
5695 safepos -= min_t(sector_t, reshape_sectors, safepos);
5698 /* Having calculated the 'writepos' possibly use it
5699 * to set 'stripe_addr' which is where we will write to.
5701 if (mddev->reshape_backwards) {
5702 BUG_ON(conf->reshape_progress == 0);
5703 stripe_addr = writepos;
5704 BUG_ON((mddev->dev_sectors &
5705 ~((sector_t)reshape_sectors - 1))
5706 - reshape_sectors - stripe_addr
5709 BUG_ON(writepos != sector_nr + reshape_sectors);
5710 stripe_addr = sector_nr;
5713 /* 'writepos' is the most advanced device address we might write.
5714 * 'readpos' is the least advanced device address we might read.
5715 * 'safepos' is the least address recorded in the metadata as having
5717 * If there is a min_offset_diff, these are adjusted either by
5718 * increasing the safepos/readpos if diff is negative, or
5719 * increasing writepos if diff is positive.
5720 * If 'readpos' is then behind 'writepos', there is no way that we can
5721 * ensure safety in the face of a crash - that must be done by userspace
5722 * making a backup of the data. So in that case there is no particular
5723 * rush to update metadata.
5724 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5725 * update the metadata to advance 'safepos' to match 'readpos' so that
5726 * we can be safe in the event of a crash.
5727 * So we insist on updating metadata if safepos is behind writepos and
5728 * readpos is beyond writepos.
5729 * In any case, update the metadata every 10 seconds.
5730 * Maybe that number should be configurable, but I'm not sure it is
5731 * worth it.... maybe it could be a multiple of safemode_delay???
5733 if (conf->min_offset_diff < 0) {
5734 safepos += -conf->min_offset_diff;
5735 readpos += -conf->min_offset_diff;
5737 writepos += conf->min_offset_diff;
5739 if ((mddev->reshape_backwards
5740 ? (safepos > writepos && readpos < writepos)
5741 : (safepos < writepos && readpos > writepos)) ||
5742 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5743 /* Cannot proceed until we've updated the superblock... */
5744 wait_event(conf->wait_for_overlap,
5745 atomic_read(&conf->reshape_stripes)==0
5746 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5747 if (atomic_read(&conf->reshape_stripes) != 0)
5749 mddev->reshape_position = conf->reshape_progress;
5750 mddev->curr_resync_completed = sector_nr;
5751 conf->reshape_checkpoint = jiffies;
5752 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5753 md_wakeup_thread(mddev->thread);
5754 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5755 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5756 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5758 spin_lock_irq(&conf->device_lock);
5759 conf->reshape_safe = mddev->reshape_position;
5760 spin_unlock_irq(&conf->device_lock);
5761 wake_up(&conf->wait_for_overlap);
5762 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5765 INIT_LIST_HEAD(&stripes);
5766 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5768 int skipped_disk = 0;
5769 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5770 set_bit(STRIPE_EXPANDING, &sh->state);
5771 atomic_inc(&conf->reshape_stripes);
5772 /* If any of this stripe is beyond the end of the old
5773 * array, then we need to zero those blocks
5775 for (j=sh->disks; j--;) {
5777 if (j == sh->pd_idx)
5779 if (conf->level == 6 &&
5782 s = raid5_compute_blocknr(sh, j, 0);
5783 if (s < raid5_size(mddev, 0, 0)) {
5787 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5788 set_bit(R5_Expanded, &sh->dev[j].flags);
5789 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5791 if (!skipped_disk) {
5792 set_bit(STRIPE_EXPAND_READY, &sh->state);
5793 set_bit(STRIPE_HANDLE, &sh->state);
5795 list_add(&sh->lru, &stripes);
5797 spin_lock_irq(&conf->device_lock);
5798 if (mddev->reshape_backwards)
5799 conf->reshape_progress -= reshape_sectors * new_data_disks;
5801 conf->reshape_progress += reshape_sectors * new_data_disks;
5802 spin_unlock_irq(&conf->device_lock);
5803 /* Ok, those stripe are ready. We can start scheduling
5804 * reads on the source stripes.
5805 * The source stripes are determined by mapping the first and last
5806 * block on the destination stripes.
5809 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5812 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5813 * new_data_disks - 1),
5815 if (last_sector >= mddev->dev_sectors)
5816 last_sector = mddev->dev_sectors - 1;
5817 while (first_sector <= last_sector) {
5818 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5819 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5820 set_bit(STRIPE_HANDLE, &sh->state);
5821 raid5_release_stripe(sh);
5822 first_sector += STRIPE_SECTORS;
5824 /* Now that the sources are clearly marked, we can release
5825 * the destination stripes
5827 while (!list_empty(&stripes)) {
5828 sh = list_entry(stripes.next, struct stripe_head, lru);
5829 list_del_init(&sh->lru);
5830 raid5_release_stripe(sh);
5832 /* If this takes us to the resync_max point where we have to pause,
5833 * then we need to write out the superblock.
5835 sector_nr += reshape_sectors;
5836 retn = reshape_sectors;
5838 if (mddev->curr_resync_completed > mddev->resync_max ||
5839 (sector_nr - mddev->curr_resync_completed) * 2
5840 >= mddev->resync_max - mddev->curr_resync_completed) {
5841 /* Cannot proceed until we've updated the superblock... */
5842 wait_event(conf->wait_for_overlap,
5843 atomic_read(&conf->reshape_stripes) == 0
5844 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5845 if (atomic_read(&conf->reshape_stripes) != 0)
5847 mddev->reshape_position = conf->reshape_progress;
5848 mddev->curr_resync_completed = sector_nr;
5849 conf->reshape_checkpoint = jiffies;
5850 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5851 md_wakeup_thread(mddev->thread);
5852 wait_event(mddev->sb_wait,
5853 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
5854 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5855 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5857 spin_lock_irq(&conf->device_lock);
5858 conf->reshape_safe = mddev->reshape_position;
5859 spin_unlock_irq(&conf->device_lock);
5860 wake_up(&conf->wait_for_overlap);
5861 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5867 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
5870 struct r5conf *conf = mddev->private;
5871 struct stripe_head *sh;
5872 sector_t max_sector = mddev->dev_sectors;
5873 sector_t sync_blocks;
5874 int still_degraded = 0;
5877 if (sector_nr >= max_sector) {
5878 /* just being told to finish up .. nothing much to do */
5880 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5885 if (mddev->curr_resync < max_sector) /* aborted */
5886 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5888 else /* completed sync */
5890 bitmap_close_sync(mddev->bitmap);
5895 /* Allow raid5_quiesce to complete */
5896 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5898 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5899 return reshape_request(mddev, sector_nr, skipped);
5901 /* No need to check resync_max as we never do more than one
5902 * stripe, and as resync_max will always be on a chunk boundary,
5903 * if the check in md_do_sync didn't fire, there is no chance
5904 * of overstepping resync_max here
5907 /* if there is too many failed drives and we are trying
5908 * to resync, then assert that we are finished, because there is
5909 * nothing we can do.
5911 if (mddev->degraded >= conf->max_degraded &&
5912 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5913 sector_t rv = mddev->dev_sectors - sector_nr;
5917 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5919 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5920 sync_blocks >= STRIPE_SECTORS) {
5921 /* we can skip this block, and probably more */
5922 sync_blocks /= STRIPE_SECTORS;
5924 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5927 bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5929 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5931 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5932 /* make sure we don't swamp the stripe cache if someone else
5933 * is trying to get access
5935 schedule_timeout_uninterruptible(1);
5937 /* Need to check if array will still be degraded after recovery/resync
5938 * Note in case of > 1 drive failures it's possible we're rebuilding
5939 * one drive while leaving another faulty drive in array.
5942 for (i = 0; i < conf->raid_disks; i++) {
5943 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5945 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5950 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5952 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5953 set_bit(STRIPE_HANDLE, &sh->state);
5955 raid5_release_stripe(sh);
5957 return STRIPE_SECTORS;
5960 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5962 /* We may not be able to submit a whole bio at once as there
5963 * may not be enough stripe_heads available.
5964 * We cannot pre-allocate enough stripe_heads as we may need
5965 * more than exist in the cache (if we allow ever large chunks).
5966 * So we do one stripe head at a time and record in
5967 * ->bi_hw_segments how many have been done.
5969 * We *know* that this entire raid_bio is in one chunk, so
5970 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5972 struct stripe_head *sh;
5974 sector_t sector, logical_sector, last_sector;
5979 logical_sector = raid_bio->bi_iter.bi_sector &
5980 ~((sector_t)STRIPE_SECTORS-1);
5981 sector = raid5_compute_sector(conf, logical_sector,
5983 last_sector = bio_end_sector(raid_bio);
5985 for (; logical_sector < last_sector;
5986 logical_sector += STRIPE_SECTORS,
5987 sector += STRIPE_SECTORS,
5990 if (scnt < raid5_bi_processed_stripes(raid_bio))
5991 /* already done this stripe */
5994 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5997 /* failed to get a stripe - must wait */
5998 raid5_set_bi_processed_stripes(raid_bio, scnt);
5999 conf->retry_read_aligned = raid_bio;
6003 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6004 raid5_release_stripe(sh);
6005 raid5_set_bi_processed_stripes(raid_bio, scnt);
6006 conf->retry_read_aligned = raid_bio;
6010 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6012 raid5_release_stripe(sh);
6015 remaining = raid5_dec_bi_active_stripes(raid_bio);
6016 if (remaining == 0) {
6017 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
6019 bio_endio(raid_bio);
6021 if (atomic_dec_and_test(&conf->active_aligned_reads))
6022 wake_up(&conf->wait_for_quiescent);
6026 static int handle_active_stripes(struct r5conf *conf, int group,
6027 struct r5worker *worker,
6028 struct list_head *temp_inactive_list)
6030 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6031 int i, batch_size = 0, hash;
6032 bool release_inactive = false;
6034 while (batch_size < MAX_STRIPE_BATCH &&
6035 (sh = __get_priority_stripe(conf, group)) != NULL)
6036 batch[batch_size++] = sh;
6038 if (batch_size == 0) {
6039 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6040 if (!list_empty(temp_inactive_list + i))
6042 if (i == NR_STRIPE_HASH_LOCKS) {
6043 spin_unlock_irq(&conf->device_lock);
6044 r5l_flush_stripe_to_raid(conf->log);
6045 spin_lock_irq(&conf->device_lock);
6048 release_inactive = true;
6050 spin_unlock_irq(&conf->device_lock);
6052 release_inactive_stripe_list(conf, temp_inactive_list,
6053 NR_STRIPE_HASH_LOCKS);
6055 r5l_flush_stripe_to_raid(conf->log);
6056 if (release_inactive) {
6057 spin_lock_irq(&conf->device_lock);
6061 for (i = 0; i < batch_size; i++)
6062 handle_stripe(batch[i]);
6063 r5l_write_stripe_run(conf->log);
6067 spin_lock_irq(&conf->device_lock);
6068 for (i = 0; i < batch_size; i++) {
6069 hash = batch[i]->hash_lock_index;
6070 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6075 static void raid5_do_work(struct work_struct *work)
6077 struct r5worker *worker = container_of(work, struct r5worker, work);
6078 struct r5worker_group *group = worker->group;
6079 struct r5conf *conf = group->conf;
6080 int group_id = group - conf->worker_groups;
6082 struct blk_plug plug;
6084 pr_debug("+++ raid5worker active\n");
6086 blk_start_plug(&plug);
6088 spin_lock_irq(&conf->device_lock);
6090 int batch_size, released;
6092 released = release_stripe_list(conf, worker->temp_inactive_list);
6094 batch_size = handle_active_stripes(conf, group_id, worker,
6095 worker->temp_inactive_list);
6096 worker->working = false;
6097 if (!batch_size && !released)
6099 handled += batch_size;
6101 pr_debug("%d stripes handled\n", handled);
6103 spin_unlock_irq(&conf->device_lock);
6104 blk_finish_plug(&plug);
6106 pr_debug("--- raid5worker inactive\n");
6110 * This is our raid5 kernel thread.
6112 * We scan the hash table for stripes which can be handled now.
6113 * During the scan, completed stripes are saved for us by the interrupt
6114 * handler, so that they will not have to wait for our next wakeup.
6116 static void raid5d(struct md_thread *thread)
6118 struct mddev *mddev = thread->mddev;
6119 struct r5conf *conf = mddev->private;
6121 struct blk_plug plug;
6123 pr_debug("+++ raid5d active\n");
6125 md_check_recovery(mddev);
6127 if (!bio_list_empty(&conf->return_bi) &&
6128 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
6129 struct bio_list tmp = BIO_EMPTY_LIST;
6130 spin_lock_irq(&conf->device_lock);
6131 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
6132 bio_list_merge(&tmp, &conf->return_bi);
6133 bio_list_init(&conf->return_bi);
6135 spin_unlock_irq(&conf->device_lock);
6139 blk_start_plug(&plug);
6141 spin_lock_irq(&conf->device_lock);
6144 int batch_size, released;
6146 released = release_stripe_list(conf, conf->temp_inactive_list);
6148 clear_bit(R5_DID_ALLOC, &conf->cache_state);
6151 !list_empty(&conf->bitmap_list)) {
6152 /* Now is a good time to flush some bitmap updates */
6154 spin_unlock_irq(&conf->device_lock);
6155 bitmap_unplug(mddev->bitmap);
6156 spin_lock_irq(&conf->device_lock);
6157 conf->seq_write = conf->seq_flush;
6158 activate_bit_delay(conf, conf->temp_inactive_list);
6160 raid5_activate_delayed(conf);
6162 while ((bio = remove_bio_from_retry(conf))) {
6164 spin_unlock_irq(&conf->device_lock);
6165 ok = retry_aligned_read(conf, bio);
6166 spin_lock_irq(&conf->device_lock);
6172 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6173 conf->temp_inactive_list);
6174 if (!batch_size && !released)
6176 handled += batch_size;
6178 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6179 spin_unlock_irq(&conf->device_lock);
6180 md_check_recovery(mddev);
6181 spin_lock_irq(&conf->device_lock);
6184 pr_debug("%d stripes handled\n", handled);
6186 spin_unlock_irq(&conf->device_lock);
6187 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6188 mutex_trylock(&conf->cache_size_mutex)) {
6189 grow_one_stripe(conf, __GFP_NOWARN);
6190 /* Set flag even if allocation failed. This helps
6191 * slow down allocation requests when mem is short
6193 set_bit(R5_DID_ALLOC, &conf->cache_state);
6194 mutex_unlock(&conf->cache_size_mutex);
6197 flush_deferred_bios(conf);
6199 r5l_flush_stripe_to_raid(conf->log);
6201 async_tx_issue_pending_all();
6202 blk_finish_plug(&plug);
6204 pr_debug("--- raid5d inactive\n");
6208 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6210 struct r5conf *conf;
6212 spin_lock(&mddev->lock);
6213 conf = mddev->private;
6215 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6216 spin_unlock(&mddev->lock);
6221 raid5_set_cache_size(struct mddev *mddev, int size)
6223 struct r5conf *conf = mddev->private;
6226 if (size <= 16 || size > 32768)
6229 conf->min_nr_stripes = size;
6230 mutex_lock(&conf->cache_size_mutex);
6231 while (size < conf->max_nr_stripes &&
6232 drop_one_stripe(conf))
6234 mutex_unlock(&conf->cache_size_mutex);
6237 err = md_allow_write(mddev);
6241 mutex_lock(&conf->cache_size_mutex);
6242 while (size > conf->max_nr_stripes)
6243 if (!grow_one_stripe(conf, GFP_KERNEL))
6245 mutex_unlock(&conf->cache_size_mutex);
6249 EXPORT_SYMBOL(raid5_set_cache_size);
6252 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6254 struct r5conf *conf;
6258 if (len >= PAGE_SIZE)
6260 if (kstrtoul(page, 10, &new))
6262 err = mddev_lock(mddev);
6265 conf = mddev->private;
6269 err = raid5_set_cache_size(mddev, new);
6270 mddev_unlock(mddev);
6275 static struct md_sysfs_entry
6276 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6277 raid5_show_stripe_cache_size,
6278 raid5_store_stripe_cache_size);
6281 raid5_show_rmw_level(struct mddev *mddev, char *page)
6283 struct r5conf *conf = mddev->private;
6285 return sprintf(page, "%d\n", conf->rmw_level);
6291 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6293 struct r5conf *conf = mddev->private;
6299 if (len >= PAGE_SIZE)
6302 if (kstrtoul(page, 10, &new))
6305 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6308 if (new != PARITY_DISABLE_RMW &&
6309 new != PARITY_ENABLE_RMW &&
6310 new != PARITY_PREFER_RMW)
6313 conf->rmw_level = new;
6317 static struct md_sysfs_entry
6318 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6319 raid5_show_rmw_level,
6320 raid5_store_rmw_level);
6324 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6326 struct r5conf *conf;
6328 spin_lock(&mddev->lock);
6329 conf = mddev->private;
6331 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6332 spin_unlock(&mddev->lock);
6337 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6339 struct r5conf *conf;
6343 if (len >= PAGE_SIZE)
6345 if (kstrtoul(page, 10, &new))
6348 err = mddev_lock(mddev);
6351 conf = mddev->private;
6354 else if (new > conf->min_nr_stripes)
6357 conf->bypass_threshold = new;
6358 mddev_unlock(mddev);
6362 static struct md_sysfs_entry
6363 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6365 raid5_show_preread_threshold,
6366 raid5_store_preread_threshold);
6369 raid5_show_skip_copy(struct mddev *mddev, char *page)
6371 struct r5conf *conf;
6373 spin_lock(&mddev->lock);
6374 conf = mddev->private;
6376 ret = sprintf(page, "%d\n", conf->skip_copy);
6377 spin_unlock(&mddev->lock);
6382 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6384 struct r5conf *conf;
6388 if (len >= PAGE_SIZE)
6390 if (kstrtoul(page, 10, &new))
6394 err = mddev_lock(mddev);
6397 conf = mddev->private;
6400 else if (new != conf->skip_copy) {
6401 mddev_suspend(mddev);
6402 conf->skip_copy = new;
6404 mddev->queue->backing_dev_info->capabilities |=
6405 BDI_CAP_STABLE_WRITES;
6407 mddev->queue->backing_dev_info->capabilities &=
6408 ~BDI_CAP_STABLE_WRITES;
6409 mddev_resume(mddev);
6411 mddev_unlock(mddev);
6415 static struct md_sysfs_entry
6416 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6417 raid5_show_skip_copy,
6418 raid5_store_skip_copy);
6421 stripe_cache_active_show(struct mddev *mddev, char *page)
6423 struct r5conf *conf = mddev->private;
6425 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6430 static struct md_sysfs_entry
6431 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6434 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6436 struct r5conf *conf;
6438 spin_lock(&mddev->lock);
6439 conf = mddev->private;
6441 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6442 spin_unlock(&mddev->lock);
6446 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6448 int *worker_cnt_per_group,
6449 struct r5worker_group **worker_groups);
6451 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6453 struct r5conf *conf;
6456 struct r5worker_group *new_groups, *old_groups;
6457 int group_cnt, worker_cnt_per_group;
6459 if (len >= PAGE_SIZE)
6461 if (kstrtoul(page, 10, &new))
6464 err = mddev_lock(mddev);
6467 conf = mddev->private;
6470 else if (new != conf->worker_cnt_per_group) {
6471 mddev_suspend(mddev);
6473 old_groups = conf->worker_groups;
6475 flush_workqueue(raid5_wq);
6477 err = alloc_thread_groups(conf, new,
6478 &group_cnt, &worker_cnt_per_group,
6481 spin_lock_irq(&conf->device_lock);
6482 conf->group_cnt = group_cnt;
6483 conf->worker_cnt_per_group = worker_cnt_per_group;
6484 conf->worker_groups = new_groups;
6485 spin_unlock_irq(&conf->device_lock);
6488 kfree(old_groups[0].workers);
6491 mddev_resume(mddev);
6493 mddev_unlock(mddev);
6498 static struct md_sysfs_entry
6499 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6500 raid5_show_group_thread_cnt,
6501 raid5_store_group_thread_cnt);
6503 static struct attribute *raid5_attrs[] = {
6504 &raid5_stripecache_size.attr,
6505 &raid5_stripecache_active.attr,
6506 &raid5_preread_bypass_threshold.attr,
6507 &raid5_group_thread_cnt.attr,
6508 &raid5_skip_copy.attr,
6509 &raid5_rmw_level.attr,
6510 &r5c_journal_mode.attr,
6513 static struct attribute_group raid5_attrs_group = {
6515 .attrs = raid5_attrs,
6518 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6520 int *worker_cnt_per_group,
6521 struct r5worker_group **worker_groups)
6525 struct r5worker *workers;
6527 *worker_cnt_per_group = cnt;
6530 *worker_groups = NULL;
6533 *group_cnt = num_possible_nodes();
6534 size = sizeof(struct r5worker) * cnt;
6535 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6536 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6537 *group_cnt, GFP_NOIO);
6538 if (!*worker_groups || !workers) {
6540 kfree(*worker_groups);
6544 for (i = 0; i < *group_cnt; i++) {
6545 struct r5worker_group *group;
6547 group = &(*worker_groups)[i];
6548 INIT_LIST_HEAD(&group->handle_list);
6550 group->workers = workers + i * cnt;
6552 for (j = 0; j < cnt; j++) {
6553 struct r5worker *worker = group->workers + j;
6554 worker->group = group;
6555 INIT_WORK(&worker->work, raid5_do_work);
6557 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6558 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6565 static void free_thread_groups(struct r5conf *conf)
6567 if (conf->worker_groups)
6568 kfree(conf->worker_groups[0].workers);
6569 kfree(conf->worker_groups);
6570 conf->worker_groups = NULL;
6574 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6576 struct r5conf *conf = mddev->private;
6579 sectors = mddev->dev_sectors;
6581 /* size is defined by the smallest of previous and new size */
6582 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6584 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6585 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6586 return sectors * (raid_disks - conf->max_degraded);
6589 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6591 safe_put_page(percpu->spare_page);
6592 if (percpu->scribble)
6593 flex_array_free(percpu->scribble);
6594 percpu->spare_page = NULL;
6595 percpu->scribble = NULL;
6598 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6600 if (conf->level == 6 && !percpu->spare_page)
6601 percpu->spare_page = alloc_page(GFP_KERNEL);
6602 if (!percpu->scribble)
6603 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6604 conf->previous_raid_disks),
6605 max(conf->chunk_sectors,
6606 conf->prev_chunk_sectors)
6610 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6611 free_scratch_buffer(conf, percpu);
6618 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6620 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6622 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6626 static void raid5_free_percpu(struct r5conf *conf)
6631 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6632 free_percpu(conf->percpu);
6635 static void free_conf(struct r5conf *conf)
6640 r5l_exit_log(conf->log);
6641 if (conf->shrinker.nr_deferred)
6642 unregister_shrinker(&conf->shrinker);
6644 free_thread_groups(conf);
6645 shrink_stripes(conf);
6646 raid5_free_percpu(conf);
6647 for (i = 0; i < conf->pool_size; i++)
6648 if (conf->disks[i].extra_page)
6649 put_page(conf->disks[i].extra_page);
6651 kfree(conf->stripe_hashtbl);
6655 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6657 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6658 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6660 if (alloc_scratch_buffer(conf, percpu)) {
6661 pr_warn("%s: failed memory allocation for cpu%u\n",
6668 static int raid5_alloc_percpu(struct r5conf *conf)
6672 conf->percpu = alloc_percpu(struct raid5_percpu);
6676 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6678 conf->scribble_disks = max(conf->raid_disks,
6679 conf->previous_raid_disks);
6680 conf->scribble_sectors = max(conf->chunk_sectors,
6681 conf->prev_chunk_sectors);
6686 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6687 struct shrink_control *sc)
6689 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6690 unsigned long ret = SHRINK_STOP;
6692 if (mutex_trylock(&conf->cache_size_mutex)) {
6694 while (ret < sc->nr_to_scan &&
6695 conf->max_nr_stripes > conf->min_nr_stripes) {
6696 if (drop_one_stripe(conf) == 0) {
6702 mutex_unlock(&conf->cache_size_mutex);
6707 static unsigned long raid5_cache_count(struct shrinker *shrink,
6708 struct shrink_control *sc)
6710 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6712 if (conf->max_nr_stripes < conf->min_nr_stripes)
6713 /* unlikely, but not impossible */
6715 return conf->max_nr_stripes - conf->min_nr_stripes;
6718 static struct r5conf *setup_conf(struct mddev *mddev)
6720 struct r5conf *conf;
6721 int raid_disk, memory, max_disks;
6722 struct md_rdev *rdev;
6723 struct disk_info *disk;
6726 int group_cnt, worker_cnt_per_group;
6727 struct r5worker_group *new_group;
6729 if (mddev->new_level != 5
6730 && mddev->new_level != 4
6731 && mddev->new_level != 6) {
6732 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6733 mdname(mddev), mddev->new_level);
6734 return ERR_PTR(-EIO);
6736 if ((mddev->new_level == 5
6737 && !algorithm_valid_raid5(mddev->new_layout)) ||
6738 (mddev->new_level == 6
6739 && !algorithm_valid_raid6(mddev->new_layout))) {
6740 pr_warn("md/raid:%s: layout %d not supported\n",
6741 mdname(mddev), mddev->new_layout);
6742 return ERR_PTR(-EIO);
6744 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6745 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6746 mdname(mddev), mddev->raid_disks);
6747 return ERR_PTR(-EINVAL);
6750 if (!mddev->new_chunk_sectors ||
6751 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6752 !is_power_of_2(mddev->new_chunk_sectors)) {
6753 pr_warn("md/raid:%s: invalid chunk size %d\n",
6754 mdname(mddev), mddev->new_chunk_sectors << 9);
6755 return ERR_PTR(-EINVAL);
6758 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6761 /* Don't enable multi-threading by default*/
6762 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6764 conf->group_cnt = group_cnt;
6765 conf->worker_cnt_per_group = worker_cnt_per_group;
6766 conf->worker_groups = new_group;
6769 spin_lock_init(&conf->device_lock);
6770 seqcount_init(&conf->gen_lock);
6771 mutex_init(&conf->cache_size_mutex);
6772 init_waitqueue_head(&conf->wait_for_quiescent);
6773 init_waitqueue_head(&conf->wait_for_stripe);
6774 init_waitqueue_head(&conf->wait_for_overlap);
6775 INIT_LIST_HEAD(&conf->handle_list);
6776 INIT_LIST_HEAD(&conf->hold_list);
6777 INIT_LIST_HEAD(&conf->delayed_list);
6778 INIT_LIST_HEAD(&conf->bitmap_list);
6779 bio_list_init(&conf->return_bi);
6780 init_llist_head(&conf->released_stripes);
6781 atomic_set(&conf->active_stripes, 0);
6782 atomic_set(&conf->preread_active_stripes, 0);
6783 atomic_set(&conf->active_aligned_reads, 0);
6784 bio_list_init(&conf->pending_bios);
6785 spin_lock_init(&conf->pending_bios_lock);
6786 conf->batch_bio_dispatch = true;
6787 rdev_for_each(rdev, mddev) {
6788 if (test_bit(Journal, &rdev->flags))
6790 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6791 conf->batch_bio_dispatch = false;
6796 conf->bypass_threshold = BYPASS_THRESHOLD;
6797 conf->recovery_disabled = mddev->recovery_disabled - 1;
6799 conf->raid_disks = mddev->raid_disks;
6800 if (mddev->reshape_position == MaxSector)
6801 conf->previous_raid_disks = mddev->raid_disks;
6803 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6804 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6806 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6812 for (i = 0; i < max_disks; i++) {
6813 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6814 if (!conf->disks[i].extra_page)
6818 conf->mddev = mddev;
6820 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6823 /* We init hash_locks[0] separately to that it can be used
6824 * as the reference lock in the spin_lock_nest_lock() call
6825 * in lock_all_device_hash_locks_irq in order to convince
6826 * lockdep that we know what we are doing.
6828 spin_lock_init(conf->hash_locks);
6829 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6830 spin_lock_init(conf->hash_locks + i);
6832 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6833 INIT_LIST_HEAD(conf->inactive_list + i);
6835 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6836 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6838 atomic_set(&conf->r5c_cached_full_stripes, 0);
6839 INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
6840 atomic_set(&conf->r5c_cached_partial_stripes, 0);
6841 INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
6842 atomic_set(&conf->r5c_flushing_full_stripes, 0);
6843 atomic_set(&conf->r5c_flushing_partial_stripes, 0);
6845 conf->level = mddev->new_level;
6846 conf->chunk_sectors = mddev->new_chunk_sectors;
6847 if (raid5_alloc_percpu(conf) != 0)
6850 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6852 rdev_for_each(rdev, mddev) {
6853 raid_disk = rdev->raid_disk;
6854 if (raid_disk >= max_disks
6855 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6857 disk = conf->disks + raid_disk;
6859 if (test_bit(Replacement, &rdev->flags)) {
6860 if (disk->replacement)
6862 disk->replacement = rdev;
6869 if (test_bit(In_sync, &rdev->flags)) {
6870 char b[BDEVNAME_SIZE];
6871 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
6872 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6873 } else if (rdev->saved_raid_disk != raid_disk)
6874 /* Cannot rely on bitmap to complete recovery */
6878 conf->level = mddev->new_level;
6879 if (conf->level == 6) {
6880 conf->max_degraded = 2;
6881 if (raid6_call.xor_syndrome)
6882 conf->rmw_level = PARITY_ENABLE_RMW;
6884 conf->rmw_level = PARITY_DISABLE_RMW;
6886 conf->max_degraded = 1;
6887 conf->rmw_level = PARITY_ENABLE_RMW;
6889 conf->algorithm = mddev->new_layout;
6890 conf->reshape_progress = mddev->reshape_position;
6891 if (conf->reshape_progress != MaxSector) {
6892 conf->prev_chunk_sectors = mddev->chunk_sectors;
6893 conf->prev_algo = mddev->layout;
6895 conf->prev_chunk_sectors = conf->chunk_sectors;
6896 conf->prev_algo = conf->algorithm;
6899 conf->min_nr_stripes = NR_STRIPES;
6900 if (mddev->reshape_position != MaxSector) {
6901 int stripes = max_t(int,
6902 ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
6903 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
6904 conf->min_nr_stripes = max(NR_STRIPES, stripes);
6905 if (conf->min_nr_stripes != NR_STRIPES)
6906 pr_info("md/raid:%s: force stripe size %d for reshape\n",
6907 mdname(mddev), conf->min_nr_stripes);
6909 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6910 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6911 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6912 if (grow_stripes(conf, conf->min_nr_stripes)) {
6913 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
6914 mdname(mddev), memory);
6917 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
6919 * Losing a stripe head costs more than the time to refill it,
6920 * it reduces the queue depth and so can hurt throughput.
6921 * So set it rather large, scaled by number of devices.
6923 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6924 conf->shrinker.scan_objects = raid5_cache_scan;
6925 conf->shrinker.count_objects = raid5_cache_count;
6926 conf->shrinker.batch = 128;
6927 conf->shrinker.flags = 0;
6928 if (register_shrinker(&conf->shrinker)) {
6929 pr_warn("md/raid:%s: couldn't register shrinker.\n",
6934 sprintf(pers_name, "raid%d", mddev->new_level);
6935 conf->thread = md_register_thread(raid5d, mddev, pers_name);
6936 if (!conf->thread) {
6937 pr_warn("md/raid:%s: couldn't allocate thread.\n",
6947 return ERR_PTR(-EIO);
6949 return ERR_PTR(-ENOMEM);
6952 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6955 case ALGORITHM_PARITY_0:
6956 if (raid_disk < max_degraded)
6959 case ALGORITHM_PARITY_N:
6960 if (raid_disk >= raid_disks - max_degraded)
6963 case ALGORITHM_PARITY_0_6:
6964 if (raid_disk == 0 ||
6965 raid_disk == raid_disks - 1)
6968 case ALGORITHM_LEFT_ASYMMETRIC_6:
6969 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6970 case ALGORITHM_LEFT_SYMMETRIC_6:
6971 case ALGORITHM_RIGHT_SYMMETRIC_6:
6972 if (raid_disk == raid_disks - 1)
6978 static int raid5_run(struct mddev *mddev)
6980 struct r5conf *conf;
6981 int working_disks = 0;
6982 int dirty_parity_disks = 0;
6983 struct md_rdev *rdev;
6984 struct md_rdev *journal_dev = NULL;
6985 sector_t reshape_offset = 0;
6987 long long min_offset_diff = 0;
6990 if (mddev->recovery_cp != MaxSector)
6991 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
6994 rdev_for_each(rdev, mddev) {
6997 if (test_bit(Journal, &rdev->flags)) {
7001 if (rdev->raid_disk < 0)
7003 diff = (rdev->new_data_offset - rdev->data_offset);
7005 min_offset_diff = diff;
7007 } else if (mddev->reshape_backwards &&
7008 diff < min_offset_diff)
7009 min_offset_diff = diff;
7010 else if (!mddev->reshape_backwards &&
7011 diff > min_offset_diff)
7012 min_offset_diff = diff;
7015 if (mddev->reshape_position != MaxSector) {
7016 /* Check that we can continue the reshape.
7017 * Difficulties arise if the stripe we would write to
7018 * next is at or after the stripe we would read from next.
7019 * For a reshape that changes the number of devices, this
7020 * is only possible for a very short time, and mdadm makes
7021 * sure that time appears to have past before assembling
7022 * the array. So we fail if that time hasn't passed.
7023 * For a reshape that keeps the number of devices the same
7024 * mdadm must be monitoring the reshape can keeping the
7025 * critical areas read-only and backed up. It will start
7026 * the array in read-only mode, so we check for that.
7028 sector_t here_new, here_old;
7030 int max_degraded = (mddev->level == 6 ? 2 : 1);
7035 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7040 if (mddev->new_level != mddev->level) {
7041 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7045 old_disks = mddev->raid_disks - mddev->delta_disks;
7046 /* reshape_position must be on a new-stripe boundary, and one
7047 * further up in new geometry must map after here in old
7049 * If the chunk sizes are different, then as we perform reshape
7050 * in units of the largest of the two, reshape_position needs
7051 * be a multiple of the largest chunk size times new data disks.
7053 here_new = mddev->reshape_position;
7054 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7055 new_data_disks = mddev->raid_disks - max_degraded;
7056 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7057 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7061 reshape_offset = here_new * chunk_sectors;
7062 /* here_new is the stripe we will write to */
7063 here_old = mddev->reshape_position;
7064 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7065 /* here_old is the first stripe that we might need to read
7067 if (mddev->delta_disks == 0) {
7068 /* We cannot be sure it is safe to start an in-place
7069 * reshape. It is only safe if user-space is monitoring
7070 * and taking constant backups.
7071 * mdadm always starts a situation like this in
7072 * readonly mode so it can take control before
7073 * allowing any writes. So just check for that.
7075 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7076 abs(min_offset_diff) >= mddev->new_chunk_sectors)
7077 /* not really in-place - so OK */;
7078 else if (mddev->ro == 0) {
7079 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7083 } else if (mddev->reshape_backwards
7084 ? (here_new * chunk_sectors + min_offset_diff <=
7085 here_old * chunk_sectors)
7086 : (here_new * chunk_sectors >=
7087 here_old * chunk_sectors + (-min_offset_diff))) {
7088 /* Reading from the same stripe as writing to - bad */
7089 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7093 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7094 /* OK, we should be able to continue; */
7096 BUG_ON(mddev->level != mddev->new_level);
7097 BUG_ON(mddev->layout != mddev->new_layout);
7098 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7099 BUG_ON(mddev->delta_disks != 0);
7102 if (mddev->private == NULL)
7103 conf = setup_conf(mddev);
7105 conf = mddev->private;
7108 return PTR_ERR(conf);
7110 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7112 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7115 set_disk_ro(mddev->gendisk, 1);
7116 } else if (mddev->recovery_cp == MaxSector)
7117 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7120 conf->min_offset_diff = min_offset_diff;
7121 mddev->thread = conf->thread;
7122 conf->thread = NULL;
7123 mddev->private = conf;
7125 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7127 rdev = conf->disks[i].rdev;
7128 if (!rdev && conf->disks[i].replacement) {
7129 /* The replacement is all we have yet */
7130 rdev = conf->disks[i].replacement;
7131 conf->disks[i].replacement = NULL;
7132 clear_bit(Replacement, &rdev->flags);
7133 conf->disks[i].rdev = rdev;
7137 if (conf->disks[i].replacement &&
7138 conf->reshape_progress != MaxSector) {
7139 /* replacements and reshape simply do not mix. */
7140 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7143 if (test_bit(In_sync, &rdev->flags)) {
7147 /* This disc is not fully in-sync. However if it
7148 * just stored parity (beyond the recovery_offset),
7149 * when we don't need to be concerned about the
7150 * array being dirty.
7151 * When reshape goes 'backwards', we never have
7152 * partially completed devices, so we only need
7153 * to worry about reshape going forwards.
7155 /* Hack because v0.91 doesn't store recovery_offset properly. */
7156 if (mddev->major_version == 0 &&
7157 mddev->minor_version > 90)
7158 rdev->recovery_offset = reshape_offset;
7160 if (rdev->recovery_offset < reshape_offset) {
7161 /* We need to check old and new layout */
7162 if (!only_parity(rdev->raid_disk,
7165 conf->max_degraded))
7168 if (!only_parity(rdev->raid_disk,
7170 conf->previous_raid_disks,
7171 conf->max_degraded))
7173 dirty_parity_disks++;
7177 * 0 for a fully functional array, 1 or 2 for a degraded array.
7179 mddev->degraded = raid5_calc_degraded(conf);
7181 if (has_failed(conf)) {
7182 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7183 mdname(mddev), mddev->degraded, conf->raid_disks);
7187 /* device size must be a multiple of chunk size */
7188 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7189 mddev->resync_max_sectors = mddev->dev_sectors;
7191 if (mddev->degraded > dirty_parity_disks &&
7192 mddev->recovery_cp != MaxSector) {
7193 if (mddev->ok_start_degraded)
7194 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7197 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7203 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7204 mdname(mddev), conf->level,
7205 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7208 print_raid5_conf(conf);
7210 if (conf->reshape_progress != MaxSector) {
7211 conf->reshape_safe = conf->reshape_progress;
7212 atomic_set(&conf->reshape_stripes, 0);
7213 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7214 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7215 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7216 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7217 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7221 /* Ok, everything is just fine now */
7222 if (mddev->to_remove == &raid5_attrs_group)
7223 mddev->to_remove = NULL;
7224 else if (mddev->kobj.sd &&
7225 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7226 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7228 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7232 bool discard_supported = true;
7233 /* read-ahead size must cover two whole stripes, which
7234 * is 2 * (datadisks) * chunksize where 'n' is the
7235 * number of raid devices
7237 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7238 int stripe = data_disks *
7239 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7240 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7241 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7243 chunk_size = mddev->chunk_sectors << 9;
7244 blk_queue_io_min(mddev->queue, chunk_size);
7245 blk_queue_io_opt(mddev->queue, chunk_size *
7246 (conf->raid_disks - conf->max_degraded));
7247 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7249 * We can only discard a whole stripe. It doesn't make sense to
7250 * discard data disk but write parity disk
7252 stripe = stripe * PAGE_SIZE;
7253 /* Round up to power of 2, as discard handling
7254 * currently assumes that */
7255 while ((stripe-1) & stripe)
7256 stripe = (stripe | (stripe-1)) + 1;
7257 mddev->queue->limits.discard_alignment = stripe;
7258 mddev->queue->limits.discard_granularity = stripe;
7261 * We use 16-bit counter of active stripes in bi_phys_segments
7262 * (minus one for over-loaded initialization)
7264 blk_queue_max_hw_sectors(mddev->queue, 0xfffe * STRIPE_SECTORS);
7265 blk_queue_max_discard_sectors(mddev->queue,
7266 0xfffe * STRIPE_SECTORS);
7269 * unaligned part of discard request will be ignored, so can't
7270 * guarantee discard_zeroes_data
7272 mddev->queue->limits.discard_zeroes_data = 0;
7274 blk_queue_max_write_same_sectors(mddev->queue, 0);
7276 rdev_for_each(rdev, mddev) {
7277 disk_stack_limits(mddev->gendisk, rdev->bdev,
7278 rdev->data_offset << 9);
7279 disk_stack_limits(mddev->gendisk, rdev->bdev,
7280 rdev->new_data_offset << 9);
7282 * discard_zeroes_data is required, otherwise data
7283 * could be lost. Consider a scenario: discard a stripe
7284 * (the stripe could be inconsistent if
7285 * discard_zeroes_data is 0); write one disk of the
7286 * stripe (the stripe could be inconsistent again
7287 * depending on which disks are used to calculate
7288 * parity); the disk is broken; The stripe data of this
7291 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7292 !bdev_get_queue(rdev->bdev)->
7293 limits.discard_zeroes_data)
7294 discard_supported = false;
7295 /* Unfortunately, discard_zeroes_data is not currently
7296 * a guarantee - just a hint. So we only allow DISCARD
7297 * if the sysadmin has confirmed that only safe devices
7298 * are in use by setting a module parameter.
7300 if (!devices_handle_discard_safely) {
7301 if (discard_supported) {
7302 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7303 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7305 discard_supported = false;
7309 if (discard_supported &&
7310 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7311 mddev->queue->limits.discard_granularity >= stripe)
7312 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7315 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7318 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7322 char b[BDEVNAME_SIZE];
7324 pr_debug("md/raid:%s: using device %s as journal\n",
7325 mdname(mddev), bdevname(journal_dev->bdev, b));
7326 if (r5l_init_log(conf, journal_dev))
7332 md_unregister_thread(&mddev->thread);
7333 print_raid5_conf(conf);
7335 mddev->private = NULL;
7336 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7340 static void raid5_free(struct mddev *mddev, void *priv)
7342 struct r5conf *conf = priv;
7345 mddev->to_remove = &raid5_attrs_group;
7348 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7350 struct r5conf *conf = mddev->private;
7353 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7354 conf->chunk_sectors / 2, mddev->layout);
7355 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7357 for (i = 0; i < conf->raid_disks; i++) {
7358 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7359 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7362 seq_printf (seq, "]");
7365 static void print_raid5_conf (struct r5conf *conf)
7368 struct disk_info *tmp;
7370 pr_debug("RAID conf printout:\n");
7372 pr_debug("(conf==NULL)\n");
7375 pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7377 conf->raid_disks - conf->mddev->degraded);
7379 for (i = 0; i < conf->raid_disks; i++) {
7380 char b[BDEVNAME_SIZE];
7381 tmp = conf->disks + i;
7383 pr_debug(" disk %d, o:%d, dev:%s\n",
7384 i, !test_bit(Faulty, &tmp->rdev->flags),
7385 bdevname(tmp->rdev->bdev, b));
7389 static int raid5_spare_active(struct mddev *mddev)
7392 struct r5conf *conf = mddev->private;
7393 struct disk_info *tmp;
7395 unsigned long flags;
7397 for (i = 0; i < conf->raid_disks; i++) {
7398 tmp = conf->disks + i;
7399 if (tmp->replacement
7400 && tmp->replacement->recovery_offset == MaxSector
7401 && !test_bit(Faulty, &tmp->replacement->flags)
7402 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7403 /* Replacement has just become active. */
7405 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7408 /* Replaced device not technically faulty,
7409 * but we need to be sure it gets removed
7410 * and never re-added.
7412 set_bit(Faulty, &tmp->rdev->flags);
7413 sysfs_notify_dirent_safe(
7414 tmp->rdev->sysfs_state);
7416 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7417 } else if (tmp->rdev
7418 && tmp->rdev->recovery_offset == MaxSector
7419 && !test_bit(Faulty, &tmp->rdev->flags)
7420 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7422 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7425 spin_lock_irqsave(&conf->device_lock, flags);
7426 mddev->degraded = raid5_calc_degraded(conf);
7427 spin_unlock_irqrestore(&conf->device_lock, flags);
7428 print_raid5_conf(conf);
7432 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7434 struct r5conf *conf = mddev->private;
7436 int number = rdev->raid_disk;
7437 struct md_rdev **rdevp;
7438 struct disk_info *p = conf->disks + number;
7440 print_raid5_conf(conf);
7441 if (test_bit(Journal, &rdev->flags) && conf->log) {
7442 struct r5l_log *log;
7444 * we can't wait pending write here, as this is called in
7445 * raid5d, wait will deadlock.
7447 if (atomic_read(&mddev->writes_pending))
7455 if (rdev == p->rdev)
7457 else if (rdev == p->replacement)
7458 rdevp = &p->replacement;
7462 if (number >= conf->raid_disks &&
7463 conf->reshape_progress == MaxSector)
7464 clear_bit(In_sync, &rdev->flags);
7466 if (test_bit(In_sync, &rdev->flags) ||
7467 atomic_read(&rdev->nr_pending)) {
7471 /* Only remove non-faulty devices if recovery
7474 if (!test_bit(Faulty, &rdev->flags) &&
7475 mddev->recovery_disabled != conf->recovery_disabled &&
7476 !has_failed(conf) &&
7477 (!p->replacement || p->replacement == rdev) &&
7478 number < conf->raid_disks) {
7483 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7485 if (atomic_read(&rdev->nr_pending)) {
7486 /* lost the race, try later */
7491 if (p->replacement) {
7492 /* We must have just cleared 'rdev' */
7493 p->rdev = p->replacement;
7494 clear_bit(Replacement, &p->replacement->flags);
7495 smp_mb(); /* Make sure other CPUs may see both as identical
7496 * but will never see neither - if they are careful
7498 p->replacement = NULL;
7499 clear_bit(WantReplacement, &rdev->flags);
7501 /* We might have just removed the Replacement as faulty-
7502 * clear the bit just in case
7504 clear_bit(WantReplacement, &rdev->flags);
7507 print_raid5_conf(conf);
7511 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7513 struct r5conf *conf = mddev->private;
7516 struct disk_info *p;
7518 int last = conf->raid_disks - 1;
7520 if (test_bit(Journal, &rdev->flags)) {
7521 char b[BDEVNAME_SIZE];
7525 rdev->raid_disk = 0;
7527 * The array is in readonly mode if journal is missing, so no
7528 * write requests running. We should be safe
7530 r5l_init_log(conf, rdev);
7531 pr_debug("md/raid:%s: using device %s as journal\n",
7532 mdname(mddev), bdevname(rdev->bdev, b));
7535 if (mddev->recovery_disabled == conf->recovery_disabled)
7538 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7539 /* no point adding a device */
7542 if (rdev->raid_disk >= 0)
7543 first = last = rdev->raid_disk;
7546 * find the disk ... but prefer rdev->saved_raid_disk
7549 if (rdev->saved_raid_disk >= 0 &&
7550 rdev->saved_raid_disk >= first &&
7551 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7552 first = rdev->saved_raid_disk;
7554 for (disk = first; disk <= last; disk++) {
7555 p = conf->disks + disk;
7556 if (p->rdev == NULL) {
7557 clear_bit(In_sync, &rdev->flags);
7558 rdev->raid_disk = disk;
7560 if (rdev->saved_raid_disk != disk)
7562 rcu_assign_pointer(p->rdev, rdev);
7566 for (disk = first; disk <= last; disk++) {
7567 p = conf->disks + disk;
7568 if (test_bit(WantReplacement, &p->rdev->flags) &&
7569 p->replacement == NULL) {
7570 clear_bit(In_sync, &rdev->flags);
7571 set_bit(Replacement, &rdev->flags);
7572 rdev->raid_disk = disk;
7575 rcu_assign_pointer(p->replacement, rdev);
7580 print_raid5_conf(conf);
7584 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7586 /* no resync is happening, and there is enough space
7587 * on all devices, so we can resize.
7588 * We need to make sure resync covers any new space.
7589 * If the array is shrinking we should possibly wait until
7590 * any io in the removed space completes, but it hardly seems
7594 struct r5conf *conf = mddev->private;
7598 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7599 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7600 if (mddev->external_size &&
7601 mddev->array_sectors > newsize)
7603 if (mddev->bitmap) {
7604 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7608 md_set_array_sectors(mddev, newsize);
7609 if (sectors > mddev->dev_sectors &&
7610 mddev->recovery_cp > mddev->dev_sectors) {
7611 mddev->recovery_cp = mddev->dev_sectors;
7612 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7614 mddev->dev_sectors = sectors;
7615 mddev->resync_max_sectors = sectors;
7619 static int check_stripe_cache(struct mddev *mddev)
7621 /* Can only proceed if there are plenty of stripe_heads.
7622 * We need a minimum of one full stripe,, and for sensible progress
7623 * it is best to have about 4 times that.
7624 * If we require 4 times, then the default 256 4K stripe_heads will
7625 * allow for chunk sizes up to 256K, which is probably OK.
7626 * If the chunk size is greater, user-space should request more
7627 * stripe_heads first.
7629 struct r5conf *conf = mddev->private;
7630 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7631 > conf->min_nr_stripes ||
7632 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7633 > conf->min_nr_stripes) {
7634 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7636 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7643 static int check_reshape(struct mddev *mddev)
7645 struct r5conf *conf = mddev->private;
7649 if (mddev->delta_disks == 0 &&
7650 mddev->new_layout == mddev->layout &&
7651 mddev->new_chunk_sectors == mddev->chunk_sectors)
7652 return 0; /* nothing to do */
7653 if (has_failed(conf))
7655 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7656 /* We might be able to shrink, but the devices must
7657 * be made bigger first.
7658 * For raid6, 4 is the minimum size.
7659 * Otherwise 2 is the minimum
7662 if (mddev->level == 6)
7664 if (mddev->raid_disks + mddev->delta_disks < min)
7668 if (!check_stripe_cache(mddev))
7671 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7672 mddev->delta_disks > 0)
7673 if (resize_chunks(conf,
7674 conf->previous_raid_disks
7675 + max(0, mddev->delta_disks),
7676 max(mddev->new_chunk_sectors,
7677 mddev->chunk_sectors)
7680 return resize_stripes(conf, (conf->previous_raid_disks
7681 + mddev->delta_disks));
7684 static int raid5_start_reshape(struct mddev *mddev)
7686 struct r5conf *conf = mddev->private;
7687 struct md_rdev *rdev;
7689 unsigned long flags;
7691 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7694 if (!check_stripe_cache(mddev))
7697 if (has_failed(conf))
7700 rdev_for_each(rdev, mddev) {
7701 if (!test_bit(In_sync, &rdev->flags)
7702 && !test_bit(Faulty, &rdev->flags))
7706 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7707 /* Not enough devices even to make a degraded array
7712 /* Refuse to reduce size of the array. Any reductions in
7713 * array size must be through explicit setting of array_size
7716 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7717 < mddev->array_sectors) {
7718 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7723 atomic_set(&conf->reshape_stripes, 0);
7724 spin_lock_irq(&conf->device_lock);
7725 write_seqcount_begin(&conf->gen_lock);
7726 conf->previous_raid_disks = conf->raid_disks;
7727 conf->raid_disks += mddev->delta_disks;
7728 conf->prev_chunk_sectors = conf->chunk_sectors;
7729 conf->chunk_sectors = mddev->new_chunk_sectors;
7730 conf->prev_algo = conf->algorithm;
7731 conf->algorithm = mddev->new_layout;
7733 /* Code that selects data_offset needs to see the generation update
7734 * if reshape_progress has been set - so a memory barrier needed.
7737 if (mddev->reshape_backwards)
7738 conf->reshape_progress = raid5_size(mddev, 0, 0);
7740 conf->reshape_progress = 0;
7741 conf->reshape_safe = conf->reshape_progress;
7742 write_seqcount_end(&conf->gen_lock);
7743 spin_unlock_irq(&conf->device_lock);
7745 /* Now make sure any requests that proceeded on the assumption
7746 * the reshape wasn't running - like Discard or Read - have
7749 mddev_suspend(mddev);
7750 mddev_resume(mddev);
7752 /* Add some new drives, as many as will fit.
7753 * We know there are enough to make the newly sized array work.
7754 * Don't add devices if we are reducing the number of
7755 * devices in the array. This is because it is not possible
7756 * to correctly record the "partially reconstructed" state of
7757 * such devices during the reshape and confusion could result.
7759 if (mddev->delta_disks >= 0) {
7760 rdev_for_each(rdev, mddev)
7761 if (rdev->raid_disk < 0 &&
7762 !test_bit(Faulty, &rdev->flags)) {
7763 if (raid5_add_disk(mddev, rdev) == 0) {
7765 >= conf->previous_raid_disks)
7766 set_bit(In_sync, &rdev->flags);
7768 rdev->recovery_offset = 0;
7770 if (sysfs_link_rdev(mddev, rdev))
7771 /* Failure here is OK */;
7773 } else if (rdev->raid_disk >= conf->previous_raid_disks
7774 && !test_bit(Faulty, &rdev->flags)) {
7775 /* This is a spare that was manually added */
7776 set_bit(In_sync, &rdev->flags);
7779 /* When a reshape changes the number of devices,
7780 * ->degraded is measured against the larger of the
7781 * pre and post number of devices.
7783 spin_lock_irqsave(&conf->device_lock, flags);
7784 mddev->degraded = raid5_calc_degraded(conf);
7785 spin_unlock_irqrestore(&conf->device_lock, flags);
7787 mddev->raid_disks = conf->raid_disks;
7788 mddev->reshape_position = conf->reshape_progress;
7789 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7791 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7792 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7793 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7794 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7795 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7796 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7798 if (!mddev->sync_thread) {
7799 mddev->recovery = 0;
7800 spin_lock_irq(&conf->device_lock);
7801 write_seqcount_begin(&conf->gen_lock);
7802 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7803 mddev->new_chunk_sectors =
7804 conf->chunk_sectors = conf->prev_chunk_sectors;
7805 mddev->new_layout = conf->algorithm = conf->prev_algo;
7806 rdev_for_each(rdev, mddev)
7807 rdev->new_data_offset = rdev->data_offset;
7809 conf->generation --;
7810 conf->reshape_progress = MaxSector;
7811 mddev->reshape_position = MaxSector;
7812 write_seqcount_end(&conf->gen_lock);
7813 spin_unlock_irq(&conf->device_lock);
7816 conf->reshape_checkpoint = jiffies;
7817 md_wakeup_thread(mddev->sync_thread);
7818 md_new_event(mddev);
7822 /* This is called from the reshape thread and should make any
7823 * changes needed in 'conf'
7825 static void end_reshape(struct r5conf *conf)
7828 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7829 struct md_rdev *rdev;
7831 spin_lock_irq(&conf->device_lock);
7832 conf->previous_raid_disks = conf->raid_disks;
7833 rdev_for_each(rdev, conf->mddev)
7834 rdev->data_offset = rdev->new_data_offset;
7836 conf->reshape_progress = MaxSector;
7837 conf->mddev->reshape_position = MaxSector;
7838 spin_unlock_irq(&conf->device_lock);
7839 wake_up(&conf->wait_for_overlap);
7841 /* read-ahead size must cover two whole stripes, which is
7842 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7844 if (conf->mddev->queue) {
7845 int data_disks = conf->raid_disks - conf->max_degraded;
7846 int stripe = data_disks * ((conf->chunk_sectors << 9)
7848 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7849 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7854 /* This is called from the raid5d thread with mddev_lock held.
7855 * It makes config changes to the device.
7857 static void raid5_finish_reshape(struct mddev *mddev)
7859 struct r5conf *conf = mddev->private;
7861 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7863 if (mddev->delta_disks > 0) {
7864 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7866 set_capacity(mddev->gendisk, mddev->array_sectors);
7867 revalidate_disk(mddev->gendisk);
7871 spin_lock_irq(&conf->device_lock);
7872 mddev->degraded = raid5_calc_degraded(conf);
7873 spin_unlock_irq(&conf->device_lock);
7874 for (d = conf->raid_disks ;
7875 d < conf->raid_disks - mddev->delta_disks;
7877 struct md_rdev *rdev = conf->disks[d].rdev;
7879 clear_bit(In_sync, &rdev->flags);
7880 rdev = conf->disks[d].replacement;
7882 clear_bit(In_sync, &rdev->flags);
7885 mddev->layout = conf->algorithm;
7886 mddev->chunk_sectors = conf->chunk_sectors;
7887 mddev->reshape_position = MaxSector;
7888 mddev->delta_disks = 0;
7889 mddev->reshape_backwards = 0;
7893 static void raid5_quiesce(struct mddev *mddev, int state)
7895 struct r5conf *conf = mddev->private;
7898 case 2: /* resume for a suspend */
7899 wake_up(&conf->wait_for_overlap);
7902 case 1: /* stop all writes */
7903 lock_all_device_hash_locks_irq(conf);
7904 /* '2' tells resync/reshape to pause so that all
7905 * active stripes can drain
7907 r5c_flush_cache(conf, INT_MAX);
7909 wait_event_cmd(conf->wait_for_quiescent,
7910 atomic_read(&conf->active_stripes) == 0 &&
7911 atomic_read(&conf->active_aligned_reads) == 0,
7912 unlock_all_device_hash_locks_irq(conf),
7913 lock_all_device_hash_locks_irq(conf));
7915 unlock_all_device_hash_locks_irq(conf);
7916 /* allow reshape to continue */
7917 wake_up(&conf->wait_for_overlap);
7920 case 0: /* re-enable writes */
7921 lock_all_device_hash_locks_irq(conf);
7923 wake_up(&conf->wait_for_quiescent);
7924 wake_up(&conf->wait_for_overlap);
7925 unlock_all_device_hash_locks_irq(conf);
7928 r5l_quiesce(conf->log, state);
7931 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7933 struct r0conf *raid0_conf = mddev->private;
7936 /* for raid0 takeover only one zone is supported */
7937 if (raid0_conf->nr_strip_zones > 1) {
7938 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7940 return ERR_PTR(-EINVAL);
7943 sectors = raid0_conf->strip_zone[0].zone_end;
7944 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7945 mddev->dev_sectors = sectors;
7946 mddev->new_level = level;
7947 mddev->new_layout = ALGORITHM_PARITY_N;
7948 mddev->new_chunk_sectors = mddev->chunk_sectors;
7949 mddev->raid_disks += 1;
7950 mddev->delta_disks = 1;
7951 /* make sure it will be not marked as dirty */
7952 mddev->recovery_cp = MaxSector;
7954 return setup_conf(mddev);
7957 static void *raid5_takeover_raid1(struct mddev *mddev)
7962 if (mddev->raid_disks != 2 ||
7963 mddev->degraded > 1)
7964 return ERR_PTR(-EINVAL);
7966 /* Should check if there are write-behind devices? */
7968 chunksect = 64*2; /* 64K by default */
7970 /* The array must be an exact multiple of chunksize */
7971 while (chunksect && (mddev->array_sectors & (chunksect-1)))
7974 if ((chunksect<<9) < STRIPE_SIZE)
7975 /* array size does not allow a suitable chunk size */
7976 return ERR_PTR(-EINVAL);
7978 mddev->new_level = 5;
7979 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7980 mddev->new_chunk_sectors = chunksect;
7982 ret = setup_conf(mddev);
7984 mddev_clear_unsupported_flags(mddev,
7985 UNSUPPORTED_MDDEV_FLAGS);
7989 static void *raid5_takeover_raid6(struct mddev *mddev)
7993 switch (mddev->layout) {
7994 case ALGORITHM_LEFT_ASYMMETRIC_6:
7995 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7997 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7998 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8000 case ALGORITHM_LEFT_SYMMETRIC_6:
8001 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8003 case ALGORITHM_RIGHT_SYMMETRIC_6:
8004 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8006 case ALGORITHM_PARITY_0_6:
8007 new_layout = ALGORITHM_PARITY_0;
8009 case ALGORITHM_PARITY_N:
8010 new_layout = ALGORITHM_PARITY_N;
8013 return ERR_PTR(-EINVAL);
8015 mddev->new_level = 5;
8016 mddev->new_layout = new_layout;
8017 mddev->delta_disks = -1;
8018 mddev->raid_disks -= 1;
8019 return setup_conf(mddev);
8022 static int raid5_check_reshape(struct mddev *mddev)
8024 /* For a 2-drive array, the layout and chunk size can be changed
8025 * immediately as not restriping is needed.
8026 * For larger arrays we record the new value - after validation
8027 * to be used by a reshape pass.
8029 struct r5conf *conf = mddev->private;
8030 int new_chunk = mddev->new_chunk_sectors;
8032 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8034 if (new_chunk > 0) {
8035 if (!is_power_of_2(new_chunk))
8037 if (new_chunk < (PAGE_SIZE>>9))
8039 if (mddev->array_sectors & (new_chunk-1))
8040 /* not factor of array size */
8044 /* They look valid */
8046 if (mddev->raid_disks == 2) {
8047 /* can make the change immediately */
8048 if (mddev->new_layout >= 0) {
8049 conf->algorithm = mddev->new_layout;
8050 mddev->layout = mddev->new_layout;
8052 if (new_chunk > 0) {
8053 conf->chunk_sectors = new_chunk ;
8054 mddev->chunk_sectors = new_chunk;
8056 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8057 md_wakeup_thread(mddev->thread);
8059 return check_reshape(mddev);
8062 static int raid6_check_reshape(struct mddev *mddev)
8064 int new_chunk = mddev->new_chunk_sectors;
8066 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8068 if (new_chunk > 0) {
8069 if (!is_power_of_2(new_chunk))
8071 if (new_chunk < (PAGE_SIZE >> 9))
8073 if (mddev->array_sectors & (new_chunk-1))
8074 /* not factor of array size */
8078 /* They look valid */
8079 return check_reshape(mddev);
8082 static void *raid5_takeover(struct mddev *mddev)
8084 /* raid5 can take over:
8085 * raid0 - if there is only one strip zone - make it a raid4 layout
8086 * raid1 - if there are two drives. We need to know the chunk size
8087 * raid4 - trivial - just use a raid4 layout.
8088 * raid6 - Providing it is a *_6 layout
8090 if (mddev->level == 0)
8091 return raid45_takeover_raid0(mddev, 5);
8092 if (mddev->level == 1)
8093 return raid5_takeover_raid1(mddev);
8094 if (mddev->level == 4) {
8095 mddev->new_layout = ALGORITHM_PARITY_N;
8096 mddev->new_level = 5;
8097 return setup_conf(mddev);
8099 if (mddev->level == 6)
8100 return raid5_takeover_raid6(mddev);
8102 return ERR_PTR(-EINVAL);
8105 static void *raid4_takeover(struct mddev *mddev)
8107 /* raid4 can take over:
8108 * raid0 - if there is only one strip zone
8109 * raid5 - if layout is right
8111 if (mddev->level == 0)
8112 return raid45_takeover_raid0(mddev, 4);
8113 if (mddev->level == 5 &&
8114 mddev->layout == ALGORITHM_PARITY_N) {
8115 mddev->new_layout = 0;
8116 mddev->new_level = 4;
8117 return setup_conf(mddev);
8119 return ERR_PTR(-EINVAL);
8122 static struct md_personality raid5_personality;
8124 static void *raid6_takeover(struct mddev *mddev)
8126 /* Currently can only take over a raid5. We map the
8127 * personality to an equivalent raid6 personality
8128 * with the Q block at the end.
8132 if (mddev->pers != &raid5_personality)
8133 return ERR_PTR(-EINVAL);
8134 if (mddev->degraded > 1)
8135 return ERR_PTR(-EINVAL);
8136 if (mddev->raid_disks > 253)
8137 return ERR_PTR(-EINVAL);
8138 if (mddev->raid_disks < 3)
8139 return ERR_PTR(-EINVAL);
8141 switch (mddev->layout) {
8142 case ALGORITHM_LEFT_ASYMMETRIC:
8143 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8145 case ALGORITHM_RIGHT_ASYMMETRIC:
8146 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8148 case ALGORITHM_LEFT_SYMMETRIC:
8149 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8151 case ALGORITHM_RIGHT_SYMMETRIC:
8152 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8154 case ALGORITHM_PARITY_0:
8155 new_layout = ALGORITHM_PARITY_0_6;
8157 case ALGORITHM_PARITY_N:
8158 new_layout = ALGORITHM_PARITY_N;
8161 return ERR_PTR(-EINVAL);
8163 mddev->new_level = 6;
8164 mddev->new_layout = new_layout;
8165 mddev->delta_disks = 1;
8166 mddev->raid_disks += 1;
8167 return setup_conf(mddev);
8170 static struct md_personality raid6_personality =
8174 .owner = THIS_MODULE,
8175 .make_request = raid5_make_request,
8178 .status = raid5_status,
8179 .error_handler = raid5_error,
8180 .hot_add_disk = raid5_add_disk,
8181 .hot_remove_disk= raid5_remove_disk,
8182 .spare_active = raid5_spare_active,
8183 .sync_request = raid5_sync_request,
8184 .resize = raid5_resize,
8186 .check_reshape = raid6_check_reshape,
8187 .start_reshape = raid5_start_reshape,
8188 .finish_reshape = raid5_finish_reshape,
8189 .quiesce = raid5_quiesce,
8190 .takeover = raid6_takeover,
8191 .congested = raid5_congested,
8193 static struct md_personality raid5_personality =
8197 .owner = THIS_MODULE,
8198 .make_request = raid5_make_request,
8201 .status = raid5_status,
8202 .error_handler = raid5_error,
8203 .hot_add_disk = raid5_add_disk,
8204 .hot_remove_disk= raid5_remove_disk,
8205 .spare_active = raid5_spare_active,
8206 .sync_request = raid5_sync_request,
8207 .resize = raid5_resize,
8209 .check_reshape = raid5_check_reshape,
8210 .start_reshape = raid5_start_reshape,
8211 .finish_reshape = raid5_finish_reshape,
8212 .quiesce = raid5_quiesce,
8213 .takeover = raid5_takeover,
8214 .congested = raid5_congested,
8217 static struct md_personality raid4_personality =
8221 .owner = THIS_MODULE,
8222 .make_request = raid5_make_request,
8225 .status = raid5_status,
8226 .error_handler = raid5_error,
8227 .hot_add_disk = raid5_add_disk,
8228 .hot_remove_disk= raid5_remove_disk,
8229 .spare_active = raid5_spare_active,
8230 .sync_request = raid5_sync_request,
8231 .resize = raid5_resize,
8233 .check_reshape = raid5_check_reshape,
8234 .start_reshape = raid5_start_reshape,
8235 .finish_reshape = raid5_finish_reshape,
8236 .quiesce = raid5_quiesce,
8237 .takeover = raid4_takeover,
8238 .congested = raid5_congested,
8241 static int __init raid5_init(void)
8245 raid5_wq = alloc_workqueue("raid5wq",
8246 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8250 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8252 raid456_cpu_up_prepare,
8255 destroy_workqueue(raid5_wq);
8258 register_md_personality(&raid6_personality);
8259 register_md_personality(&raid5_personality);
8260 register_md_personality(&raid4_personality);
8264 static void raid5_exit(void)
8266 unregister_md_personality(&raid6_personality);
8267 unregister_md_personality(&raid5_personality);
8268 unregister_md_personality(&raid4_personality);
8269 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8270 destroy_workqueue(raid5_wq);
8273 module_init(raid5_init);
8274 module_exit(raid5_exit);
8275 MODULE_LICENSE("GPL");
8276 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8277 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8278 MODULE_ALIAS("md-raid5");
8279 MODULE_ALIAS("md-raid4");
8280 MODULE_ALIAS("md-level-5");
8281 MODULE_ALIAS("md-level-4");
8282 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8283 MODULE_ALIAS("md-raid6");
8284 MODULE_ALIAS("md-level-6");
8286 /* This used to be two separate modules, they were: */
8287 MODULE_ALIAS("raid5");
8288 MODULE_ALIAS("raid6");